Function (File:Line)	Call count	Line coverage	Branch coverage	Block coverage
(anonymous namespace)::print_usage(char*) (benchmark/main.cpp:28)	not called	0.0%	0.0%	0.0%
_ZZN3kai4testL60gtest_MatMulMatMulTest_f16_qai8dxp_qsi4cxp_EvalGenerateName_ERKN7testing13TestParamInfoINSt3__15tupleIJmNS0_11MatMulShapeENS0_13MatrixPortionEbEEEEEENK3$_1clIS8_EEDaRKT_ (test/tests/matmul_clamp_f16_qai8dxp_qsi4cxp_test.cpp:187)	called 896 times	100.0%	30.6%	50.0%
_ZZN3kai4testL60gtest_MatMulMatMulTest_f16_qai8dxp_qsi8cxp_EvalGenerateName_ERKN7testing13TestParamInfoINSt3__15tupleIJmNS0_11MatMulShapeENS0_13MatrixPortionEbEEEEEENK3$_1clIS8_EEDaRKT_ (test/tests/matmul_clamp_f16_qai8dxp_qsi8cxp_test.cpp:182)	called 896 times	100.0%	30.6%	50.0%
_ZZN3kai4testL60gtest_MatMulMatMulTest_f32_qai8dxp_qsi4cxp_EvalGenerateName_ERKN7testing13TestParamInfoINSt3__15tupleIJmNS0_11MatMulShapeENS0_13MatrixPortionEEEEEEENK3$_1clIS8_EEDaRKT_ (test/tests/matmul_clamp_f32_qai8dxp_qsi4cxp_test.cpp:794)	called 4800 times	100.0%	32.6%	50.0%
_ZZN3kai4testL60gtest_MatMulMatMulTest_f32_qai8dxp_qsi8cxp_EvalGenerateName_ERKN7testing13TestParamInfoINSt3__15tupleIJmNS0_11MatMulShapeENS0_13MatrixPortionEEEEEEENK3$_1clIS8_EEDaRKT_ (test/tests/matmul_clamp_f32_qai8dxp_qsi8cxp_test.cpp:378)	called 2464 times	100.0%	31.0%	50.0%
_ZZN3kai4testL61gtest_MatMulMatMulTest_bf16_qai8dxp_qsi4cxp_EvalGenerateName_ERKN7testing13TestParamInfoINSt3__15tupleIJmNS0_11MatMulShapeENS0_13MatrixPortionEbEEEEEENK3$_2clIS8_EENS3_12basic_stringIcNS3_11char_traitsIcEENS3_9allocatorIcEEEERKT_ (test/tests/matmul_clamp_bf16_qai8dxp_qsi4cxp_test.cpp:328)	called 784 times	100.0%	30.0%	100.0%
_ZZN3kai4testL62gtest_MatMulMatMulTest_f16_qsi8d32p_qai4c32p_EvalGenerateName_ERKN7testing13TestParamInfoINSt3__15tupleIJmNS0_11MatMulShapeEmNS0_13MatrixPortionEbEEEEEENK3$_1clIS8_EEDaRKT_ (test/tests/matmul_clamp_f16_qsi8d32p_qai4c32p_test.cpp:384)	called 11200 times	100.0%	48.2%	84.0%
_ZZN3kai4testL62gtest_MatMulMatMulTest_f32_qsi8d32p_qai4c32p_EvalGenerateName_ERKN7testing13TestParamInfoINSt3__15tupleIJmNS0_11MatMulShapeEmNS0_13MatrixPortionEbEEEEEENK3$_1clIS8_EEDaRKT_ (test/tests/matmul_clamp_f32_qsi8d32p_qai4c32p_test.cpp:330)	called 11200 times	100.0%	38.5%	84.0%
_ZZN3kai4testL62gtest_MatMulMatMulTest_f32_qsi8d32p_qsi4c32p_EvalGenerateName_ERKN7testing13TestParamInfoINSt3__15tupleIJmNS0_11MatMulShapeENS0_13MatrixPortionEEEEEEENK3$_1clIS8_EEDaRKT_ (test/tests/matmul_clamp_f32_qsi8d32p_qsi4c32p_test.cpp:281)	called 576 times	100.0%	32.6%	50.0%
_ZZN3kai4testL73gtest_matmul_clamp_qai8_qai8_qsi8cxpMatMulQuantizedTest_EvalGenerateName_ERKN7testing13TestParamInfoINSt3__15tupleIJNS0_12_GLOBAL__N_113MatMulVariantENS0_11MatMulShapeENS0_13MatrixPortionEfEEEEEENK3$_2clISA_EENS3_12basic_stringIcNS3_11char_traitsIcEENS3_9allocatorIcEEEERKT_ (test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp:1055)	called 168 times	100.0%	31.2%	100.0%
_ZZN3kai4testL74gtest_matmul_clamp_qai8_qai8p_qsi8cxpMatMulQuantizedTest_EvalGenerateName_ERKN7testing13TestParamInfoINSt3__15tupleIJNS0_12_GLOBAL__N_113MatMulVariantENS0_11MatMulShapeENS0_13MatrixPortionEfEEEEEENK3$_2clISA_EENS3_12basic_stringIcNS3_11char_traitsIcEENS3_9allocatorIcEEEERKT_ (test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp:1034)	called 666 times	100.0%	31.8%	100.0%
_ZZN3kai4testL75gtest_MatMulMatMulTest_qmatmul_clamp_f32_qai8dxp_qsi4c32p_EvalGenerateName_ERKN7testing13TestParamInfoINSt3__15tupleIJmNS0_11MatMulShapeEmNS0_13MatrixPortionENS0_11RhsPackTypeEEEEEEENK3$_1clIS9_EEDaRKT_ (test/tests/matmul_clamp_qai8dxp_qsi4c32p_test.cpp:518)	called 1232 times	100.0%	45.7%	77.0%
_ZZN3kai4testL76gtest_MatMulMatMulTest_qmatmul_clamp_bf16_qai8dxp_qsi4c32p_EvalGenerateName_ERKN7testing13TestParamInfoINSt3__15tupleIJmNS0_11MatMulShapeEmNS0_13MatrixPortionENS0_11RhsPackTypeEEEEEEENK3$_1clIS9_EEDaRKT_ (test/tests/matmul_clamp_qai8dxp_qsi4c32p_test.cpp:540)	called 224 times	100.0%	45.7%	77.0%
_ZZN3kai4testL91gtest_indirect_matmul_clamp_qai8_qai8p_qsi8cxpIndirectMatMulQuantizedTest_EvalGenerateName_ERKN7testing13TestParamInfoINSt3__15tupleIJNS0_12_GLOBAL__N_121IndirectMatMulVariantENS0_11MatMulShapeENS0_13MatrixPortionEmfEEEEEENK3$_2clISA_EENS3_12basic_stringIcNS3_11char_traitsIcEENS3_9allocatorIcEEEERKT_ (test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp:1095)	called 9324 times	100.0%	30.0%	100.0%
bool kai::test::(anonymous namespace)::compare_per_row<float, std::nullptr_t, float>(void const*, void const*, kai::test::DataFormat const&, unsigned long, unsigned long, kai::test::Rect const&, kai::test::MismatchHandler&) (test/common/compare.cpp:127)	called 257 times	100.0%	-%	72.0%
bool kai::test::(anonymous namespace)::compare_per_row<kai::test::BFloat16<true>, std::nullptr_t, float>(void const*, void const*, kai::test::DataFormat const&, unsigned long, unsigned long, kai::test::Rect const&, kai::test::MismatchHandler&) (test/common/compare.cpp:127)	not called	100.0%	-%	0.0%
bool kai::test::(anonymous namespace)::compare_per_row<kai::test::Float16, std::nullptr_t, kai::test::Float16>(void const*, void const*, kai::test::DataFormat const&, unsigned long, unsigned long, kai::test::Rect const&, kai::test::MismatchHandler&) (test/common/compare.cpp:127)	called 211 times	100.0%	-%	72.0%
bool kai::test::(anonymous namespace)::compare_per_row<kai::test::Int4, float, int>(void const*, void const*, kai::test::DataFormat const&, unsigned long, unsigned long, kai::test::Rect const&, kai::test::MismatchHandler&) (test/common/compare.cpp:127)	not called	61.5%	25.8%	0.0%
bool kai::test::(anonymous namespace)::compare_per_row<signed char, float, int>(void const*, void const*, kai::test::DataFormat const&, unsigned long, unsigned long, kai::test::Rect const&, kai::test::MismatchHandler&) (test/common/compare.cpp:127)	not called	100.0%	-%	0.0%
bool kai::test::(anonymous namespace)::compare_raw<float>(void const*, void const*, kai::test::DataFormat const&, unsigned long, unsigned long, kai::test::Rect const&, kai::test::MismatchHandler&) (test/common/compare.cpp:51)	called 16744 times	100.0%	-%	73.0%
bool kai::test::(anonymous namespace)::compare_raw<kai::test::BFloat16<true>>(void const*, void const*, kai::test::DataFormat const&, unsigned long, unsigned long, kai::test::Rect const&, kai::test::MismatchHandler&) (test/common/compare.cpp:51)	called 1068 times	93.5%	51.1%	50.0%
bool kai::test::(anonymous namespace)::compare_raw<kai::test::Float16>(void const*, void const*, kai::test::DataFormat const&, unsigned long, unsigned long, kai::test::Rect const&, kai::test::MismatchHandler&) (test/common/compare.cpp:51)	called 15378 times	100.0%	-%	74.0%
convert_s1s0_s16s0 (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon.c:21)	called 1956 times	100.0%	100.0%	100.0%
float kai::test::(anonymous namespace)::scalar_binary_elementwise<(kai::test::(anonymous namespace)::BinaryElementwiseOperator)0, float>(float, float) (test/reference/binary_elementwise.cpp:42)	called 2892060 times	100.0%	-%	100.0%
float kai::test::(anonymous namespace)::scalar_binary_elementwise<(kai::test::(anonymous namespace)::BinaryElementwiseOperator)1, float>(float, float) (test/reference/binary_elementwise.cpp:42)	not called	100.0%	-%	0.0%
float kai::test::(anonymous namespace)::scalar_binary_elementwise<(kai::test::(anonymous namespace)::BinaryElementwiseOperator)2, float>(float, float) (test/reference/binary_elementwise.cpp:42)	called 210756 times	100.0%	-%	100.0%
float kai::test::(anonymous namespace)::scalar_binary_elementwise<(kai::test::(anonymous namespace)::BinaryElementwiseOperator)3, float>(float, float) (test/reference/binary_elementwise.cpp:42)	not called	100.0%	-%	0.0%
float kai::test::read_array<float>(void const*, unsigned long) (test/common/memory.hpp:39)	called 4482014863 times	100.0%	100.0%	100.0%
float kai::test::reduce_max<float>(void const*, unsigned long) (test/reference/reduce.cpp:148)	called 927 times	100.0%	100.0%	75.0%
float kai::test::reduce_min<float>(void const*, unsigned long) (test/reference/reduce.cpp:133)	called 927 times	100.0%	100.0%	75.0%
get_block_height (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_f32p2vlx1biasf32_f32_f32_sme.c:30)	called 206 times	100.0%	-%	100.0%
get_block_height (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_x16p2vlx2b_x16_x16_sme.c:30)	called 206 times	100.0%	-%	100.0%
int kai::test::(anonymous namespace)::scalar_binary_elementwise<(kai::test::(anonymous namespace)::BinaryElementwiseOperator)0, int>(int, int) (test/reference/binary_elementwise.cpp:42)	not called	100.0%	-%	0.0%
int kai::test::(anonymous namespace)::scalar_binary_elementwise<(kai::test::(anonymous namespace)::BinaryElementwiseOperator)1, int>(int, int) (test/reference/binary_elementwise.cpp:42)	called 105378 times	100.0%	-%	100.0%
int kai::test::(anonymous namespace)::scalar_binary_elementwise<(kai::test::(anonymous namespace)::BinaryElementwiseOperator)2, int>(int, int) (test/reference/binary_elementwise.cpp:42)	called 105378 times	100.0%	-%	100.0%
int kai::test::(anonymous namespace)::scalar_binary_elementwise<(kai::test::(anonymous namespace)::BinaryElementwiseOperator)3, int>(int, int) (test/reference/binary_elementwise.cpp:42)	not called	100.0%	-%	0.0%
int kai::test::(anonymous namespace)::scalar_reduce<(kai::test::ReductionOperator)0, int>(int, int) (test/reference/reduce.cpp:26)	not called	0.0%	-%	0.0%
int kai::test::quantize_symmetric<int>(float, float) (test/reference/quantize.cpp:60)	called 105378 times	100.0%	-%	75.0%
int kai::test::read_array<int>(void const*, unsigned long) (test/common/memory.hpp:39)	called 1909563957 times	100.0%	100.0%	100.0%
int kai::test::round_to_nearest_even<int>(float) (test/common/round.cpp:44)	called 36898147 times	100.0%	-%	100.0%
kai::benchmark::MatMulRunner<kai::benchmark::MatMulBaseInterface>::MatMulRunner(kai::benchmark::MatMulBaseInterface const&, kai::test::DataType) (benchmark/matmul/matmul_runner.hpp:35)	called 36 times	100.0%	-%	100.0%
kai::benchmark::MatMulRunner<kai::benchmark::MatMulBaseInterface>::run(void const*, void const*, void*) (benchmark/matmul/matmul_runner.hpp:89)	called 18 times	100.0%	-%	100.0%
kai::benchmark::MatMulRunner<kai::benchmark::MatMulBaseInterface>::set_bl(unsigned long) (benchmark/matmul/matmul_runner.hpp:57)	called 18 times	100.0%	-%	100.0%
kai::benchmark::MatMulRunner<kai::benchmark::MatMulBaseInterface>::set_mnk(unsigned long, unsigned long, unsigned long) (benchmark/matmul/matmul_runner.hpp:44)	called 18 times	100.0%	-%	100.0%
kai::benchmark::MatMulRunner<kai::benchmark::MatMulBlockwiseDynamicQuantGenericDstInterface>::MatMulRunner(kai::benchmark::MatMulBlockwiseDynamicQuantGenericDstInterface const&, kai::test::DataType) (benchmark/matmul/matmul_runner.hpp:35)	called 16 times	100.0%	-%	100.0%
kai::benchmark::MatMulRunner<kai::benchmark::MatMulBlockwiseDynamicQuantGenericDstInterface>::run(void const*, void const*, void*) (benchmark/matmul/matmul_runner.hpp:151)	called 8 times	100.0%	-%	100.0%
kai::benchmark::MatMulRunner<kai::benchmark::MatMulBlockwiseDynamicQuantGenericDstInterface>::set_bl(unsigned long) (benchmark/matmul/matmul_runner.hpp:57)	called 8 times	100.0%	-%	100.0%
kai::benchmark::MatMulRunner<kai::benchmark::MatMulBlockwiseDynamicQuantGenericDstInterface>::set_mnk(unsigned long, unsigned long, unsigned long) (benchmark/matmul/matmul_runner.hpp:44)	called 8 times	100.0%	-%	100.0%
kai::benchmark::MatMulRunner<kai::benchmark::MatMulBlockwiseDynamicQuantInterface>::MatMulRunner(kai::benchmark::MatMulBlockwiseDynamicQuantInterface const&, kai::test::DataType) (benchmark/matmul/matmul_runner.hpp:35)	called 48 times	100.0%	-%	100.0%
kai::benchmark::MatMulRunner<kai::benchmark::MatMulBlockwiseDynamicQuantInterface>::run(void const*, void const*, void*) (benchmark/matmul/matmul_runner.hpp:167)	called 24 times	100.0%	-%	100.0%
kai::benchmark::MatMulRunner<kai::benchmark::MatMulBlockwiseDynamicQuantInterface>::set_bl(unsigned long) (benchmark/matmul/matmul_runner.hpp:57)	called 24 times	100.0%	-%	100.0%
kai::benchmark::MatMulRunner<kai::benchmark::MatMulBlockwiseDynamicQuantInterface>::set_mnk(unsigned long, unsigned long, unsigned long) (benchmark/matmul/matmul_runner.hpp:44)	called 24 times	100.0%	-%	100.0%
kai::benchmark::MatMulRunner<kai::benchmark::MatMulFloatInterface>::MatMulRunner(kai::benchmark::MatMulFloatInterface const&, kai::test::DataType) (benchmark/matmul/matmul_runner.hpp:35)	called 38 times	100.0%	-%	100.0%
kai::benchmark::MatMulRunner<kai::benchmark::MatMulFloatInterface>::run(void const*, void const*, void*) (benchmark/matmul/matmul_runner.hpp:119)	called 19 times	100.0%	-%	100.0%
kai::benchmark::MatMulRunner<kai::benchmark::MatMulFloatInterface>::set_bl(unsigned long) (benchmark/matmul/matmul_runner.hpp:57)	called 19 times	100.0%	-%	100.0%
kai::benchmark::MatMulRunner<kai::benchmark::MatMulFloatInterface>::set_mnk(unsigned long, unsigned long, unsigned long) (benchmark/matmul/matmul_runner.hpp:44)	called 19 times	100.0%	-%	100.0%
kai::benchmark::MatMulRunner<kai::benchmark::MatMulStaticQuantInterface>::MatMulRunner(kai::benchmark::MatMulStaticQuantInterface const&, kai::test::DataType) (benchmark/matmul/matmul_runner.hpp:35)	called 6 times	100.0%	-%	100.0%
kai::benchmark::MatMulRunner<kai::benchmark::MatMulStaticQuantInterface>::run(void const*, void const*, void*) (benchmark/matmul/matmul_runner.hpp:134)	called 3 times	100.0%	-%	100.0%
kai::benchmark::MatMulRunner<kai::benchmark::MatMulStaticQuantInterface>::set_bl(unsigned long) (benchmark/matmul/matmul_runner.hpp:57)	called 3 times	100.0%	-%	100.0%
kai::benchmark::MatMulRunner<kai::benchmark::MatMulStaticQuantInterface>::set_mnk(unsigned long, unsigned long, unsigned long) (benchmark/matmul/matmul_runner.hpp:44)	called 3 times	100.0%	-%	100.0%
kai::benchmark::MatMulRunner<kai::benchmark::MatMulStridedLhsInterface>::MatMulRunner(kai::benchmark::MatMulStridedLhsInterface const&, kai::test::DataType) (benchmark/matmul/matmul_runner.hpp:35)	called 22 times	100.0%	-%	100.0%
kai::benchmark::MatMulRunner<kai::benchmark::MatMulStridedLhsInterface>::run(void const*, void const*, void*) (benchmark/matmul/matmul_runner.hpp:104)	called 11 times	100.0%	-%	100.0%
kai::benchmark::MatMulRunner<kai::benchmark::MatMulStridedLhsInterface>::set_bl(unsigned long) (benchmark/matmul/matmul_runner.hpp:57)	called 11 times	100.0%	-%	100.0%
kai::benchmark::MatMulRunner<kai::benchmark::MatMulStridedLhsInterface>::set_mnk(unsigned long, unsigned long, unsigned long) (benchmark/matmul/matmul_runner.hpp:44)	called 11 times	100.0%	-%	100.0%
kai::benchmark::RegisterMatMulBenchmarks(kai::test::MatMulShape const&, unsigned long) (benchmark/matmul/matmul_registry.cpp:939)	called 1 time	100.0%	56.2%	48.0%
kai::test::(anonymous namespace)::CpuInfo::CpuInfo() (test/common/cpu_info.cpp:206)	called 4 times	100.0%	-%	100.0%
kai::test::(anonymous namespace)::CpuInfo::current() (test/common/cpu_info.cpp:219)	called 76601 times	75.0%	66.7%	71.0%
kai::test::(anonymous namespace)::Depthwise::Depthwise(kai::test::(anonymous namespace)::Depthwise const&) (test/tests/dwconv_test.cpp:55)	called 604 times	100.0%	-%	100.0%
kai::test::(anonymous namespace)::Depthwise::~Depthwise() (test/tests/dwconv_test.cpp:55)	called 486 times	100.0%	50.0%	100.0%
kai::test::(anonymous namespace)::DepthwisePlanarKernel::DepthwisePlanarKernel(kai::test::(anonymous namespace)::DepthwisePlanarKernel const&) (test/tests/dwconv_test.cpp:34)	called 604 times	100.0%	-%	100.0%
kai::test::(anonymous namespace)::DepthwisePlanarKernel::~DepthwisePlanarKernel() (test/tests/dwconv_test.cpp:34)	called 486 times	100.0%	50.0%	100.0%
kai::test::(anonymous namespace)::IndirectMatMul::Format::Hash::operator()(kai::test::(anonymous namespace)::IndirectMatMul::Format const&) const (test/tests/imatmul_test.cpp:103)	called 29172 times	100.0%	-%	100.0%
kai::test::(anonymous namespace)::IndirectMatMul::IndirectMatMul(kai::test::(anonymous namespace)::IndirectMatMul const&) (test/tests/imatmul_test.cpp:91)	called 269296 times	100.0%	-%	100.0%
kai::test::(anonymous namespace)::IndirectMatMul::~IndirectMatMul() (test/tests/imatmul_test.cpp:91)	called 215448 times	100.0%	50.0%	100.0%
kai::test::(anonymous namespace)::IndirectMatMulVariant::IndirectMatMulVariant() (test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp:276)	called 4 times	100.0%	50.0%	100.0%
kai::test::(anonymous namespace)::IndirectMatMulVariant::IndirectMatMulVariant(kai::test::(anonymous namespace)::IndirectMatMulVariant const&) (test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp:276)	called 93248 times	100.0%	-%	100.0%
kai::test::(anonymous namespace)::IndirectMatMulVariant::~IndirectMatMulVariant() (test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp:276)	called 74604 times	100.0%	-%	100.0%
kai::test::(anonymous namespace)::LhsPackIndirectKernel::LhsPackIndirectKernel() (test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp:72)	called 4 times	100.0%	50.0%	100.0%
kai::test::(anonymous namespace)::LhsPackIndirectKernel::LhsPackIndirectKernel(kai::test::(anonymous namespace)::LhsPackIndirectKernel const&) (test/tests/imatmul_test.cpp:50)	called 269296 times	100.0%	-%	100.0%
kai::test::(anonymous namespace)::LhsPackIndirectKernel::LhsPackIndirectKernel(kai::test::(anonymous namespace)::LhsPackIndirectKernel const&) (test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp:72)	called 93248 times	100.0%	-%	100.0%
kai::test::(anonymous namespace)::LhsPackIndirectKernel::~LhsPackIndirectKernel() (test/tests/imatmul_test.cpp:50)	called 215448 times	100.0%	50.0%	100.0%
kai::test::(anonymous namespace)::LhsPackIndirectKernel::~LhsPackIndirectKernel() (test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp:72)	called 74604 times	100.0%	-%	100.0%
kai::test::(anonymous namespace)::LhsPackKernel::LhsPackKernel() (test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp:61)	called 2 times	100.0%	-%	100.0%
kai::test::(anonymous namespace)::LhsPackKernel::LhsPackKernel(kai::test::(anonymous namespace)::LhsPackKernel const&) (test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp:61)	called 6672 times	100.0%	-%	100.0%
kai::test::(anonymous namespace)::LhsPackKernel::operator=(kai::test::(anonymous namespace)::LhsPackKernel const&) (test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp:61)	not called	100.0%	50.0%	0.0%
kai::test::(anonymous namespace)::LhsPackKernel::~LhsPackKernel() (test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp:61)	called 5342 times	100.0%	-%	100.0%
kai::test::(anonymous namespace)::MatMulIndirectKernel::MatMulIndirectKernel() (test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp:126)	called 4 times	100.0%	50.0%	100.0%
kai::test::(anonymous namespace)::MatMulIndirectKernel::MatMulIndirectKernel(kai::test::(anonymous namespace)::MatMulIndirectKernel const&) (test/tests/imatmul_test.cpp:74)	called 269296 times	100.0%	-%	100.0%
kai::test::(anonymous namespace)::MatMulIndirectKernel::MatMulIndirectKernel(kai::test::(anonymous namespace)::MatMulIndirectKernel const&) (test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp:126)	called 93248 times	100.0%	-%	100.0%
kai::test::(anonymous namespace)::MatMulIndirectKernel::~MatMulIndirectKernel() (test/tests/imatmul_test.cpp:74)	called 215448 times	100.0%	50.0%	100.0%
kai::test::(anonymous namespace)::MatMulIndirectKernel::~MatMulIndirectKernel() (test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp:126)	called 74604 times	100.0%	-%	100.0%
kai::test::(anonymous namespace)::MatMulKernel::MatMulKernel() (test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp:109)	called 6 times	100.0%	50.0%	100.0%
kai::test::(anonymous namespace)::MatMulKernel::MatMulKernel(kai::test::(anonymous namespace)::MatMulKernel const&) (test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp:109)	called 8352 times	100.0%	-%	100.0%
kai::test::(anonymous namespace)::MatMulKernel::~MatMulKernel() (test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp:109)	called 6690 times	100.0%	-%	100.0%
kai::test::(anonymous namespace)::MatMulVariant::MatMulVariant() (test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp:264)	called 6 times	100.0%	50.0%	100.0%
kai::test::(anonymous namespace)::MatMulVariant::MatMulVariant(kai::test::(anonymous namespace)::MatMulVariant const&) (test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp:264)	called 8352 times	100.0%	-%	100.0%
kai::test::(anonymous namespace)::MatMulVariant::~MatMulVariant() (test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp:264)	called 6690 times	100.0%	-%	100.0%
kai::test::(anonymous namespace)::ReferenceGenerator::generate_reference(kai::test::(anonymous namespace)::TestDataId const&) (test/tests/imatmul_test.cpp:414)	called 2244 times	100.0%	58.2%	57.0%
kai::test::(anonymous namespace)::ReferenceGenerator::generate_reference(kai::test::(anonymous namespace)::TestDataId const&, kai::test::MatMulShape const&) (test/tests/dwconv_test.cpp:165)	called 60 times	100.0%	50.0%	56.0%
kai::test::(anonymous namespace)::ReferenceGenerator::get_seed() (test/tests/dwconv_test.cpp:158)	called 180 times	100.0%	-%	100.0%
kai::test::(anonymous namespace)::ReferenceGenerator::get_seed() (test/tests/imatmul_test.cpp:407)	called 6732 times	100.0%	-%	100.0%
kai::test::(anonymous namespace)::ReferenceGenerator::get_test_reference(kai::test::(anonymous namespace)::TestDataId const&) (test/tests/imatmul_test.cpp:396)	called 26928 times	100.0%	72.7%	76.0%
kai::test::(anonymous namespace)::ReferenceGenerator::get_test_reference(kai::test::(anonymous namespace)::TestDataId, kai::test::MatMulShape const&) (test/tests/dwconv_test.cpp:147)	called 60 times	83.3%	54.5%	69.0%
kai::test::(anonymous namespace)::RhsPackDepthwiseKernel::RhsPackDepthwiseKernel(kai::test::(anonymous namespace)::RhsPackDepthwiseKernel const&) (test/tests/dwconv_test.cpp:46)	called 604 times	100.0%	-%	100.0%
kai::test::(anonymous namespace)::RhsPackDepthwiseKernel::~RhsPackDepthwiseKernel() (test/tests/dwconv_test.cpp:46)	called 486 times	100.0%	-%	100.0%
kai::test::(anonymous namespace)::RhsPackIndirectKernel::RhsPackIndirectKernel() (test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp:96)	called 4 times	100.0%	50.0%	100.0%
kai::test::(anonymous namespace)::RhsPackIndirectKernel::RhsPackIndirectKernel(kai::test::(anonymous namespace)::RhsPackIndirectKernel const&) (test/tests/imatmul_test.cpp:61)	called 269296 times	100.0%	-%	100.0%
kai::test::(anonymous namespace)::RhsPackIndirectKernel::RhsPackIndirectKernel(kai::test::(anonymous namespace)::RhsPackIndirectKernel const&) (test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp:96)	called 93248 times	100.0%	-%	100.0%
kai::test::(anonymous namespace)::RhsPackIndirectKernel::~RhsPackIndirectKernel() (test/tests/imatmul_test.cpp:61)	called 215448 times	100.0%	50.0%	100.0%
kai::test::(anonymous namespace)::RhsPackIndirectKernel::~RhsPackIndirectKernel() (test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp:96)	called 74604 times	100.0%	-%	100.0%
kai::test::(anonymous namespace)::RhsPackKernel::RhsPackKernel() (test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp:82)	called 8 times	100.0%	50.0%	100.0%
kai::test::(anonymous namespace)::RhsPackKernel::RhsPackKernel(kai::test::(anonymous namespace)::RhsPackKernel const&) (test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp:82)	called 8352 times	100.0%	-%	100.0%
kai::test::(anonymous namespace)::RhsPackKernel::operator=(kai::test::(anonymous namespace)::RhsPackKernel const&) (test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp:82)	called 3 times	100.0%	-%	100.0%
kai::test::(anonymous namespace)::RhsPackKernel::~RhsPackKernel() (test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp:82)	called 6692 times	100.0%	-%	100.0%
kai::test::(anonymous namespace)::TestData::TestData() (test/tests/dwconv_test.cpp:135)	called 240 times	100.0%	-%	100.0%
kai::test::(anonymous namespace)::TestData::TestData() (test/tests/imatmul_test.cpp:380)	called 8976 times	100.0%	-%	100.0%
kai::test::(anonymous namespace)::TestData::operator=(kai::test::(anonymous namespace)::TestData&&) (test/tests/dwconv_test.cpp:135)	called 60 times	100.0%	-%	100.0%
kai::test::(anonymous namespace)::TestData::operator=(kai::test::(anonymous namespace)::TestData&&) (test/tests/imatmul_test.cpp:380)	called 2244 times	100.0%	-%	100.0%
kai::test::(anonymous namespace)::TestData::~TestData() (test/tests/dwconv_test.cpp:135)	called 240 times	100.0%	-%	100.0%
kai::test::(anonymous namespace)::TestData::~TestData() (test/tests/imatmul_test.cpp:380)	called 8976 times	100.0%	-%	100.0%
kai::test::(anonymous namespace)::TestDataId::Hash::operator()(kai::test::(anonymous namespace)::TestDataId const&) const (test/tests/dwconv_test.cpp:111)	called 120 times	100.0%	-%	100.0%
kai::test::(anonymous namespace)::TestDataId::Hash::operator()(kai::test::(anonymous namespace)::TestDataId const&) const (test/tests/imatmul_test.cpp:358)	called 29172 times	100.0%	-%	100.0%
kai::test::(anonymous namespace)::TestDataId::Hash::operator()(kai::test::(anonymous namespace)::TestDataId const&) const (test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp:487)	called 11085 times	100.0%	-%	100.0%
kai::test::(anonymous namespace)::TestReference::TestReference() (test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp:449)	called 1854 times	100.0%	-%	100.0%
kai::test::(anonymous namespace)::TestReference::~TestReference() (test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp:449)	not called	100.0%	-%	0.0%
kai::test::(anonymous namespace)::bits(kai::test::DataType) (test/common/data_type.cpp:34)	called 67361 times	100.0%	-%	100.0%
kai::test::(anonymous namespace)::compare_matmul_result(kai::test::MatMulShape const&, kai::test::Rect const&, kai::test::Buffer const&, kai::test::Buffer const&) (test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp:745)	called 10158 times	73.3%	37.1%	42.0%
kai::test::(anonymous namespace)::convert(unsigned char const*, kai::test::DataType, kai::test::DataType) (test/reference/pack.cpp:29)	called 918733 times	83.3%	66.7%	56.0%
kai::test::(anonymous namespace)::dwconv(kai::test::(anonymous namespace)::DepthwisePlanarKernel const&, kai::test::Rect const&, kai::test::MatMulShape const&, kai::test::MatMulShape const&, kai::test::Padding2D, kai::test::(anonymous namespace)::TestData const&, kai::test::Buffer const&, kai::test::Range<float>, kai::test::DataType) (test/tests/dwconv_test.cpp:204)	called 60 times	100.0%	62.5%	74.0%
kai::test::(anonymous namespace)::get_cap_support(kai::test::(anonymous namespace)::CpuFeatures) (test/common/cpu_info.cpp:120)	called 18 times	100.0%	100.0%	53.0%
kai::test::(anonymous namespace)::get_depthwise_methods() (test/tests/dwconv_test.cpp:80)	called 1 time	100.0%	50.0%	85.0%
kai::test::(anonymous namespace)::get_dwconv_clamp_f32_f32_f32p1vlx1b_3x3_s1_4xc_sme2_mla() (test/tests/dwconv_test.cpp:71)	called 1 time	100.0%	-%	100.0%
kai::test::(anonymous namespace)::get_gemm_methods() (test/tests/matmul_clamp_f32_bf16p_bf16p_test.cpp:54)	called 1 time	100.0%	50.0%	85.0%
kai::test::(anonymous namespace)::get_gemm_variants() (test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp:288)	called 1 time	98.0%	50.0%	80.0%
kai::test::(anonymous namespace)::get_gemv_methods() (test/tests/matmul_clamp_f32_bf16p_bf16p_test.cpp:238)	called 1 time	100.0%	50.0%	85.0%
kai::test::(anonymous namespace)::get_gemv_variants() (test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp:410)	called 1 time	100.0%	50.0%	80.0%
kai::test::(anonymous namespace)::get_gemv_variants()::$_0::operator()() const (test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp:427)	called 168 times	90.9%	-%	100.0%
kai::test::(anonymous namespace)::get_imatmul_clamp_f16_f16p2vlx2_f16p2vlx2_2vlx2vl_sme2_mopa() (test/tests/imatmul_test.cpp:137)	called 1 time	100.0%	-%	100.0%
kai::test::(anonymous namespace)::get_imatmul_clamp_f16_f16p2vlx2_f16p2vlx2b_2vlx2vl_sme_mopa() (test/tests/imatmul_test.cpp:149)	called 1 time	100.0%	-%	100.0%
kai::test::(anonymous namespace)::get_imatmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme2_mopa() (test/tests/imatmul_test.cpp:162)	called 1 time	100.0%	-%	100.0%
kai::test::(anonymous namespace)::get_imatmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme_mopa() (test/tests/imatmul_test.cpp:174)	called 1 time	100.0%	-%	100.0%
kai::test::(anonymous namespace)::get_imatmul_clamp_qai8_qai8p2vlx4_qsi8cxp2vlx4sb_2vlx2vl_sme_mopa_interface() (test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp:222)	called 1 time	100.0%	-%	100.0%
kai::test::(anonymous namespace)::get_imatmul_clamp_qai8_qai8p2vlx4_qsi8cxpsb2vlx4_2vlx2vl_sme2_mopa_interface() (test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp:204)	called 1 time	100.0%	-%	100.0%
kai::test::(anonymous namespace)::get_indirect_gemm_variants() (test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp:342)	called 1 time	98.3%	50.0%	80.0%
kai::test::(anonymous namespace)::get_indirect_matmul_methods() (test/tests/imatmul_test.cpp:187)	called 1 time	100.0%	50.0%	85.0%
kai::test::(anonymous namespace)::get_lhs_pack() (test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp:252)	called 2 times	100.0%	75.0%	100.0%
kai::test::(anonymous namespace)::get_matmul_clamp_qai8_qai8_qsi8cxp2vlx4sb_1x16vl_sme2_dot_interface() (test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp:185)	called 1 time	100.0%	-%	100.0%
kai::test::(anonymous namespace)::get_matmul_clamp_qai8_qai8p2vlx4_qsi8cxp2vlx4sb_2vlx2vl_sme_mopa_interface() (test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp:163)	called 1 time	100.0%	-%	100.0%
kai::test::(anonymous namespace)::get_matmul_clamp_qai8_qai8p2vlx4_qsi8cxpsb2vlx4_2vlx2vl_sme2_mopa_interface() (test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp:141)	called 1 time	100.0%	-%	100.0%
kai::test::(anonymous namespace)::get_matmul_methods() (test/tests/matmul_clamp_f16_bf16p_bf16p_test.cpp:42)	called 1 time	100.0%	50.0%	85.0%
kai::test::(anonymous namespace)::get_rhs_pack() (test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp:238)	called 3 times	100.0%	75.0%	100.0%
kai::test::(anonymous namespace)::get_test_reference(kai::test::(anonymous namespace)::TestDataId const&) (test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp:513)	called 10158 times	100.0%	53.5%	55.0%
kai::test::(anonymous namespace)::has_a(kai::test::DataType) (test/common/data_type.cpp:30)	not called	0.0%	-%	0.0%
kai::test::(anonymous namespace)::has_i(kai::test::DataType) (test/common/data_type.cpp:18)	called 2595 times	100.0%	-%	100.0%
kai::test::(anonymous namespace)::has_q(kai::test::DataType) (test/common/data_type.cpp:26)	not called	0.0%	-%	0.0%
kai::test::(anonymous namespace)::has_s(kai::test::DataType) (test/common/data_type.cpp:22)	not called	0.0%	-%	0.0%
kai::test::(anonymous namespace)::imatmul(kai::test::(anonymous namespace)::MatMulIndirectKernel const&, kai::test::Rect const&, kai::test::MatMulShape const&, kai::test::(anonymous namespace)::KChunk const&, kai::test::Buffer const&, kai::test::Buffer const&, kai::test::Range<float>, kai::test::DataType) (test/tests/imatmul_test.cpp:539)	called 22704 times	100.0%	50.0%	68.0%
kai::test::(anonymous namespace)::operator==(kai::test::(anonymous namespace)::IndirectMatMul::Format const&, kai::test::(anonymous namespace)::IndirectMatMul::Format const&) (test/tests/imatmul_test.cpp:113)	called 24684 times	100.0%	50.0%	100.0%
kai::test::(anonymous namespace)::operator==(kai::test::(anonymous namespace)::TestDataId const&, kai::test::(anonymous namespace)::TestDataId const&) (test/tests/dwconv_test.cpp:123)	not called	0.0%	0.0%	0.0%
kai::test::(anonymous namespace)::operator==(kai::test::(anonymous namespace)::TestDataId const&, kai::test::(anonymous namespace)::TestDataId const&) (test/tests/imatmul_test.cpp:369)	called 31212 times	100.0%	62.5%	100.0%
kai::test::(anonymous namespace)::operator==(kai::test::(anonymous namespace)::TestDataId const&, kai::test::(anonymous namespace)::TestDataId const&) (test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp:498)	called 13342 times	100.0%	75.0%	100.0%
kai::test::(anonymous namespace)::pack_bias_per_row(kai::test::DataType, kai::test::DataType, kai::test::DataType, unsigned long, unsigned long, unsigned long, void const*, void const*, unsigned long, unsigned long, unsigned long, unsigned long, unsigned long, unsigned long) (test/reference/pack.cpp:100)	called 125 times	73.6%	53.6%	57.0%
kai::test::(anonymous namespace)::pack_block(void const*, kai::test::DataType, kai::test::DataType, unsigned long, unsigned long, unsigned long, unsigned long, unsigned long, unsigned long, unsigned long, unsigned long) (test/reference/pack.cpp:42)	called 116 times	100.0%	88.5%	80.0%
kai::test::(anonymous namespace)::pack_lhs(kai::test::(anonymous namespace)::LhsPackIndirectKernel const&, kai::test::Rect const&, kai::test::(anonymous namespace)::TestData const&, unsigned long, kai::test::(anonymous namespace)::KChunk const&) (test/tests/imatmul_test.cpp:489)	called 22704 times	100.0%	50.0%	64.0%
kai::test::(anonymous namespace)::pack_rhs(kai::test::(anonymous namespace)::RhsPackDepthwiseKernel const&, kai::test::MatMulShape const&, kai::test::(anonymous namespace)::TestData const&) (test/tests/dwconv_test.cpp:193)	called 60 times	100.0%	50.0%	66.0%
kai::test::(anonymous namespace)::pack_rhs(kai::test::(anonymous namespace)::RhsPackIndirectKernel const&, kai::test::Rect const&, kai::test::(anonymous namespace)::TestData const&, unsigned long, kai::test::(anonymous namespace)::KChunk const&, kai::test::DataType) (test/tests/imatmul_test.cpp:513)	called 22704 times	100.0%	50.0%	65.0%
kai::test::(anonymous namespace)::print_data(std::__1::basic_ostream<char, std::__1::char_traits<char>>&, unsigned char const*, unsigned long, kai::test::DataType) (test/common/printer.cpp:25)	not called	0.0%	0.0%	0.0%
kai::test::(anonymous namespace)::print_matrix_per_row(std::__1::basic_ostream<char, std::__1::char_traits<char>>&, unsigned char const*, kai::test::DataFormat const&, unsigned long, unsigned long) (test/common/printer.cpp:126)	not called	0.0%	0.0%	0.0%
kai::test::(anonymous namespace)::print_matrix_raw(std::__1::basic_ostream<char, std::__1::char_traits<char>>&, unsigned char const*, kai::test::DataFormat const&, unsigned long, unsigned long) (test/common/printer.cpp:68)	not called	0.0%	0.0%	0.0%
kai::test::(anonymous namespace)::test_lhs_pack(kai::test::MatMulShape const&, kai::test::(anonymous namespace)::MatMulVariant const&, kai::test::Rect const&, kai::test::(anonymous namespace)::TestReference const&) (test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp:666)	called 666 times	100.0%	38.6%	47.0%
kai::test::(anonymous namespace)::test_matmul(kai::test::MatMulShape const&, kai::test::(anonymous namespace)::MatMulVariant const&, kai::test::Rect const&, kai::test::(anonymous namespace)::TestReference const&) (test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp:779)	called 834 times	100.0%	16.7%	43.0%
kai::test::(anonymous namespace)::test_matmul(kai::test::MatMulShape const&, kai::test::(anonymous namespace)::MatMulVariant const&, kai::test::Rect const&, kai::test::(anonymous namespace)::TestReference const&)::$_0::operator()() const (test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp:785)	called 834 times	100.0%	42.5%	100.0%
kai::test::(anonymous namespace)::test_rhs_pack(kai::test::MatMulShape const&, kai::test::(anonymous namespace)::MatMulVariant const&, kai::test::Rect const&, kai::test::(anonymous namespace)::TestReference const&) (test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp:702)	called 834 times	87.9%	42.4%	48.0%
kai::test::BFloat16<false> kai::test::read_array<kai::test::BFloat16<false>>(void const*, unsigned long) (test/common/memory.hpp:39)	called 1394600792 times	100.0%	-%	100.0%
kai::test::BFloat16<false>::BFloat16(float) (test/common/bfloat16.hpp:39)	called 2798529236 times	100.0%	-%	100.0%
kai::test::BFloat16<false>::float_to_bfloat16_round_towards_zero(float) (test/common/bfloat16.hpp:90)	called 1399264618 times	100.0%	-%	100.0%
kai::test::BFloat16<false>::operator float() const (test/common/bfloat16.hpp:62)	called 1394600792 times	100.0%	-%	100.0%
kai::test::BFloat16<true> kai::test::read_array<kai::test::BFloat16<true>>(void const*, unsigned long) (test/common/memory.hpp:39)	called 290269930 times	100.0%	-%	100.0%
kai::test::BFloat16<true>::BFloat16(float) (test/common/bfloat16.hpp:39)	called 593584734 times	100.0%	-%	100.0%
kai::test::BFloat16<true>::float_to_bfloat16_round_towards_zero(float) (test/common/bfloat16.hpp:90)	not called	0.0%	-%	0.0%
kai::test::BFloat16<true>::operator float() const (test/common/bfloat16.hpp:62)	called 291804637 times	100.0%	-%	100.0%
kai::test::BFloat16_NumericLimitTest_Test::BFloat16_NumericLimitTest_Test() (test/tests/bfloat16_test.cpp:36)	called 2 times	100.0%	-%	100.0%
kai::test::BFloat16_NumericLimitTest_Test::TestBody() (test/tests/bfloat16_test.cpp:36)	called 1 time	100.0%	-%	24.0%
kai::test::BFloat16_NumericLimitTest_Test::~BFloat16_NumericLimitTest_Test() (test/tests/bfloat16_test.cpp:36)	called 3 times	83.3%	28.0%	100.0%
kai::test::BFloat16_SimpleTest_Test::BFloat16_SimpleTest_Test() (test/tests/bfloat16_test.cpp:16)	called 2 times	100.0%	-%	100.0%
kai::test::BFloat16_SimpleTest_Test::TestBody() (test/tests/bfloat16_test.cpp:16)	called 1 time	100.0%	-%	25.0%
kai::test::BFloat16_SimpleTest_Test::~BFloat16_SimpleTest_Test() (test/tests/bfloat16_test.cpp:16)	called 3 times	93.3%	23.3%	100.0%
kai::test::Buffer kai::test::(anonymous namespace)::binary_elementwise_any_op_type<(kai::test::(anonymous namespace)::BinaryElementwiseOperator)0, float>(void const*, void const*, unsigned long, unsigned long, unsigned long, unsigned long) (test/reference/binary_elementwise.cpp:70)	called 1191 times	100.0%	-%	72.0%
kai::test::Buffer kai::test::(anonymous namespace)::binary_elementwise_any_op_type<(kai::test::(anonymous namespace)::BinaryElementwiseOperator)0, int>(void const*, void const*, unsigned long, unsigned long, unsigned long, unsigned long) (test/reference/binary_elementwise.cpp:70)	not called	100.0%	-%	0.0%
kai::test::Buffer kai::test::(anonymous namespace)::binary_elementwise_any_op_type<(kai::test::(anonymous namespace)::BinaryElementwiseOperator)0, kai::test::Float16>(void const*, void const*, unsigned long, unsigned long, unsigned long, unsigned long) (test/reference/binary_elementwise.cpp:70)	called 1187 times	100.0%	-%	72.0%
kai::test::Buffer kai::test::(anonymous namespace)::binary_elementwise_any_op_type<(kai::test::(anonymous namespace)::BinaryElementwiseOperator)0, kai::test::UInt4>(void const*, void const*, unsigned long, unsigned long, unsigned long, unsigned long) (test/reference/binary_elementwise.cpp:70)	not called	100.0%	-%	0.0%
kai::test::Buffer kai::test::(anonymous namespace)::binary_elementwise_any_op_type<(kai::test::(anonymous namespace)::BinaryElementwiseOperator)1, float>(void const*, void const*, unsigned long, unsigned long, unsigned long, unsigned long) (test/reference/binary_elementwise.cpp:70)	not called	100.0%	-%	0.0%
kai::test::Buffer kai::test::(anonymous namespace)::binary_elementwise_any_op_type<(kai::test::(anonymous namespace)::BinaryElementwiseOperator)1, int>(void const*, void const*, unsigned long, unsigned long, unsigned long, unsigned long) (test/reference/binary_elementwise.cpp:70)	called 927 times	100.0%	-%	75.0%
kai::test::Buffer kai::test::(anonymous namespace)::binary_elementwise_any_op_type<(kai::test::(anonymous namespace)::BinaryElementwiseOperator)1, kai::test::Float16>(void const*, void const*, unsigned long, unsigned long, unsigned long, unsigned long) (test/reference/binary_elementwise.cpp:70)	not called	100.0%	-%	0.0%
kai::test::Buffer kai::test::(anonymous namespace)::binary_elementwise_any_op_type<(kai::test::(anonymous namespace)::BinaryElementwiseOperator)1, kai::test::UInt4>(void const*, void const*, unsigned long, unsigned long, unsigned long, unsigned long) (test/reference/binary_elementwise.cpp:70)	not called	100.0%	-%	0.0%
kai::test::Buffer kai::test::(anonymous namespace)::binary_elementwise_any_op_type<(kai::test::(anonymous namespace)::BinaryElementwiseOperator)2, float>(void const*, void const*, unsigned long, unsigned long, unsigned long, unsigned long) (test/reference/binary_elementwise.cpp:70)	called 1854 times	100.0%	-%	75.0%
kai::test::Buffer kai::test::(anonymous namespace)::binary_elementwise_any_op_type<(kai::test::(anonymous namespace)::BinaryElementwiseOperator)2, int>(void const*, void const*, unsigned long, unsigned long, unsigned long, unsigned long) (test/reference/binary_elementwise.cpp:70)	called 927 times	100.0%	-%	72.0%
kai::test::Buffer kai::test::(anonymous namespace)::binary_elementwise_any_op_type<(kai::test::(anonymous namespace)::BinaryElementwiseOperator)2, kai::test::Float16>(void const*, void const*, unsigned long, unsigned long, unsigned long, unsigned long) (test/reference/binary_elementwise.cpp:70)	not called	100.0%	-%	0.0%
kai::test::Buffer kai::test::(anonymous namespace)::binary_elementwise_any_op_type<(kai::test::(anonymous namespace)::BinaryElementwiseOperator)2, kai::test::UInt4>(void const*, void const*, unsigned long, unsigned long, unsigned long, unsigned long) (test/reference/binary_elementwise.cpp:70)	not called	100.0%	-%	0.0%
kai::test::Buffer kai::test::(anonymous namespace)::binary_elementwise_any_op_type<(kai::test::(anonymous namespace)::BinaryElementwiseOperator)3, float>(void const*, void const*, unsigned long, unsigned long, unsigned long, unsigned long) (test/reference/binary_elementwise.cpp:70)	not called	100.0%	-%	0.0%
kai::test::Buffer kai::test::(anonymous namespace)::binary_elementwise_any_op_type<(kai::test::(anonymous namespace)::BinaryElementwiseOperator)3, int>(void const*, void const*, unsigned long, unsigned long, unsigned long, unsigned long) (test/reference/binary_elementwise.cpp:70)	not called	100.0%	-%	0.0%
kai::test::Buffer kai::test::(anonymous namespace)::binary_elementwise_any_op_type<(kai::test::(anonymous namespace)::BinaryElementwiseOperator)3, kai::test::Float16>(void const*, void const*, unsigned long, unsigned long, unsigned long, unsigned long) (test/reference/binary_elementwise.cpp:70)	not called	100.0%	-%	0.0%
kai::test::Buffer kai::test::(anonymous namespace)::binary_elementwise_any_op_type<(kai::test::(anonymous namespace)::BinaryElementwiseOperator)3, kai::test::UInt4>(void const*, void const*, unsigned long, unsigned long, unsigned long, unsigned long) (test/reference/binary_elementwise.cpp:70)	not called	100.0%	23.5%	0.0%
kai::test::Buffer kai::test::(anonymous namespace)::binary_elementwise_any_type<(kai::test::(anonymous namespace)::BinaryElementwiseOperator)0>(void const*, kai::test::DataType, unsigned long, unsigned long, void const*, kai::test::DataType, unsigned long, unsigned long) (test/reference/binary_elementwise.cpp:97)	called 2378 times	100.0%	-%	50.0%
kai::test::Buffer kai::test::(anonymous namespace)::binary_elementwise_any_type<(kai::test::(anonymous namespace)::BinaryElementwiseOperator)1>(void const*, kai::test::DataType, unsigned long, unsigned long, void const*, kai::test::DataType, unsigned long, unsigned long) (test/reference/binary_elementwise.cpp:97)	not called	100.0%	-%	0.0%
kai::test::Buffer kai::test::(anonymous namespace)::binary_elementwise_any_type<(kai::test::(anonymous namespace)::BinaryElementwiseOperator)2>(void const*, kai::test::DataType, unsigned long, unsigned long, void const*, kai::test::DataType, unsigned long, unsigned long) (test/reference/binary_elementwise.cpp:97)	not called	100.0%	-%	0.0%
kai::test::Buffer kai::test::(anonymous namespace)::binary_elementwise_any_type<(kai::test::(anonymous namespace)::BinaryElementwiseOperator)3>(void const*, kai::test::DataType, unsigned long, unsigned long, void const*, kai::test::DataType, unsigned long, unsigned long) (test/reference/binary_elementwise.cpp:97)	not called	62.5%	10.0%	0.0%
kai::test::Buffer kai::test::(anonymous namespace)::fill_matrix_random_raw<float>(unsigned long, unsigned long, unsigned int) (test/reference/fill.cpp:29)	called 35739 times	100.0%	-%	50.0%
kai::test::Buffer kai::test::(anonymous namespace)::fill_matrix_random_raw<float>(unsigned long, unsigned long, unsigned int)::'lambda'(unsigned long, unsigned long)::operator()(unsigned long, unsigned long) const (test/reference/fill.cpp:36)	called 133017626 times	100.0%	50.0%	100.0%
kai::test::Buffer kai::test::(anonymous namespace)::fill_matrix_random_raw<kai::test::BFloat16<false>>(unsigned long, unsigned long, unsigned int) (test/reference/fill.cpp:57)	called 980 times	100.0%	50.0%	50.0%
kai::test::Buffer kai::test::(anonymous namespace)::fill_matrix_random_raw<kai::test::BFloat16<false>>(unsigned long, unsigned long, unsigned int)::$_0::operator()(unsigned long, unsigned long) const (test/reference/fill.cpp:62)	called 1942616 times	100.0%	-%	100.0%
kai::test::Buffer kai::test::(anonymous namespace)::fill_matrix_random_raw<kai::test::BFloat16<true>>(unsigned long, unsigned long, unsigned int) (test/reference/fill.cpp:48)	not called	0.0%	0.0%	0.0%
kai::test::Buffer kai::test::(anonymous namespace)::fill_matrix_random_raw<kai::test::BFloat16<true>>(unsigned long, unsigned long, unsigned int)::$_0::operator()(unsigned long, unsigned long) const (test/reference/fill.cpp:53)	not called	0.0%	-%	0.0%
kai::test::Buffer kai::test::(anonymous namespace)::fill_matrix_random_raw<kai::test::Float16>(unsigned long, unsigned long, unsigned int) (test/reference/fill.cpp:40)	called 7355 times	100.0%	-%	50.0%
kai::test::Buffer kai::test::(anonymous namespace)::fill_matrix_random_raw<kai::test::Float16>(unsigned long, unsigned long, unsigned int)::$_0::operator()(unsigned long, unsigned long) const (test/reference/fill.cpp:44)	called 28428890 times	100.0%	50.0%	100.0%
kai::test::Buffer kai::test::(anonymous namespace)::fill_matrix_random_raw<kai::test::Int4>(unsigned long, unsigned long, unsigned int) (test/reference/fill.cpp:66)	not called	0.0%	-%	0.0%
kai::test::Buffer kai::test::(anonymous namespace)::fill_matrix_random_raw<kai::test::Int4>(unsigned long, unsigned long, unsigned int)::$_0::operator()(unsigned long, unsigned long) const (test/reference/fill.cpp:70)	not called	0.0%	0.0%	0.0%
kai::test::Buffer kai::test::(anonymous namespace)::fill_matrix_random_raw<kai::test::UInt4>(unsigned long, unsigned long, unsigned int) (test/reference/fill.cpp:74)	not called	0.0%	-%	0.0%
kai::test::Buffer kai::test::(anonymous namespace)::fill_matrix_random_raw<kai::test::UInt4>(unsigned long, unsigned long, unsigned int)::$_0::operator()(unsigned long, unsigned long) const (test/reference/fill.cpp:78)	not called	0.0%	0.0%	0.0%
kai::test::Buffer kai::test::(anonymous namespace)::matmul_any_type<float>(void const*, void const*, unsigned long, unsigned long, unsigned long, bool, bool) (test/reference/matmul.cpp:45)	called 1191 times	100.0%	-%	64.0%
kai::test::Buffer kai::test::(anonymous namespace)::matmul_any_type<kai::test::Float16>(void const*, void const*, unsigned long, unsigned long, unsigned long, bool, bool) (test/reference/matmul.cpp:45)	called 1196 times	100.0%	63.0%	66.0%
kai::test::Buffer kai::test::(anonymous namespace)::reduce_any_op<(kai::test::ReductionOperator)0>(void const*, kai::test::DataFormat const&, unsigned long, unsigned long, kai::test::DataFormat const&, unsigned long) (test/reference/reduce.cpp:77)	not called	0.0%	0.0%	0.0%
kai::test::Buffer kai::test::(anonymous namespace)::reduce_any_op_type<(kai::test::ReductionOperator)0, kai::test::UInt4, int>(void const*, unsigned long, unsigned long, unsigned long) (test/reference/reduce.cpp:33)	not called	0.0%	0.0%	0.0%
kai::test::Buffer kai::test::cast<float, kai::test::BFloat16<false>>(void const*, unsigned long) (test/reference/cast.cpp:23)	called 980 times	100.0%	-%	68.0%
kai::test::Buffer kai::test::cast<float, kai::test::BFloat16<true>>(void const*, unsigned long) (test/reference/cast.cpp:23)	not called	100.0%	50.0%	0.0%
kai::test::Buffer kai::test::cast<float, kai::test::Float16>(void const*, unsigned long) (test/reference/cast.cpp:23)	called 3736 times	100.0%	-%	68.0%
kai::test::Buffer kai::test::cast<kai::test::BFloat16<false>, float>(void const*, unsigned long) (test/reference/cast.cpp:23)	called 2384 times	100.0%	-%	78.0%
kai::test::Buffer kai::test::cast<kai::test::BFloat16<false>, kai::test::Float16>(void const*, unsigned long) (test/reference/cast.cpp:34)	not called	0.0%	0.0%	0.0%
kai::test::Buffer kai::test::cast<kai::test::BFloat16<true>, float>(void const*, unsigned long) (test/reference/cast.cpp:23)	called 108 times	100.0%	-%	78.0%
kai::test::Buffer kai::test::cast<kai::test::BFloat16<true>, kai::test::Float16>(void const*, unsigned long) (test/reference/cast.cpp:46)	called 40 times	100.0%	58.3%	70.0%
kai::test::Buffer kai::test::cast<kai::test::Float16, float>(void const*, unsigned long) (test/reference/cast.cpp:23)	called 3984 times	100.0%	-%	78.0%
kai::test::Buffer kai::test::clamp<float>(void const*, unsigned long, float, float) (test/reference/clamp.cpp:66)	called 9447 times	100.0%	-%	68.0%
kai::test::Buffer kai::test::clamp<kai::test::Float16>(void const*, unsigned long, kai::test::Float16, kai::test::Float16) (test/reference/clamp.cpp:66)	not called	100.0%	30.0%	0.0%
kai::test::Buffer kai::test::compute_symmetric_per_block_quantization_info<float, int, float>(void const*, unsigned long, unsigned long, unsigned long) (test/reference/quantize.cpp:82)	not called	100.0%	53.1%	0.0%
kai::test::Buffer kai::test::compute_symmetric_per_block_quantization_info<float, kai::test::Int4, float>(void const*, unsigned long, unsigned long, unsigned long) (test/reference/quantize.cpp:82)	called 4452 times	100.0%	-%	75.0%
kai::test::Buffer kai::test::compute_symmetric_per_block_quantization_info<float, signed char, float>(void const*, unsigned long, unsigned long, unsigned long) (test/reference/quantize.cpp:82)	called 16363 times	100.0%	-%	79.0%
kai::test::Buffer kai::test::depthwise_reference<float>(unsigned long, unsigned long, unsigned long, unsigned long, unsigned long, unsigned long, void const*, void const*, void const*, kai::test::Padding2D const&) (test/reference/dwconv.cpp:23)	called 60 times	100.0%	81.2%	80.0%
kai::test::Buffer kai::test::fill_matrix_raw<float>(unsigned long, unsigned long, std::__1::function<float (unsigned long, unsigned long)>) (test/reference/fill.cpp:84)	called 37339 times	100.0%	-%	67.0%
kai::test::Buffer kai::test::fill_matrix_raw<kai::test::BFloat16<false>>(unsigned long, unsigned long, std::__1::function<kai::test::BFloat16<false> (unsigned long, unsigned long)>) (test/reference/fill.cpp:84)	called 980 times	100.0%	-%	72.0%
kai::test::Buffer kai::test::fill_matrix_raw<kai::test::BFloat16<true>>(unsigned long, unsigned long, std::__1::function<kai::test::BFloat16<true> (unsigned long, unsigned long)>) (test/reference/fill.cpp:84)	not called	100.0%	-%	0.0%
kai::test::Buffer kai::test::fill_matrix_raw<kai::test::Float16>(unsigned long, unsigned long, std::__1::function<kai::test::Float16 (unsigned long, unsigned long)>) (test/reference/fill.cpp:84)	called 7355 times	100.0%	-%	72.0%
kai::test::Buffer kai::test::fill_matrix_raw<kai::test::Int4>(unsigned long, unsigned long, std::__1::function<kai::test::Int4 (unsigned long, unsigned long)>) (test/reference/fill.cpp:84)	not called	100.0%	38.0%	0.0%
kai::test::Buffer kai::test::fill_matrix_raw<kai::test::UInt4>(unsigned long, unsigned long, std::__1::function<kai::test::UInt4 (unsigned long, unsigned long)>) (test/reference/fill.cpp:84)	not called	100.0%	-%	0.0%
kai::test::Buffer kai::test::fill_random<float>(unsigned long, unsigned int) (test/reference/fill.cpp:131)	called 31803 times	100.0%	-%	100.0%
kai::test::Buffer kai::test::fill_random<kai::test::BFloat16<false>>(unsigned long, unsigned int) (test/reference/fill.cpp:131)	called 980 times	100.0%	-%	100.0%
kai::test::Buffer kai::test::fill_random<kai::test::Float16>(unsigned long, unsigned int) (test/reference/fill.cpp:131)	called 3736 times	100.0%	-%	100.0%
kai::test::Buffer kai::test::indirect_matmul_nt_t_quantized<signed char, float, int, signed char, float, int, int, float, int, float>(unsigned long, unsigned long, unsigned long, unsigned long, void const* const*, unsigned long, void const*, void const*, void const*, unsigned long, unsigned long, void const*, void const*, void const*, unsigned long, unsigned long, void const*, void const*, void const*, unsigned long) (test/reference/matmul.cpp:219)	called 927 times	95.7%	56.5%	59.0%
kai::test::Buffer kai::test::matmul_clamp_nt_nt<signed char, float, int, kai::test::Int4, float, int, float, int, float>(unsigned long, unsigned long, unsigned long, void const*, void const*, void const*, unsigned long, void const*, void const*, void const*, unsigned long, void const*, float, float) (test/reference/matmul.cpp:588)	called 1112 times	100.0%	-%	64.0%
kai::test::Buffer kai::test::matmul_clamp_nt_nt<signed char, float, int, kai::test::Int4, kai::test::BFloat16<false>, int, float, int, float>(unsigned long, unsigned long, unsigned long, void const*, void const*, void const*, unsigned long, void const*, void const*, void const*, unsigned long, void const*, float, float) (test/reference/matmul.cpp:588)	not called	97.2%	29.8%	0.0%
kai::test::Buffer kai::test::matmul_clamp_nt_nt<signed char, float, int, signed char, float, int, float, int, float>(unsigned long, unsigned long, unsigned long, void const*, void const*, void const*, unsigned long, void const*, void const*, void const*, unsigned long, void const*, float, float) (test/reference/matmul.cpp:588)	called 616 times	100.0%	-%	66.0%
kai::test::Buffer kai::test::matmul_clamp_nt_nt<signed char, kai::test::Float16, int, kai::test::Int4, kai::test::Float16, int, float, int, float>(unsigned long, unsigned long, unsigned long, void const*, void const*, void const*, unsigned long, void const*, void const*, void const*, unsigned long, void const*, float, float) (test/reference/matmul.cpp:588)	not called	100.0%	-%	0.0%
kai::test::Buffer kai::test::matmul_clamp_nt_t<signed char, float, int, kai::test::Int4, float, int, float, int, float>(unsigned long, unsigned long, unsigned long, void const*, void const*, void const*, unsigned long, void const*, void const*, void const*, unsigned long, void const*, float, float) (test/reference/matmul.cpp:502)	called 880 times	100.0%	-%	64.0%
kai::test::Buffer kai::test::matmul_clamp_nt_t<signed char, float, int, kai::test::Int4, kai::test::BFloat16<false>, int, float, int, float>(unsigned long, unsigned long, unsigned long, void const*, void const*, void const*, unsigned long, void const*, void const*, void const*, unsigned long, void const*, float, float) (test/reference/matmul.cpp:502)	not called	100.0%	-%	0.0%
kai::test::Buffer kai::test::matmul_clamp_nt_t<signed char, float, int, signed char, float, int, float, int, float>(unsigned long, unsigned long, unsigned long, void const*, void const*, void const*, unsigned long, void const*, void const*, void const*, unsigned long, void const*, float, float) (test/reference/matmul.cpp:502)	called 616 times	97.2%	45.7%	66.0%
kai::test::Buffer kai::test::matmul_clamp_nt_t<signed char, kai::test::Float16, int, kai::test::Int4, kai::test::Float16, int, float, int, float>(unsigned long, unsigned long, unsigned long, void const*, void const*, void const*, unsigned long, void const*, void const*, void const*, unsigned long, void const*, float, float) (test/reference/matmul.cpp:502)	called 124 times	100.0%	-%	63.0%
kai::test::Buffer kai::test::matmul_nt_nt_quantized<kai::test::BFloat16<true>, float, float, kai::test::BFloat16<true>, float, float, float, float, float, float>(unsigned long, unsigned long, unsigned long, void const*, void const*, void const*, unsigned long, unsigned long, void const*, void const*, void const*, unsigned long, unsigned long, void const*, void const*, void const*, unsigned long) (test/reference/matmul.cpp:404)	called 74 times	90.0%	46.8%	53.0%
kai::test::Buffer kai::test::matmul_nt_nt_quantized<signed char, float, int, kai::test::Int4, kai::test::BFloat16<false>, int, float, float, int, float>(unsigned long, unsigned long, unsigned long, void const*, void const*, void const*, unsigned long, unsigned long, void const*, void const*, void const*, unsigned long, unsigned long, void const*, void const*, void const*, unsigned long) (test/reference/matmul.cpp:404)	called 702 times	100.0%	-%	55.0%
kai::test::Buffer kai::test::matmul_nt_t_quantized<signed char, float, int, kai::test::Int4, float, int, float, float, int, float>(unsigned long, unsigned long, unsigned long, void const*, void const*, void const*, unsigned long, unsigned long, void const*, void const*, void const*, unsigned long, unsigned long, void const*, void const*, void const*, unsigned long) (test/reference/matmul.cpp:300)	called 6164 times	100.0%	-%	58.0%
kai::test::Buffer kai::test::matmul_nt_t_quantized<signed char, float, int, kai::test::Int4, kai::test::BFloat16<false>, int, float, float, int, float>(unsigned long, unsigned long, unsigned long, void const*, void const*, void const*, unsigned long, unsigned long, void const*, void const*, void const*, unsigned long, unsigned long, void const*, void const*, void const*, unsigned long) (test/reference/matmul.cpp:300)	called 702 times	92.5%	40.3%	55.0%
kai::test::Buffer kai::test::matmul_nt_t_quantized<signed char, float, int, signed char, float, int, float, float, int, float>(unsigned long, unsigned long, unsigned long, void const*, void const*, void const*, unsigned long, unsigned long, void const*, void const*, void const*, unsigned long, unsigned long, void const*, void const*, void const*, unsigned long) (test/reference/matmul.cpp:300)	called 560 times	100.0%	-%	55.0%
kai::test::Buffer kai::test::matmul_nt_t_quantized<signed char, float, int, signed char, float, int, int, float, int, float>(unsigned long, unsigned long, unsigned long, void const*, void const*, void const*, unsigned long, unsigned long, void const*, void const*, void const*, unsigned long, unsigned long, void const*, void const*, void const*, unsigned long) (test/reference/matmul.cpp:300)	not called	100.0%	-%	0.0%
kai::test::Buffer kai::test::matmul_pack_rhs_nxk_static_quantized<signed char, float, int>(void const*, void const*, float, float, void const*, int, unsigned long, unsigned long, unsigned long, unsigned long) (test/reference/matmul_pack.cpp:22)	called 927 times	100.0%	50.0%	57.0%
kai::test::Buffer kai::test::mul<float>(void const*, unsigned long, unsigned long, void const*, unsigned long, unsigned long) (test/reference/binary_elementwise.cpp:158)	called 1854 times	100.0%	-%	100.0%
kai::test::Buffer kai::test::mul<int>(void const*, unsigned long, unsigned long, void const*, unsigned long, unsigned long) (test/reference/binary_elementwise.cpp:158)	called 927 times	100.0%	-%	100.0%
kai::test::Buffer kai::test::pack_data_scales_interleave_block<kai::test::UInt4, kai::test::Float16>(void const*, void const*, unsigned long, unsigned long, unsigned long) (test/reference/pack.cpp:309)	called 124 times	100.0%	-%	64.0%
kai::test::Buffer kai::test::pack_data_scales_interleave_block<kai::test::UInt4, std::nullptr_t>(void const*, void const*, unsigned long, unsigned long, unsigned long) (test/reference/pack.cpp:309)	not called	100.0%	35.0%	0.0%
kai::test::Buffer kai::test::pack_zero_points_data_scales_per_block<int, signed char, float>(void const*, void const*, void const*, unsigned long, unsigned long, unsigned long, unsigned long) (test/reference/pack.cpp:259)	called 927 times	100.0%	66.7%	70.0%
kai::test::Buffer kai::test::pad_matrix<float>(void const*, unsigned long, unsigned long, unsigned long, unsigned long, unsigned long, unsigned long, float) (test/reference/pad.cpp:44)	called 927 times	100.0%	-%	83.0%
kai::test::Buffer kai::test::pad_matrix<int>(void const*, unsigned long, unsigned long, unsigned long, unsigned long, unsigned long, unsigned long, int) (test/reference/pad.cpp:44)	called 927 times	100.0%	63.6%	83.0%
kai::test::Buffer kai::test::pad_row<kai::test::Int4>(void const*, unsigned long, unsigned long, unsigned long, unsigned long, unsigned long, unsigned char) (test/reference/pad.cpp:22)	called 2104 times	100.0%	-%	83.0%
kai::test::Buffer kai::test::pad_row<kai::test::UInt4>(void const*, unsigned long, unsigned long, unsigned long, unsigned long, unsigned long, unsigned char) (test/reference/pad.cpp:22)	called 2906 times	100.0%	70.0%	83.0%
kai::test::Buffer kai::test::quantize_asymmetric_per_block<float, kai::test::Int4, float, int>(void const*, void const*, void const*, unsigned long, unsigned long, unsigned long) (test/reference/quantize.cpp:236)	called 5232 times	100.0%	75.0%	87.0%
kai::test::Buffer kai::test::quantize_asymmetric_per_block<float, signed char, float, int>(void const*, void const*, void const*, unsigned long, unsigned long, unsigned long) (test/reference/quantize.cpp:236)	called 7974 times	100.0%	-%	80.0%
kai::test::Buffer kai::test::quantize_symmetric_per_block<float, int, float>(void const*, void const*, unsigned long, unsigned long, unsigned long) (test/reference/quantize.cpp:121)	called 927 times	100.0%	-%	82.0%
kai::test::Buffer kai::test::quantize_symmetric_per_block<float, kai::test::Int4, float>(void const*, void const*, unsigned long, unsigned long, unsigned long) (test/reference/quantize.cpp:121)	called 4452 times	100.0%	-%	87.0%
kai::test::Buffer kai::test::quantize_symmetric_per_block<float, signed char, float>(void const*, void const*, unsigned long, unsigned long, unsigned long) (test/reference/quantize.cpp:121)	called 16363 times	100.0%	72.5%	87.0%
kai::test::Buffer kai::test::reduce_add_x<signed char, int>(void const*, unsigned long, unsigned long) (test/reference/reduce.cpp:114)	called 927 times	100.0%	70.0%	75.0%
kai::test::Buffer kai::test::reorder_block<signed char>(void const*, unsigned long, unsigned long, unsigned long, unsigned long) (test/reference/reorder.cpp:20)	called 1854 times	100.0%	-%	90.0%
kai::test::Buffer kai::test::reorder_block<void const*>(void const*, unsigned long, unsigned long, unsigned long, unsigned long) (test/reference/reorder.cpp:20)	called 3171 times	100.0%	80.6%	90.0%
kai::test::Buffer kai::test::sub<int>(void const*, unsigned long, unsigned long, void const*, unsigned long, unsigned long) (test/reference/binary_elementwise.cpp:139)	called 927 times	100.0%	-%	100.0%
kai::test::Buffer kai::test::transpose<float>(void const*, unsigned long, unsigned long) (test/reference/transpose.cpp:65)	called 927 times	100.0%	-%	83.0%
kai::test::Buffer kai::test::transpose<signed char>(void const*, unsigned long, unsigned long) (test/reference/transpose.cpp:65)	called 927 times	100.0%	70.0%	83.0%
kai::test::Buffer kai::test::transpose_with_padding<kai::test::Int4>(void const*, unsigned long, unsigned long, unsigned long, unsigned long, unsigned long) (test/reference/transpose.cpp:41)	called 1814 times	100.0%	-%	83.0%
kai::test::Buffer kai::test::transpose_with_padding<signed char>(void const*, unsigned long, unsigned long, unsigned long, unsigned long, unsigned long) (test/reference/transpose.cpp:41)	called 616 times	100.0%	70.0%	83.0%
kai::test::Buffer::Buffer() (test/common/buffer.hpp:32)	called 122746 times	100.0%	-%	100.0%
kai::test::Buffer::Buffer(kai::test::Buffer&&) (test/common/buffer.hpp:37)	called 246626 times	100.0%	-%	100.0%
kai::test::Buffer::Buffer(unsigned long) (test/common/buffer.cpp:24)	called 321254 times	100.0%	-%	100.0%
kai::test::Buffer::Buffer(unsigned long, unsigned char) (test/common/buffer.cpp:27)	called 733698 times	100.0%	-%	100.0%
kai::test::Buffer::allocate() (test/common/buffer.cpp:67)	called 366649 times	100.0%	-%	53.0%
kai::test::Buffer::allocate_with_guard_pages() (test/common/buffer.cpp:74)	called 200 times	100.0%	100.0%	38.0%
kai::test::Buffer::allocate_with_guard_pages()::$_0::operator()(void*) const (test/common/buffer.cpp:95)	called 200 times	73.3%	50.0%	40.0%
kai::test::Buffer::data() const (test/common/buffer.hpp:46)	called 418798584 times	100.0%	-%	100.0%
kai::test::Buffer::operator=(kai::test::Buffer&&) (test/common/buffer.hpp:39)	called 55044 times	100.0%	-%	100.0%
kai::test::Buffer::size() const (test/common/buffer.hpp:56)	called 4098469 times	100.0%	-%	100.0%
kai::test::Buffer::~Buffer() (test/common/buffer.hpp:41)	called 1077534 times	100.0%	-%	100.0%
kai::test::BufferDeathTest_InvalidPolicy_Test::BufferDeathTest_InvalidPolicy_Test() (test/tests/buffer_test.cpp:59)	called 2 times	100.0%	-%	100.0%
kai::test::BufferDeathTest_InvalidPolicy_Test::TestBody() (test/tests/buffer_test.cpp:59)	called 1 time	100.0%	-%	28.0%
kai::test::BufferDeathTest_InvalidPolicy_Test::~BufferDeathTest_InvalidPolicy_Test() (test/tests/buffer_test.cpp:59)	called 3 times	83.3%	26.8%	100.0%
kai::test::BufferDeathTest_ProtectOverflowPolicy_Test::BufferDeathTest_ProtectOverflowPolicy_Test() (test/tests/buffer_test.cpp:140)	called 2 times	100.0%	-%	100.0%
kai::test::BufferDeathTest_ProtectOverflowPolicy_Test::TestBody() (test/tests/buffer_test.cpp:140)	called 1 time	100.0%	-%	30.0%
kai::test::BufferDeathTest_ProtectOverflowPolicy_Test::TestBody()::$_0::operator()(unsigned long) const (test/tests/buffer_test.cpp:166)	called 100 times	85.7%	19.8%	50.0%
kai::test::BufferDeathTest_ProtectOverflowPolicy_Test::~BufferDeathTest_ProtectOverflowPolicy_Test() (test/tests/buffer_test.cpp:140)	called 3 times	86.7%	36.4%	100.0%
kai::test::BufferDeathTest_ProtectUnderflowPolicy_Test::BufferDeathTest_ProtectUnderflowPolicy_Test() (test/tests/buffer_test.cpp:93)	called 2 times	100.0%	-%	100.0%
kai::test::BufferDeathTest_ProtectUnderflowPolicy_Test::TestBody() (test/tests/buffer_test.cpp:93)	called 1 time	100.0%	-%	30.0%
kai::test::BufferDeathTest_ProtectUnderflowPolicy_Test::TestBody()::$_0::operator()(unsigned long) const (test/tests/buffer_test.cpp:119)	called 100 times	85.7%	19.8%	50.0%
kai::test::BufferDeathTest_ProtectUnderflowPolicy_Test::~BufferDeathTest_ProtectUnderflowPolicy_Test() (test/tests/buffer_test.cpp:93)	called 3 times	86.7%	36.4%	100.0%
kai::test::BufferTest_NonePolicy_Test::BufferTest_NonePolicy_Test() (test/tests/buffer_test.cpp:27)	called 2 times	100.0%	-%	100.0%
kai::test::BufferTest_NonePolicy_Test::TestBody() (test/tests/buffer_test.cpp:27)	called 1 time	100.0%	-%	31.0%
kai::test::BufferTest_NonePolicy_Test::~BufferTest_NonePolicy_Test() (test/tests/buffer_test.cpp:27)	called 3 times	85.0%	30.9%	100.0%
kai::test::CacheData const& kai::test::getV<std::__1::tuple<kai::test::MatMulShape, kai::test::DataFormat, kai::test::DataFormat, kai::test::DataFormat>, kai::test::CacheData, kai::test::ReferenceGenerator<std::__1::tuple<kai::test::MatMulShape, kai::test::DataFormat, kai::test::DataFormat, kai::test::DataFormat>, kai::test::CacheData>>(std::__1::tuple<kai::test::MatMulShape, kai::test::DataFormat, kai::test::DataFormat, kai::test::DataFormat> const&) (test/common/cache.hpp:91)	called 1232 times	100.0%	-%	100.0%
kai::test::CacheData::CacheData() (test/tests/matmul_clamp_f32_qai8dxp_qsi8cxp_test.cpp:45)	called 44 times	100.0%	-%	100.0%
kai::test::CacheData::operator=(kai::test::CacheData&&) (test/tests/matmul_clamp_f32_qai8dxp_qsi8cxp_test.cpp:45)	called 11 times	100.0%	-%	100.0%
kai::test::CacheData::~CacheData() (test/tests/matmul_clamp_f32_qai8dxp_qsi8cxp_test.cpp:45)	called 44 times	100.0%	-%	100.0%
kai::test::DataFormat::DataFormat(kai::test::DataType, unsigned long, unsigned long, kai::test::DataFormat::PackFormat, kai::test::DataType, kai::test::DataType, unsigned long, unsigned long) (test/common/data_format.cpp:19)	called 72180 times	100.0%	-%	100.0%
kai::test::DataFormat::Hash::operator()(kai::test::DataFormat const&) const (test/common/data_format.cpp:179)	called 120417 times	100.0%	-%	100.0%
kai::test::DataFormat::actual_block_height(unsigned long) const (test/common/data_format.cpp:78)	called 67113 times	100.0%	75.0%	85.0%
kai::test::DataFormat::actual_block_width(unsigned long) const (test/common/data_format.cpp:83)	called 75599 times	100.0%	100.0%	100.0%
kai::test::DataFormat::actual_subblock_height(unsigned long) const (test/common/data_format.cpp:87)	called 33899 times	100.0%	100.0%	100.0%
kai::test::DataFormat::actual_subblock_width(unsigned long) const (test/common/data_format.cpp:91)	called 33899 times	100.0%	100.0%	100.0%
kai::test::DataFormat::block_height() const (test/common/data_format.cpp:62)	not called	0.0%	-%	0.0%
kai::test::DataFormat::block_width() const (test/common/data_format.cpp:66)	not called	0.0%	-%	0.0%
kai::test::DataFormat::data_type() const (test/common/data_format.cpp:41)	called 121184 times	100.0%	-%	100.0%
kai::test::DataFormat::default_offset_in_bytes(unsigned long, unsigned long, unsigned long) const (test/common/data_format.cpp:151)	called 4138 times	88.9%	80.0%	50.0%
kai::test::DataFormat::default_row_stride(unsigned long) const (test/common/data_format.cpp:128)	called 8486 times	63.6%	66.7%	37.0%
kai::test::DataFormat::default_size_in_bytes(unsigned long, unsigned long) const (test/common/data_format.cpp:173)	called 1332 times	100.0%	100.0%	100.0%
kai::test::DataFormat::is_raw() const (test/common/data_format.cpp:57)	called 776 times	100.0%	50.0%	100.0%
kai::test::DataFormat::operator!=(kai::test::DataFormat const&) const (test/common/data_format.cpp:37)	not called	0.0%	-%	0.0%
kai::test::DataFormat::operator==(kai::test::DataFormat const&) const (test/common/data_format.cpp:32)	called 102399 times	100.0%	50.0%	100.0%
kai::test::DataFormat::pack_format() const (test/common/data_format.cpp:45)	called 42413 times	100.0%	-%	100.0%
kai::test::DataFormat::scale_data_type() const (test/common/data_format.cpp:49)	called 34126 times	100.0%	-%	100.0%
kai::test::DataFormat::scheduler_block_height(unsigned long) const (test/common/data_format.cpp:95)	called 1124 times	85.7%	71.4%	51.0%
kai::test::DataFormat::scheduler_block_width(unsigned long) const (test/common/data_format.cpp:112)	called 5118 times	85.7%	85.7%	57.0%
kai::test::DataFormat::subblock_height() const (test/common/data_format.cpp:70)	not called	0.0%	-%	0.0%
kai::test::DataFormat::subblock_width() const (test/common/data_format.cpp:74)	not called	0.0%	-%	0.0%
kai::test::DataFormat::zero_point_data_type() const (test/common/data_format.cpp:53)	called 33908 times	100.0%	-%	100.0%
kai::test::DefaultMismatchHandler::DefaultMismatchHandler(float, float, unsigned long, float) (test/common/compare.cpp:289)	called 65912 times	100.0%	-%	100.0%
kai::test::DefaultMismatchHandler::DefaultMismatchHandler(kai::test::DefaultMismatchHandler const&) (test/common/compare.cpp:300)	not called	0.0%	-%	0.0%
kai::test::DefaultMismatchHandler::handle_data(float, float) (test/common/compare.cpp:327)	called 37870175 times	100.0%	75.0%	100.0%
kai::test::DefaultMismatchHandler::mark_as_failed() (test/common/compare.cpp:337)	not called	0.0%	-%	0.0%
kai::test::DefaultMismatchHandler::operator=(kai::test::DefaultMismatchHandler const&) (test/common/compare.cpp:312)	not called	0.0%	0.0%	0.0%
kai::test::DefaultMismatchHandler::success(unsigned long) const (test/common/compare.cpp:341)	called 33658 times	83.3%	75.0%	83.0%
kai::test::DefaultMismatchHandler::~DefaultMismatchHandler() (test/common/compare.hpp:97)	called 65912 times	100.0%	-%	100.0%
kai::test::DepthwisePlanarTest_Output_Test::AddToRegistry() (test/tests/dwconv_test.cpp:235)	called 1 time	100.0%	-%	44.0%
kai::test::DepthwisePlanarTest_Output_Test::DepthwisePlanarTest_Output_Test() (test/tests/dwconv_test.cpp:235)	called 120 times	100.0%	-%	100.0%
kai::test::DepthwisePlanarTest_Output_Test::TestBody() (test/tests/dwconv_test.cpp:235)	called 60 times	100.0%	-%	33.0%
kai::test::DepthwisePlanarTest_Output_Test::~DepthwisePlanarTest_Output_Test() (test/tests/dwconv_test.cpp:235)	called 180 times	95.8%	36.2%	100.0%
kai::test::Float16 kai::test::(anonymous namespace)::scalar_binary_elementwise<(kai::test::(anonymous namespace)::BinaryElementwiseOperator)0, kai::test::Float16>(kai::test::Float16, kai::test::Float16) (test/reference/binary_elementwise.cpp:42)	called 2902781 times	100.0%	-%	100.0%
kai::test::Float16 kai::test::(anonymous namespace)::scalar_binary_elementwise<(kai::test::(anonymous namespace)::BinaryElementwiseOperator)1, kai::test::Float16>(kai::test::Float16, kai::test::Float16) (test/reference/binary_elementwise.cpp:42)	not called	100.0%	-%	0.0%
kai::test::Float16 kai::test::(anonymous namespace)::scalar_binary_elementwise<(kai::test::(anonymous namespace)::BinaryElementwiseOperator)2, kai::test::Float16>(kai::test::Float16, kai::test::Float16) (test/reference/binary_elementwise.cpp:42)	not called	100.0%	-%	0.0%
kai::test::Float16 kai::test::(anonymous namespace)::scalar_binary_elementwise<(kai::test::(anonymous namespace)::BinaryElementwiseOperator)3, kai::test::Float16>(kai::test::Float16, kai::test::Float16) (test/reference/binary_elementwise.cpp:42)	not called	100.0%	-%	0.0%
kai::test::Float16 kai::test::read_array<kai::test::Float16>(void const*, unsigned long) (test/common/memory.hpp:39)	called 1279391564 times	100.0%	100.0%	100.0%
kai::test::Float16::Float16() (test/common/float16.hpp:98)	called 1207246214 times	100.0%	-%	100.0%
kai::test::Float16::Float16(float) (test/common/float16.hpp:104)	called 103717472 times	100.0%	-%	100.0%
kai::test::Float16::operator float() const (test/common/float16.hpp:127)	called 114399181 times	100.0%	-%	100.0%
kai::test::Float16::operator*=(kai::test::Float16) (test/common/float16.hpp:144)	called 1 time	100.0%	-%	100.0%
kai::test::Float16::operator+=(kai::test::Float16) (test/common/float16.hpp:132)	called 600720322 times	100.0%	-%	100.0%
kai::test::Float16::operator-=(kai::test::Float16) (test/common/float16.hpp:138)	called 1 time	100.0%	-%	100.0%
kai::test::Float16::operator/=(kai::test::Float16) (test/common/float16.hpp:150)	called 1 time	100.0%	-%	100.0%
kai::test::Float16_NumericLimitTest_Test::Float16_NumericLimitTest_Test() (test/tests/float16_test.cpp:66)	called 2 times	100.0%	-%	100.0%
kai::test::Float16_NumericLimitTest_Test::TestBody() (test/tests/float16_test.cpp:66)	called 1 time	100.0%	-%	27.0%
kai::test::Float16_NumericLimitTest_Test::~Float16_NumericLimitTest_Test() (test/tests/float16_test.cpp:66)	called 3 times	100.0%	31.0%	100.0%
kai::test::Float16_SimpleTest_Test::Float16_SimpleTest_Test() (test/tests/float16_test.cpp:16)	called 2 times	100.0%	-%	100.0%
kai::test::Float16_SimpleTest_Test::TestBody() (test/tests/float16_test.cpp:16)	called 1 time	100.0%	-%	26.0%
kai::test::Float16_SimpleTest_Test::~Float16_SimpleTest_Test() (test/tests/float16_test.cpp:16)	called 3 times	97.4%	22.3%	100.0%
kai::test::IndirectMatMulQuantizedTest_EndToEnd_Test::AddToRegistry() (test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp:969)	called 1 time	100.0%	-%	44.0%
kai::test::IndirectMatMulQuantizedTest_EndToEnd_Test::IndirectMatMulQuantizedTest_EndToEnd_Test() (test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp:969)	called 18648 times	100.0%	-%	100.0%
kai::test::IndirectMatMulQuantizedTest_EndToEnd_Test::TestBody() (test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp:969)	called 9324 times	100.0%	-%	28.0%
kai::test::IndirectMatMulQuantizedTest_EndToEnd_Test::~IndirectMatMulQuantizedTest_EndToEnd_Test() (test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp:969)	called 27972 times	92.9%	40.6%	100.0%
kai::test::IndirectMatMulTest_Output_Test::AddToRegistry() (test/tests/imatmul_test.cpp:566)	called 1 time	100.0%	-%	44.0%
kai::test::IndirectMatMulTest_Output_Test::IndirectMatMulTest_Output_Test() (test/tests/imatmul_test.cpp:566)	called 53856 times	100.0%	-%	100.0%
kai::test::IndirectMatMulTest_Output_Test::TestBody() (test/tests/imatmul_test.cpp:566)	called 26928 times	100.0%	-%	40.0%
kai::test::IndirectMatMulTest_Output_Test::~IndirectMatMulTest_Output_Test() (test/tests/imatmul_test.cpp:566)	called 80784 times	95.0%	44.4%	100.0%
kai::test::Int4 kai::test::quantize_asymmetric<float, kai::test::Int4, int>(float, float, int) (test/reference/quantize.cpp:72)	called 21518336 times	100.0%	50.0%	75.0%
kai::test::Int4 kai::test::quantize_symmetric<kai::test::Int4>(float, float) (test/reference/quantize.cpp:60)	called 25768996 times	100.0%	-%	75.0%
kai::test::Int4 kai::test::read_array<kai::test::Int4>(void const*, unsigned long) (test/common/memory.hpp:39)	called 2119608532 times	80.0%	50.0%	100.0%
kai::test::Int4::Int4(signed char) (test/common/int4.hpp:74)	called 8833010184 times	100.0%	-%	100.0%
kai::test::Int4::operator float() const (test/common/int4.cpp:84)	called 1711103008 times	100.0%	-%	100.0%
kai::test::Int4::operator int() const (test/common/int4.cpp:80)	called 486417508 times	100.0%	-%	100.0%
kai::test::Int4::operator*(kai::test::Int4) const (test/common/int4.cpp:96)	not called	0.0%	-%	0.0%
kai::test::Int4::operator+(kai::test::Int4) const (test/common/int4.cpp:88)	not called	0.0%	-%	0.0%
kai::test::Int4::operator-(kai::test::Int4) const (test/common/int4.cpp:92)	not called	0.0%	-%	0.0%
kai::test::Int4::operator/(kai::test::Int4) const (test/common/int4.cpp:100)	not called	0.0%	-%	0.0%
kai::test::Int4::operator=(int) (test/common/int4.cpp:74)	not called	0.0%	-%	0.0%
kai::test::Int4::operator=(signed char) (test/common/int4.cpp:68)	not called	0.0%	-%	0.0%
kai::test::Int4::pack_u8(kai::test::Int4, kai::test::Int4) (test/common/int4.cpp:104)	called 65000348 times	100.0%	-%	100.0%
kai::test::Int4::unpack_u8(unsigned char) (test/common/int4.cpp:110)	called 2184608880 times	100.0%	-%	100.0%
kai::test::MatMulMethod::MatMulMethod(kai::test::MatMulMethod const&) (test/common/matmul_test_common.hpp:61)	called 22236 times	100.0%	-%	100.0%
kai::test::MatMulMethod::has_main_kernel() const (test/common/matmul_test_common.hpp:450)	called 876 times	100.0%	100.0%	100.0%
kai::test::MatMulMethod::is_pack_lhs_needed() const (test/common/matmul_test_common.hpp:173)	called 830 times	100.0%	-%	100.0%
kai::test::MatMulMethod::is_pack_rhs_needed() const (test/common/matmul_test_common.hpp:395)	called 830 times	100.0%	-%	100.0%
kai::test::MatMulMethod::is_pack_rhs_nxk_needed() const (test/common/matmul_test_common.hpp:400)	called 440 times	100.0%	-%	100.0%
kai::test::MatMulMethod::main_kernel(unsigned long, unsigned long, unsigned long, void const*, void const*, void const*, void*, unsigned long, unsigned long, unsigned long, float, float) const (test/common/matmul_test_common.hpp:459)	called 796 times	100.0%	91.7%	85.0%
kai::test::MatMulMethod::pack_rhs(unsigned long, unsigned long, void const*, unsigned long, void const*, void const*, void*) const (test/common/matmul_test_common.hpp:413)	called 492 times	100.0%	50.0%	50.0%
kai::test::MatMulMethod::pack_rhs_nxk(unsigned long, unsigned long, void const*, unsigned long, void const*, void const*, void*) const (test/common/matmul_test_common.hpp:436)	called 68 times	100.0%	50.0%	50.0%
kai::test::MatMulMethod::~MatMulMethod() (test/common/matmul_test_common.hpp:61)	called 17892 times	100.0%	50.0%	100.0%
kai::test::MatMulQuantizedTest_EndToEnd_Test::AddToRegistry() (test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp:840)	called 1 time	100.0%	-%	44.0%
kai::test::MatMulQuantizedTest_EndToEnd_Test::MatMulQuantizedTest_EndToEnd_Test() (test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp:840)	called 1668 times	100.0%	-%	100.0%
kai::test::MatMulQuantizedTest_EndToEnd_Test::TestBody() (test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp:840)	called 834 times	100.0%	-%	30.0%
kai::test::MatMulQuantizedTest_EndToEnd_Test::~MatMulQuantizedTest_EndToEnd_Test() (test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp:840)	called 2502 times	96.3%	28.0%	100.0%
kai::test::MatMulShape::Hash::operator()(kai::test::MatMulShape const&) const (test/common/matmul_test_common.hpp:29)	called 81997 times	100.0%	-%	100.0%
kai::test::MatMulTest::MatMulTest() (test/tests/matmul_test.cpp:615)	called 1760 times	100.0%	-%	100.0%
kai::test::MatMulTest::TestData::TestData() (test/tests/matmul_test.cpp:622)	called 250 times	100.0%	-%	100.0%
kai::test::MatMulTest::TestData::operator=(kai::test::MatMulTest::TestData&&) (test/tests/matmul_test.cpp:622)	called 125 times	100.0%	-%	100.0%
kai::test::MatMulTest::TestData::~TestData() (test/tests/matmul_test.cpp:622)	called 250 times	100.0%	-%	100.0%
kai::test::MatMulTest::test_data() (test/tests/matmul_test.cpp:636)	called 1024 times	95.0%	49.1%	55.0%
kai::test::MatMulTest::~MatMulTest() (test/tests/matmul_test.cpp:615)	called 1760 times	100.0%	-%	100.0%
kai::test::MatMulTestBf16::MatMulTestBf16() (test/tests/matmul_clamp_f32_bf16p_bf16p_test.cpp:318)	called 346 times	100.0%	-%	100.0%
kai::test::MatMulTestBf16::TestData::TestData() (test/tests/matmul_clamp_f32_bf16p_bf16p_test.cpp:325)	called 148 times	100.0%	-%	100.0%
kai::test::MatMulTestBf16::TestData::operator=(kai::test::MatMulTestBf16::TestData&&) (test/tests/matmul_clamp_f32_bf16p_bf16p_test.cpp:325)	called 74 times	100.0%	-%	100.0%
kai::test::MatMulTestBf16::TestData::~TestData() (test/tests/matmul_clamp_f32_bf16p_bf16p_test.cpp:325)	called 148 times	100.0%	-%	100.0%
kai::test::MatMulTestBf16::test_data() (test/tests/matmul_clamp_f32_bf16p_bf16p_test.cpp:336)	called 346 times	95.9%	52.6%	57.0%
kai::test::MatMulTestBf16::~MatMulTestBf16() (test/tests/matmul_clamp_f32_bf16p_bf16p_test.cpp:318)	called 346 times	100.0%	-%	100.0%
kai::test::MatMulTestBf16OutFp16::MatMulTestBf16OutFp16() (test/tests/matmul_clamp_f16_bf16p_bf16p_test.cpp:121)	called 90 times	100.0%	-%	100.0%
kai::test::MatMulTestBf16OutFp16::TestData::TestData() (test/tests/matmul_clamp_f16_bf16p_bf16p_test.cpp:128)	called 36 times	100.0%	-%	100.0%
kai::test::MatMulTestBf16OutFp16::TestData::operator=(kai::test::MatMulTestBf16OutFp16::TestData&&) (test/tests/matmul_clamp_f16_bf16p_bf16p_test.cpp:128)	called 18 times	100.0%	-%	100.0%
kai::test::MatMulTestBf16OutFp16::TestData::~TestData() (test/tests/matmul_clamp_f16_bf16p_bf16p_test.cpp:128)	called 36 times	100.0%	-%	100.0%
kai::test::MatMulTestBf16OutFp16::test_data() (test/tests/matmul_clamp_f16_bf16p_bf16p_test.cpp:139)	called 90 times	94.5%	51.1%	59.0%
kai::test::MatMulTestBf16OutFp16::~MatMulTestBf16OutFp16() (test/tests/matmul_clamp_f16_bf16p_bf16p_test.cpp:121)	called 90 times	100.0%	-%	100.0%
kai::test::MatMulTestBf16OutFp16_Output_Test::AddToRegistry() (test/tests/matmul_clamp_f16_bf16p_bf16p_test.cpp:226)	called 1 time	100.0%	-%	44.0%
kai::test::MatMulTestBf16OutFp16_Output_Test::MatMulTestBf16OutFp16_Output_Test() (test/tests/matmul_clamp_f16_bf16p_bf16p_test.cpp:226)	called 180 times	100.0%	-%	100.0%
kai::test::MatMulTestBf16OutFp16_Output_Test::TestBody() (test/tests/matmul_clamp_f16_bf16p_bf16p_test.cpp:226)	called 90 times	100.0%	-%	40.0%
kai::test::MatMulTestBf16OutFp16_Output_Test::~MatMulTestBf16OutFp16_Output_Test() (test/tests/matmul_clamp_f16_bf16p_bf16p_test.cpp:226)	called 270 times	96.8%	40.1%	100.0%
kai::test::MatMulTestBf16_Output_Test::AddToRegistry() (test/tests/matmul_clamp_f32_bf16p_bf16p_test.cpp:448)	called 1 time	100.0%	-%	44.0%
kai::test::MatMulTestBf16_Output_Test::MatMulTestBf16_Output_Test() (test/tests/matmul_clamp_f32_bf16p_bf16p_test.cpp:448)	called 692 times	100.0%	-%	100.0%
kai::test::MatMulTestBf16_Output_Test::TestBody() (test/tests/matmul_clamp_f32_bf16p_bf16p_test.cpp:448)	called 346 times	100.0%	-%	40.0%
kai::test::MatMulTestBf16_Output_Test::~MatMulTestBf16_Output_Test() (test/tests/matmul_clamp_f32_bf16p_bf16p_test.cpp:448)	called 1038 times	97.1%	40.8%	100.0%
kai::test::MatMulTest_Output_Test::AddToRegistry() (test/tests/matmul_test.cpp:916)	called 1 time	100.0%	-%	44.0%
kai::test::MatMulTest_Output_Test::MatMulTest_Output_Test() (test/tests/matmul_test.cpp:916)	called 880 times	100.0%	-%	100.0%
kai::test::MatMulTest_Output_Test::TestBody() (test/tests/matmul_test.cpp:916)	called 440 times	100.0%	-%	37.0%
kai::test::MatMulTest_Output_Test::~MatMulTest_Output_Test() (test/tests/matmul_test.cpp:916)	called 1320 times	89.9%	32.5%	100.0%
kai::test::MatMulTest_PackedLhs_Test::AddToRegistry() (test/tests/matmul_test.cpp:731)	called 1 time	100.0%	-%	44.0%
kai::test::MatMulTest_PackedLhs_Test::MatMulTest_PackedLhs_Test() (test/tests/matmul_test.cpp:731)	called 880 times	100.0%	-%	100.0%
kai::test::MatMulTest_PackedLhs_Test::TestBody() (test/tests/matmul_test.cpp:731)	called 440 times	100.0%	-%	39.0%
kai::test::MatMulTest_PackedLhs_Test::~MatMulTest_PackedLhs_Test() (test/tests/matmul_test.cpp:731)	called 1320 times	97.2%	36.9%	100.0%
kai::test::MatMulTest_PackedRhs_Test::AddToRegistry() (test/tests/matmul_test.cpp:783)	called 1 time	100.0%	-%	44.0%
kai::test::MatMulTest_PackedRhs_Test::MatMulTest_PackedRhs_Test() (test/tests/matmul_test.cpp:783)	called 880 times	100.0%	-%	100.0%
kai::test::MatMulTest_PackedRhs_Test::TestBody() (test/tests/matmul_test.cpp:783)	called 440 times	100.0%	-%	35.0%
kai::test::MatMulTest_PackedRhs_Test::~MatMulTest_PackedRhs_Test() (test/tests/matmul_test.cpp:783)	called 1320 times	95.9%	32.5%	100.0%
kai::test::MatMulTest_PackedTransposedRhs_Test::AddToRegistry() (test/tests/matmul_test.cpp:855)	called 1 time	100.0%	-%	44.0%
kai::test::MatMulTest_PackedTransposedRhs_Test::MatMulTest_PackedTransposedRhs_Test() (test/tests/matmul_test.cpp:855)	called 880 times	100.0%	-%	100.0%
kai::test::MatMulTest_PackedTransposedRhs_Test::TestBody() (test/tests/matmul_test.cpp:855)	called 440 times	100.0%	-%	38.0%
kai::test::MatMulTest_PackedTransposedRhs_Test::~MatMulTest_PackedTransposedRhs_Test() (test/tests/matmul_test.cpp:855)	called 1320 times	97.6%	35.1%	100.0%
kai::test::MatMulTest_bf16_qai8dxp_qsi4cxp::MatMulTest_bf16_qai8dxp_qsi4cxp() (test/tests/matmul_clamp_bf16_qai8dxp_qsi4cxp_test.cpp:57)	called 784 times	100.0%	-%	100.0%
kai::test::MatMulTest_bf16_qai8dxp_qsi4cxp::~MatMulTest_bf16_qai8dxp_qsi4cxp() (test/tests/matmul_clamp_bf16_qai8dxp_qsi4cxp_test.cpp:57)	called 784 times	100.0%	-%	100.0%
kai::test::MatMulTest_bf16_qai8dxp_qsi4cxp_EndToEnd_RHS_KxN_Test::AddToRegistry() (test/tests/matmul_clamp_bf16_qai8dxp_qsi4cxp_test.cpp:189)	called 1 time	100.0%	-%	44.0%
kai::test::MatMulTest_bf16_qai8dxp_qsi4cxp_EndToEnd_RHS_KxN_Test::MatMulTest_bf16_qai8dxp_qsi4cxp_EndToEnd_RHS_KxN_Test() (test/tests/matmul_clamp_bf16_qai8dxp_qsi4cxp_test.cpp:189)	called 784 times	100.0%	-%	100.0%
kai::test::MatMulTest_bf16_qai8dxp_qsi4cxp_EndToEnd_RHS_KxN_Test::TestBody() (test/tests/matmul_clamp_bf16_qai8dxp_qsi4cxp_test.cpp:189)	called 392 times	100.0%	-%	43.0%
kai::test::MatMulTest_bf16_qai8dxp_qsi4cxp_EndToEnd_RHS_KxN_Test::~MatMulTest_bf16_qai8dxp_qsi4cxp_EndToEnd_RHS_KxN_Test() (test/tests/matmul_clamp_bf16_qai8dxp_qsi4cxp_test.cpp:189)	called 1176 times	98.9%	42.3%	100.0%
kai::test::MatMulTest_bf16_qai8dxp_qsi4cxp_EndToEnd_RHS_NxK_Test::AddToRegistry() (test/tests/matmul_clamp_bf16_qai8dxp_qsi4cxp_test.cpp:59)	called 1 time	100.0%	-%	44.0%
kai::test::MatMulTest_bf16_qai8dxp_qsi4cxp_EndToEnd_RHS_NxK_Test::MatMulTest_bf16_qai8dxp_qsi4cxp_EndToEnd_RHS_NxK_Test() (test/tests/matmul_clamp_bf16_qai8dxp_qsi4cxp_test.cpp:59)	called 784 times	100.0%	-%	100.0%
kai::test::MatMulTest_bf16_qai8dxp_qsi4cxp_EndToEnd_RHS_NxK_Test::TestBody() (test/tests/matmul_clamp_bf16_qai8dxp_qsi4cxp_test.cpp:59)	called 392 times	100.0%	-%	43.0%
kai::test::MatMulTest_bf16_qai8dxp_qsi4cxp_EndToEnd_RHS_NxK_Test::~MatMulTest_bf16_qai8dxp_qsi4cxp_EndToEnd_RHS_NxK_Test() (test/tests/matmul_clamp_bf16_qai8dxp_qsi4cxp_test.cpp:59)	called 1176 times	98.8%	41.7%	100.0%
kai::test::MatMulTest_f16_qai8dxp_qsi4cxp::MatMulTest_f16_qai8dxp_qsi4cxp() (test/tests/matmul_clamp_f16_qai8dxp_qsi4cxp_test.cpp:57)	called 896 times	100.0%	-%	100.0%
kai::test::MatMulTest_f16_qai8dxp_qsi4cxp::~MatMulTest_f16_qai8dxp_qsi4cxp() (test/tests/matmul_clamp_f16_qai8dxp_qsi4cxp_test.cpp:57)	called 896 times	100.0%	-%	100.0%
kai::test::MatMulTest_f16_qai8dxp_qsi4cxp_EndToEnd_Test::AddToRegistry() (test/tests/matmul_clamp_f16_qai8dxp_qsi4cxp_test.cpp:59)	called 1 time	100.0%	-%	44.0%
kai::test::MatMulTest_f16_qai8dxp_qsi4cxp_EndToEnd_Test::MatMulTest_f16_qai8dxp_qsi4cxp_EndToEnd_Test() (test/tests/matmul_clamp_f16_qai8dxp_qsi4cxp_test.cpp:59)	called 1792 times	100.0%	-%	100.0%
kai::test::MatMulTest_f16_qai8dxp_qsi4cxp_EndToEnd_Test::TestBody() (test/tests/matmul_clamp_f16_qai8dxp_qsi4cxp_test.cpp:59)	called 896 times	100.0%	-%	43.0%
kai::test::MatMulTest_f16_qai8dxp_qsi4cxp_EndToEnd_Test::~MatMulTest_f16_qai8dxp_qsi4cxp_EndToEnd_Test() (test/tests/matmul_clamp_f16_qai8dxp_qsi4cxp_test.cpp:59)	called 2688 times	98.8%	42.6%	100.0%
kai::test::MatMulTest_f16_qai8dxp_qsi8cxp::MatMulTest_f16_qai8dxp_qsi8cxp() (test/tests/matmul_clamp_f16_qai8dxp_qsi8cxp_test.cpp:58)	called 896 times	100.0%	-%	100.0%
kai::test::MatMulTest_f16_qai8dxp_qsi8cxp::~MatMulTest_f16_qai8dxp_qsi8cxp() (test/tests/matmul_clamp_f16_qai8dxp_qsi8cxp_test.cpp:58)	called 896 times	100.0%	-%	100.0%
kai::test::MatMulTest_f16_qai8dxp_qsi8cxp_EndToEnd_Test::AddToRegistry() (test/tests/matmul_clamp_f16_qai8dxp_qsi8cxp_test.cpp:60)	called 1 time	100.0%	-%	44.0%
kai::test::MatMulTest_f16_qai8dxp_qsi8cxp_EndToEnd_Test::MatMulTest_f16_qai8dxp_qsi8cxp_EndToEnd_Test() (test/tests/matmul_clamp_f16_qai8dxp_qsi8cxp_test.cpp:60)	called 1792 times	100.0%	-%	100.0%
kai::test::MatMulTest_f16_qai8dxp_qsi8cxp_EndToEnd_Test::TestBody() (test/tests/matmul_clamp_f16_qai8dxp_qsi8cxp_test.cpp:60)	called 896 times	100.0%	-%	42.0%
kai::test::MatMulTest_f16_qai8dxp_qsi8cxp_EndToEnd_Test::~MatMulTest_f16_qai8dxp_qsi8cxp_EndToEnd_Test() (test/tests/matmul_clamp_f16_qai8dxp_qsi8cxp_test.cpp:60)	called 2688 times	98.8%	42.4%	100.0%
kai::test::MatMulTest_f16_qsi8d32p_qai4c32p::MatMulTest_f16_qsi8d32p_qai4c32p() (test/tests/matmul_clamp_f16_qsi8d32p_qai4c32p_test.cpp:188)	called 11200 times	100.0%	-%	100.0%
kai::test::MatMulTest_f16_qsi8d32p_qai4c32p::~MatMulTest_f16_qsi8d32p_qai4c32p() (test/tests/matmul_clamp_f16_qsi8d32p_qai4c32p_test.cpp:188)	called 11200 times	100.0%	-%	100.0%
kai::test::MatMulTest_f16_qsi8d32p_qai4c32p_EndToEnd_Test::AddToRegistry() (test/tests/matmul_clamp_f16_qsi8d32p_qai4c32p_test.cpp:258)	called 1 time	100.0%	-%	44.0%
kai::test::MatMulTest_f16_qsi8d32p_qai4c32p_EndToEnd_Test::MatMulTest_f16_qsi8d32p_qai4c32p_EndToEnd_Test() (test/tests/matmul_clamp_f16_qsi8d32p_qai4c32p_test.cpp:258)	called 11200 times	100.0%	-%	100.0%
kai::test::MatMulTest_f16_qsi8d32p_qai4c32p_EndToEnd_Test::TestBody() (test/tests/matmul_clamp_f16_qsi8d32p_qai4c32p_test.cpp:258)	called 5600 times	100.0%	-%	42.0%
kai::test::MatMulTest_f16_qsi8d32p_qai4c32p_EndToEnd_Test::~MatMulTest_f16_qsi8d32p_qai4c32p_EndToEnd_Test() (test/tests/matmul_clamp_f16_qsi8d32p_qai4c32p_test.cpp:258)	called 16800 times	98.8%	43.4%	100.0%
kai::test::MatMulTest_f16_qsi8d32p_qai4c32p_LhsPackedWithSameBlockdepth_Test::AddToRegistry() (test/tests/matmul_clamp_f16_qsi8d32p_qai4c32p_test.cpp:190)	called 1 time	100.0%	-%	44.0%
kai::test::MatMulTest_f16_qsi8d32p_qai4c32p_LhsPackedWithSameBlockdepth_Test::MatMulTest_f16_qsi8d32p_qai4c32p_LhsPackedWithSameBlockdepth_Test() (test/tests/matmul_clamp_f16_qsi8d32p_qai4c32p_test.cpp:190)	called 11200 times	100.0%	-%	100.0%
kai::test::MatMulTest_f16_qsi8d32p_qai4c32p_LhsPackedWithSameBlockdepth_Test::TestBody() (test/tests/matmul_clamp_f16_qsi8d32p_qai4c32p_test.cpp:190)	called 5600 times	100.0%	-%	36.0%
kai::test::MatMulTest_f16_qsi8d32p_qai4c32p_LhsPackedWithSameBlockdepth_Test::~MatMulTest_f16_qsi8d32p_qai4c32p_LhsPackedWithSameBlockdepth_Test() (test/tests/matmul_clamp_f16_qsi8d32p_qai4c32p_test.cpp:190)	called 16800 times	97.6%	37.9%	100.0%
kai::test::MatMulTest_f32_qai8dxp_qsi4cxp::MatMulTest_f32_qai8dxp_qsi4cxp() (test/tests/matmul_clamp_f32_qai8dxp_qsi4cxp_test.cpp:205)	called 4800 times	100.0%	-%	100.0%
kai::test::MatMulTest_f32_qai8dxp_qsi4cxp::~MatMulTest_f32_qai8dxp_qsi4cxp() (test/tests/matmul_clamp_f32_qai8dxp_qsi4cxp_test.cpp:205)	called 4800 times	100.0%	-%	100.0%
kai::test::MatMulTest_f32_qai8dxp_qsi4cxp_EndToEnd_RHS_kxn_qsi4cx_Test::AddToRegistry() (test/tests/matmul_clamp_f32_qai8dxp_qsi4cxp_test.cpp:530)	called 1 time	100.0%	-%	44.0%
kai::test::MatMulTest_f32_qai8dxp_qsi4cxp_EndToEnd_RHS_kxn_qsi4cx_Test::MatMulTest_f32_qai8dxp_qsi4cxp_EndToEnd_RHS_kxn_qsi4cx_Test() (test/tests/matmul_clamp_f32_qai8dxp_qsi4cxp_test.cpp:530)	called 1600 times	100.0%	-%	100.0%
kai::test::MatMulTest_f32_qai8dxp_qsi4cxp_EndToEnd_RHS_kxn_qsi4cx_Test::TestBody() (test/tests/matmul_clamp_f32_qai8dxp_qsi4cxp_test.cpp:530)	called 800 times	100.0%	-%	39.0%
kai::test::MatMulTest_f32_qai8dxp_qsi4cxp_EndToEnd_RHS_kxn_qsi4cx_Test::TestBody()::$_0::operator()(unsigned long, unsigned long) const (test/tests/matmul_clamp_f32_qai8dxp_qsi4cxp_test.cpp:557)	called 1546560 times	95.9%	36.5%	100.0%
kai::test::MatMulTest_f32_qai8dxp_qsi4cxp_EndToEnd_RHS_kxn_qsi4cx_Test::~MatMulTest_f32_qai8dxp_qsi4cxp_EndToEnd_RHS_kxn_qsi4cx_Test() (test/tests/matmul_clamp_f32_qai8dxp_qsi4cxp_test.cpp:530)	called 2400 times	94.4%	44.4%	100.0%
kai::test::MatMulTest_f32_qai8dxp_qsi4cxp_EndToEnd_RHS_kxn_qsu4cx_Test::AddToRegistry() (test/tests/matmul_clamp_f32_qai8dxp_qsi4cxp_test.cpp:665)	called 1 time	100.0%	-%	44.0%
kai::test::MatMulTest_f32_qai8dxp_qsi4cxp_EndToEnd_RHS_kxn_qsu4cx_Test::MatMulTest_f32_qai8dxp_qsi4cxp_EndToEnd_RHS_kxn_qsu4cx_Test() (test/tests/matmul_clamp_f32_qai8dxp_qsi4cxp_test.cpp:665)	called 1600 times	100.0%	-%	100.0%
kai::test::MatMulTest_f32_qai8dxp_qsi4cxp_EndToEnd_RHS_kxn_qsu4cx_Test::TestBody() (test/tests/matmul_clamp_f32_qai8dxp_qsi4cxp_test.cpp:665)	called 800 times	100.0%	-%	38.0%
kai::test::MatMulTest_f32_qai8dxp_qsi4cxp_EndToEnd_RHS_kxn_qsu4cx_Test::TestBody()::$_0::operator()(unsigned long, unsigned long) const (test/tests/matmul_clamp_f32_qai8dxp_qsi4cxp_test.cpp:692)	called 1546560 times	98.5%	36.2%	100.0%
kai::test::MatMulTest_f32_qai8dxp_qsi4cxp_EndToEnd_RHS_kxn_qsu4cx_Test::~MatMulTest_f32_qai8dxp_qsi4cxp_EndToEnd_RHS_kxn_qsu4cx_Test() (test/tests/matmul_clamp_f32_qai8dxp_qsi4cxp_test.cpp:665)	called 2400 times	94.4%	44.4%	100.0%
kai::test::MatMulTest_f32_qai8dxp_qsi4cxp_EndToEnd_RHS_nxk_qsi4cx_Test::AddToRegistry() (test/tests/matmul_clamp_f32_qai8dxp_qsi4cxp_test.cpp:268)	called 1 time	100.0%	-%	44.0%
kai::test::MatMulTest_f32_qai8dxp_qsi4cxp_EndToEnd_RHS_nxk_qsi4cx_Test::MatMulTest_f32_qai8dxp_qsi4cxp_EndToEnd_RHS_nxk_qsi4cx_Test() (test/tests/matmul_clamp_f32_qai8dxp_qsi4cxp_test.cpp:268)	called 1600 times	100.0%	-%	100.0%
kai::test::MatMulTest_f32_qai8dxp_qsi4cxp_EndToEnd_RHS_nxk_qsi4cx_Test::TestBody() (test/tests/matmul_clamp_f32_qai8dxp_qsi4cxp_test.cpp:268)	called 800 times	100.0%	-%	40.0%
kai::test::MatMulTest_f32_qai8dxp_qsi4cxp_EndToEnd_RHS_nxk_qsi4cx_Test::TestBody()::$_0::operator()(unsigned long, unsigned long) const (test/tests/matmul_clamp_f32_qai8dxp_qsi4cxp_test.cpp:296)	called 1890240 times	95.8%	36.4%	100.0%
kai::test::MatMulTest_f32_qai8dxp_qsi4cxp_EndToEnd_RHS_nxk_qsi4cx_Test::~MatMulTest_f32_qai8dxp_qsi4cxp_EndToEnd_RHS_nxk_qsi4cx_Test() (test/tests/matmul_clamp_f32_qai8dxp_qsi4cxp_test.cpp:268)	called 2400 times	94.7%	44.7%	100.0%
kai::test::MatMulTest_f32_qai8dxp_qsi4cxp_EndToEnd_RHS_nxk_qsu4cx_Test::AddToRegistry() (test/tests/matmul_clamp_f32_qai8dxp_qsi4cxp_test.cpp:399)	called 1 time	100.0%	-%	44.0%
kai::test::MatMulTest_f32_qai8dxp_qsi4cxp_EndToEnd_RHS_nxk_qsu4cx_Test::MatMulTest_f32_qai8dxp_qsi4cxp_EndToEnd_RHS_nxk_qsu4cx_Test() (test/tests/matmul_clamp_f32_qai8dxp_qsi4cxp_test.cpp:399)	called 1600 times	100.0%	-%	100.0%
kai::test::MatMulTest_f32_qai8dxp_qsi4cxp_EndToEnd_RHS_nxk_qsu4cx_Test::TestBody() (test/tests/matmul_clamp_f32_qai8dxp_qsi4cxp_test.cpp:399)	called 800 times	100.0%	-%	40.0%
kai::test::MatMulTest_f32_qai8dxp_qsi4cxp_EndToEnd_RHS_nxk_qsu4cx_Test::TestBody()::$_0::operator()(unsigned long, unsigned long) const (test/tests/matmul_clamp_f32_qai8dxp_qsi4cxp_test.cpp:426)	called 1890240 times	95.9%	36.8%	100.0%
kai::test::MatMulTest_f32_qai8dxp_qsi4cxp_EndToEnd_RHS_nxk_qsu4cx_Test::~MatMulTest_f32_qai8dxp_qsi4cxp_EndToEnd_RHS_nxk_qsu4cx_Test() (test/tests/matmul_clamp_f32_qai8dxp_qsi4cxp_test.cpp:399)	called 2400 times	94.4%	44.4%	100.0%
kai::test::MatMulTest_f32_qai8dxp_qsi4cxp_Offset_LHS_Test::AddToRegistry() (test/tests/matmul_clamp_f32_qai8dxp_qsi4cxp_test.cpp:237)	called 1 time	100.0%	-%	44.0%
kai::test::MatMulTest_f32_qai8dxp_qsi4cxp_Offset_LHS_Test::MatMulTest_f32_qai8dxp_qsi4cxp_Offset_LHS_Test() (test/tests/matmul_clamp_f32_qai8dxp_qsi4cxp_test.cpp:237)	called 1600 times	100.0%	-%	100.0%
kai::test::MatMulTest_f32_qai8dxp_qsi4cxp_Offset_LHS_Test::TestBody() (test/tests/matmul_clamp_f32_qai8dxp_qsi4cxp_test.cpp:237)	called 800 times	100.0%	-%	23.0%
kai::test::MatMulTest_f32_qai8dxp_qsi4cxp_Offset_LHS_Test::~MatMulTest_f32_qai8dxp_qsi4cxp_Offset_LHS_Test() (test/tests/matmul_clamp_f32_qai8dxp_qsi4cxp_test.cpp:237)	called 2400 times	90.5%	23.3%	100.0%
kai::test::MatMulTest_f32_qai8dxp_qsi4cxp_Offset_RHS_Test::AddToRegistry() (test/tests/matmul_clamp_f32_qai8dxp_qsi4cxp_test.cpp:207)	called 1 time	100.0%	-%	44.0%
kai::test::MatMulTest_f32_qai8dxp_qsi4cxp_Offset_RHS_Test::MatMulTest_f32_qai8dxp_qsi4cxp_Offset_RHS_Test() (test/tests/matmul_clamp_f32_qai8dxp_qsi4cxp_test.cpp:207)	called 1600 times	100.0%	-%	100.0%
kai::test::MatMulTest_f32_qai8dxp_qsi4cxp_Offset_RHS_Test::TestBody() (test/tests/matmul_clamp_f32_qai8dxp_qsi4cxp_test.cpp:207)	called 800 times	100.0%	-%	23.0%
kai::test::MatMulTest_f32_qai8dxp_qsi4cxp_Offset_RHS_Test::~MatMulTest_f32_qai8dxp_qsi4cxp_Offset_RHS_Test() (test/tests/matmul_clamp_f32_qai8dxp_qsi4cxp_test.cpp:207)	called 2400 times	90.5%	23.3%	100.0%
kai::test::MatMulTest_f32_qai8dxp_qsi8cxp::MatMulTest_f32_qai8dxp_qsi8cxp() (test/tests/matmul_clamp_f32_qai8dxp_qsi8cxp_test.cpp:91)	called 2464 times	100.0%	-%	100.0%
kai::test::MatMulTest_f32_qai8dxp_qsi8cxp::~MatMulTest_f32_qai8dxp_qsi8cxp() (test/tests/matmul_clamp_f32_qai8dxp_qsi8cxp_test.cpp:91)	called 2464 times	100.0%	-%	100.0%
kai::test::MatMulTest_f32_qai8dxp_qsi8cxp_EndToEnd_RHS_kxn_qsi8cx_Test::AddToRegistry() (test/tests/matmul_clamp_f32_qai8dxp_qsi8cxp_test.cpp:253)	called 1 time	100.0%	-%	44.0%
kai::test::MatMulTest_f32_qai8dxp_qsi8cxp_EndToEnd_RHS_kxn_qsi8cx_Test::MatMulTest_f32_qai8dxp_qsi8cxp_EndToEnd_RHS_kxn_qsi8cx_Test() (test/tests/matmul_clamp_f32_qai8dxp_qsi8cxp_test.cpp:253)	called 1232 times	100.0%	-%	100.0%
kai::test::MatMulTest_f32_qai8dxp_qsi8cxp_EndToEnd_RHS_kxn_qsi8cx_Test::TestBody() (test/tests/matmul_clamp_f32_qai8dxp_qsi8cxp_test.cpp:253)	called 616 times	100.0%	-%	42.0%
kai::test::MatMulTest_f32_qai8dxp_qsi8cxp_EndToEnd_RHS_kxn_qsi8cx_Test::~MatMulTest_f32_qai8dxp_qsi8cxp_EndToEnd_RHS_kxn_qsi8cx_Test() (test/tests/matmul_clamp_f32_qai8dxp_qsi8cxp_test.cpp:253)	called 1848 times	96.5%	39.3%	100.0%
kai::test::MatMulTest_f32_qai8dxp_qsi8cxp_EndToEnd_RHS_nxk_qsi8cx_Test::AddToRegistry() (test/tests/matmul_clamp_f32_qai8dxp_qsi8cxp_test.cpp:138)	called 1 time	100.0%	-%	44.0%
kai::test::MatMulTest_f32_qai8dxp_qsi8cxp_EndToEnd_RHS_nxk_qsi8cx_Test::MatMulTest_f32_qai8dxp_qsi8cxp_EndToEnd_RHS_nxk_qsi8cx_Test() (test/tests/matmul_clamp_f32_qai8dxp_qsi8cxp_test.cpp:138)	called 1232 times	100.0%	-%	100.0%
kai::test::MatMulTest_f32_qai8dxp_qsi8cxp_EndToEnd_RHS_nxk_qsi8cx_Test::TestBody() (test/tests/matmul_clamp_f32_qai8dxp_qsi8cxp_test.cpp:138)	called 616 times	100.0%	-%	42.0%
kai::test::MatMulTest_f32_qai8dxp_qsi8cxp_EndToEnd_RHS_nxk_qsi8cx_Test::~MatMulTest_f32_qai8dxp_qsi8cxp_EndToEnd_RHS_nxk_qsi8cx_Test() (test/tests/matmul_clamp_f32_qai8dxp_qsi8cxp_test.cpp:138)	called 1848 times	96.2%	39.3%	100.0%
kai::test::MatMulTest_f32_qai8dxp_qsi8cxp_Offset_LHS_Test::AddToRegistry() (test/tests/matmul_clamp_f32_qai8dxp_qsi8cxp_test.cpp:117)	called 1 time	100.0%	-%	44.0%
kai::test::MatMulTest_f32_qai8dxp_qsi8cxp_Offset_LHS_Test::MatMulTest_f32_qai8dxp_qsi8cxp_Offset_LHS_Test() (test/tests/matmul_clamp_f32_qai8dxp_qsi8cxp_test.cpp:117)	called 1232 times	100.0%	-%	100.0%
kai::test::MatMulTest_f32_qai8dxp_qsi8cxp_Offset_LHS_Test::TestBody() (test/tests/matmul_clamp_f32_qai8dxp_qsi8cxp_test.cpp:117)	called 616 times	100.0%	-%	30.0%
kai::test::MatMulTest_f32_qai8dxp_qsi8cxp_Offset_LHS_Test::~MatMulTest_f32_qai8dxp_qsi8cxp_Offset_LHS_Test() (test/tests/matmul_clamp_f32_qai8dxp_qsi8cxp_test.cpp:117)	called 1848 times	92.9%	31.6%	100.0%
kai::test::MatMulTest_f32_qai8dxp_qsi8cxp_Offset_RHS_Test::AddToRegistry() (test/tests/matmul_clamp_f32_qai8dxp_qsi8cxp_test.cpp:93)	called 1 time	100.0%	-%	44.0%
kai::test::MatMulTest_f32_qai8dxp_qsi8cxp_Offset_RHS_Test::MatMulTest_f32_qai8dxp_qsi8cxp_Offset_RHS_Test() (test/tests/matmul_clamp_f32_qai8dxp_qsi8cxp_test.cpp:93)	called 1232 times	100.0%	-%	100.0%
kai::test::MatMulTest_f32_qai8dxp_qsi8cxp_Offset_RHS_Test::TestBody() (test/tests/matmul_clamp_f32_qai8dxp_qsi8cxp_test.cpp:93)	called 616 times	100.0%	-%	30.0%
kai::test::MatMulTest_f32_qai8dxp_qsi8cxp_Offset_RHS_Test::~MatMulTest_f32_qai8dxp_qsi8cxp_Offset_RHS_Test() (test/tests/matmul_clamp_f32_qai8dxp_qsi8cxp_test.cpp:93)	called 1848 times	93.8%	28.8%	100.0%
kai::test::MatMulTest_f32_qsi8d32p_qai4c32p::MatMulTest_f32_qsi8d32p_qai4c32p() (test/tests/matmul_clamp_f32_qsi8d32p_qai4c32p_test.cpp:146)	called 11200 times	100.0%	-%	100.0%
kai::test::MatMulTest_f32_qsi8d32p_qai4c32p::~MatMulTest_f32_qsi8d32p_qai4c32p() (test/tests/matmul_clamp_f32_qsi8d32p_qai4c32p_test.cpp:146)	called 11200 times	100.0%	-%	100.0%
kai::test::MatMulTest_f32_qsi8d32p_qai4c32p_EndToEnd_Test::AddToRegistry() (test/tests/matmul_clamp_f32_qsi8d32p_qai4c32p_test.cpp:213)	called 1 time	100.0%	-%	44.0%
kai::test::MatMulTest_f32_qsi8d32p_qai4c32p_EndToEnd_Test::MatMulTest_f32_qsi8d32p_qai4c32p_EndToEnd_Test() (test/tests/matmul_clamp_f32_qsi8d32p_qai4c32p_test.cpp:213)	called 11200 times	100.0%	-%	100.0%
kai::test::MatMulTest_f32_qsi8d32p_qai4c32p_EndToEnd_Test::TestBody() (test/tests/matmul_clamp_f32_qsi8d32p_qai4c32p_test.cpp:213)	called 5600 times	100.0%	-%	42.0%
kai::test::MatMulTest_f32_qsi8d32p_qai4c32p_EndToEnd_Test::~MatMulTest_f32_qsi8d32p_qai4c32p_EndToEnd_Test() (test/tests/matmul_clamp_f32_qsi8d32p_qai4c32p_test.cpp:213)	called 16800 times	98.8%	43.1%	100.0%
kai::test::MatMulTest_f32_qsi8d32p_qai4c32p_LhsPackedWithSameBlockdepth_Test::AddToRegistry() (test/tests/matmul_clamp_f32_qsi8d32p_qai4c32p_test.cpp:148)	called 1 time	100.0%	-%	44.0%
kai::test::MatMulTest_f32_qsi8d32p_qai4c32p_LhsPackedWithSameBlockdepth_Test::MatMulTest_f32_qsi8d32p_qai4c32p_LhsPackedWithSameBlockdepth_Test() (test/tests/matmul_clamp_f32_qsi8d32p_qai4c32p_test.cpp:148)	called 11200 times	100.0%	-%	100.0%
kai::test::MatMulTest_f32_qsi8d32p_qai4c32p_LhsPackedWithSameBlockdepth_Test::TestBody() (test/tests/matmul_clamp_f32_qsi8d32p_qai4c32p_test.cpp:148)	called 5600 times	100.0%	-%	36.0%
kai::test::MatMulTest_f32_qsi8d32p_qai4c32p_LhsPackedWithSameBlockdepth_Test::~MatMulTest_f32_qsi8d32p_qai4c32p_LhsPackedWithSameBlockdepth_Test() (test/tests/matmul_clamp_f32_qsi8d32p_qai4c32p_test.cpp:148)	called 16800 times	97.4%	37.7%	100.0%
kai::test::MatMulTest_f32_qsi8d32p_qsi4c32p::MatMulTest_f32_qsi8d32p_qsi4c32p() (test/tests/matmul_clamp_f32_qsi8d32p_qsi4c32p_test.cpp:103)	called 576 times	100.0%	-%	100.0%
kai::test::MatMulTest_f32_qsi8d32p_qsi4c32p::~MatMulTest_f32_qsi8d32p_qsi4c32p() (test/tests/matmul_clamp_f32_qsi8d32p_qsi4c32p_test.cpp:103)	called 576 times	100.0%	-%	100.0%
kai::test::MatMulTest_f32_qsi8d32p_qsi4c32p_EndToEnd_Test::AddToRegistry() (test/tests/matmul_clamp_f32_qsi8d32p_qsi4c32p_test.cpp:167)	called 1 time	100.0%	-%	44.0%
kai::test::MatMulTest_f32_qsi8d32p_qsi4c32p_EndToEnd_Test::MatMulTest_f32_qsi8d32p_qsi4c32p_EndToEnd_Test() (test/tests/matmul_clamp_f32_qsi8d32p_qsi4c32p_test.cpp:167)	called 384 times	100.0%	-%	100.0%
kai::test::MatMulTest_f32_qsi8d32p_qsi4c32p_EndToEnd_Test::TestBody() (test/tests/matmul_clamp_f32_qsi8d32p_qsi4c32p_test.cpp:167)	called 192 times	100.0%	-%	42.0%
kai::test::MatMulTest_f32_qsi8d32p_qsi4c32p_EndToEnd_Test::~MatMulTest_f32_qsi8d32p_qsi4c32p_EndToEnd_Test() (test/tests/matmul_clamp_f32_qsi8d32p_qsi4c32p_test.cpp:167)	called 576 times	96.3%	39.8%	100.0%
kai::test::MatMulTest_f32_qsi8d32p_qsi4c32p_Offset_LHS_Test::AddToRegistry() (test/tests/matmul_clamp_f32_qsi8d32p_qsi4c32p_test.cpp:135)	called 1 time	100.0%	-%	44.0%
kai::test::MatMulTest_f32_qsi8d32p_qsi4c32p_Offset_LHS_Test::MatMulTest_f32_qsi8d32p_qsi4c32p_Offset_LHS_Test() (test/tests/matmul_clamp_f32_qsi8d32p_qsi4c32p_test.cpp:135)	called 384 times	100.0%	-%	100.0%
kai::test::MatMulTest_f32_qsi8d32p_qsi4c32p_Offset_LHS_Test::TestBody() (test/tests/matmul_clamp_f32_qsi8d32p_qsi4c32p_test.cpp:135)	called 192 times	100.0%	-%	40.0%
kai::test::MatMulTest_f32_qsi8d32p_qsi4c32p_Offset_LHS_Test::~MatMulTest_f32_qsi8d32p_qsi4c32p_Offset_LHS_Test() (test/tests/matmul_clamp_f32_qsi8d32p_qsi4c32p_test.cpp:135)	called 576 times	95.5%	40.0%	100.0%
kai::test::MatMulTest_f32_qsi8d32p_qsi4c32p_Offset_RHS_Test::AddToRegistry() (test/tests/matmul_clamp_f32_qsi8d32p_qsi4c32p_test.cpp:105)	called 1 time	100.0%	-%	44.0%
kai::test::MatMulTest_f32_qsi8d32p_qsi4c32p_Offset_RHS_Test::MatMulTest_f32_qsi8d32p_qsi4c32p_Offset_RHS_Test() (test/tests/matmul_clamp_f32_qsi8d32p_qsi4c32p_test.cpp:105)	called 384 times	100.0%	-%	100.0%
kai::test::MatMulTest_f32_qsi8d32p_qsi4c32p_Offset_RHS_Test::TestBody() (test/tests/matmul_clamp_f32_qsi8d32p_qsi4c32p_test.cpp:105)	called 192 times	100.0%	-%	40.0%
kai::test::MatMulTest_f32_qsi8d32p_qsi4c32p_Offset_RHS_Test::~MatMulTest_f32_qsi8d32p_qsi4c32p_Offset_RHS_Test() (test/tests/matmul_clamp_f32_qsi8d32p_qsi4c32p_test.cpp:105)	called 576 times	95.2%	40.0%	100.0%
kai::test::MatMulTest_qmatmul_clamp_bf16_qai8dxp_qsi4c32p::MatMulTest_qmatmul_clamp_bf16_qai8dxp_qsi4c32p() (test/tests/matmul_clamp_qai8dxp_qsi4c32p_test.cpp:242)	called 224 times	100.0%	-%	100.0%
kai::test::MatMulTest_qmatmul_clamp_bf16_qai8dxp_qsi4c32p::~MatMulTest_qmatmul_clamp_bf16_qai8dxp_qsi4c32p() (test/tests/matmul_clamp_qai8dxp_qsi4c32p_test.cpp:242)	called 224 times	100.0%	-%	100.0%
kai::test::MatMulTest_qmatmul_clamp_bf16_qai8dxp_qsi4c32p_EndToEnd_Test::AddToRegistry() (test/tests/matmul_clamp_qai8dxp_qsi4c32p_test.cpp:379)	called 1 time	100.0%	-%	44.0%
kai::test::MatMulTest_qmatmul_clamp_bf16_qai8dxp_qsi4c32p_EndToEnd_Test::MatMulTest_qmatmul_clamp_bf16_qai8dxp_qsi4c32p_EndToEnd_Test() (test/tests/matmul_clamp_qai8dxp_qsi4c32p_test.cpp:379)	called 448 times	100.0%	-%	100.0%
kai::test::MatMulTest_qmatmul_clamp_bf16_qai8dxp_qsi4c32p_EndToEnd_Test::TestBody() (test/tests/matmul_clamp_qai8dxp_qsi4c32p_test.cpp:379)	called 224 times	100.0%	-%	42.0%
kai::test::MatMulTest_qmatmul_clamp_bf16_qai8dxp_qsi4c32p_EndToEnd_Test::~MatMulTest_qmatmul_clamp_bf16_qai8dxp_qsi4c32p_EndToEnd_Test() (test/tests/matmul_clamp_qai8dxp_qsi4c32p_test.cpp:379)	called 672 times	98.9%	39.7%	100.0%
kai::test::MatMulTest_qmatmul_clamp_f32_qai8dxp_qsi4c32p::MatMulTest_qmatmul_clamp_f32_qai8dxp_qsi4c32p() (test/tests/matmul_clamp_qai8dxp_qsi4c32p_test.cpp:240)	called 1232 times	100.0%	-%	100.0%
kai::test::MatMulTest_qmatmul_clamp_f32_qai8dxp_qsi4c32p::~MatMulTest_qmatmul_clamp_f32_qai8dxp_qsi4c32p() (test/tests/matmul_clamp_qai8dxp_qsi4c32p_test.cpp:240)	called 1232 times	100.0%	-%	100.0%
kai::test::MatMulTest_qmatmul_clamp_f32_qai8dxp_qsi4c32p_EndToEnd_Test::AddToRegistry() (test/tests/matmul_clamp_qai8dxp_qsi4c32p_test.cpp:245)	called 1 time	100.0%	-%	44.0%
kai::test::MatMulTest_qmatmul_clamp_f32_qai8dxp_qsi4c32p_EndToEnd_Test::MatMulTest_qmatmul_clamp_f32_qai8dxp_qsi4c32p_EndToEnd_Test() (test/tests/matmul_clamp_qai8dxp_qsi4c32p_test.cpp:245)	called 2464 times	100.0%	-%	100.0%
kai::test::MatMulTest_qmatmul_clamp_f32_qai8dxp_qsi4c32p_EndToEnd_Test::TestBody() (test/tests/matmul_clamp_qai8dxp_qsi4c32p_test.cpp:245)	called 1232 times	100.0%	-%	42.0%
kai::test::MatMulTest_qmatmul_clamp_f32_qai8dxp_qsi4c32p_EndToEnd_Test::~MatMulTest_qmatmul_clamp_f32_qai8dxp_qsi4c32p_EndToEnd_Test() (test/tests/matmul_clamp_qai8dxp_qsi4c32p_test.cpp:245)	called 3696 times	98.9%	39.9%	100.0%
kai::test::MatrixPortion::MatrixPortion(float, float, float, float) (test/common/matrix_portion.cpp:18)	called 476 times	100.0%	-%	100.0%
kai::test::MatrixPortion::compute_portion(unsigned long, unsigned long, unsigned long, unsigned long) const (test/common/matrix_portion.cpp:38)	called 61860 times	100.0%	-%	51.0%
kai::test::MatrixPortion::height() const (test/common/matrix_portion.cpp:30)	called 147108 times	100.0%	-%	100.0%
kai::test::MatrixPortion::start_col() const (test/common/matrix_portion.cpp:26)	called 147108 times	100.0%	-%	100.0%
kai::test::MatrixPortion::start_row() const (test/common/matrix_portion.cpp:22)	called 147108 times	100.0%	-%	100.0%
kai::test::MatrixPortion::width() const (test/common/matrix_portion.cpp:34)	called 147108 times	100.0%	-%	100.0%
kai::test::MismatchHandler::MismatchHandler() (test/common/compare.hpp:36)	called 32956 times	100.0%	-%	100.0%
kai::test::MismatchHandler::~MismatchHandler() (test/common/compare.hpp:39)	called 32956 times	100.0%	-%	100.0%
kai::test::Padding2D::Hash::operator()(kai::test::Padding2D) const (test/reference/dwconv.hpp:23)	called 120 times	100.0%	-%	100.0%
kai::test::PrintTo(kai::test::MatMulShape const&, std::__1::basic_ostream<char, std::__1::char_traits<char>>*) (test/common/matmul_test_common.cpp:26)	called 137904 times	100.0%	-%	100.0%
kai::test::PrintTo(kai::test::MatrixPortion const&, std::__1::basic_ostream<char, std::__1::char_traits<char>>*) (test/common/matmul_test_common.cpp:30)	called 147108 times	100.0%	-%	100.0%
kai::test::PrintTo(kai::test::Padding2D const&, std::__1::basic_ostream<char, std::__1::char_traits<char>>*) (test/reference/dwconv.cpp:18)	called 120 times	100.0%	-%	100.0%
kai::test::PrintTo(std::__1::tuple<kai::test::(anonymous namespace)::Depthwise, kai::test::MatMulShape, kai::test::Padding2D, float> const&, std::__1::basic_ostream<char, std::__1::char_traits<char>>*) (test/tests/dwconv_test.cpp:273)	called 120 times	100.0%	-%	100.0%
kai::test::PrintTo(std::__1::tuple<kai::test::(anonymous namespace)::IndirectMatMul, kai::test::MatMulShape, unsigned long, kai::test::MatrixPortion, float> const&, std::__1::basic_ostream<char, std::__1::char_traits<char>>*) (test/tests/imatmul_test.cpp:598)	called 53856 times	100.0%	-%	100.0%
kai::test::PrintTo(std::__1::tuple<kai::test::MatMulMethod, kai::test::MatMulShape, kai::test::MatrixPortion> const&, std::__1::basic_ostream<char, std::__1::char_traits<char>>*) (test/common/matmul_test_common.cpp:17)	called 4392 times	100.0%	-%	100.0%
kai::test::Rect::Rect(unsigned long, unsigned long, unsigned long, unsigned long) (test/common/rect.cpp:14)	called 123720 times	100.0%	-%	100.0%
kai::test::Rect::contains(unsigned long, unsigned long) const (test/common/rect.cpp:42)	called 116108514 times	100.0%	100.0%	100.0%
kai::test::Rect::end_col() const (test/common/rect.cpp:38)	called 101374232 times	100.0%	-%	100.0%
kai::test::Rect::end_row() const (test/common/rect.cpp:34)	called 124064690 times	100.0%	-%	100.0%
kai::test::Rect::height() const (test/common/rect.cpp:26)	called 203912 times	100.0%	-%	100.0%
kai::test::Rect::start_col() const (test/common/rect.cpp:22)	called 20827906 times	100.0%	-%	100.0%
kai::test::Rect::start_row() const (test/common/rect.cpp:18)	called 25246676 times	100.0%	-%	100.0%
kai::test::Rect::width() const (test/common/rect.cpp:30)	called 3115706 times	100.0%	-%	100.0%
kai::test::ReferenceGenerator<std::__1::tuple<kai::test::MatMulShape, kai::test::DataFormat, kai::test::DataFormat, kai::test::DataFormat>, kai::test::CacheData>::ReferenceGenerator() (test/common/cache.hpp:80)	called 2 times	100.0%	-%	100.0%
kai::test::ReferenceGenerator<std::__1::tuple<kai::test::MatMulShape, kai::test::DataFormat, kai::test::DataFormat, kai::test::DataFormat>, kai::test::CacheData>::generate_reference(std::__1::tuple<kai::test::MatMulShape, kai::test::DataFormat, kai::test::DataFormat, kai::test::DataFormat> const&) (test/tests/matmul_clamp_f32_qai8dxp_qsi8cxp_test.cpp:52)	called 11 times	100.0%	48.1%	44.0%
kai::test::ReferenceGenerator<std::__1::tuple<kai::test::MatMulShape, kai::test::DataFormat, kai::test::DataFormat, kai::test::DataFormat>, kai::test::CacheData>::getRG() (test/common/cache.hpp:66)	called 1232 times	100.0%	75.0%	100.0%
kai::test::ReferenceGenerator<std::__1::tuple<kai::test::MatMulShape, kai::test::DataFormat, kai::test::DataFormat, kai::test::DataFormat>, kai::test::CacheData>::get_test_reference(std::__1::tuple<kai::test::MatMulShape, kai::test::DataFormat, kai::test::DataFormat, kai::test::DataFormat> const&) (test/common/cache.hpp:71)	called 1232 times	100.0%	71.4%	70.0%
kai::test::ReferenceGenerator<std::__1::tuple<kai::test::MatMulShape, kai::test::DataFormat, kai::test::DataFormat, kai::test::DataFormat>, kai::test::CacheData>::~ReferenceGenerator() (test/common/cache.hpp:65)	called 2 times	100.0%	-%	100.0%
kai::test::UInt4 kai::test::(anonymous namespace)::scalar_binary_elementwise<(kai::test::(anonymous namespace)::BinaryElementwiseOperator)0, kai::test::UInt4>(kai::test::UInt4, kai::test::UInt4) (test/reference/binary_elementwise.cpp:42)	not called	100.0%	-%	0.0%
kai::test::UInt4 kai::test::(anonymous namespace)::scalar_binary_elementwise<(kai::test::(anonymous namespace)::BinaryElementwiseOperator)1, kai::test::UInt4>(kai::test::UInt4, kai::test::UInt4) (test/reference/binary_elementwise.cpp:42)	not called	100.0%	-%	0.0%
kai::test::UInt4 kai::test::(anonymous namespace)::scalar_binary_elementwise<(kai::test::(anonymous namespace)::BinaryElementwiseOperator)2, kai::test::UInt4>(kai::test::UInt4, kai::test::UInt4) (test/reference/binary_elementwise.cpp:42)	not called	100.0%	-%	0.0%
kai::test::UInt4 kai::test::(anonymous namespace)::scalar_binary_elementwise<(kai::test::(anonymous namespace)::BinaryElementwiseOperator)3, kai::test::UInt4>(kai::test::UInt4, kai::test::UInt4) (test/reference/binary_elementwise.cpp:42)	not called	80.0%	-%	0.0%
kai::test::UInt4 kai::test::read_array<kai::test::UInt4>(void const*, unsigned long) (test/common/memory.hpp:39)	called 28125312 times	100.0%	100.0%	100.0%
kai::test::UInt4::UInt4(unsigned char) (test/common/int4.hpp:22)	called 555541008 times	100.0%	-%	100.0%
kai::test::UInt4::operator float() const (test/common/int4.cpp:35)	not called	0.0%	-%	0.0%
kai::test::UInt4::operator int() const (test/common/int4.cpp:31)	not called	0.0%	-%	0.0%
kai::test::UInt4::operator*(kai::test::UInt4) const (test/common/int4.cpp:47)	not called	0.0%	-%	0.0%
kai::test::UInt4::operator+(kai::test::UInt4) const (test/common/int4.cpp:39)	not called	0.0%	-%	0.0%
kai::test::UInt4::operator-(kai::test::UInt4) const (test/common/int4.cpp:43)	not called	0.0%	-%	0.0%
kai::test::UInt4::operator/(kai::test::UInt4) const (test/common/int4.cpp:51)	not called	0.0%	-%	0.0%
kai::test::UInt4::operator=(int) (test/common/int4.cpp:25)	not called	0.0%	-%	0.0%
kai::test::UInt4::operator=(unsigned char) (test/common/int4.cpp:19)	not called	0.0%	-%	0.0%
kai::test::UInt4::pack_u8(kai::test::UInt4, kai::test::UInt4) (test/common/int4.cpp:55)	called 85904856 times	100.0%	-%	100.0%
kai::test::UInt4::unpack_u8(unsigned char) (test/common/int4.cpp:59)	called 114030168 times	100.0%	-%	100.0%
kai::test::UkernelVariant<kai_matmul_clamp_bf16_qai8dxp_qsi4c32p_ukernel>::UkernelVariant(kai_matmul_clamp_bf16_qai8dxp_qsi4c32p_ukernel, std::__1::basic_string_view<char, std::__1::char_traits<char>>, std::__1::function<bool ()> const&) (test/common/test_suite.hpp:76)	called 4 times	100.0%	-%	100.0%
kai::test::UkernelVariant<kai_matmul_clamp_bf16_qai8dxp_qsi4c32p_ukernel>::~UkernelVariant() (test/common/test_suite.hpp:64)	not called	100.0%	-%	0.0%
kai::test::UkernelVariant<kai_matmul_clamp_bf16_qai8dxp_qsi4cxp_ukernel>::UkernelVariant(kai_matmul_clamp_bf16_qai8dxp_qsi4cxp_ukernel, std::__1::basic_string_view<char, std::__1::char_traits<char>>, std::__1::function<bool ()> const&) (test/common/test_suite.hpp:76)	called 4 times	100.0%	-%	100.0%
kai::test::UkernelVariant<kai_matmul_clamp_bf16_qai8dxp_qsi4cxp_ukernel>::~UkernelVariant() (test/common/test_suite.hpp:64)	not called	100.0%	-%	0.0%
kai::test::UkernelVariant<kai_matmul_clamp_f16_qai8dxp_qsi4cxp_ukernel>::UkernelVariant(kai_matmul_clamp_f16_qai8dxp_qsi4cxp_ukernel, std::__1::basic_string_view<char, std::__1::char_traits<char>>, std::__1::function<bool ()> const&) (test/common/test_suite.hpp:76)	called 8 times	100.0%	-%	100.0%
kai::test::UkernelVariant<kai_matmul_clamp_f16_qai8dxp_qsi4cxp_ukernel>::~UkernelVariant() (test/common/test_suite.hpp:64)	not called	100.0%	-%	0.0%
kai::test::UkernelVariant<kai_matmul_clamp_f16_qai8dxp_qsi8cxp_ukernel>::UkernelVariant(kai_matmul_clamp_f16_qai8dxp_qsi8cxp_ukernel, std::__1::basic_string_view<char, std::__1::char_traits<char>>, std::__1::function<bool ()> const&) (test/common/test_suite.hpp:76)	called 8 times	100.0%	-%	100.0%
kai::test::UkernelVariant<kai_matmul_clamp_f16_qai8dxp_qsi8cxp_ukernel>::~UkernelVariant() (test/common/test_suite.hpp:64)	not called	100.0%	-%	0.0%
kai::test::UkernelVariant<kai_matmul_clamp_f16_qsi8d32p_qai4c32p_ukernel>::UkernelVariant(kai_matmul_clamp_f16_qsi8d32p_qai4c32p_ukernel, std::__1::basic_string_view<char, std::__1::char_traits<char>>, std::__1::function<bool ()> const&) (test/common/test_suite.hpp:76)	called 16 times	100.0%	-%	100.0%
kai::test::UkernelVariant<kai_matmul_clamp_f16_qsi8d32p_qai4c32p_ukernel>::~UkernelVariant() (test/common/test_suite.hpp:64)	not called	100.0%	-%	0.0%
kai::test::UkernelVariant<kai_matmul_clamp_f32_qai8dxp_qsi4c32p_ukernel>::UkernelVariant(kai_matmul_clamp_f32_qai8dxp_qsi4c32p_ukernel, std::__1::basic_string_view<char, std::__1::char_traits<char>>, std::__1::function<bool ()> const&) (test/common/test_suite.hpp:76)	called 22 times	100.0%	-%	100.0%
kai::test::UkernelVariant<kai_matmul_clamp_f32_qai8dxp_qsi4c32p_ukernel>::~UkernelVariant() (test/common/test_suite.hpp:64)	not called	100.0%	-%	0.0%
kai::test::UkernelVariant<kai_matmul_clamp_f32_qai8dxp_qsi4cxp_ukernel>::UkernelVariant(kai_matmul_clamp_f32_qai8dxp_qsi4cxp_ukernel, std::__1::basic_string_view<char, std::__1::char_traits<char>>, std::__1::function<bool ()> const&) (test/common/test_suite.hpp:76)	called 20 times	100.0%	-%	100.0%
kai::test::UkernelVariant<kai_matmul_clamp_f32_qai8dxp_qsi4cxp_ukernel>::~UkernelVariant() (test/common/test_suite.hpp:64)	not called	100.0%	-%	0.0%
kai::test::UkernelVariant<kai_matmul_clamp_f32_qai8dxp_qsi8cxp_ukernel>::UkernelVariant(kai_matmul_clamp_f32_qai8dxp_qsi8cxp_ukernel, std::__1::basic_string_view<char, std::__1::char_traits<char>>, std::__1::function<bool ()> const&) (test/common/test_suite.hpp:76)	called 16 times	100.0%	-%	100.0%
kai::test::UkernelVariant<kai_matmul_clamp_f32_qai8dxp_qsi8cxp_ukernel>::~UkernelVariant() (test/common/test_suite.hpp:64)	not called	100.0%	-%	0.0%
kai::test::UkernelVariant<kai_matmul_clamp_f32_qsi8d32p_qai4c32p_ukernel>::UkernelVariant(kai_matmul_clamp_f32_qsi8d32p_qai4c32p_ukernel, std::__1::basic_string_view<char, std::__1::char_traits<char>>, std::__1::function<bool ()> const&) (test/common/test_suite.hpp:76)	called 16 times	100.0%	-%	100.0%
kai::test::UkernelVariant<kai_matmul_clamp_f32_qsi8d32p_qai4c32p_ukernel>::~UkernelVariant() (test/common/test_suite.hpp:64)	not called	100.0%	-%	0.0%
kai::test::UkernelVariant<kai_matmul_clamp_f32_qsi8d32p_qsi4c32p_ukernel>::UkernelVariant(kai_matmul_clamp_f32_qsi8d32p_qsi4c32p_ukernel, std::__1::basic_string_view<char, std::__1::char_traits<char>>, std::__1::function<bool ()> const&) (test/common/test_suite.hpp:76)	called 16 times	100.0%	-%	100.0%
kai::test::UkernelVariant<kai_matmul_clamp_f32_qsi8d32p_qsi4c32p_ukernel>::~UkernelVariant() (test/common/test_suite.hpp:64)	not called	100.0%	-%	0.0%
kai::test::UkernelVariantCustom<kai_matmul_clamp_f32_qai8dxp_qsi4cxp_ukernel>::UkernelVariantCustom(kai_matmul_clamp_f32_qai8dxp_qsi4cxp_ukernel, std::__1::basic_string_view<char, std::__1::char_traits<char>>, std::__1::function<bool ()> const&, std::__1::function<void (unsigned long, unsigned long, unsigned long, unsigned long, unsigned long, unsigned long, unsigned char const*, float const*, float const*, void*, unsigned long, kai_rhs_pack_qs4cxs1s0_param const*)>, std::__1::function<unsigned long (unsigned long, unsigned long, unsigned long, unsigned long, unsigned long)>, std::__1::function<unsigned long (unsigned long, unsigned long, unsigned long, unsigned long, unsigned long)>, std::__1::function<unsigned long (unsigned long, unsigned long)>, kai::test::RhsPackType) (test/tests/matmul_clamp_f32_qai8dxp_qsi4cxp_test.cpp:72)	called 40 times	100.0%	49.6%	100.0%
kai::test::UkernelVariantCustom<kai_matmul_clamp_f32_qai8dxp_qsi4cxp_ukernel>::~UkernelVariantCustom() (test/tests/matmul_clamp_f32_qai8dxp_qsi4cxp_test.cpp:63)	not called	100.0%	-%	0.0%
kai::test::add(void const*, kai::test::DataType, unsigned long, unsigned long, void const*, kai::test::DataType, unsigned long, unsigned long) (test/reference/binary_elementwise.cpp:124)	called 2378 times	100.0%	-%	100.0%
kai::test::cast(void const*, kai::test::DataType, kai::test::DataType, unsigned long, unsigned long) (test/reference/cast.cpp:64)	called 148 times	88.9%	50.0%	47.0%
kai::test::cast_qsu4_qsi4(void const*, unsigned long) (test/reference/cast.cpp:78)	called 8262 times	100.0%	58.3%	70.0%
kai::test::clamp(kai::test::DataType, void const*, unsigned long, float, float) (test/reference/clamp.cpp:79)	called 2429 times	100.0%	60.0%	68.0%
kai::test::compare(void const*, void const*, kai::test::DataFormat const&, unsigned long, unsigned long, kai::test::Rect const&, kai::test::MismatchHandler&) (test/common/compare.cpp:230)	called 33658 times	53.3%	38.5%	31.0%
kai::test::convert_s0s1_s1s0(kai::test::Buffer const&) (test/common/int4.cpp:121)	called 3924 times	100.0%	58.3%	63.0%
kai::test::cpu_has_advsimd() (test/common/cpu_info.cpp:238)	called 219 times	100.0%	-%	100.0%
kai::test::cpu_has_bf16() (test/common/cpu_info.cpp:262)	called 1402 times	100.0%	-%	100.0%
kai::test::cpu_has_dotprod() (test/common/cpu_info.cpp:242)	called 14605 times	100.0%	-%	100.0%
kai::test::cpu_has_dotprod_and_bf16() (test/common/cpu_info.cpp:282)	called 506 times	100.0%	50.0%	100.0%
kai::test::cpu_has_dotprod_and_fp16() (test/common/cpu_info.cpp:246)	called 1353 times	100.0%	50.0%	100.0%
kai::test::cpu_has_fp16() (test/common/cpu_info.cpp:258)	called 2024 times	100.0%	-%	100.0%
kai::test::cpu_has_i8mm() (test/common/cpu_info.cpp:250)	called 6662 times	100.0%	-%	100.0%
kai::test::cpu_has_i8mm_and_bf16() (test/common/cpu_info.cpp:286)	called 506 times	100.0%	50.0%	100.0%
kai::test::cpu_has_i8mm_and_fp16() (test/common/cpu_info.cpp:254)	called 451 times	100.0%	50.0%	100.0%
kai::test::cpu_has_sme() (test/common/cpu_info.cpp:274)	called 19726 times	100.0%	-%	100.0%
kai::test::cpu_has_sme2() (test/common/cpu_info.cpp:278)	called 31963 times	100.0%	-%	100.0%
kai::test::cpu_has_sve() (test/common/cpu_info.cpp:266)	not called	0.0%	-%	0.0%
kai::test::cpu_has_sve2() (test/common/cpu_info.cpp:270)	not called	0.0%	-%	0.0%
kai::test::data_type_is_float(kai::test::DataType) (test/common/data_type.cpp:48)	not called	0.0%	-%	0.0%
kai::test::data_type_is_float_bf(kai::test::DataType) (test/common/data_type.cpp:58)	not called	0.0%	-%	0.0%
kai::test::data_type_is_float_fp(kai::test::DataType) (test/common/data_type.cpp:53)	not called	0.0%	-%	0.0%
kai::test::data_type_is_integral(kai::test::DataType) (test/common/data_type.cpp:44)	called 2595 times	100.0%	-%	100.0%
kai::test::data_type_is_quantized(kai::test::DataType) (test/common/data_type.cpp:71)	called 2595 times	100.0%	50.0%	66.0%
kai::test::data_type_is_quantized_asymm(kai::test::DataType) (test/common/data_type.cpp:75)	not called	0.0%	0.0%	0.0%
kai::test::data_type_is_signed(kai::test::DataType) (test/common/data_type.cpp:63)	not called	0.0%	0.0%	0.0%
kai::test::data_type_size_in_bits(kai::test::DataType) (test/common/data_type.cpp:40)	called 67361 times	100.0%	-%	100.0%
kai::test::div(void const*, kai::test::DataType, unsigned long, unsigned long, void const*, kai::test::DataType, unsigned long, unsigned long) (test/reference/binary_elementwise.cpp:173)	not called	0.0%	-%	0.0%
kai::test::fill_matrix_random(unsigned long, unsigned long, kai::test::DataFormat const&, unsigned int) (test/reference/fill.cpp:100)	called 7555 times	50.0%	37.5%	31.0%
kai::test::find_clamp_range(kai::test::DataType, void const*, unsigned long, float) (test/reference/clamp.cpp:51)	called 2429 times	100.0%	100.0%	100.0%
kai::test::get_matmul_methods() (test/tests/matmul_test.cpp:71)	called 4 times	100.0%	75.0%	85.0%
kai::test::get_vecmul_methods() (test/tests/matmul_test.cpp:434)	called 4 times	100.0%	75.0%	85.0%
kai::test::gtest_DepthwiseDepthwisePlanarTest_EvalGenerateName_(testing::TestParamInfo<std::__1::tuple<kai::test::(anonymous namespace)::Depthwise, kai::test::MatMulShape, kai::test::Padding2D, float>> const&) (test/tests/dwconv_test.cpp:284)	called 60 times	100.0%	34.4%	11.0%
kai::test::gtest_DepthwiseDepthwisePlanarTest_EvalGenerator_() (test/tests/dwconv_test.cpp:284)	called 1 time	100.0%	-%	37.0%
kai::test::gtest_IndirectMatMulIndirectMatMulTest_EvalGenerateName_(testing::TestParamInfo<std::__1::tuple<kai::test::(anonymous namespace)::IndirectMatMul, kai::test::MatMulShape, unsigned long, kai::test::MatrixPortion, float>> const&) (test/tests/imatmul_test.cpp:608)	called 26928 times	100.0%	30.0%	7.0%
kai::test::gtest_IndirectMatMulIndirectMatMulTest_EvalGenerator_() (test/tests/imatmul_test.cpp:608)	called 1 time	100.0%	-%	52.0%
kai::test::gtest_MatMulGemmMatMulTestBf16_EvalGenerateName_(testing::TestParamInfo<std::__1::tuple<kai::test::MatMulMethod, kai::test::MatMulShape, kai::test::MatrixPortion>> const&) (test/tests/matmul_clamp_f32_bf16p_bf16p_test.cpp:550)	called 250 times	100.0%	31.2%	11.0%
kai::test::gtest_MatMulGemmMatMulTestBf16_EvalGenerator_() (test/tests/matmul_clamp_f32_bf16p_bf16p_test.cpp:550)	called 1 time	100.0%	-%	62.0%
kai::test::gtest_MatMulGemvMatMulTestBf16_EvalGenerateName_(testing::TestParamInfo<std::__1::tuple<kai::test::MatMulMethod, kai::test::MatMulShape, kai::test::MatrixPortion>> const&) (test/tests/matmul_clamp_f32_bf16p_bf16p_test.cpp:575)	called 96 times	100.0%	31.8%	11.0%
kai::test::gtest_MatMulGemvMatMulTestBf16_EvalGenerator_() (test/tests/matmul_clamp_f32_bf16p_bf16p_test.cpp:575)	called 1 time	100.0%	-%	60.0%
kai::test::gtest_MatMulMatMulTestBf16OutFp16_EvalGenerateName_(testing::TestParamInfo<std::__1::tuple<kai::test::MatMulMethod, kai::test::MatMulShape, kai::test::MatrixPortion>> const&) (test/tests/matmul_clamp_f16_bf16p_bf16p_test.cpp:321)	called 90 times	100.0%	31.2%	11.0%
kai::test::gtest_MatMulMatMulTestBf16OutFp16_EvalGenerator_() (test/tests/matmul_clamp_f16_bf16p_bf16p_test.cpp:321)	called 1 time	100.0%	-%	62.0%
kai::test::gtest_MatMulMatMulTest_EvalGenerateName_(testing::TestParamInfo<std::__1::tuple<kai::test::MatMulMethod, kai::test::MatMulShape, kai::test::MatrixPortion>> const&) (test/tests/matmul_test.cpp:1013)	called 720 times	100.0%	32.5%	12.0%
kai::test::gtest_MatMulMatMulTest_EvalGenerator_() (test/tests/matmul_test.cpp:1013)	called 4 times	100.0%	-%	57.0%
kai::test::gtest_MatMulMatMulTest_bf16_qai8dxp_qsi4cxp_EvalGenerateName_(testing::TestParamInfo<std::__1::tuple<unsigned long, kai::test::MatMulShape, kai::test::MatrixPortion, bool>> const&) (test/tests/matmul_clamp_bf16_qai8dxp_qsi4cxp_test.cpp:328)	called 784 times	100.0%	-%	8.0%
kai::test::gtest_MatMulMatMulTest_bf16_qai8dxp_qsi4cxp_EvalGenerator_() (test/tests/matmul_clamp_bf16_qai8dxp_qsi4cxp_test.cpp:328)	called 2 times	100.0%	-%	56.0%
kai::test::gtest_MatMulMatMulTest_f16_qai8dxp_qsi4cxp_EvalGenerateName_(testing::TestParamInfo<std::__1::tuple<unsigned long, kai::test::MatMulShape, kai::test::MatrixPortion, bool>> const&) (test/tests/matmul_clamp_f16_qai8dxp_qsi4cxp_test.cpp:187)	called 896 times	100.0%	-%	8.0%
kai::test::gtest_MatMulMatMulTest_f16_qai8dxp_qsi4cxp_EvalGenerator_() (test/tests/matmul_clamp_f16_qai8dxp_qsi4cxp_test.cpp:187)	called 1 time	100.0%	-%	56.0%
kai::test::gtest_MatMulMatMulTest_f16_qai8dxp_qsi8cxp_EvalGenerateName_(testing::TestParamInfo<std::__1::tuple<unsigned long, kai::test::MatMulShape, kai::test::MatrixPortion, bool>> const&) (test/tests/matmul_clamp_f16_qai8dxp_qsi8cxp_test.cpp:182)	called 896 times	100.0%	-%	8.0%
kai::test::gtest_MatMulMatMulTest_f16_qai8dxp_qsi8cxp_EvalGenerator_() (test/tests/matmul_clamp_f16_qai8dxp_qsi8cxp_test.cpp:182)	called 1 time	100.0%	-%	56.0%
kai::test::gtest_MatMulMatMulTest_f16_qsi8d32p_qai4c32p_EvalGenerateName_(testing::TestParamInfo<std::__1::tuple<unsigned long, kai::test::MatMulShape, unsigned long, kai::test::MatrixPortion, bool>> const&) (test/tests/matmul_clamp_f16_qsi8d32p_qai4c32p_test.cpp:384)	called 11200 times	100.0%	-%	16.0%
kai::test::gtest_MatMulMatMulTest_f16_qsi8d32p_qai4c32p_EvalGenerator_() (test/tests/matmul_clamp_f16_qsi8d32p_qai4c32p_test.cpp:384)	called 2 times	100.0%	-%	40.0%
kai::test::gtest_MatMulMatMulTest_f32_qai8dxp_qsi4cxp_EvalGenerateName_(testing::TestParamInfo<std::__1::tuple<unsigned long, kai::test::MatMulShape, kai::test::MatrixPortion>> const&) (test/tests/matmul_clamp_f32_qai8dxp_qsi4cxp_test.cpp:794)	called 4800 times	100.0%	-%	11.0%
kai::test::gtest_MatMulMatMulTest_f32_qai8dxp_qsi4cxp_EvalGenerator_() (test/tests/matmul_clamp_f32_qai8dxp_qsi4cxp_test.cpp:794)	called 6 times	100.0%	-%	60.0%
kai::test::gtest_MatMulMatMulTest_f32_qai8dxp_qsi8cxp_EvalGenerateName_(testing::TestParamInfo<std::__1::tuple<unsigned long, kai::test::MatMulShape, kai::test::MatrixPortion>> const&) (test/tests/matmul_clamp_f32_qai8dxp_qsi8cxp_test.cpp:378)	called 2464 times	100.0%	-%	10.0%
kai::test::gtest_MatMulMatMulTest_f32_qai8dxp_qsi8cxp_EvalGenerator_() (test/tests/matmul_clamp_f32_qai8dxp_qsi8cxp_test.cpp:378)	called 4 times	100.0%	-%	66.0%
kai::test::gtest_MatMulMatMulTest_f32_qsi8d32p_qai4c32p_EvalGenerateName_(testing::TestParamInfo<std::__1::tuple<unsigned long, kai::test::MatMulShape, unsigned long, kai::test::MatrixPortion, bool>> const&) (test/tests/matmul_clamp_f32_qsi8d32p_qai4c32p_test.cpp:330)	called 11200 times	100.0%	-%	7.0%
kai::test::gtest_MatMulMatMulTest_f32_qsi8d32p_qai4c32p_EvalGenerator_() (test/tests/matmul_clamp_f32_qsi8d32p_qai4c32p_test.cpp:330)	called 2 times	100.0%	-%	53.0%
kai::test::gtest_MatMulMatMulTest_f32_qsi8d32p_qsi4c32p_EvalGenerateName_(testing::TestParamInfo<std::__1::tuple<unsigned long, kai::test::MatMulShape, kai::test::MatrixPortion>> const&) (test/tests/matmul_clamp_f32_qsi8d32p_qsi4c32p_test.cpp:281)	called 576 times	100.0%	-%	11.0%
kai::test::gtest_MatMulMatMulTest_f32_qsi8d32p_qsi4c32p_EvalGenerator_() (test/tests/matmul_clamp_f32_qsi8d32p_qsi4c32p_test.cpp:281)	called 3 times	100.0%	-%	60.0%
kai::test::gtest_MatMulMatMulTest_qmatmul_clamp_bf16_qai8dxp_qsi4c32p_EvalGenerateName_(testing::TestParamInfo<std::__1::tuple<unsigned long, kai::test::MatMulShape, unsigned long, kai::test::MatrixPortion, kai::test::RhsPackType>> const&) (test/tests/matmul_clamp_qai8dxp_qsi4c32p_test.cpp:540)	called 224 times	100.0%	-%	20.0%
kai::test::gtest_MatMulMatMulTest_qmatmul_clamp_bf16_qai8dxp_qsi4c32p_EvalGenerator_() (test/tests/matmul_clamp_qai8dxp_qsi4c32p_test.cpp:540)	called 1 time	100.0%	-%	50.0%
kai::test::gtest_MatMulMatMulTest_qmatmul_clamp_f32_qai8dxp_qsi4c32p_EvalGenerateName_(testing::TestParamInfo<std::__1::tuple<unsigned long, kai::test::MatMulShape, unsigned long, kai::test::MatrixPortion, kai::test::RhsPackType>> const&) (test/tests/matmul_clamp_qai8dxp_qsi4c32p_test.cpp:518)	called 1232 times	100.0%	-%	20.0%
kai::test::gtest_MatMulMatMulTest_qmatmul_clamp_f32_qai8dxp_qsi4c32p_EvalGenerator_() (test/tests/matmul_clamp_qai8dxp_qsi4c32p_test.cpp:518)	called 1 time	100.0%	-%	50.0%
kai::test::gtest_VecMulMatMulTest_EvalGenerateName_(testing::TestParamInfo<std::__1::tuple<kai::test::MatMulMethod, kai::test::MatMulShape, kai::test::MatrixPortion>> const&) (test/tests/matmul_test.cpp:1032)	called 1040 times	100.0%	31.8%	11.0%
kai::test::gtest_VecMulMatMulTest_EvalGenerator_() (test/tests/matmul_test.cpp:1032)	called 4 times	100.0%	-%	60.0%
kai::test::gtest_indirect_matmul_clamp_qai8_qai8p_qsi8cxpIndirectMatMulQuantizedTest_EvalGenerateName_(testing::TestParamInfo<std::__1::tuple<kai::test::(anonymous namespace)::IndirectMatMulVariant, kai::test::MatMulShape, kai::test::MatrixPortion, unsigned long, float>> const&) (test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp:1095)	called 9324 times	100.0%	-%	7.0%
kai::test::gtest_indirect_matmul_clamp_qai8_qai8p_qsi8cxpIndirectMatMulQuantizedTest_EvalGenerator_() (test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp:1095)	called 1 time	100.0%	-%	52.0%
kai::test::gtest_matmul_clamp_qai8_qai8_qsi8cxpMatMulQuantizedTest_EvalGenerateName_(testing::TestParamInfo<std::__1::tuple<kai::test::(anonymous namespace)::MatMulVariant, kai::test::MatMulShape, kai::test::MatrixPortion, float>> const&) (test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp:1055)	called 168 times	100.0%	-%	9.0%
kai::test::gtest_matmul_clamp_qai8_qai8_qsi8cxpMatMulQuantizedTest_EvalGenerator_() (test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp:1055)	called 1 time	100.0%	-%	50.0%
kai::test::gtest_matmul_clamp_qai8_qai8p_qsi8cxpMatMulQuantizedTest_EvalGenerateName_(testing::TestParamInfo<std::__1::tuple<kai::test::(anonymous namespace)::MatMulVariant, kai::test::MatMulShape, kai::test::MatrixPortion, float>> const&) (test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp:1034)	called 666 times	100.0%	-%	9.0%
kai::test::gtest_matmul_clamp_qai8_qai8p_qsi8cxpMatMulQuantizedTest_EvalGenerator_() (test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp:1034)	called 1 time	100.0%	-%	47.0%
kai::test::imatmul::lhs_pack(kai::test::(anonymous namespace)::LhsPackIndirectKernel const&, kai::test::Rect const&, kai::test::(anonymous namespace)::TestReference const&, unsigned long, kai::test::(anonymous namespace)::KChunk const&) (test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp:881)	called 9324 times	100.0%	50.0%	58.0%
kai::test::imatmul::matmul(kai::test::(anonymous namespace)::MatMulIndirectKernel const&, kai::test::Rect const&, kai::test::(anonymous namespace)::TestReference const&, kai::test::Buffer const&, kai::test::Buffer const&, kai::test::MatMulShape const&, kai::test::(anonymous namespace)::KChunk const&) (test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp:939)	called 9324 times	100.0%	50.0%	62.0%
kai::test::imatmul::rhs_pack(kai::test::(anonymous namespace)::RhsPackIndirectKernel const&, kai::test::Rect const&, kai::test::(anonymous namespace)::TestReference const&, unsigned long, kai::test::(anonymous namespace)::KChunk const&) (test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp:906)	called 9324 times	100.0%	50.0%	57.0%
kai::test::impl::TupleHash<std::__1::tuple<kai::test::MatMulShape, kai::test::DataFormat, kai::test::DataFormat, kai::test::DataFormat>, 0ul>::combine(std::__1::tuple<kai::test::MatMulShape, kai::test::DataFormat, kai::test::DataFormat, kai::test::DataFormat> const&) (test/common/cache.hpp:34)	called 1243 times	100.0%	50.0%	66.0%
kai::test::impl::TupleHash<std::__1::tuple<kai::test::MatMulShape, kai::test::DataFormat, kai::test::DataFormat, kai::test::DataFormat>, 1ul>::combine(std::__1::tuple<kai::test::MatMulShape, kai::test::DataFormat, kai::test::DataFormat, kai::test::DataFormat> const&) (test/common/cache.hpp:22)	called 1243 times	100.0%	50.0%	66.0%
kai::test::impl::TupleHash<std::__1::tuple<kai::test::MatMulShape, kai::test::DataFormat, kai::test::DataFormat, kai::test::DataFormat>, 2ul>::combine(std::__1::tuple<kai::test::MatMulShape, kai::test::DataFormat, kai::test::DataFormat, kai::test::DataFormat> const&) (test/common/cache.hpp:22)	called 1243 times	100.0%	-%	66.0%
kai::test::impl::TupleHash<std::__1::tuple<kai::test::MatMulShape, kai::test::DataFormat, kai::test::DataFormat, kai::test::DataFormat>, 3ul>::combine(std::__1::tuple<kai::test::MatMulShape, kai::test::DataFormat, kai::test::DataFormat, kai::test::DataFormat> const&) (test/common/cache.hpp:22)	called 1243 times	100.0%	-%	66.0%
kai::test::impl::hash_combine(unsigned long, unsigned long) (test/common/cache.hpp:14)	called 3729 times	100.0%	-%	100.0%
kai::test::indirect_matmul(void const* const*, unsigned long, void const*, void const*, void const*, kai::test::DataType, void const*, void const*, void const*, kai::test::DataType, void const*, void const*, void const*, kai::test::DataType, kai::test::DataType, unsigned long, unsigned long, unsigned long, unsigned long) (test/reference/matmul.cpp:186)	called 2244 times	100.0%	70.0%	73.0%
kai::test::matmul(void const*, void const*, void const*, kai::test::DataType, void const*, void const*, void const*, kai::test::DataType, void const*, void const*, void const*, kai::test::DataType, kai::test::DataType, unsigned long, unsigned long, unsigned long, bool, bool) (test/reference/matmul.cpp:127)	called 2387 times	68.8%	38.5%	44.0%
kai::test::matmul_pack_rhs(void const*, void const*, void const*, kai::test::DataFormat const&, kai::test::DataFormat const&, unsigned long, unsigned long, bool) (test/reference/matmul.cpp:77)	called 125 times	50.0%	10.9%	11.0%
kai::test::mul(void const*, kai::test::DataType, unsigned long, unsigned long, void const*, kai::test::DataType, unsigned long, unsigned long) (test/reference/binary_elementwise.cpp:150)	not called	0.0%	-%	0.0%
kai::test::operator!=(kai::test::BFloat16<true>, kai::test::BFloat16<true>) (test/common/bfloat16.hpp:78)	called 3 times	100.0%	-%	100.0%
kai::test::operator!=(kai::test::Float16, kai::test::Float16) (test/common/float16.hpp:190)	called 3 times	100.0%	-%	100.0%
kai::test::operator*(kai::test::Float16, kai::test::Float16) (test/common/float16.hpp:171)	called 600720322 times	100.0%	-%	100.0%
kai::test::operator+(kai::test::Float16, kai::test::Float16) (test/common/float16.hpp:157)	called 2902782 times	100.0%	-%	100.0%
kai::test::operator-(kai::test::Float16, kai::test::Float16) (test/common/float16.hpp:164)	called 1 time	100.0%	-%	100.0%
kai::test::operator/(kai::test::Float16, kai::test::Float16) (test/common/float16.hpp:178)	called 1 time	100.0%	-%	100.0%
kai::test::operator<(kai::test::Float16, kai::test::Float16) (test/common/float16.hpp:195)	called 6 times	66.7%	-%	100.0%
kai::test::operator<<(std::__1::basic_ostream<char, std::__1::char_traits<char>>&, kai::test::BFloat16<true>) (test/common/bfloat16.cpp:29)	not called	0.0%	-%	0.0%
kai::test::operator<<(std::__1::basic_ostream<char, std::__1::char_traits<char>>&, kai::test::Float16) (test/common/float16.cpp:29)	not called	0.0%	-%	0.0%
kai::test::operator<<(std::__1::basic_ostream<char, std::__1::char_traits<char>>&, kai::test::MatMulShape const&) (test/common/matmul_test_common.cpp:13)	not called	0.0%	-%	0.0%
kai::test::operator<<(std::__1::basic_ostream<char, std::__1::char_traits<char>>&, kai::test::Padding2D const&) (test/reference/dwconv.cpp:13)	not called	0.0%	-%	0.0%
kai::test::operator<<(std::__1::basic_ostream<char, std::__1::char_traits<char>>&, kai::test::Rect const&) (test/common/rect.cpp:46)	not called	0.0%	-%	0.0%
kai::test::operator<=(kai::test::Float16, kai::test::Float16) (test/common/float16.hpp:205)	called 3 times	100.0%	-%	100.0%
kai::test::operator==(kai::test::BFloat16<true>, kai::test::BFloat16<true>) (test/common/bfloat16.hpp:73)	called 3 times	100.0%	-%	100.0%
kai::test::operator==(kai::test::Float16, kai::test::Float16) (test/common/float16.hpp:185)	called 11 times	100.0%	-%	100.0%
kai::test::operator==(kai::test::MatMulShape const&, kai::test::MatMulShape const&) (test/common/matmul_test_common.hpp:38)	called 79726 times	100.0%	100.0%	100.0%
kai::test::operator==(kai::test::Padding2D const&, kai::test::Padding2D const&) (test/reference/dwconv.hpp:33)	not called	0.0%	0.0%	0.0%
kai::test::operator>(kai::test::Float16, kai::test::Float16) (test/common/float16.hpp:200)	called 6 times	100.0%	-%	100.0%
kai::test::operator>=(kai::test::Float16, kai::test::Float16) (test/common/float16.hpp:210)	called 3 times	100.0%	-%	100.0%
kai::test::pack(kai::test::DataFormat const&, void const*, void const*, void const*, kai::test::DataFormat const&, unsigned long, unsigned long) (test/reference/pack.cpp:173)	called 241 times	100.0%	66.7%	50.0%
kai::test::pack_lhs_qsi8d32p(kai::test::kai_qsi8d32p_pack_functions const&, unsigned long, unsigned long, unsigned long, unsigned long, unsigned long, unsigned long, kai::test::Buffer const&, unsigned long, unsigned long, unsigned long) (test/tests/matmul_clamp_f32_qsi8d32p_qai4c32p_test.cpp:105)	called 10904 times	100.0%	50.0%	50.0%
kai::test::pack_lhs_qsi8d32p_f16(kai::test::kai_qsi8d32p_f16_pack_functions const&, unsigned long, unsigned long, unsigned long, unsigned long, unsigned long, unsigned long, kai::test::Buffer const&, unsigned long, unsigned long, unsigned long) (test/tests/matmul_clamp_f16_qsi8d32p_qai4c32p_test.cpp:148)	called 10904 times	100.0%	50.0%	50.0%
kai::test::pack_rhs_qai4c32p(kai::test::kai_qai4c32p_pack_functions const&, unsigned long, unsigned long, unsigned long, unsigned long, unsigned long, unsigned long, kai::test::Buffer const&, bool, kai::test::Buffer const&, kai::test::Buffer const&, kai::test::Buffer const&, bool) (test/tests/matmul_clamp_f16_qsi8d32p_qai4c32p_test.cpp:165)	called 2616 times	100.0%	60.0%	59.0%
kai::test::pack_rhs_qai4c32p(kai::test::kai_qai4c32p_pack_functions const&, unsigned long, unsigned long, unsigned long, unsigned long, unsigned long, unsigned long, kai::test::Buffer const&, bool, kai::test::Buffer const&, kai::test::Buffer const&, kai::test::Buffer const&, bool, unsigned long) (test/tests/matmul_clamp_f32_qsi8d32p_qai4c32p_test.cpp:122)	called 2616 times	100.0%	59.1%	53.0%
kai::test::pack_rhs_qsi4c32pscalebf16(unsigned long, unsigned long, unsigned long, unsigned long, unsigned long, unsigned long, kai::test::Buffer const&, kai::test::Buffer const&, unsigned long, kai::test::Buffer const&, kai::test::RhsPackType, unsigned long, unsigned long) (test/tests/matmul_clamp_qai8dxp_qsi4c32p_test.cpp:119)	called 1404 times	100.0%	63.3%	66.0%
kai::test::pack_rhs_qsi4c32pscalebf16_neon(unsigned long, unsigned long, unsigned long, unsigned long, unsigned long, unsigned long, kai::test::Buffer const&, kai::test::Buffer const&, unsigned long, kai::test::Buffer const&, kai::test::RhsPackType, unsigned long, unsigned long) (test/tests/matmul_clamp_qai8dxp_qsi4c32p_test.cpp:177)	called 702 times	100.0%	57.1%	63.0%
kai::test::print_matrix(std::__1::basic_ostream<char, std::__1::char_traits<char>>&, std::__1::basic_string_view<char, std::__1::char_traits<char>>, void const*, kai::test::DataFormat const&, unsigned long, unsigned long) (test/common/printer.cpp:160)	not called	0.0%	0.0%	0.0%
kai::test::read_array(kai::test::DataType, void const*, unsigned long) (test/common/memory.cpp:16)	called 68574998 times	38.5%	22.2%	28.0%
kai::test::reduce_add(void const*, kai::test::DataFormat const&, unsigned long, unsigned long, kai::test::DataFormat const&, unsigned long) (test/reference/reduce.cpp:107)	not called	0.0%	-%	0.0%
kai::test::round_down_multiple(unsigned long, unsigned long) (test/common/round.cpp:63)	called 247440 times	100.0%	-%	57.0%
kai::test::round_to_nearest_even_i32(float) (test/common/round.cpp:33)	called 92897202 times	100.0%	-%	100.0%
kai::test::round_to_nearest_even_usize(float) (test/common/round.cpp:37)	called 247440 times	100.0%	-%	57.0%
kai::test::round_up_division(unsigned long, unsigned long) (test/common/round.cpp:58)	called 241298 times	100.0%	-%	57.0%
kai::test::round_up_multiple(unsigned long, unsigned long) (test/common/round.cpp:53)	called 304115 times	100.0%	-%	57.0%
kai::test::sub(void const*, kai::test::DataType, unsigned long, unsigned long, void const*, kai::test::DataType, unsigned long, unsigned long) (test/reference/binary_elementwise.cpp:131)	not called	0.0%	-%	0.0%
kai::test::test_description(kai::test::(anonymous namespace)::IndirectMatMulVariant const&, kai::test::MatMulShape const&, kai::test::MatrixPortion const&, unsigned long, float) (test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp:827)	called 9324 times	100.0%	50.0%	88.0%
kai::test::test_description(kai::test::(anonymous namespace)::MatMulVariant const&, kai::test::MatMulShape const&, kai::test::MatrixPortion const&, float) (test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp:815)	called 834 times	100.0%	50.0%	50.0%
kai::test::test_description(std::__1::basic_string_view<char, std::__1::char_traits<char>> const&, kai::test::MatMulShape const&, kai::test::MatrixPortion const&, bool) (test/common/matmul_test_common.cpp:37)	called 11250 times	100.0%	56.2%	83.0%
kai::test::transpose(void const*, kai::test::DataType, unsigned long, unsigned long) (test/reference/transpose.cpp:21)	called 250 times	100.0%	83.3%	72.0%
kai::test::write_array(kai::test::DataType, void*, unsigned long, double) (test/common/memory.cpp:39)	called 34336255 times	36.8%	22.2%	30.0%
kai_cast_bf16_f32 (kai/kai_common.h:118)	called 128 times	100.0%	-%	100.0%
kai_cast_f16_f32 (kai/kai_common.h:134)	called 4540 times	100.0%	-%	100.0%
kai_cast_f32_bf16 (kai/kai_common.h:107)	called 993160 times	100.0%	-%	100.0%
kai_cast_f32_bf16(unsigned short) (kai/kai_common.h:107)	called 1684870722 times	100.0%	-%	100.0%
kai_ext_sign_i8_i4 (kai/kai_common.h:165)	called 5788242 times	100.0%	-%	57.0%
kai_get_bias_offset_rhs_imatmul_pack_kxn_qsi8cxp2vlx4sb_qs8cx_f32_i32_sme (kai/ukernels/matmul/pack/kai_rhs_imatmul_pack_kxn_qsi8cxp2vlx4sb_qs8cx_f32_i32_sme.c:56)	called 9324 times	100.0%	-%	100.0%
kai_get_bias_offset_rhs_imatmul_pack_kxn_x16p2vlx2b_x16_x16_sme (kai/ukernels/matmul/pack/kai_rhs_imatmul_pack_kxn_x16p2vlx2b_x16_x16_sme.c:52)	called 11352 times	100.0%	-%	100.0%
kai_get_bias_offset_rhs_imatmul_pack_kxn_x32p2vlx1b_x32_x32_sme (kai/ukernels/matmul/pack/kai_rhs_imatmul_pack_kxn_x32p2vlx1b_x32_x32_sme.c:49)	called 11352 times	100.0%	-%	100.0%
kai_get_bias_offset_rhs_pack_kxn_bf16p12x4biasf16_f16_neon (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_bf16p12x4biasf16_f16_neon.c:39)	not called	0.0%	-%	0.0%
kai_get_bias_offset_rhs_pack_kxn_bf16p12x4biasf32_f16_neon (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_bf16p12x4biasf32_f16_neon.c:37)	not called	0.0%	-%	0.0%
kai_get_bias_offset_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme.c:35)	not called	0.0%	-%	0.0%
kai_get_bias_offset_rhs_pack_kxn_f16p16x1biasf16_f16_f16_neon (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_f16p16x1biasf16_f16_f16_neon.c:31)	called 17 times	100.0%	-%	100.0%
kai_get_bias_offset_rhs_pack_kxn_f32p16vlx1b_f32_f32_sme (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_f32p16vlx1b_f32_f32_sme.c:48)	called 46 times	100.0%	-%	100.0%
kai_get_bias_offset_rhs_pack_kxn_f32p2vlx1biasf32_f32_f32_sme (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_f32p2vlx1biasf32_f32_f32_sme.c:48)	called 160 times	100.0%	-%	100.0%
kai_get_bias_offset_rhs_pack_kxn_f32p8x1biasf32_f32_f32_neon (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_f32p8x1biasf32_f32_f32_neon.c:31)	called 17 times	100.0%	-%	100.0%
kai_get_bias_offset_rhs_pack_kxn_qsi8cxp2vlx4sb_qs8cx_f32_i32_sme (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_qsi8cxp2vlx4sb_qs8cx_f32_i32_sme.c:55)	called 834 times	100.0%	-%	100.0%
kai_get_bias_offset_rhs_pack_kxn_x16p2vlx2b_x16_x16_sme (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_x16p2vlx2b_x16_x16_sme.c:51)	called 160 times	100.0%	-%	100.0%
kai_get_bias_offset_rhs_pack_kxn_x16p32x1b_x16_x16_neon (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_x16p32x1b_x16_x16_neon.c:49)	called 34 times	100.0%	-%	100.0%
kai_get_bias_offset_rhs_pack_kxn_x32p16x1b_x32_x32_neon (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_x32p16x1b_x32_x32_neon.c:49)	called 34 times	100.0%	-%	100.0%
kai_get_bias_offset_rhs_pack_nxk_f32p2vlx1biasf32_f32_f32_sme (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_f32p2vlx1biasf32_f32_f32_sme.c:45)	not called	0.0%	-%	0.0%
kai_get_bias_offset_rhs_pack_nxk_x16p2vlx2b_x16_x16_sme (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_x16p2vlx2b_x16_x16_sme.c:45)	not called	0.0%	-%	0.0%
kai_get_bias_offset_rhs_quant_pack_kxn_bf16p12x4biasf32_f32_neon (kai/ukernels/matmul/pack/kai_rhs_quant_pack_kxn_bf16p12x4biasf32_f32_neon.c:38)	not called	0.0%	-%	0.0%
kai_get_datatype_size_in_bytes (kai/kai_common.h:87)	called 7020 times	100.0%	-%	100.0%
kai_get_datatype_size_in_bytes(kai_datatype) (kai/kai_common.h:87)	called 2106 times	100.0%	-%	100.0%
kai_get_dst_offset_dwconv_clamp_f32_f32_f32p1vlx1b_3x3_s1_4xc_sme2_mla (kai/ukernels/dwconv/dwconv_f32_f32_f32p/kai_dwconv_clamp_f32_f32_f32p1vlx1b_3x3_s1_4xc_sme2_mla.c:66)	not called	0.0%	-%	0.0%
kai_get_dst_offset_imatmul_clamp_f16_f16p2vlx2_f16p2vlx2_2vlx2vl_sme2_mopa (kai/ukernels/matmul/imatmul_clamp_f16_f16p_f16p/kai_imatmul_clamp_f16_f16p2vlx2_f16p2vlx2_2vlx2vl_sme2_mopa.c:73)	called 5676 times	100.0%	-%	53.0%
kai_get_dst_offset_imatmul_clamp_f16_f16p2vlx2_f16p2vlx2b_2vlx2vl_sme_mopa (kai/ukernels/matmul/imatmul_clamp_f16_f16p_f16p/kai_imatmul_clamp_f16_f16p2vlx2_f16p2vlx2b_2vlx2vl_sme_mopa.c:73)	called 5676 times	100.0%	-%	53.0%
kai_get_dst_offset_imatmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme2_mopa (kai/ukernels/matmul/imatmul_clamp_f32_f32p_f32p/kai_imatmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme2_mopa.c:72)	called 5676 times	100.0%	-%	53.0%
kai_get_dst_offset_imatmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme_mopa (kai/ukernels/matmul/imatmul_clamp_f32_f32p_f32p/kai_imatmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme_mopa.c:72)	called 5676 times	100.0%	-%	53.0%
kai_get_dst_offset_imatmul_clamp_qai8_qai8p2vlx4_qsi8cxp2vlx4sb_2vlx2vl_sme_mopa (kai/ukernels/matmul/imatmul_clamp_qai8_qai8p_qsi8cxp/kai_imatmul_clamp_qai8_qai8p2vlx4_qsi8cxp2vlx4sb_2vlx2vl_sme_mopa.c:73)	called 4662 times	100.0%	-%	53.0%
kai_get_dst_offset_imatmul_clamp_qai8_qai8p2vlx4_qsi8cxpsb2vlx4_2vlx2vl_sme2_mopa (kai/ukernels/matmul/imatmul_clamp_qai8_qai8p_qsi8cxp/kai_imatmul_clamp_qai8_qai8p2vlx4_qsi8cxpsb2vlx4_2vlx2vl_sme2_mopa.c:73)	called 4662 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_bf16_qai8dxp1x8_qsi4c32p4x8_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_bf16_qai8dxp_qsi4c32p/kai_matmul_clamp_bf16_qai8dxp1x8_qsi4c32p4x8_1x4_neon_dotprod.c:136)	called 108 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_bf16_qai8dxp1x8_qsi4cxp8x8_1x8_neon_dotprod (kai/ukernels/matmul/matmul_clamp_bf16_qai8dxp_qsi4cxp/kai_matmul_clamp_bf16_qai8dxp1x8_qsi4cxp8x8_1x8_neon_dotprod.c:120)	called 356 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_bf16_qai8dxp4x8_qsi4c32p4x8_16x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_bf16_qai8dxp_qsi4c32p/kai_matmul_clamp_bf16_qai8dxp4x8_qsi4c32p4x8_16x4_neon_i8mm.c:137)	called 108 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_bf16_qai8dxp4x8_qsi4cxp8x8_8x8_neon_i8mm (kai/ukernels/matmul/matmul_clamp_bf16_qai8dxp_qsi4cxp/kai_matmul_clamp_bf16_qai8dxp4x8_qsi4cxp8x8_8x8_neon_i8mm.c:121)	called 388 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_f16_bf16p8x4_bf16p12x4b_8x12_neon_mmla (kai/ukernels/matmul/matmul_clamp_f16_bf16p_bf16p/kai_matmul_clamp_f16_bf16p8x4_bf16p12x4b_8x12_neon_mmla.c:63)	called 84 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_f16_f16_f16p16x1biasf16_6x16x8_neon_mla (kai/ukernels/matmul/matmul_clamp_f16_f16_f16p/kai_matmul_clamp_f16_f16_f16p16x1biasf16_6x16x8_neon_mla.c:60)	called 16 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_f16_f16_f16p2vlx2b_1x16vl_sme2_dot (kai/ukernels/matmul/matmul_clamp_f16_f16_f16p/kai_matmul_clamp_f16_f16_f16p2vlx2b_1x16vl_sme2_dot.c:77)	called 46 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_f16_f16_f16p2vlx2b_1x8vl_sme_mla (kai/ukernels/matmul/matmul_clamp_f16_f16_f16p/kai_matmul_clamp_f16_f16_f16p2vlx2b_1x8vl_sme_mla.c:77)	called 46 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_f16_f16_f16p32x1b_6x32_neon_mla (kai/ukernels/matmul/matmul_clamp_f16_f16_f16p/kai_matmul_clamp_f16_f16_f16p32x1b_6x32_neon_mla.c:76)	called 16 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_f16_f16_f16p32x1b_6x32_neon_mla_cortexa55 (kai/ukernels/matmul/matmul_clamp_f16_f16_f16p/kai_matmul_clamp_f16_f16_f16p32x1b_6x32_neon_mla_cortexa55.c:76)	called 16 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_f16_f16p2vlx2_f16p2vlx2_2vlx2vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f16_f16p_f16p/kai_matmul_clamp_f16_f16p2vlx2_f16p2vlx2_2vlx2vl_sme2_mopa.c:85)	called 16 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_f16_f16p2vlx2_f16p2vlx2b_2vlx2vl_sme_mopa (kai/ukernels/matmul/matmul_clamp_f16_f16p_f16p/kai_matmul_clamp_f16_f16p2vlx2_f16p2vlx2b_2vlx2vl_sme_mopa.c:85)	called 16 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_f16_qai8dxp1x4_qsi4cxp4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi4cxp/kai_matmul_clamp_f16_qai8dxp1x4_qsi4cxp4x4_1x4_neon_dotprod.c:120)	called 58 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_f16_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi8cxp/kai_matmul_clamp_f16_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod.c:117)	called 58 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_f16_qai8dxp1x8_qsi4cxp4x8_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi4cxp/kai_matmul_clamp_f16_qai8dxp1x8_qsi4cxp4x8_1x4_neon_dotprod.c:120)	called 58 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_f16_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi8cxp/kai_matmul_clamp_f16_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod.c:117)	called 58 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_f16_qai8dxp4x4_qsi4cxp4x4_16x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi4cxp/kai_matmul_clamp_f16_qai8dxp4x4_qsi4cxp4x4_16x4_neon_dotprod.c:120)	called 222 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_f16_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi8cxp/kai_matmul_clamp_f16_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod.c:117)	called 222 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_f16_qai8dxp4x8_qsi4cxp4x8_16x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi4cxp/kai_matmul_clamp_f16_qai8dxp4x8_qsi4cxp4x8_16x4_neon_i8mm.c:120)	called 222 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_f16_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi8cxp/kai_matmul_clamp_f16_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm.c:117)	called 222 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_f16_qsi8d32p1vlx4_qai4c32p4vlx4_1vlx4vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p1vlx4_qai4c32p4vlx4_1vlx4vl_sme2_mopa.c:143)	called 1032 times	100.0%	-%	52.0%
kai_get_dst_offset_matmul_clamp_f16_qsi8d32p1x4_qai4c32p4vlx4_1x4vl_sme2_dot (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p1x4_qai4c32p4vlx4_1x4vl_sme2_dot.c:139)	called 276 times	100.0%	-%	52.0%
kai_get_dst_offset_matmul_clamp_f16_qsi8d32p1x4_qai4c32p4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p1x4_qai4c32p4x4_1x4_neon_dotprod.c:129)	called 138 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_f16_qsi8d32p1x8_qai4c32p4x8_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p1x8_qai4c32p4x8_1x4_neon_dotprod.c:129)	called 138 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_f16_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod.c:129)	called 516 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_f16_qsi8d32p4x8_qai4c32p4x8_8x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p4x8_qai4c32p4x8_8x4_neon_i8mm.c:129)	called 516 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_f32_bf16p1x4_bf16p12x4b_1x36_neon_dot (kai/ukernels/matmul/matmul_clamp_f32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p1x4_bf16p12x4b_1x36_neon_dot.c:64)	called 92 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa.c:84)	called 46 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_f32_bf16p8x4_bf16p12x4b_8x12_neon_mmla (kai/ukernels/matmul/matmul_clamp_f32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p8x4_bf16p12x4b_8x12_neon_mmla.c:64)	called 184 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_f32_f32_f32p16vlx1b_1x16vl_sme2_mla (kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p16vlx1b_1x16vl_sme2_mla.c:75)	called 46 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_f32_f32_f32p16x1b_6x16_neon_mla (kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p16x1b_6x16_neon_mla.c:72)	called 16 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_f32_f32_f32p16x1b_6x16_neon_mla_cortexa55 (kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p16x1b_6x16_neon_mla_cortexa55.c:72)	called 16 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla (kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla.c:75)	called 46 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_f32_f32_f32p2vlx1b_1x8vl_sme_mla (kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p2vlx1b_1x8vl_sme_mla.c:75)	called 46 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_f32_f32_f32p8x1biasf32_6x8x4_neon_mla (kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p8x1biasf32_6x8x4_neon_mla.c:73)	called 16 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme_mopa (kai/ukernels/matmul/matmul_clamp_f32_f32p_f32p/kai_matmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme_mopa.c:84)	called 16 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f32_f32p_f32p/kai_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa.c:84)	called 16 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa.c:130)	called 154 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme_mopa (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme_mopa.c:130)	called 154 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_f32_qai8dxp1vlx8_qsi4cxp4vlx8_1vlx4vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1vlx8_qsi4cxp4vlx8_1vlx4vl_sme2_mopa.c:100)	called 80 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_f32_qai8dxp1x4_qsi4c32p4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x4_qsi4c32p4x4_1x4_neon_dotprod.c:135)	called 108 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_f32_qai8dxp1x4_qsi4c32p8x4_1x8_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x4_qsi4c32p8x4_1x8_neon_dotprod.c:135)	called 108 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_f32_qai8dxp1x4_qsi4cxp4vlx4_1x4vl_sme2_sdot (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi4cxp4vlx4_1x4vl_sme2_sdot.c:94)	called 80 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_f32_qai8dxp1x4_qsi4cxp4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi4cxp4x4_1x4_neon_dotprod.c:105)	called 160 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot.c:129)	called 134 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme_dot (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme_dot.c:129)	called 134 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod.c:103)	called 134 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_f32_qai8dxp1x8_qsi4c32p4x8_1x4x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x8_qsi4c32p4x8_1x4x32_neon_dotprod.c:135)	called 108 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_f32_qai8dxp1x8_qsi4c32p8x8_1x8_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x8_qsi4c32p8x8_1x8_neon_dotprod.c:135)	called 108 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_f32_qai8dxp1x8_qsi4c32p8x8_1x8x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x8_qsi4c32p8x8_1x8x32_neon_dotprod.c:135)	called 108 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_f32_qai8dxp1x8_qsi4cxp4x8_1x4x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi4cxp4x8_1x4x32_neon_dotprod.c:89)	called 160 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_f32_qai8dxp1x8_qsi4cxp8x8_1x8x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi4cxp8x8_1x8x32_neon_dotprod.c:89)	called 160 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_f32_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod.c:103)	called 134 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_f32_qai8dxp4x4_qsi4c32p4x4_16x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x4_qsi4c32p4x4_16x4_neon_dotprod.c:136)	called 108 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod.c:135)	called 108 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_f32_qai8dxp4x4_qsi4cxp8x4_8x8x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x4_qsi4cxp8x4_8x8x32_neon_dotprod.c:91)	called 160 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_f32_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod.c:107)	called 154 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_f32_qai8dxp4x8_qsi4c32p4x8_16x4x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p4x8_16x4x32_neon_i8mm.c:136)	called 108 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_f32_qai8dxp4x8_qsi4c32p4x8_8x4x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p4x8_8x4x32_neon_i8mm.c:135)	called 108 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_f32_qai8dxp4x8_qsi4c32p8x8_4x8_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p8x8_4x8_neon_i8mm.c:135)	called 108 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_f32_qai8dxp4x8_qsi4c32p8x8_4x8x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p8x8_4x8x32_neon_i8mm.c:135)	called 108 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x4_16x4x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x4_16x4x32_neon_dotprod.c:91)	called 160 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_4x4x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_4x4x32_neon_i8mm.c:88)	called 160 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_8x4x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_8x4x32_neon_i8mm.c:92)	called 160 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_4x8x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_4x8x32_neon_i8mm.c:92)	called 160 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_8x8x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_8x8x32_neon_i8mm.c:92)	called 160 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_f32_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm.c:107)	called 154 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_f32_qsi8d32p1vlx4_qai4c32p4vlx4_1vlx4vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p1vlx4_qai4c32p4vlx4_1vlx4vl_sme2_mopa.c:143)	called 1032 times	100.0%	-%	52.0%
kai_get_dst_offset_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa.c:119)	called 23 times	100.0%	-%	52.0%
kai_get_dst_offset_matmul_clamp_f32_qsi8d32p1x4_qai4c32p4vlx4_1x4vl_sme2_dot (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qai4c32p4vlx4_1x4vl_sme2_dot.c:139)	called 276 times	100.0%	-%	52.0%
kai_get_dst_offset_matmul_clamp_f32_qsi8d32p1x4_qai4c32p4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qai4c32p4x4_1x4_neon_dotprod.c:127)	called 138 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot.c:120)	called 3 times	100.0%	-%	52.0%
kai_get_dst_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod.c:108)	called 3 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_f32_qsi8d32p1x8_qai4c32p4x8_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p1x8_qai4c32p4x8_1x4_neon_dotprod.c:127)	called 138 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod.c:101)	called 3 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_f32_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod.c:127)	called 516 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod.c:112)	called 23 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_f32_qsi8d32p4x8_qai4c32p4x8_8x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qai4c32p4x8_8x4_neon_i8mm.c:127)	called 516 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm.c:104)	called 46 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_8x4x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_8x4x32_neon_i8mm.c:104)	called 23 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_qai8_qai8_qsi8cxp2vlx4sb_1x16vl_sme2_dot (kai/ukernels/matmul/matmul_clamp_qai8_qai8_qsi8cxp/kai_matmul_clamp_qai8_qai8_qsi8cxp2vlx4sb_1x16vl_sme2_dot.c:76)	called 168 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_qai8_qai8p2vlx4_qsi8cxp2vlx4sb_2vlx2vl_sme_mopa (kai/ukernels/matmul/matmul_clamp_qai8_qai8p_qsi8cxp/kai_matmul_clamp_qai8_qai8p2vlx4_qsi8cxp2vlx4sb_2vlx2vl_sme_mopa.c:85)	called 333 times	100.0%	-%	53.0%
kai_get_dst_offset_matmul_clamp_qai8_qai8p2vlx4_qsi8cxpsb2vlx4_2vlx2vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_qai8_qai8p_qsi8cxp/kai_matmul_clamp_qai8_qai8p2vlx4_qsi8cxpsb2vlx4_2vlx2vl_sme2_mopa.c:85)	called 333 times	100.0%	-%	53.0%
kai_get_dst_size_dwconv_clamp_f32_f32_f32p1vlx1b_3x3_s1_4xc_sme2_mla (kai/ukernels/dwconv/dwconv_f32_f32_f32p/kai_dwconv_clamp_f32_f32_f32p1vlx1b_3x3_s1_4xc_sme2_mla.c:61)	called 60 times	100.0%	-%	100.0%
kai_get_dst_size_imatmul_clamp_f16_f16p2vlx2_f16p2vlx2_2vlx2vl_sme2_mopa (kai/ukernels/matmul/imatmul_clamp_f16_f16p_f16p/kai_imatmul_clamp_f16_f16p2vlx2_f16p2vlx2_2vlx2vl_sme2_mopa.c:81)	called 5676 times	100.0%	-%	100.0%
kai_get_dst_size_imatmul_clamp_f16_f16p2vlx2_f16p2vlx2b_2vlx2vl_sme_mopa (kai/ukernels/matmul/imatmul_clamp_f16_f16p_f16p/kai_imatmul_clamp_f16_f16p2vlx2_f16p2vlx2b_2vlx2vl_sme_mopa.c:81)	called 5676 times	100.0%	-%	100.0%
kai_get_dst_size_imatmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme2_mopa (kai/ukernels/matmul/imatmul_clamp_f32_f32p_f32p/kai_imatmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme2_mopa.c:80)	called 5676 times	100.0%	-%	100.0%
kai_get_dst_size_imatmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme_mopa (kai/ukernels/matmul/imatmul_clamp_f32_f32p_f32p/kai_imatmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme_mopa.c:80)	called 5676 times	100.0%	-%	100.0%
kai_get_dst_size_imatmul_clamp_qai8_qai8p2vlx4_qsi8cxp2vlx4sb_2vlx2vl_sme_mopa (kai/ukernels/matmul/imatmul_clamp_qai8_qai8p_qsi8cxp/kai_imatmul_clamp_qai8_qai8p2vlx4_qsi8cxp2vlx4sb_2vlx2vl_sme_mopa.c:81)	called 4662 times	100.0%	-%	100.0%
kai_get_dst_size_imatmul_clamp_qai8_qai8p2vlx4_qsi8cxpsb2vlx4_2vlx2vl_sme2_mopa (kai/ukernels/matmul/imatmul_clamp_qai8_qai8p_qsi8cxp/kai_imatmul_clamp_qai8_qai8p2vlx4_qsi8cxpsb2vlx4_2vlx2vl_sme2_mopa.c:81)	called 4662 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_bf16_qai8dxp1x8_qsi4c32p4x8_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_bf16_qai8dxp_qsi4c32p/kai_matmul_clamp_bf16_qai8dxp1x8_qsi4c32p4x8_1x4_neon_dotprod.c:144)	called 108 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_bf16_qai8dxp1x8_qsi4cxp8x8_1x8_neon_dotprod (kai/ukernels/matmul/matmul_clamp_bf16_qai8dxp_qsi4cxp/kai_matmul_clamp_bf16_qai8dxp1x8_qsi4cxp8x8_1x8_neon_dotprod.c:128)	called 356 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_bf16_qai8dxp4x8_qsi4c32p4x8_16x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_bf16_qai8dxp_qsi4c32p/kai_matmul_clamp_bf16_qai8dxp4x8_qsi4c32p4x8_16x4_neon_i8mm.c:145)	called 108 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_bf16_qai8dxp4x8_qsi4cxp8x8_8x8_neon_i8mm (kai/ukernels/matmul/matmul_clamp_bf16_qai8dxp_qsi4cxp/kai_matmul_clamp_bf16_qai8dxp4x8_qsi4cxp8x8_8x8_neon_i8mm.c:129)	called 388 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f16_bf16p8x4_bf16p12x4b_8x12_neon_mmla (kai/ukernels/matmul/matmul_clamp_f16_bf16p_bf16p/kai_matmul_clamp_f16_bf16p8x4_bf16p12x4b_8x12_neon_mmla.c:71)	called 84 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f16_f16_f16p16x1biasf16_6x16x8_neon_mla (kai/ukernels/matmul/matmul_clamp_f16_f16_f16p/kai_matmul_clamp_f16_f16_f16p16x1biasf16_6x16x8_neon_mla.c:68)	called 16 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f16_f16_f16p2vlx2b_1x16vl_sme2_dot (kai/ukernels/matmul/matmul_clamp_f16_f16_f16p/kai_matmul_clamp_f16_f16_f16p2vlx2b_1x16vl_sme2_dot.c:85)	called 46 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f16_f16_f16p2vlx2b_1x8vl_sme_mla (kai/ukernels/matmul/matmul_clamp_f16_f16_f16p/kai_matmul_clamp_f16_f16_f16p2vlx2b_1x8vl_sme_mla.c:84)	called 46 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f16_f16_f16p32x1b_6x32_neon_mla (kai/ukernels/matmul/matmul_clamp_f16_f16_f16p/kai_matmul_clamp_f16_f16_f16p32x1b_6x32_neon_mla.c:83)	called 16 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f16_f16_f16p32x1b_6x32_neon_mla_cortexa55 (kai/ukernels/matmul/matmul_clamp_f16_f16_f16p/kai_matmul_clamp_f16_f16_f16p32x1b_6x32_neon_mla_cortexa55.c:84)	called 16 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f16_f16p2vlx2_f16p2vlx2_2vlx2vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f16_f16p_f16p/kai_matmul_clamp_f16_f16p2vlx2_f16p2vlx2_2vlx2vl_sme2_mopa.c:93)	called 16 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f16_f16p2vlx2_f16p2vlx2b_2vlx2vl_sme_mopa (kai/ukernels/matmul/matmul_clamp_f16_f16p_f16p/kai_matmul_clamp_f16_f16p2vlx2_f16p2vlx2b_2vlx2vl_sme_mopa.c:93)	called 16 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f16_qai8dxp1x4_qsi4cxp4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi4cxp/kai_matmul_clamp_f16_qai8dxp1x4_qsi4cxp4x4_1x4_neon_dotprod.c:128)	called 58 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f16_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi8cxp/kai_matmul_clamp_f16_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod.c:125)	called 58 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f16_qai8dxp1x8_qsi4cxp4x8_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi4cxp/kai_matmul_clamp_f16_qai8dxp1x8_qsi4cxp4x8_1x4_neon_dotprod.c:128)	called 58 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f16_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi8cxp/kai_matmul_clamp_f16_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod.c:125)	called 58 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f16_qai8dxp4x4_qsi4cxp4x4_16x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi4cxp/kai_matmul_clamp_f16_qai8dxp4x4_qsi4cxp4x4_16x4_neon_dotprod.c:128)	called 222 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f16_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi8cxp/kai_matmul_clamp_f16_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod.c:125)	called 222 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f16_qai8dxp4x8_qsi4cxp4x8_16x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi4cxp/kai_matmul_clamp_f16_qai8dxp4x8_qsi4cxp4x8_16x4_neon_i8mm.c:128)	called 222 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f16_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi8cxp/kai_matmul_clamp_f16_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm.c:125)	called 222 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f16_qsi8d32p1vlx4_qai4c32p4vlx4_1vlx4vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p1vlx4_qai4c32p4vlx4_1vlx4vl_sme2_mopa.c:153)	called 1032 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f16_qsi8d32p1x4_qai4c32p4vlx4_1x4vl_sme2_dot (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p1x4_qai4c32p4vlx4_1x4vl_sme2_dot.c:149)	called 276 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f16_qsi8d32p1x4_qai4c32p4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p1x4_qai4c32p4x4_1x4_neon_dotprod.c:137)	called 138 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f16_qsi8d32p1x8_qai4c32p4x8_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p1x8_qai4c32p4x8_1x4_neon_dotprod.c:137)	called 138 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f16_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod.c:137)	called 516 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f16_qsi8d32p4x8_qai4c32p4x8_8x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p4x8_qai4c32p4x8_8x4_neon_i8mm.c:137)	called 516 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f32_bf16p1x4_bf16p12x4b_1x36_neon_dot (kai/ukernels/matmul/matmul_clamp_f32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p1x4_bf16p12x4b_1x36_neon_dot.c:72)	called 92 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa.c:92)	called 46 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f32_bf16p8x4_bf16p12x4b_8x12_neon_mmla (kai/ukernels/matmul/matmul_clamp_f32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p8x4_bf16p12x4b_8x12_neon_mmla.c:72)	called 184 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f32_f32_f32p16vlx1b_1x16vl_sme2_mla (kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p16vlx1b_1x16vl_sme2_mla.c:83)	called 46 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f32_f32_f32p16x1b_6x16_neon_mla (kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p16x1b_6x16_neon_mla.c:79)	called 16 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f32_f32_f32p16x1b_6x16_neon_mla_cortexa55 (kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p16x1b_6x16_neon_mla_cortexa55.c:80)	called 16 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla (kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla.c:83)	called 46 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f32_f32_f32p2vlx1b_1x8vl_sme_mla (kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p2vlx1b_1x8vl_sme_mla.c:82)	called 46 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f32_f32_f32p8x1biasf32_6x8x4_neon_mla (kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p8x1biasf32_6x8x4_neon_mla.c:81)	called 16 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme_mopa (kai/ukernels/matmul/matmul_clamp_f32_f32p_f32p/kai_matmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme_mopa.c:92)	called 16 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f32_f32p_f32p/kai_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa.c:92)	called 16 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa.c:138)	called 154 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme_mopa (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme_mopa.c:138)	called 154 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f32_qai8dxp1vlx8_qsi4cxp4vlx8_1vlx4vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1vlx8_qsi4cxp4vlx8_1vlx4vl_sme2_mopa.c:108)	called 80 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f32_qai8dxp1x4_qsi4c32p4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x4_qsi4c32p4x4_1x4_neon_dotprod.c:143)	called 108 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f32_qai8dxp1x4_qsi4c32p8x4_1x8_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x4_qsi4c32p8x4_1x8_neon_dotprod.c:143)	called 108 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f32_qai8dxp1x4_qsi4cxp4vlx4_1x4vl_sme2_sdot (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi4cxp4vlx4_1x4vl_sme2_sdot.c:102)	called 80 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f32_qai8dxp1x4_qsi4cxp4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi4cxp4x4_1x4_neon_dotprod.c:113)	called 160 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot.c:137)	called 134 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme_dot (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme_dot.c:137)	called 134 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod.c:111)	called 134 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f32_qai8dxp1x8_qsi4c32p4x8_1x4x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x8_qsi4c32p4x8_1x4x32_neon_dotprod.c:143)	called 108 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f32_qai8dxp1x8_qsi4c32p8x8_1x8_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x8_qsi4c32p8x8_1x8_neon_dotprod.c:143)	called 108 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f32_qai8dxp1x8_qsi4c32p8x8_1x8x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x8_qsi4c32p8x8_1x8x32_neon_dotprod.c:143)	called 108 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f32_qai8dxp1x8_qsi4cxp4x8_1x4x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi4cxp4x8_1x4x32_neon_dotprod.c:97)	called 160 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f32_qai8dxp1x8_qsi4cxp8x8_1x8x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi4cxp8x8_1x8x32_neon_dotprod.c:97)	called 160 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f32_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod.c:111)	called 134 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f32_qai8dxp4x4_qsi4c32p4x4_16x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x4_qsi4c32p4x4_16x4_neon_dotprod.c:144)	called 108 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod.c:143)	called 108 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f32_qai8dxp4x4_qsi4cxp8x4_8x8x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x4_qsi4cxp8x4_8x8x32_neon_dotprod.c:99)	called 160 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f32_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod.c:115)	called 154 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f32_qai8dxp4x8_qsi4c32p4x8_16x4x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p4x8_16x4x32_neon_i8mm.c:144)	called 108 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f32_qai8dxp4x8_qsi4c32p4x8_8x4x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p4x8_8x4x32_neon_i8mm.c:143)	called 108 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f32_qai8dxp4x8_qsi4c32p8x8_4x8_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p8x8_4x8_neon_i8mm.c:143)	called 108 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f32_qai8dxp4x8_qsi4c32p8x8_4x8x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p8x8_4x8x32_neon_i8mm.c:143)	called 108 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x4_16x4x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x4_16x4x32_neon_dotprod.c:99)	called 160 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_4x4x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_4x4x32_neon_i8mm.c:96)	called 160 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_8x4x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_8x4x32_neon_i8mm.c:100)	called 160 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_4x8x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_4x8x32_neon_i8mm.c:100)	called 160 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_8x8x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_8x8x32_neon_i8mm.c:100)	called 160 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f32_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm.c:115)	called 154 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f32_qsi8d32p1vlx4_qai4c32p4vlx4_1vlx4vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p1vlx4_qai4c32p4vlx4_1vlx4vl_sme2_mopa.c:153)	called 1032 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa.c:129)	called 23 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f32_qsi8d32p1x4_qai4c32p4vlx4_1x4vl_sme2_dot (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qai4c32p4vlx4_1x4vl_sme2_dot.c:149)	called 276 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f32_qsi8d32p1x4_qai4c32p4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qai4c32p4x4_1x4_neon_dotprod.c:135)	called 138 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot.c:130)	called 3 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod.c:116)	called 3 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f32_qsi8d32p1x8_qai4c32p4x8_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p1x8_qai4c32p4x8_1x4_neon_dotprod.c:135)	called 138 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod.c:109)	called 3 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f32_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod.c:135)	called 516 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod.c:120)	called 23 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f32_qsi8d32p4x8_qai4c32p4x8_8x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qai4c32p4x8_8x4_neon_i8mm.c:135)	called 516 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm.c:112)	called 46 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_8x4x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_8x4x32_neon_i8mm.c:112)	called 23 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_qai8_qai8_qsi8cxp2vlx4sb_1x16vl_sme2_dot (kai/ukernels/matmul/matmul_clamp_qai8_qai8_qsi8cxp/kai_matmul_clamp_qai8_qai8_qsi8cxp2vlx4sb_1x16vl_sme2_dot.c:84)	called 168 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_qai8_qai8p2vlx4_qsi8cxp2vlx4sb_2vlx2vl_sme_mopa (kai/ukernels/matmul/matmul_clamp_qai8_qai8p_qsi8cxp/kai_matmul_clamp_qai8_qai8p2vlx4_qsi8cxp2vlx4sb_2vlx2vl_sme_mopa.c:93)	called 333 times	100.0%	-%	100.0%
kai_get_dst_size_matmul_clamp_qai8_qai8p2vlx4_qsi8cxpsb2vlx4_2vlx2vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_qai8_qai8p_qsi8cxp/kai_matmul_clamp_qai8_qai8p2vlx4_qsi8cxpsb2vlx4_2vlx2vl_sme2_mopa.c:93)	called 333 times	100.0%	-%	100.0%
kai_get_filter_height_dwconv_clamp_f32_f32_f32p1vlx1b_3x3_s1_4xc_sme2_mla (kai/ukernels/dwconv/dwconv_f32_f32_f32p/kai_dwconv_clamp_f32_f32_f32p1vlx1b_3x3_s1_4xc_sme2_mla.c:49)	not called	0.0%	-%	0.0%
kai_get_filter_width_dwconv_clamp_f32_f32_f32p1vlx1b_3x3_s1_4xc_sme2_mla (kai/ukernels/dwconv/dwconv_f32_f32_f32p/kai_dwconv_clamp_f32_f32_f32p1vlx1b_3x3_s1_4xc_sme2_mla.c:53)	not called	0.0%	-%	0.0%
kai_get_k_roundedup (kai/ukernels/matmul/matmul_clamp_bf16_qai8dxp_qsi4c32p/kai_matmul_clamp_bf16_qai8dxp1x8_qsi4c32p4x8_1x4_neon_dotprod.c:56)	called 108 times	100.0%	-%	100.0%
kai_get_k_roundedup (kai/ukernels/matmul/matmul_clamp_bf16_qai8dxp_qsi4c32p/kai_matmul_clamp_bf16_qai8dxp4x8_qsi4c32p4x8_16x4_neon_i8mm.c:57)	called 108 times	100.0%	-%	100.0%
kai_get_k_roundedup (kai/ukernels/matmul/matmul_clamp_bf16_qai8dxp_qsi4cxp/kai_matmul_clamp_bf16_qai8dxp1x8_qsi4cxp8x8_1x8_neon_dotprod.c:56)	called 1069 times	100.0%	-%	100.0%
kai_get_k_roundedup (kai/ukernels/matmul/matmul_clamp_bf16_qai8dxp_qsi4cxp/kai_matmul_clamp_bf16_qai8dxp4x8_qsi4cxp8x8_8x8_neon_i8mm.c:57)	called 1165 times	100.0%	-%	100.0%
kai_get_k_roundedup (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi4cxp/kai_matmul_clamp_f16_qai8dxp1x4_qsi4cxp4x4_1x4_neon_dotprod.c:56)	called 175 times	100.0%	-%	100.0%
kai_get_k_roundedup (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi4cxp/kai_matmul_clamp_f16_qai8dxp1x8_qsi4cxp4x8_1x4_neon_dotprod.c:56)	called 175 times	100.0%	-%	100.0%
kai_get_k_roundedup (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi4cxp/kai_matmul_clamp_f16_qai8dxp4x4_qsi4cxp4x4_16x4_neon_dotprod.c:56)	called 667 times	100.0%	-%	100.0%
kai_get_k_roundedup (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi4cxp/kai_matmul_clamp_f16_qai8dxp4x8_qsi4cxp4x8_16x4_neon_i8mm.c:56)	called 667 times	100.0%	-%	100.0%
kai_get_k_roundedup (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi8cxp/kai_matmul_clamp_f16_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod.c:53)	called 175 times	100.0%	-%	100.0%
kai_get_k_roundedup (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi8cxp/kai_matmul_clamp_f16_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod.c:53)	called 175 times	100.0%	-%	100.0%
kai_get_k_roundedup (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi8cxp/kai_matmul_clamp_f16_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod.c:53)	called 667 times	100.0%	-%	100.0%
kai_get_k_roundedup (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi8cxp/kai_matmul_clamp_f16_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm.c:53)	called 667 times	100.0%	-%	100.0%
kai_get_k_roundedup (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x4_qsi4c32p4x4_1x4_neon_dotprod.c:55)	called 108 times	100.0%	-%	100.0%
kai_get_k_roundedup (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x4_qsi4c32p8x4_1x8_neon_dotprod.c:55)	called 108 times	100.0%	-%	100.0%
kai_get_k_roundedup (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x8_qsi4c32p4x8_1x4x32_neon_dotprod.c:55)	called 108 times	100.0%	-%	100.0%
kai_get_k_roundedup (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x8_qsi4c32p8x8_1x8_neon_dotprod.c:55)	called 108 times	100.0%	-%	100.0%
kai_get_k_roundedup (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x8_qsi4c32p8x8_1x8x32_neon_dotprod.c:55)	called 108 times	100.0%	-%	100.0%
kai_get_k_roundedup (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x4_qsi4c32p4x4_16x4_neon_dotprod.c:56)	called 108 times	100.0%	-%	100.0%
kai_get_k_roundedup (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod.c:55)	called 108 times	100.0%	-%	100.0%
kai_get_k_roundedup (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p4x8_16x4x32_neon_i8mm.c:56)	called 108 times	100.0%	-%	100.0%
kai_get_k_roundedup (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p4x8_8x4x32_neon_i8mm.c:55)	called 108 times	100.0%	-%	100.0%
kai_get_k_roundedup (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p8x8_4x8_neon_i8mm.c:55)	called 108 times	100.0%	-%	100.0%
kai_get_k_roundedup (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p8x8_4x8x32_neon_i8mm.c:55)	called 108 times	100.0%	-%	100.0%
kai_get_k_roundedup (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi4cxp4x4_1x4_neon_dotprod.c:41)	called 641 times	100.0%	-%	100.0%
kai_get_kernel_vec_length_constant (kai/ukernels/matmul/imatmul_clamp_f16_f16p_f16p/kai_imatmul_clamp_f16_f16p2vlx2_f16p2vlx2_2vlx2vl_sme2_mopa.c:39)	called 34056 times	100.0%	-%	100.0%
kai_get_kernel_vec_length_constant (kai/ukernels/matmul/imatmul_clamp_f16_f16p_f16p/kai_imatmul_clamp_f16_f16p2vlx2_f16p2vlx2b_2vlx2vl_sme_mopa.c:39)	called 34056 times	100.0%	-%	100.0%
kai_get_kernel_vec_length_constant (kai/ukernels/matmul/imatmul_clamp_f32_f32p_f32p/kai_imatmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme2_mopa.c:38)	called 34056 times	100.0%	-%	100.0%
kai_get_kernel_vec_length_constant (kai/ukernels/matmul/imatmul_clamp_f32_f32p_f32p/kai_imatmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme_mopa.c:38)	called 34056 times	100.0%	-%	100.0%
kai_get_kernel_vec_length_constant (kai/ukernels/matmul/imatmul_clamp_qai8_qai8p_qsi8cxp/kai_imatmul_clamp_qai8_qai8p2vlx4_qsi8cxp2vlx4sb_2vlx2vl_sme_mopa.c:39)	called 27972 times	100.0%	-%	100.0%
kai_get_kernel_vec_length_constant (kai/ukernels/matmul/imatmul_clamp_qai8_qai8p_qsi8cxp/kai_imatmul_clamp_qai8_qai8p2vlx4_qsi8cxpsb2vlx4_2vlx2vl_sme2_mopa.c:39)	called 27972 times	100.0%	-%	100.0%
kai_get_kernel_vec_length_constant (kai/ukernels/matmul/matmul_clamp_f16_f16p_f16p/kai_matmul_clamp_f16_f16p2vlx2_f16p2vlx2_2vlx2vl_sme2_mopa.c:40)	called 200 times	100.0%	-%	100.0%
kai_get_kernel_vec_length_constant (kai/ukernels/matmul/matmul_clamp_f16_f16p_f16p/kai_matmul_clamp_f16_f16p2vlx2_f16p2vlx2b_2vlx2vl_sme_mopa.c:40)	called 200 times	100.0%	-%	100.0%
kai_get_kernel_vec_length_constant (kai/ukernels/matmul/matmul_clamp_f32_f32p_f32p/kai_matmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme_mopa.c:39)	called 200 times	100.0%	-%	100.0%
kai_get_kernel_vec_length_constant (kai/ukernels/matmul/matmul_clamp_f32_f32p_f32p/kai_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa.c:39)	called 200 times	100.0%	-%	100.0%
kai_get_kernel_vec_length_constant (kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa.c:39)	called 386 times	100.0%	-%	100.0%
kai_get_kernel_vec_length_constant (kai/ukernels/matmul/matmul_clamp_qai8_qai8p_qsi8cxp/kai_matmul_clamp_qai8_qai8p2vlx4_qsi8cxp2vlx4sb_2vlx2vl_sme_mopa.c:40)	called 3330 times	100.0%	-%	100.0%
kai_get_kernel_vec_length_constant (kai/ukernels/matmul/matmul_clamp_qai8_qai8p_qsi8cxp/kai_matmul_clamp_qai8_qai8p2vlx4_qsi8cxpsb2vlx4_2vlx2vl_sme2_mopa.c:40)	called 3330 times	100.0%	-%	100.0%
kai_get_kr_dwconv_clamp_f32_f32_f32p1vlx1b_3x3_s1_4xc_sme2_mla (kai/ukernels/dwconv/dwconv_f32_f32_f32p/kai_dwconv_clamp_f32_f32_f32p1vlx1b_3x3_s1_4xc_sme2_mla.c:57)	not called	0.0%	-%	0.0%
kai_get_kr_matmul_clamp_bf16_qai8dxp1x8_qsi4c32p4x8_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_bf16_qai8dxp_qsi4c32p/kai_matmul_clamp_bf16_qai8dxp1x8_qsi4c32p4x8_1x4_neon_dotprod.c:114)	called 112 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_bf16_qai8dxp1x8_qsi4cxp8x8_1x8_neon_dotprod (kai/ukernels/matmul/matmul_clamp_bf16_qai8dxp_qsi4cxp/kai_matmul_clamp_bf16_qai8dxp1x8_qsi4cxp8x8_1x8_neon_dotprod.c:100)	called 392 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_bf16_qai8dxp4x8_qsi4c32p4x8_16x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_bf16_qai8dxp_qsi4c32p/kai_matmul_clamp_bf16_qai8dxp4x8_qsi4c32p4x8_16x4_neon_i8mm.c:115)	called 112 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_bf16_qai8dxp4x8_qsi4cxp8x8_8x8_neon_i8mm (kai/ukernels/matmul/matmul_clamp_bf16_qai8dxp_qsi4cxp/kai_matmul_clamp_bf16_qai8dxp4x8_qsi4cxp8x8_8x8_neon_i8mm.c:101)	called 392 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f16_bf16p8x4_bf16p12x4b_8x12_neon_mmla (kai/ukernels/matmul/matmul_clamp_f16_bf16p_bf16p/kai_matmul_clamp_f16_bf16p8x4_bf16p12x4b_8x12_neon_mmla.c:43)	called 18 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f16_f16_f16p16x1biasf16_6x16x8_neon_mla (kai/ukernels/matmul/matmul_clamp_f16_f16_f16p/kai_matmul_clamp_f16_f16_f16p16x1biasf16_6x16x8_neon_mla.c:40)	called 17 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f16_f16_f16p2vlx2b_1x16vl_sme2_dot (kai/ukernels/matmul/matmul_clamp_f16_f16_f16p/kai_matmul_clamp_f16_f16_f16p2vlx2b_1x16vl_sme2_dot.c:51)	called 46 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f16_f16_f16p2vlx2b_1x8vl_sme_mla (kai/ukernels/matmul/matmul_clamp_f16_f16_f16p/kai_matmul_clamp_f16_f16_f16p2vlx2b_1x8vl_sme_mla.c:51)	called 46 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f16_f16_f16p32x1b_6x32_neon_mla (kai/ukernels/matmul/matmul_clamp_f16_f16_f16p/kai_matmul_clamp_f16_f16_f16p32x1b_6x32_neon_mla.c:56)	called 17 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f16_f16_f16p32x1b_6x32_neon_mla_cortexa55 (kai/ukernels/matmul/matmul_clamp_f16_f16_f16p/kai_matmul_clamp_f16_f16_f16p32x1b_6x32_neon_mla_cortexa55.c:56)	called 17 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f16_f16p2vlx2_f16p2vlx2_2vlx2vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f16_f16p_f16p/kai_matmul_clamp_f16_f16p2vlx2_f16p2vlx2_2vlx2vl_sme2_mopa.c:61)	called 51 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f16_f16p2vlx2_f16p2vlx2b_2vlx2vl_sme_mopa (kai/ukernels/matmul/matmul_clamp_f16_f16p_f16p/kai_matmul_clamp_f16_f16p2vlx2_f16p2vlx2b_2vlx2vl_sme_mopa.c:61)	called 51 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f16_qai8dxp1x4_qsi4cxp4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi4cxp/kai_matmul_clamp_f16_qai8dxp1x4_qsi4cxp4x4_1x4_neon_dotprod.c:100)	called 224 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f16_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi8cxp/kai_matmul_clamp_f16_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod.c:97)	called 224 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f16_qai8dxp1x8_qsi4cxp4x8_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi4cxp/kai_matmul_clamp_f16_qai8dxp1x8_qsi4cxp4x8_1x4_neon_dotprod.c:100)	called 224 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f16_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi8cxp/kai_matmul_clamp_f16_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod.c:97)	called 224 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f16_qai8dxp4x4_qsi4cxp4x4_16x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi4cxp/kai_matmul_clamp_f16_qai8dxp4x4_qsi4cxp4x4_16x4_neon_dotprod.c:100)	called 224 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f16_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi8cxp/kai_matmul_clamp_f16_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod.c:97)	called 224 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f16_qai8dxp4x8_qsi4cxp4x8_16x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi4cxp/kai_matmul_clamp_f16_qai8dxp4x8_qsi4cxp4x8_16x4_neon_i8mm.c:100)	called 224 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f16_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi8cxp/kai_matmul_clamp_f16_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm.c:97)	called 224 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f16_qsi8d32p1vlx4_qai4c32p4vlx4_1vlx4vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p1vlx4_qai4c32p4vlx4_1vlx4vl_sme2_mopa.c:114)	called 2072 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f16_qsi8d32p1x4_qai4c32p4vlx4_1x4vl_sme2_dot (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p1x4_qai4c32p4vlx4_1x4vl_sme2_dot.c:110)	called 2072 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f16_qsi8d32p1x4_qai4c32p4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p1x4_qai4c32p4x4_1x4_neon_dotprod.c:106)	called 1036 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f16_qsi8d32p1x8_qai4c32p4x8_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p1x8_qai4c32p4x8_1x4_neon_dotprod.c:106)	called 1036 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f16_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod.c:106)	called 1036 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f16_qsi8d32p4x8_qai4c32p4x8_8x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p4x8_qai4c32p4x8_8x4_neon_i8mm.c:106)	called 1036 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f32_bf16p1x4_bf16p12x4b_1x36_neon_dot (kai/ukernels/matmul/matmul_clamp_f32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p1x4_bf16p12x4b_1x36_neon_dot.c:45)	called 24 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa.c:60)	called 10 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f32_bf16p8x4_bf16p12x4b_8x12_neon_mmla (kai/ukernels/matmul/matmul_clamp_f32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p8x4_bf16p12x4b_8x12_neon_mmla.c:44)	called 40 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f32_f32_f32p16vlx1b_1x16vl_sme2_mla (kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p16vlx1b_1x16vl_sme2_mla.c:49)	called 46 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f32_f32_f32p16x1b_6x16_neon_mla (kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p16x1b_6x16_neon_mla.c:52)	called 17 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f32_f32_f32p16x1b_6x16_neon_mla_cortexa55 (kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p16x1b_6x16_neon_mla_cortexa55.c:52)	called 17 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla (kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla.c:49)	called 46 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f32_f32_f32p2vlx1b_1x8vl_sme_mla (kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p2vlx1b_1x8vl_sme_mla.c:49)	called 46 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f32_f32_f32p8x1biasf32_6x8x4_neon_mla (kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p8x1biasf32_6x8x4_neon_mla.c:53)	called 17 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme_mopa (kai/ukernels/matmul/matmul_clamp_f32_f32p_f32p/kai_matmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme_mopa.c:60)	called 51 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f32_f32p_f32p/kai_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa.c:60)	called 51 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa.c:106)	called 308 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme_mopa (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme_mopa.c:106)	called 308 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f32_qai8dxp1vlx8_qsi4cxp4vlx8_1vlx4vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1vlx8_qsi4cxp4vlx8_1vlx4vl_sme2_mopa.c:76)	called 160 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f32_qai8dxp1x4_qsi4c32p4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x4_qsi4c32p4x4_1x4_neon_dotprod.c:113)	called 112 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f32_qai8dxp1x4_qsi4c32p8x4_1x8_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x4_qsi4c32p8x4_1x8_neon_dotprod.c:113)	called 112 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f32_qai8dxp1x4_qsi4cxp4vlx4_1x4vl_sme2_sdot (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi4cxp4vlx4_1x4vl_sme2_sdot.c:74)	called 160 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f32_qai8dxp1x4_qsi4cxp4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi4cxp4x4_1x4_neon_dotprod.c:85)	called 320 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot.c:109)	called 308 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme_dot (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme_dot.c:109)	called 308 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod.c:83)	called 308 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f32_qai8dxp1x8_qsi4c32p4x8_1x4x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x8_qsi4c32p4x8_1x4x32_neon_dotprod.c:113)	called 112 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f32_qai8dxp1x8_qsi4c32p8x8_1x8_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x8_qsi4c32p8x8_1x8_neon_dotprod.c:113)	called 112 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f32_qai8dxp1x8_qsi4c32p8x8_1x8x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x8_qsi4c32p8x8_1x8x32_neon_dotprod.c:113)	called 112 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f32_qai8dxp1x8_qsi4cxp4x8_1x4x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi4cxp4x8_1x4x32_neon_dotprod.c:69)	called 320 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f32_qai8dxp1x8_qsi4cxp8x8_1x8x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi4cxp8x8_1x8x32_neon_dotprod.c:69)	called 320 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f32_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod.c:83)	called 308 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f32_qai8dxp4x4_qsi4c32p4x4_16x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x4_qsi4c32p4x4_16x4_neon_dotprod.c:114)	called 112 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod.c:113)	called 112 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f32_qai8dxp4x4_qsi4cxp8x4_8x8x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x4_qsi4cxp8x4_8x8x32_neon_dotprod.c:71)	called 320 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f32_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod.c:87)	called 308 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f32_qai8dxp4x8_qsi4c32p4x8_16x4x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p4x8_16x4x32_neon_i8mm.c:114)	called 112 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f32_qai8dxp4x8_qsi4c32p4x8_8x4x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p4x8_8x4x32_neon_i8mm.c:113)	called 112 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f32_qai8dxp4x8_qsi4c32p8x8_4x8_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p8x8_4x8_neon_i8mm.c:113)	called 112 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f32_qai8dxp4x8_qsi4c32p8x8_4x8x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p8x8_4x8x32_neon_i8mm.c:113)	called 112 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x4_16x4x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x4_16x4x32_neon_dotprod.c:71)	called 320 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_4x4x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_4x4x32_neon_i8mm.c:68)	called 320 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_8x4x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_8x4x32_neon_i8mm.c:72)	called 320 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_4x8x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_4x8x32_neon_i8mm.c:72)	called 320 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_8x8x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_8x8x32_neon_i8mm.c:72)	called 320 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f32_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm.c:87)	called 308 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f32_qsi8d32p1vlx4_qai4c32p4vlx4_1vlx4vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p1vlx4_qai4c32p4vlx4_1vlx4vl_sme2_mopa.c:114)	called 2072 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa.c:92)	called 72 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f32_qsi8d32p1x4_qai4c32p4vlx4_1x4vl_sme2_dot (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qai4c32p4vlx4_1x4vl_sme2_dot.c:110)	called 2072 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f32_qsi8d32p1x4_qai4c32p4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qai4c32p4x4_1x4_neon_dotprod.c:104)	called 1036 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot.c:92)	called 72 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod.c:86)	called 72 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f32_qsi8d32p1x8_qai4c32p4x8_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p1x8_qai4c32p4x8_1x4_neon_dotprod.c:104)	called 1036 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod.c:71)	called 72 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f32_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod.c:104)	called 1036 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod.c:90)	called 72 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f32_qsi8d32p4x8_qai4c32p4x8_8x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qai4c32p4x8_8x4_neon_i8mm.c:104)	called 1036 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm.c:74)	called 144 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_8x4x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_8x4x32_neon_i8mm.c:74)	called 72 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_qai8_qai8_qsi8cxp2vlx4sb_1x16vl_sme2_dot (kai/ukernels/matmul/matmul_clamp_qai8_qai8_qsi8cxp/kai_matmul_clamp_qai8_qai8_qsi8cxp2vlx4sb_1x16vl_sme2_dot.c:50)	called 168 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_qai8_qai8p2vlx4_qsi8cxp2vlx4sb_2vlx2vl_sme_mopa (kai/ukernels/matmul/matmul_clamp_qai8_qai8p_qsi8cxp/kai_matmul_clamp_qai8_qai8p2vlx4_qsi8cxp2vlx4sb_2vlx2vl_sme_mopa.c:61)	called 333 times	100.0%	-%	100.0%
kai_get_kr_matmul_clamp_qai8_qai8p2vlx4_qsi8cxpsb2vlx4_2vlx2vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_qai8_qai8p_qsi8cxp/kai_matmul_clamp_qai8_qai8p2vlx4_qsi8cxpsb2vlx4_2vlx2vl_sme2_mopa.c:61)	called 333 times	100.0%	-%	100.0%
kai_get_lhs_offset_lhs_pack_bf16p2vlx2_f32_sme (kai/ukernels/matmul/pack/kai_lhs_pack_bf16p2vlx2_f32_sme.c:36)	called 46 times	100.0%	-%	57.0%
kai_get_lhs_offset_lhs_pack_bf16p8x4_f16_neon (kai/ukernels/matmul/pack/kai_lhs_pack_bf16p8x4_f16_neon.c:32)	called 176 times	100.0%	-%	57.0%
kai_get_lhs_offset_lhs_pack_f32p2vlx1_f32_sme (kai/ukernels/matmul/pack/kai_lhs_pack_f32p2vlx1_f32_sme.c:55)	called 34 times	100.0%	-%	57.0%
kai_get_lhs_offset_lhs_pack_x16p2vlx2_x16_sme (kai/ukernels/matmul/pack/kai_lhs_pack_x16p2vlx2_x16_sme.c:55)	called 126 times	100.0%	-%	57.0%
kai_get_lhs_offset_lhs_pack_x8p2vlx4_x8_sme (kai/ukernels/matmul/pack/kai_lhs_pack_x8p2vlx4_x8_sme.c:55)	called 666 times	100.0%	-%	57.0%
kai_get_lhs_offset_lhs_quant_pack_bf16p1x4_f32_neon (kai/ukernels/matmul/pack/kai_lhs_quant_pack_bf16p1x4_f32_neon.c:28)	called 92 times	100.0%	-%	57.0%
kai_get_lhs_offset_lhs_quant_pack_bf16p8x4_f32_neon (kai/ukernels/matmul/pack/kai_lhs_quant_pack_bf16p8x4_f32_neon.c:30)	called 92 times	100.0%	-%	100.0%
kai_get_lhs_offset_lhs_quant_pack_qai8dxp_bf16_neon (kai/ukernels/matmul/pack/kai_lhs_quant_pack_qai8dxp_bf16_neon.c:43)	called 960 times	100.0%	-%	100.0%
kai_get_lhs_offset_lhs_quant_pack_qai8dxp_f16_neon (kai/ukernels/matmul/pack/kai_lhs_quant_pack_qai8dxp_f16_neon.c:50)	called 1120 times	100.0%	-%	100.0%
kai_get_lhs_offset_lhs_quant_pack_qai8dxp_f32 (kai/ukernels/matmul/pack/kai_lhs_quant_pack_qai8dxp_f32.c:43)	called 3940 times	100.0%	-%	100.0%
kai_get_lhs_offset_lhs_quant_pack_qsi8d32p4x8sb_f32_neon (kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p4x8sb_f32_neon.c:35)	called 23 times	100.0%	-%	100.0%
kai_get_lhs_offset_lhs_quant_pack_qsi8d32p_f32 (kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p_f32.c:35)	called 75 times	100.0%	-%	100.0%
kai_get_lhs_offset_lhs_quant_pack_qsi8d32p_f32_neon (kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p_f32_neon.c:41)	called 26 times	100.0%	-%	100.0%
kai_get_lhs_offset_lhs_quant_pack_qsi8d32pscalef32_f16_neon (kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32pscalef32_f16_neon.c:44)	called 10904 times	100.0%	-%	100.0%
kai_get_lhs_offset_lhs_quant_pack_qsi8d32pscalef32_f32_neon (kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32pscalef32_f32_neon.c:44)	called 10904 times	100.0%	-%	100.0%
kai_get_lhs_offset_matmul_clamp_f16_f16_f16p16x1biasf16_6x16x8_neon_mla (kai/ukernels/matmul/matmul_clamp_f16_f16_f16p/kai_matmul_clamp_f16_f16_f16p16x1biasf16_6x16x8_neon_mla.c:48)	called 16 times	100.0%	-%	57.0%
kai_get_lhs_offset_matmul_clamp_f16_f16_f16p2vlx2b_1x8vl_sme_mla (kai/ukernels/matmul/matmul_clamp_f16_f16_f16p/kai_matmul_clamp_f16_f16_f16p2vlx2b_1x8vl_sme_mla.c:59)	not called	0.0%	-%	0.0%
kai_get_lhs_offset_matmul_clamp_f16_f16_f16p32x1b_6x32_neon_mla (kai/ukernels/matmul/matmul_clamp_f16_f16_f16p/kai_matmul_clamp_f16_f16_f16p32x1b_6x32_neon_mla.c:64)	called 16 times	100.0%	-%	57.0%
kai_get_lhs_offset_matmul_clamp_f16_f16_f16p32x1b_6x32_neon_mla_cortexa55 (kai/ukernels/matmul/matmul_clamp_f16_f16_f16p/kai_matmul_clamp_f16_f16_f16p32x1b_6x32_neon_mla_cortexa55.c:64)	called 16 times	100.0%	-%	57.0%
kai_get_lhs_offset_matmul_clamp_f32_f32_f32p16vlx1b_1x16vl_sme2_mla (kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p16vlx1b_1x16vl_sme2_mla.c:57)	called 46 times	100.0%	-%	57.0%
kai_get_lhs_offset_matmul_clamp_f32_f32_f32p16x1b_6x16_neon_mla (kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p16x1b_6x16_neon_mla.c:60)	called 16 times	100.0%	-%	57.0%
kai_get_lhs_offset_matmul_clamp_f32_f32_f32p16x1b_6x16_neon_mla_cortexa55 (kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p16x1b_6x16_neon_mla_cortexa55.c:60)	called 16 times	100.0%	-%	57.0%
kai_get_lhs_offset_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla (kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla.c:57)	called 46 times	100.0%	-%	57.0%
kai_get_lhs_offset_matmul_clamp_f32_f32_f32p2vlx1b_1x8vl_sme_mla (kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p2vlx1b_1x8vl_sme_mla.c:57)	called 46 times	100.0%	-%	57.0%
kai_get_lhs_offset_matmul_clamp_f32_f32_f32p8x1biasf32_6x8x4_neon_mla (kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p8x1biasf32_6x8x4_neon_mla.c:61)	called 16 times	100.0%	-%	57.0%
kai_get_lhs_offset_matmul_clamp_qai8_qai8_qsi8cxp2vlx4sb_1x16vl_sme2_dot (kai/ukernels/matmul/matmul_clamp_qai8_qai8_qsi8cxp/kai_matmul_clamp_qai8_qai8_qsi8cxp2vlx4sb_1x16vl_sme2_dot.c:58)	not called	0.0%	-%	0.0%
kai_get_lhs_packed_offset_imatmul_clamp_f16_f16p2vlx2_f16p2vlx2_2vlx2vl_sme2_mopa (kai/ukernels/matmul/imatmul_clamp_f16_f16p_f16p/kai_imatmul_clamp_f16_f16p2vlx2_f16p2vlx2_2vlx2vl_sme2_mopa.c:52)	called 5676 times	100.0%	-%	57.0%
kai_get_lhs_packed_offset_imatmul_clamp_f16_f16p2vlx2_f16p2vlx2b_2vlx2vl_sme_mopa (kai/ukernels/matmul/imatmul_clamp_f16_f16p_f16p/kai_imatmul_clamp_f16_f16p2vlx2_f16p2vlx2b_2vlx2vl_sme_mopa.c:52)	called 5676 times	100.0%	-%	57.0%
kai_get_lhs_packed_offset_imatmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme2_mopa (kai/ukernels/matmul/imatmul_clamp_f32_f32p_f32p/kai_imatmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme2_mopa.c:51)	called 5676 times	100.0%	-%	57.0%
kai_get_lhs_packed_offset_imatmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme_mopa (kai/ukernels/matmul/imatmul_clamp_f32_f32p_f32p/kai_imatmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme_mopa.c:51)	called 5676 times	100.0%	-%	57.0%
kai_get_lhs_packed_offset_imatmul_clamp_qai8_qai8p2vlx4_qsi8cxp2vlx4sb_2vlx2vl_sme_mopa (kai/ukernels/matmul/imatmul_clamp_qai8_qai8p_qsi8cxp/kai_imatmul_clamp_qai8_qai8p2vlx4_qsi8cxp2vlx4sb_2vlx2vl_sme_mopa.c:52)	called 4662 times	100.0%	-%	57.0%
kai_get_lhs_packed_offset_imatmul_clamp_qai8_qai8p2vlx4_qsi8cxpsb2vlx4_2vlx2vl_sme2_mopa (kai/ukernels/matmul/imatmul_clamp_qai8_qai8p_qsi8cxp/kai_imatmul_clamp_qai8_qai8p2vlx4_qsi8cxpsb2vlx4_2vlx2vl_sme2_mopa.c:52)	called 4662 times	100.0%	-%	57.0%
kai_get_lhs_packed_offset_lhs_imatmul_pack_x16p2vlx2_x16p_sme (kai/ukernels/matmul/pack/kai_lhs_imatmul_pack_x16p2vlx2_x16p_sme.c:34)	called 22704 times	100.0%	-%	57.0%
kai_get_lhs_packed_offset_lhs_imatmul_pack_x32p2vlx1_x32p_sme (kai/ukernels/matmul/pack/kai_lhs_imatmul_pack_x32p2vlx1_x32p_sme.c:34)	called 22704 times	100.0%	-%	57.0%
kai_get_lhs_packed_offset_lhs_imatmul_pack_x8p2vlx4_x8p_sme (kai/ukernels/matmul/pack/kai_lhs_imatmul_pack_x8p2vlx4_x8p_sme.c:34)	called 18648 times	100.0%	-%	57.0%
kai_get_lhs_packed_offset_lhs_pack_bf16p2vlx2_f32_sme (kai/ukernels/matmul/pack/kai_lhs_pack_bf16p2vlx2_f32_sme.c:42)	not called	0.0%	-%	0.0%
kai_get_lhs_packed_offset_lhs_pack_bf16p8x4_f16_neon (kai/ukernels/matmul/pack/kai_lhs_pack_bf16p8x4_f16_neon.c:38)	not called	0.0%	-%	0.0%
kai_get_lhs_packed_offset_lhs_pack_f32p2vlx1_f32_sme (kai/ukernels/matmul/pack/kai_lhs_pack_f32p2vlx1_f32_sme.c:61)	not called	0.0%	-%	0.0%
kai_get_lhs_packed_offset_lhs_pack_x16p2vlx2_x16_sme (kai/ukernels/matmul/pack/kai_lhs_pack_x16p2vlx2_x16_sme.c:61)	not called	0.0%	-%	0.0%
kai_get_lhs_packed_offset_lhs_pack_x8p2vlx4_x8_sme (kai/ukernels/matmul/pack/kai_lhs_pack_x8p2vlx4_x8_sme.c:61)	called 798 times	100.0%	-%	52.0%
kai_get_lhs_packed_offset_lhs_quant_pack_bf16p1x4_f32_neon (kai/ukernels/matmul/pack/kai_lhs_quant_pack_bf16p1x4_f32_neon.c:33)	not called	0.0%	-%	0.0%
kai_get_lhs_packed_offset_lhs_quant_pack_bf16p8x4_f32_neon (kai/ukernels/matmul/pack/kai_lhs_quant_pack_bf16p8x4_f32_neon.c:34)	not called	0.0%	-%	0.0%
kai_get_lhs_packed_offset_lhs_quant_pack_qai8dxp_bf16_neon (kai/ukernels/matmul/pack/kai_lhs_quant_pack_qai8dxp_bf16_neon.c:47)	called 960 times	100.0%	-%	100.0%
kai_get_lhs_packed_offset_lhs_quant_pack_qai8dxp_f16_neon (kai/ukernels/matmul/pack/kai_lhs_quant_pack_qai8dxp_f16_neon.c:54)	called 1120 times	100.0%	-%	100.0%
kai_get_lhs_packed_offset_lhs_quant_pack_qai8dxp_f32 (kai/ukernels/matmul/pack/kai_lhs_quant_pack_qai8dxp_f32.c:47)	called 5356 times	100.0%	-%	100.0%
kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p4x8sb_f32_neon (kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p4x8sb_f32_neon.c:39)	called 46 times	100.0%	-%	52.0%
kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32 (kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p_f32.c:39)	called 190 times	100.0%	-%	52.0%
kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32_neon (kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p_f32_neon.c:45)	called 72 times	100.0%	-%	52.0%
kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32pscalef32_f16_neon (kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32pscalef32_f16_neon.c:47)	called 21808 times	100.0%	-%	52.0%
kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32pscalef32_f32_neon (kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32pscalef32_f32_neon.c:48)	called 10904 times	100.0%	-%	52.0%
kai_get_lhs_packed_offset_matmul_clamp_bf16_qai8dxp1x8_qsi4c32p4x8_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_bf16_qai8dxp_qsi4c32p/kai_matmul_clamp_bf16_qai8dxp1x8_qsi4c32p4x8_1x4_neon_dotprod.c:122)	called 108 times	100.0%	-%	57.0%
kai_get_lhs_packed_offset_matmul_clamp_bf16_qai8dxp1x8_qsi4cxp8x8_1x8_neon_dotprod (kai/ukernels/matmul/matmul_clamp_bf16_qai8dxp_qsi4cxp/kai_matmul_clamp_bf16_qai8dxp1x8_qsi4cxp8x8_1x8_neon_dotprod.c:108)	called 356 times	100.0%	-%	57.0%
kai_get_lhs_packed_offset_matmul_clamp_bf16_qai8dxp4x8_qsi4c32p4x8_16x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_bf16_qai8dxp_qsi4c32p/kai_matmul_clamp_bf16_qai8dxp4x8_qsi4c32p4x8_16x4_neon_i8mm.c:123)	called 108 times	100.0%	-%	57.0%
kai_get_lhs_packed_offset_matmul_clamp_bf16_qai8dxp4x8_qsi4cxp8x8_8x8_neon_i8mm (kai/ukernels/matmul/matmul_clamp_bf16_qai8dxp_qsi4cxp/kai_matmul_clamp_bf16_qai8dxp4x8_qsi4cxp8x8_8x8_neon_i8mm.c:109)	called 388 times	100.0%	-%	57.0%
kai_get_lhs_packed_offset_matmul_clamp_f16_bf16p8x4_bf16p12x4b_8x12_neon_mmla (kai/ukernels/matmul/matmul_clamp_f16_bf16p_bf16p/kai_matmul_clamp_f16_bf16p8x4_bf16p12x4b_8x12_neon_mmla.c:51)	called 84 times	100.0%	-%	57.0%
kai_get_lhs_packed_offset_matmul_clamp_f16_f16_f16p2vlx2b_1x16vl_sme2_dot (kai/ukernels/matmul/matmul_clamp_f16_f16_f16p/kai_matmul_clamp_f16_f16_f16p2vlx2b_1x16vl_sme2_dot.c:59)	not called	0.0%	-%	0.0%
kai_get_lhs_packed_offset_matmul_clamp_f16_f16p2vlx2_f16p2vlx2_2vlx2vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f16_f16p_f16p/kai_matmul_clamp_f16_f16p2vlx2_f16p2vlx2_2vlx2vl_sme2_mopa.c:69)	called 33 times	100.0%	-%	57.0%
kai_get_lhs_packed_offset_matmul_clamp_f16_f16p2vlx2_f16p2vlx2b_2vlx2vl_sme_mopa (kai/ukernels/matmul/matmul_clamp_f16_f16p_f16p/kai_matmul_clamp_f16_f16p2vlx2_f16p2vlx2b_2vlx2vl_sme_mopa.c:69)	called 33 times	100.0%	-%	57.0%
kai_get_lhs_packed_offset_matmul_clamp_f16_qai8dxp1x4_qsi4cxp4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi4cxp/kai_matmul_clamp_f16_qai8dxp1x4_qsi4cxp4x4_1x4_neon_dotprod.c:108)	called 58 times	100.0%	-%	57.0%
kai_get_lhs_packed_offset_matmul_clamp_f16_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi8cxp/kai_matmul_clamp_f16_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod.c:105)	called 58 times	100.0%	-%	57.0%
kai_get_lhs_packed_offset_matmul_clamp_f16_qai8dxp1x8_qsi4cxp4x8_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi4cxp/kai_matmul_clamp_f16_qai8dxp1x8_qsi4cxp4x8_1x4_neon_dotprod.c:108)	called 58 times	100.0%	-%	57.0%
kai_get_lhs_packed_offset_matmul_clamp_f16_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi8cxp/kai_matmul_clamp_f16_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod.c:105)	called 58 times	100.0%	-%	57.0%
kai_get_lhs_packed_offset_matmul_clamp_f16_qai8dxp4x4_qsi4cxp4x4_16x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi4cxp/kai_matmul_clamp_f16_qai8dxp4x4_qsi4cxp4x4_16x4_neon_dotprod.c:108)	called 222 times	100.0%	-%	57.0%
kai_get_lhs_packed_offset_matmul_clamp_f16_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi8cxp/kai_matmul_clamp_f16_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod.c:105)	called 222 times	100.0%	-%	57.0%
kai_get_lhs_packed_offset_matmul_clamp_f16_qai8dxp4x8_qsi4cxp4x8_16x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi4cxp/kai_matmul_clamp_f16_qai8dxp4x8_qsi4cxp4x8_16x4_neon_i8mm.c:108)	called 222 times	100.0%	-%	57.0%
kai_get_lhs_packed_offset_matmul_clamp_f16_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi8cxp/kai_matmul_clamp_f16_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm.c:105)	called 222 times	100.0%	-%	57.0%
kai_get_lhs_packed_offset_matmul_clamp_f16_qsi8d32p1vlx4_qai4c32p4vlx4_1vlx4vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p1vlx4_qai4c32p4vlx4_1vlx4vl_sme2_mopa.c:122)	called 1032 times	100.0%	-%	55.0%
kai_get_lhs_packed_offset_matmul_clamp_f16_qsi8d32p1x4_qai4c32p4vlx4_1x4vl_sme2_dot (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p1x4_qai4c32p4vlx4_1x4vl_sme2_dot.c:118)	called 276 times	100.0%	-%	55.0%
kai_get_lhs_packed_offset_matmul_clamp_f16_qsi8d32p1x4_qai4c32p4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p1x4_qai4c32p4x4_1x4_neon_dotprod.c:114)	called 138 times	100.0%	-%	57.0%
kai_get_lhs_packed_offset_matmul_clamp_f16_qsi8d32p1x8_qai4c32p4x8_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p1x8_qai4c32p4x8_1x4_neon_dotprod.c:114)	called 138 times	100.0%	-%	57.0%
kai_get_lhs_packed_offset_matmul_clamp_f16_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod.c:114)	called 516 times	100.0%	-%	57.0%
kai_get_lhs_packed_offset_matmul_clamp_f16_qsi8d32p4x8_qai4c32p4x8_8x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p4x8_qai4c32p4x8_8x4_neon_i8mm.c:114)	called 516 times	100.0%	-%	57.0%
kai_get_lhs_packed_offset_matmul_clamp_f32_bf16p1x4_bf16p12x4b_1x36_neon_dot (kai/ukernels/matmul/matmul_clamp_f32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p1x4_bf16p12x4b_1x36_neon_dot.c:53)	called 92 times	100.0%	-%	57.0%
kai_get_lhs_packed_offset_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa.c:68)	called 46 times	100.0%	-%	57.0%
kai_get_lhs_packed_offset_matmul_clamp_f32_bf16p8x4_bf16p12x4b_8x12_neon_mmla (kai/ukernels/matmul/matmul_clamp_f32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p8x4_bf16p12x4b_8x12_neon_mmla.c:52)	called 184 times	100.0%	-%	57.0%
kai_get_lhs_packed_offset_matmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme_mopa (kai/ukernels/matmul/matmul_clamp_f32_f32p_f32p/kai_matmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme_mopa.c:68)	called 33 times	100.0%	-%	57.0%
kai_get_lhs_packed_offset_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f32_f32p_f32p/kai_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa.c:68)	called 33 times	100.0%	-%	57.0%
kai_get_lhs_packed_offset_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa.c:114)	called 231 times	100.0%	-%	57.0%
kai_get_lhs_packed_offset_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme_mopa (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme_mopa.c:114)	called 231 times	100.0%	-%	57.0%
kai_get_lhs_packed_offset_matmul_clamp_f32_qai8dxp1vlx8_qsi4cxp4vlx8_1vlx4vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1vlx8_qsi4cxp4vlx8_1vlx4vl_sme2_mopa.c:84)	called 120 times	100.0%	-%	57.0%
kai_get_lhs_packed_offset_matmul_clamp_f32_qai8dxp1x4_qsi4c32p4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x4_qsi4c32p4x4_1x4_neon_dotprod.c:121)	called 108 times	100.0%	-%	57.0%
kai_get_lhs_packed_offset_matmul_clamp_f32_qai8dxp1x4_qsi4c32p8x4_1x8_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x4_qsi4c32p8x4_1x8_neon_dotprod.c:121)	called 108 times	100.0%	-%	57.0%
kai_get_lhs_packed_offset_matmul_clamp_f32_qai8dxp1x4_qsi4cxp4vlx4_1x4vl_sme2_sdot (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi4cxp4vlx4_1x4vl_sme2_sdot.c:82)	called 120 times	100.0%	-%	57.0%
kai_get_lhs_packed_offset_matmul_clamp_f32_qai8dxp1x4_qsi4cxp4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi4cxp4x4_1x4_neon_dotprod.c:93)	called 240 times	100.0%	-%	57.0%
kai_get_lhs_packed_offset_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot.c:117)	called 211 times	100.0%	-%	57.0%
kai_get_lhs_packed_offset_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme_dot (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme_dot.c:117)	called 211 times	100.0%	-%	57.0%
kai_get_lhs_packed_offset_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod.c:91)	called 211 times	100.0%	-%	57.0%
kai_get_lhs_packed_offset_matmul_clamp_f32_qai8dxp1x8_qsi4c32p4x8_1x4x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x8_qsi4c32p4x8_1x4x32_neon_dotprod.c:121)	called 108 times	100.0%	-%	57.0%
kai_get_lhs_packed_offset_matmul_clamp_f32_qai8dxp1x8_qsi4c32p8x8_1x8_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x8_qsi4c32p8x8_1x8_neon_dotprod.c:121)	called 108 times	100.0%	-%	57.0%
kai_get_lhs_packed_offset_matmul_clamp_f32_qai8dxp1x8_qsi4c32p8x8_1x8x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x8_qsi4c32p8x8_1x8x32_neon_dotprod.c:121)	called 108 times	100.0%	-%	57.0%
kai_get_lhs_packed_offset_matmul_clamp_f32_qai8dxp1x8_qsi4cxp4x8_1x4x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi4cxp4x8_1x4x32_neon_dotprod.c:77)	called 240 times	100.0%	-%	57.0%
kai_get_lhs_packed_offset_matmul_clamp_f32_qai8dxp1x8_qsi4cxp8x8_1x8x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi4cxp8x8_1x8x32_neon_dotprod.c:77)	called 240 times	100.0%	-%	57.0%
kai_get_lhs_packed_offset_matmul_clamp_f32_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod.c:91)	called 211 times	100.0%	-%	57.0%
kai_get_lhs_packed_offset_matmul_clamp_f32_qai8dxp4x4_qsi4c32p4x4_16x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x4_qsi4c32p4x4_16x4_neon_dotprod.c:122)	called 108 times	100.0%	-%	57.0%
kai_get_lhs_packed_offset_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod.c:121)	called 108 times	100.0%	-%	57.0%
kai_get_lhs_packed_offset_matmul_clamp_f32_qai8dxp4x4_qsi4cxp8x4_8x8x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x4_qsi4cxp8x4_8x8x32_neon_dotprod.c:79)	called 240 times	100.0%	-%	57.0%
kai_get_lhs_packed_offset_matmul_clamp_f32_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod.c:95)	called 231 times	100.0%	-%	57.0%
kai_get_lhs_packed_offset_matmul_clamp_f32_qai8dxp4x8_qsi4c32p4x8_16x4x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p4x8_16x4x32_neon_i8mm.c:122)	called 108 times	100.0%	-%	57.0%
kai_get_lhs_packed_offset_matmul_clamp_f32_qai8dxp4x8_qsi4c32p4x8_8x4x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p4x8_8x4x32_neon_i8mm.c:121)	called 108 times	100.0%	-%	57.0%
kai_get_lhs_packed_offset_matmul_clamp_f32_qai8dxp4x8_qsi4c32p8x8_4x8_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p8x8_4x8_neon_i8mm.c:121)	called 108 times	100.0%	-%	57.0%
kai_get_lhs_packed_offset_matmul_clamp_f32_qai8dxp4x8_qsi4c32p8x8_4x8x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p8x8_4x8x32_neon_i8mm.c:121)	called 108 times	100.0%	-%	57.0%
kai_get_lhs_packed_offset_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x4_16x4x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x4_16x4x32_neon_dotprod.c:79)	called 240 times	100.0%	-%	57.0%
kai_get_lhs_packed_offset_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_4x4x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_4x4x32_neon_i8mm.c:76)	called 240 times	100.0%	-%	57.0%
kai_get_lhs_packed_offset_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_8x4x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_8x4x32_neon_i8mm.c:80)	called 240 times	100.0%	-%	57.0%
kai_get_lhs_packed_offset_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_4x8x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_4x8x32_neon_i8mm.c:80)	called 240 times	100.0%	-%	57.0%
kai_get_lhs_packed_offset_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_8x8x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_8x8x32_neon_i8mm.c:80)	called 240 times	100.0%	-%	57.0%
kai_get_lhs_packed_offset_matmul_clamp_f32_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm.c:95)	called 231 times	100.0%	-%	57.0%
kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1vlx4_qai4c32p4vlx4_1vlx4vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p1vlx4_qai4c32p4vlx4_1vlx4vl_sme2_mopa.c:122)	called 1032 times	100.0%	-%	55.0%
kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa.c:100)	called 46 times	100.0%	-%	55.0%
kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1x4_qai4c32p4vlx4_1x4vl_sme2_dot (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qai4c32p4vlx4_1x4vl_sme2_dot.c:118)	called 276 times	100.0%	-%	55.0%
kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1x4_qai4c32p4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qai4c32p4x4_1x4_neon_dotprod.c:112)	called 138 times	100.0%	-%	57.0%
kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot.c:100)	called 26 times	100.0%	-%	55.0%
kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod.c:94)	called 26 times	100.0%	-%	57.0%
kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1x8_qai4c32p4x8_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p1x8_qai4c32p4x8_1x4_neon_dotprod.c:112)	called 138 times	100.0%	-%	57.0%
kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod.c:79)	called 26 times	100.0%	-%	51.0%
kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod.c:112)	called 516 times	100.0%	-%	57.0%
kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod.c:98)	called 46 times	100.0%	-%	57.0%
kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p4x8_qai4c32p4x8_8x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qai4c32p4x8_8x4_neon_i8mm.c:112)	called 516 times	100.0%	-%	57.0%
kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm.c:82)	called 92 times	100.0%	-%	51.0%
kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_8x4x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_8x4x32_neon_i8mm.c:82)	called 46 times	100.0%	-%	51.0%
kai_get_lhs_packed_offset_matmul_clamp_qai8_qai8p2vlx4_qsi8cxp2vlx4sb_2vlx2vl_sme_mopa (kai/ukernels/matmul/matmul_clamp_qai8_qai8p_qsi8cxp/kai_matmul_clamp_qai8_qai8p2vlx4_qsi8cxp2vlx4sb_2vlx2vl_sme_mopa.c:69)	called 333 times	100.0%	-%	57.0%
kai_get_lhs_packed_offset_matmul_clamp_qai8_qai8p2vlx4_qsi8cxpsb2vlx4_2vlx2vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_qai8_qai8p_qsi8cxp/kai_matmul_clamp_qai8_qai8p2vlx4_qsi8cxpsb2vlx4_2vlx2vl_sme2_mopa.c:69)	called 333 times	100.0%	-%	57.0%
kai_get_lhs_packed_size_lhs_imatmul_pack_x16p2vlx2_x16p_sme (kai/ukernels/matmul/pack/kai_lhs_imatmul_pack_x16p2vlx2_x16p_sme.c:41)	called 11352 times	100.0%	-%	100.0%
kai_get_lhs_packed_size_lhs_imatmul_pack_x32p2vlx1_x32p_sme (kai/ukernels/matmul/pack/kai_lhs_imatmul_pack_x32p2vlx1_x32p_sme.c:41)	called 11352 times	100.0%	-%	100.0%
kai_get_lhs_packed_size_lhs_imatmul_pack_x8p2vlx4_x8p_sme (kai/ukernels/matmul/pack/kai_lhs_imatmul_pack_x8p2vlx4_x8p_sme.c:41)	called 9324 times	100.0%	-%	100.0%
kai_get_lhs_packed_size_lhs_pack_bf16p2vlx2_f32_sme (kai/ukernels/matmul/pack/kai_lhs_pack_bf16p2vlx2_f32_sme.c:55)	called 46 times	100.0%	-%	52.0%
kai_get_lhs_packed_size_lhs_pack_bf16p8x4_f16_neon (kai/ukernels/matmul/pack/kai_lhs_pack_bf16p8x4_f16_neon.c:47)	called 176 times	100.0%	-%	52.0%
kai_get_lhs_packed_size_lhs_pack_f32p2vlx1_f32_sme (kai/ukernels/matmul/pack/kai_lhs_pack_f32p2vlx1_f32_sme.c:74)	called 34 times	100.0%	-%	52.0%
kai_get_lhs_packed_size_lhs_pack_x16p2vlx2_x16_sme (kai/ukernels/matmul/pack/kai_lhs_pack_x16p2vlx2_x16_sme.c:74)	called 34 times	100.0%	-%	52.0%
kai_get_lhs_packed_size_lhs_pack_x8p2vlx4_x8_sme (kai/ukernels/matmul/pack/kai_lhs_pack_x8p2vlx4_x8_sme.c:74)	called 666 times	100.0%	-%	52.0%
kai_get_lhs_packed_size_lhs_quant_pack_bf16p1x4_f32_neon (kai/ukernels/matmul/pack/kai_lhs_quant_pack_bf16p1x4_f32_neon.c:44)	called 92 times	100.0%	-%	52.0%
kai_get_lhs_packed_size_lhs_quant_pack_bf16p8x4_f32_neon (kai/ukernels/matmul/pack/kai_lhs_quant_pack_bf16p8x4_f32_neon.c:45)	called 92 times	100.0%	-%	52.0%
kai_get_lhs_packed_size_lhs_quant_pack_qai8dxp_bf16_neon (kai/ukernels/matmul/pack/kai_lhs_quant_pack_qai8dxp_bf16_neon.c:55)	called 960 times	100.0%	-%	100.0%
kai_get_lhs_packed_size_lhs_quant_pack_qai8dxp_f16_neon (kai/ukernels/matmul/pack/kai_lhs_quant_pack_qai8dxp_f16_neon.c:60)	called 1120 times	100.0%	-%	100.0%
kai_get_lhs_packed_size_lhs_quant_pack_qai8dxp_f32 (kai/ukernels/matmul/pack/kai_lhs_quant_pack_qai8dxp_f32.c:52)	called 3940 times	100.0%	-%	100.0%
kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p4x8sb_f32_neon (kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p4x8sb_f32_neon.c:51)	called 23 times	100.0%	-%	52.0%
kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32 (kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p_f32.c:51)	called 75 times	100.0%	-%	52.0%
kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32_neon (kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p_f32_neon.c:59)	called 26 times	100.0%	-%	52.0%
kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32pscalef32_f16_neon (kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32pscalef32_f16_neon.c:55)	called 4762504 times	100.0%	-%	52.0%
kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32pscalef32_f32_neon (kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32pscalef32_f32_neon.c:59)	called 4762504 times	100.0%	-%	52.0%
kai_get_lhs_packed_stride (kai/ukernels/matmul/matmul_clamp_bf16_qai8dxp_qsi4c32p/kai_matmul_clamp_bf16_qai8dxp1x8_qsi4c32p4x8_1x4_neon_dotprod.c:72)	called 108 times	100.0%	-%	100.0%
kai_get_lhs_packed_stride (kai/ukernels/matmul/matmul_clamp_bf16_qai8dxp_qsi4c32p/kai_matmul_clamp_bf16_qai8dxp4x8_qsi4c32p4x8_16x4_neon_i8mm.c:73)	called 108 times	100.0%	-%	100.0%
kai_get_lhs_packed_stride (kai/ukernels/matmul/matmul_clamp_bf16_qai8dxp_qsi4cxp/kai_matmul_clamp_bf16_qai8dxp1x8_qsi4cxp8x8_1x8_neon_dotprod.c:60)	called 356 times	100.0%	-%	100.0%
kai_get_lhs_packed_stride (kai/ukernels/matmul/matmul_clamp_bf16_qai8dxp_qsi4cxp/kai_matmul_clamp_bf16_qai8dxp4x8_qsi4cxp8x8_8x8_neon_i8mm.c:61)	called 388 times	100.0%	-%	100.0%
kai_get_lhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi4cxp/kai_matmul_clamp_f16_qai8dxp1x4_qsi4cxp4x4_1x4_neon_dotprod.c:60)	called 58 times	100.0%	-%	100.0%
kai_get_lhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi4cxp/kai_matmul_clamp_f16_qai8dxp1x8_qsi4cxp4x8_1x4_neon_dotprod.c:60)	called 58 times	100.0%	-%	100.0%
kai_get_lhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi4cxp/kai_matmul_clamp_f16_qai8dxp4x4_qsi4cxp4x4_16x4_neon_dotprod.c:60)	called 222 times	100.0%	-%	100.0%
kai_get_lhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi4cxp/kai_matmul_clamp_f16_qai8dxp4x8_qsi4cxp4x8_16x4_neon_i8mm.c:60)	called 222 times	100.0%	-%	100.0%
kai_get_lhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi8cxp/kai_matmul_clamp_f16_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod.c:57)	called 58 times	100.0%	-%	100.0%
kai_get_lhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi8cxp/kai_matmul_clamp_f16_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod.c:57)	called 58 times	100.0%	-%	100.0%
kai_get_lhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi8cxp/kai_matmul_clamp_f16_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod.c:57)	called 222 times	100.0%	-%	100.0%
kai_get_lhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi8cxp/kai_matmul_clamp_f16_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm.c:57)	called 222 times	100.0%	-%	100.0%
kai_get_lhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p1vlx4_qai4c32p4vlx4_1vlx4vl_sme2_mopa.c:78)	called 2065 times	100.0%	-%	100.0%
kai_get_lhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p1x4_qai4c32p4vlx4_1x4vl_sme2_dot.c:74)	called 276 times	100.0%	-%	100.0%
kai_get_lhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p1x4_qai4c32p4x4_1x4_neon_dotprod.c:73)	called 138 times	100.0%	-%	100.0%
kai_get_lhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p1x8_qai4c32p4x8_1x4_neon_dotprod.c:73)	called 138 times	100.0%	-%	100.0%
kai_get_lhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod.c:73)	called 516 times	100.0%	-%	100.0%
kai_get_lhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p4x8_qai4c32p4x8_8x4_neon_i8mm.c:73)	called 516 times	100.0%	-%	100.0%
kai_get_lhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x4_qsi4c32p4x4_1x4_neon_dotprod.c:71)	called 108 times	100.0%	-%	100.0%
kai_get_lhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x4_qsi4c32p8x4_1x8_neon_dotprod.c:71)	called 108 times	100.0%	-%	100.0%
kai_get_lhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x8_qsi4c32p4x8_1x4x32_neon_dotprod.c:71)	called 108 times	100.0%	-%	100.0%
kai_get_lhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x8_qsi4c32p8x8_1x8_neon_dotprod.c:71)	called 108 times	100.0%	-%	100.0%
kai_get_lhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x8_qsi4c32p8x8_1x8x32_neon_dotprod.c:71)	called 108 times	100.0%	-%	100.0%
kai_get_lhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x4_qsi4c32p4x4_16x4_neon_dotprod.c:72)	called 108 times	100.0%	-%	100.0%
kai_get_lhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod.c:71)	called 108 times	100.0%	-%	100.0%
kai_get_lhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p4x8_16x4x32_neon_i8mm.c:72)	called 108 times	100.0%	-%	100.0%
kai_get_lhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p4x8_8x4x32_neon_i8mm.c:71)	called 108 times	100.0%	-%	100.0%
kai_get_lhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p8x8_4x8_neon_i8mm.c:71)	called 108 times	100.0%	-%	100.0%
kai_get_lhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p8x8_4x8x32_neon_i8mm.c:71)	called 108 times	100.0%	-%	100.0%
kai_get_lhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1vlx8_qsi4cxp4vlx8_1vlx4vl_sme2_mopa.c:42)	called 201 times	100.0%	-%	55.0%
kai_get_lhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi4cxp4vlx4_1x4vl_sme2_sdot.c:39)	called 201 times	100.0%	-%	55.0%
kai_get_lhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi4cxp4x4_1x4_neon_dotprod.c:45)	called 240 times	100.0%	-%	100.0%
kai_get_lhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa.c:63)	called 386 times	100.0%	-%	55.0%
kai_get_lhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme_mopa.c:63)	called 386 times	100.0%	-%	55.0%
kai_get_lhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot.c:64)	called 346 times	100.0%	-%	55.0%
kai_get_lhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme_dot.c:64)	called 346 times	100.0%	-%	55.0%
kai_get_lhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p1vlx4_qai4c32p4vlx4_1vlx4vl_sme2_mopa.c:78)	called 2065 times	100.0%	-%	100.0%
kai_get_lhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qai4c32p4vlx4_1x4vl_sme2_dot.c:74)	called 276 times	100.0%	-%	100.0%
kai_get_lhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qai4c32p4x4_1x4_neon_dotprod.c:71)	called 138 times	100.0%	-%	100.0%
kai_get_lhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p1x8_qai4c32p4x8_1x4_neon_dotprod.c:71)	called 138 times	100.0%	-%	100.0%
kai_get_lhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod.c:71)	called 516 times	100.0%	-%	100.0%
kai_get_lhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qai4c32p4x8_8x4_neon_i8mm.c:71)	called 516 times	100.0%	-%	100.0%
kai_get_lhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa.c:58)	called 70 times	100.0%	-%	100.0%
kai_get_lhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot.c:58)	called 30 times	100.0%	-%	100.0%
kai_get_lhs_packed_stride (kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32pscalef32_f16_neon.c:35)	called 4794768 times	100.0%	-%	100.0%
kai_get_lhs_packed_stride (kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32pscalef32_f32_neon.c:33)	called 4783864 times	100.0%	-%	100.0%
kai_get_m_step_dwconv_clamp_f32_f32_f32p1vlx1b_3x3_s1_4xc_sme2_mla (kai/ukernels/dwconv/dwconv_f32_f32_f32p/kai_dwconv_clamp_f32_f32_f32p1vlx1b_3x3_s1_4xc_sme2_mla.c:45)	called 1098 times	100.0%	-%	100.0%
kai_get_m_step_imatmul_clamp_f16_f16p2vlx2_f16p2vlx2_2vlx2vl_sme2_mopa (kai/ukernels/matmul/imatmul_clamp_f16_f16p_f16p/kai_imatmul_clamp_f16_f16p2vlx2_f16p2vlx2_2vlx2vl_sme2_mopa.c:44)	called 11352 times	100.0%	-%	100.0%
kai_get_m_step_imatmul_clamp_f16_f16p2vlx2_f16p2vlx2b_2vlx2vl_sme_mopa (kai/ukernels/matmul/imatmul_clamp_f16_f16p_f16p/kai_imatmul_clamp_f16_f16p2vlx2_f16p2vlx2b_2vlx2vl_sme_mopa.c:44)	called 11352 times	100.0%	-%	100.0%
kai_get_m_step_imatmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme2_mopa (kai/ukernels/matmul/imatmul_clamp_f32_f32p_f32p/kai_imatmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme2_mopa.c:43)	called 11352 times	100.0%	-%	100.0%
kai_get_m_step_imatmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme_mopa (kai/ukernels/matmul/imatmul_clamp_f32_f32p_f32p/kai_imatmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme_mopa.c:43)	called 11352 times	100.0%	-%	100.0%
kai_get_m_step_imatmul_clamp_qai8_qai8p2vlx4_qsi8cxp2vlx4sb_2vlx2vl_sme_mopa (kai/ukernels/matmul/imatmul_clamp_qai8_qai8p_qsi8cxp/kai_imatmul_clamp_qai8_qai8p2vlx4_qsi8cxp2vlx4sb_2vlx2vl_sme_mopa.c:44)	called 9324 times	100.0%	-%	100.0%
kai_get_m_step_imatmul_clamp_qai8_qai8p2vlx4_qsi8cxpsb2vlx4_2vlx2vl_sme2_mopa (kai/ukernels/matmul/imatmul_clamp_qai8_qai8p_qsi8cxp/kai_imatmul_clamp_qai8_qai8p2vlx4_qsi8cxpsb2vlx4_2vlx2vl_sme2_mopa.c:44)	called 9324 times	100.0%	-%	100.0%
kai_get_m_step_lhs_imatmul_pack_x16p2vlx2_x16p_sme (kai/ukernels/matmul/pack/kai_lhs_imatmul_pack_x16p2vlx2_x16p_sme.c:30)	called 22704 times	100.0%	-%	100.0%
kai_get_m_step_lhs_imatmul_pack_x32p2vlx1_x32p_sme (kai/ukernels/matmul/pack/kai_lhs_imatmul_pack_x32p2vlx1_x32p_sme.c:30)	called 22704 times	100.0%	-%	100.0%
kai_get_m_step_lhs_imatmul_pack_x8p2vlx4_x8p_sme (kai/ukernels/matmul/pack/kai_lhs_imatmul_pack_x8p2vlx4_x8p_sme.c:30)	called 18648 times	100.0%	-%	100.0%
kai_get_m_step_lhs_pack_bf16p2vlx2_f32_sme (kai/ukernels/matmul/pack/kai_lhs_pack_bf16p2vlx2_f32_sme.c:29)	not called	0.0%	-%	0.0%
kai_get_m_step_lhs_pack_bf16p8x4_f16_neon (kai/ukernels/matmul/pack/kai_lhs_pack_bf16p8x4_f16_neon.c:26)	not called	0.0%	-%	0.0%
kai_get_m_step_lhs_pack_f32p2vlx1_f32_sme (kai/ukernels/matmul/pack/kai_lhs_pack_f32p2vlx1_f32_sme.c:48)	not called	0.0%	-%	0.0%
kai_get_m_step_lhs_pack_x16p2vlx2_x16_sme (kai/ukernels/matmul/pack/kai_lhs_pack_x16p2vlx2_x16_sme.c:48)	not called	0.0%	-%	0.0%
kai_get_m_step_lhs_pack_x8p2vlx4_x8_sme (kai/ukernels/matmul/pack/kai_lhs_pack_x8p2vlx4_x8_sme.c:48)	called 798 times	100.0%	-%	57.0%
kai_get_m_step_lhs_quant_pack_bf16p1x4_f32_neon (kai/ukernels/matmul/pack/kai_lhs_quant_pack_bf16p1x4_f32_neon.c:23)	not called	0.0%	-%	0.0%
kai_get_m_step_lhs_quant_pack_bf16p8x4_f32_neon (kai/ukernels/matmul/pack/kai_lhs_quant_pack_bf16p8x4_f32_neon.c:25)	not called	0.0%	-%	0.0%
kai_get_m_step_lhs_quant_pack_qai8dxp_bf16_neon (kai/ukernels/matmul/pack/kai_lhs_quant_pack_qai8dxp_bf16_neon.c:38)	not called	0.0%	-%	0.0%
kai_get_m_step_lhs_quant_pack_qai8dxp_f16_neon (kai/ukernels/matmul/pack/kai_lhs_quant_pack_qai8dxp_f16_neon.c:45)	not called	0.0%	-%	0.0%
kai_get_m_step_lhs_quant_pack_qai8dxp_f32 (kai/ukernels/matmul/pack/kai_lhs_quant_pack_qai8dxp_f32.c:38)	not called	0.0%	-%	0.0%
kai_get_m_step_lhs_quant_pack_qsi8d32p4x8sb_f32_neon (kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p4x8sb_f32_neon.c:30)	not called	0.0%	-%	0.0%
kai_get_m_step_lhs_quant_pack_qsi8d32p_f32 (kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p_f32.c:30)	not called	0.0%	-%	0.0%
kai_get_m_step_lhs_quant_pack_qsi8d32p_f32_neon (kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p_f32_neon.c:36)	not called	0.0%	-%	0.0%
kai_get_m_step_lhs_quant_pack_qsi8d32pscalef32_f16_neon (kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32pscalef32_f16_neon.c:40)	not called	0.0%	-%	0.0%
kai_get_m_step_lhs_quant_pack_qsi8d32pscalef32_f32_neon (kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32pscalef32_f32_neon.c:39)	not called	0.0%	-%	0.0%
kai_get_m_step_matmul_clamp_bf16_qai8dxp1x8_qsi4c32p4x8_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_bf16_qai8dxp_qsi4c32p/kai_matmul_clamp_bf16_qai8dxp1x8_qsi4c32p4x8_1x4_neon_dotprod.c:98)	called 112 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_bf16_qai8dxp1x8_qsi4cxp8x8_1x8_neon_dotprod (kai/ukernels/matmul/matmul_clamp_bf16_qai8dxp_qsi4cxp/kai_matmul_clamp_bf16_qai8dxp1x8_qsi4cxp8x8_1x8_neon_dotprod.c:84)	called 392 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_bf16_qai8dxp4x8_qsi4c32p4x8_16x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_bf16_qai8dxp_qsi4c32p/kai_matmul_clamp_bf16_qai8dxp4x8_qsi4c32p4x8_16x4_neon_i8mm.c:99)	called 112 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_bf16_qai8dxp4x8_qsi4cxp8x8_8x8_neon_i8mm (kai/ukernels/matmul/matmul_clamp_bf16_qai8dxp_qsi4cxp/kai_matmul_clamp_bf16_qai8dxp4x8_qsi4cxp8x8_8x8_neon_i8mm.c:85)	called 392 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f16_bf16p8x4_bf16p12x4b_8x12_neon_mmla (kai/ukernels/matmul/matmul_clamp_f16_bf16p_bf16p/kai_matmul_clamp_f16_bf16p8x4_bf16p12x4b_8x12_neon_mmla.c:27)	called 90 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f16_f16_f16p16x1biasf16_6x16x8_neon_mla (kai/ukernels/matmul/matmul_clamp_f16_f16_f16p/kai_matmul_clamp_f16_f16_f16p16x1biasf16_6x16x8_neon_mla.c:28)	called 18 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f16_f16_f16p2vlx2b_1x16vl_sme2_dot (kai/ukernels/matmul/matmul_clamp_f16_f16_f16p/kai_matmul_clamp_f16_f16_f16p2vlx2b_1x16vl_sme2_dot.c:39)	called 52 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f16_f16_f16p2vlx2b_1x8vl_sme_mla (kai/ukernels/matmul/matmul_clamp_f16_f16_f16p/kai_matmul_clamp_f16_f16_f16p2vlx2b_1x8vl_sme_mla.c:39)	called 52 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f16_f16_f16p32x1b_6x32_neon_mla (kai/ukernels/matmul/matmul_clamp_f16_f16_f16p/kai_matmul_clamp_f16_f16_f16p32x1b_6x32_neon_mla.c:44)	called 50 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f16_f16_f16p32x1b_6x32_neon_mla_cortexa55 (kai/ukernels/matmul/matmul_clamp_f16_f16_f16p/kai_matmul_clamp_f16_f16_f16p32x1b_6x32_neon_mla_cortexa55.c:44)	called 50 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f16_f16p2vlx2_f16p2vlx2_2vlx2vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f16_f16p_f16p/kai_matmul_clamp_f16_f16p2vlx2_f16p2vlx2_2vlx2vl_sme2_mopa.c:45)	called 67 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f16_f16p2vlx2_f16p2vlx2b_2vlx2vl_sme_mopa (kai/ukernels/matmul/matmul_clamp_f16_f16p_f16p/kai_matmul_clamp_f16_f16p2vlx2_f16p2vlx2b_2vlx2vl_sme_mopa.c:45)	called 67 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f16_qai8dxp1x4_qsi4cxp4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi4cxp/kai_matmul_clamp_f16_qai8dxp1x4_qsi4cxp4x4_1x4_neon_dotprod.c:84)	called 84 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f16_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi8cxp/kai_matmul_clamp_f16_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod.c:81)	called 84 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f16_qai8dxp1x8_qsi4cxp4x8_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi4cxp/kai_matmul_clamp_f16_qai8dxp1x8_qsi4cxp4x8_1x4_neon_dotprod.c:84)	called 84 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f16_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi8cxp/kai_matmul_clamp_f16_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod.c:81)	called 84 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f16_qai8dxp4x4_qsi4cxp4x4_16x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi4cxp/kai_matmul_clamp_f16_qai8dxp4x4_qsi4cxp4x4_16x4_neon_dotprod.c:84)	called 224 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f16_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi8cxp/kai_matmul_clamp_f16_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod.c:81)	called 224 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f16_qai8dxp4x8_qsi4cxp4x8_16x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi4cxp/kai_matmul_clamp_f16_qai8dxp4x8_qsi4cxp4x8_16x4_neon_i8mm.c:84)	called 224 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f16_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi8cxp/kai_matmul_clamp_f16_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm.c:81)	called 224 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f16_qsi8d32p1vlx4_qai4c32p4vlx4_1vlx4vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p1vlx4_qai4c32p4vlx4_1vlx4vl_sme2_mopa.c:98)	called 4136 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f16_qsi8d32p1x4_qai4c32p4vlx4_1x4vl_sme2_dot (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p1x4_qai4c32p4vlx4_1x4vl_sme2_dot.c:94)	called 1980 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f16_qsi8d32p1x4_qai4c32p4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p1x4_qai4c32p4x4_1x4_neon_dotprod.c:90)	called 714 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f16_qsi8d32p1x8_qai4c32p4x8_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p1x8_qai4c32p4x8_1x4_neon_dotprod.c:90)	called 714 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f16_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod.c:90)	called 1036 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f16_qsi8d32p4x8_qai4c32p4x8_8x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p4x8_qai4c32p4x8_8x4_neon_i8mm.c:90)	called 1036 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f32_bf16p1x4_bf16p12x4b_1x36_neon_dot (kai/ukernels/matmul/matmul_clamp_f32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p1x4_bf16p12x4b_1x36_neon_dot.c:29)	called 188 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa.c:44)	called 142 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f32_bf16p8x4_bf16p12x4b_8x12_neon_mmla (kai/ukernels/matmul/matmul_clamp_f32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p8x4_bf16p12x4b_8x12_neon_mmla.c:28)	called 200 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f32_f32_f32p16vlx1b_1x16vl_sme2_mla (kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p16vlx1b_1x16vl_sme2_mla.c:37)	called 52 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f32_f32_f32p16x1b_6x16_neon_mla (kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p16x1b_6x16_neon_mla.c:40)	called 50 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f32_f32_f32p16x1b_6x16_neon_mla_cortexa55 (kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p16x1b_6x16_neon_mla_cortexa55.c:40)	called 50 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla (kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla.c:37)	called 52 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f32_f32_f32p2vlx1b_1x8vl_sme_mla (kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p2vlx1b_1x8vl_sme_mla.c:37)	called 52 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f32_f32_f32p8x1biasf32_6x8x4_neon_mla (kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p8x1biasf32_6x8x4_neon_mla.c:41)	called 18 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme_mopa (kai/ukernels/matmul/matmul_clamp_f32_f32p_f32p/kai_matmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme_mopa.c:44)	called 67 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f32_f32p_f32p/kai_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa.c:44)	called 67 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa.c:90)	called 616 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme_mopa (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme_mopa.c:90)	called 616 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f32_qai8dxp1vlx8_qsi4cxp4vlx8_1vlx4vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1vlx8_qsi4cxp4vlx8_1vlx4vl_sme2_mopa.c:60)	called 360 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f32_qai8dxp1x4_qsi4c32p4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x4_qsi4c32p4x4_1x4_neon_dotprod.c:97)	called 112 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f32_qai8dxp1x4_qsi4c32p8x4_1x8_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x4_qsi4c32p8x4_1x8_neon_dotprod.c:97)	called 112 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f32_qai8dxp1x4_qsi4cxp4vlx4_1x4vl_sme2_sdot (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi4cxp4vlx4_1x4vl_sme2_sdot.c:57)	called 160 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f32_qai8dxp1x4_qsi4cxp4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi4cxp4x4_1x4_neon_dotprod.c:69)	called 320 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot.c:93)	called 576 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme_dot (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme_dot.c:93)	called 576 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod.c:67)	called 231 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f32_qai8dxp1x8_qsi4c32p4x8_1x4x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x8_qsi4c32p4x8_1x4x32_neon_dotprod.c:97)	called 112 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f32_qai8dxp1x8_qsi4c32p8x8_1x8_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x8_qsi4c32p8x8_1x8_neon_dotprod.c:97)	called 112 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f32_qai8dxp1x8_qsi4c32p8x8_1x8x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x8_qsi4c32p8x8_1x8x32_neon_dotprod.c:97)	called 112 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f32_qai8dxp1x8_qsi4cxp4x8_1x4x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi4cxp4x8_1x4x32_neon_dotprod.c:53)	called 320 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f32_qai8dxp1x8_qsi4cxp8x8_1x8x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi4cxp8x8_1x8x32_neon_dotprod.c:53)	called 320 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f32_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod.c:67)	called 231 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f32_qai8dxp4x4_qsi4c32p4x4_16x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x4_qsi4c32p4x4_16x4_neon_dotprod.c:98)	called 112 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod.c:97)	called 112 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f32_qai8dxp4x4_qsi4cxp8x4_8x8x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x4_qsi4cxp8x4_8x8x32_neon_dotprod.c:55)	called 320 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f32_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod.c:71)	called 231 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f32_qai8dxp4x8_qsi4c32p4x8_16x4x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p4x8_16x4x32_neon_i8mm.c:98)	called 112 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f32_qai8dxp4x8_qsi4c32p4x8_8x4x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p4x8_8x4x32_neon_i8mm.c:97)	called 112 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f32_qai8dxp4x8_qsi4c32p8x8_4x8_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p8x8_4x8_neon_i8mm.c:97)	called 112 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f32_qai8dxp4x8_qsi4c32p8x8_4x8x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p8x8_4x8x32_neon_i8mm.c:97)	called 112 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x4_16x4x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x4_16x4x32_neon_dotprod.c:55)	called 320 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_4x4x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_4x4x32_neon_i8mm.c:52)	called 320 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_8x4x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_8x4x32_neon_i8mm.c:56)	called 320 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_4x8x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_4x8x32_neon_i8mm.c:56)	called 320 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_8x8x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_8x8x32_neon_i8mm.c:56)	called 320 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f32_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm.c:71)	called 231 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f32_qsi8d32p1vlx4_qai4c32p4vlx4_1vlx4vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p1vlx4_qai4c32p4vlx4_1vlx4vl_sme2_mopa.c:98)	called 4136 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa.c:76)	called 141 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f32_qsi8d32p1x4_qai4c32p4vlx4_1x4vl_sme2_dot (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qai4c32p4vlx4_1x4vl_sme2_dot.c:94)	called 1980 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f32_qsi8d32p1x4_qai4c32p4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qai4c32p4x4_1x4_neon_dotprod.c:88)	called 714 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot.c:76)	called 81 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod.c:70)	called 52 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f32_qsi8d32p1x8_qai4c32p4x8_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p1x8_qai4c32p4x8_1x4_neon_dotprod.c:88)	called 714 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod.c:55)	called 52 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f32_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod.c:88)	called 1036 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod.c:74)	called 72 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f32_qsi8d32p4x8_qai4c32p4x8_8x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qai4c32p4x8_8x4_neon_i8mm.c:88)	called 1036 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm.c:58)	called 144 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_8x4x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_8x4x32_neon_i8mm.c:58)	called 72 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_qai8_qai8_qsi8cxp2vlx4sb_1x16vl_sme2_dot (kai/ukernels/matmul/matmul_clamp_qai8_qai8_qsi8cxp/kai_matmul_clamp_qai8_qai8_qsi8cxp2vlx4sb_1x16vl_sme2_dot.c:38)	called 168 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_qai8_qai8p2vlx4_qsi8cxp2vlx4sb_2vlx2vl_sme_mopa (kai/ukernels/matmul/matmul_clamp_qai8_qai8p_qsi8cxp/kai_matmul_clamp_qai8_qai8p2vlx4_qsi8cxp2vlx4sb_2vlx2vl_sme_mopa.c:45)	called 999 times	100.0%	-%	100.0%
kai_get_m_step_matmul_clamp_qai8_qai8p2vlx4_qsi8cxpsb2vlx4_2vlx2vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_qai8_qai8p_qsi8cxp/kai_matmul_clamp_qai8_qai8p2vlx4_qsi8cxpsb2vlx4_2vlx2vl_sme2_mopa.c:45)	called 999 times	100.0%	-%	100.0%
kai_get_mr_lhs_imatmul_pack_x16p2vlx2_x16p_sme (kai/ukernels/matmul/pack/kai_lhs_imatmul_pack_x16p2vlx2_x16p_sme.c:26)	called 45408 times	100.0%	-%	100.0%
kai_get_mr_lhs_imatmul_pack_x32p2vlx1_x32p_sme (kai/ukernels/matmul/pack/kai_lhs_imatmul_pack_x32p2vlx1_x32p_sme.c:26)	called 45408 times	100.0%	-%	100.0%
kai_get_mr_lhs_imatmul_pack_x8p2vlx4_x8p_sme (kai/ukernels/matmul/pack/kai_lhs_imatmul_pack_x8p2vlx4_x8p_sme.c:26)	called 37296 times	100.0%	-%	100.0%
kai_get_mr_lhs_pack_bf16p2vlx2_f32_sme (kai/ukernels/matmul/pack/kai_lhs_pack_bf16p2vlx2_f32_sme.c:25)	called 184 times	100.0%	-%	100.0%
kai_get_mr_lhs_pack_f32p2vlx1_f32_sme (kai/ukernels/matmul/pack/kai_lhs_pack_f32p2vlx1_f32_sme.c:44)	called 170 times	100.0%	-%	100.0%
kai_get_mr_lhs_pack_x16p2vlx2_x16_sme (kai/ukernels/matmul/pack/kai_lhs_pack_x16p2vlx2_x16_sme.c:44)	called 262 times	100.0%	-%	100.0%
kai_get_mr_lhs_pack_x8p2vlx4_x8_sme (kai/ukernels/matmul/pack/kai_lhs_pack_x8p2vlx4_x8_sme.c:44)	called 5724 times	100.0%	-%	100.0%
kai_get_mr_matmul_clamp_bf16_qai8dxp1x8_qsi4c32p4x8_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_bf16_qai8dxp_qsi4c32p/kai_matmul_clamp_bf16_qai8dxp1x8_qsi4c32p4x8_1x4_neon_dotprod.c:106)	called 112 times	100.0%	-%	100.0%
kai_get_mr_matmul_clamp_bf16_qai8dxp1x8_qsi4cxp8x8_1x8_neon_dotprod (kai/ukernels/matmul/matmul_clamp_bf16_qai8dxp_qsi4cxp/kai_matmul_clamp_bf16_qai8dxp1x8_qsi4cxp8x8_1x8_neon_dotprod.c:92)	called 392 times	100.0%	-%	100.0%
kai_get_mr_matmul_clamp_bf16_qai8dxp4x8_qsi4c32p4x8_16x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_bf16_qai8dxp_qsi4c32p/kai_matmul_clamp_bf16_qai8dxp4x8_qsi4c32p4x8_16x4_neon_i8mm.c:107)	called 112 times	100.0%	-%	100.0%
kai_get_mr_matmul_clamp_bf16_qai8dxp4x8_qsi4cxp8x8_8x8_neon_i8mm (kai/ukernels/matmul/matmul_clamp_bf16_qai8dxp_qsi4cxp/kai_matmul_clamp_bf16_qai8dxp4x8_qsi4cxp8x8_8x8_neon_i8mm.c:93)	called 392 times	100.0%	-%	100.0%
kai_get_mr_matmul_clamp_f16_bf16p8x4_bf16p12x4b_8x12_neon_mmla (kai/ukernels/matmul/matmul_clamp_f16_bf16p_bf16p/kai_matmul_clamp_f16_bf16p8x4_bf16p12x4b_8x12_neon_mmla.c:35)	not called	0.0%	-%	0.0%
kai_get_mr_matmul_clamp_f16_f16p2vlx2_f16p2vlx2_2vlx2vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f16_f16p_f16p/kai_matmul_clamp_f16_f16p2vlx2_f16p2vlx2_2vlx2vl_sme2_mopa.c:53)	called 17 times	100.0%	-%	100.0%
kai_get_mr_matmul_clamp_f16_f16p2vlx2_f16p2vlx2b_2vlx2vl_sme_mopa (kai/ukernels/matmul/matmul_clamp_f16_f16p_f16p/kai_matmul_clamp_f16_f16p2vlx2_f16p2vlx2b_2vlx2vl_sme_mopa.c:53)	called 17 times	100.0%	-%	100.0%
kai_get_mr_matmul_clamp_f16_qai8dxp1x4_qsi4cxp4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi4cxp/kai_matmul_clamp_f16_qai8dxp1x4_qsi4cxp4x4_1x4_neon_dotprod.c:92)	called 224 times	100.0%	-%	100.0%
kai_get_mr_matmul_clamp_f16_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi8cxp/kai_matmul_clamp_f16_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod.c:89)	called 224 times	100.0%	-%	100.0%
kai_get_mr_matmul_clamp_f16_qai8dxp1x8_qsi4cxp4x8_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi4cxp/kai_matmul_clamp_f16_qai8dxp1x8_qsi4cxp4x8_1x4_neon_dotprod.c:92)	called 224 times	100.0%	-%	100.0%
kai_get_mr_matmul_clamp_f16_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi8cxp/kai_matmul_clamp_f16_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod.c:89)	called 224 times	100.0%	-%	100.0%
kai_get_mr_matmul_clamp_f16_qai8dxp4x4_qsi4cxp4x4_16x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi4cxp/kai_matmul_clamp_f16_qai8dxp4x4_qsi4cxp4x4_16x4_neon_dotprod.c:92)	called 224 times	100.0%	-%	100.0%
kai_get_mr_matmul_clamp_f16_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi8cxp/kai_matmul_clamp_f16_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod.c:89)	called 224 times	100.0%	-%	100.0%
kai_get_mr_matmul_clamp_f16_qai8dxp4x8_qsi4cxp4x8_16x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi4cxp/kai_matmul_clamp_f16_qai8dxp4x8_qsi4cxp4x8_16x4_neon_i8mm.c:92)	called 224 times	100.0%	-%	100.0%
kai_get_mr_matmul_clamp_f16_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi8cxp/kai_matmul_clamp_f16_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm.c:89)	called 224 times	100.0%	-%	100.0%
kai_get_mr_matmul_clamp_f16_qsi8d32p1vlx4_qai4c32p4vlx4_1vlx4vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p1vlx4_qai4c32p4vlx4_1vlx4vl_sme2_mopa.c:106)	called 5169 times	100.0%	-%	100.0%
kai_get_mr_matmul_clamp_f16_qsi8d32p1x4_qai4c32p4vlx4_1x4vl_sme2_dot (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p1x4_qai4c32p4vlx4_1x4vl_sme2_dot.c:102)	called 2624 times	100.0%	-%	100.0%
kai_get_mr_matmul_clamp_f16_qsi8d32p1x4_qai4c32p4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p1x4_qai4c32p4x4_1x4_neon_dotprod.c:98)	called 1036 times	100.0%	-%	100.0%
kai_get_mr_matmul_clamp_f16_qsi8d32p1x8_qai4c32p4x8_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p1x8_qai4c32p4x8_1x4_neon_dotprod.c:98)	called 1036 times	100.0%	-%	100.0%
kai_get_mr_matmul_clamp_f16_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod.c:98)	called 1036 times	100.0%	-%	100.0%
kai_get_mr_matmul_clamp_f16_qsi8d32p4x8_qai4c32p4x8_8x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p4x8_qai4c32p4x8_8x4_neon_i8mm.c:98)	called 1036 times	100.0%	-%	100.0%
kai_get_mr_matmul_clamp_f32_bf16p1x4_bf16p12x4b_1x36_neon_dot (kai/ukernels/matmul/matmul_clamp_f32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p1x4_bf16p12x4b_1x36_neon_dot.c:37)	not called	0.0%	-%	0.0%
kai_get_mr_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa.c:52)	not called	0.0%	-%	0.0%
kai_get_mr_matmul_clamp_f32_bf16p8x4_bf16p12x4b_8x12_neon_mmla (kai/ukernels/matmul/matmul_clamp_f32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p8x4_bf16p12x4b_8x12_neon_mmla.c:36)	not called	0.0%	-%	0.0%
kai_get_mr_matmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme_mopa (kai/ukernels/matmul/matmul_clamp_f32_f32p_f32p/kai_matmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme_mopa.c:52)	called 17 times	100.0%	-%	100.0%
kai_get_mr_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f32_f32p_f32p/kai_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa.c:52)	called 17 times	100.0%	-%	100.0%
kai_get_mr_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa.c:98)	called 1003 times	100.0%	-%	100.0%
kai_get_mr_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme_mopa (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme_mopa.c:98)	called 1003 times	100.0%	-%	100.0%
kai_get_mr_matmul_clamp_f32_qai8dxp1vlx8_qsi4cxp4vlx8_1vlx4vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1vlx8_qsi4cxp4vlx8_1vlx4vl_sme2_mopa.c:68)	called 522 times	100.0%	-%	100.0%
kai_get_mr_matmul_clamp_f32_qai8dxp1x4_qsi4c32p4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x4_qsi4c32p4x4_1x4_neon_dotprod.c:105)	called 112 times	100.0%	-%	100.0%
kai_get_mr_matmul_clamp_f32_qai8dxp1x4_qsi4c32p8x4_1x8_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x4_qsi4c32p8x4_1x8_neon_dotprod.c:105)	called 112 times	100.0%	-%	100.0%
kai_get_mr_matmul_clamp_f32_qai8dxp1x4_qsi4cxp4vlx4_1x4vl_sme2_sdot (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi4cxp4vlx4_1x4vl_sme2_sdot.c:69)	called 321 times	100.0%	-%	100.0%
kai_get_mr_matmul_clamp_f32_qai8dxp1x4_qsi4cxp4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi4cxp4x4_1x4_neon_dotprod.c:77)	called 240 times	100.0%	-%	100.0%
kai_get_mr_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot.c:101)	called 577 times	100.0%	-%	100.0%
kai_get_mr_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme_dot (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme_dot.c:101)	called 577 times	100.0%	-%	100.0%
kai_get_mr_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod.c:75)	called 231 times	100.0%	-%	100.0%
kai_get_mr_matmul_clamp_f32_qai8dxp1x8_qsi4c32p4x8_1x4x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x8_qsi4c32p4x8_1x4x32_neon_dotprod.c:105)	called 112 times	100.0%	-%	100.0%
kai_get_mr_matmul_clamp_f32_qai8dxp1x8_qsi4c32p8x8_1x8_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x8_qsi4c32p8x8_1x8_neon_dotprod.c:105)	called 112 times	100.0%	-%	100.0%
kai_get_mr_matmul_clamp_f32_qai8dxp1x8_qsi4c32p8x8_1x8x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x8_qsi4c32p8x8_1x8x32_neon_dotprod.c:105)	called 112 times	100.0%	-%	100.0%
kai_get_mr_matmul_clamp_f32_qai8dxp1x8_qsi4cxp4x8_1x4x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi4cxp4x8_1x4x32_neon_dotprod.c:61)	called 240 times	100.0%	-%	100.0%
kai_get_mr_matmul_clamp_f32_qai8dxp1x8_qsi4cxp8x8_1x8x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi4cxp8x8_1x8x32_neon_dotprod.c:61)	called 240 times	100.0%	-%	100.0%
kai_get_mr_matmul_clamp_f32_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod.c:75)	called 231 times	100.0%	-%	100.0%
kai_get_mr_matmul_clamp_f32_qai8dxp4x4_qsi4c32p4x4_16x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x4_qsi4c32p4x4_16x4_neon_dotprod.c:106)	called 112 times	100.0%	-%	100.0%
kai_get_mr_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod.c:105)	called 112 times	100.0%	-%	100.0%
kai_get_mr_matmul_clamp_f32_qai8dxp4x4_qsi4cxp8x4_8x8x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x4_qsi4cxp8x4_8x8x32_neon_dotprod.c:63)	called 240 times	100.0%	-%	100.0%
kai_get_mr_matmul_clamp_f32_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod.c:79)	called 231 times	100.0%	-%	100.0%
kai_get_mr_matmul_clamp_f32_qai8dxp4x8_qsi4c32p4x8_16x4x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p4x8_16x4x32_neon_i8mm.c:106)	called 112 times	100.0%	-%	100.0%
kai_get_mr_matmul_clamp_f32_qai8dxp4x8_qsi4c32p4x8_8x4x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p4x8_8x4x32_neon_i8mm.c:105)	called 112 times	100.0%	-%	100.0%
kai_get_mr_matmul_clamp_f32_qai8dxp4x8_qsi4c32p8x8_4x8_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p8x8_4x8_neon_i8mm.c:105)	called 112 times	100.0%	-%	100.0%
kai_get_mr_matmul_clamp_f32_qai8dxp4x8_qsi4c32p8x8_4x8x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p8x8_4x8x32_neon_i8mm.c:105)	called 112 times	100.0%	-%	100.0%
kai_get_mr_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x4_16x4x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x4_16x4x32_neon_dotprod.c:63)	called 240 times	100.0%	-%	100.0%
kai_get_mr_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_4x4x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_4x4x32_neon_i8mm.c:60)	called 240 times	100.0%	-%	100.0%
kai_get_mr_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_8x4x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_8x4x32_neon_i8mm.c:64)	called 240 times	100.0%	-%	100.0%
kai_get_mr_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_4x8x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_4x8x32_neon_i8mm.c:64)	called 240 times	100.0%	-%	100.0%
kai_get_mr_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_8x8x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_8x8x32_neon_i8mm.c:64)	called 240 times	100.0%	-%	100.0%
kai_get_mr_matmul_clamp_f32_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm.c:79)	called 231 times	100.0%	-%	100.0%
kai_get_mr_matmul_clamp_f32_qsi8d32p1vlx4_qai4c32p4vlx4_1vlx4vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p1vlx4_qai4c32p4vlx4_1vlx4vl_sme2_mopa.c:106)	called 5169 times	100.0%	-%	100.0%
kai_get_mr_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa.c:84)	called 188 times	100.0%	-%	100.0%
kai_get_mr_matmul_clamp_f32_qsi8d32p1x4_qai4c32p4vlx4_1x4vl_sme2_dot (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qai4c32p4vlx4_1x4vl_sme2_dot.c:102)	called 2624 times	100.0%	-%	100.0%
kai_get_mr_matmul_clamp_f32_qsi8d32p1x4_qai4c32p4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qai4c32p4x4_1x4_neon_dotprod.c:96)	called 1036 times	100.0%	-%	100.0%
kai_get_mr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot.c:84)	called 108 times	100.0%	-%	100.0%
kai_get_mr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod.c:78)	called 48 times	100.0%	-%	100.0%
kai_get_mr_matmul_clamp_f32_qsi8d32p1x8_qai4c32p4x8_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p1x8_qai4c32p4x8_1x4_neon_dotprod.c:96)	called 1036 times	100.0%	-%	100.0%
kai_get_mr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod.c:63)	called 48 times	100.0%	-%	100.0%
kai_get_mr_matmul_clamp_f32_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod.c:96)	called 1036 times	100.0%	-%	100.0%
kai_get_mr_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod.c:82)	called 48 times	100.0%	-%	100.0%
kai_get_mr_matmul_clamp_f32_qsi8d32p4x8_qai4c32p4x8_8x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qai4c32p4x8_8x4_neon_i8mm.c:96)	called 1036 times	100.0%	-%	100.0%
kai_get_mr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm.c:66)	called 96 times	100.0%	-%	100.0%
kai_get_mr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_8x4x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_8x4x32_neon_i8mm.c:66)	called 48 times	100.0%	-%	100.0%
kai_get_mr_matmul_clamp_qai8_qai8p2vlx4_qsi8cxp2vlx4sb_2vlx2vl_sme_mopa (kai/ukernels/matmul/matmul_clamp_qai8_qai8p_qsi8cxp/kai_matmul_clamp_qai8_qai8p2vlx4_qsi8cxp2vlx4sb_2vlx2vl_sme_mopa.c:53)	called 333 times	100.0%	-%	100.0%
kai_get_mr_matmul_clamp_qai8_qai8p2vlx4_qsi8cxpsb2vlx4_2vlx2vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_qai8_qai8p_qsi8cxp/kai_matmul_clamp_qai8_qai8p2vlx4_qsi8cxpsb2vlx4_2vlx2vl_sme2_mopa.c:53)	called 333 times	100.0%	-%	100.0%
kai_get_n_step_imatmul_clamp_f16_f16p2vlx2_f16p2vlx2_2vlx2vl_sme2_mopa (kai/ukernels/matmul/imatmul_clamp_f16_f16p_f16p/kai_imatmul_clamp_f16_f16p2vlx2_f16p2vlx2_2vlx2vl_sme2_mopa.c:48)	called 22704 times	100.0%	-%	100.0%
kai_get_n_step_imatmul_clamp_f16_f16p2vlx2_f16p2vlx2b_2vlx2vl_sme_mopa (kai/ukernels/matmul/imatmul_clamp_f16_f16p_f16p/kai_imatmul_clamp_f16_f16p2vlx2_f16p2vlx2b_2vlx2vl_sme_mopa.c:48)	called 22704 times	100.0%	-%	100.0%
kai_get_n_step_imatmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme2_mopa (kai/ukernels/matmul/imatmul_clamp_f32_f32p_f32p/kai_imatmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme2_mopa.c:47)	called 22704 times	100.0%	-%	100.0%
kai_get_n_step_imatmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme_mopa (kai/ukernels/matmul/imatmul_clamp_f32_f32p_f32p/kai_imatmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme_mopa.c:47)	called 22704 times	100.0%	-%	100.0%
kai_get_n_step_imatmul_clamp_qai8_qai8p2vlx4_qsi8cxp2vlx4sb_2vlx2vl_sme_mopa (kai/ukernels/matmul/imatmul_clamp_qai8_qai8p_qsi8cxp/kai_imatmul_clamp_qai8_qai8p2vlx4_qsi8cxp2vlx4sb_2vlx2vl_sme_mopa.c:48)	called 18648 times	100.0%	-%	100.0%
kai_get_n_step_imatmul_clamp_qai8_qai8p2vlx4_qsi8cxpsb2vlx4_2vlx2vl_sme2_mopa (kai/ukernels/matmul/imatmul_clamp_qai8_qai8p_qsi8cxp/kai_imatmul_clamp_qai8_qai8p2vlx4_qsi8cxpsb2vlx4_2vlx2vl_sme2_mopa.c:48)	called 18648 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_bf16_qai8dxp1x8_qsi4c32p4x8_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_bf16_qai8dxp_qsi4c32p/kai_matmul_clamp_bf16_qai8dxp1x8_qsi4c32p4x8_1x4_neon_dotprod.c:102)	called 112 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_bf16_qai8dxp1x8_qsi4cxp8x8_1x8_neon_dotprod (kai/ukernels/matmul/matmul_clamp_bf16_qai8dxp_qsi4cxp/kai_matmul_clamp_bf16_qai8dxp1x8_qsi4cxp8x8_1x8_neon_dotprod.c:88)	called 392 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_bf16_qai8dxp4x8_qsi4c32p4x8_16x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_bf16_qai8dxp_qsi4c32p/kai_matmul_clamp_bf16_qai8dxp4x8_qsi4c32p4x8_16x4_neon_i8mm.c:103)	called 112 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_bf16_qai8dxp4x8_qsi4cxp8x8_8x8_neon_i8mm (kai/ukernels/matmul/matmul_clamp_bf16_qai8dxp_qsi4cxp/kai_matmul_clamp_bf16_qai8dxp4x8_qsi4cxp8x8_8x8_neon_i8mm.c:89)	called 392 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f16_bf16p8x4_bf16p12x4b_8x12_neon_mmla (kai/ukernels/matmul/matmul_clamp_f16_bf16p_bf16p/kai_matmul_clamp_f16_bf16p8x4_bf16p12x4b_8x12_neon_mmla.c:31)	called 90 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f16_f16_f16p16x1biasf16_6x16x8_neon_mla (kai/ukernels/matmul/matmul_clamp_f16_f16_f16p/kai_matmul_clamp_f16_f16_f16p16x1biasf16_6x16x8_neon_mla.c:32)	called 18 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f16_f16_f16p2vlx2b_1x16vl_sme2_dot (kai/ukernels/matmul/matmul_clamp_f16_f16_f16p/kai_matmul_clamp_f16_f16_f16p2vlx2b_1x16vl_sme2_dot.c:43)	called 236 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f16_f16_f16p2vlx2b_1x8vl_sme_mla (kai/ukernels/matmul/matmul_clamp_f16_f16_f16p/kai_matmul_clamp_f16_f16_f16p2vlx2b_1x8vl_sme_mla.c:43)	called 236 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f16_f16_f16p32x1b_6x32_neon_mla (kai/ukernels/matmul/matmul_clamp_f16_f16_f16p/kai_matmul_clamp_f16_f16_f16p32x1b_6x32_neon_mla.c:48)	called 50 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f16_f16_f16p32x1b_6x32_neon_mla_cortexa55 (kai/ukernels/matmul/matmul_clamp_f16_f16_f16p/kai_matmul_clamp_f16_f16_f16p32x1b_6x32_neon_mla_cortexa55.c:48)	called 50 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f16_f16p2vlx2_f16p2vlx2_2vlx2vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f16_f16p_f16p/kai_matmul_clamp_f16_f16p2vlx2_f16p2vlx2_2vlx2vl_sme2_mopa.c:49)	called 82 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f16_f16p2vlx2_f16p2vlx2b_2vlx2vl_sme_mopa (kai/ukernels/matmul/matmul_clamp_f16_f16p_f16p/kai_matmul_clamp_f16_f16p2vlx2_f16p2vlx2b_2vlx2vl_sme_mopa.c:49)	called 82 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f16_qai8dxp1x4_qsi4cxp4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi4cxp/kai_matmul_clamp_f16_qai8dxp1x4_qsi4cxp4x4_1x4_neon_dotprod.c:88)	called 84 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f16_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi8cxp/kai_matmul_clamp_f16_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod.c:85)	called 84 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f16_qai8dxp1x8_qsi4cxp4x8_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi4cxp/kai_matmul_clamp_f16_qai8dxp1x8_qsi4cxp4x8_1x4_neon_dotprod.c:88)	called 84 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f16_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi8cxp/kai_matmul_clamp_f16_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod.c:85)	called 84 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f16_qai8dxp4x4_qsi4cxp4x4_16x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi4cxp/kai_matmul_clamp_f16_qai8dxp4x4_qsi4cxp4x4_16x4_neon_dotprod.c:88)	called 224 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f16_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi8cxp/kai_matmul_clamp_f16_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod.c:85)	called 224 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f16_qai8dxp4x8_qsi4cxp4x8_16x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi4cxp/kai_matmul_clamp_f16_qai8dxp4x8_qsi4cxp4x8_16x4_neon_i8mm.c:88)	called 224 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f16_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi8cxp/kai_matmul_clamp_f16_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm.c:85)	called 224 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f16_qsi8d32p1vlx4_qai4c32p4vlx4_1vlx4vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p1vlx4_qai4c32p4vlx4_1vlx4vl_sme2_mopa.c:102)	called 4136 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f16_qsi8d32p1x4_qai4c32p4vlx4_1x4vl_sme2_dot (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p1x4_qai4c32p4vlx4_1x4vl_sme2_dot.c:98)	called 1980 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f16_qsi8d32p1x4_qai4c32p4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p1x4_qai4c32p4x4_1x4_neon_dotprod.c:94)	called 714 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f16_qsi8d32p1x8_qai4c32p4x8_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p1x8_qai4c32p4x8_1x4_neon_dotprod.c:94)	called 714 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f16_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod.c:94)	called 1036 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f16_qsi8d32p4x8_qai4c32p4x8_8x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p4x8_qai4c32p4x8_8x4_neon_i8mm.c:94)	called 1036 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f32_bf16p1x4_bf16p12x4b_1x36_neon_dot (kai/ukernels/matmul/matmul_clamp_f32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p1x4_bf16p12x4b_1x36_neon_dot.c:33)	called 280 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa.c:48)	called 234 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f32_bf16p8x4_bf16p12x4b_8x12_neon_mmla (kai/ukernels/matmul/matmul_clamp_f32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p8x4_bf16p12x4b_8x12_neon_mmla.c:32)	called 200 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f32_f32_f32p16vlx1b_1x16vl_sme2_mla (kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p16vlx1b_1x16vl_sme2_mla.c:41)	called 236 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f32_f32_f32p16x1b_6x16_neon_mla (kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p16x1b_6x16_neon_mla.c:44)	called 50 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f32_f32_f32p16x1b_6x16_neon_mla_cortexa55 (kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p16x1b_6x16_neon_mla_cortexa55.c:44)	called 50 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla (kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla.c:41)	called 236 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f32_f32_f32p2vlx1b_1x8vl_sme_mla (kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p2vlx1b_1x8vl_sme_mla.c:41)	called 236 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f32_f32_f32p8x1biasf32_6x8x4_neon_mla (kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p8x1biasf32_6x8x4_neon_mla.c:45)	called 18 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme_mopa (kai/ukernels/matmul/matmul_clamp_f32_f32p_f32p/kai_matmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme_mopa.c:48)	called 82 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f32_f32p_f32p/kai_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa.c:48)	called 82 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa.c:94)	called 616 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme_mopa (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme_mopa.c:94)	called 616 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f32_qai8dxp1vlx8_qsi4cxp4vlx8_1vlx4vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1vlx8_qsi4cxp4vlx8_1vlx4vl_sme2_mopa.c:64)	called 360 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f32_qai8dxp1x4_qsi4c32p4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x4_qsi4c32p4x4_1x4_neon_dotprod.c:101)	called 112 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f32_qai8dxp1x4_qsi4c32p8x4_1x8_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x4_qsi4c32p8x4_1x8_neon_dotprod.c:101)	called 112 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f32_qai8dxp1x4_qsi4cxp4vlx4_1x4vl_sme2_sdot (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi4cxp4vlx4_1x4vl_sme2_sdot.c:65)	called 160 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f32_qai8dxp1x4_qsi4cxp4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi4cxp4x4_1x4_neon_dotprod.c:73)	called 320 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot.c:97)	called 576 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme_dot (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme_dot.c:97)	called 576 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod.c:71)	called 231 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f32_qai8dxp1x8_qsi4c32p4x8_1x4x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x8_qsi4c32p4x8_1x4x32_neon_dotprod.c:101)	called 112 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f32_qai8dxp1x8_qsi4c32p8x8_1x8_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x8_qsi4c32p8x8_1x8_neon_dotprod.c:101)	called 112 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f32_qai8dxp1x8_qsi4c32p8x8_1x8x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x8_qsi4c32p8x8_1x8x32_neon_dotprod.c:101)	called 112 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f32_qai8dxp1x8_qsi4cxp4x8_1x4x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi4cxp4x8_1x4x32_neon_dotprod.c:57)	called 320 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f32_qai8dxp1x8_qsi4cxp8x8_1x8x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi4cxp8x8_1x8x32_neon_dotprod.c:57)	called 320 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f32_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod.c:71)	called 231 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f32_qai8dxp4x4_qsi4c32p4x4_16x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x4_qsi4c32p4x4_16x4_neon_dotprod.c:102)	called 112 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod.c:101)	called 112 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f32_qai8dxp4x4_qsi4cxp8x4_8x8x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x4_qsi4cxp8x4_8x8x32_neon_dotprod.c:59)	called 320 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f32_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod.c:75)	called 231 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f32_qai8dxp4x8_qsi4c32p4x8_16x4x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p4x8_16x4x32_neon_i8mm.c:102)	called 112 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f32_qai8dxp4x8_qsi4c32p4x8_8x4x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p4x8_8x4x32_neon_i8mm.c:101)	called 112 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f32_qai8dxp4x8_qsi4c32p8x8_4x8_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p8x8_4x8_neon_i8mm.c:101)	called 112 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f32_qai8dxp4x8_qsi4c32p8x8_4x8x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p8x8_4x8x32_neon_i8mm.c:101)	called 112 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x4_16x4x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x4_16x4x32_neon_dotprod.c:59)	called 320 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_4x4x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_4x4x32_neon_i8mm.c:56)	called 320 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_8x4x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_8x4x32_neon_i8mm.c:60)	called 320 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_4x8x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_4x8x32_neon_i8mm.c:60)	called 320 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_8x8x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_8x8x32_neon_i8mm.c:60)	called 320 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f32_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm.c:75)	called 231 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f32_qsi8d32p1vlx4_qai4c32p4vlx4_1vlx4vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p1vlx4_qai4c32p4vlx4_1vlx4vl_sme2_mopa.c:102)	called 4136 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa.c:80)	called 141 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f32_qsi8d32p1x4_qai4c32p4vlx4_1x4vl_sme2_dot (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qai4c32p4vlx4_1x4vl_sme2_dot.c:98)	called 1980 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f32_qsi8d32p1x4_qai4c32p4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qai4c32p4x4_1x4_neon_dotprod.c:92)	called 714 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot.c:80)	called 81 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod.c:74)	called 52 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f32_qsi8d32p1x8_qai4c32p4x8_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p1x8_qai4c32p4x8_1x4_neon_dotprod.c:92)	called 714 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod.c:59)	called 52 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f32_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod.c:92)	called 1036 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod.c:78)	called 72 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f32_qsi8d32p4x8_qai4c32p4x8_8x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qai4c32p4x8_8x4_neon_i8mm.c:92)	called 1036 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm.c:62)	called 144 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_8x4x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_8x4x32_neon_i8mm.c:62)	called 72 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_qai8_qai8_qsi8cxp2vlx4sb_1x16vl_sme2_dot (kai/ukernels/matmul/matmul_clamp_qai8_qai8_qsi8cxp/kai_matmul_clamp_qai8_qai8_qsi8cxp2vlx4sb_1x16vl_sme2_dot.c:42)	called 840 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_qai8_qai8p2vlx4_qsi8cxp2vlx4sb_2vlx2vl_sme_mopa (kai/ukernels/matmul/matmul_clamp_qai8_qai8p_qsi8cxp/kai_matmul_clamp_qai8_qai8p2vlx4_qsi8cxp2vlx4sb_2vlx2vl_sme_mopa.c:49)	called 1665 times	100.0%	-%	100.0%
kai_get_n_step_matmul_clamp_qai8_qai8p2vlx4_qsi8cxpsb2vlx4_2vlx2vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_qai8_qai8p_qsi8cxp/kai_matmul_clamp_qai8_qai8p2vlx4_qsi8cxpsb2vlx4_2vlx2vl_sme2_mopa.c:49)	called 1665 times	100.0%	-%	100.0%
kai_get_n_step_rhs_imatmul_pack_kxn_qsi8cxp2vlx4sb_qs8cx_f32_i32_sme (kai/ukernels/matmul/pack/kai_rhs_imatmul_pack_kxn_qsi8cxp2vlx4sb_qs8cx_f32_i32_sme.c:46)	called 93240 times	100.0%	-%	100.0%
kai_get_n_step_rhs_imatmul_pack_kxn_x16p2vlx2b_x16_x16_sme (kai/ukernels/matmul/pack/kai_rhs_imatmul_pack_kxn_x16p2vlx2b_x16_x16_sme.c:42)	called 113520 times	100.0%	-%	100.0%
kai_get_n_step_rhs_imatmul_pack_kxn_x32p2vlx1b_x32_x32_sme (kai/ukernels/matmul/pack/kai_rhs_imatmul_pack_kxn_x32p2vlx1b_x32_x32_sme.c:39)	called 102168 times	100.0%	-%	100.0%
kai_get_n_step_rhs_pack_kxn_bf16p12x4biasf16_f16_neon (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_bf16p12x4biasf16_f16_neon.c:29)	not called	0.0%	-%	0.0%
kai_get_n_step_rhs_pack_kxn_bf16p12x4biasf32_f16_neon (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_bf16p12x4biasf32_f16_neon.c:27)	not called	0.0%	-%	0.0%
kai_get_n_step_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme.c:25)	called 382 times	100.0%	-%	100.0%
kai_get_n_step_rhs_pack_kxn_f16p16x1biasf16_f16_f16_neon (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_f16p16x1biasf16_f16_f16_neon.c:21)	called 17 times	100.0%	-%	100.0%
kai_get_n_step_rhs_pack_kxn_f32p16vlx1b_f32_f32_sme (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_f32p16vlx1b_f32_f32_sme.c:38)	called 506 times	100.0%	-%	100.0%
kai_get_n_step_rhs_pack_kxn_f32p2vlx1biasf32_f32_f32_sme (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_f32p2vlx1biasf32_f32_f32_sme.c:38)	called 1386 times	100.0%	-%	100.0%
kai_get_n_step_rhs_pack_kxn_f32p8x1biasf32_f32_f32_neon (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_f32p8x1biasf32_f32_f32_neon.c:21)	called 17 times	100.0%	-%	100.0%
kai_get_n_step_rhs_pack_kxn_qsi4c32p_qsu4c32s1s0 (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_qsi4c32p_qsu4c32s1s0.c:41)	not called	0.0%	-%	0.0%
kai_get_n_step_rhs_pack_kxn_qsi4cxp_qs4cxs1s0 (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_qsi4cxp_qs4cxs1s0.c:24)	not called	0.0%	-%	0.0%
kai_get_n_step_rhs_pack_kxn_qsi8cxp2vlx4sb_qs8cx_f32_i32_sme (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_qsi8cxp2vlx4sb_qs8cx_f32_i32_sme.c:45)	called 9696 times	100.0%	-%	100.0%
kai_get_n_step_rhs_pack_kxn_qsi8cxp_qsi8cx_neon (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_qsi8cxp_qsi8cx_neon.c:24)	not called	0.0%	-%	0.0%
kai_get_n_step_rhs_pack_kxn_x16p2vlx2b_x16_x16_sme (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_x16p2vlx2b_x16_x16_sme.c:41)	called 1512 times	100.0%	-%	100.0%
kai_get_n_step_rhs_pack_kxn_x16p32x1b_x16_x16_neon (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_x16p32x1b_x16_x16_neon.c:39)	called 374 times	100.0%	-%	100.0%
kai_get_n_step_rhs_pack_kxn_x32p16x1b_x32_x32_neon (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_x32p16x1b_x32_x32_neon.c:39)	called 374 times	100.0%	-%	100.0%
kai_get_n_step_rhs_pack_nxk_f32p2vlx1biasf32_f32_f32_sme (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_f32p2vlx1biasf32_f32_f32_sme.c:35)	called 36 times	100.0%	-%	100.0%
kai_get_n_step_rhs_pack_nxk_qsi4c32p_qsu4c32s1s0 (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32p_qsu4c32s1s0.c:41)	not called	0.0%	-%	0.0%
kai_get_n_step_rhs_pack_nxk_qsi4c32pnrx4_qsu4c32s1s0_neon (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32pnrx4_qsu4c32s1s0_neon.c:50)	not called	0.0%	-%	0.0%
kai_get_n_step_rhs_pack_nxk_qsi4c32pnrx8_qsu4c32s1s0_neon (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32pnrx8_qsu4c32s1s0_neon.c:50)	not called	0.0%	-%	0.0%
kai_get_n_step_rhs_pack_nxk_qsi4cxp_qs4cxs1s0 (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4cxp_qs4cxs1s0.c:24)	not called	0.0%	-%	0.0%
kai_get_n_step_rhs_pack_nxk_qsi4cxps1s0_qsu4cxs1s0_neon (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4cxps1s0_qsu4cxs1s0_neon.c:27)	not called	0.0%	-%	0.0%
kai_get_n_step_rhs_pack_nxk_qsi8cxp_qsi8cx_neon (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi8cxp_qsi8cx_neon.c:24)	not called	0.0%	-%	0.0%
kai_get_n_step_rhs_pack_nxk_x16p2vlx2b_x16_x16_sme (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_x16p2vlx2b_x16_x16_sme.c:35)	called 36 times	100.0%	-%	100.0%
kai_get_n_step_rhs_quant_pack_kxn_bf16p12x4biasf32_f32_neon (kai/ukernels/matmul/pack/kai_rhs_quant_pack_kxn_bf16p12x4biasf32_f32_neon.c:29)	not called	0.0%	-%	0.0%
kai_get_nr_matmul_clamp_bf16_qai8dxp1x8_qsi4c32p4x8_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_bf16_qai8dxp_qsi4c32p/kai_matmul_clamp_bf16_qai8dxp1x8_qsi4c32p4x8_1x4_neon_dotprod.c:110)	called 112 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_bf16_qai8dxp1x8_qsi4cxp8x8_1x8_neon_dotprod (kai/ukernels/matmul/matmul_clamp_bf16_qai8dxp_qsi4cxp/kai_matmul_clamp_bf16_qai8dxp1x8_qsi4cxp8x8_1x8_neon_dotprod.c:96)	called 392 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_bf16_qai8dxp4x8_qsi4c32p4x8_16x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_bf16_qai8dxp_qsi4c32p/kai_matmul_clamp_bf16_qai8dxp4x8_qsi4c32p4x8_16x4_neon_i8mm.c:111)	called 112 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_bf16_qai8dxp4x8_qsi4cxp8x8_8x8_neon_i8mm (kai/ukernels/matmul/matmul_clamp_bf16_qai8dxp_qsi4cxp/kai_matmul_clamp_bf16_qai8dxp4x8_qsi4cxp8x8_8x8_neon_i8mm.c:97)	called 392 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f16_bf16p8x4_bf16p12x4b_8x12_neon_mmla (kai/ukernels/matmul/matmul_clamp_f16_bf16p_bf16p/kai_matmul_clamp_f16_bf16p8x4_bf16p12x4b_8x12_neon_mmla.c:39)	called 18 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f16_f16_f16p16x1biasf16_6x16x8_neon_mla (kai/ukernels/matmul/matmul_clamp_f16_f16_f16p/kai_matmul_clamp_f16_f16_f16p16x1biasf16_6x16x8_neon_mla.c:36)	called 17 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f16_f16_f16p2vlx2b_1x16vl_sme2_dot (kai/ukernels/matmul/matmul_clamp_f16_f16_f16p/kai_matmul_clamp_f16_f16_f16p2vlx2b_1x16vl_sme2_dot.c:47)	called 46 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f16_f16_f16p2vlx2b_1x8vl_sme_mla (kai/ukernels/matmul/matmul_clamp_f16_f16_f16p/kai_matmul_clamp_f16_f16_f16p2vlx2b_1x8vl_sme_mla.c:47)	called 46 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f16_f16_f16p32x1b_6x32_neon_mla (kai/ukernels/matmul/matmul_clamp_f16_f16_f16p/kai_matmul_clamp_f16_f16_f16p32x1b_6x32_neon_mla.c:52)	called 17 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f16_f16_f16p32x1b_6x32_neon_mla_cortexa55 (kai/ukernels/matmul/matmul_clamp_f16_f16_f16p/kai_matmul_clamp_f16_f16_f16p32x1b_6x32_neon_mla_cortexa55.c:52)	called 17 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f16_f16p2vlx2_f16p2vlx2_2vlx2vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f16_f16p_f16p/kai_matmul_clamp_f16_f16p2vlx2_f16p2vlx2_2vlx2vl_sme2_mopa.c:57)	called 34 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f16_f16p2vlx2_f16p2vlx2b_2vlx2vl_sme_mopa (kai/ukernels/matmul/matmul_clamp_f16_f16p_f16p/kai_matmul_clamp_f16_f16p2vlx2_f16p2vlx2b_2vlx2vl_sme_mopa.c:57)	called 34 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f16_qai8dxp1x4_qsi4cxp4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi4cxp/kai_matmul_clamp_f16_qai8dxp1x4_qsi4cxp4x4_1x4_neon_dotprod.c:96)	called 224 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f16_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi8cxp/kai_matmul_clamp_f16_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod.c:93)	called 224 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f16_qai8dxp1x8_qsi4cxp4x8_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi4cxp/kai_matmul_clamp_f16_qai8dxp1x8_qsi4cxp4x8_1x4_neon_dotprod.c:96)	called 224 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f16_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi8cxp/kai_matmul_clamp_f16_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod.c:93)	called 224 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f16_qai8dxp4x4_qsi4cxp4x4_16x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi4cxp/kai_matmul_clamp_f16_qai8dxp4x4_qsi4cxp4x4_16x4_neon_dotprod.c:96)	called 224 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f16_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi8cxp/kai_matmul_clamp_f16_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod.c:93)	called 224 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f16_qai8dxp4x8_qsi4cxp4x8_16x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi4cxp/kai_matmul_clamp_f16_qai8dxp4x8_qsi4cxp4x8_16x4_neon_i8mm.c:96)	called 224 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f16_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi8cxp/kai_matmul_clamp_f16_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm.c:93)	called 224 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f16_qsi8d32p1vlx4_qai4c32p4vlx4_1vlx4vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p1vlx4_qai4c32p4vlx4_1vlx4vl_sme2_mopa.c:110)	called 6202 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f16_qsi8d32p1x4_qai4c32p4vlx4_1x4vl_sme2_dot (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p1x4_qai4c32p4vlx4_1x4vl_sme2_dot.c:106)	called 2901 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f16_qsi8d32p1x4_qai4c32p4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p1x4_qai4c32p4x4_1x4_neon_dotprod.c:102)	called 1036 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f16_qsi8d32p1x8_qai4c32p4x8_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p1x8_qai4c32p4x8_1x4_neon_dotprod.c:102)	called 1036 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f16_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod.c:102)	called 1036 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f16_qsi8d32p4x8_qai4c32p4x8_8x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p4x8_qai4c32p4x8_8x4_neon_i8mm.c:102)	called 1036 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f32_bf16p1x4_bf16p12x4b_1x36_neon_dot (kai/ukernels/matmul/matmul_clamp_f32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p1x4_bf16p12x4b_1x36_neon_dot.c:41)	called 24 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa.c:56)	called 10 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f32_bf16p8x4_bf16p12x4b_8x12_neon_mmla (kai/ukernels/matmul/matmul_clamp_f32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p8x4_bf16p12x4b_8x12_neon_mmla.c:40)	called 40 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f32_f32_f32p16vlx1b_1x16vl_sme2_mla (kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p16vlx1b_1x16vl_sme2_mla.c:45)	called 46 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f32_f32_f32p16x1b_6x16_neon_mla (kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p16x1b_6x16_neon_mla.c:48)	called 17 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f32_f32_f32p16x1b_6x16_neon_mla_cortexa55 (kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p16x1b_6x16_neon_mla_cortexa55.c:48)	called 17 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla (kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla.c:45)	called 46 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f32_f32_f32p2vlx1b_1x8vl_sme_mla (kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p2vlx1b_1x8vl_sme_mla.c:45)	called 46 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f32_f32_f32p8x1biasf32_6x8x4_neon_mla (kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p8x1biasf32_6x8x4_neon_mla.c:49)	called 17 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme_mopa (kai/ukernels/matmul/matmul_clamp_f32_f32p_f32p/kai_matmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme_mopa.c:56)	called 34 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f32_f32p_f32p/kai_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa.c:56)	called 34 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa.c:102)	called 1003 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme_mopa (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme_mopa.c:102)	called 1003 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f32_qai8dxp1vlx8_qsi4cxp4vlx8_1vlx4vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1vlx8_qsi4cxp4vlx8_1vlx4vl_sme2_mopa.c:72)	called 522 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f32_qai8dxp1x4_qsi4c32p4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x4_qsi4c32p4x4_1x4_neon_dotprod.c:109)	called 112 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f32_qai8dxp1x4_qsi4c32p8x4_1x8_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x4_qsi4c32p8x4_1x8_neon_dotprod.c:109)	called 112 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f32_qai8dxp1x4_qsi4cxp4vlx4_1x4vl_sme2_sdot (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi4cxp4vlx4_1x4vl_sme2_sdot.c:61)	called 682 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f32_qai8dxp1x4_qsi4cxp4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi4cxp4x4_1x4_neon_dotprod.c:81)	called 240 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot.c:105)	called 923 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme_dot (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme_dot.c:105)	called 923 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod.c:79)	called 231 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f32_qai8dxp1x8_qsi4c32p4x8_1x4x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x8_qsi4c32p4x8_1x4x32_neon_dotprod.c:109)	called 112 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f32_qai8dxp1x8_qsi4c32p8x8_1x8_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x8_qsi4c32p8x8_1x8_neon_dotprod.c:109)	called 112 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f32_qai8dxp1x8_qsi4c32p8x8_1x8x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x8_qsi4c32p8x8_1x8x32_neon_dotprod.c:109)	called 112 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f32_qai8dxp1x8_qsi4cxp4x8_1x4x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi4cxp4x8_1x4x32_neon_dotprod.c:65)	called 240 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f32_qai8dxp1x8_qsi4cxp8x8_1x8x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi4cxp8x8_1x8x32_neon_dotprod.c:65)	called 240 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f32_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod.c:79)	called 231 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f32_qai8dxp4x4_qsi4c32p4x4_16x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x4_qsi4c32p4x4_16x4_neon_dotprod.c:110)	called 112 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod.c:109)	called 112 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f32_qai8dxp4x4_qsi4cxp8x4_8x8x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x4_qsi4cxp8x4_8x8x32_neon_dotprod.c:67)	called 240 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f32_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod.c:83)	called 231 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f32_qai8dxp4x8_qsi4c32p4x8_16x4x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p4x8_16x4x32_neon_i8mm.c:110)	called 112 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f32_qai8dxp4x8_qsi4c32p4x8_8x4x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p4x8_8x4x32_neon_i8mm.c:109)	called 112 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f32_qai8dxp4x8_qsi4c32p8x8_4x8_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p8x8_4x8_neon_i8mm.c:109)	called 112 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f32_qai8dxp4x8_qsi4c32p8x8_4x8x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p8x8_4x8x32_neon_i8mm.c:109)	called 112 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x4_16x4x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x4_16x4x32_neon_dotprod.c:67)	called 240 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_4x4x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_4x4x32_neon_i8mm.c:64)	called 240 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_8x4x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_8x4x32_neon_i8mm.c:68)	called 240 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_4x8x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_4x8x32_neon_i8mm.c:68)	called 240 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_8x8x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_8x8x32_neon_i8mm.c:68)	called 240 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f32_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm.c:83)	called 231 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f32_qsi8d32p1vlx4_qai4c32p4vlx4_1vlx4vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p1vlx4_qai4c32p4vlx4_1vlx4vl_sme2_mopa.c:110)	called 6202 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa.c:88)	called 188 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f32_qsi8d32p1x4_qai4c32p4vlx4_1x4vl_sme2_dot (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qai4c32p4vlx4_1x4vl_sme2_dot.c:106)	called 2901 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f32_qsi8d32p1x4_qai4c32p4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qai4c32p4x4_1x4_neon_dotprod.c:100)	called 1036 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot.c:88)	called 108 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod.c:82)	called 48 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f32_qsi8d32p1x8_qai4c32p4x8_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p1x8_qai4c32p4x8_1x4_neon_dotprod.c:100)	called 1036 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod.c:67)	called 48 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f32_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod.c:100)	called 1036 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod.c:86)	called 48 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f32_qsi8d32p4x8_qai4c32p4x8_8x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qai4c32p4x8_8x4_neon_i8mm.c:100)	called 1036 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm.c:70)	called 96 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_8x4x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_8x4x32_neon_i8mm.c:70)	called 48 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_qai8_qai8_qsi8cxp2vlx4sb_1x16vl_sme2_dot (kai/ukernels/matmul/matmul_clamp_qai8_qai8_qsi8cxp/kai_matmul_clamp_qai8_qai8_qsi8cxp2vlx4sb_1x16vl_sme2_dot.c:46)	called 168 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_qai8_qai8p2vlx4_qsi8cxp2vlx4sb_2vlx2vl_sme_mopa (kai/ukernels/matmul/matmul_clamp_qai8_qai8p_qsi8cxp/kai_matmul_clamp_qai8_qai8p2vlx4_qsi8cxp2vlx4sb_2vlx2vl_sme_mopa.c:57)	called 333 times	100.0%	-%	100.0%
kai_get_nr_matmul_clamp_qai8_qai8p2vlx4_qsi8cxpsb2vlx4_2vlx2vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_qai8_qai8p_qsi8cxp/kai_matmul_clamp_qai8_qai8p2vlx4_qsi8cxpsb2vlx4_2vlx2vl_sme2_mopa.c:57)	called 333 times	100.0%	-%	100.0%
kai_get_num_blocks_per_row (kai/ukernels/matmul/matmul_clamp_bf16_qai8dxp_qsi4c32p/kai_matmul_clamp_bf16_qai8dxp1x8_qsi4c32p4x8_1x4_neon_dotprod.c:66)	called 271 times	100.0%	-%	57.0%
kai_get_num_blocks_per_row (kai/ukernels/matmul/matmul_clamp_bf16_qai8dxp_qsi4c32p/kai_matmul_clamp_bf16_qai8dxp4x8_qsi4c32p4x8_16x4_neon_i8mm.c:67)	called 271 times	100.0%	-%	57.0%
kai_get_num_blocks_per_row (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p1vlx4_qai4c32p4vlx4_1vlx4vl_sme2_mopa.c:72)	called 4130 times	100.0%	-%	57.0%
kai_get_num_blocks_per_row (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p1x4_qai4c32p4vlx4_1x4vl_sme2_dot.c:68)	called 829 times	100.0%	-%	57.0%
kai_get_num_blocks_per_row (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p1x4_qai4c32p4x4_1x4_neon_dotprod.c:67)	called 415 times	100.0%	-%	57.0%
kai_get_num_blocks_per_row (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p1x8_qai4c32p4x8_1x4_neon_dotprod.c:67)	called 415 times	100.0%	-%	57.0%
kai_get_num_blocks_per_row (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod.c:67)	called 1549 times	100.0%	-%	57.0%
kai_get_num_blocks_per_row (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p4x8_qai4c32p4x8_8x4_neon_i8mm.c:67)	called 1549 times	100.0%	-%	57.0%
kai_get_num_blocks_per_row (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x4_qsi4c32p4x4_1x4_neon_dotprod.c:65)	called 271 times	100.0%	-%	57.0%
kai_get_num_blocks_per_row (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x4_qsi4c32p8x4_1x8_neon_dotprod.c:65)	called 271 times	100.0%	-%	57.0%
kai_get_num_blocks_per_row (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x8_qsi4c32p4x8_1x4x32_neon_dotprod.c:65)	called 271 times	100.0%	-%	57.0%
kai_get_num_blocks_per_row (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x8_qsi4c32p8x8_1x8_neon_dotprod.c:65)	called 271 times	100.0%	-%	57.0%
kai_get_num_blocks_per_row (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x8_qsi4c32p8x8_1x8x32_neon_dotprod.c:65)	called 271 times	100.0%	-%	57.0%
kai_get_num_blocks_per_row (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x4_qsi4c32p4x4_16x4_neon_dotprod.c:66)	called 271 times	100.0%	-%	57.0%
kai_get_num_blocks_per_row (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod.c:65)	called 271 times	100.0%	-%	57.0%
kai_get_num_blocks_per_row (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p4x8_16x4x32_neon_i8mm.c:66)	called 271 times	100.0%	-%	57.0%
kai_get_num_blocks_per_row (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p4x8_8x4x32_neon_i8mm.c:65)	called 271 times	100.0%	-%	57.0%
kai_get_num_blocks_per_row (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p8x8_4x8_neon_i8mm.c:65)	called 271 times	100.0%	-%	57.0%
kai_get_num_blocks_per_row (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p8x8_4x8x32_neon_i8mm.c:65)	called 271 times	100.0%	-%	57.0%
kai_get_num_blocks_per_row (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p1vlx4_qai4c32p4vlx4_1vlx4vl_sme2_mopa.c:72)	called 4130 times	100.0%	-%	57.0%
kai_get_num_blocks_per_row (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qai4c32p4vlx4_1x4vl_sme2_dot.c:68)	called 829 times	100.0%	-%	57.0%
kai_get_num_blocks_per_row (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qai4c32p4x4_1x4_neon_dotprod.c:65)	called 415 times	100.0%	-%	57.0%
kai_get_num_blocks_per_row (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p1x8_qai4c32p4x8_1x4_neon_dotprod.c:65)	called 415 times	100.0%	-%	57.0%
kai_get_num_blocks_per_row (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod.c:65)	called 1549 times	100.0%	-%	57.0%
kai_get_num_blocks_per_row (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qai4c32p4x8_8x4_neon_i8mm.c:65)	called 1549 times	100.0%	-%	57.0%
kai_get_num_blocks_per_row (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa.c:51)	called 164 times	100.0%	-%	53.0%
kai_get_num_blocks_per_row (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot.c:51)	called 64 times	100.0%	-%	53.0%
kai_get_num_blocks_per_row (kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32pscalef32_f16_neon.c:30)	called 4805224 times	100.0%	-%	57.0%
kai_get_num_blocks_per_row (kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32pscalef32_f32_neon.c:28)	called 4794320 times	100.0%	-%	57.0%
kai_get_num_blocks_per_row (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_qsi4c32p_qsu4c32s1s0.c:19)	called 3510 times	100.0%	-%	57.0%
kai_get_num_blocks_per_row (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qai4c32p_qau4c32s0s1_f32_f32_f32_neon.c:22)	called 7848 times	100.0%	-%	52.0%
kai_get_num_blocks_per_row (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qai4c32ps1s0nrx4_qau4c32s0s1_f32_f32_f32_neon.c:24)	called 3924 times	100.0%	-%	52.0%
kai_get_num_blocks_per_row (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qai4c32ps1s0nrx4_qau4c32s1s0_f32_f32_f32_neon.c:24)	called 3924 times	100.0%	-%	52.0%
kai_get_num_blocks_per_row (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32p_qsu4c32s1s0.c:19)	called 2808 times	100.0%	-%	57.0%
kai_get_num_blocks_per_row (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32pnrx4_qsu4c32s1s0_neon.c:28)	called 864 times	100.0%	-%	57.0%
kai_get_num_blocks_per_row (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32pnrx8_qsu4c32s1s0_neon.c:28)	called 1944 times	100.0%	-%	57.0%
kai_get_num_bytes_per_block (kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32pscalef32_f16_neon.c:26)	called 4805224 times	100.0%	-%	100.0%
kai_get_num_bytes_per_block (kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32pscalef32_f32_neon.c:24)	called 4794320 times	100.0%	-%	100.0%
kai_get_num_bytes_per_block (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_qsi4c32p_qsu4c32s1s0.c:24)	called 3510 times	100.0%	-%	57.0%
kai_get_num_bytes_per_block (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qai4c32p_qau4c32s0s1_f32_f32_f32_neon.c:29)	called 10464 times	100.0%	-%	100.0%
kai_get_num_bytes_per_block (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qai4c32ps1s0nrx4_qau4c32s0s1_f32_f32_f32_neon.c:31)	called 5232 times	100.0%	-%	100.0%
kai_get_num_bytes_per_block (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qai4c32ps1s0nrx4_qau4c32s1s0_f32_f32_f32_neon.c:31)	called 5232 times	100.0%	-%	100.0%
kai_get_num_bytes_per_block (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32p_qsu4c32s1s0.c:24)	called 2106 times	100.0%	-%	57.0%
kai_get_num_bytes_per_block (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32pnrx4_qsu4c32s1s0_neon.c:33)	called 648 times	100.0%	-%	57.0%
kai_get_num_bytes_per_block (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32pnrx8_qsu4c32s1s0_neon.c:33)	called 1458 times	100.0%	-%	57.0%
kai_get_num_bytes_per_block_lhs (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p1vlx4_qai4c32p4vlx4_1vlx4vl_sme2_mopa.c:61)	called 2065 times	100.0%	-%	100.0%
kai_get_num_bytes_per_block_lhs (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p1x4_qai4c32p4vlx4_1x4vl_sme2_dot.c:57)	called 276 times	100.0%	-%	100.0%
kai_get_num_bytes_per_block_lhs (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p1x4_qai4c32p4x4_1x4_neon_dotprod.c:56)	called 138 times	100.0%	-%	100.0%
kai_get_num_bytes_per_block_lhs (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p1x8_qai4c32p4x8_1x4_neon_dotprod.c:56)	called 138 times	100.0%	-%	100.0%
kai_get_num_bytes_per_block_lhs (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod.c:56)	called 516 times	100.0%	-%	100.0%
kai_get_num_bytes_per_block_lhs (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p4x8_qai4c32p4x8_8x4_neon_i8mm.c:56)	called 516 times	100.0%	-%	100.0%
kai_get_num_bytes_per_block_lhs (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p1vlx4_qai4c32p4vlx4_1vlx4vl_sme2_mopa.c:61)	called 2065 times	100.0%	-%	100.0%
kai_get_num_bytes_per_block_lhs (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qai4c32p4vlx4_1x4vl_sme2_dot.c:57)	called 276 times	100.0%	-%	100.0%
kai_get_num_bytes_per_block_lhs (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qai4c32p4x4_1x4_neon_dotprod.c:54)	called 138 times	100.0%	-%	100.0%
kai_get_num_bytes_per_block_lhs (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p1x8_qai4c32p4x8_1x4_neon_dotprod.c:54)	called 138 times	100.0%	-%	100.0%
kai_get_num_bytes_per_block_lhs (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod.c:54)	called 516 times	100.0%	-%	100.0%
kai_get_num_bytes_per_block_lhs (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qai4c32p4x8_8x4_neon_i8mm.c:54)	called 516 times	100.0%	-%	100.0%
kai_get_num_bytes_per_block_lhs (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa.c:41)	called 70 times	100.0%	-%	100.0%
kai_get_num_bytes_per_block_lhs (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot.c:41)	called 30 times	100.0%	-%	100.0%
kai_get_num_bytes_per_block_rhs (kai/ukernels/matmul/matmul_clamp_bf16_qai8dxp_qsi4c32p/kai_matmul_clamp_bf16_qai8dxp1x8_qsi4c32p4x8_1x4_neon_dotprod.c:60)	called 108 times	100.0%	-%	57.0%
kai_get_num_bytes_per_block_rhs (kai/ukernels/matmul/matmul_clamp_bf16_qai8dxp_qsi4c32p/kai_matmul_clamp_bf16_qai8dxp4x8_qsi4c32p4x8_16x4_neon_i8mm.c:61)	called 108 times	100.0%	-%	57.0%
kai_get_num_bytes_per_block_rhs (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p1vlx4_qai4c32p4vlx4_1vlx4vl_sme2_mopa.c:65)	called 2065 times	100.0%	-%	57.0%
kai_get_num_bytes_per_block_rhs (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p1x4_qai4c32p4vlx4_1x4vl_sme2_dot.c:61)	called 553 times	100.0%	-%	57.0%
kai_get_num_bytes_per_block_rhs (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p1x4_qai4c32p4x4_1x4_neon_dotprod.c:60)	called 138 times	100.0%	-%	57.0%
kai_get_num_bytes_per_block_rhs (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p1x8_qai4c32p4x8_1x4_neon_dotprod.c:60)	called 138 times	100.0%	-%	57.0%
kai_get_num_bytes_per_block_rhs (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod.c:60)	called 516 times	100.0%	-%	57.0%
kai_get_num_bytes_per_block_rhs (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p4x8_qai4c32p4x8_8x4_neon_i8mm.c:60)	called 516 times	100.0%	-%	57.0%
kai_get_num_bytes_per_block_rhs (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x4_qsi4c32p4x4_1x4_neon_dotprod.c:59)	called 108 times	100.0%	-%	57.0%
kai_get_num_bytes_per_block_rhs (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x4_qsi4c32p8x4_1x8_neon_dotprod.c:59)	called 108 times	100.0%	-%	57.0%
kai_get_num_bytes_per_block_rhs (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x8_qsi4c32p4x8_1x4x32_neon_dotprod.c:59)	called 108 times	100.0%	-%	57.0%
kai_get_num_bytes_per_block_rhs (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x8_qsi4c32p8x8_1x8_neon_dotprod.c:59)	called 108 times	100.0%	-%	57.0%
kai_get_num_bytes_per_block_rhs (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x8_qsi4c32p8x8_1x8x32_neon_dotprod.c:59)	called 108 times	100.0%	-%	57.0%
kai_get_num_bytes_per_block_rhs (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x4_qsi4c32p4x4_16x4_neon_dotprod.c:60)	called 108 times	100.0%	-%	57.0%
kai_get_num_bytes_per_block_rhs (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod.c:59)	called 108 times	100.0%	-%	57.0%
kai_get_num_bytes_per_block_rhs (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p4x8_16x4x32_neon_i8mm.c:60)	called 108 times	100.0%	-%	57.0%
kai_get_num_bytes_per_block_rhs (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p4x8_8x4x32_neon_i8mm.c:59)	called 108 times	100.0%	-%	57.0%
kai_get_num_bytes_per_block_rhs (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p8x8_4x8_neon_i8mm.c:59)	called 108 times	100.0%	-%	57.0%
kai_get_num_bytes_per_block_rhs (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p8x8_4x8x32_neon_i8mm.c:59)	called 108 times	100.0%	-%	57.0%
kai_get_num_bytes_per_block_rhs (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p1vlx4_qai4c32p4vlx4_1vlx4vl_sme2_mopa.c:65)	called 2065 times	100.0%	-%	57.0%
kai_get_num_bytes_per_block_rhs (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qai4c32p4vlx4_1x4vl_sme2_dot.c:61)	called 553 times	100.0%	-%	57.0%
kai_get_num_bytes_per_block_rhs (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qai4c32p4x4_1x4_neon_dotprod.c:58)	called 138 times	100.0%	-%	57.0%
kai_get_num_bytes_per_block_rhs (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p1x8_qai4c32p4x8_1x4_neon_dotprod.c:58)	called 138 times	100.0%	-%	57.0%
kai_get_num_bytes_per_block_rhs (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod.c:58)	called 516 times	100.0%	-%	57.0%
kai_get_num_bytes_per_block_rhs (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qai4c32p4x8_8x4_neon_i8mm.c:58)	called 516 times	100.0%	-%	57.0%
kai_get_num_bytes_per_block_rhs (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa.c:45)	called 70 times	100.0%	-%	57.0%
kai_get_num_bytes_per_block_rhs (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot.c:45)	called 30 times	100.0%	-%	57.0%
kai_get_rhs_offset_rhs_imatmul_pack_kxn_qsi8cxp2vlx4sb_qs8cx_f32_i32_sme (kai/ukernels/matmul/pack/kai_rhs_imatmul_pack_kxn_qsi8cxp2vlx4sb_qs8cx_f32_i32_sme.c:50)	called 9324 times	100.0%	-%	57.0%
kai_get_rhs_offset_rhs_imatmul_pack_kxn_x16p2vlx2b_x16_x16_sme (kai/ukernels/matmul/pack/kai_rhs_imatmul_pack_kxn_x16p2vlx2b_x16_x16_sme.c:46)	called 11352 times	100.0%	-%	57.0%
kai_get_rhs_offset_rhs_imatmul_pack_kxn_x32p2vlx1b_x32_x32_sme (kai/ukernels/matmul/pack/kai_rhs_imatmul_pack_kxn_x32p2vlx1b_x32_x32_sme.c:43)	called 11352 times	100.0%	-%	57.0%
kai_get_rhs_offset_rhs_pack_kxn_bf16p12x4biasf16_f16_neon (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_bf16p12x4biasf16_f16_neon.c:33)	called 84 times	100.0%	-%	57.0%
kai_get_rhs_offset_rhs_pack_kxn_bf16p12x4biasf32_f16_neon (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_bf16p12x4biasf32_f16_neon.c:31)	called 92 times	100.0%	-%	57.0%
kai_get_rhs_offset_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme.c:29)	called 46 times	100.0%	-%	57.0%
kai_get_rhs_offset_rhs_pack_kxn_f16p16x1biasf16_f16_f16_neon (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_f16p16x1biasf16_f16_f16_neon.c:25)	called 17 times	100.0%	-%	57.0%
kai_get_rhs_offset_rhs_pack_kxn_f32p16vlx1b_f32_f32_sme (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_f32p16vlx1b_f32_f32_sme.c:42)	called 46 times	100.0%	-%	57.0%
kai_get_rhs_offset_rhs_pack_kxn_f32p2vlx1biasf32_f32_f32_sme (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_f32p2vlx1biasf32_f32_f32_sme.c:42)	called 126 times	100.0%	-%	57.0%
kai_get_rhs_offset_rhs_pack_kxn_f32p8x1biasf32_f32_f32_neon (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_f32p8x1biasf32_f32_f32_neon.c:25)	called 17 times	100.0%	-%	57.0%
kai_get_rhs_offset_rhs_pack_kxn_qsi4c32p_qsu4c32s1s0 (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_qsi4c32p_qsu4c32s1s0.c:45)	called 702 times	100.0%	-%	57.0%
kai_get_rhs_offset_rhs_pack_kxn_qsi4cxp_qs4cxs1s0 (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_qsi4cxp_qs4cxs1s0.c:28)	not called	0.0%	-%	0.0%
kai_get_rhs_offset_rhs_pack_kxn_qsi8cxp2vlx4sb_qs8cx_f32_i32_sme (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_qsi8cxp2vlx4sb_qs8cx_f32_i32_sme.c:49)	called 834 times	100.0%	-%	57.0%
kai_get_rhs_offset_rhs_pack_kxn_qsi8cxp_qsi8cx_neon (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_qsi8cxp_qsi8cx_neon.c:28)	not called	0.0%	-%	0.0%
kai_get_rhs_offset_rhs_pack_kxn_x16p2vlx2b_x16_x16_sme (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_x16p2vlx2b_x16_x16_sme.c:45)	called 126 times	100.0%	-%	57.0%
kai_get_rhs_offset_rhs_pack_kxn_x16p32x1b_x16_x16_neon (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_x16p32x1b_x16_x16_neon.c:43)	called 34 times	100.0%	-%	57.0%
kai_get_rhs_offset_rhs_pack_kxn_x32p16x1b_x32_x32_neon (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_x32p16x1b_x32_x32_neon.c:43)	called 34 times	100.0%	-%	57.0%
kai_get_rhs_offset_rhs_pack_nxk_f32p2vlx1biasf32_f32_f32_sme (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_f32p2vlx1biasf32_f32_f32_sme.c:39)	called 34 times	100.0%	-%	57.0%
kai_get_rhs_offset_rhs_pack_nxk_qai4c32p_qau4c32s0s1_f32_f32_f32_neon (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qai4c32p_qau4c32s0s1_f32_f32_f32_neon.c:44)	not called	0.0%	-%	0.0%
kai_get_rhs_offset_rhs_pack_nxk_qai4c32ps1s0nrx4_qau4c32s0s1_f32_f32_f32_neon (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qai4c32ps1s0nrx4_qau4c32s0s1_f32_f32_f32_neon.c:46)	not called	0.0%	-%	0.0%
kai_get_rhs_offset_rhs_pack_nxk_qai4c32ps1s0nrx4_qau4c32s1s0_f32_f32_f32_neon (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qai4c32ps1s0nrx4_qau4c32s1s0_f32_f32_f32_neon.c:46)	not called	0.0%	-%	0.0%
kai_get_rhs_offset_rhs_pack_nxk_qsi4c32p_qsu4c32s1s0 (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32p_qsu4c32s1s0.c:45)	called 702 times	100.0%	-%	100.0%
kai_get_rhs_offset_rhs_pack_nxk_qsi4c32pnrx4_qsu4c32s1s0_neon (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32pnrx4_qsu4c32s1s0_neon.c:54)	called 216 times	100.0%	-%	100.0%
kai_get_rhs_offset_rhs_pack_nxk_qsi4c32pnrx8_qsu4c32s1s0_neon (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32pnrx8_qsu4c32s1s0_neon.c:54)	called 486 times	100.0%	-%	100.0%
kai_get_rhs_offset_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon.c:71)	not called	0.0%	-%	0.0%
kai_get_rhs_offset_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0 (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0.c:51)	not called	0.0%	-%	0.0%
kai_get_rhs_offset_rhs_pack_nxk_qsi4cxp_qs4cxs1s0 (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4cxp_qs4cxs1s0.c:28)	called 720 times	100.0%	-%	100.0%
kai_get_rhs_offset_rhs_pack_nxk_qsi4cxps1s0_qsu4cxs1s0_neon (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4cxps1s0_qsu4cxs1s0_neon.c:31)	called 160 times	100.0%	-%	100.0%
kai_get_rhs_offset_rhs_pack_nxk_qsi8cxp_qsi8cx_neon (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi8cxp_qsi8cx_neon.c:28)	not called	0.0%	-%	0.0%
kai_get_rhs_offset_rhs_pack_nxk_x16p2vlx2b_x16_x16_sme (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_x16p2vlx2b_x16_x16_sme.c:39)	called 34 times	100.0%	-%	57.0%
kai_get_rhs_offset_rhs_quant_pack_kxn_bf16p12x4biasf32_f32_neon (kai/ukernels/matmul/pack/kai_rhs_quant_pack_kxn_bf16p12x4biasf32_f32_neon.c:33)	called 184 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_end_of_all_blocks (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_qsi4c32p_qsu4c32s1s0.c:29)	called 702 times	100.0%	-%	52.0%
kai_get_rhs_packed_offset_end_of_all_blocks (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32p_qsu4c32s1s0.c:29)	called 702 times	100.0%	-%	52.0%
kai_get_rhs_packed_offset_end_of_all_blocks (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32pnrx4_qsu4c32s1s0_neon.c:38)	called 216 times	100.0%	-%	52.0%
kai_get_rhs_packed_offset_end_of_all_blocks (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32pnrx8_qsu4c32s1s0_neon.c:38)	called 486 times	100.0%	-%	52.0%
kai_get_rhs_packed_offset_imatmul_clamp_f16_f16p2vlx2_f16p2vlx2_2vlx2vl_sme2_mopa (kai/ukernels/matmul/imatmul_clamp_f16_f16p_f16p/kai_imatmul_clamp_f16_f16p2vlx2_f16p2vlx2_2vlx2vl_sme2_mopa.c:64)	called 5676 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_imatmul_clamp_f16_f16p2vlx2_f16p2vlx2b_2vlx2vl_sme_mopa (kai/ukernels/matmul/imatmul_clamp_f16_f16p_f16p/kai_imatmul_clamp_f16_f16p2vlx2_f16p2vlx2b_2vlx2vl_sme_mopa.c:64)	called 5676 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_imatmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme2_mopa (kai/ukernels/matmul/imatmul_clamp_f32_f32p_f32p/kai_imatmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme2_mopa.c:63)	called 5676 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_imatmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme_mopa (kai/ukernels/matmul/imatmul_clamp_f32_f32p_f32p/kai_imatmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme_mopa.c:63)	called 5676 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_imatmul_clamp_qai8_qai8p2vlx4_qsi8cxp2vlx4sb_2vlx2vl_sme_mopa (kai/ukernels/matmul/imatmul_clamp_qai8_qai8p_qsi8cxp/kai_imatmul_clamp_qai8_qai8p2vlx4_qsi8cxp2vlx4sb_2vlx2vl_sme_mopa.c:64)	called 4662 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_imatmul_clamp_qai8_qai8p2vlx4_qsi8cxpsb2vlx4_2vlx2vl_sme2_mopa (kai/ukernels/matmul/imatmul_clamp_qai8_qai8p_qsi8cxp/kai_imatmul_clamp_qai8_qai8p2vlx4_qsi8cxpsb2vlx4_2vlx2vl_sme2_mopa.c:64)	called 4662 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_matmul_clamp_bf16_qai8dxp1x8_qsi4c32p4x8_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_bf16_qai8dxp_qsi4c32p/kai_matmul_clamp_bf16_qai8dxp1x8_qsi4c32p4x8_1x4_neon_dotprod.c:128)	called 108 times	100.0%	-%	53.0%
kai_get_rhs_packed_offset_matmul_clamp_bf16_qai8dxp1x8_qsi4cxp8x8_1x8_neon_dotprod (kai/ukernels/matmul/matmul_clamp_bf16_qai8dxp_qsi4cxp/kai_matmul_clamp_bf16_qai8dxp1x8_qsi4cxp8x8_1x8_neon_dotprod.c:114)	called 356 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_matmul_clamp_bf16_qai8dxp4x8_qsi4c32p4x8_16x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_bf16_qai8dxp_qsi4c32p/kai_matmul_clamp_bf16_qai8dxp4x8_qsi4c32p4x8_16x4_neon_i8mm.c:129)	called 108 times	100.0%	-%	53.0%
kai_get_rhs_packed_offset_matmul_clamp_bf16_qai8dxp4x8_qsi4cxp8x8_8x8_neon_i8mm (kai/ukernels/matmul/matmul_clamp_bf16_qai8dxp_qsi4cxp/kai_matmul_clamp_bf16_qai8dxp4x8_qsi4cxp8x8_8x8_neon_i8mm.c:115)	called 388 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_matmul_clamp_f16_bf16p8x4_bf16p12x4b_8x12_neon_mmla (kai/ukernels/matmul/matmul_clamp_f16_bf16p_bf16p/kai_matmul_clamp_f16_bf16p8x4_bf16p12x4b_8x12_neon_mmla.c:57)	called 84 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_matmul_clamp_f16_f16_f16p16x1biasf16_6x16x8_neon_mla (kai/ukernels/matmul/matmul_clamp_f16_f16_f16p/kai_matmul_clamp_f16_f16_f16p16x1biasf16_6x16x8_neon_mla.c:54)	called 16 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_matmul_clamp_f16_f16_f16p2vlx2b_1x16vl_sme2_dot (kai/ukernels/matmul/matmul_clamp_f16_f16_f16p/kai_matmul_clamp_f16_f16_f16p2vlx2b_1x16vl_sme2_dot.c:70)	called 46 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_matmul_clamp_f16_f16_f16p2vlx2b_1x8vl_sme_mla (kai/ukernels/matmul/matmul_clamp_f16_f16_f16p/kai_matmul_clamp_f16_f16_f16p2vlx2b_1x8vl_sme_mla.c:70)	called 46 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_matmul_clamp_f16_f16_f16p32x1b_6x32_neon_mla (kai/ukernels/matmul/matmul_clamp_f16_f16_f16p/kai_matmul_clamp_f16_f16_f16p32x1b_6x32_neon_mla.c:70)	called 16 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_matmul_clamp_f16_f16_f16p32x1b_6x32_neon_mla_cortexa55 (kai/ukernels/matmul/matmul_clamp_f16_f16_f16p/kai_matmul_clamp_f16_f16_f16p32x1b_6x32_neon_mla_cortexa55.c:70)	called 16 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_matmul_clamp_f16_f16p2vlx2_f16p2vlx2_2vlx2vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f16_f16p_f16p/kai_matmul_clamp_f16_f16p2vlx2_f16p2vlx2_2vlx2vl_sme2_mopa.c:79)	called 16 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_matmul_clamp_f16_f16p2vlx2_f16p2vlx2b_2vlx2vl_sme_mopa (kai/ukernels/matmul/matmul_clamp_f16_f16p_f16p/kai_matmul_clamp_f16_f16p2vlx2_f16p2vlx2b_2vlx2vl_sme_mopa.c:79)	called 16 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_matmul_clamp_f16_qai8dxp1x4_qsi4cxp4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi4cxp/kai_matmul_clamp_f16_qai8dxp1x4_qsi4cxp4x4_1x4_neon_dotprod.c:114)	called 58 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_matmul_clamp_f16_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi8cxp/kai_matmul_clamp_f16_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod.c:111)	called 58 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_matmul_clamp_f16_qai8dxp1x8_qsi4cxp4x8_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi4cxp/kai_matmul_clamp_f16_qai8dxp1x8_qsi4cxp4x8_1x4_neon_dotprod.c:114)	called 58 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_matmul_clamp_f16_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi8cxp/kai_matmul_clamp_f16_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod.c:111)	called 58 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_matmul_clamp_f16_qai8dxp4x4_qsi4cxp4x4_16x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi4cxp/kai_matmul_clamp_f16_qai8dxp4x4_qsi4cxp4x4_16x4_neon_dotprod.c:114)	called 222 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_matmul_clamp_f16_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi8cxp/kai_matmul_clamp_f16_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod.c:111)	called 222 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_matmul_clamp_f16_qai8dxp4x8_qsi4cxp4x8_16x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi4cxp/kai_matmul_clamp_f16_qai8dxp4x8_qsi4cxp4x8_16x4_neon_i8mm.c:114)	called 222 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_matmul_clamp_f16_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi8cxp/kai_matmul_clamp_f16_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm.c:111)	called 222 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_matmul_clamp_f16_qsi8d32p1vlx4_qai4c32p4vlx4_1vlx4vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p1vlx4_qai4c32p4vlx4_1vlx4vl_sme2_mopa.c:132)	called 1032 times	100.0%	-%	53.0%
kai_get_rhs_packed_offset_matmul_clamp_f16_qsi8d32p1x4_qai4c32p4vlx4_1x4vl_sme2_dot (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p1x4_qai4c32p4vlx4_1x4vl_sme2_dot.c:128)	called 276 times	100.0%	-%	53.0%
kai_get_rhs_packed_offset_matmul_clamp_f16_qsi8d32p1x4_qai4c32p4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p1x4_qai4c32p4x4_1x4_neon_dotprod.c:121)	called 138 times	100.0%	-%	53.0%
kai_get_rhs_packed_offset_matmul_clamp_f16_qsi8d32p1x8_qai4c32p4x8_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p1x8_qai4c32p4x8_1x4_neon_dotprod.c:121)	called 138 times	100.0%	-%	53.0%
kai_get_rhs_packed_offset_matmul_clamp_f16_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod.c:121)	called 516 times	100.0%	-%	53.0%
kai_get_rhs_packed_offset_matmul_clamp_f16_qsi8d32p4x8_qai4c32p4x8_8x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p4x8_qai4c32p4x8_8x4_neon_i8mm.c:121)	called 516 times	100.0%	-%	53.0%
kai_get_rhs_packed_offset_matmul_clamp_f32_bf16p1x4_bf16p12x4b_1x36_neon_dot (kai/ukernels/matmul/matmul_clamp_f32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p1x4_bf16p12x4b_1x36_neon_dot.c:59)	called 92 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa.c:78)	called 46 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_matmul_clamp_f32_bf16p8x4_bf16p12x4b_8x12_neon_mmla (kai/ukernels/matmul/matmul_clamp_f32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p8x4_bf16p12x4b_8x12_neon_mmla.c:58)	called 184 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_matmul_clamp_f32_f32_f32p16vlx1b_1x16vl_sme2_mla (kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p16vlx1b_1x16vl_sme2_mla.c:68)	called 46 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_matmul_clamp_f32_f32_f32p16x1b_6x16_neon_mla (kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p16x1b_6x16_neon_mla.c:66)	called 16 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_matmul_clamp_f32_f32_f32p16x1b_6x16_neon_mla_cortexa55 (kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p16x1b_6x16_neon_mla_cortexa55.c:66)	called 16 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla (kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla.c:68)	called 46 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_matmul_clamp_f32_f32_f32p2vlx1b_1x8vl_sme_mla (kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p2vlx1b_1x8vl_sme_mla.c:68)	called 46 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_matmul_clamp_f32_f32_f32p8x1biasf32_6x8x4_neon_mla (kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p8x1biasf32_6x8x4_neon_mla.c:67)	called 16 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_matmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme_mopa (kai/ukernels/matmul/matmul_clamp_f32_f32p_f32p/kai_matmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme_mopa.c:78)	called 16 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f32_f32p_f32p/kai_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa.c:78)	called 16 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa.c:122)	called 231 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme_mopa (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme_mopa.c:122)	called 231 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_matmul_clamp_f32_qai8dxp1vlx8_qsi4cxp4vlx8_1vlx4vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1vlx8_qsi4cxp4vlx8_1vlx4vl_sme2_mopa.c:92)	called 120 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_matmul_clamp_f32_qai8dxp1x4_qsi4c32p4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x4_qsi4c32p4x4_1x4_neon_dotprod.c:127)	called 108 times	100.0%	-%	53.0%
kai_get_rhs_packed_offset_matmul_clamp_f32_qai8dxp1x4_qsi4c32p8x4_1x8_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x4_qsi4c32p8x4_1x8_neon_dotprod.c:127)	called 108 times	100.0%	-%	53.0%
kai_get_rhs_packed_offset_matmul_clamp_f32_qai8dxp1x4_qsi4cxp4vlx4_1x4vl_sme2_sdot (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi4cxp4vlx4_1x4vl_sme2_sdot.c:88)	called 120 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_matmul_clamp_f32_qai8dxp1x4_qsi4cxp4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi4cxp4x4_1x4_neon_dotprod.c:99)	called 240 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot.c:123)	called 211 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme_dot (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme_dot.c:123)	called 211 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod.c:97)	called 211 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_matmul_clamp_f32_qai8dxp1x8_qsi4c32p4x8_1x4x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x8_qsi4c32p4x8_1x4x32_neon_dotprod.c:127)	called 108 times	100.0%	-%	53.0%
kai_get_rhs_packed_offset_matmul_clamp_f32_qai8dxp1x8_qsi4c32p8x8_1x8_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x8_qsi4c32p8x8_1x8_neon_dotprod.c:127)	called 108 times	100.0%	-%	53.0%
kai_get_rhs_packed_offset_matmul_clamp_f32_qai8dxp1x8_qsi4c32p8x8_1x8x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x8_qsi4c32p8x8_1x8x32_neon_dotprod.c:127)	called 108 times	100.0%	-%	53.0%
kai_get_rhs_packed_offset_matmul_clamp_f32_qai8dxp1x8_qsi4cxp4x8_1x4x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi4cxp4x8_1x4x32_neon_dotprod.c:83)	called 240 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_matmul_clamp_f32_qai8dxp1x8_qsi4cxp8x8_1x8x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi4cxp8x8_1x8x32_neon_dotprod.c:83)	called 240 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_matmul_clamp_f32_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod.c:97)	called 211 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_matmul_clamp_f32_qai8dxp4x4_qsi4c32p4x4_16x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x4_qsi4c32p4x4_16x4_neon_dotprod.c:128)	called 108 times	100.0%	-%	53.0%
kai_get_rhs_packed_offset_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod.c:127)	called 108 times	100.0%	-%	53.0%
kai_get_rhs_packed_offset_matmul_clamp_f32_qai8dxp4x4_qsi4cxp8x4_8x8x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x4_qsi4cxp8x4_8x8x32_neon_dotprod.c:85)	called 240 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_matmul_clamp_f32_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod.c:101)	called 231 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_matmul_clamp_f32_qai8dxp4x8_qsi4c32p4x8_16x4x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p4x8_16x4x32_neon_i8mm.c:128)	called 108 times	100.0%	-%	53.0%
kai_get_rhs_packed_offset_matmul_clamp_f32_qai8dxp4x8_qsi4c32p4x8_8x4x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p4x8_8x4x32_neon_i8mm.c:127)	called 108 times	100.0%	-%	53.0%
kai_get_rhs_packed_offset_matmul_clamp_f32_qai8dxp4x8_qsi4c32p8x8_4x8_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p8x8_4x8_neon_i8mm.c:127)	called 108 times	100.0%	-%	53.0%
kai_get_rhs_packed_offset_matmul_clamp_f32_qai8dxp4x8_qsi4c32p8x8_4x8x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p8x8_4x8x32_neon_i8mm.c:127)	called 108 times	100.0%	-%	53.0%
kai_get_rhs_packed_offset_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x4_16x4x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x4_16x4x32_neon_dotprod.c:85)	called 240 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_4x4x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_4x4x32_neon_i8mm.c:82)	called 240 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_8x4x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_8x4x32_neon_i8mm.c:86)	called 240 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_4x8x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_4x8x32_neon_i8mm.c:86)	called 240 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_8x8x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_8x8x32_neon_i8mm.c:86)	called 240 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_matmul_clamp_f32_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm.c:101)	called 231 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1vlx4_qai4c32p4vlx4_1vlx4vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p1vlx4_qai4c32p4vlx4_1vlx4vl_sme2_mopa.c:132)	called 1032 times	100.0%	-%	53.0%
kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa.c:109)	called 46 times	100.0%	-%	55.0%
kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1x4_qai4c32p4vlx4_1x4vl_sme2_dot (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qai4c32p4vlx4_1x4vl_sme2_dot.c:128)	called 276 times	100.0%	-%	53.0%
kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1x4_qai4c32p4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qai4c32p4x4_1x4_neon_dotprod.c:119)	called 138 times	100.0%	-%	53.0%
kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot.c:110)	called 26 times	100.0%	-%	55.0%
kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod.c:101)	called 26 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1x8_qai4c32p4x8_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p1x8_qai4c32p4x8_1x4_neon_dotprod.c:119)	called 138 times	100.0%	-%	53.0%
kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod.c:90)	called 26 times	100.0%	-%	51.0%
kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod.c:119)	called 516 times	100.0%	-%	53.0%
kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod.c:105)	called 46 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p4x8_qai4c32p4x8_8x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qai4c32p4x8_8x4_neon_i8mm.c:119)	called 516 times	100.0%	-%	53.0%
kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm.c:93)	called 92 times	100.0%	-%	51.0%
kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_8x4x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_8x4x32_neon_i8mm.c:93)	called 46 times	100.0%	-%	51.0%
kai_get_rhs_packed_offset_matmul_clamp_qai8_qai8_qsi8cxp2vlx4sb_1x16vl_sme2_dot (kai/ukernels/matmul/matmul_clamp_qai8_qai8_qsi8cxp/kai_matmul_clamp_qai8_qai8_qsi8cxp2vlx4sb_1x16vl_sme2_dot.c:69)	called 168 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_matmul_clamp_qai8_qai8p2vlx4_qsi8cxp2vlx4sb_2vlx2vl_sme_mopa (kai/ukernels/matmul/matmul_clamp_qai8_qai8p_qsi8cxp/kai_matmul_clamp_qai8_qai8p2vlx4_qsi8cxp2vlx4sb_2vlx2vl_sme_mopa.c:79)	called 333 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_matmul_clamp_qai8_qai8p2vlx4_qsi8cxpsb2vlx4_2vlx2vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_qai8_qai8p_qsi8cxp/kai_matmul_clamp_qai8_qai8p2vlx4_qsi8cxpsb2vlx4_2vlx2vl_sme2_mopa.c:79)	called 333 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_rhs_imatmul_pack_kxn_qsi8cxp2vlx4sb_qs8cx_f32_i32_sme (kai/ukernels/matmul/pack/kai_rhs_imatmul_pack_kxn_qsi8cxp2vlx4sb_qs8cx_f32_i32_sme.c:71)	called 18648 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_rhs_imatmul_pack_kxn_x16p2vlx2b_x16_x16_sme (kai/ukernels/matmul/pack/kai_rhs_imatmul_pack_kxn_x16p2vlx2b_x16_x16_sme.c:62)	called 22704 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_rhs_imatmul_pack_kxn_x32p2vlx1b_x32_x32_sme (kai/ukernels/matmul/pack/kai_rhs_imatmul_pack_kxn_x32p2vlx1b_x32_x32_sme.c:59)	called 22704 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_rhs_pack_kxn_bf16p12x4biasf16_f16_neon (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_bf16p12x4biasf16_f16_neon.c:47)	called 102 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_rhs_pack_kxn_bf16p12x4biasf32_f16_neon (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_bf16p12x4biasf32_f16_neon.c:41)	called 112 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme.c:44)	called 56 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_rhs_pack_kxn_f16p16x1biasf16_f16_f16_neon (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_f16p16x1biasf16_f16_f16_neon.c:35)	called 34 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_rhs_pack_kxn_f32p16vlx1b_f32_f32_sme (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_f32p16vlx1b_f32_f32_sme.c:57)	called 92 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_rhs_pack_kxn_f32p2vlx1biasf32_f32_f32_sme (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_f32p2vlx1biasf32_f32_f32_sme.c:57)	called 252 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_rhs_pack_kxn_f32p8x1biasf32_f32_f32_neon (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_f32p8x1biasf32_f32_f32_neon.c:35)	called 34 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_rhs_pack_kxn_qsi4c32p_qsu4c32s1s0 (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_qsi4c32p_qsu4c32s1s0.c:76)	called 702 times	100.0%	-%	51.0%
kai_get_rhs_packed_offset_rhs_pack_kxn_qsi4cxp_qs4cxs1s0 (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_qsi4cxp_qs4cxs1s0.c:45)	called 1092 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_rhs_pack_kxn_qsi8cxp2vlx4sb_qs8cx_f32_i32_sme (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_qsi8cxp2vlx4sb_qs8cx_f32_i32_sme.c:68)	called 1842 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_rhs_pack_kxn_qsi8cxp_qsi8cx_neon (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_qsi8cxp_qsi8cx_neon.c:41)	called 1192 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_rhs_pack_kxn_x16p2vlx2b_x16_x16_sme (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_x16p2vlx2b_x16_x16_sme.c:60)	called 252 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_rhs_pack_kxn_x16p32x1b_x16_x16_neon (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_x16p32x1b_x16_x16_neon.c:58)	called 68 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_rhs_pack_kxn_x32p16x1b_x32_x32_neon (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_x32p16x1b_x32_x32_neon.c:58)	called 68 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_rhs_pack_nxk_f32p2vlx1biasf32_f32_f32_sme (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_f32p2vlx1biasf32_f32_f32_sme.c:55)	called 34 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_rhs_pack_nxk_qai4c32p_qau4c32s0s1_f32_f32_f32_neon (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qai4c32p_qau4c32s0s1_f32_f32_f32_neon.c:48)	called 2616 times	100.0%	-%	52.0%
kai_get_rhs_packed_offset_rhs_pack_nxk_qai4c32ps1s0nrx4_qau4c32s0s1_f32_f32_f32_neon (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qai4c32ps1s0nrx4_qau4c32s0s1_f32_f32_f32_neon.c:50)	called 1308 times	100.0%	-%	52.0%
kai_get_rhs_packed_offset_rhs_pack_nxk_qai4c32ps1s0nrx4_qau4c32s1s0_f32_f32_f32_neon (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qai4c32ps1s0nrx4_qau4c32s1s0_f32_f32_f32_neon.c:50)	called 1308 times	100.0%	-%	52.0%
kai_get_rhs_packed_offset_rhs_pack_nxk_qsi4c32p_qsu4c32s1s0 (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32p_qsu4c32s1s0.c:74)	called 702 times	100.0%	-%	51.0%
kai_get_rhs_packed_offset_rhs_pack_nxk_qsi4c32pnrx4_qsu4c32s1s0_neon (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32pnrx4_qsu4c32s1s0_neon.c:83)	called 216 times	100.0%	-%	51.0%
kai_get_rhs_packed_offset_rhs_pack_nxk_qsi4c32pnrx8_qsu4c32s1s0_neon (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32pnrx8_qsu4c32s1s0_neon.c:83)	called 486 times	100.0%	-%	51.0%
kai_get_rhs_packed_offset_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon.c:75)	called 72 times	100.0%	-%	51.0%
kai_get_rhs_packed_offset_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0 (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0.c:55)	called 236 times	100.0%	-%	51.0%
kai_get_rhs_packed_offset_rhs_pack_nxk_qsi4cxp_qs4cxs1s0 (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4cxp_qs4cxs1s0.c:42)	called 2372 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_rhs_pack_nxk_qsi4cxps1s0_qsu4cxs1s0_neon (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4cxps1s0_qsu4cxs1s0_neon.c:47)	called 240 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_rhs_pack_nxk_qsi8cxp_qsi8cx_neon (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi8cxp_qsi8cx_neon.c:40)	called 1752 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_rhs_pack_nxk_x16p2vlx2b_x16_x16_sme (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_x16p2vlx2b_x16_x16_sme.c:55)	called 34 times	100.0%	-%	57.0%
kai_get_rhs_packed_offset_rhs_quant_pack_kxn_bf16p12x4biasf32_f32_neon (kai/ukernels/matmul/pack/kai_rhs_quant_pack_kxn_bf16p12x4biasf32_f32_neon.c:43)	called 228 times	100.0%	-%	52.0%
kai_get_rhs_packed_size_rhs_imatmul_pack_kxn_qsi8cxp2vlx4sb_qs8cx_f32_i32_sme (kai/ukernels/matmul/pack/kai_rhs_imatmul_pack_kxn_qsi8cxp2vlx4sb_qs8cx_f32_i32_sme.c:80)	called 9324 times	100.0%	-%	100.0%
kai_get_rhs_packed_size_rhs_imatmul_pack_kxn_x16p2vlx2b_x16_x16_sme (kai/ukernels/matmul/pack/kai_rhs_imatmul_pack_kxn_x16p2vlx2b_x16_x16_sme.c:71)	called 11352 times	100.0%	-%	100.0%
kai_get_rhs_packed_size_rhs_imatmul_pack_kxn_x32p2vlx1b_x32_x32_sme (kai/ukernels/matmul/pack/kai_rhs_imatmul_pack_kxn_x32p2vlx1b_x32_x32_sme.c:68)	called 11352 times	100.0%	-%	100.0%
kai_get_rhs_packed_size_rhs_pack_kxn_bf16p12x4biasf16_f16_neon (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_bf16p12x4biasf16_f16_neon.c:53)	called 102 times	100.0%	-%	100.0%
kai_get_rhs_packed_size_rhs_pack_kxn_bf16p12x4biasf32_f16_neon (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_bf16p12x4biasf32_f16_neon.c:47)	called 112 times	100.0%	-%	100.0%
kai_get_rhs_packed_size_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme.c:51)	called 56 times	100.0%	-%	100.0%
kai_get_rhs_packed_size_rhs_pack_kxn_f16p16x1biasf16_f16_f16_neon (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_f16p16x1biasf16_f16_f16_neon.c:41)	called 17 times	100.0%	-%	100.0%
kai_get_rhs_packed_size_rhs_pack_kxn_f32p16vlx1b_f32_f32_sme (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_f32p16vlx1b_f32_f32_sme.c:64)	called 46 times	100.0%	-%	100.0%
kai_get_rhs_packed_size_rhs_pack_kxn_f32p2vlx1biasf32_f32_f32_sme (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_f32p2vlx1biasf32_f32_f32_sme.c:64)	called 126 times	100.0%	-%	100.0%
kai_get_rhs_packed_size_rhs_pack_kxn_f32p8x1biasf32_f32_f32_neon (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_f32p8x1biasf32_f32_f32_neon.c:41)	called 17 times	100.0%	-%	100.0%
kai_get_rhs_packed_size_rhs_pack_kxn_qsi4c32p_qsu4c32s1s0 (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_qsi4c32p_qsu4c32s1s0.c:94)	called 702 times	100.0%	-%	51.0%
kai_get_rhs_packed_size_rhs_pack_kxn_qsi4cxp_qs4cxs1s0 (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_qsi4cxp_qs4cxs1s0.c:52)	called 720 times	100.0%	-%	100.0%
kai_get_rhs_packed_size_rhs_pack_kxn_qsi8cxp2vlx4sb_qs8cx_f32_i32_sme (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_qsi8cxp2vlx4sb_qs8cx_f32_i32_sme.c:75)	called 834 times	100.0%	-%	100.0%
kai_get_rhs_packed_size_rhs_pack_kxn_qsi8cxp_qsi8cx_neon (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_qsi8cxp_qsi8cx_neon.c:48)	called 576 times	100.0%	-%	100.0%
kai_get_rhs_packed_size_rhs_pack_kxn_x16p2vlx2b_x16_x16_sme (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_x16p2vlx2b_x16_x16_sme.c:67)	called 126 times	100.0%	-%	100.0%
kai_get_rhs_packed_size_rhs_pack_kxn_x16p32x1b_x16_x16_neon (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_x16p32x1b_x16_x16_neon.c:65)	called 34 times	100.0%	-%	100.0%
kai_get_rhs_packed_size_rhs_pack_kxn_x32p16x1b_x32_x32_neon (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_x32p16x1b_x32_x32_neon.c:65)	called 34 times	100.0%	-%	100.0%
kai_get_rhs_packed_size_rhs_pack_nxk_f32p2vlx1biasf32_f32_f32_sme (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_f32p2vlx1biasf32_f32_f32_sme.c:61)	called 34 times	100.0%	-%	100.0%
kai_get_rhs_packed_size_rhs_pack_nxk_qai4c32p_qau4c32s0s1_f32_f32_f32_neon (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qai4c32p_qau4c32s0s1_f32_f32_f32_neon.c:58)	called 2616 times	100.0%	-%	52.0%
kai_get_rhs_packed_size_rhs_pack_nxk_qai4c32ps1s0nrx4_qau4c32s0s1_f32_f32_f32_neon (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qai4c32ps1s0nrx4_qau4c32s0s1_f32_f32_f32_neon.c:60)	called 1308 times	100.0%	-%	52.0%
kai_get_rhs_packed_size_rhs_pack_nxk_qai4c32ps1s0nrx4_qau4c32s1s0_f32_f32_f32_neon (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qai4c32ps1s0nrx4_qau4c32s1s0_f32_f32_f32_neon.c:60)	called 1308 times	100.0%	-%	52.0%
kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32p_qsu4c32s1s0 (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32p_qsu4c32s1s0.c:92)	called 702 times	100.0%	-%	51.0%
kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pnrx4_qsu4c32s1s0_neon (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32pnrx4_qsu4c32s1s0_neon.c:102)	called 216 times	100.0%	-%	51.0%
kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pnrx8_qsu4c32s1s0_neon (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32pnrx8_qsu4c32s1s0_neon.c:102)	called 486 times	100.0%	-%	51.0%
kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon.c:87)	called 26 times	100.0%	-%	52.0%
kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0 (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0.c:66)	called 98 times	100.0%	-%	52.0%
kai_get_rhs_packed_size_rhs_pack_nxk_qsi4cxp_qs4cxs1s0 (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4cxp_qs4cxs1s0.c:49)	called 2024 times	100.0%	-%	100.0%
kai_get_rhs_packed_size_rhs_pack_nxk_qsi4cxps1s0_qsu4cxs1s0_neon (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4cxps1s0_qsu4cxs1s0_neon.c:54)	called 160 times	100.0%	-%	100.0%
kai_get_rhs_packed_size_rhs_pack_nxk_qsi8cxp_qsi8cx_neon (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi8cxp_qsi8cx_neon.c:47)	called 1136 times	100.0%	-%	100.0%
kai_get_rhs_packed_size_rhs_pack_nxk_x16p2vlx2b_x16_x16_sme (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_x16p2vlx2b_x16_x16_sme.c:61)	called 34 times	100.0%	-%	100.0%
kai_get_rhs_packed_size_rhs_quant_pack_kxn_bf16p12x4biasf32_f32_neon (kai/ukernels/matmul/pack/kai_rhs_quant_pack_kxn_bf16p12x4biasf32_f32_neon.c:52)	called 228 times	100.0%	-%	100.0%
kai_get_rhs_packed_stride (kai/ukernels/matmul/matmul_clamp_bf16_qai8dxp_qsi4c32p/kai_matmul_clamp_bf16_qai8dxp1x8_qsi4c32p4x8_1x4_neon_dotprod.c:82)	called 108 times	100.0%	-%	57.0%
kai_get_rhs_packed_stride (kai/ukernels/matmul/matmul_clamp_bf16_qai8dxp_qsi4c32p/kai_matmul_clamp_bf16_qai8dxp4x8_qsi4c32p4x8_16x4_neon_i8mm.c:83)	called 108 times	100.0%	-%	57.0%
kai_get_rhs_packed_stride (kai/ukernels/matmul/matmul_clamp_bf16_qai8dxp_qsi4cxp/kai_matmul_clamp_bf16_qai8dxp1x8_qsi4cxp8x8_1x8_neon_dotprod.c:70)	called 356 times	100.0%	-%	100.0%
kai_get_rhs_packed_stride (kai/ukernels/matmul/matmul_clamp_bf16_qai8dxp_qsi4cxp/kai_matmul_clamp_bf16_qai8dxp4x8_qsi4cxp8x8_8x8_neon_i8mm.c:71)	called 388 times	100.0%	-%	100.0%
kai_get_rhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi4cxp/kai_matmul_clamp_f16_qai8dxp1x4_qsi4cxp4x4_1x4_neon_dotprod.c:70)	called 58 times	100.0%	-%	100.0%
kai_get_rhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi4cxp/kai_matmul_clamp_f16_qai8dxp1x8_qsi4cxp4x8_1x4_neon_dotprod.c:70)	called 58 times	100.0%	-%	100.0%
kai_get_rhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi4cxp/kai_matmul_clamp_f16_qai8dxp4x4_qsi4cxp4x4_16x4_neon_dotprod.c:70)	called 222 times	100.0%	-%	100.0%
kai_get_rhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi4cxp/kai_matmul_clamp_f16_qai8dxp4x8_qsi4cxp4x8_16x4_neon_i8mm.c:70)	called 222 times	100.0%	-%	100.0%
kai_get_rhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi8cxp/kai_matmul_clamp_f16_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod.c:67)	called 58 times	100.0%	-%	100.0%
kai_get_rhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi8cxp/kai_matmul_clamp_f16_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod.c:67)	called 58 times	100.0%	-%	100.0%
kai_get_rhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi8cxp/kai_matmul_clamp_f16_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod.c:67)	called 222 times	100.0%	-%	100.0%
kai_get_rhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi8cxp/kai_matmul_clamp_f16_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm.c:67)	called 222 times	100.0%	-%	100.0%
kai_get_rhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p1vlx4_qai4c32p4vlx4_1vlx4vl_sme2_mopa.c:83)	called 2065 times	100.0%	-%	53.0%
kai_get_rhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p1x4_qai4c32p4vlx4_1x4vl_sme2_dot.c:79)	called 553 times	100.0%	-%	53.0%
kai_get_rhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p1x4_qai4c32p4x4_1x4_neon_dotprod.c:77)	called 138 times	100.0%	-%	57.0%
kai_get_rhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p1x8_qai4c32p4x8_1x4_neon_dotprod.c:77)	called 138 times	100.0%	-%	57.0%
kai_get_rhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod.c:77)	called 516 times	100.0%	-%	57.0%
kai_get_rhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p4x8_qai4c32p4x8_8x4_neon_i8mm.c:77)	called 516 times	100.0%	-%	57.0%
kai_get_rhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x4_qsi4c32p4x4_1x4_neon_dotprod.c:81)	called 108 times	100.0%	-%	57.0%
kai_get_rhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x4_qsi4c32p8x4_1x8_neon_dotprod.c:81)	called 108 times	100.0%	-%	57.0%
kai_get_rhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x8_qsi4c32p4x8_1x4x32_neon_dotprod.c:81)	called 108 times	100.0%	-%	57.0%
kai_get_rhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x8_qsi4c32p8x8_1x8_neon_dotprod.c:81)	called 108 times	100.0%	-%	57.0%
kai_get_rhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x8_qsi4c32p8x8_1x8x32_neon_dotprod.c:81)	called 108 times	100.0%	-%	57.0%
kai_get_rhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x4_qsi4c32p4x4_16x4_neon_dotprod.c:82)	called 108 times	100.0%	-%	57.0%
kai_get_rhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod.c:81)	called 108 times	100.0%	-%	57.0%
kai_get_rhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p4x8_16x4x32_neon_i8mm.c:82)	called 108 times	100.0%	-%	57.0%
kai_get_rhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p4x8_8x4x32_neon_i8mm.c:81)	called 108 times	100.0%	-%	57.0%
kai_get_rhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p8x8_4x8_neon_i8mm.c:81)	called 108 times	100.0%	-%	57.0%
kai_get_rhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p8x8_4x8x32_neon_i8mm.c:81)	called 108 times	100.0%	-%	57.0%
kai_get_rhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1vlx8_qsi4cxp4vlx8_1vlx4vl_sme2_mopa.c:51)	called 201 times	100.0%	-%	55.0%
kai_get_rhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi4cxp4vlx4_1x4vl_sme2_sdot.c:48)	called 201 times	100.0%	-%	55.0%
kai_get_rhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi4cxp4x4_1x4_neon_dotprod.c:55)	called 240 times	100.0%	-%	100.0%
kai_get_rhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa.c:75)	called 386 times	100.0%	-%	55.0%
kai_get_rhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme_mopa.c:75)	called 386 times	100.0%	-%	55.0%
kai_get_rhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot.c:76)	called 346 times	100.0%	-%	55.0%
kai_get_rhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme_dot.c:76)	called 346 times	100.0%	-%	55.0%
kai_get_rhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p1vlx4_qai4c32p4vlx4_1vlx4vl_sme2_mopa.c:83)	called 2065 times	100.0%	-%	53.0%
kai_get_rhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qai4c32p4vlx4_1x4vl_sme2_dot.c:79)	called 553 times	100.0%	-%	53.0%
kai_get_rhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qai4c32p4x4_1x4_neon_dotprod.c:75)	called 138 times	100.0%	-%	57.0%
kai_get_rhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p1x8_qai4c32p4x8_1x4_neon_dotprod.c:75)	called 138 times	100.0%	-%	57.0%
kai_get_rhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod.c:75)	called 516 times	100.0%	-%	57.0%
kai_get_rhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qai4c32p4x8_8x4_neon_i8mm.c:75)	called 516 times	100.0%	-%	57.0%
kai_get_rhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa.c:63)	called 70 times	100.0%	-%	53.0%
kai_get_rhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot.c:63)	called 30 times	100.0%	-%	53.0%
kai_get_rhs_packed_stride (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qai4c32p_qau4c32s0s1_f32_f32_f32_neon.c:33)	called 7848 times	100.0%	-%	51.0%
kai_get_rhs_packed_stride (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qai4c32ps1s0nrx4_qau4c32s0s1_f32_f32_f32_neon.c:35)	called 3924 times	100.0%	-%	51.0%
kai_get_rhs_packed_stride (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qai4c32ps1s0nrx4_qau4c32s1s0_f32_f32_f32_neon.c:35)	called 3924 times	100.0%	-%	51.0%
kai_get_rhs_packed_stride_imatmul_clamp_f16_f16p2vlx2_f16p2vlx2_2vlx2vl_sme2_mopa (kai/ukernels/matmul/imatmul_clamp_f16_f16p_f16p/kai_imatmul_clamp_f16_f16p2vlx2_f16p2vlx2_2vlx2vl_sme2_mopa.c:58)	called 5676 times	100.0%	-%	100.0%
kai_get_rhs_packed_stride_imatmul_clamp_f16_f16p2vlx2_f16p2vlx2b_2vlx2vl_sme_mopa (kai/ukernels/matmul/imatmul_clamp_f16_f16p_f16p/kai_imatmul_clamp_f16_f16p2vlx2_f16p2vlx2b_2vlx2vl_sme_mopa.c:58)	called 5676 times	100.0%	-%	100.0%
kai_get_rhs_packed_stride_imatmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme2_mopa (kai/ukernels/matmul/imatmul_clamp_f32_f32p_f32p/kai_imatmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme2_mopa.c:57)	called 5676 times	100.0%	-%	100.0%
kai_get_rhs_packed_stride_imatmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme_mopa (kai/ukernels/matmul/imatmul_clamp_f32_f32p_f32p/kai_imatmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme_mopa.c:57)	called 5676 times	100.0%	-%	100.0%
kai_get_rhs_packed_stride_imatmul_clamp_qai8_qai8p2vlx4_qsi8cxp2vlx4sb_2vlx2vl_sme_mopa (kai/ukernels/matmul/imatmul_clamp_qai8_qai8p_qsi8cxp/kai_imatmul_clamp_qai8_qai8p2vlx4_qsi8cxp2vlx4sb_2vlx2vl_sme_mopa.c:58)	called 4662 times	100.0%	-%	100.0%
kai_get_rhs_packed_stride_imatmul_clamp_qai8_qai8p2vlx4_qsi8cxpsb2vlx4_2vlx2vl_sme2_mopa (kai/ukernels/matmul/imatmul_clamp_qai8_qai8p_qsi8cxp/kai_imatmul_clamp_qai8_qai8p2vlx4_qsi8cxpsb2vlx4_2vlx2vl_sme2_mopa.c:58)	called 4662 times	100.0%	-%	100.0%
kai_get_rhs_packed_stride_matmul_clamp_f16_f16_f16p2vlx2b_1x16vl_sme2_dot (kai/ukernels/matmul/matmul_clamp_f16_f16_f16p/kai_matmul_clamp_f16_f16_f16p2vlx2b_1x16vl_sme2_dot.c:65)	called 46 times	100.0%	-%	100.0%
kai_get_rhs_packed_stride_matmul_clamp_f16_f16_f16p2vlx2b_1x8vl_sme_mla (kai/ukernels/matmul/matmul_clamp_f16_f16_f16p/kai_matmul_clamp_f16_f16_f16p2vlx2b_1x8vl_sme_mla.c:65)	called 46 times	100.0%	-%	100.0%
kai_get_rhs_packed_stride_matmul_clamp_f16_f16p2vlx2_f16p2vlx2_2vlx2vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f16_f16p_f16p/kai_matmul_clamp_f16_f16p2vlx2_f16p2vlx2_2vlx2vl_sme2_mopa.c:74)	called 16 times	100.0%	-%	100.0%
kai_get_rhs_packed_stride_matmul_clamp_f16_f16p2vlx2_f16p2vlx2b_2vlx2vl_sme_mopa (kai/ukernels/matmul/matmul_clamp_f16_f16p_f16p/kai_matmul_clamp_f16_f16p2vlx2_f16p2vlx2b_2vlx2vl_sme_mopa.c:74)	called 16 times	100.0%	-%	100.0%
kai_get_rhs_packed_stride_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa.c:73)	called 46 times	100.0%	-%	100.0%
kai_get_rhs_packed_stride_matmul_clamp_f32_f32_f32p16vlx1b_1x16vl_sme2_mla (kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p16vlx1b_1x16vl_sme2_mla.c:63)	called 46 times	100.0%	-%	100.0%
kai_get_rhs_packed_stride_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla (kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla.c:63)	called 46 times	100.0%	-%	100.0%
kai_get_rhs_packed_stride_matmul_clamp_f32_f32_f32p2vlx1b_1x8vl_sme_mla (kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p2vlx1b_1x8vl_sme_mla.c:63)	called 46 times	100.0%	-%	100.0%
kai_get_rhs_packed_stride_matmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme_mopa (kai/ukernels/matmul/matmul_clamp_f32_f32p_f32p/kai_matmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme_mopa.c:73)	called 16 times	100.0%	-%	100.0%
kai_get_rhs_packed_stride_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f32_f32p_f32p/kai_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa.c:73)	called 16 times	100.0%	-%	100.0%
kai_get_rhs_packed_stride_matmul_clamp_qai8_qai8_qsi8cxp2vlx4sb_1x16vl_sme2_dot (kai/ukernels/matmul/matmul_clamp_qai8_qai8_qsi8cxp/kai_matmul_clamp_qai8_qai8_qsi8cxp2vlx4sb_1x16vl_sme2_dot.c:64)	called 168 times	100.0%	-%	100.0%
kai_get_rhs_packed_stride_matmul_clamp_qai8_qai8p2vlx4_qsi8cxp2vlx4sb_2vlx2vl_sme_mopa (kai/ukernels/matmul/matmul_clamp_qai8_qai8p_qsi8cxp/kai_matmul_clamp_qai8_qai8p2vlx4_qsi8cxp2vlx4sb_2vlx2vl_sme_mopa.c:74)	called 333 times	100.0%	-%	100.0%
kai_get_rhs_packed_stride_matmul_clamp_qai8_qai8p2vlx4_qsi8cxpsb2vlx4_2vlx2vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_qai8_qai8p_qsi8cxp/kai_matmul_clamp_qai8_qai8p2vlx4_qsi8cxpsb2vlx4_2vlx2vl_sme2_mopa.c:74)	called 333 times	100.0%	-%	100.0%
kai_get_rhs_packed_stride_rhs_imatmul_pack_kxn_qsi8cxp2vlx4sb_qs8cx_f32_i32_sme (kai/ukernels/matmul/pack/kai_rhs_imatmul_pack_kxn_qsi8cxp2vlx4sb_qs8cx_f32_i32_sme.c:64)	called 27972 times	100.0%	-%	100.0%
kai_get_rhs_packed_stride_rhs_imatmul_pack_kxn_x16p2vlx2b_x16_x16_sme (kai/ukernels/matmul/pack/kai_rhs_imatmul_pack_kxn_x16p2vlx2b_x16_x16_sme.c:56)	called 34056 times	100.0%	-%	100.0%
kai_get_rhs_packed_stride_rhs_imatmul_pack_kxn_x32p2vlx1b_x32_x32_sme (kai/ukernels/matmul/pack/kai_rhs_imatmul_pack_kxn_x32p2vlx1b_x32_x32_sme.c:53)	called 34056 times	100.0%	-%	100.0%
kai_get_rhs_packed_stride_rhs_pack_kxn_bf16p12x4biasf16_f16_neon (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_bf16p12x4biasf16_f16_neon.c:43)	called 102 times	100.0%	-%	100.0%
kai_get_rhs_packed_stride_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme.c:39)	called 112 times	100.0%	-%	100.0%
kai_get_rhs_packed_stride_rhs_pack_kxn_f32p16vlx1b_f32_f32_sme (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_f32p16vlx1b_f32_f32_sme.c:52)	called 138 times	100.0%	-%	100.0%
kai_get_rhs_packed_stride_rhs_pack_kxn_f32p2vlx1biasf32_f32_f32_sme (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_f32p2vlx1biasf32_f32_f32_sme.c:52)	called 378 times	100.0%	-%	100.0%
kai_get_rhs_packed_stride_rhs_pack_kxn_qsi4c32p_qsu4c32s1s0 (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_qsi4c32p_qsu4c32s1s0.c:53)	called 2106 times	100.0%	-%	51.0%
kai_get_rhs_packed_stride_rhs_pack_kxn_qsi4cxp_qs4cxs1s0 (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_qsi4cxp_qs4cxs1s0.c:34)	called 2904 times	100.0%	-%	55.0%
kai_get_rhs_packed_stride_rhs_pack_kxn_qsi8cxp2vlx4sb_qs8cx_f32_i32_sme (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_qsi8cxp2vlx4sb_qs8cx_f32_i32_sme.c:63)	called 2676 times	100.0%	-%	100.0%
kai_get_rhs_packed_stride_rhs_pack_kxn_qsi8cxp_qsi8cx_neon (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_qsi8cxp_qsi8cx_neon.c:33)	called 2344 times	100.0%	-%	100.0%
kai_get_rhs_packed_stride_rhs_pack_kxn_x16p2vlx2b_x16_x16_sme (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_x16p2vlx2b_x16_x16_sme.c:55)	called 378 times	100.0%	-%	100.0%
kai_get_rhs_packed_stride_rhs_pack_kxn_x16p32x1b_x16_x16_neon (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_x16p32x1b_x16_x16_neon.c:53)	called 102 times	100.0%	-%	100.0%
kai_get_rhs_packed_stride_rhs_pack_kxn_x32p16x1b_x32_x32_neon (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_x32p16x1b_x32_x32_neon.c:53)	called 102 times	100.0%	-%	100.0%
kai_get_rhs_packed_stride_rhs_pack_nxk_f32p2vlx1biasf32_f32_f32_sme (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_f32p2vlx1biasf32_f32_f32_sme.c:51)	called 68 times	100.0%	-%	100.0%
kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32p_qsu4c32s1s0 (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32p_qsu4c32s1s0.c:51)	called 1404 times	100.0%	-%	51.0%
kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pnrx4_qsu4c32s1s0_neon (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32pnrx4_qsu4c32s1s0_neon.c:60)	called 432 times	100.0%	-%	51.0%
kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pnrx8_qsu4c32s1s0_neon (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32pnrx8_qsu4c32s1s0_neon.c:60)	called 972 times	100.0%	-%	51.0%
kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon.c:58)	called 124 times	100.0%	-%	52.0%
kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0 (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0.c:39)	called 432 times	100.0%	-%	52.0%
kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4cxp_qs4cxs1s0 (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4cxp_qs4cxs1s0.c:32)	called 6048 times	100.0%	-%	55.0%
kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4cxps1s0_qsu4cxs1s0_neon (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4cxps1s0_qsu4cxs1s0_neon.c:35)	called 560 times	100.0%	-%	55.0%
kai_get_rhs_packed_stride_rhs_pack_nxk_qsi8cxp_qsi8cx_neon (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi8cxp_qsi8cx_neon.c:32)	called 4024 times	100.0%	-%	100.0%
kai_get_rhs_packed_stride_rhs_pack_nxk_x16p2vlx2b_x16_x16_sme (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_x16p2vlx2b_x16_x16_sme.c:51)	called 68 times	100.0%	-%	100.0%
kai_get_scale_offset_rhs_imatmul_pack_kxn_qsi8cxp2vlx4sb_qs8cx_f32_i32_sme (kai/ukernels/matmul/pack/kai_rhs_imatmul_pack_kxn_qsi8cxp2vlx4sb_qs8cx_f32_i32_sme.c:60)	called 9324 times	100.0%	-%	100.0%
kai_get_scale_offset_rhs_pack_kxn_qsi8cxp2vlx4sb_qs8cx_f32_i32_sme (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_qsi8cxp2vlx4sb_qs8cx_f32_i32_sme.c:59)	called 834 times	100.0%	-%	100.0%
kai_get_sme_vector_length_u16 (kai/kai_common.h:151)	called 230550 times	100.0%	-%	100.0%
kai_get_sme_vector_length_u32 (kai/kai_common.h:156)	called 273087 times	100.0%	-%	100.0%
kai_get_sme_vector_length_u32() (kai/kai_common.h:156)	called 249 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_bf16_qai8dxp1x8_qsi4c32p4x8_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_bf16_qai8dxp_qsi4c32p/kai_matmul_clamp_bf16_qai8dxp1x8_qsi4c32p4x8_1x4_neon_dotprod.c:118)	called 112 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_bf16_qai8dxp1x8_qsi4cxp8x8_1x8_neon_dotprod (kai/ukernels/matmul/matmul_clamp_bf16_qai8dxp_qsi4cxp/kai_matmul_clamp_bf16_qai8dxp1x8_qsi4cxp8x8_1x8_neon_dotprod.c:104)	called 392 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_bf16_qai8dxp4x8_qsi4c32p4x8_16x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_bf16_qai8dxp_qsi4c32p/kai_matmul_clamp_bf16_qai8dxp4x8_qsi4c32p4x8_16x4_neon_i8mm.c:119)	called 112 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_bf16_qai8dxp4x8_qsi4cxp8x8_8x8_neon_i8mm (kai/ukernels/matmul/matmul_clamp_bf16_qai8dxp_qsi4cxp/kai_matmul_clamp_bf16_qai8dxp4x8_qsi4cxp8x8_8x8_neon_i8mm.c:105)	called 392 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f16_bf16p8x4_bf16p12x4b_8x12_neon_mmla (kai/ukernels/matmul/matmul_clamp_f16_bf16p_bf16p/kai_matmul_clamp_f16_bf16p8x4_bf16p12x4b_8x12_neon_mmla.c:47)	called 18 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f16_f16_f16p16x1biasf16_6x16x8_neon_mla (kai/ukernels/matmul/matmul_clamp_f16_f16_f16p/kai_matmul_clamp_f16_f16_f16p16x1biasf16_6x16x8_neon_mla.c:44)	called 17 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f16_f16_f16p2vlx2b_1x16vl_sme2_dot (kai/ukernels/matmul/matmul_clamp_f16_f16_f16p/kai_matmul_clamp_f16_f16_f16p2vlx2b_1x16vl_sme2_dot.c:55)	called 46 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f16_f16_f16p2vlx2b_1x8vl_sme_mla (kai/ukernels/matmul/matmul_clamp_f16_f16_f16p/kai_matmul_clamp_f16_f16_f16p2vlx2b_1x8vl_sme_mla.c:55)	called 46 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f16_f16_f16p32x1b_6x32_neon_mla (kai/ukernels/matmul/matmul_clamp_f16_f16_f16p/kai_matmul_clamp_f16_f16_f16p32x1b_6x32_neon_mla.c:60)	called 17 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f16_f16_f16p32x1b_6x32_neon_mla_cortexa55 (kai/ukernels/matmul/matmul_clamp_f16_f16_f16p/kai_matmul_clamp_f16_f16_f16p32x1b_6x32_neon_mla_cortexa55.c:60)	called 17 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f16_f16p2vlx2_f16p2vlx2_2vlx2vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f16_f16p_f16p/kai_matmul_clamp_f16_f16p2vlx2_f16p2vlx2_2vlx2vl_sme2_mopa.c:65)	called 51 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f16_f16p2vlx2_f16p2vlx2b_2vlx2vl_sme_mopa (kai/ukernels/matmul/matmul_clamp_f16_f16p_f16p/kai_matmul_clamp_f16_f16p2vlx2_f16p2vlx2b_2vlx2vl_sme_mopa.c:65)	called 51 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f16_qai8dxp1x4_qsi4cxp4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi4cxp/kai_matmul_clamp_f16_qai8dxp1x4_qsi4cxp4x4_1x4_neon_dotprod.c:104)	called 224 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f16_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi8cxp/kai_matmul_clamp_f16_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod.c:101)	called 224 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f16_qai8dxp1x8_qsi4cxp4x8_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi4cxp/kai_matmul_clamp_f16_qai8dxp1x8_qsi4cxp4x8_1x4_neon_dotprod.c:104)	called 224 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f16_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi8cxp/kai_matmul_clamp_f16_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod.c:101)	called 224 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f16_qai8dxp4x4_qsi4cxp4x4_16x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi4cxp/kai_matmul_clamp_f16_qai8dxp4x4_qsi4cxp4x4_16x4_neon_dotprod.c:104)	called 224 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f16_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi8cxp/kai_matmul_clamp_f16_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod.c:101)	called 224 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f16_qai8dxp4x8_qsi4cxp4x8_16x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi4cxp/kai_matmul_clamp_f16_qai8dxp4x8_qsi4cxp4x8_16x4_neon_i8mm.c:104)	called 224 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f16_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi8cxp/kai_matmul_clamp_f16_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm.c:101)	called 224 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f16_qsi8d32p1vlx4_qai4c32p4vlx4_1vlx4vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p1vlx4_qai4c32p4vlx4_1vlx4vl_sme2_mopa.c:118)	called 2072 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f16_qsi8d32p1x4_qai4c32p4vlx4_1x4vl_sme2_dot (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p1x4_qai4c32p4vlx4_1x4vl_sme2_dot.c:114)	called 2072 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f16_qsi8d32p1x4_qai4c32p4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p1x4_qai4c32p4x4_1x4_neon_dotprod.c:110)	called 1036 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f16_qsi8d32p1x8_qai4c32p4x8_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p1x8_qai4c32p4x8_1x4_neon_dotprod.c:110)	called 1036 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f16_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod.c:110)	called 1036 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f16_qsi8d32p4x8_qai4c32p4x8_8x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p4x8_qai4c32p4x8_8x4_neon_i8mm.c:110)	called 1036 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f32_bf16p1x4_bf16p12x4b_1x36_neon_dot (kai/ukernels/matmul/matmul_clamp_f32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p1x4_bf16p12x4b_1x36_neon_dot.c:49)	called 24 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa.c:64)	called 10 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f32_bf16p8x4_bf16p12x4b_8x12_neon_mmla (kai/ukernels/matmul/matmul_clamp_f32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p8x4_bf16p12x4b_8x12_neon_mmla.c:48)	called 40 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f32_f32_f32p16vlx1b_1x16vl_sme2_mla (kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p16vlx1b_1x16vl_sme2_mla.c:53)	called 46 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f32_f32_f32p16x1b_6x16_neon_mla (kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p16x1b_6x16_neon_mla.c:56)	called 17 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f32_f32_f32p16x1b_6x16_neon_mla_cortexa55 (kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p16x1b_6x16_neon_mla_cortexa55.c:56)	called 17 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla (kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla.c:53)	called 46 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f32_f32_f32p2vlx1b_1x8vl_sme_mla (kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p2vlx1b_1x8vl_sme_mla.c:53)	called 46 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f32_f32_f32p8x1biasf32_6x8x4_neon_mla (kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p8x1biasf32_6x8x4_neon_mla.c:57)	called 17 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme_mopa (kai/ukernels/matmul/matmul_clamp_f32_f32p_f32p/kai_matmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme_mopa.c:64)	called 51 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f32_f32p_f32p/kai_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa.c:64)	called 51 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa.c:110)	called 308 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme_mopa (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme_mopa.c:110)	called 308 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f32_qai8dxp1vlx8_qsi4cxp4vlx8_1vlx4vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1vlx8_qsi4cxp4vlx8_1vlx4vl_sme2_mopa.c:80)	called 160 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f32_qai8dxp1x4_qsi4c32p4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x4_qsi4c32p4x4_1x4_neon_dotprod.c:117)	called 112 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f32_qai8dxp1x4_qsi4c32p8x4_1x8_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x4_qsi4c32p8x4_1x8_neon_dotprod.c:117)	called 112 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f32_qai8dxp1x4_qsi4cxp4vlx4_1x4vl_sme2_sdot (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi4cxp4vlx4_1x4vl_sme2_sdot.c:78)	called 160 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f32_qai8dxp1x4_qsi4cxp4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi4cxp4x4_1x4_neon_dotprod.c:89)	called 320 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot.c:113)	called 308 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme_dot (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme_dot.c:113)	called 308 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod.c:87)	called 308 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f32_qai8dxp1x8_qsi4c32p4x8_1x4x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x8_qsi4c32p4x8_1x4x32_neon_dotprod.c:117)	called 112 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f32_qai8dxp1x8_qsi4c32p8x8_1x8_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x8_qsi4c32p8x8_1x8_neon_dotprod.c:117)	called 112 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f32_qai8dxp1x8_qsi4c32p8x8_1x8x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x8_qsi4c32p8x8_1x8x32_neon_dotprod.c:117)	called 112 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f32_qai8dxp1x8_qsi4cxp4x8_1x4x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi4cxp4x8_1x4x32_neon_dotprod.c:73)	called 320 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f32_qai8dxp1x8_qsi4cxp8x8_1x8x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi4cxp8x8_1x8x32_neon_dotprod.c:73)	called 320 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f32_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod.c:87)	called 308 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f32_qai8dxp4x4_qsi4c32p4x4_16x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x4_qsi4c32p4x4_16x4_neon_dotprod.c:118)	called 112 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod.c:117)	called 112 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f32_qai8dxp4x4_qsi4cxp8x4_8x8x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x4_qsi4cxp8x4_8x8x32_neon_dotprod.c:75)	called 320 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f32_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod.c:91)	called 308 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f32_qai8dxp4x8_qsi4c32p4x8_16x4x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p4x8_16x4x32_neon_i8mm.c:118)	called 112 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f32_qai8dxp4x8_qsi4c32p4x8_8x4x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p4x8_8x4x32_neon_i8mm.c:117)	called 112 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f32_qai8dxp4x8_qsi4c32p8x8_4x8_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p8x8_4x8_neon_i8mm.c:117)	called 112 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f32_qai8dxp4x8_qsi4c32p8x8_4x8x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p8x8_4x8x32_neon_i8mm.c:117)	called 112 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x4_16x4x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x4_16x4x32_neon_dotprod.c:75)	called 320 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_4x4x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_4x4x32_neon_i8mm.c:72)	called 320 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_8x4x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_8x4x32_neon_i8mm.c:76)	called 320 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_4x8x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_4x8x32_neon_i8mm.c:76)	called 320 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_8x8x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_8x8x32_neon_i8mm.c:76)	called 320 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f32_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm.c:91)	called 308 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f32_qsi8d32p1vlx4_qai4c32p4vlx4_1vlx4vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p1vlx4_qai4c32p4vlx4_1vlx4vl_sme2_mopa.c:118)	called 2072 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa.c:96)	called 48 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f32_qsi8d32p1x4_qai4c32p4vlx4_1x4vl_sme2_dot (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qai4c32p4vlx4_1x4vl_sme2_dot.c:114)	called 2072 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f32_qsi8d32p1x4_qai4c32p4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qai4c32p4x4_1x4_neon_dotprod.c:108)	called 1036 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot.c:96)	called 48 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod.c:90)	called 48 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f32_qsi8d32p1x8_qai4c32p4x8_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p1x8_qai4c32p4x8_1x4_neon_dotprod.c:108)	called 1036 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod.c:75)	called 48 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f32_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod.c:108)	called 1036 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod.c:94)	called 48 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f32_qsi8d32p4x8_qai4c32p4x8_8x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qai4c32p4x8_8x4_neon_i8mm.c:108)	called 1036 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm.c:78)	called 96 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_8x4x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_8x4x32_neon_i8mm.c:78)	called 48 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_qai8_qai8_qsi8cxp2vlx4sb_1x16vl_sme2_dot (kai/ukernels/matmul/matmul_clamp_qai8_qai8_qsi8cxp/kai_matmul_clamp_qai8_qai8_qsi8cxp2vlx4sb_1x16vl_sme2_dot.c:54)	called 168 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_qai8_qai8p2vlx4_qsi8cxp2vlx4sb_2vlx2vl_sme_mopa (kai/ukernels/matmul/matmul_clamp_qai8_qai8p_qsi8cxp/kai_matmul_clamp_qai8_qai8p2vlx4_qsi8cxp2vlx4sb_2vlx2vl_sme_mopa.c:65)	called 333 times	100.0%	-%	100.0%
kai_get_sr_matmul_clamp_qai8_qai8p2vlx4_qsi8cxpsb2vlx4_2vlx2vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_qai8_qai8p_qsi8cxp/kai_matmul_clamp_qai8_qai8p2vlx4_qsi8cxpsb2vlx4_2vlx2vl_sme2_mopa.c:65)	called 333 times	100.0%	-%	100.0%
kai_get_src_offset_dwconv_clamp_f32_f32_f32p1vlx1b_3x3_s1_4xc_sme2_mla (kai/ukernels/dwconv/dwconv_f32_f32_f32p/kai_dwconv_clamp_f32_f32_f32p1vlx1b_3x3_s1_4xc_sme2_mla.c:72)	not called	0.0%	-%	0.0%
kai_get_version (kai/kai_common.h:63)	called 1 time	100.0%	-%	100.0%
kai_k_roundedup (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1vlx8_qsi4cxp4vlx8_1vlx4vl_sme2_mopa.c:37)	called 483 times	100.0%	-%	100.0%
kai_k_roundedup (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi4cxp4vlx4_1x4vl_sme2_sdot.c:34)	called 483 times	100.0%	-%	100.0%
kai_k_roundedup (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi4cxp4x8_1x4x32_neon_dotprod.c:30)	called 641 times	100.0%	-%	100.0%
kai_k_roundedup (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi4cxp8x8_1x8x32_neon_dotprod.c:30)	called 641 times	100.0%	-%	100.0%
kai_k_roundedup (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x4_qsi4cxp8x4_8x8x32_neon_dotprod.c:33)	called 641 times	100.0%	-%	100.0%
kai_k_roundedup (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x4_16x4x32_neon_dotprod.c:33)	called 641 times	100.0%	-%	100.0%
kai_k_roundedup (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_4x4x32_neon_i8mm.c:29)	called 641 times	100.0%	-%	100.0%
kai_k_roundedup (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_8x4x32_neon_i8mm.c:33)	called 641 times	100.0%	-%	100.0%
kai_k_roundedup (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_4x8x32_neon_i8mm.c:33)	called 641 times	100.0%	-%	100.0%
kai_k_roundedup (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_8x8x32_neon_i8mm.c:33)	called 641 times	100.0%	-%	100.0%
kai_k_roundedup (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa.c:58)	called 927 times	100.0%	-%	100.0%
kai_k_roundedup (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme_mopa.c:58)	called 927 times	100.0%	-%	100.0%
kai_k_roundedup (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot.c:59)	called 827 times	100.0%	-%	100.0%
kai_k_roundedup (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme_dot.c:59)	called 827 times	100.0%	-%	100.0%
kai_k_roundedup (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod.c:40)	called 557 times	100.0%	-%	100.0%
kai_k_roundedup (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod.c:40)	called 557 times	100.0%	-%	100.0%
kai_k_roundedup (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod.c:44)	called 617 times	100.0%	-%	100.0%
kai_k_roundedup (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm.c:44)	called 617 times	100.0%	-%	100.0%
kai_k_roundedup (kai/ukernels/matmul/pack/kai_lhs_quant_pack_qai8dxp_bf16_neon.c:24)	called 3840 times	100.0%	-%	100.0%
kai_k_roundedup (kai/ukernels/matmul/pack/kai_lhs_quant_pack_qai8dxp_f16_neon.c:28)	called 4480 times	100.0%	-%	100.0%
kai_k_roundedup (kai/ukernels/matmul/pack/kai_lhs_quant_pack_qai8dxp_f32.c:21)	called 17176 times	100.0%	-%	100.0%
kai_k_roundedup (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_qsi4cxp_qs4cxs1s0.c:18)	called 5088 times	100.0%	-%	100.0%
kai_k_roundedup (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_qsi8cxp_qsi8cx_neon.c:18)	called 2920 times	100.0%	-%	100.0%
kai_k_roundedup (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4cxp_qs4cxs1s0.c:18)	called 9352 times	100.0%	-%	100.0%
kai_k_roundedup (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4cxps1s0_qsu4cxs1s0_neon.c:21)	called 720 times	100.0%	-%	100.0%
kai_k_roundedup (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi8cxp_qsi8cx_neon.c:18)	called 5160 times	100.0%	-%	100.0%
kai_lhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi4cxp4x8_1x4x32_neon_dotprod.c:37)	called 240 times	100.0%	-%	55.0%
kai_lhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi4cxp8x8_1x8x32_neon_dotprod.c:37)	called 240 times	100.0%	-%	55.0%
kai_lhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x4_qsi4cxp8x4_8x8x32_neon_dotprod.c:39)	called 240 times	100.0%	-%	55.0%
kai_lhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x4_16x4x32_neon_dotprod.c:39)	called 240 times	100.0%	-%	55.0%
kai_lhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_4x4x32_neon_i8mm.c:36)	called 240 times	100.0%	-%	55.0%
kai_lhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_8x4x32_neon_i8mm.c:40)	called 240 times	100.0%	-%	55.0%
kai_lhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_4x8x32_neon_i8mm.c:40)	called 240 times	100.0%	-%	55.0%
kai_lhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_8x8x32_neon_i8mm.c:40)	called 240 times	100.0%	-%	55.0%
kai_lhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod.c:44)	called 211 times	100.0%	-%	100.0%
kai_lhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod.c:44)	called 211 times	100.0%	-%	100.0%
kai_lhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod.c:48)	called 231 times	100.0%	-%	100.0%
kai_lhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm.c:48)	called 231 times	100.0%	-%	100.0%
kai_lhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod.c:54)	called 26 times	100.0%	-%	100.0%
kai_lhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod.c:40)	called 26 times	100.0%	-%	100.0%
kai_lhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod.c:58)	called 46 times	100.0%	-%	100.0%
kai_lhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm.c:43)	called 92 times	100.0%	-%	100.0%
kai_lhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_8x4x32_neon_i8mm.c:43)	called 46 times	100.0%	-%	100.0%
kai_lhs_packed_stride (kai/ukernels/matmul/pack/kai_lhs_quant_pack_qai8dxp_bf16_neon.c:30)	called 2880 times	100.0%	-%	55.0%
kai_lhs_packed_stride (kai/ukernels/matmul/pack/kai_lhs_quant_pack_qai8dxp_f16_neon.c:34)	called 3360 times	100.0%	-%	55.0%
kai_lhs_packed_stride (kai/ukernels/matmul/pack/kai_lhs_quant_pack_qai8dxp_f32.c:27)	called 13236 times	100.0%	-%	55.0%
kai_lhs_packed_stride (kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p4x8sb_f32_neon.c:25)	called 92 times	100.0%	-%	100.0%
kai_lhs_packed_stride (kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p_f32.c:25)	called 340 times	100.0%	-%	100.0%
kai_lhs_packed_stride (kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p_f32_neon.c:30)	called 124 times	100.0%	-%	100.0%
kai_num_blocks_per_row (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod.c:47)	called 56 times	100.0%	-%	53.0%
kai_num_blocks_per_row (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod.c:35)	called 52 times	100.0%	-%	57.0%
kai_num_blocks_per_row (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod.c:51)	called 116 times	100.0%	-%	53.0%
kai_num_blocks_per_row (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm.c:38)	called 184 times	100.0%	-%	57.0%
kai_num_blocks_per_row (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_8x4x32_neon_i8mm.c:38)	called 92 times	100.0%	-%	57.0%
kai_num_blocks_per_row (kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p4x8sb_f32_neon.c:20)	called 115 times	100.0%	-%	57.0%
kai_num_blocks_per_row (kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p_f32.c:20)	called 415 times	100.0%	-%	57.0%
kai_num_blocks_per_row (kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p_f32_neon.c:25)	called 150 times	100.0%	-%	57.0%
kai_num_blocks_per_row (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon.c:35)	called 150 times	100.0%	-%	53.0%
kai_num_blocks_per_row (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0.c:17)	called 628 times	100.0%	-%	53.0%
kai_num_bytes_per_block (kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p4x8sb_f32_neon.c:16)	called 115 times	100.0%	-%	100.0%
kai_num_bytes_per_block (kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p_f32.c:16)	called 415 times	100.0%	-%	100.0%
kai_num_bytes_per_block (kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p_f32_neon.c:21)	called 124 times	100.0%	-%	100.0%
kai_num_bytes_per_block (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon.c:41)	called 176 times	100.0%	-%	57.0%
kai_num_bytes_per_block (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0.c:23)	called 530 times	100.0%	-%	57.0%
kai_num_bytes_per_block_lhs (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod.c:37)	called 26 times	100.0%	-%	100.0%
kai_num_bytes_per_block_lhs (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod.c:27)	called 26 times	100.0%	-%	100.0%
kai_num_bytes_per_block_lhs (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod.c:41)	called 46 times	100.0%	-%	100.0%
kai_num_bytes_per_block_lhs (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm.c:30)	called 92 times	100.0%	-%	100.0%
kai_num_bytes_per_block_lhs (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_8x4x32_neon_i8mm.c:30)	called 46 times	100.0%	-%	100.0%
kai_num_bytes_per_block_rhs (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod.c:41)	called 26 times	100.0%	-%	57.0%
kai_num_bytes_per_block_rhs (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod.c:31)	called 26 times	100.0%	-%	100.0%
kai_num_bytes_per_block_rhs (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod.c:45)	called 46 times	100.0%	-%	57.0%
kai_num_bytes_per_block_rhs (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm.c:34)	called 92 times	100.0%	-%	100.0%
kai_num_bytes_per_block_rhs (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_8x4x32_neon_i8mm.c:34)	called 46 times	100.0%	-%	100.0%
kai_rhs_get_dst_size_dwconv_pack_x32p1vlx1b_x32_x32_sme (kai/ukernels/dwconv/pack/kai_rhs_dwconv_pack_x32p1vlx1b_x32_x32_sme.c:14)	called 60 times	100.0%	-%	100.0%
kai_rhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi4cxp4x8_1x4x32_neon_dotprod.c:45)	called 240 times	100.0%	-%	55.0%
kai_rhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi4cxp8x8_1x8x32_neon_dotprod.c:45)	called 240 times	100.0%	-%	55.0%
kai_rhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x4_qsi4cxp8x4_8x8x32_neon_dotprod.c:47)	called 240 times	100.0%	-%	55.0%
kai_rhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x4_16x4x32_neon_dotprod.c:47)	called 240 times	100.0%	-%	55.0%
kai_rhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_4x4x32_neon_i8mm.c:44)	called 240 times	100.0%	-%	55.0%
kai_rhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_8x4x32_neon_i8mm.c:48)	called 240 times	100.0%	-%	55.0%
kai_rhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_4x8x32_neon_i8mm.c:48)	called 240 times	100.0%	-%	55.0%
kai_rhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_8x8x32_neon_i8mm.c:48)	called 240 times	100.0%	-%	55.0%
kai_rhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod.c:54)	called 211 times	100.0%	-%	100.0%
kai_rhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod.c:54)	called 211 times	100.0%	-%	100.0%
kai_rhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod.c:58)	called 231 times	100.0%	-%	100.0%
kai_rhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm.c:58)	called 231 times	100.0%	-%	100.0%
kai_rhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod.c:58)	called 26 times	100.0%	-%	53.0%
kai_rhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod.c:44)	called 26 times	100.0%	-%	52.0%
kai_rhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod.c:62)	called 46 times	100.0%	-%	53.0%
kai_rhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm.c:47)	called 92 times	100.0%	-%	52.0%
kai_rhs_packed_stride (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_8x4x32_neon_i8mm.c:47)	called 46 times	100.0%	-%	52.0%
kai_rhs_stride (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon.c:47)	called 26 times	100.0%	-%	52.0%
kai_rhs_stride (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0.c:28)	called 98 times	100.0%	-%	52.0%
kai_roundup (kai/kai_common.h:142)	called 10735877 times	100.0%	-%	100.0%
kai_roundup(unsigned long, unsigned long) (kai/kai_common.h:142)	called 200 times	100.0%	-%	100.0%
kai_run_dwconv_clamp_f32_f32_f32p1vlx1b_3x3_s1_4xc_sme2_mla (kai/ukernels/dwconv/dwconv_f32_f32_f32p/kai_dwconv_clamp_f32_f32_f32p1vlx1b_3x3_s1_4xc_sme2_mla.c:76)	called 1038 times	100.0%	100.0%	61.0%
kai_run_imatmul_clamp_f16_f16p2vlx2_f16p2vlx2_2vlx2vl_sme2_mopa (kai/ukernels/matmul/imatmul_clamp_f16_f16p_f16p/kai_imatmul_clamp_f16_f16p2vlx2_f16p2vlx2_2vlx2vl_sme2_mopa.c:85)	called 5676 times	100.0%	-%	100.0%
kai_run_imatmul_clamp_f16_f16p2vlx2_f16p2vlx2b_2vlx2vl_sme_mopa (kai/ukernels/matmul/imatmul_clamp_f16_f16p_f16p/kai_imatmul_clamp_f16_f16p2vlx2_f16p2vlx2b_2vlx2vl_sme_mopa.c:85)	called 5676 times	100.0%	-%	100.0%
kai_run_imatmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme2_mopa (kai/ukernels/matmul/imatmul_clamp_f32_f32p_f32p/kai_imatmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme2_mopa.c:84)	called 5676 times	100.0%	-%	100.0%
kai_run_imatmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme_mopa (kai/ukernels/matmul/imatmul_clamp_f32_f32p_f32p/kai_imatmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme_mopa.c:84)	called 5676 times	100.0%	-%	100.0%
kai_run_imatmul_clamp_qai8_qai8p2vlx4_qsi8cxp2vlx4sb_2vlx2vl_sme_mopa (kai/ukernels/matmul/imatmul_clamp_qai8_qai8p_qsi8cxp/kai_imatmul_clamp_qai8_qai8p2vlx4_qsi8cxp2vlx4sb_2vlx2vl_sme_mopa.c:85)	called 4662 times	100.0%	-%	100.0%
kai_run_imatmul_clamp_qai8_qai8p2vlx4_qsi8cxpsb2vlx4_2vlx2vl_sme2_mopa (kai/ukernels/matmul/imatmul_clamp_qai8_qai8p_qsi8cxp/kai_imatmul_clamp_qai8_qai8p2vlx4_qsi8cxpsb2vlx4_2vlx2vl_sme2_mopa.c:85)	called 4662 times	100.0%	-%	100.0%
kai_run_lhs_imatmul_pack_x16p2vlx2_x16p_sme (kai/ukernels/matmul/pack/kai_lhs_imatmul_pack_x16p2vlx2_x16p_sme.c:47)	called 11352 times	100.0%	100.0%	71.0%
kai_run_lhs_imatmul_pack_x32p2vlx1_x32p_sme (kai/ukernels/matmul/pack/kai_lhs_imatmul_pack_x32p2vlx1_x32p_sme.c:47)	called 11352 times	100.0%	100.0%	71.0%
kai_run_lhs_imatmul_pack_x8p2vlx4_x8p_sme (kai/ukernels/matmul/pack/kai_lhs_imatmul_pack_x8p2vlx4_x8p_sme.c:47)	called 9324 times	100.0%	100.0%	71.0%
kai_run_lhs_pack_bf16p2vlx2_f32_sme (kai/ukernels/matmul/pack/kai_lhs_pack_bf16p2vlx2_f32_sme.c:67)	called 46 times	100.0%	100.0%	64.0%
kai_run_lhs_pack_bf16p8x4_f16_neon (kai/ukernels/matmul/pack/kai_lhs_pack_bf16p8x4_f16_neon.c:55)	called 176 times	100.0%	100.0%	64.0%
kai_run_lhs_pack_f32p2vlx1_f32_sme (kai/ukernels/matmul/pack/kai_lhs_pack_f32p2vlx1_f32_sme.c:86)	called 34 times	100.0%	100.0%	62.0%
kai_run_lhs_pack_x16p2vlx2_x16_sme (kai/ukernels/matmul/pack/kai_lhs_pack_x16p2vlx2_x16_sme.c:86)	called 34 times	100.0%	100.0%	62.0%
kai_run_lhs_pack_x8p2vlx4_x8_sme (kai/ukernels/matmul/pack/kai_lhs_pack_x8p2vlx4_x8_sme.c:86)	called 666 times	100.0%	100.0%	62.0%
kai_run_lhs_quant_pack_bf16p1x4_f32_neon (kai/ukernels/matmul/pack/kai_lhs_quant_pack_bf16p1x4_f32_neon.c:53)	called 92 times	100.0%	100.0%	66.0%
kai_run_lhs_quant_pack_bf16p8x4_f32_neon (kai/ukernels/matmul/pack/kai_lhs_quant_pack_bf16p8x4_f32_neon.c:54)	called 92 times	100.0%	100.0%	62.0%
kai_run_lhs_quant_pack_qai8dxp_bf16_neon (kai/ukernels/matmul/pack/kai_lhs_quant_pack_qai8dxp_bf16_neon.c:65)	called 960 times	98.6%	77.4%	83.0%
kai_run_lhs_quant_pack_qai8dxp_f16_neon (kai/ukernels/matmul/pack/kai_lhs_quant_pack_qai8dxp_f16_neon.c:66)	called 1120 times	98.4%	75.6%	83.0%
kai_run_lhs_quant_pack_qai8dxp_f32 (kai/ukernels/matmul/pack/kai_lhs_quant_pack_qai8dxp_f32.c:58)	called 3940 times	99.1%	89.5%	90.0%
kai_run_lhs_quant_pack_qsi8d32p4x8sb_f32_neon (kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p4x8sb_f32_neon.c:65)	called 23 times	98.5%	88.5%	80.0%
kai_run_lhs_quant_pack_qsi8d32p_f32 (kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p_f32.c:65)	called 75 times	97.5%	90.0%	95.0%
kai_run_lhs_quant_pack_qsi8d32p_f32_neon (kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p_f32_neon.c:73)	called 26 times	97.4%	88.9%	77.0%
kai_run_lhs_quant_pack_qsi8d32pscalef32_f16_neon (kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32pscalef32_f16_neon.c:67)	called 10904 times	100.0%	87.5%	77.0%
kai_run_lhs_quant_pack_qsi8d32pscalef32_f32_neon (kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32pscalef32_f32_neon.c:71)	called 10904 times	100.0%	81.8%	73.0%
kai_run_matmul_clamp_bf16_qai8dxp1x8_qsi4c32p4x8_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_bf16_qai8dxp_qsi4c32p/kai_matmul_clamp_bf16_qai8dxp1x8_qsi4c32p4x8_1x4_neon_dotprod.c:148)	called 163 times	94.4%	50.0%	54.0%
kai_run_matmul_clamp_bf16_qai8dxp1x8_qsi4cxp8x8_1x8_neon_dotprod (kai/ukernels/matmul/matmul_clamp_bf16_qai8dxp_qsi4cxp/kai_matmul_clamp_bf16_qai8dxp1x8_qsi4cxp8x8_1x8_neon_dotprod.c:132)	called 357 times	94.1%	50.0%	60.0%
kai_run_matmul_clamp_bf16_qai8dxp4x8_qsi4c32p4x8_16x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_bf16_qai8dxp_qsi4c32p/kai_matmul_clamp_bf16_qai8dxp4x8_qsi4c32p4x8_16x4_neon_i8mm.c:149)	called 163 times	94.4%	50.0%	54.0%
kai_run_matmul_clamp_bf16_qai8dxp4x8_qsi4cxp8x8_8x8_neon_i8mm (kai/ukernels/matmul/matmul_clamp_bf16_qai8dxp_qsi4cxp/kai_matmul_clamp_bf16_qai8dxp4x8_qsi4cxp8x8_8x8_neon_i8mm.c:133)	called 389 times	94.1%	50.0%	60.0%
kai_run_matmul_clamp_f16_bf16p8x4_bf16p12x4b_8x12_neon_mmla (kai/ukernels/matmul/matmul_clamp_f16_bf16p_bf16p/kai_matmul_clamp_f16_bf16p8x4_bf16p12x4b_8x12_neon_mmla.c:75)	called 85 times	100.0%	-%	57.0%
kai_run_matmul_clamp_f16_f16_f16p16x1biasf16_6x16x8_neon_mla (kai/ukernels/matmul/matmul_clamp_f16_f16_f16p/kai_matmul_clamp_f16_f16_f16p16x1biasf16_6x16x8_neon_mla.c:72)	called 17 times	100.0%	-%	57.0%
kai_run_matmul_clamp_f16_f16_f16p2vlx2b_1x16vl_sme2_dot (kai/ukernels/matmul/matmul_clamp_f16_f16_f16p/kai_matmul_clamp_f16_f16_f16p2vlx2b_1x16vl_sme2_dot.c:89)	called 47 times	100.0%	-%	57.0%
kai_run_matmul_clamp_f16_f16_f16p2vlx2b_1x8vl_sme_mla (kai/ukernels/matmul/matmul_clamp_f16_f16_f16p/kai_matmul_clamp_f16_f16_f16p2vlx2b_1x8vl_sme_mla.c:88)	called 47 times	100.0%	-%	57.0%
kai_run_matmul_clamp_f16_f16_f16p32x1b_6x32_neon_mla (kai/ukernels/matmul/matmul_clamp_f16_f16_f16p/kai_matmul_clamp_f16_f16_f16p32x1b_6x32_neon_mla.c:87)	called 17 times	100.0%	-%	100.0%
kai_run_matmul_clamp_f16_f16_f16p32x1b_6x32_neon_mla_cortexa55 (kai/ukernels/matmul/matmul_clamp_f16_f16_f16p/kai_matmul_clamp_f16_f16_f16p32x1b_6x32_neon_mla_cortexa55.c:88)	called 17 times	100.0%	-%	100.0%
kai_run_matmul_clamp_f16_f16p2vlx2_f16p2vlx2_2vlx2vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f16_f16p_f16p/kai_matmul_clamp_f16_f16p2vlx2_f16p2vlx2_2vlx2vl_sme2_mopa.c:97)	called 17 times	100.0%	-%	100.0%
kai_run_matmul_clamp_f16_f16p2vlx2_f16p2vlx2b_2vlx2vl_sme_mopa (kai/ukernels/matmul/matmul_clamp_f16_f16p_f16p/kai_matmul_clamp_f16_f16p2vlx2_f16p2vlx2b_2vlx2vl_sme_mopa.c:97)	called 17 times	100.0%	-%	100.0%
kai_run_matmul_clamp_f16_qai8dxp1x4_qsi4cxp4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi4cxp/kai_matmul_clamp_f16_qai8dxp1x4_qsi4cxp4x4_1x4_neon_dotprod.c:132)	called 59 times	94.1%	50.0%	60.0%
kai_run_matmul_clamp_f16_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi8cxp/kai_matmul_clamp_f16_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod.c:129)	called 59 times	93.8%	50.0%	60.0%
kai_run_matmul_clamp_f16_qai8dxp1x8_qsi4cxp4x8_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi4cxp/kai_matmul_clamp_f16_qai8dxp1x8_qsi4cxp4x8_1x4_neon_dotprod.c:132)	called 59 times	94.1%	50.0%	60.0%
kai_run_matmul_clamp_f16_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi8cxp/kai_matmul_clamp_f16_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod.c:129)	called 59 times	93.8%	50.0%	60.0%
kai_run_matmul_clamp_f16_qai8dxp4x4_qsi4cxp4x4_16x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi4cxp/kai_matmul_clamp_f16_qai8dxp4x4_qsi4cxp4x4_16x4_neon_dotprod.c:132)	called 223 times	94.1%	50.0%	60.0%
kai_run_matmul_clamp_f16_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi8cxp/kai_matmul_clamp_f16_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod.c:129)	called 223 times	93.8%	50.0%	60.0%
kai_run_matmul_clamp_f16_qai8dxp4x8_qsi4cxp4x8_16x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi4cxp/kai_matmul_clamp_f16_qai8dxp4x8_qsi4cxp4x8_16x4_neon_i8mm.c:132)	called 223 times	94.1%	50.0%	60.0%
kai_run_matmul_clamp_f16_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi8cxp/kai_matmul_clamp_f16_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm.c:129)	called 223 times	93.8%	50.0%	60.0%
kai_run_matmul_clamp_f16_qsi8d32p1vlx4_qai4c32p4vlx4_1vlx4vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p1vlx4_qai4c32p4vlx4_1vlx4vl_sme2_mopa.c:157)	called 1033 times	95.2%	50.0%	54.0%
kai_run_matmul_clamp_f16_qsi8d32p1x4_qai4c32p4vlx4_1x4vl_sme2_dot (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p1x4_qai4c32p4vlx4_1x4vl_sme2_dot.c:153)	called 277 times	94.1%	50.0%	53.0%
kai_run_matmul_clamp_f16_qsi8d32p1x4_qai4c32p4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p1x4_qai4c32p4x4_1x4_neon_dotprod.c:141)	called 139 times	94.4%	50.0%	54.0%
kai_run_matmul_clamp_f16_qsi8d32p1x8_qai4c32p4x8_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p1x8_qai4c32p4x8_1x4_neon_dotprod.c:141)	called 139 times	94.4%	50.0%	54.0%
kai_run_matmul_clamp_f16_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod.c:141)	called 517 times	94.4%	50.0%	54.0%
kai_run_matmul_clamp_f16_qsi8d32p4x8_qai4c32p4x8_8x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p4x8_qai4c32p4x8_8x4_neon_i8mm.c:141)	called 517 times	94.4%	50.0%	54.0%
kai_run_matmul_clamp_f32_bf16p1x4_bf16p12x4b_1x36_neon_dot (kai/ukernels/matmul/matmul_clamp_f32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p1x4_bf16p12x4b_1x36_neon_dot.c:76)	called 93 times	100.0%	-%	57.0%
kai_run_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa.c:96)	called 47 times	100.0%	-%	100.0%
kai_run_matmul_clamp_f32_bf16p8x4_bf16p12x4b_8x12_neon_mmla (kai/ukernels/matmul/matmul_clamp_f32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p8x4_bf16p12x4b_8x12_neon_mmla.c:76)	called 185 times	100.0%	-%	57.0%
kai_run_matmul_clamp_f32_f32_f32p16vlx1b_1x16vl_sme2_mla (kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p16vlx1b_1x16vl_sme2_mla.c:87)	called 47 times	100.0%	-%	57.0%
kai_run_matmul_clamp_f32_f32_f32p16x1b_6x16_neon_mla (kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p16x1b_6x16_neon_mla.c:83)	called 17 times	100.0%	-%	100.0%
kai_run_matmul_clamp_f32_f32_f32p16x1b_6x16_neon_mla_cortexa55 (kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p16x1b_6x16_neon_mla_cortexa55.c:84)	called 17 times	100.0%	-%	100.0%
kai_run_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla (kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla.c:87)	called 47 times	100.0%	-%	57.0%
kai_run_matmul_clamp_f32_f32_f32p2vlx1b_1x8vl_sme_mla (kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p2vlx1b_1x8vl_sme_mla.c:86)	called 47 times	100.0%	-%	57.0%
kai_run_matmul_clamp_f32_f32_f32p8x1biasf32_6x8x4_neon_mla (kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p8x1biasf32_6x8x4_neon_mla.c:85)	called 17 times	100.0%	-%	57.0%
kai_run_matmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme_mopa (kai/ukernels/matmul/matmul_clamp_f32_f32p_f32p/kai_matmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme_mopa.c:96)	called 17 times	100.0%	-%	100.0%
kai_run_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f32_f32p_f32p/kai_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa.c:96)	called 17 times	100.0%	-%	100.0%
kai_run_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa.c:142)	called 155 times	100.0%	-%	52.0%
kai_run_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme_mopa (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme_mopa.c:142)	called 155 times	100.0%	-%	53.0%
kai_run_matmul_clamp_f32_qai8dxp1vlx8_qsi4cxp4vlx8_1vlx4vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1vlx8_qsi4cxp4vlx8_1vlx4vl_sme2_mopa.c:112)	called 81 times	100.0%	-%	52.0%
kai_run_matmul_clamp_f32_qai8dxp1x4_qsi4c32p4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x4_qsi4c32p4x4_1x4_neon_dotprod.c:147)	called 163 times	94.4%	50.0%	54.0%
kai_run_matmul_clamp_f32_qai8dxp1x4_qsi4c32p8x4_1x8_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x4_qsi4c32p8x4_1x8_neon_dotprod.c:147)	called 163 times	94.4%	50.0%	54.0%
kai_run_matmul_clamp_f32_qai8dxp1x4_qsi4cxp4vlx4_1x4vl_sme2_sdot (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi4cxp4vlx4_1x4vl_sme2_sdot.c:113)	called 81 times	93.3%	50.0%	66.0%
kai_run_matmul_clamp_f32_qai8dxp1x4_qsi4cxp4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi4cxp4x4_1x4_neon_dotprod.c:117)	called 161 times	90.0%	50.0%	60.0%
kai_run_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot.c:141)	called 135 times	96.2%	50.0%	60.0%
kai_run_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme_dot (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme_dot.c:141)	called 135 times	96.3%	50.0%	75.0%
kai_run_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod.c:115)	called 135 times	90.0%	50.0%	60.0%
kai_run_matmul_clamp_f32_qai8dxp1x8_qsi4c32p4x8_1x4x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x8_qsi4c32p4x8_1x4x32_neon_dotprod.c:147)	called 163 times	94.4%	50.0%	54.0%
kai_run_matmul_clamp_f32_qai8dxp1x8_qsi4c32p8x8_1x8_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x8_qsi4c32p8x8_1x8_neon_dotprod.c:147)	called 163 times	94.4%	50.0%	54.0%
kai_run_matmul_clamp_f32_qai8dxp1x8_qsi4c32p8x8_1x8x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x8_qsi4c32p8x8_1x8x32_neon_dotprod.c:147)	called 163 times	94.4%	50.0%	54.0%
kai_run_matmul_clamp_f32_qai8dxp1x8_qsi4cxp4x8_1x4x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi4cxp4x8_1x4x32_neon_dotprod.c:101)	called 161 times	90.0%	50.0%	60.0%
kai_run_matmul_clamp_f32_qai8dxp1x8_qsi4cxp8x8_1x8x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi4cxp8x8_1x8x32_neon_dotprod.c:101)	called 161 times	90.0%	50.0%	60.0%
kai_run_matmul_clamp_f32_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod.c:115)	called 135 times	90.0%	50.0%	60.0%
kai_run_matmul_clamp_f32_qai8dxp4x4_qsi4c32p4x4_16x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x4_qsi4c32p4x4_16x4_neon_dotprod.c:148)	called 163 times	95.2%	75.0%	60.0%
kai_run_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod.c:147)	called 163 times	94.4%	50.0%	54.0%
kai_run_matmul_clamp_f32_qai8dxp4x4_qsi4cxp8x4_8x8x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x4_qsi4cxp8x4_8x8x32_neon_dotprod.c:103)	called 161 times	90.0%	50.0%	60.0%
kai_run_matmul_clamp_f32_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod.c:119)	called 155 times	90.0%	50.0%	60.0%
kai_run_matmul_clamp_f32_qai8dxp4x8_qsi4c32p4x8_16x4x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p4x8_16x4x32_neon_i8mm.c:148)	called 163 times	95.2%	75.0%	60.0%
kai_run_matmul_clamp_f32_qai8dxp4x8_qsi4c32p4x8_8x4x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p4x8_8x4x32_neon_i8mm.c:147)	called 163 times	94.4%	50.0%	54.0%
kai_run_matmul_clamp_f32_qai8dxp4x8_qsi4c32p8x8_4x8_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p8x8_4x8_neon_i8mm.c:147)	called 163 times	94.4%	50.0%	54.0%
kai_run_matmul_clamp_f32_qai8dxp4x8_qsi4c32p8x8_4x8x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p8x8_4x8x32_neon_i8mm.c:147)	called 163 times	94.4%	50.0%	54.0%
kai_run_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x4_16x4x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x4_16x4x32_neon_dotprod.c:103)	called 161 times	90.0%	50.0%	60.0%
kai_run_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_4x4x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_4x4x32_neon_i8mm.c:100)	called 161 times	90.0%	50.0%	60.0%
kai_run_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_8x4x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_8x4x32_neon_i8mm.c:104)	called 161 times	90.0%	50.0%	60.0%
kai_run_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_4x8x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_4x8x32_neon_i8mm.c:104)	called 161 times	90.0%	50.0%	60.0%
kai_run_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_8x8x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_8x8x32_neon_i8mm.c:104)	called 161 times	90.0%	50.0%	60.0%
kai_run_matmul_clamp_f32_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm.c:119)	called 155 times	90.0%	50.0%	60.0%
kai_run_matmul_clamp_f32_qsi8d32p1vlx4_qai4c32p4vlx4_1vlx4vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p1vlx4_qai4c32p4vlx4_1vlx4vl_sme2_mopa.c:157)	called 1033 times	95.2%	50.0%	54.0%
kai_run_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa.c:133)	called 24 times	95.0%	50.0%	60.0%
kai_run_matmul_clamp_f32_qsi8d32p1x4_qai4c32p4vlx4_1x4vl_sme2_dot (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qai4c32p4vlx4_1x4vl_sme2_dot.c:153)	called 277 times	94.1%	50.0%	53.0%
kai_run_matmul_clamp_f32_qsi8d32p1x4_qai4c32p4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qai4c32p4x4_1x4_neon_dotprod.c:139)	called 139 times	94.4%	50.0%	53.0%
kai_run_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot.c:134)	called 4 times	95.0%	50.0%	56.0%
kai_run_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod.c:120)	called 4 times	90.0%	50.0%	56.0%
kai_run_matmul_clamp_f32_qsi8d32p1x8_qai4c32p4x8_1x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p1x8_qai4c32p4x8_1x4_neon_dotprod.c:139)	called 139 times	94.4%	50.0%	53.0%
kai_run_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod.c:113)	called 4 times	88.9%	50.0%	54.0%
kai_run_matmul_clamp_f32_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod.c:139)	called 517 times	94.4%	50.0%	54.0%
kai_run_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod.c:124)	called 24 times	90.0%	50.0%	60.0%
kai_run_matmul_clamp_f32_qsi8d32p4x8_qai4c32p4x8_8x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qai4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qai4c32p4x8_8x4_neon_i8mm.c:139)	called 517 times	94.4%	50.0%	54.0%
kai_run_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm.c:116)	called 47 times	88.9%	50.0%	54.0%
kai_run_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_8x4x32_neon_i8mm (kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_8x4x32_neon_i8mm.c:116)	called 24 times	88.9%	50.0%	54.0%
kai_run_matmul_clamp_qai8_qai8_qsi8cxp2vlx4sb_1x16vl_sme2_dot (kai/ukernels/matmul/matmul_clamp_qai8_qai8_qsi8cxp/kai_matmul_clamp_qai8_qai8_qsi8cxp2vlx4sb_1x16vl_sme2_dot.c:88)	called 169 times	100.0%	-%	57.0%
kai_run_matmul_clamp_qai8_qai8p2vlx4_qsi8cxp2vlx4sb_2vlx2vl_sme_mopa (kai/ukernels/matmul/matmul_clamp_qai8_qai8p_qsi8cxp/kai_matmul_clamp_qai8_qai8p2vlx4_qsi8cxp2vlx4sb_2vlx2vl_sme_mopa.c:97)	called 334 times	100.0%	-%	100.0%
kai_run_matmul_clamp_qai8_qai8p2vlx4_qsi8cxpsb2vlx4_2vlx2vl_sme2_mopa (kai/ukernels/matmul/matmul_clamp_qai8_qai8p_qsi8cxp/kai_matmul_clamp_qai8_qai8p2vlx4_qsi8cxpsb2vlx4_2vlx2vl_sme2_mopa.c:97)	called 334 times	100.0%	-%	100.0%
kai_run_rhs_dwconv_pack_x32p1vlx1b_x32_x32_sme (kai/ukernels/dwconv/pack/kai_rhs_dwconv_pack_x32p1vlx1b_x32_x32_sme.c:20)	called 60 times	100.0%	100.0%	71.0%
kai_run_rhs_imatmul_pack_kxn_qsi8cxp2vlx4sb_qs8cx_f32_i32_sme (kai/ukernels/matmul/pack/kai_rhs_imatmul_pack_kxn_qsi8cxp2vlx4sb_qs8cx_f32_i32_sme.c:87)	called 9324 times	100.0%	-%	51.0%
kai_run_rhs_imatmul_pack_kxn_x16p2vlx2b_x16_x16_sme (kai/ukernels/matmul/pack/kai_rhs_imatmul_pack_kxn_x16p2vlx2b_x16_x16_sme.c:78)	called 11352 times	100.0%	-%	52.0%
kai_run_rhs_imatmul_pack_kxn_x32p2vlx1b_x32_x32_sme (kai/ukernels/matmul/pack/kai_rhs_imatmul_pack_kxn_x32p2vlx1b_x32_x32_sme.c:75)	called 11352 times	100.0%	-%	52.0%
kai_run_rhs_pack_kxn_bf16p12x4biasf16_f16_neon (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_bf16p12x4biasf16_f16_neon.c:57)	called 102 times	100.0%	100.0%	55.0%
kai_run_rhs_pack_kxn_bf16p12x4biasf32_f16_neon (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_bf16p12x4biasf32_f16_neon.c:51)	called 112 times	100.0%	100.0%	55.0%
kai_run_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme.c:56)	called 56 times	100.0%	-%	50.0%
kai_run_rhs_pack_kxn_f16p16x1biasf16_f16_f16_neon (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_f16p16x1biasf16_f16_f16_neon.c:45)	called 17 times	100.0%	-%	50.0%
kai_run_rhs_pack_kxn_f32p16vlx1b_f32_f32_sme (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_f32p16vlx1b_f32_f32_sme.c:69)	called 46 times	100.0%	-%	50.0%
kai_run_rhs_pack_kxn_f32p2vlx1biasf32_f32_f32_sme (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_f32p2vlx1biasf32_f32_f32_sme.c:69)	called 126 times	100.0%	-%	50.0%
kai_run_rhs_pack_kxn_f32p8x1biasf32_f32_f32_neon (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_f32p8x1biasf32_f32_f32_neon.c:45)	called 17 times	100.0%	-%	50.0%
kai_run_rhs_pack_kxn_qsi4c32p_qsu4c32s1s0 (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_qsi4c32p_qsu4c32s1s0.c:113)	called 702 times	97.6%	88.5%	67.0%
kai_run_rhs_pack_kxn_qsi4cxp_qs4cxs1s0 (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_qsi4cxp_qs4cxs1s0.c:58)	called 1092 times	100.0%	100.0%	74.0%
kai_run_rhs_pack_kxn_qsi8cxp2vlx4sb_qs8cx_f32_i32_sme (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_qsi8cxp2vlx4sb_qs8cx_f32_i32_sme.c:80)	called 834 times	100.0%	-%	50.0%
kai_run_rhs_pack_kxn_qsi8cxp_qsi8cx_neon (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_qsi8cxp_qsi8cx_neon.c:54)	called 576 times	95.9%	95.5%	74.0%
kai_run_rhs_pack_kxn_x16p2vlx2b_x16_x16_sme (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_x16p2vlx2b_x16_x16_sme.c:72)	called 126 times	100.0%	-%	50.0%
kai_run_rhs_pack_kxn_x16p32x1b_x16_x16_neon (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_x16p32x1b_x16_x16_neon.c:70)	called 34 times	100.0%	-%	50.0%
kai_run_rhs_pack_kxn_x32p16x1b_x32_x32_neon (kai/ukernels/matmul/pack/kai_rhs_pack_kxn_x32p16x1b_x32_x32_neon.c:70)	called 34 times	100.0%	-%	50.0%
kai_run_rhs_pack_nxk_f32p2vlx1biasf32_f32_f32_sme (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_f32p2vlx1biasf32_f32_f32_sme.c:65)	called 34 times	100.0%	100.0%	59.0%
kai_run_rhs_pack_nxk_qai4c32p_qau4c32s0s1_f32_f32_f32_neon (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qai4c32p_qau4c32s0s1_f32_f32_f32_neon.c:68)	called 2616 times	100.0%	90.9%	66.0%
kai_run_rhs_pack_nxk_qai4c32ps1s0nrx4_qau4c32s0s1_f32_f32_f32_neon (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qai4c32ps1s0nrx4_qau4c32s0s1_f32_f32_f32_neon.c:70)	called 1308 times	100.0%	100.0%	65.0%
kai_run_rhs_pack_nxk_qai4c32ps1s0nrx4_qau4c32s1s0_f32_f32_f32_neon (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qai4c32ps1s0nrx4_qau4c32s1s0_f32_f32_f32_neon.c:70)	called 1308 times	100.0%	100.0%	65.0%
kai_run_rhs_pack_nxk_qsi4c32p_qsu4c32s1s0 (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32p_qsu4c32s1s0.c:111)	called 702 times	97.7%	84.4%	70.0%
kai_run_rhs_pack_nxk_qsi4c32pnrx4_qsu4c32s1s0_neon (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32pnrx4_qsu4c32s1s0_neon.c:121)	called 216 times	98.6%	96.2%	67.0%
kai_run_rhs_pack_nxk_qsi4c32pnrx8_qsu4c32s1s0_neon (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32pnrx8_qsu4c32s1s0_neon.c:121)	called 486 times	98.3%	96.2%	67.0%
kai_run_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon.c:99)	called 26 times	100.0%	100.0%	58.0%
kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0 (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0.c:78)	called 98 times	88.6%	85.0%	62.0%
kai_run_rhs_pack_nxk_qsi4cxp_qs4cxs1s0 (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4cxp_qs4cxs1s0.c:55)	called 1652 times	100.0%	100.0%	74.0%
kai_run_rhs_pack_nxk_qsi4cxps1s0_qsu4cxs1s0_neon (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4cxps1s0_qsu4cxs1s0_neon.c:61)	called 160 times	97.1%	96.9%	69.0%
kai_run_rhs_pack_nxk_qsi8cxp_qsi8cx_neon (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi8cxp_qsi8cx_neon.c:53)	called 1136 times	100.0%	100.0%	75.0%
kai_run_rhs_pack_nxk_x16p2vlx2b_x16_x16_sme (kai/ukernels/matmul/pack/kai_rhs_pack_nxk_x16p2vlx2b_x16_x16_sme.c:65)	called 34 times	100.0%	100.0%	59.0%
kai_run_rhs_quant_pack_kxn_bf16p12x4biasf32_f32_neon (kai/ukernels/matmul/pack/kai_rhs_quant_pack_kxn_bf16p12x4biasf32_f32_neon.c:56)	called 228 times	100.0%	100.0%	54.0%
main (benchmark/main.cpp:43)	called 1 time	81.0%	43.1%	56.0%
signed char kai::test::quantize_asymmetric<float, signed char, int>(float, float, int) (test/reference/quantize.cpp:72)	called 14734805 times	100.0%	-%	75.0%
signed char kai::test::quantize_symmetric<signed char>(float, float) (test/reference/quantize.cpp:60)	called 30124681 times	100.0%	50.0%	75.0%
signed char kai::test::read_array<signed char>(void const*, unsigned long) (test/common/memory.hpp:39)	called 2602434964 times	100.0%	-%	100.0%
std::__1::hash<kai::test::DataFormat::PackFormat>::operator()(kai::test::DataFormat::PackFormat const&) const (test/common/data_format.hpp:172)	called 120417 times	100.0%	-%	100.0%
std::__1::hash<kai::test::DataFormat>::operator()(kai::test::DataFormat const&) const (test/common/data_format.hpp:165)	called 3729 times	100.0%	-%	100.0%
std::__1::hash<kai::test::DataType>::operator()(kai::test::DataType const&) const (test/common/data_type.hpp:117)	called 361251 times	100.0%	-%	100.0%
std::__1::hash<kai::test::MatMulShape>::operator()(kai::test::MatMulShape const&) const (test/common/matmul_test_common.hpp:507)	called 1243 times	100.0%	-%	100.0%
std::__1::tuple<float, float> kai::test::(anonymous namespace)::calculate_error<float>(float, float) (test/common/compare.cpp:39)	called 78372476 times	100.0%	-%	100.0%
std::__1::tuple<float, float> kai::test::(anonymous namespace)::calculate_error<int>(int, int) (test/common/compare.cpp:39)	not called	100.0%	-%	0.0%
std::__1::tuple<float, float> kai::test::(anonymous namespace)::calculate_error<kai::test::BFloat16<true>>(kai::test::BFloat16<true>, kai::test::BFloat16<true>) (test/common/compare.cpp:39)	called 2776060 times	100.0%	-%	100.0%
std::__1::tuple<float, float> kai::test::(anonymous namespace)::calculate_error<kai::test::Float16>(kai::test::Float16, kai::test::Float16) (test/common/compare.cpp:39)	called 38971162 times	100.0%	-%	100.0%
std::__1::tuple<float, float> kai::test::(anonymous namespace)::calculate_error<kai::test::Int4>(kai::test::Int4, kai::test::Int4) (test/common/compare.cpp:39)	not called	100.0%	50.0%	0.0%
std::__1::tuple<float, float> kai::test::(anonymous namespace)::calculate_error<signed char>(signed char, signed char) (test/common/compare.cpp:39)	not called	100.0%	-%	0.0%
std::__1::tuple<float, float> kai::test::find_clamp_range<float>(void const*, unsigned long, float) (test/reference/clamp.cpp:22)	called 8520 times	100.0%	-%	66.0%
std::__1::tuple<float, int> kai::test::(anonymous namespace)::get_scale_zero_point_from_range<float, kai::test::Int4, int>(float, float) (test/reference/quantize.cpp:30)	called 525168 times	88.2%	50.0%	72.0%
std::__1::tuple<float, int> kai::test::(anonymous namespace)::get_scale_zero_point_from_range<float, signed char, int>(float, float) (test/reference/quantize.cpp:30)	called 119838 times	100.0%	-%	72.0%
std::__1::tuple<kai::test::Buffer, kai::test::Buffer, kai::test::Buffer> kai::test::quantize_asymmetric_per_block_dynamic<float, kai::test::Int4, float, int>(void const*, unsigned long, unsigned long, unsigned long) (test/reference/quantize.cpp:273)	called 5232 times	66.7%	26.7%	66.0%
std::__1::tuple<kai::test::Buffer, kai::test::Buffer, kai::test::Buffer> kai::test::quantize_asymmetric_per_block_dynamic<float, signed char, float, int>(void const*, unsigned long, unsigned long, unsigned long) (test/reference/quantize.cpp:273)	called 7047 times	100.0%	-%	66.0%
std::__1::tuple<kai::test::Buffer, kai::test::Buffer, kai::test::Buffer> kai::test::quantize_asymmetric_per_block_dynamic<float, signed char, kai::test::BFloat16<true>, int>(void const*, unsigned long, unsigned long, unsigned long) (test/reference/quantize.cpp:273)	not called	100.0%	-%	0.0%
std::__1::tuple<kai::test::Buffer, kai::test::Buffer> kai::test::compute_asymmetric_per_block_quantization_info<float, kai::test::Int4, float, int>(void const*, unsigned long, unsigned long, unsigned long) (test/reference/quantize.cpp:189)	called 5232 times	100.0%	68.8%	73.0%
std::__1::tuple<kai::test::Buffer, kai::test::Buffer> kai::test::compute_asymmetric_per_block_quantization_info<float, signed char, float, int>(void const*, unsigned long, unsigned long, unsigned long) (test/reference/quantize.cpp:189)	called 7974 times	100.0%	-%	73.0%
std::__1::tuple<kai::test::Buffer, kai::test::Buffer> kai::test::quantize_rhs_qsi4c32p<float, kai::test::BFloat16<false>>(unsigned long, unsigned long, unsigned long, kai::test::Buffer const&, bool) (test/reference/quantize.cpp:305)	called 1404 times	100.0%	68.8%	68.0%
std::__1::tuple<kai::test::Buffer, kai::test::Buffer> kai::test::quantize_symmetric_per_block_dynamic<float, int, float>(void const*, unsigned long, unsigned long, unsigned long) (test/reference/quantize.cpp:156)	not called	100.0%	36.5%	0.0%
std::__1::tuple<kai::test::Buffer, kai::test::Buffer> kai::test::quantize_symmetric_per_block_dynamic<float, kai::test::Int4, float>(void const*, unsigned long, unsigned long, unsigned long) (test/reference/quantize.cpp:156)	called 2924 times	100.0%	-%	60.0%
std::__1::tuple<kai::test::Buffer, kai::test::Buffer> kai::test::quantize_symmetric_per_block_dynamic<float, kai::test::Int4, kai::test::BFloat16<false>>(void const*, unsigned long, unsigned long, unsigned long) (test/reference/quantize.cpp:156)	called 1404 times	100.0%	-%	60.0%
std::__1::tuple<kai::test::Buffer, kai::test::Buffer> kai::test::quantize_symmetric_per_block_dynamic<float, kai::test::Int4, kai::test::BFloat16<true>>(void const*, unsigned long, unsigned long, unsigned long) (test/reference/quantize.cpp:156)	not called	100.0%	-%	0.0%
std::__1::tuple<kai::test::Buffer, kai::test::Buffer> kai::test::quantize_symmetric_per_block_dynamic<float, kai::test::Int4, kai::test::Float16>(void const*, unsigned long, unsigned long, unsigned long) (test/reference/quantize.cpp:156)	called 124 times	100.0%	-%	60.0%
std::__1::tuple<kai::test::Buffer, kai::test::Buffer> kai::test::quantize_symmetric_per_block_dynamic<float, signed char, float>(void const*, unsigned long, unsigned long, unsigned long) (test/reference/quantize.cpp:156)	called 16239 times	100.0%	-%	60.0%
std::__1::tuple<kai::test::Buffer, kai::test::Buffer> kai::test::quantize_symmetric_per_block_dynamic<float, signed char, kai::test::Float16>(void const*, unsigned long, unsigned long, unsigned long) (test/reference/quantize.cpp:156)	called 124 times	100.0%	-%	60.0%
std::__1::tuple<kai::test::Float16, kai::test::Float16> kai::test::find_clamp_range<kai::test::Float16>(void const*, unsigned long, float) (test/reference/clamp.cpp:22)	not called	100.0%	50.0%	0.0%
unsigned char kai::test::read_array<unsigned char>(void const*, unsigned long) (test/common/memory.hpp:39)	called 8069376 times	100.0%	-%	100.0%
unsigned long kai::test::TupleHash::operator()<kai::test::MatMulShape, kai::test::DataFormat, kai::test::DataFormat, kai::test::DataFormat>(std::__1::tuple<kai::test::MatMulShape, kai::test::DataFormat, kai::test::DataFormat, kai::test::DataFormat> const&) const (test/common/cache.hpp:43)	called 1243 times	100.0%	-%	100.0%
unsigned long kai::test::round_to_nearest_even<unsigned long>(float) (test/common/round.cpp:49)	not called	0.0%	-%	0.0%
unsigned long long kai::test::get_sme_vector_length<1ul>() (test/common/sme.cpp:18)	not called	0.0%	0.0%	0.0%
unsigned long long kai::test::get_sme_vector_length<2ul>() (test/common/sme.cpp:41)	not called	0.0%	0.0%	0.0%
unsigned long long kai::test::get_sme_vector_length<4ul>() (test/common/sme.cpp:64)	called 172 times	87.5%	75.0%	83.0%
unsigned long long kai::test::get_sme_vector_length<float>() (test/common/sme.hpp:20)	called 146 times	100.0%	-%	100.0%
unsigned long long kai::test::get_sme_vector_length<int>() (test/common/sme.hpp:20)	called 26 times	100.0%	-%	100.0%
void const* kai::test::read_array<void const*>(void const*, unsigned long) (test/common/memory.hpp:39)	called 1130571 times	100.0%	-%	100.0%
void kai::benchmark::kai_benchmark_matmul<kai::benchmark::MatMulBaseInterface>(benchmark::State&, kai::benchmark::MatMulBaseInterface, kai::test::DataType, kai::benchmark::MatMulOp, std::__1::function<bool ()> const&) (benchmark/matmul/matmul_benchmark_logic.hpp:50)	called 18 times	100.0%	-%	50.0%
void kai::benchmark::kai_benchmark_matmul<kai::benchmark::MatMulBlockwiseDynamicQuantGenericDstInterface>(benchmark::State&, kai::benchmark::MatMulBlockwiseDynamicQuantGenericDstInterface, kai::test::DataType, kai::benchmark::MatMulOp, std::__1::function<bool ()> const&) (benchmark/matmul/matmul_benchmark_logic.hpp:50)	called 8 times	81.2%	40.7%	45.0%
void kai::benchmark::kai_benchmark_matmul<kai::benchmark::MatMulBlockwiseDynamicQuantInterface>(benchmark::State&, kai::benchmark::MatMulBlockwiseDynamicQuantInterface, kai::test::DataType, kai::benchmark::MatMulOp, std::__1::function<bool ()> const&) (benchmark/matmul/matmul_benchmark_logic.hpp:50)	called 24 times	100.0%	-%	45.0%
void kai::benchmark::kai_benchmark_matmul<kai::benchmark::MatMulFloatInterface>(benchmark::State&, kai::benchmark::MatMulFloatInterface, kai::test::DataType, kai::benchmark::MatMulOp, std::__1::function<bool ()> const&) (benchmark/matmul/matmul_benchmark_logic.hpp:50)	called 19 times	100.0%	-%	47.0%
void kai::benchmark::kai_benchmark_matmul<kai::benchmark::MatMulStaticQuantInterface>(benchmark::State&, kai::benchmark::MatMulStaticQuantInterface, kai::test::DataType, kai::benchmark::MatMulOp, std::__1::function<bool ()> const&) (benchmark/matmul/matmul_benchmark_logic.hpp:50)	called 3 times	100.0%	-%	47.0%
void kai::benchmark::kai_benchmark_matmul<kai::benchmark::MatMulStridedLhsInterface>(benchmark::State&, kai::benchmark::MatMulStridedLhsInterface, kai::test::DataType, kai::benchmark::MatMulOp, std::__1::function<bool ()> const&) (benchmark/matmul/matmul_benchmark_logic.hpp:50)	called 11 times	100.0%	-%	47.0%
void kai::test::detail::log<char const (&) [13], std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>>>(std::__1::basic_string_view<char, std::__1::char_traits<char>>, char const (&) [13], std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>>&&) (test/common/logging.hpp:62)	not called	100.0%	-%	0.0%
void kai::test::detail::log<char const (&) [17], unsigned long const&, char const (&) [12], float const&, char const (&) [13], float const&>(std::__1::basic_string_view<char, std::__1::char_traits<char>>, char const (&) [17], unsigned long const&, char const (&) [12], float const&, char const (&) [13], float const&) (test/common/logging.hpp:62)	not called	100.0%	-%	0.0%
void kai::test::detail::log<char const (&) [17], unsigned long const&, char const (&) [12], kai::test::BFloat16<true> const&, char const (&) [13], kai::test::BFloat16<true> const&>(std::__1::basic_string_view<char, std::__1::char_traits<char>>, char const (&) [17], unsigned long const&, char const (&) [12], kai::test::BFloat16<true> const&, char const (&) [13], kai::test::BFloat16<true> const&) (test/common/logging.hpp:62)	not called	100.0%	-%	0.0%
void kai::test::detail::log<char const (&) [17], unsigned long const&, char const (&) [12], kai::test::Float16 const&, char const (&) [13], kai::test::Float16 const&>(std::__1::basic_string_view<char, std::__1::char_traits<char>>, char const (&) [17], unsigned long const&, char const (&) [12], kai::test::Float16 const&, char const (&) [13], kai::test::Float16 const&) (test/common/logging.hpp:62)	not called	100.0%	-%	0.0%
void kai::test::detail::log<char const (&) [17], unsigned long const&, char const (&) [12], kai::test::Int4 const&, char const (&) [13], kai::test::Int4 const&>(std::__1::basic_string_view<char, std::__1::char_traits<char>>, char const (&) [17], unsigned long const&, char const (&) [12], kai::test::Int4 const&, char const (&) [13], kai::test::Int4 const&) (test/common/logging.hpp:62)	not called	0.0%	-%	0.0%
void kai::test::detail::log<char const (&) [17], unsigned long const&, char const (&) [12], signed char const&, char const (&) [13], signed char const&>(std::__1::basic_string_view<char, std::__1::char_traits<char>>, char const (&) [17], unsigned long const&, char const (&) [12], signed char const&, char const (&) [13], signed char const&) (test/common/logging.hpp:62)	not called	100.0%	-%	0.0%
void kai::test::detail::log<char const (&) [23], unsigned long const&, char const (&) [12], float const&, char const (&) [13], float const&>(std::__1::basic_string_view<char, std::__1::char_traits<char>>, char const (&) [23], unsigned long const&, char const (&) [12], float const&, char const (&) [13], float const&) (test/common/logging.hpp:62)	not called	100.0%	-%	0.0%
void kai::test::detail::log<char const (&) [23], unsigned long const&, char const (&) [12], int const&, char const (&) [13], int const&>(std::__1::basic_string_view<char, std::__1::char_traits<char>>, char const (&) [23], unsigned long const&, char const (&) [12], int const&, char const (&) [13], int const&) (test/common/logging.hpp:62)	not called	100.0%	-%	0.0%
void kai::test::detail::log<char const (&) [23], unsigned long const&, char const (&) [12], kai::test::Float16 const&, char const (&) [13], kai::test::Float16 const&>(std::__1::basic_string_view<char, std::__1::char_traits<char>>, char const (&) [23], unsigned long const&, char const (&) [12], kai::test::Float16 const&, char const (&) [13], kai::test::Float16 const&) (test/common/logging.hpp:62)	not called	100.0%	-%	0.0%
void kai::test::detail::log<char const (&) [31], unsigned long const&, char const (&) [12], float const&, char const (&) [13], float const&>(std::__1::basic_string_view<char, std::__1::char_traits<char>>, char const (&) [31], unsigned long const&, char const (&) [12], float const&, char const (&) [13], float const&) (test/common/logging.hpp:62)	not called	100.0%	-%	0.0%
void kai::test::detail::write_log_content<char const (&) [12], float const&, char const (&) [13], float const&>(char const (&) [12], float const&, char const (&) [13], float const&) (test/common/logging.hpp:50)	not called	100.0%	-%	0.0%
void kai::test::detail::write_log_content<char const (&) [12], int const&, char const (&) [13], int const&>(char const (&) [12], int const&, char const (&) [13], int const&) (test/common/logging.hpp:50)	not called	100.0%	-%	0.0%
void kai::test::detail::write_log_content<char const (&) [12], kai::test::BFloat16<true> const&, char const (&) [13], kai::test::BFloat16<true> const&>(char const (&) [12], kai::test::BFloat16<true> const&, char const (&) [13], kai::test::BFloat16<true> const&) (test/common/logging.hpp:50)	not called	100.0%	-%	0.0%
void kai::test::detail::write_log_content<char const (&) [12], kai::test::Float16 const&, char const (&) [13], kai::test::Float16 const&>(char const (&) [12], kai::test::Float16 const&, char const (&) [13], kai::test::Float16 const&) (test/common/logging.hpp:50)	not called	100.0%	-%	0.0%
void kai::test::detail::write_log_content<char const (&) [12], kai::test::Int4 const&, char const (&) [13], kai::test::Int4 const&>(char const (&) [12], kai::test::Int4 const&, char const (&) [13], kai::test::Int4 const&) (test/common/logging.hpp:50)	not called	100.0%	-%	0.0%
void kai::test::detail::write_log_content<char const (&) [12], signed char const&, char const (&) [13], signed char const&>(char const (&) [12], signed char const&, char const (&) [13], signed char const&) (test/common/logging.hpp:50)	not called	100.0%	-%	0.0%
void kai::test::detail::write_log_content<char const (&) [12]>(char const (&) [12]) (test/common/logging.hpp:26)	not called	100.0%	-%	0.0%
void kai::test::detail::write_log_content<char const (&) [13], float const&>(char const (&) [13], float const&) (test/common/logging.hpp:50)	not called	100.0%	-%	0.0%
void kai::test::detail::write_log_content<char const (&) [13], int const&>(char const (&) [13], int const&) (test/common/logging.hpp:50)	not called	100.0%	-%	0.0%
void kai::test::detail::write_log_content<char const (&) [13], kai::test::BFloat16<true> const&>(char const (&) [13], kai::test::BFloat16<true> const&) (test/common/logging.hpp:50)	not called	100.0%	-%	0.0%
void kai::test::detail::write_log_content<char const (&) [13], kai::test::Float16 const&>(char const (&) [13], kai::test::Float16 const&) (test/common/logging.hpp:50)	not called	100.0%	-%	0.0%
void kai::test::detail::write_log_content<char const (&) [13], kai::test::Int4 const&>(char const (&) [13], kai::test::Int4 const&) (test/common/logging.hpp:50)	not called	0.0%	-%	0.0%
void kai::test::detail::write_log_content<char const (&) [13], signed char const&>(char const (&) [13], signed char const&) (test/common/logging.hpp:50)	not called	100.0%	-%	0.0%
void kai::test::detail::write_log_content<char const (&) [13], std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>>>(char const (&) [13], std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>>&&) (test/common/logging.hpp:50)	not called	100.0%	-%	0.0%
void kai::test::detail::write_log_content<char const (&) [13]>(char const (&) [13]) (test/common/logging.hpp:26)	not called	100.0%	-%	0.0%
void kai::test::detail::write_log_content<char const (&) [17], unsigned long const&, char const (&) [12], float const&, char const (&) [13], float const&>(char const (&) [17], unsigned long const&, char const (&) [12], float const&, char const (&) [13], float const&) (test/common/logging.hpp:50)	not called	100.0%	-%	0.0%
void kai::test::detail::write_log_content<char const (&) [17], unsigned long const&, char const (&) [12], kai::test::BFloat16<true> const&, char const (&) [13], kai::test::BFloat16<true> const&>(char const (&) [17], unsigned long const&, char const (&) [12], kai::test::BFloat16<true> const&, char const (&) [13], kai::test::BFloat16<true> const&) (test/common/logging.hpp:50)	not called	100.0%	-%	0.0%
void kai::test::detail::write_log_content<char const (&) [17], unsigned long const&, char const (&) [12], kai::test::Float16 const&, char const (&) [13], kai::test::Float16 const&>(char const (&) [17], unsigned long const&, char const (&) [12], kai::test::Float16 const&, char const (&) [13], kai::test::Float16 const&) (test/common/logging.hpp:50)	not called	100.0%	-%	0.0%
void kai::test::detail::write_log_content<char const (&) [17], unsigned long const&, char const (&) [12], kai::test::Int4 const&, char const (&) [13], kai::test::Int4 const&>(char const (&) [17], unsigned long const&, char const (&) [12], kai::test::Int4 const&, char const (&) [13], kai::test::Int4 const&) (test/common/logging.hpp:50)	not called	100.0%	-%	0.0%
void kai::test::detail::write_log_content<char const (&) [17], unsigned long const&, char const (&) [12], signed char const&, char const (&) [13], signed char const&>(char const (&) [17], unsigned long const&, char const (&) [12], signed char const&, char const (&) [13], signed char const&) (test/common/logging.hpp:50)	not called	100.0%	-%	0.0%
void kai::test::detail::write_log_content<char const (&) [17]>(char const (&) [17]) (test/common/logging.hpp:26)	not called	100.0%	-%	0.0%
void kai::test::detail::write_log_content<char const (&) [23], unsigned long const&, char const (&) [12], float const&, char const (&) [13], float const&>(char const (&) [23], unsigned long const&, char const (&) [12], float const&, char const (&) [13], float const&) (test/common/logging.hpp:50)	not called	100.0%	-%	0.0%
void kai::test::detail::write_log_content<char const (&) [23], unsigned long const&, char const (&) [12], int const&, char const (&) [13], int const&>(char const (&) [23], unsigned long const&, char const (&) [12], int const&, char const (&) [13], int const&) (test/common/logging.hpp:50)	not called	100.0%	-%	0.0%
void kai::test::detail::write_log_content<char const (&) [23], unsigned long const&, char const (&) [12], kai::test::Float16 const&, char const (&) [13], kai::test::Float16 const&>(char const (&) [23], unsigned long const&, char const (&) [12], kai::test::Float16 const&, char const (&) [13], kai::test::Float16 const&) (test/common/logging.hpp:50)	not called	100.0%	-%	0.0%
void kai::test::detail::write_log_content<char const (&) [23]>(char const (&) [23]) (test/common/logging.hpp:26)	not called	100.0%	-%	0.0%
void kai::test::detail::write_log_content<char const (&) [31], unsigned long const&, char const (&) [12], float const&, char const (&) [13], float const&>(char const (&) [31], unsigned long const&, char const (&) [12], float const&, char const (&) [13], float const&) (test/common/logging.hpp:50)	not called	100.0%	-%	0.0%
void kai::test::detail::write_log_content<char const (&) [31]>(char const (&) [31]) (test/common/logging.hpp:26)	not called	100.0%	-%	0.0%
void kai::test::detail::write_log_content<float const&, char const (&) [13], float const&>(float const&, char const (&) [13], float const&) (test/common/logging.hpp:50)	not called	100.0%	-%	0.0%
void kai::test::detail::write_log_content<float const&>(float const&) (test/common/logging.hpp:26)	not called	100.0%	-%	0.0%
void kai::test::detail::write_log_content<int const&, char const (&) [13], int const&>(int const&, char const (&) [13], int const&) (test/common/logging.hpp:50)	not called	100.0%	-%	0.0%
void kai::test::detail::write_log_content<int const&>(int const&) (test/common/logging.hpp:26)	not called	100.0%	-%	0.0%
void kai::test::detail::write_log_content<kai::test::BFloat16<true> const&, char const (&) [13], kai::test::BFloat16<true> const&>(kai::test::BFloat16<true> const&, char const (&) [13], kai::test::BFloat16<true> const&) (test/common/logging.hpp:50)	not called	100.0%	-%	0.0%
void kai::test::detail::write_log_content<kai::test::BFloat16<true> const&>(kai::test::BFloat16<true> const&) (test/common/logging.hpp:26)	not called	100.0%	-%	0.0%
void kai::test::detail::write_log_content<kai::test::Float16 const&, char const (&) [13], kai::test::Float16 const&>(kai::test::Float16 const&, char const (&) [13], kai::test::Float16 const&) (test/common/logging.hpp:50)	not called	100.0%	-%	0.0%
void kai::test::detail::write_log_content<kai::test::Float16 const&>(kai::test::Float16 const&) (test/common/logging.hpp:26)	not called	100.0%	-%	0.0%
void kai::test::detail::write_log_content<kai::test::Int4 const&, char const (&) [13], kai::test::Int4 const&>(kai::test::Int4 const&, char const (&) [13], kai::test::Int4 const&) (test/common/logging.hpp:50)	not called	100.0%	-%	0.0%
void kai::test::detail::write_log_content<kai::test::Int4 const&>(kai::test::Int4 const&) (test/common/logging.hpp:26)	not called	0.0%	-%	0.0%
void kai::test::detail::write_log_content<signed char const&, char const (&) [13], signed char const&>(signed char const&, char const (&) [13], signed char const&) (test/common/logging.hpp:50)	not called	100.0%	-%	0.0%
void kai::test::detail::write_log_content<signed char const&>(signed char const&) (test/common/logging.hpp:26)	not called	100.0%	-%	0.0%
void kai::test::detail::write_log_content<std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>>>(std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>>&&) (test/common/logging.hpp:26)	not called	100.0%	-%	0.0%
void kai::test::detail::write_log_content<unsigned long const&, char const (&) [12], float const&, char const (&) [13], float const&>(unsigned long const&, char const (&) [12], float const&, char const (&) [13], float const&) (test/common/logging.hpp:50)	not called	100.0%	-%	0.0%
void kai::test::detail::write_log_content<unsigned long const&, char const (&) [12], int const&, char const (&) [13], int const&>(unsigned long const&, char const (&) [12], int const&, char const (&) [13], int const&) (test/common/logging.hpp:50)	not called	100.0%	-%	0.0%
void kai::test::detail::write_log_content<unsigned long const&, char const (&) [12], kai::test::BFloat16<true> const&, char const (&) [13], kai::test::BFloat16<true> const&>(unsigned long const&, char const (&) [12], kai::test::BFloat16<true> const&, char const (&) [13], kai::test::BFloat16<true> const&) (test/common/logging.hpp:50)	not called	100.0%	-%	0.0%
void kai::test::detail::write_log_content<unsigned long const&, char const (&) [12], kai::test::Float16 const&, char const (&) [13], kai::test::Float16 const&>(unsigned long const&, char const (&) [12], kai::test::Float16 const&, char const (&) [13], kai::test::Float16 const&) (test/common/logging.hpp:50)	not called	100.0%	-%	0.0%
void kai::test::detail::write_log_content<unsigned long const&, char const (&) [12], kai::test::Int4 const&, char const (&) [13], kai::test::Int4 const&>(unsigned long const&, char const (&) [12], kai::test::Int4 const&, char const (&) [13], kai::test::Int4 const&) (test/common/logging.hpp:50)	not called	100.0%	-%	0.0%
void kai::test::detail::write_log_content<unsigned long const&, char const (&) [12], signed char const&, char const (&) [13], signed char const&>(unsigned long const&, char const (&) [12], signed char const&, char const (&) [13], signed char const&) (test/common/logging.hpp:50)	not called	100.0%	-%	0.0%
void kai::test::detail::write_log_content<unsigned long const&>(unsigned long const&) (test/common/logging.hpp:26)	not called	100.0%	-%	0.0%
void kai::test::write_array<float>(void*, unsigned long, float) (test/common/memory.hpp:72)	called 256128150 times	100.0%	100.0%	100.0%
void kai::test::write_array<int>(void*, unsigned long, int) (test/common/memory.hpp:72)	called 1300182 times	100.0%	100.0%	100.0%
void kai::test::write_array<kai::test::BFloat16<false>>(void*, unsigned long, kai::test::BFloat16<false>) (test/common/memory.hpp:72)	called 4663826 times	100.0%	-%	100.0%
void kai::test::write_array<kai::test::BFloat16<true>>(void*, unsigned long, kai::test::BFloat16<true>) (test/common/memory.hpp:72)	called 4068988 times	100.0%	-%	100.0%
void kai::test::write_array<kai::test::Float16>(void*, unsigned long, kai::test::Float16) (test/common/memory.hpp:72)	called 40302159 times	100.0%	-%	100.0%
void kai::test::write_array<kai::test::Int4>(void*, unsigned long, kai::test::Int4) (test/common/memory.hpp:72)	called 65000348 times	17.6%	0.0%	100.0%
void kai::test::write_array<kai::test::UInt4>(void*, unsigned long, kai::test::UInt4) (test/common/memory.hpp:72)	called 77835480 times	100.0%	100.0%	100.0%
void kai::test::write_array<signed char>(void*, unsigned long, signed char) (test/common/memory.hpp:72)	called 100151175 times	100.0%	100.0%	100.0%
void kai::test::write_array<std::nullptr_t>(void*, unsigned long, std::nullptr_t) (test/common/memory.hpp:72)	not called	100.0%	100.0%	0.0%
void kai::test::write_array<unsigned char>(void*, unsigned long, unsigned char) (test/common/memory.hpp:72)	called 8069376 times	100.0%	-%	100.0%
void kai::test::write_array<void const*>(void*, unsigned long, void const*) (test/common/memory.hpp:72)	called 1130571 times	100.0%	-%	100.0%

---