kai/ukernels/matmul/pack/kai_rhs_pack_kxn_qsi4c32p_qsu4c32s1s0.c

Directory:	./
Coverage:	low: ≥ 0% medium: ≥ 75.0% high: ≥ 90.0%

	Coverage	Exec / Excl / Total
Lines:	96.5%	109 / 36 / 149
Functions:	88.9%	8 / 0 / 9
Branches:	88.5%	23 / 72 / 98

    kai/ukernels/matmul/pack/kai_rhs_pack_kxn_qsi4c32p_qsu4c32s1s0.c
    
        Line
        Branch
        Exec
        Source
      
        //
      
        // SPDX-FileCopyrightText: Copyright 2024-2025 Arm Limited and/or its affiliates <open-source-office@arm.com>
      
        //
      
        // SPDX-License-Identifier: Apache-2.0
      
        //
      
        #include "kai_rhs_pack_kxn_qsi4c32p_qsu4c32s1s0.h"
      
        #include <stddef.h>
      
        #include <stdint.h>
      
        #include <string.h>
      
        #include "kai/kai_common.h"
      
        static const size_t kai_num_bytes_sum_rhs = sizeof(float);
      
        static const size_t kai_num_bytes_bias = sizeof(float);
      
        static const size_t kai_nr_multiple_of = 4;
      
        static const size_t kai_bl_multiple_of = 32;
      
        7676
        inline static size_t kai_get_num_blocks_per_row(size_t k, size_t bl) {
      
        −
            KAI_ASSERT((bl % kai_bl_multiple_of) == 0);
      
        7676
            return kai_roundup(k, bl) / bl;
      
        }
      
        7676
        inline static size_t kai_get_num_bytes_per_block(size_t bl, size_t num_bytes_multiplier_rhs) {
      
        −
            KAI_ASSERT((bl % kai_bl_multiple_of) == 0);
      
        7676
            return (bl / 2) + num_bytes_multiplier_rhs;
      
        }
      
        1108
        inline static size_t kai_get_rhs_packed_offset_end_of_all_blocks(
      
            size_t k, size_t nr, size_t kr, size_t bl, size_t num_bytes_multiplier_rhs) {
      
        −
            KAI_ASSERT((bl % kr) == 0);
      
        −
            KAI_ASSERT((nr % kai_nr_multiple_of) == 0);
      
        −
            KAI_ASSERT((bl % kai_bl_multiple_of) == 0);
      
        1108
            const size_t num_blocks_per_row = kai_get_num_blocks_per_row(k, bl);
      
        1108
            const size_t num_bytes_per_block = kai_get_num_bytes_per_block(bl, num_bytes_multiplier_rhs);
      
        2216
            return (nr * num_bytes_per_block * num_blocks_per_row);
      
        1108
        }
      
        ✗
        size_t kai_get_n_step_rhs_pack_kxn_qsi4c32p_qsu4c32s1s0(size_t nr) {
      
        ✗
            return nr;
      
        }
      
        1108
        size_t kai_get_rhs_offset_rhs_pack_kxn_qsi4c32p_qsu4c32s1s0(
      
            size_t n_idx,  //
      
            size_t rhs_stride) {
      
        1108
            KAI_UNUSED(rhs_stride);
      
        −
            KAI_ASSERT((n_idx % 2) == 0);
      
        1108
            return (n_idx / 2) * sizeof(int8_t);
      
        }
      
        5460
        size_t kai_get_rhs_packed_stride_rhs_pack_kxn_qsi4c32p_qsu4c32s1s0(
      
            size_t k,   //
      
            size_t nr,  //
      
            size_t kr,  //
      
            size_t sr,  //
      
            size_t bl,  //
      
            enum kai_datatype scale_dt) {
      
        −
            KAI_ASSERT((k % bl) == 0);
      
        −
            KAI_ASSERT((bl % kr) == 0);
      
        −
            KAI_ASSERT((nr % kai_nr_multiple_of) == 0);
      
        −
            KAI_ASSERT((bl % kai_bl_multiple_of) == 0);
      
        −
            KAI_ASSERT(scale_dt == kai_dt_bf16);
      
        5460
            KAI_UNUSED(kr);
      
        5460
            KAI_UNUSED(sr);
      
        5460
            const size_t num_bytes_multiplier_rhs = kai_get_datatype_size_in_bytes(scale_dt);
      
        5460
            const size_t num_blocks_per_row = kai_get_num_blocks_per_row(k, bl);
      
        5460
            const size_t num_bytes_per_block = kai_get_num_bytes_per_block(bl, num_bytes_multiplier_rhs);
      
        10920
            return nr * ((num_bytes_per_block * num_blocks_per_row) + kai_num_bytes_sum_rhs + kai_num_bytes_bias);
      
        5460
        }
      
        3244
        size_t kai_get_rhs_packed_offset_rhs_pack_kxn_qsi4c32p_qsu4c32s1s0(
      
            size_t n_idx,  //
      
            size_t k,      //
      
            size_t nr,     //
      
            size_t kr,     //
      
            size_t sr,     //
      
            size_t bl,     //
      
            enum kai_datatype scale_dt) {
      
        −
            KAI_ASSERT((n_idx % nr) == 0);
      
        −
            KAI_ASSERT((k % bl) == 0);
      
        −
            KAI_ASSERT((bl % kr) == 0);
      
        −
            KAI_ASSERT((nr % kai_nr_multiple_of) == 0);
      
        −
            KAI_ASSERT((bl % kai_bl_multiple_of) == 0);
      
        −
            KAI_ASSERT(scale_dt == kai_dt_bf16);
      
        3244
            return (n_idx / nr) * kai_get_rhs_packed_stride_rhs_pack_kxn_qsi4c32p_qsu4c32s1s0(k, nr, kr, sr, bl, scale_dt);
      
        }
      
        1108
        size_t kai_get_rhs_packed_size_rhs_pack_kxn_qsi4c32p_qsu4c32s1s0(
      
            size_t n,   //
      
            size_t k,   //
      
            size_t nr,  //
      
            size_t kr,  //
      
            size_t sr,  //
      
            size_t bl,  //
      
            enum kai_datatype scale_dt) {
      
        −
            KAI_ASSERT((k % bl) == 0);
      
        −
            KAI_ASSERT((bl % kr) == 0);
      
        −
            KAI_ASSERT((nr % kai_nr_multiple_of) == 0);
      
        −
            KAI_ASSERT((bl % kai_bl_multiple_of) == 0);
      
        −
            KAI_ASSERT(scale_dt == kai_dt_bf16);
      
        1108
            const size_t num_rows = kai_roundup(n, nr) / nr;
      
        2216
            return num_rows * kai_get_rhs_packed_stride_rhs_pack_kxn_qsi4c32p_qsu4c32s1s0(k, nr, kr, sr, bl, scale_dt);
      
        1108
        }
      
        1108
        void kai_run_rhs_pack_kxn_qsi4c32p_qsu4c32s1s0(
      
            size_t num_groups,    //
      
            size_t n,             //
      
            size_t k,             //
      
            size_t nr,            //
      
            size_t kr,            //
      
            size_t sr,            //
      
            size_t bl,            //
      
            const uint8_t* rhs,   //
      
            size_t rhs_stride,    //
      
            const float* bias,    //
      
            const void* scale,    //
      
            size_t scale_stride,  //
      
            void* rhs_packed,     //
      
            size_t extra_bytes,   //
      
            const struct kai_rhs_pack_kxn_qsi4c32p_qsu4c32s1s0_params* params) {
      
        −
            KAI_ASSERT(num_groups == 1);
      
        −
            KAI_ASSERT(extra_bytes == 0);
      
        −
            KAI_ASSERT(rhs != NULL);
      
        −
            KAI_ASSERT(scale != NULL);
      
        −
            KAI_ASSERT(rhs_packed != NULL);
      
        −
            KAI_ASSERT(params != NULL);
      
        −
            KAI_ASSERT(params->rhs_zero_point == 8);
      
        −
            KAI_ASSERT(params->lhs_zero_point == 1);
      
        −
            KAI_ASSERT((k % bl) == 0);
      
        −
            KAI_ASSERT((bl % kr) == 0);
      
        −
            KAI_ASSERT((kr % sr) == 0);
      
        −
            KAI_ASSERT((nr % kai_nr_multiple_of) == 0);
      
        −
            KAI_ASSERT((bl % kai_bl_multiple_of) == 0);
      
        −
            KAI_ASSERT(params->scale_dt == kai_dt_bf16);
      
            // Note: The input matrix (rhs) is expected with:
      
            // "n" columns and "k" rows (kxn)
      
        1108
            const size_t num_bytes_multiplier_rhs = kai_get_datatype_size_in_bytes(params->scale_dt);
      
        2216
            const size_t rhs_packed_stride =
      
        1108
                kai_get_rhs_packed_stride_rhs_pack_kxn_qsi4c32p_qsu4c32s1s0(k, nr, kr, sr, bl, params->scale_dt);
      
        2216
            const size_t rhs_packed_offset_end_of_all_blocks =
      
        1108
                kai_get_rhs_packed_offset_end_of_all_blocks(k, nr, kr, bl, num_bytes_multiplier_rhs);
      
        1108
            const size_t num_qblocks_per_row = kai_get_num_blocks_per_row(k, bl);
      
        1108
            const size_t num_bytes_per_block = kai_get_num_bytes_per_block(bl, num_bytes_multiplier_rhs);
      
        1108
            const size_t num_bytes_per_block_k = bl / 2;
      
        1108
            const size_t dst_num_rows = kai_roundup(n, nr) / nr;
      
        1108
            const size_t k_interleaved_v = 16U;
      
        1108
            const size_t block_length_in_bytes = kr / sr;
      
        1108
            const int32_t rhs_zero_point = params->rhs_zero_point;
      
          2/2✓ Branch 0 taken 1108 times.
✓ Branch 1 taken 17166 times.

        18274
            for (size_t dst_row_idx = 0; dst_row_idx < dst_num_rows; ++dst_row_idx) {
      
                // Before packing, it keeps the pointer to the first quantized block
      
        17166
                uint8_t* dst_row = (uint8_t*)rhs_packed + dst_row_idx * rhs_packed_stride;
      
        17166
                float* sums = (float*)(dst_row + rhs_packed_offset_end_of_all_blocks);
      
                // Initialize the RHS reduction sums to zero
      
        17166
                memset(sums, 0, nr * kai_num_bytes_sum_rhs);
      
                // Iterate over the quantized blocks
      
          2/2✓ Branch 0 taken 136876 times.
✓ Branch 1 taken 17166 times.

        154042
                for (size_t dst_qblock_idx = 0; dst_qblock_idx < num_qblocks_per_row; ++dst_qblock_idx) {
      
                    // Store the scales after packing all K values
      
        136876
                    uint8_t* rhs_packed_scale = dst_row + num_bytes_per_block_k * nr;
      
        136876
                    const uint8_t* scale_ptr = scale;
      
          2/2✓ Branch 0 taken 738160 times.
✓ Branch 1 taken 136876 times.

        875036
                    for (size_t i = 0; i < nr; ++i) {
      
          2/2✓ Branch 0 taken 727132 times.
✓ Branch 1 taken 11028 times.

        738160
                        const size_t src_row_idx = KAI_MIN(dst_row_idx * nr + i, n - 1);
      
        738160
                        void* dst_scales_ptr = rhs_packed_scale + i * num_bytes_multiplier_rhs;
      
        1476320
                        const void* src_scales_ptr = scale_ptr + dst_qblock_idx * num_bytes_multiplier_rhs +  //
      
        738160
                            (src_row_idx * scale_stride);                                                     //
      
        738160
                        memcpy(
      
        369080
                            dst_scales_ptr,             //
      
        369080
                            src_scales_ptr,             //
      
        369080
                            num_bytes_multiplier_rhs);  //
      
        738160
                    }
      
        136876
                    size_t kr_block_idx = 0;
      
          2/2✓ Branch 0 taken 2199520 times.
✓ Branch 1 taken 136876 times.

        2336396
                    for (size_t dst_byte_idx = 0; dst_byte_idx < nr * num_bytes_per_block_k;
      
        2199520
                         dst_byte_idx += block_length_in_bytes) {
      
        2199520
                        const size_t super_kr_block_idx = kr_block_idx / nr;
      
        2199520
                        const size_t nr_idx = kr_block_idx % nr;
      
        2199520
                        const size_t n0_idx = dst_row_idx * nr + nr_idx;
      
                        // Clamp the index to avoid out-of-bound reads
      
          2/2✓ Branch 0 taken 2167792 times.
✓ Branch 1 taken 31728 times.

        2199520
                        const size_t n0_valid_idx = KAI_MIN(n0_idx, n - 1);
      
        2199520
                        float d = kai_cast_f32_bf16(((uint16_t*)rhs_packed_scale)[nr_idx]);
      
        4399040
                        const size_t k_adjustment =
      
        2199520
                            ((super_kr_block_idx * block_length_in_bytes) / k_interleaved_v) * k_interleaved_v;
      
        2199520
                        size_t k0_idx = dst_qblock_idx * bl + super_kr_block_idx * block_length_in_bytes + k_adjustment;
      
        2199520
                        size_t k1_idx = k0_idx + k_interleaved_v;
      
        2199520
                        float partial_sum = 0.0F;
      
          2/2✓ Branch 0 taken 13077760 times.
✓ Branch 1 taken 2199520 times.

        15277280
                        for (size_t block_byte_idx = 0; block_byte_idx < block_length_in_bytes; ++block_byte_idx) {
      
        13077760
                            const size_t src_addr_byte0 = (n0_valid_idx / 2) + k0_idx * rhs_stride;
      
        13077760
                            const size_t src_addr_byte1 = (n0_valid_idx / 2) + k1_idx * rhs_stride;
      
        13077760
                            uint8_t byte0 = rhs_zero_point | rhs_zero_point << 4;
      
        13077760
                            uint8_t byte1 = rhs_zero_point | rhs_zero_point << 4;
      
          1/2✗ Branch 0 not taken.
✓ Branch 1 taken 13077760 times.

        13077760
                            if (k0_idx < k) {
      
        13077760
                                byte0 = rhs[src_addr_byte0];
      
        13077760
                            }
      
          1/2✗ Branch 0 not taken.
✓ Branch 1 taken 13077760 times.

        13077760
                            if (k1_idx < k) {
      
        13077760
                                byte1 = rhs[src_addr_byte1];
      
        13077760
                            }
      
          2/2✓ Branch 0 taken 6538880 times.
✓ Branch 1 taken 6538880 times.

        13077760
                            if ((n0_idx % 2) == 0) {
      
        6538880
                                const uint8_t src_x0_lo = (byte0 & 0x0F);
      
        6538880
                                const uint8_t src_x0_hi = (byte1 & 0x0F);
      
        6538880
                                partial_sum += (float)((int32_t)src_x0_lo + (int32_t)src_x0_hi - 2 * rhs_zero_point) * d;
      
        6538880
                                const uint8_t dst_qs0 = src_x0_lo | (src_x0_hi << 4);
      
        6538880
                                dst_row[dst_byte_idx + block_byte_idx] = dst_qs0 ^ 0x88;
      
        6538880
                            } else {
      
        6538880
                                const uint8_t src_x1_lo = (byte0 >> 4);
      
        6538880
                                const uint8_t src_x1_hi = (byte1 >> 4);
      
        6538880
                                partial_sum += (float)((int32_t)src_x1_lo + (int32_t)src_x1_hi - 2 * rhs_zero_point) * d;
      
        6538880
                                const uint8_t dst_qs1 = src_x1_lo | (src_x1_hi << 4);
      
        6538880
                                dst_row[dst_byte_idx + block_byte_idx] = dst_qs1 ^ 0x88;
      
        6538880
                            }
      
        13077760
                            k0_idx++;
      
        13077760
                            k1_idx++;
      
        13077760
                        }
      
        2199520
                        sums[nr_idx] += partial_sum;
      
                        // Increment the Kr block index
      
        2199520
                        kr_block_idx++;
      
        2199520
                    }
      
                    // Move the pointer after K values
      
        136876
                    dst_row += num_bytes_per_block * nr;
      
        136876
                }
      
                // Move the pointer after the row sum
      
        17166
                dst_row += kai_num_bytes_sum_rhs * nr;
      
                // Set the bias
      
          1/2✓ Branch 0 taken 17166 times.
✗ Branch 1 not taken.

        17166
                if (bias == NULL) {
      
        ✗
                    memset(dst_row, 0, nr * kai_num_bytes_bias);
      
        ✗
                } else {
      
          2/2✓ Branch 0 taken 92088 times.
✓ Branch 1 taken 17166 times.

        109254
                    for (size_t i = 0; i < nr; ++i) {
      
                        // Clamp the row index to avoid out-of-bound reads
      
          2/2✓ Branch 0 taken 90364 times.
✓ Branch 1 taken 1724 times.

        92088
                        const size_t src_row_idx = KAI_MIN(dst_row_idx * nr + i, n - 1);
      
        92088
                        ((float*)dst_row)[i] = bias[src_row_idx];
      
        92088
                    }
      
                }
      
        17166
            }
      
        1108
        }

Line	Branch	Exec	Source
1			//
2			// SPDX-FileCopyrightText: Copyright 2024-2025 Arm Limited and/or its affiliates <open-source-office@arm.com>
3			//
4			// SPDX-License-Identifier: Apache-2.0
5			//
6			#include "kai_rhs_pack_kxn_qsi4c32p_qsu4c32s1s0.h"
7
8			#include <stddef.h>
9			#include <stdint.h>
10			#include <string.h>
11
12			#include "kai/kai_common.h"
13
14			static const size_t kai_num_bytes_sum_rhs = sizeof(float);
15			static const size_t kai_num_bytes_bias = sizeof(float);
16			static const size_t kai_nr_multiple_of = 4;
17			static const size_t kai_bl_multiple_of = 32;
18
19		7676	inline static size_t kai_get_num_blocks_per_row(size_t k, size_t bl) {
20		−	KAI_ASSERT((bl % kai_bl_multiple_of) == 0);
21		7676	return kai_roundup(k, bl) / bl;
22			}
23
24		7676	inline static size_t kai_get_num_bytes_per_block(size_t bl, size_t num_bytes_multiplier_rhs) {
25		−	KAI_ASSERT((bl % kai_bl_multiple_of) == 0);
26		7676	return (bl / 2) + num_bytes_multiplier_rhs;
27			}
28
29		1108	inline static size_t kai_get_rhs_packed_offset_end_of_all_blocks(
30			size_t k, size_t nr, size_t kr, size_t bl, size_t num_bytes_multiplier_rhs) {
31		−	KAI_ASSERT((bl % kr) == 0);
32		−	KAI_ASSERT((nr % kai_nr_multiple_of) == 0);
33		−	KAI_ASSERT((bl % kai_bl_multiple_of) == 0);
34
35		1108	const size_t num_blocks_per_row = kai_get_num_blocks_per_row(k, bl);
36		1108	const size_t num_bytes_per_block = kai_get_num_bytes_per_block(bl, num_bytes_multiplier_rhs);
37
38		2216	return (nr * num_bytes_per_block * num_blocks_per_row);
39		1108	}
40
41		✗	size_t kai_get_n_step_rhs_pack_kxn_qsi4c32p_qsu4c32s1s0(size_t nr) {
42		✗	return nr;
43			}
44
45		1108	size_t kai_get_rhs_offset_rhs_pack_kxn_qsi4c32p_qsu4c32s1s0(
46			size_t n_idx, //
47			size_t rhs_stride) {
48		1108	KAI_UNUSED(rhs_stride);
49		−	KAI_ASSERT((n_idx % 2) == 0);
50		1108	return (n_idx / 2) * sizeof(int8_t);
51			}
52
53		5460	size_t kai_get_rhs_packed_stride_rhs_pack_kxn_qsi4c32p_qsu4c32s1s0(
54			size_t k, //
55			size_t nr, //
56			size_t kr, //
57			size_t sr, //
58			size_t bl, //
59			enum kai_datatype scale_dt) {
60		−	KAI_ASSERT((k % bl) == 0);
61		−	KAI_ASSERT((bl % kr) == 0);
62		−	KAI_ASSERT((nr % kai_nr_multiple_of) == 0);
63		−	KAI_ASSERT((bl % kai_bl_multiple_of) == 0);
64		−	KAI_ASSERT(scale_dt == kai_dt_bf16);
65
66		5460	KAI_UNUSED(kr);
67		5460	KAI_UNUSED(sr);
68
69		5460	const size_t num_bytes_multiplier_rhs = kai_get_datatype_size_in_bytes(scale_dt);
70		5460	const size_t num_blocks_per_row = kai_get_num_blocks_per_row(k, bl);
71		5460	const size_t num_bytes_per_block = kai_get_num_bytes_per_block(bl, num_bytes_multiplier_rhs);
72
73		10920	return nr * ((num_bytes_per_block * num_blocks_per_row) + kai_num_bytes_sum_rhs + kai_num_bytes_bias);
74		5460	}
75
76		3244	size_t kai_get_rhs_packed_offset_rhs_pack_kxn_qsi4c32p_qsu4c32s1s0(
77			size_t n_idx, //
78			size_t k, //
79			size_t nr, //
80			size_t kr, //
81			size_t sr, //
82			size_t bl, //
83			enum kai_datatype scale_dt) {
84		−	KAI_ASSERT((n_idx % nr) == 0);
85		−	KAI_ASSERT((k % bl) == 0);
86		−	KAI_ASSERT((bl % kr) == 0);
87		−	KAI_ASSERT((nr % kai_nr_multiple_of) == 0);
88		−	KAI_ASSERT((bl % kai_bl_multiple_of) == 0);
89		−	KAI_ASSERT(scale_dt == kai_dt_bf16);
90
91		3244	return (n_idx / nr) * kai_get_rhs_packed_stride_rhs_pack_kxn_qsi4c32p_qsu4c32s1s0(k, nr, kr, sr, bl, scale_dt);
92			}
93
94		1108	size_t kai_get_rhs_packed_size_rhs_pack_kxn_qsi4c32p_qsu4c32s1s0(
95			size_t n, //
96			size_t k, //
97			size_t nr, //
98			size_t kr, //
99			size_t sr, //
100			size_t bl, //
101			enum kai_datatype scale_dt) {
102		−	KAI_ASSERT((k % bl) == 0);
103		−	KAI_ASSERT((bl % kr) == 0);
104		−	KAI_ASSERT((nr % kai_nr_multiple_of) == 0);
105		−	KAI_ASSERT((bl % kai_bl_multiple_of) == 0);
106		−	KAI_ASSERT(scale_dt == kai_dt_bf16);
107
108		1108	const size_t num_rows = kai_roundup(n, nr) / nr;
109
110		2216	return num_rows * kai_get_rhs_packed_stride_rhs_pack_kxn_qsi4c32p_qsu4c32s1s0(k, nr, kr, sr, bl, scale_dt);
111		1108	}
112
113		1108	void kai_run_rhs_pack_kxn_qsi4c32p_qsu4c32s1s0(
114			size_t num_groups, //
115			size_t n, //
116			size_t k, //
117			size_t nr, //
118			size_t kr, //
119			size_t sr, //
120			size_t bl, //
121			const uint8_t* rhs, //
122			size_t rhs_stride, //
123			const float* bias, //
124			const void* scale, //
125			size_t scale_stride, //
126			void* rhs_packed, //
127			size_t extra_bytes, //
128			const struct kai_rhs_pack_kxn_qsi4c32p_qsu4c32s1s0_params* params) {
129		−	KAI_ASSERT(num_groups == 1);
130		−	KAI_ASSERT(extra_bytes == 0);
131		−	KAI_ASSERT(rhs != NULL);
132		−	KAI_ASSERT(scale != NULL);
133		−	KAI_ASSERT(rhs_packed != NULL);
134		−	KAI_ASSERT(params != NULL);
135		−	KAI_ASSERT(params->rhs_zero_point == 8);
136		−	KAI_ASSERT(params->lhs_zero_point == 1);
137
138		−	KAI_ASSERT((k % bl) == 0);
139		−	KAI_ASSERT((bl % kr) == 0);
140		−	KAI_ASSERT((kr % sr) == 0);
141		−	KAI_ASSERT((nr % kai_nr_multiple_of) == 0);
142		−	KAI_ASSERT((bl % kai_bl_multiple_of) == 0);
143		−	KAI_ASSERT(params->scale_dt == kai_dt_bf16);
144
145			// Note: The input matrix (rhs) is expected with:
146			// "n" columns and "k" rows (kxn)
147
148		1108	const size_t num_bytes_multiplier_rhs = kai_get_datatype_size_in_bytes(params->scale_dt);
149		2216	const size_t rhs_packed_stride =
150		1108	kai_get_rhs_packed_stride_rhs_pack_kxn_qsi4c32p_qsu4c32s1s0(k, nr, kr, sr, bl, params->scale_dt);
151		2216	const size_t rhs_packed_offset_end_of_all_blocks =
152		1108	kai_get_rhs_packed_offset_end_of_all_blocks(k, nr, kr, bl, num_bytes_multiplier_rhs);
153		1108	const size_t num_qblocks_per_row = kai_get_num_blocks_per_row(k, bl);
154		1108	const size_t num_bytes_per_block = kai_get_num_bytes_per_block(bl, num_bytes_multiplier_rhs);
155		1108	const size_t num_bytes_per_block_k = bl / 2;
156		1108	const size_t dst_num_rows = kai_roundup(n, nr) / nr;
157		1108	const size_t k_interleaved_v = 16U;
158		1108	const size_t block_length_in_bytes = kr / sr;
159
160		1108	const int32_t rhs_zero_point = params->rhs_zero_point;
161
162	2/2 ✓ Branch 0 taken 1108 times. ✓ Branch 1 taken 17166 times.	18274	for (size_t dst_row_idx = 0; dst_row_idx < dst_num_rows; ++dst_row_idx) {
163			// Before packing, it keeps the pointer to the first quantized block
164		17166	uint8_t* dst_row = (uint8_t)rhs_packed + dst_row_idx rhs_packed_stride;
165
166		17166	float* sums = (float*)(dst_row + rhs_packed_offset_end_of_all_blocks);
167
168			// Initialize the RHS reduction sums to zero
169		17166	memset(sums, 0, nr * kai_num_bytes_sum_rhs);
170
171			// Iterate over the quantized blocks
172	2/2 ✓ Branch 0 taken 136876 times. ✓ Branch 1 taken 17166 times.	154042	for (size_t dst_qblock_idx = 0; dst_qblock_idx < num_qblocks_per_row; ++dst_qblock_idx) {
173			// Store the scales after packing all K values
174		136876	uint8_t* rhs_packed_scale = dst_row + num_bytes_per_block_k * nr;
175		136876	const uint8_t* scale_ptr = scale;
176
177	2/2 ✓ Branch 0 taken 738160 times. ✓ Branch 1 taken 136876 times.	875036	for (size_t i = 0; i < nr; ++i) {
178	2/2 ✓ Branch 0 taken 727132 times. ✓ Branch 1 taken 11028 times.	738160	const size_t src_row_idx = KAI_MIN(dst_row_idx * nr + i, n - 1);
179
180		738160	void* dst_scales_ptr = rhs_packed_scale + i * num_bytes_multiplier_rhs;
181		1476320	const void* src_scales_ptr = scale_ptr + dst_qblock_idx * num_bytes_multiplier_rhs + //
182		738160	(src_row_idx * scale_stride); //
183
184		738160	memcpy(
185		369080	dst_scales_ptr, //
186		369080	src_scales_ptr, //
187		369080	num_bytes_multiplier_rhs); //
188		738160	}
189
190		136876	size_t kr_block_idx = 0;
191	2/2 ✓ Branch 0 taken 2199520 times. ✓ Branch 1 taken 136876 times.	2336396	for (size_t dst_byte_idx = 0; dst_byte_idx < nr * num_bytes_per_block_k;
192		2199520	dst_byte_idx += block_length_in_bytes) {
193		2199520	const size_t super_kr_block_idx = kr_block_idx / nr;
194		2199520	const size_t nr_idx = kr_block_idx % nr;
195		2199520	const size_t n0_idx = dst_row_idx * nr + nr_idx;
196
197			// Clamp the index to avoid out-of-bound reads
198	2/2 ✓ Branch 0 taken 2167792 times. ✓ Branch 1 taken 31728 times.	2199520	const size_t n0_valid_idx = KAI_MIN(n0_idx, n - 1);
199
200		2199520	float d = kai_cast_f32_bf16(((uint16_t*)rhs_packed_scale)[nr_idx]);
201
202		4399040	const size_t k_adjustment =
203		2199520	((super_kr_block_idx * block_length_in_bytes) / k_interleaved_v) * k_interleaved_v;
204		2199520	size_t k0_idx = dst_qblock_idx * bl + super_kr_block_idx * block_length_in_bytes + k_adjustment;
205		2199520	size_t k1_idx = k0_idx + k_interleaved_v;
206
207		2199520	float partial_sum = 0.0F;
208
209	2/2 ✓ Branch 0 taken 13077760 times. ✓ Branch 1 taken 2199520 times.	15277280	for (size_t block_byte_idx = 0; block_byte_idx < block_length_in_bytes; ++block_byte_idx) {
210		13077760	const size_t src_addr_byte0 = (n0_valid_idx / 2) + k0_idx * rhs_stride;
211		13077760	const size_t src_addr_byte1 = (n0_valid_idx / 2) + k1_idx * rhs_stride;
212
213		13077760	uint8_t byte0 = rhs_zero_point \| rhs_zero_point << 4;
214		13077760	uint8_t byte1 = rhs_zero_point \| rhs_zero_point << 4;
215
216	1/2 ✗ Branch 0 not taken. ✓ Branch 1 taken 13077760 times.	13077760	if (k0_idx < k) {
217		13077760	byte0 = rhs[src_addr_byte0];
218		13077760	}
219
220	1/2 ✗ Branch 0 not taken. ✓ Branch 1 taken 13077760 times.	13077760	if (k1_idx < k) {
221		13077760	byte1 = rhs[src_addr_byte1];
222		13077760	}
223
224	2/2 ✓ Branch 0 taken 6538880 times. ✓ Branch 1 taken 6538880 times.	13077760	if ((n0_idx % 2) == 0) {
225		6538880	const uint8_t src_x0_lo = (byte0 & 0x0F);
226		6538880	const uint8_t src_x0_hi = (byte1 & 0x0F);
227
228		6538880	partial_sum += (float)((int32_t)src_x0_lo + (int32_t)src_x0_hi - 2 * rhs_zero_point) * d;
229
230		6538880	const uint8_t dst_qs0 = src_x0_lo \| (src_x0_hi << 4);
231
232		6538880	dst_row[dst_byte_idx + block_byte_idx] = dst_qs0 ^ 0x88;
233		6538880	} else {
234		6538880	const uint8_t src_x1_lo = (byte0 >> 4);
235		6538880	const uint8_t src_x1_hi = (byte1 >> 4);
236
237		6538880	partial_sum += (float)((int32_t)src_x1_lo + (int32_t)src_x1_hi - 2 * rhs_zero_point) * d;
238
239		6538880	const uint8_t dst_qs1 = src_x1_lo \| (src_x1_hi << 4);
240
241		6538880	dst_row[dst_byte_idx + block_byte_idx] = dst_qs1 ^ 0x88;
242		6538880	}
243		13077760	k0_idx++;
244		13077760	k1_idx++;
245		13077760	}
246		2199520	sums[nr_idx] += partial_sum;
247
248			// Increment the Kr block index
249		2199520	kr_block_idx++;
250		2199520	}
251			// Move the pointer after K values
252		136876	dst_row += num_bytes_per_block * nr;
253		136876	}
254
255			// Move the pointer after the row sum
256		17166	dst_row += kai_num_bytes_sum_rhs * nr;
257
258			// Set the bias
259	1/2 ✓ Branch 0 taken 17166 times. ✗ Branch 1 not taken.	17166	if (bias == NULL) {
260		✗	memset(dst_row, 0, nr * kai_num_bytes_bias);
261		✗	} else {
262	2/2 ✓ Branch 0 taken 92088 times. ✓ Branch 1 taken 17166 times.	109254	for (size_t i = 0; i < nr; ++i) {
263			// Clamp the row index to avoid out-of-bound reads
264	2/2 ✓ Branch 0 taken 90364 times. ✓ Branch 1 taken 1724 times.	92088	const size_t src_row_idx = KAI_MIN(dst_row_idx * nr + i, n - 1);
265		92088	((float*)dst_row)[i] = bias[src_row_idx];
266		92088	}
267			}
268		17166	}
269		1108	}
270

KleidiAI Coverage Report