kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi4cxp8x8_1x8x32_neon_dotprod.c

Directory:	./
File:	kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi4cxp8x8_1x8x32_neon_dotprod.c
Date:	2025-10-20 13:18:31

	Coverage	Exec	Excl	Total
Lines:	97.6%	41	7	49
Functions:	100.0%	14	0	14
Branches:	50.0%	1	14	16

  
      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
    
      //
    
      #if !defined(__ARM_FEATURE_DOTPROD)
    
      #error "Dotprod extension required to compile this micro-kernel"
    
      #else  // Architectural features check.
    
      #include "kai_matmul_clamp_f32_qai8dxp1x8_qsi4cxp8x8_1x8x32_neon_dotprod.h"
    
      #include <arm_neon.h>
    
      #include <stddef.h>
    
      #include <stdint.h>
    
      #include "kai/kai_common.h"
    
      static const size_t kai_m_step = 1;
    
      static const size_t kai_n_step = 8;
    
      static const size_t kai_mr = 1;
    
      static const size_t kai_nr = 8;
    
      static const size_t kai_kr = 16;
    
      static const size_t kai_sr = 2;
    
      static const size_t kai_num_bytes_multiplier_lhs = sizeof(float);
    
      static const size_t kai_num_bytes_multiplier_rhs = sizeof(float);
    
      static const size_t kai_num_bytes_offset_lhs = sizeof(int32_t);
    
      static const size_t kai_num_bytes_sum_rhs = sizeof(int32_t);
    
      static const size_t kai_num_bytes_bias = sizeof(float);
    
      641
      inline static size_t kai_k_roundedup(size_t k) {
    
          // Since we pack a float and int32 value at the end of the row,
    
          // we must make sure that k is a multiple of 4 for alignment
    
      641
          size_t kr_sr_roundedup4 = kai_roundup(kai_kr * kai_sr, 4);
    
      1282
          return kai_roundup(k, kr_sr_roundedup4);
    
      641
      }
    
      240
      inline static size_t kai_lhs_packed_stride(size_t k) {
    
      240
          const size_t k_internal = kai_k_roundedup(k);
    
      −
          KAI_ASSERT((k_internal % 2) == 0);
    
      480
          return kai_mr * (k_internal * sizeof(int8_t) + kai_num_bytes_multiplier_lhs + kai_num_bytes_offset_lhs);
    
      240
      }
    
      240
      inline static size_t kai_rhs_packed_stride(size_t k) {
    
      240
          const size_t k_internal = kai_k_roundedup(k);
    
      −
          KAI_ASSERT((k_internal % 2) == 0);
    
      480
          return kai_nr * ((k_internal / 2) + kai_num_bytes_multiplier_rhs + kai_num_bytes_sum_rhs + kai_num_bytes_bias);
    
      240
      }
    
      320
      size_t kai_get_m_step_matmul_clamp_f32_qai8dxp1x8_qsi4cxp8x8_1x8x32_neon_dotprod(void) {
    
      320
          return kai_m_step;
    
      }
    
      320
      size_t kai_get_n_step_matmul_clamp_f32_qai8dxp1x8_qsi4cxp8x8_1x8x32_neon_dotprod(void) {
    
      320
          return kai_n_step;
    
      }
    
      240
      size_t kai_get_mr_matmul_clamp_f32_qai8dxp1x8_qsi4cxp8x8_1x8x32_neon_dotprod(void) {
    
      240
          return kai_mr;
    
      }
    
      240
      size_t kai_get_nr_matmul_clamp_f32_qai8dxp1x8_qsi4cxp8x8_1x8x32_neon_dotprod(void) {
    
      240
          return kai_nr;
    
      }
    
      320
      size_t kai_get_kr_matmul_clamp_f32_qai8dxp1x8_qsi4cxp8x8_1x8x32_neon_dotprod(void) {
    
      320
          return kai_kr;
    
      }
    
      320
      size_t kai_get_sr_matmul_clamp_f32_qai8dxp1x8_qsi4cxp8x8_1x8x32_neon_dotprod(void) {
    
      320
          return kai_sr;
    
      }
    
      240
      size_t kai_get_lhs_packed_offset_matmul_clamp_f32_qai8dxp1x8_qsi4cxp8x8_1x8x32_neon_dotprod(size_t m_idx, size_t k) {
    
      −
          KAI_ASSERT((m_idx % kai_m_step) == 0);
    
      240
          return (m_idx / kai_mr) * kai_lhs_packed_stride(k);
    
      }
    
      240
      size_t kai_get_rhs_packed_offset_matmul_clamp_f32_qai8dxp1x8_qsi4cxp8x8_1x8x32_neon_dotprod(size_t n_idx, size_t k) {
    
      −
          KAI_ASSERT((n_idx % kai_n_step) == 0);
    
      240
          return (n_idx / kai_nr) * kai_rhs_packed_stride(k);
    
      }
    
      160
      size_t kai_get_dst_offset_matmul_clamp_f32_qai8dxp1x8_qsi4cxp8x8_1x8x32_neon_dotprod(
    
          size_t m_idx, size_t n_idx, size_t dst_stride) {
    
      −
          KAI_ASSERT((m_idx % kai_m_step) == 0);
    
      −
          KAI_ASSERT((n_idx % kai_n_step) == 0);
    
      160
          return (n_idx * sizeof(float)) + m_idx * dst_stride;
    
      }
    
      160
      size_t kai_get_dst_size_matmul_clamp_f32_qai8dxp1x8_qsi4cxp8x8_1x8x32_neon_dotprod(size_t m, size_t n) {
    
      160
          return m * n * sizeof(float);
    
      }
    
      161
      void kai_run_matmul_clamp_f32_qai8dxp1x8_qsi4cxp8x8_1x8x32_neon_dotprod(
    
          size_t m, size_t n, size_t k, const void* restrict lhs_packed, const void* restrict rhs_packed,
    
          float* restrict dst,  // NOLINT(readability-non-const-parameter)
    
          size_t dst_stride_row, size_t dst_stride_col, float scalar_min, float scalar_max) {
    
      −
          KAI_ASSERT(dst_stride_col == sizeof(float));
    
        1/2✓ Branch 0 taken 161 times.
✗ Branch 1 not taken.

      161
          if (m == 0) {
    
      ✗
              return;
    
          }
    
      161
          const size_t k_internal = kai_k_roundedup(k);
    
      161
          size_t num_blocks = k_internal / 32;
    
      161
          float clamp_vals[2] = {scalar_min, scalar_max};
    
      322
          __asm__ __volatile__(
    
              "mov x26, #0x20\n"
    
              "mov x20, #0x8\n"
    
              "movi v5.16b, #0xf0\n"
    
              "mov x25, %x[m]\n"
    
              "madd x26, %x[num_blocks], x26, x20\n"
    
              "1:"  // Row loop
    
              "mov x24, %x[rhs_packed]\n"
    
              "mov x23, %x[n]\n"
    
              "add x22, %x[dst], %x[dst_stride_row]\n"
    
              "2:"  // Column loop
    
              "mov x21, %x[lhs_packed]\n"
    
              "movi v4.4s, #0x0\n"
    
              "movi v3.4s, #0x0\n"
    
              "mov x20, %x[num_blocks]\n"
    
              "movi v2.4s, #0x0\n"
    
              "movi v1.4s, #0x0\n"
    
              "3:"  // Sub block loop
    
              "ldr q0, [x24, #0x0]\n"
    
              "ldr q31, [x24, #0x10]\n"
    
              "subs x20, x20, #0x1\n"
    
              "ldr q30, [x24, #0x20]\n"
    
              "ldr q29, [x24, #0x30]\n"
    
              "ld1r { v28.2d }, [x21], #0x8\n"
    
              "ldr q27, [x24, #0x40]\n"
    
              "ldr q26, [x24, #0x50]\n"
    
              "ldr q25, [x24, #0x60]\n"
    
              "shl v24.16b, v0.16b, #0x4\n"
    
              "shl v18.16b, v31.16b, #0x4\n"
    
              "ldr q23, [x24, #0x70]\n"
    
              "shl v17.16b, v30.16b, #0x4\n"
    
              "shl v16.16b, v29.16b, #0x4\n"
    
              "add x24, x24, #0x80\n"
    
              "ld1r { v22.2d }, [x21], #0x8\n"
    
              "shl v21.16b, v27.16b, #0x4\n"
    
              "and v0.16b, v0.16b, v5.16b\n"
    
              "ld1r { v20.2d }, [x21], #0x8\n"
    
              "ld1r { v19.2d }, [x21], #0x8\n"
    
              ".inst 0x4e9c9704  // sdot v4.4s, v24.16b, v28.16b\n"
    
              ".inst 0x4e9c9643  // sdot v3.4s, v18.16b, v28.16b\n"
    
              "shl v18.16b, v26.16b, #0x4\n"
    
              ".inst 0x4e9c9622  // sdot v2.4s, v17.16b, v28.16b\n"
    
              ".inst 0x4e9c9601  // sdot v1.4s, v16.16b, v28.16b\n"
    
              "shl v17.16b, v25.16b, #0x4\n"
    
              "shl v16.16b, v23.16b, #0x4\n"
    
              "and v31.16b, v31.16b, v5.16b\n"
    
              "and v30.16b, v30.16b, v5.16b\n"
    
              "and v29.16b, v29.16b, v5.16b\n"
    
              ".inst 0x4e9696a4  // sdot v4.4s, v21.16b, v22.16b\n"
    
              ".inst 0x4e969643  // sdot v3.4s, v18.16b, v22.16b\n"
    
              "and v27.16b, v27.16b, v5.16b\n"
    
              ".inst 0x4e969622  // sdot v2.4s, v17.16b, v22.16b\n"
    
              ".inst 0x4e969601  // sdot v1.4s, v16.16b, v22.16b\n"
    
              "and v26.16b, v26.16b, v5.16b\n"
    
              "and v25.16b, v25.16b, v5.16b\n"
    
              "and v23.16b, v23.16b, v5.16b\n"
    
              ".inst 0x4e949404  // sdot v4.4s, v0.16b, v20.16b\n"
    
              ".inst 0x4e9497e3  // sdot v3.4s, v31.16b, v20.16b\n"
    
              ".inst 0x4e9497c2  // sdot v2.4s, v30.16b, v20.16b\n"
    
              ".inst 0x4e9497a1  // sdot v1.4s, v29.16b, v20.16b\n"
    
              ".inst 0x4e939764  // sdot v4.4s, v27.16b, v19.16b\n"
    
              ".inst 0x4e939743  // sdot v3.4s, v26.16b, v19.16b\n"
    
              ".inst 0x4e939722  // sdot v2.4s, v25.16b, v19.16b\n"
    
              ".inst 0x4e9396e1  // sdot v1.4s, v23.16b, v19.16b\n"
    
              "bgt 3b\n"
    
              "ldr q25, [x24, #0x0]\n"
    
              "ldr q24, [x24, #0x10]\n"
    
              "addp v4.4s, v4.4s, v3.4s\n"
    
              "addp v2.4s, v2.4s, v1.4s\n"
    
              "ld1r { v23.4s }, [x21]\n"
    
              "ldr q22, [x24, #0x20]\n"
    
              "add x21, x21, #0x4\n"
    
              "add x20, %x[clamp_vals], #0x4\n"
    
              "ld1r { v17.4s }, [x21]\n"
    
              "ldr q16, [x24, #0x30]\n"
    
              "cmp x23, #0x8\n"
    
              "ldr q21, [x24, #0x40]\n"
    
              "ldr q20, [x24, #0x50]\n"
    
              "add x24, x24, #0x60\n"
    
              "ld1r { v19.4s }, [%x[clamp_vals]]\n"
    
              "ld1r { v18.4s }, [x20]\n"
    
              "mla v4.4s, v25.4s, v23.s[0]\n"
    
              "mla v2.4s, v24.4s, v23.s[0]\n"
    
              "fmul v22.4s, v22.4s, v17.4s\n"
    
              "fmul v16.4s, v16.4s, v17.4s\n"
    
              "scvtf v4.4s, v4.4s\n"
    
              "scvtf v2.4s, v2.4s\n"
    
              "fmul v17.4s, v4.4s, v22.4s\n"
    
              "fmul v16.4s, v2.4s, v16.4s\n"
    
              "fadd v17.4s, v17.4s, v21.4s\n"
    
              "fadd v16.4s, v16.4s, v20.4s\n"
    
              "fmax v17.4s, v17.4s, v19.4s\n"
    
              "fmax v16.4s, v16.4s, v19.4s\n"
    
              "fmin v17.4s, v17.4s, v18.4s\n"
    
              "fmin v16.4s, v16.4s, v18.4s\n"
    
              "blt 4f\n"
    
              "str q17, [%x[dst], #0x0]\n"
    
              "str q16, [%x[dst], #0x10]\n"
    
              "b 9f\n"
    
              "4:"  // Partial output
    
              "mov x20, %x[dst]\n"
    
              "tbz x23, #2, 6f\n"
    
              "st1 { v17.4s }, [x20], #0x10\n"
    
              "tbz x23, #1, 5f\n"
    
              "st1 { v16.d }[0], [x20], #0x8\n"
    
              "tbz x23, #0, 8f\n"
    
              "st1 { v16.s }[2], [x20]\n"
    
              "b 8f\n"
    
              "5:"  // Output block 0: partial_1_4
    
              "tbz x23, #0, 8f\n"
    
              "st1 { v16.s }[0], [x20]\n"
    
              "b 8f\n"
    
              "6:"  // Output block 0: partial_2_0
    
              "tbz x23, #1, 7f\n"
    
              "st1 { v17.d }[0], [x20], #0x8\n"
    
              "tbz x23, #0, 8f\n"
    
              "st1 { v17.s }[2], [x20]\n"
    
              "b 8f\n"
    
              "7:"  // Output block 0: partial_1_0
    
              "st1 { v17.s }[0], [x20]\n"
    
              "8:"  // Output block 0: Done
    
              "9:"  // Stores done
    
              "subs x23, x23, #0x8\n"
    
              "add %x[dst], %x[dst], #0x20\n"
    
              "bgt 2b\n"
    
              "subs x25, x25, #0x1\n"
    
              "add %x[lhs_packed], %x[lhs_packed], x26\n"
    
              "mov %x[dst], x22\n"
    
              "bgt 1b\n"
    
              : [dst] "+&r"(dst), [lhs_packed] "+&r"(lhs_packed)
    
      161
              : [clamp_vals] "r"(clamp_vals), [dst_stride_row] "r"(dst_stride_row), [m] "r"(m), [n] "r"(n),
    
      161
                [num_blocks] "r"(num_blocks), [rhs_packed] "r"(rhs_packed)
    
              : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23",
    
                "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x20", "x21", "x22", "x23", "x24", "x25", "x26");
    
      161
      }
    
      #endif  // Architectural features check.

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
7			#if !defined(__ARM_FEATURE_DOTPROD)
8			#error "Dotprod extension required to compile this micro-kernel"
9			#else // Architectural features check.
10			#include "kai_matmul_clamp_f32_qai8dxp1x8_qsi4cxp8x8_1x8x32_neon_dotprod.h"
11
12			#include <arm_neon.h>
13			#include <stddef.h>
14			#include <stdint.h>
15
16			#include "kai/kai_common.h"
17
18			static const size_t kai_m_step = 1;
19			static const size_t kai_n_step = 8;
20			static const size_t kai_mr = 1;
21			static const size_t kai_nr = 8;
22			static const size_t kai_kr = 16;
23			static const size_t kai_sr = 2;
24			static const size_t kai_num_bytes_multiplier_lhs = sizeof(float);
25			static const size_t kai_num_bytes_multiplier_rhs = sizeof(float);
26			static const size_t kai_num_bytes_offset_lhs = sizeof(int32_t);
27			static const size_t kai_num_bytes_sum_rhs = sizeof(int32_t);
28			static const size_t kai_num_bytes_bias = sizeof(float);
29
30		641	inline static size_t kai_k_roundedup(size_t k) {
31			// Since we pack a float and int32 value at the end of the row,
32			// we must make sure that k is a multiple of 4 for alignment
33		641	size_t kr_sr_roundedup4 = kai_roundup(kai_kr * kai_sr, 4);
34		1282	return kai_roundup(k, kr_sr_roundedup4);
35		641	}
36
37		240	inline static size_t kai_lhs_packed_stride(size_t k) {
38		240	const size_t k_internal = kai_k_roundedup(k);
39
40		−	KAI_ASSERT((k_internal % 2) == 0);
41
42		480	return kai_mr * (k_internal * sizeof(int8_t) + kai_num_bytes_multiplier_lhs + kai_num_bytes_offset_lhs);
43		240	}
44
45		240	inline static size_t kai_rhs_packed_stride(size_t k) {
46		240	const size_t k_internal = kai_k_roundedup(k);
47
48		−	KAI_ASSERT((k_internal % 2) == 0);
49
50		480	return kai_nr * ((k_internal / 2) + kai_num_bytes_multiplier_rhs + kai_num_bytes_sum_rhs + kai_num_bytes_bias);
51		240	}
52
53		320	size_t kai_get_m_step_matmul_clamp_f32_qai8dxp1x8_qsi4cxp8x8_1x8x32_neon_dotprod(void) {
54		320	return kai_m_step;
55			}
56
57		320	size_t kai_get_n_step_matmul_clamp_f32_qai8dxp1x8_qsi4cxp8x8_1x8x32_neon_dotprod(void) {
58		320	return kai_n_step;
59			}
60
61		240	size_t kai_get_mr_matmul_clamp_f32_qai8dxp1x8_qsi4cxp8x8_1x8x32_neon_dotprod(void) {
62		240	return kai_mr;
63			}
64
65		240	size_t kai_get_nr_matmul_clamp_f32_qai8dxp1x8_qsi4cxp8x8_1x8x32_neon_dotprod(void) {
66		240	return kai_nr;
67			}
68
69		320	size_t kai_get_kr_matmul_clamp_f32_qai8dxp1x8_qsi4cxp8x8_1x8x32_neon_dotprod(void) {
70		320	return kai_kr;
71			}
72
73		320	size_t kai_get_sr_matmul_clamp_f32_qai8dxp1x8_qsi4cxp8x8_1x8x32_neon_dotprod(void) {
74		320	return kai_sr;
75			}
76
77		240	size_t kai_get_lhs_packed_offset_matmul_clamp_f32_qai8dxp1x8_qsi4cxp8x8_1x8x32_neon_dotprod(size_t m_idx, size_t k) {
78		−	KAI_ASSERT((m_idx % kai_m_step) == 0);
79
80		240	return (m_idx / kai_mr) * kai_lhs_packed_stride(k);
81			}
82
83		240	size_t kai_get_rhs_packed_offset_matmul_clamp_f32_qai8dxp1x8_qsi4cxp8x8_1x8x32_neon_dotprod(size_t n_idx, size_t k) {
84		−	KAI_ASSERT((n_idx % kai_n_step) == 0);
85
86		240	return (n_idx / kai_nr) * kai_rhs_packed_stride(k);
87			}
88
89		160	size_t kai_get_dst_offset_matmul_clamp_f32_qai8dxp1x8_qsi4cxp8x8_1x8x32_neon_dotprod(
90			size_t m_idx, size_t n_idx, size_t dst_stride) {
91		−	KAI_ASSERT((m_idx % kai_m_step) == 0);
92		−	KAI_ASSERT((n_idx % kai_n_step) == 0);
93
94		160	return (n_idx * sizeof(float)) + m_idx * dst_stride;
95			}
96
97		160	size_t kai_get_dst_size_matmul_clamp_f32_qai8dxp1x8_qsi4cxp8x8_1x8x32_neon_dotprod(size_t m, size_t n) {
98		160	return m * n * sizeof(float);
99			}
100
101		161	void kai_run_matmul_clamp_f32_qai8dxp1x8_qsi4cxp8x8_1x8x32_neon_dotprod(
102			size_t m, size_t n, size_t k, const void* restrict lhs_packed, const void* restrict rhs_packed,
103			float* restrict dst, // NOLINT(readability-non-const-parameter)
104			size_t dst_stride_row, size_t dst_stride_col, float scalar_min, float scalar_max) {
105		−	KAI_ASSERT(dst_stride_col == sizeof(float));
106
107	1/2 ✓ Branch 0 taken 161 times. ✗ Branch 1 not taken.	161	if (m == 0) {
108		✗	return;
109			}
110
111		161	const size_t k_internal = kai_k_roundedup(k);
112
113		161	size_t num_blocks = k_internal / 32;
114
115		161	float clamp_vals[2] = {scalar_min, scalar_max};
116		322	__asm__ __volatile__(
117			"mov x26, #0x20\n"
118			"mov x20, #0x8\n"
119			"movi v5.16b, #0xf0\n"
120			"mov x25, %x[m]\n"
121			"madd x26, %x[num_blocks], x26, x20\n"
122			"1:" // Row loop
123			"mov x24, %x[rhs_packed]\n"
124			"mov x23, %x[n]\n"
125			"add x22, %x[dst], %x[dst_stride_row]\n"
126			"2:" // Column loop
127			"mov x21, %x[lhs_packed]\n"
128			"movi v4.4s, #0x0\n"
129			"movi v3.4s, #0x0\n"
130			"mov x20, %x[num_blocks]\n"
131			"movi v2.4s, #0x0\n"
132			"movi v1.4s, #0x0\n"
133			"3:" // Sub block loop
134			"ldr q0, [x24, #0x0]\n"
135			"ldr q31, [x24, #0x10]\n"
136			"subs x20, x20, #0x1\n"
137			"ldr q30, [x24, #0x20]\n"
138			"ldr q29, [x24, #0x30]\n"
139			"ld1r { v28.2d }, [x21], #0x8\n"
140			"ldr q27, [x24, #0x40]\n"
141			"ldr q26, [x24, #0x50]\n"
142			"ldr q25, [x24, #0x60]\n"
143			"shl v24.16b, v0.16b, #0x4\n"
144			"shl v18.16b, v31.16b, #0x4\n"
145			"ldr q23, [x24, #0x70]\n"
146			"shl v17.16b, v30.16b, #0x4\n"
147			"shl v16.16b, v29.16b, #0x4\n"
148			"add x24, x24, #0x80\n"
149			"ld1r { v22.2d }, [x21], #0x8\n"
150			"shl v21.16b, v27.16b, #0x4\n"
151			"and v0.16b, v0.16b, v5.16b\n"
152			"ld1r { v20.2d }, [x21], #0x8\n"
153			"ld1r { v19.2d }, [x21], #0x8\n"
154			".inst 0x4e9c9704 // sdot v4.4s, v24.16b, v28.16b\n"
155			".inst 0x4e9c9643 // sdot v3.4s, v18.16b, v28.16b\n"
156			"shl v18.16b, v26.16b, #0x4\n"
157			".inst 0x4e9c9622 // sdot v2.4s, v17.16b, v28.16b\n"
158			".inst 0x4e9c9601 // sdot v1.4s, v16.16b, v28.16b\n"
159			"shl v17.16b, v25.16b, #0x4\n"
160			"shl v16.16b, v23.16b, #0x4\n"
161			"and v31.16b, v31.16b, v5.16b\n"
162			"and v30.16b, v30.16b, v5.16b\n"
163			"and v29.16b, v29.16b, v5.16b\n"
164			".inst 0x4e9696a4 // sdot v4.4s, v21.16b, v22.16b\n"
165			".inst 0x4e969643 // sdot v3.4s, v18.16b, v22.16b\n"
166			"and v27.16b, v27.16b, v5.16b\n"
167			".inst 0x4e969622 // sdot v2.4s, v17.16b, v22.16b\n"
168			".inst 0x4e969601 // sdot v1.4s, v16.16b, v22.16b\n"
169			"and v26.16b, v26.16b, v5.16b\n"
170			"and v25.16b, v25.16b, v5.16b\n"
171			"and v23.16b, v23.16b, v5.16b\n"
172			".inst 0x4e949404 // sdot v4.4s, v0.16b, v20.16b\n"
173			".inst 0x4e9497e3 // sdot v3.4s, v31.16b, v20.16b\n"
174			".inst 0x4e9497c2 // sdot v2.4s, v30.16b, v20.16b\n"
175			".inst 0x4e9497a1 // sdot v1.4s, v29.16b, v20.16b\n"
176			".inst 0x4e939764 // sdot v4.4s, v27.16b, v19.16b\n"
177			".inst 0x4e939743 // sdot v3.4s, v26.16b, v19.16b\n"
178			".inst 0x4e939722 // sdot v2.4s, v25.16b, v19.16b\n"
179			".inst 0x4e9396e1 // sdot v1.4s, v23.16b, v19.16b\n"
180			"bgt 3b\n"
181			"ldr q25, [x24, #0x0]\n"
182			"ldr q24, [x24, #0x10]\n"
183			"addp v4.4s, v4.4s, v3.4s\n"
184			"addp v2.4s, v2.4s, v1.4s\n"
185			"ld1r { v23.4s }, [x21]\n"
186			"ldr q22, [x24, #0x20]\n"
187			"add x21, x21, #0x4\n"
188			"add x20, %x[clamp_vals], #0x4\n"
189			"ld1r { v17.4s }, [x21]\n"
190			"ldr q16, [x24, #0x30]\n"
191			"cmp x23, #0x8\n"
192			"ldr q21, [x24, #0x40]\n"
193			"ldr q20, [x24, #0x50]\n"
194			"add x24, x24, #0x60\n"
195			"ld1r { v19.4s }, [%x[clamp_vals]]\n"
196			"ld1r { v18.4s }, [x20]\n"
197			"mla v4.4s, v25.4s, v23.s[0]\n"
198			"mla v2.4s, v24.4s, v23.s[0]\n"
199			"fmul v22.4s, v22.4s, v17.4s\n"
200			"fmul v16.4s, v16.4s, v17.4s\n"
201			"scvtf v4.4s, v4.4s\n"
202			"scvtf v2.4s, v2.4s\n"
203			"fmul v17.4s, v4.4s, v22.4s\n"
204			"fmul v16.4s, v2.4s, v16.4s\n"
205			"fadd v17.4s, v17.4s, v21.4s\n"
206			"fadd v16.4s, v16.4s, v20.4s\n"
207			"fmax v17.4s, v17.4s, v19.4s\n"
208			"fmax v16.4s, v16.4s, v19.4s\n"
209			"fmin v17.4s, v17.4s, v18.4s\n"
210			"fmin v16.4s, v16.4s, v18.4s\n"
211			"blt 4f\n"
212			"str q17, [%x[dst], #0x0]\n"
213			"str q16, [%x[dst], #0x10]\n"
214			"b 9f\n"
215			"4:" // Partial output
216			"mov x20, %x[dst]\n"
217			"tbz x23, #2, 6f\n"
218			"st1 { v17.4s }, [x20], #0x10\n"
219			"tbz x23, #1, 5f\n"
220			"st1 { v16.d }[0], [x20], #0x8\n"
221			"tbz x23, #0, 8f\n"
222			"st1 { v16.s }[2], [x20]\n"
223			"b 8f\n"
224			"5:" // Output block 0: partial_1_4
225			"tbz x23, #0, 8f\n"
226			"st1 { v16.s }[0], [x20]\n"
227			"b 8f\n"
228			"6:" // Output block 0: partial_2_0
229			"tbz x23, #1, 7f\n"
230			"st1 { v17.d }[0], [x20], #0x8\n"
231			"tbz x23, #0, 8f\n"
232			"st1 { v17.s }[2], [x20]\n"
233			"b 8f\n"
234			"7:" // Output block 0: partial_1_0
235			"st1 { v17.s }[0], [x20]\n"
236			"8:" // Output block 0: Done
237			"9:" // Stores done
238			"subs x23, x23, #0x8\n"
239			"add %x[dst], %x[dst], #0x20\n"
240			"bgt 2b\n"
241			"subs x25, x25, #0x1\n"
242			"add %x[lhs_packed], %x[lhs_packed], x26\n"
243			"mov %x[dst], x22\n"
244			"bgt 1b\n"
245			: [dst] "+&r"(dst), [lhs_packed] "+&r"(lhs_packed)
246		161	: [clamp_vals] "r"(clamp_vals), [dst_stride_row] "r"(dst_stride_row), [m] "r"(m), [n] "r"(n),
247		161	[num_blocks] "r"(num_blocks), [rhs_packed] "r"(rhs_packed)
248			: "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23",
249			"v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x20", "x21", "x22", "x23", "x24", "x25", "x26");
250		161	}
251
252			#endif // Architectural features check.
253