| 1 |
// Auto-generated file. Do not edit! |
| 2 |
// Template: src/qs8-gemm/MRx8c8-avxvnni.c.in |
| 3 |
// Generator: tools/xngen |
| 4 |
// |
| 5 |
// Copyright 2024 Google LLC |
| 6 |
// |
| 7 |
// This source code is licensed under the BSD-style license found in the |
| 8 |
// LICENSE file in the root directory of this source tree. |
| 9 |
|
| 10 |
#include <assert.h> |
| 11 |
|
| 12 |
#include <immintrin.h> |
| 13 |
|
| 14 |
#include "xnnpack/gemm.h" |
| 15 |
#include "xnnpack/intrinsics-polyfill.h" |
| 16 |
#include "xnnpack/math.h" |
| 17 |
#include "xnnpack/unaligned.h" |
| 18 |
|
| 19 |
|
| 20 |
void xnn_qd8_f16_qc4w_gemm_minmax_ukernel_1x8c8__avxvnni( |
| 21 |
size_t mr, |
| 22 |
size_t nc, |
| 23 |
size_t kc, |
| 24 |
const int8_t* restrict a, |
| 25 |
size_t a_stride, |
| 26 |
const void* restrict w, |
| 27 |
xnn_float16* restrict c, |
| 28 |
size_t cm_stride, |
| 29 |
size_t cn_stride, |
| 30 |
const struct xnn_f16_qc4w_minmax_params params[restrict XNN_MIN_ELEMENTS(1)], |
| 31 |
const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS |
| 32 |
{ |
| 33 |
assert(mr != 0); |
| 34 |
assert(mr <= 1); |
| 35 |
assert(nc != 0); |
| 36 |
assert(kc != 0); |
| 37 |
assert(kc % sizeof(int8_t) == 0); |
| 38 |
assert(a != NULL); |
| 39 |
assert(w != NULL); |
| 40 |
assert(c != NULL); |
| 41 |
|
| 42 |
kc = round_up_po2(kc, 8 * sizeof(int8_t)); |
| 43 |
const int8_t* a0 = a; |
| 44 |
uint16_t* c0 = (uint16_t*) c; |
| 45 |
|
| 46 |
const __m256i vsign_mask = _mm256_set1_epi8(0x80); |
| 47 |
XNN_FORCE_REALIZATION(vsign_mask); |
| 48 |
const __m256i vinput_zero_point0 = _mm256_set1_epi32((int) quantization_params[0].zero_point + 128); |
| 49 |
const __m256 voutput_min = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) ¶ms->scalar.min)); |
| 50 |
const __m256 voutput_max = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) ¶ms->scalar.max)); |
| 51 |
// XNN_FORCE_REALIZATION(voutput_min); |
| 52 |
// XNN_FORCE_REALIZATION(voutput_max); |
| 53 |
const __m256i vmask = _mm256_set1_epi8(0xF0); |
| 54 |
XNN_FORCE_REALIZATION(vmask); |
| 55 |
do { |
| 56 |
const __m256i vksum01234567 = _mm256_load_si256(w); |
| 57 |
__m256i vsum0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point0); |
| 58 |
__m256i vacc0x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 0)); |
| 59 |
__m256i vacc0x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 1)); |
| 60 |
__m256i vacc1x0x0123 = _mm256_setzero_si256(); |
| 61 |
__m256i vacc1x0x4567 = _mm256_setzero_si256(); |
| 62 |
w = (const int32_t*) w + 8; |
| 63 |
|
| 64 |
size_t k = kc; |
| 65 |
while (k >= 16 * sizeof(int8_t)) { |
| 66 |
const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask); |
| 67 |
const __m256i va0x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask); |
| 68 |
a0 += 16; |
| 69 |
|
| 70 |
const __m256i vbb01234567x01234567 = _mm256_load_si256(w); |
| 71 |
const __m256i vbb89ABCDEFx01234567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32)); |
| 72 |
const __m256i vbs01234567x0123 = _mm256_slli_epi32(vbb01234567x01234567, 4); |
| 73 |
const __m256i vbs89ABCDEFx0123 = _mm256_slli_epi32(vbb89ABCDEFx01234567, 4); |
| 74 |
const __m256i vb01234567x4567 = _mm256_and_si256(vbb01234567x01234567, vmask); |
| 75 |
const __m256i vb89ABCDEFx4567 = _mm256_and_si256(vbb89ABCDEFx01234567, vmask); |
| 76 |
const __m256i vb01234567x0123 = _mm256_and_si256(vbs01234567x0123, vmask); |
| 77 |
const __m256i vb89ABCDEFx0123 = _mm256_and_si256(vbs89ABCDEFx0123, vmask); |
| 78 |
|
| 79 |
vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x01234567, vb01234567x0123); |
| 80 |
vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123); |
| 81 |
vacc1x0x0123 = _mm256_dpbusd_avx_epi32(vacc1x0x0123, va0x89ABCDEF, vb01234567x4567); |
| 82 |
vacc1x0x4567 = _mm256_dpbusd_avx_epi32(vacc1x0x4567, va0x89ABCDEF, vb89ABCDEFx4567); |
| 83 |
|
| 84 |
w = (const int8_t*) w + 64; |
| 85 |
k -= 16 * sizeof(int8_t); |
| 86 |
} |
| 87 |
|
| 88 |
if (k != 0) { |
| 89 |
const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask); |
| 90 |
a0 += 8; |
| 91 |
|
| 92 |
const __m256i vbb01234567x01234567 = _mm256_load_si256(w); |
| 93 |
const __m256i vbb89ABCDEFx01234567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32)); |
| 94 |
const __m256i vb01234567x0123 = _mm256_slli_epi32(vbb01234567x01234567, 4); |
| 95 |
const __m256i vb89ABCDEFx0123 = _mm256_slli_epi32(vbb89ABCDEFx01234567, 4); |
| 96 |
|
| 97 |
vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x01234567, vb01234567x0123); |
| 98 |
vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123); |
| 99 |
|
| 100 |
w = (const int8_t*) w + 64; |
| 101 |
k -= 8 * sizeof(int8_t); |
| 102 |
} |
| 103 |
vacc0x0123 = _mm256_add_epi32(vacc0x0123, vacc1x0x0123); |
| 104 |
vacc0x4567 = _mm256_add_epi32(vacc0x4567, vacc1x0x4567); |
| 105 |
|
| 106 |
// Add adjacent pairs |
| 107 |
const __m256i vsum0x02134657 = _mm256_hadd_epi32(vacc0x0123, vacc0x4567); |
| 108 |
__m256i vacc0x01234567 = _mm256_permute4x64_epi64(vsum0x02134657, _MM_SHUFFLE(3, 1, 2, 0)); |
| 109 |
|
| 110 |
vacc0x01234567 = _mm256_srai_epi32(vacc0x01234567, 4); |
| 111 |
__m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567); |
| 112 |
|
| 113 |
vout0x01234567 = _mm256_mul_ps(vout0x01234567, _mm256_set1_ps(quantization_params[0].inv_scale)); |
| 114 |
|
| 115 |
const __m256 vfilter_output_scale01234567 = _mm256_load_ps((const float*) w); |
| 116 |
const __m256 vbias01234567 = _mm256_load_ps((const float*) w + 8); |
| 117 |
w = (const float*) w + 16; |
| 118 |
|
| 119 |
vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vfilter_output_scale01234567, vbias01234567); |
| 120 |
|
| 121 |
vout0x01234567 = _mm256_max_ps(vout0x01234567, voutput_min); |
| 122 |
|
| 123 |
vout0x01234567 = _mm256_min_ps(vout0x01234567, voutput_max); |
| 124 |
|
| 125 |
__m128i vfp16out0x01234567 = _mm256_cvtps_ph(vout0x01234567, _MM_FROUND_TO_NEAREST_INT); |
| 126 |
if XNN_LIKELY(nc >= 8) { |
| 127 |
_mm_storeu_si128((__m128i*) c0, vfp16out0x01234567); |
| 128 |
a0 = (const int8_t*) ((uintptr_t) a0 - kc); |
| 129 |
c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride); |
| 130 |
nc -= 8; |
| 131 |
} else { |
| 132 |
if (nc & 4) { |
| 133 |
_mm_storel_epi64((__m128i*) c0, vfp16out0x01234567); |
| 134 |
c0 += 4; |
| 135 |
vfp16out0x01234567 = _mm_unpackhi_epi64(vfp16out0x01234567, vfp16out0x01234567); |
| 136 |
} |
| 137 |
if (nc & 2) { |
| 138 |
_mm_storeu_si32(c0, vfp16out0x01234567); |
| 139 |
c0 += 2; |
| 140 |
vfp16out0x01234567 = _mm_srli_epi64(vfp16out0x01234567, 32); |
| 141 |
} |
| 142 |
if (nc & 1) { |
| 143 |
*c0 = (uint16_t) _mm_extract_epi16(vfp16out0x01234567, 0); |
| 144 |
} |
| 145 |
nc = 0; |
| 146 |
} |
| 147 |
} while (nc != 0); |
| 148 |
} |