| 1 |
// Auto-generated file. Do not edit! |
| 2 |
// Template: src/qs8-gemm/MRx4c8-sse.c.in |
| 3 |
// Generator: tools/xngen |
| 4 |
// |
| 5 |
// Copyright 2020 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 <smmintrin.h> |
| 13 |
|
| 14 |
#include "xnnpack/common.h" |
| 15 |
#include "xnnpack/gemm.h" |
| 16 |
#include "xnnpack/math.h" |
| 17 |
|
| 18 |
|
| 19 |
void xnn_qd8_f32_qc8w_gemm_minmax_ukernel_1x4c8__sse41_ld128( |
| 20 |
size_t mr, |
| 21 |
size_t nc, |
| 22 |
size_t kc, |
| 23 |
const int8_t* restrict a, |
| 24 |
size_t a_stride, |
| 25 |
const void* restrict w, |
| 26 |
float* restrict c, |
| 27 |
size_t cm_stride, |
| 28 |
size_t cn_stride, |
| 29 |
const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)], |
| 30 |
const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS |
| 31 |
{ |
| 32 |
assert(mr != 0); |
| 33 |
assert(mr <= 1); |
| 34 |
assert(nc != 0); |
| 35 |
assert(kc != 0); |
| 36 |
assert(kc % sizeof(int8_t) == 0); |
| 37 |
assert(a != NULL); |
| 38 |
assert(w != NULL); |
| 39 |
assert(c != NULL); |
| 40 |
|
| 41 |
kc = round_up_po2(kc, 8 * sizeof(int8_t)); |
| 42 |
const int8_t* a0 = a; |
| 43 |
float* c0 = c; |
| 44 |
|
| 45 |
const __m128 vmin = _mm_set1_ps(params->scalar.min); |
| 46 |
const __m128 vmax = _mm_set1_ps(params->scalar.max); |
| 47 |
XNN_FORCE_REALIZATION(vmin); |
| 48 |
XNN_FORCE_REALIZATION(vmax); |
| 49 |
|
| 50 |
|
| 51 |
do { |
| 52 |
const __m128i vksum = _mm_load_si128((const __m128i*) w); |
| 53 |
__m128i vinput_zero_point0 = _mm_cvtsi32_si128(*((const int*) &quantization_params[0].zero_point)); |
| 54 |
vinput_zero_point0 = _mm_shuffle_epi32(vinput_zero_point0, _MM_SHUFFLE(0, 0, 0, 0)); |
| 55 |
const __m128i vzero = _mm_setzero_si128(); |
| 56 |
const __m128i vinit0 = _mm_mullo_epi32(vksum, vinput_zero_point0); |
| 57 |
__m128i vacc0x0 = _mm_blend_epi16(vinit0, vzero, 0xFC); |
| 58 |
__m128i vacc0x1 = _mm_blend_epi16(vinit0, vzero, 0xF3); |
| 59 |
__m128i vacc0x2 = _mm_blend_epi16(vinit0, vzero, 0xCF); |
| 60 |
__m128i vacc0x3 = _mm_blend_epi16(vinit0, vzero, 0x3F); |
| 61 |
w = (const int32_t*) w + 4; |
| 62 |
|
| 63 |
size_t k = kc; |
| 64 |
|
| 65 |
|
| 66 |
while (k >= 8 * sizeof(int8_t)) { |
| 67 |
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0); |
| 68 |
const __m128i vxa0 = _mm_cvtepi8_epi16(va0); |
| 69 |
a0 += 8; |
| 70 |
|
| 71 |
const __m128i vb01 = _mm_load_si128((const __m128i*) w); |
| 72 |
|
| 73 |
const __m128i vxbm1 = _mm_unpackhi_epi8(vb01, vb01); |
| 74 |
const __m128i vxb0 = _mm_cvtepi8_epi16(vb01); |
| 75 |
const __m128i vxb1 = _mm_srai_epi16(vxbm1, 8); |
| 76 |
|
| 77 |
vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0)); |
| 78 |
vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1)); |
| 79 |
const __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16)); |
| 80 |
|
| 81 |
const __m128i vxbm3 = _mm_unpackhi_epi8(vb23, vb23); |
| 82 |
const __m128i vxb2 = _mm_cvtepi8_epi16(vb23); |
| 83 |
const __m128i vxb3 = _mm_srai_epi16(vxbm3, 8); |
| 84 |
|
| 85 |
vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2)); |
| 86 |
vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3)); |
| 87 |
|
| 88 |
w = (const int8_t*) w + 32; |
| 89 |
k -= 8 * sizeof(int8_t); |
| 90 |
} |
| 91 |
|
| 92 |
const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1); |
| 93 |
const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3); |
| 94 |
|
| 95 |
__m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23); |
| 96 |
|
| 97 |
__m128 vout0x0123 = _mm_cvtepi32_ps(vacc0x0123); |
| 98 |
|
| 99 |
const __m128 vinput_scale0 = _mm_load1_ps(&quantization_params[0].inv_scale); |
| 100 |
|
| 101 |
vout0x0123 = _mm_mul_ps(vout0x0123, vinput_scale0); |
| 102 |
|
| 103 |
const __m128 vfilter_output_scale0123 = _mm_load_ps((const float*) w); |
| 104 |
vout0x0123 = _mm_mul_ps(vout0x0123, vfilter_output_scale0123); |
| 105 |
|
| 106 |
const __m128 vbias0123 = _mm_load_ps((const float*) w + 4); |
| 107 |
w = (const float*) w + 8; |
| 108 |
vout0x0123 = _mm_add_ps(vout0x0123, vbias0123); |
| 109 |
|
| 110 |
vout0x0123 = _mm_max_ps(vout0x0123, vmin); |
| 111 |
|
| 112 |
vout0x0123 = _mm_min_ps(vout0x0123, vmax); |
| 113 |
|
| 114 |
if XNN_LIKELY(nc >= 4) { |
| 115 |
_mm_storeu_ps(c0, vout0x0123); |
| 116 |
|
| 117 |
a0 = (const int8_t*) ((uintptr_t) a0 - kc); |
| 118 |
|
| 119 |
c0 = (float*) ((uintptr_t) c0 + cn_stride); |
| 120 |
|
| 121 |
nc -= 4; |
| 122 |
} else { |
| 123 |
if (nc & 2) { |
| 124 |
_mm_storel_pi((__m64*) c0, vout0x0123); |
| 125 |
vout0x0123 = _mm_unpackhi_ps(vout0x0123, vout0x0123); |
| 126 |
c0 += 2; |
| 127 |
} |
| 128 |
if (nc & 1) { |
| 129 |
_mm_store_ss(c0, vout0x0123); |
| 130 |
} |
| 131 |
nc = 0; |
| 132 |
} |
| 133 |
} while (nc != 0); |
| 134 |
} |