| 1 |
// Auto-generated file. Do not edit! |
| 2 |
// Template: src/qs8-gemm/c8-wasmdot.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 <wasm_simd128.h> |
| 13 |
|
| 14 |
#include "xnnpack/gemm.h" |
| 15 |
#include "xnnpack/math.h" |
| 16 |
|
| 17 |
|
| 18 |
void xnn_qd8_f32_qc8w_gemm_minmax_ukernel_3x8c8__wasmsdot( |
| 19 |
size_t mr, |
| 20 |
size_t nc, |
| 21 |
size_t kc, |
| 22 |
const int8_t* restrict a, |
| 23 |
size_t a_stride, |
| 24 |
const void* restrict w, |
| 25 |
float* restrict c, |
| 26 |
size_t cm_stride, |
| 27 |
size_t cn_stride, |
| 28 |
const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)], |
| 29 |
const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS |
| 30 |
{ |
| 31 |
assert(mr != 0); |
| 32 |
assert(mr <= 3); |
| 33 |
assert(nc != 0); |
| 34 |
assert(kc != 0); |
| 35 |
assert(kc % sizeof(int8_t) == 0); |
| 36 |
assert(a != NULL); |
| 37 |
assert(w != NULL); |
| 38 |
assert(c != NULL); |
| 39 |
|
| 40 |
kc = round_up_po2(kc, 8 * sizeof(int8_t)); |
| 41 |
const int8_t* a0 = a; |
| 42 |
float* c0 = c; |
| 43 |
const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride); |
| 44 |
float* c1 = (float*) ((uintptr_t) c0 + cm_stride); |
| 45 |
if XNN_UNPREDICTABLE(mr < 2) { |
| 46 |
a1 = a0; |
| 47 |
c1 = c0; |
| 48 |
} |
| 49 |
const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride); |
| 50 |
float* c2 = (float*) ((uintptr_t) c1 + cm_stride); |
| 51 |
if XNN_UNPREDICTABLE(mr <= 2) { |
| 52 |
a2 = a1; |
| 53 |
c2 = c1; |
| 54 |
} |
| 55 |
|
| 56 |
const v128_t vmin = wasm_v128_load32_splat(¶ms->scalar.min); |
| 57 |
const v128_t vmax = wasm_v128_load32_splat(¶ms->scalar.max); |
| 58 |
XNN_FORCE_REALIZATION(vmin); |
| 59 |
XNN_FORCE_REALIZATION(vmax); |
| 60 |
|
| 61 |
do { |
| 62 |
const v128_t vinput_zero_point0 = wasm_v128_load32_splat(&quantization_params[0].zero_point); |
| 63 |
const v128_t vinput_zero_point1 = wasm_v128_load32_splat(&quantization_params[1].zero_point); |
| 64 |
const v128_t vinput_zero_point2 = wasm_v128_load32_splat(&quantization_params[2].zero_point); |
| 65 |
v128_t vksum0123 = wasm_v128_load(w); w = (const int32_t*) w + 4; |
| 66 |
v128_t vsum0x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point0); |
| 67 |
v128_t vacc0x01 = wasm_u64x2_extend_low_u32x4(vsum0x0123); |
| 68 |
v128_t vacc0x23 = wasm_u64x2_extend_high_u32x4(vsum0x0123); |
| 69 |
v128_t vsum1x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point1); |
| 70 |
v128_t vacc1x01 = wasm_u64x2_extend_low_u32x4(vsum1x0123); |
| 71 |
v128_t vacc1x23 = wasm_u64x2_extend_high_u32x4(vsum1x0123); |
| 72 |
v128_t vsum2x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point2); |
| 73 |
v128_t vacc2x01 = wasm_u64x2_extend_low_u32x4(vsum2x0123); |
| 74 |
v128_t vacc2x23 = wasm_u64x2_extend_high_u32x4(vsum2x0123); |
| 75 |
v128_t vksum4567 = wasm_v128_load(w); w = (const int32_t*) w + 4; |
| 76 |
v128_t vsum0x4567 = wasm_i32x4_mul(vksum4567, vinput_zero_point0); |
| 77 |
v128_t vacc0x45 = wasm_u64x2_extend_low_u32x4(vsum0x4567); |
| 78 |
v128_t vacc0x67 = wasm_u64x2_extend_high_u32x4(vsum0x4567); |
| 79 |
v128_t vsum1x4567 = wasm_i32x4_mul(vksum4567, vinput_zero_point1); |
| 80 |
v128_t vacc1x45 = wasm_u64x2_extend_low_u32x4(vsum1x4567); |
| 81 |
v128_t vacc1x67 = wasm_u64x2_extend_high_u32x4(vsum1x4567); |
| 82 |
v128_t vsum2x4567 = wasm_i32x4_mul(vksum4567, vinput_zero_point2); |
| 83 |
v128_t vacc2x45 = wasm_u64x2_extend_low_u32x4(vsum2x4567); |
| 84 |
v128_t vacc2x67 = wasm_u64x2_extend_high_u32x4(vsum2x4567); |
| 85 |
|
| 86 |
size_t k = kc; |
| 87 |
while (k != 0) { |
| 88 |
const v128_t va0x01 = wasm_v128_load64_splat(a0); |
| 89 |
a0 += 8; |
| 90 |
const v128_t va1x01 = wasm_v128_load64_splat(a1); |
| 91 |
a1 += 8; |
| 92 |
const v128_t va2x01 = wasm_v128_load64_splat(a2); |
| 93 |
a2 += 8; |
| 94 |
|
| 95 |
const v128_t vb01x01 = wasm_v128_load(w); w = (const int8_t*) w + 16; |
| 96 |
const v128_t vb23x01 = wasm_v128_load(w); w = (const int8_t*) w + 16; |
| 97 |
const v128_t vb45x01 = wasm_v128_load(w); w = (const int8_t*) w + 16; |
| 98 |
const v128_t vb67x01 = wasm_v128_load(w); w = (const int8_t*) w + 16; |
| 99 |
|
| 100 |
vacc0x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x01, va0x01, vacc0x01); |
| 101 |
vacc0x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x01, va0x01, vacc0x23); |
| 102 |
vacc0x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x01, va0x01, vacc0x45); |
| 103 |
vacc0x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x01, va0x01, vacc0x67); |
| 104 |
vacc1x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x01, va1x01, vacc1x01); |
| 105 |
vacc1x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x01, va1x01, vacc1x23); |
| 106 |
vacc1x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x01, va1x01, vacc1x45); |
| 107 |
vacc1x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x01, va1x01, vacc1x67); |
| 108 |
vacc2x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x01, va2x01, vacc2x01); |
| 109 |
vacc2x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x01, va2x01, vacc2x23); |
| 110 |
vacc2x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x01, va2x01, vacc2x45); |
| 111 |
vacc2x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x01, va2x01, vacc2x67); |
| 112 |
|
| 113 |
k -= 8 * sizeof(int8_t); |
| 114 |
} |
| 115 |
|
| 116 |
|
| 117 |
v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x01, vacc0x23, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc0x01, vacc0x23, 1, 3, 5, 7)); |
| 118 |
v128_t vacc0x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x45, vacc0x67, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc0x45, vacc0x67, 1, 3, 5, 7)); |
| 119 |
v128_t vacc1x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x01, vacc1x23, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc1x01, vacc1x23, 1, 3, 5, 7)); |
| 120 |
v128_t vacc1x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x45, vacc1x67, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc1x45, vacc1x67, 1, 3, 5, 7)); |
| 121 |
v128_t vacc2x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x01, vacc2x23, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc2x01, vacc2x23, 1, 3, 5, 7)); |
| 122 |
v128_t vacc2x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x45, vacc2x67, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc2x45, vacc2x67, 1, 3, 5, 7)); |
| 123 |
|
| 124 |
vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123); |
| 125 |
vacc0x4567 = wasm_f32x4_convert_i32x4(vacc0x4567); |
| 126 |
vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123); |
| 127 |
vacc1x4567 = wasm_f32x4_convert_i32x4(vacc1x4567); |
| 128 |
vacc2x0123 = wasm_f32x4_convert_i32x4(vacc2x0123); |
| 129 |
vacc2x4567 = wasm_f32x4_convert_i32x4(vacc2x4567); |
| 130 |
|
| 131 |
const v128_t vinput_scale0 = wasm_v128_load32_splat(&quantization_params[0].inv_scale); |
| 132 |
const v128_t vinput_scale1 = wasm_v128_load32_splat(&quantization_params[1].inv_scale); |
| 133 |
const v128_t vinput_scale2 = wasm_v128_load32_splat(&quantization_params[2].inv_scale); |
| 134 |
|
| 135 |
vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vinput_scale0); |
| 136 |
vacc0x4567 = wasm_f32x4_mul(vacc0x4567, vinput_scale0); |
| 137 |
vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vinput_scale1); |
| 138 |
vacc1x4567 = wasm_f32x4_mul(vacc1x4567, vinput_scale1); |
| 139 |
vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vinput_scale2); |
| 140 |
vacc2x4567 = wasm_f32x4_mul(vacc2x4567, vinput_scale2); |
| 141 |
|
| 142 |
const v128_t vfilter_output_scale0123 = wasm_v128_load(w); |
| 143 |
w = (const float*) w + 4; |
| 144 |
const v128_t vfilter_output_scale4567 = wasm_v128_load(w); |
| 145 |
w = (const float*) w + 4; |
| 146 |
vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vfilter_output_scale0123); |
| 147 |
vacc0x4567 = wasm_f32x4_mul(vacc0x4567, vfilter_output_scale4567); |
| 148 |
vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vfilter_output_scale0123); |
| 149 |
vacc1x4567 = wasm_f32x4_mul(vacc1x4567, vfilter_output_scale4567); |
| 150 |
vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vfilter_output_scale0123); |
| 151 |
vacc2x4567 = wasm_f32x4_mul(vacc2x4567, vfilter_output_scale4567); |
| 152 |
|
| 153 |
const v128_t vbias0123 = wasm_v128_load(w); |
| 154 |
w = (const float*) w + 4; |
| 155 |
const v128_t vbias4567 = wasm_v128_load(w); |
| 156 |
w = (const float*) w + 4; |
| 157 |
vacc0x0123 = wasm_f32x4_add(vacc0x0123, vbias0123); |
| 158 |
vacc0x4567 = wasm_f32x4_add(vacc0x4567, vbias4567); |
| 159 |
vacc1x0123 = wasm_f32x4_add(vacc1x0123, vbias0123); |
| 160 |
vacc1x4567 = wasm_f32x4_add(vacc1x4567, vbias4567); |
| 161 |
vacc2x0123 = wasm_f32x4_add(vacc2x0123, vbias0123); |
| 162 |
vacc2x4567 = wasm_f32x4_add(vacc2x4567, vbias4567); |
| 163 |
|
| 164 |
vacc0x0123 = wasm_f32x4_pmax(vacc0x0123, vmin); |
| 165 |
vacc0x4567 = wasm_f32x4_pmax(vacc0x4567, vmin); |
| 166 |
vacc1x0123 = wasm_f32x4_pmax(vacc1x0123, vmin); |
| 167 |
vacc1x4567 = wasm_f32x4_pmax(vacc1x4567, vmin); |
| 168 |
vacc2x0123 = wasm_f32x4_pmax(vacc2x0123, vmin); |
| 169 |
vacc2x4567 = wasm_f32x4_pmax(vacc2x4567, vmin); |
| 170 |
|
| 171 |
vacc0x0123 = wasm_f32x4_pmin(vacc0x0123, vmax); |
| 172 |
vacc0x4567 = wasm_f32x4_pmin(vacc0x4567, vmax); |
| 173 |
vacc1x0123 = wasm_f32x4_pmin(vacc1x0123, vmax); |
| 174 |
vacc1x4567 = wasm_f32x4_pmin(vacc1x4567, vmax); |
| 175 |
vacc2x0123 = wasm_f32x4_pmin(vacc2x0123, vmax); |
| 176 |
vacc2x4567 = wasm_f32x4_pmin(vacc2x4567, vmax); |
| 177 |
|
| 178 |
if XNN_LIKELY(nc >= 8) { |
| 179 |
wasm_v128_store(c0, vacc0x0123); |
| 180 |
wasm_v128_store(c0 + 4, vacc0x4567); |
| 181 |
wasm_v128_store(c1, vacc1x0123); |
| 182 |
wasm_v128_store(c1 + 4, vacc1x4567); |
| 183 |
wasm_v128_store(c2, vacc2x0123); |
| 184 |
wasm_v128_store(c2 + 4, vacc2x4567); |
| 185 |
|
| 186 |
a0 = (const int8_t*) ((uintptr_t) a0 - kc); |
| 187 |
a1 = (const int8_t*) ((uintptr_t) a1 - kc); |
| 188 |
a2 = (const int8_t*) ((uintptr_t) a2 - kc); |
| 189 |
|
| 190 |
c0 = (float*) ((uintptr_t) c0 + cn_stride); |
| 191 |
c1 = (float*) ((uintptr_t) c1 + cn_stride); |
| 192 |
c2 = (float*) ((uintptr_t) c2 + cn_stride); |
| 193 |
|
| 194 |
nc -= 8; |
| 195 |
} else { |
| 196 |
if (nc & 4) { |
| 197 |
wasm_v128_store(c0, vacc0x0123); |
| 198 |
vacc0x0123 = vacc0x4567; |
| 199 |
c0 += 4; |
| 200 |
wasm_v128_store(c1, vacc1x0123); |
| 201 |
vacc1x0123 = vacc1x4567; |
| 202 |
c1 += 4; |
| 203 |
wasm_v128_store(c2, vacc2x0123); |
| 204 |
vacc2x0123 = vacc2x4567; |
| 205 |
c2 += 4; |
| 206 |
} |
| 207 |
if (nc & 2) { |
| 208 |
wasm_v128_store64_lane(c0, vacc0x0123, 0); |
| 209 |
vacc0x0123 = wasm_v64x2_shuffle(vacc0x0123, vacc0x0123, 1, 1); |
| 210 |
c0 += 2; |
| 211 |
wasm_v128_store64_lane(c1, vacc1x0123, 0); |
| 212 |
vacc1x0123 = wasm_v64x2_shuffle(vacc1x0123, vacc1x0123, 1, 1); |
| 213 |
c1 += 2; |
| 214 |
wasm_v128_store64_lane(c2, vacc2x0123, 0); |
| 215 |
vacc2x0123 = wasm_v64x2_shuffle(vacc2x0123, vacc2x0123, 1, 1); |
| 216 |
c2 += 2; |
| 217 |
} |
| 218 |
if (nc & 1) { |
| 219 |
wasm_v128_store32_lane(c0, vacc0x0123, 0); |
| 220 |
wasm_v128_store32_lane(c1, vacc1x0123, 0); |
| 221 |
wasm_v128_store32_lane(c2, vacc2x0123, 0); |
| 222 |
} |
| 223 |
nc = 0; |
| 224 |
} |
| 225 |
} while (nc != 0); |
| 226 |
} |