| 1 |
// Auto-generated file. Do not edit! |
| 2 |
// Template: src/qs8-gemm/scalar.c.in |
| 3 |
// Generator: tools/xngen |
| 4 |
// |
| 5 |
// Copyright 2021 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 "xnnpack/gemm.h" |
| 13 |
#include "xnnpack/math.h" |
| 14 |
#include "xnnpack/unaligned.h" |
| 15 |
|
| 16 |
|
| 17 |
void xnn_qd8_f32_qc8w_gemm_minmax_ukernel_2x2__wasm( |
| 18 |
size_t mr, |
| 19 |
size_t nc, |
| 20 |
size_t kc, |
| 21 |
const int8_t* restrict a, |
| 22 |
size_t a_stride, |
| 23 |
const void* restrict w, |
| 24 |
float* restrict c, |
| 25 |
size_t cm_stride, |
| 26 |
size_t cn_stride, |
| 27 |
const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)], |
| 28 |
const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) |
| 29 |
{ |
| 30 |
assert(mr != 0); |
| 31 |
assert(mr <= 2); |
| 32 |
assert(nc != 0); |
| 33 |
assert(kc != 0); |
| 34 |
|
| 35 |
const int8_t* a0 = a; |
| 36 |
float* c0 = c; |
| 37 |
const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride); |
| 38 |
float* c1 = (float*) ((uintptr_t) c0 + cm_stride); |
| 39 |
if XNN_UNPREDICTABLE(mr != 2) { |
| 40 |
a1 = a0; |
| 41 |
c1 = c0; |
| 42 |
} |
| 43 |
|
| 44 |
|
| 45 |
do { |
| 46 |
const int32_t vksum0 = unaligned_indexed_load_s32(w, 0); |
| 47 |
const int32_t vksum1 = unaligned_indexed_load_s32(w, 1); |
| 48 |
const int32_t vinput_zero_point0 = quantization_params[0].zero_point; |
| 49 |
int32_t vacc0x0 = vksum0 * vinput_zero_point0; |
| 50 |
int32_t vacc0x1 = vksum1 * vinput_zero_point0; |
| 51 |
const int32_t vinput_zero_point1 = quantization_params[1].zero_point; |
| 52 |
int32_t vacc1x0 = vksum0 * vinput_zero_point1; |
| 53 |
int32_t vacc1x1 = vksum1 * vinput_zero_point1; |
| 54 |
w = (const int32_t*) w + 2; |
| 55 |
|
| 56 |
size_t k = kc; |
| 57 |
do { |
| 58 |
const int32_t va0 = (int32_t) *a0++; |
| 59 |
const int32_t va1 = (int32_t) *a1++; |
| 60 |
|
| 61 |
const int32_t vb0 = (int32_t) ((const int8_t*) w)[0]; |
| 62 |
const int32_t vb1 = (int32_t) ((const int8_t*) w)[1]; |
| 63 |
w = (const int8_t*) w + 2; |
| 64 |
|
| 65 |
vacc0x0 += va0 * vb0; |
| 66 |
vacc0x1 += va0 * vb1; |
| 67 |
vacc1x0 += va1 * vb0; |
| 68 |
vacc1x1 += va1 * vb1; |
| 69 |
|
| 70 |
k -= sizeof(int8_t); |
| 71 |
} while (k != 0); |
| 72 |
|
| 73 |
float vout0x0 = (float) vacc0x0; |
| 74 |
float vout0x1 = (float) vacc0x1; |
| 75 |
float vout1x0 = (float) vacc1x0; |
| 76 |
float vout1x1 = (float) vacc1x1; |
| 77 |
|
| 78 |
const float vinput_scale0 = quantization_params[0].inv_scale; |
| 79 |
vout0x0 *= vinput_scale0; |
| 80 |
vout0x1 *= vinput_scale0; |
| 81 |
const float vinput_scale1 = quantization_params[1].inv_scale; |
| 82 |
vout1x0 *= vinput_scale1; |
| 83 |
vout1x1 *= vinput_scale1; |
| 84 |
|
| 85 |
const float vfilter_output_scale0 = unaligned_indexed_load_f32(w, 0); |
| 86 |
vout0x0 *= vfilter_output_scale0; |
| 87 |
vout1x0 *= vfilter_output_scale0; |
| 88 |
const float vfilter_output_scale1 = unaligned_indexed_load_f32(w, 1); |
| 89 |
vout0x1 *= vfilter_output_scale1; |
| 90 |
vout1x1 *= vfilter_output_scale1; |
| 91 |
|
| 92 |
const float vbias0 = unaligned_indexed_load_f32(w, 2); |
| 93 |
vout0x0 += vbias0; |
| 94 |
vout1x0 += vbias0; |
| 95 |
const float vbias1 = unaligned_indexed_load_f32(w, 3); |
| 96 |
vout0x1 += vbias1; |
| 97 |
vout1x1 += vbias1; |
| 98 |
|
| 99 |
w = (const float*) w + 4; |
| 100 |
|
| 101 |
const float voutput_min = params->scalar.min; |
| 102 |
vout0x0 = __builtin_wasm_max_f32(vout0x0, voutput_min); |
| 103 |
vout1x0 = __builtin_wasm_max_f32(vout1x0, voutput_min); |
| 104 |
vout0x1 = __builtin_wasm_max_f32(vout0x1, voutput_min); |
| 105 |
vout1x1 = __builtin_wasm_max_f32(vout1x1, voutput_min); |
| 106 |
|
| 107 |
const float voutput_max = params->scalar.max; |
| 108 |
vout0x0 = __builtin_wasm_min_f32(vout0x0, voutput_max); |
| 109 |
vout1x0 = __builtin_wasm_min_f32(vout1x0, voutput_max); |
| 110 |
vout0x1 = __builtin_wasm_min_f32(vout0x1, voutput_max); |
| 111 |
vout1x1 = __builtin_wasm_min_f32(vout1x1, voutput_max); |
| 112 |
|
| 113 |
if XNN_LIKELY(nc >= 2) { |
| 114 |
c0[0] = vout0x0; |
| 115 |
c0[1] = vout0x1; |
| 116 |
c1[0] = vout1x0; |
| 117 |
c1[1] = vout1x1; |
| 118 |
|
| 119 |
a0 = (const int8_t*) ((uintptr_t) a0 - kc); |
| 120 |
a1 = (const int8_t*) ((uintptr_t) a1 - kc); |
| 121 |
|
| 122 |
c0 = (float*) ((uintptr_t) c0 + cn_stride); |
| 123 |
c1 = (float*) ((uintptr_t) c1 + cn_stride); |
| 124 |
|
| 125 |
nc -= 2; |
| 126 |
} else { |
| 127 |
if (nc & 1) { |
| 128 |
c0[0] = vout0x0; |
| 129 |
c1[0] = vout1x0; |
| 130 |
} |
| 131 |
|
| 132 |
nc = 0; |
| 133 |
} |
| 134 |
} while (nc != 0); |
| 135 |
} |