| 1 |
// Auto-generated file. Do not edit! |
| 2 |
// Template: src/x32-packw/gio-scalar.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 |
#include <stddef.h> |
| 12 |
#include <stdint.h> |
| 13 |
|
| 14 |
#include "xnnpack/packw.h" |
| 15 |
|
| 16 |
|
| 17 |
void xnn_x32_packw_gemm_gio_ukernel_x32__scalar( |
| 18 |
size_t g, |
| 19 |
size_t nc, |
| 20 |
size_t kc, |
| 21 |
size_t nr, |
| 22 |
size_t kr, |
| 23 |
size_t sr, |
| 24 |
size_t k_stride, |
| 25 |
const uint32_t* weights, |
| 26 |
const uint32_t* bias, |
| 27 |
const void* scale, |
| 28 |
uint32_t* packed_weights, |
| 29 |
size_t extra_bytes, |
| 30 |
const void* params) |
| 31 |
{ |
| 32 |
assert(g != 0); |
| 33 |
assert(nc != 0); |
| 34 |
assert(kc != 0); |
| 35 |
assert(nr == 32); // This kernel is for NR=32 |
| 36 |
assert(kr == 1); |
| 37 |
assert(sr == 1); |
| 38 |
assert(k_stride != 0); |
| 39 |
assert(weights != NULL); |
| 40 |
assert(packed_weights != NULL); |
| 41 |
|
| 42 |
const float* b = (const float*) bias; |
| 43 |
float* packed_w = (float*) packed_weights; |
| 44 |
do { |
| 45 |
// NC main loop multiple of 32 |
| 46 |
const float* w = (const float*) weights; |
| 47 |
size_t n = nc; |
| 48 |
|
| 49 |
for (; n >= 32; n -= 32) { |
| 50 |
if XNN_LIKELY(b != NULL) { |
| 51 |
const uint64_t v0 = ((const uint64_t*)b)[0]; |
| 52 |
const uint64_t v1 = ((const uint64_t*)b)[1]; |
| 53 |
const uint64_t v2 = ((const uint64_t*)b)[2]; |
| 54 |
const uint64_t v3 = ((const uint64_t*)b)[3]; |
| 55 |
const uint64_t v4 = ((const uint64_t*)b)[4]; |
| 56 |
const uint64_t v5 = ((const uint64_t*)b)[5]; |
| 57 |
const uint64_t v6 = ((const uint64_t*)b)[6]; |
| 58 |
const uint64_t v7 = ((const uint64_t*)b)[7]; |
| 59 |
const uint64_t v8 = ((const uint64_t*)b)[8]; |
| 60 |
const uint64_t v9 = ((const uint64_t*)b)[9]; |
| 61 |
const uint64_t v10 = ((const uint64_t*)b)[10]; |
| 62 |
const uint64_t v11 = ((const uint64_t*)b)[11]; |
| 63 |
const uint64_t v12 = ((const uint64_t*)b)[12]; |
| 64 |
const uint64_t v13 = ((const uint64_t*)b)[13]; |
| 65 |
const uint64_t v14 = ((const uint64_t*)b)[14]; |
| 66 |
const uint64_t v15 = ((const uint64_t*)b)[15]; |
| 67 |
((uint64_t*)packed_w)[0] = v0; |
| 68 |
((uint64_t*)packed_w)[1] = v1; |
| 69 |
((uint64_t*)packed_w)[2] = v2; |
| 70 |
((uint64_t*)packed_w)[3] = v3; |
| 71 |
((uint64_t*)packed_w)[4] = v4; |
| 72 |
((uint64_t*)packed_w)[5] = v5; |
| 73 |
((uint64_t*)packed_w)[6] = v6; |
| 74 |
((uint64_t*)packed_w)[7] = v7; |
| 75 |
((uint64_t*)packed_w)[8] = v8; |
| 76 |
((uint64_t*)packed_w)[9] = v9; |
| 77 |
((uint64_t*)packed_w)[10] = v10; |
| 78 |
((uint64_t*)packed_w)[11] = v11; |
| 79 |
((uint64_t*)packed_w)[12] = v12; |
| 80 |
((uint64_t*)packed_w)[13] = v13; |
| 81 |
((uint64_t*)packed_w)[14] = v14; |
| 82 |
((uint64_t*)packed_w)[15] = v15; |
| 83 |
b += 32; |
| 84 |
} else { |
| 85 |
((uint64_t*)packed_w)[0] = 0; |
| 86 |
((uint64_t*)packed_w)[1] = 0; |
| 87 |
((uint64_t*)packed_w)[2] = 0; |
| 88 |
((uint64_t*)packed_w)[3] = 0; |
| 89 |
((uint64_t*)packed_w)[4] = 0; |
| 90 |
((uint64_t*)packed_w)[5] = 0; |
| 91 |
((uint64_t*)packed_w)[6] = 0; |
| 92 |
((uint64_t*)packed_w)[7] = 0; |
| 93 |
((uint64_t*)packed_w)[8] = 0; |
| 94 |
((uint64_t*)packed_w)[9] = 0; |
| 95 |
((uint64_t*)packed_w)[10] = 0; |
| 96 |
((uint64_t*)packed_w)[11] = 0; |
| 97 |
((uint64_t*)packed_w)[12] = 0; |
| 98 |
((uint64_t*)packed_w)[13] = 0; |
| 99 |
((uint64_t*)packed_w)[14] = 0; |
| 100 |
((uint64_t*)packed_w)[15] = 0; |
| 101 |
} |
| 102 |
packed_w += 32; |
| 103 |
|
| 104 |
// KC main loop |
| 105 |
for (size_t k = kc; k > 0; --k) { |
| 106 |
const uint64_t v0 = ((const uint64_t*)w)[0]; |
| 107 |
const uint64_t v1 = ((const uint64_t*)w)[1]; |
| 108 |
const uint64_t v2 = ((const uint64_t*)w)[2]; |
| 109 |
const uint64_t v3 = ((const uint64_t*)w)[3]; |
| 110 |
const uint64_t v4 = ((const uint64_t*)w)[4]; |
| 111 |
const uint64_t v5 = ((const uint64_t*)w)[5]; |
| 112 |
const uint64_t v6 = ((const uint64_t*)w)[6]; |
| 113 |
const uint64_t v7 = ((const uint64_t*)w)[7]; |
| 114 |
const uint64_t v8 = ((const uint64_t*)w)[8]; |
| 115 |
const uint64_t v9 = ((const uint64_t*)w)[9]; |
| 116 |
const uint64_t v10 = ((const uint64_t*)w)[10]; |
| 117 |
const uint64_t v11 = ((const uint64_t*)w)[11]; |
| 118 |
const uint64_t v12 = ((const uint64_t*)w)[12]; |
| 119 |
const uint64_t v13 = ((const uint64_t*)w)[13]; |
| 120 |
const uint64_t v14 = ((const uint64_t*)w)[14]; |
| 121 |
const uint64_t v15 = ((const uint64_t*)w)[15]; |
| 122 |
((uint64_t*)packed_w)[0] = v0; |
| 123 |
((uint64_t*)packed_w)[1] = v1; |
| 124 |
((uint64_t*)packed_w)[2] = v2; |
| 125 |
((uint64_t*)packed_w)[3] = v3; |
| 126 |
((uint64_t*)packed_w)[4] = v4; |
| 127 |
((uint64_t*)packed_w)[5] = v5; |
| 128 |
((uint64_t*)packed_w)[6] = v6; |
| 129 |
((uint64_t*)packed_w)[7] = v7; |
| 130 |
((uint64_t*)packed_w)[8] = v8; |
| 131 |
((uint64_t*)packed_w)[9] = v9; |
| 132 |
((uint64_t*)packed_w)[10] = v10; |
| 133 |
((uint64_t*)packed_w)[11] = v11; |
| 134 |
((uint64_t*)packed_w)[12] = v12; |
| 135 |
((uint64_t*)packed_w)[13] = v13; |
| 136 |
((uint64_t*)packed_w)[14] = v14; |
| 137 |
((uint64_t*)packed_w)[15] = v15; |
| 138 |
w += k_stride; |
| 139 |
packed_w += 32; |
| 140 |
} |
| 141 |
w = w - kc * k_stride + 32; // Advance to next column of 32 floats |
| 142 |
} |
| 143 |
|
| 144 |
// NC remainder (1..31) |
| 145 |
if XNN_UNLIKELY(n != 0) { |
| 146 |
assert(n >= 1); |
| 147 |
assert(n <= 31); |
| 148 |
|
| 149 |
if XNN_LIKELY(b != NULL) { |
| 150 |
for (size_t i = 0; i < n; ++i) { |
| 151 |
packed_w[i] = b[i]; |
| 152 |
} |
| 153 |
b += n; |
| 154 |
} else { |
| 155 |
((uint64_t*)packed_w)[0] = 0; |
| 156 |
((uint64_t*)packed_w)[1] = 0; |
| 157 |
((uint64_t*)packed_w)[2] = 0; |
| 158 |
((uint64_t*)packed_w)[3] = 0; |
| 159 |
((uint64_t*)packed_w)[4] = 0; |
| 160 |
((uint64_t*)packed_w)[5] = 0; |
| 161 |
((uint64_t*)packed_w)[6] = 0; |
| 162 |
((uint64_t*)packed_w)[7] = 0; |
| 163 |
((uint64_t*)packed_w)[8] = 0; |
| 164 |
((uint64_t*)packed_w)[9] = 0; |
| 165 |
((uint64_t*)packed_w)[10] = 0; |
| 166 |
((uint64_t*)packed_w)[11] = 0; |
| 167 |
((uint64_t*)packed_w)[12] = 0; |
| 168 |
((uint64_t*)packed_w)[13] = 0; |
| 169 |
((uint64_t*)packed_w)[14] = 0; |
| 170 |
((uint64_t*)packed_w)[15] = 0; |
| 171 |
} |
| 172 |
packed_w += 32; |
| 173 |
|
| 174 |
// KC main loop |
| 175 |
for (size_t k = kc; k > 0; --k) { |
| 176 |
for (size_t i = 0; i < n; ++i) { |
| 177 |
packed_w[i] = w[i]; |
| 178 |
} |
| 179 |
w += k_stride; |
| 180 |
packed_w += 32; |
| 181 |
} |
| 182 |
} |
| 183 |
weights += nc * kc; |
| 184 |
} while (--g != 0); |
| 185 |
} |