| 1 |
// Auto-generated file. Do not edit! |
| 2 |
// Template: src/x32-packw/c4-sse2.c.in |
| 3 |
// Generator: tools/xngen |
| 4 |
// |
| 5 |
// Copyright 2023 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 |
|
| 11 |
#include <assert.h> |
| 12 |
#include <stddef.h> |
| 13 |
#include <stdint.h> |
| 14 |
|
| 15 |
#include <emmintrin.h> |
| 16 |
|
| 17 |
#include "xnnpack/packw.h" |
| 18 |
|
| 19 |
|
| 20 |
void xnn_x32_packw_gemm_goi_ukernel_x2c4__sse2_u4( |
| 21 |
size_t g, |
| 22 |
size_t nc, |
| 23 |
size_t kc, |
| 24 |
size_t nr, |
| 25 |
size_t kr, |
| 26 |
size_t sr, |
| 27 |
const uint32_t* weights, |
| 28 |
const uint32_t* bias, |
| 29 |
const void* scale, |
| 30 |
uint32_t* packed_weights, |
| 31 |
size_t extra_bytes, |
| 32 |
const void* params) |
| 33 |
{ |
| 34 |
assert(g != 0); |
| 35 |
assert(nc != 0); |
| 36 |
assert(kc != 0); |
| 37 |
assert(nr == 2); |
| 38 |
assert(kr == 4); |
| 39 |
assert(sr == 1); |
| 40 |
assert(weights != NULL); |
| 41 |
assert(packed_weights != NULL); |
| 42 |
__m128 v0; |
| 43 |
__m128 v1; |
| 44 |
|
| 45 |
const float* b = (const float*) bias; |
| 46 |
float* packed_w = (float*) packed_weights; |
| 47 |
do { |
| 48 |
// NC main loop multiple of 2 |
| 49 |
const float* w0 = (const float*) weights; |
| 50 |
size_t n = nc; |
| 51 |
|
| 52 |
for (; n >= 2; n -= 2) { |
| 53 |
if XNN_LIKELY(b != NULL) { |
| 54 |
packed_w[0] = b[0]; |
| 55 |
packed_w[1] = b[1]; |
| 56 |
b += 2; |
| 57 |
} else { |
| 58 |
packed_w[0] = 0.0f; |
| 59 |
packed_w[1] = 0.0f; |
| 60 |
} |
| 61 |
packed_w += 2; |
| 62 |
|
| 63 |
const float* w1 = w0 + kc; |
| 64 |
|
| 65 |
// KC main loop multiple of 2x4 |
| 66 |
size_t k = kc; |
| 67 |
for (; k >= 4; k -= 4) { |
| 68 |
// Read blocks of 2x4 |
| 69 |
// a b c d |
| 70 |
// e f g h |
| 71 |
v0 = _mm_loadu_ps(w0); |
| 72 |
w0 += 4; |
| 73 |
v1 = _mm_loadu_ps(w1); |
| 74 |
w1 += 4; |
| 75 |
_mm_storeu_ps(packed_w, v0); |
| 76 |
_mm_storeu_ps(packed_w + 4, v1); |
| 77 |
packed_w += 8; |
| 78 |
} |
| 79 |
|
| 80 |
// KC remainder (1..3) |
| 81 |
if XNN_UNLIKELY(k != 0) { |
| 82 |
assert(k >= 1); |
| 83 |
assert(k <= 3); |
| 84 |
switch (k) { |
| 85 |
case 1: |
| 86 |
// Read blocks of 2x1 |
| 87 |
// a |
| 88 |
// e |
| 89 |
v0 = _mm_load_ss(w0); |
| 90 |
w0 += 1; |
| 91 |
v1 = _mm_load_ss(w1); |
| 92 |
w1 += 1; |
| 93 |
break; |
| 94 |
case 2: |
| 95 |
// Read blocks of 2x2 |
| 96 |
// a b |
| 97 |
// e f |
| 98 |
v0 = _mm_castpd_ps(_mm_load_sd((const double*) w0)); |
| 99 |
w0 += 2; |
| 100 |
v1 = _mm_castpd_ps(_mm_load_sd((const double*) w1)); |
| 101 |
w1 += 2; |
| 102 |
break; |
| 103 |
case 3: |
| 104 |
{ |
| 105 |
// Read blocks of 2x3 |
| 106 |
// a b c |
| 107 |
// e f g |
| 108 |
const __m128 v0lo = _mm_castpd_ps(_mm_load_sd((const double*) w0)); |
| 109 |
const __m128 v0hi = _mm_load_ss(w0 + 2); |
| 110 |
v0 = _mm_movelh_ps(v0lo, v0hi); |
| 111 |
w0 += 3; |
| 112 |
const __m128 v1lo = _mm_castpd_ps(_mm_load_sd((const double*) w1)); |
| 113 |
const __m128 v1hi = _mm_load_ss(w1 + 2); |
| 114 |
v1 = _mm_movelh_ps(v1lo, v1hi); |
| 115 |
w1 += 3; |
| 116 |
break; |
| 117 |
} |
| 118 |
default: |
| 119 |
XNN_UNREACHABLE; |
| 120 |
} |
| 121 |
_mm_storeu_ps(packed_w, v0); |
| 122 |
_mm_storeu_ps(packed_w + 4, v1); |
| 123 |
packed_w += 8; |
| 124 |
} |
| 125 |
packed_w = (float*) ((uintptr_t) packed_w + extra_bytes); |
| 126 |
w0 = w1; |
| 127 |
} |
| 128 |
|
| 129 |
// NC remainder (1..1) |
| 130 |
if XNN_UNLIKELY(n != 0) { |
| 131 |
assert(n >= 1); |
| 132 |
assert(n <= 1); |
| 133 |
if XNN_LIKELY(b != NULL) { |
| 134 |
size_t nb = n; |
| 135 |
do { |
| 136 |
*packed_w++ = *b++; |
| 137 |
} while (--nb != 0); |
| 138 |
packed_w += (2 - n); |
| 139 |
} else { |
| 140 |
packed_w[0] = 0.0f; |
| 141 |
packed_w[1] = 0.0f; |
| 142 |
packed_w += 2; |
| 143 |
} |
| 144 |
|
| 145 |
// NR remainder has less than 2 rows so last row is not loaded |
| 146 |
|
| 147 |
// KC main loop multiple of 2x4 |
| 148 |
size_t k = kc; |
| 149 |
for (; k >= 4; k -= 4) { |
| 150 |
// Read blocks of 2x4 |
| 151 |
// a b c d |
| 152 |
// e f g h |
| 153 |
v0 = _mm_loadu_ps(w0); |
| 154 |
w0 += 4; |
| 155 |
_mm_storeu_ps(packed_w, v0); |
| 156 |
_mm_storeu_ps(packed_w + 4, v0); |
| 157 |
packed_w += 8; |
| 158 |
} |
| 159 |
|
| 160 |
// KC remainder (1..3) |
| 161 |
if XNN_UNLIKELY(k != 0) { |
| 162 |
assert(k >= 1); |
| 163 |
assert(k <= 3); |
| 164 |
switch (k) { |
| 165 |
case 1: |
| 166 |
// Read blocks of 1x1 |
| 167 |
// a |
| 168 |
v0 = _mm_load_ss(w0); |
| 169 |
w0 += 1; |
| 170 |
break; |
| 171 |
case 2: |
| 172 |
// Read blocks of 1x2 |
| 173 |
// a b |
| 174 |
v0 = _mm_castpd_ps(_mm_load_sd((const double*) w0)); |
| 175 |
w0 += 2; |
| 176 |
break; |
| 177 |
case 3: |
| 178 |
{ |
| 179 |
// Read blocks of 1x3 |
| 180 |
// a b c |
| 181 |
const __m128 v0lo = _mm_castpd_ps(_mm_load_sd((const double*) w0)); |
| 182 |
const __m128 v0hi = _mm_load_ss(w0 + 2); |
| 183 |
v0 = _mm_movelh_ps(v0lo, v0hi); |
| 184 |
w0 += 3; |
| 185 |
break; |
| 186 |
} |
| 187 |
default: |
| 188 |
XNN_UNREACHABLE; |
| 189 |
} |
| 190 |
_mm_storeu_ps(packed_w, v0); |
| 191 |
_mm_storeu_ps(packed_w + 4, v0); |
| 192 |
packed_w += 8; |
| 193 |
} |
| 194 |
packed_w = (float*) ((uintptr_t) packed_w + extra_bytes); |
| 195 |
} |
| 196 |
weights += nc * kc; |
| 197 |
} while (--g != 0); |
| 198 |
} |