File size: 10,471 Bytes
8b7c501 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 |
// Copyright 2020 Google LLC
//
// This source code is licensed under the BSD-style license found in the
// LICENSE file in the root directory of this source tree.
$assert ROW_TILE >= 1
$assert ACCUMULATORS >= 1
#include <assert.h>
#include <xmmintrin.h>
#include <xnnpack/dwconv.h>
#include <xnnpack/math.h>
void xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_${ROW_TILE}x4${"_acc%d" % ACCUMULATORS if ACCUMULATORS > 1 else ""}(
size_t input_height,
size_t input_width,
const float* input,
const float* weights,
const float* zero,
float* output,
uint32_t padding_top,
const union xnn_f32_chw_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
assert(input_height != 0);
assert(input_width != 0);
assert(input_width % sizeof(float) == 0);
assert(padding_top >= 1);
assert(padding_top <= 2);
const __m128 vmask_even = _mm_load_ps((const float*) params->sse_stride2.mask_even);
const __m128 vmask_odd = _mm_load_ps((const float*) params->sse_stride2.mask_odd);
const __m128 vmax = _mm_load_ps(params->sse_stride2.max);
const __m128 vmin = _mm_load_ps(params->sse_stride2.min);
const __m128 vbias = _mm_load1_ps(weights);
$for R in range(5):
$for S in range(5):
const __m128 vk${R}${S} = _mm_load1_ps(weights + ${R*5+S+1});
const uint32_t padding_top_less_1 = padding_top - 1;
const size_t input_decrement = round_up_po2(input_width, 8 * sizeof(float));
const float* i0 = zero;
const float* i1 = (const float*) ((uintptr_t) input - ((-padding_top_less_1) & input_width));
const float* i2 = (const float*) ((uintptr_t) i1 + input_width);
if XNN_UNPREDICTABLE(padding_top_less_1 != 0) {
i1 = zero;
}
$for M in range(3, 3 + 2 * ROW_TILE):
const float* i${M} = (const float*) ((uintptr_t) i${M-1} + input_width);
$if ROW_TILE > 1:
const size_t output_width = round_down_po2((input_width + (2 /* padding */ - 3 /* kernel size */ + 2 /* subsampling */) * sizeof(float)) / 2, sizeof(float));
float* o0 = output;
$for M in range(1, ROW_TILE):
float* o${M} = (float*) ((uintptr_t) o${M-1} + output_width);
size_t padded_input_height = input_height + (padding_top_less_1 + 1) + 2 /* padding bottom */;
size_t output_height = (padded_input_height - 5 /* kernel size */ + 2 /* subsampling */) / 2;
do {
$for M in range(3, 3 + 2 * ROW_TILE):
if XNN_UNPREDICTABLE(padded_input_height < ${3 + M}) {
i${M} = zero;
$if M % 2 == 0 and M <= 2 * ROW_TILE + 1:
o${M // 2 - 1} = o${M // 2 - 2};
}
$for M in range(3 + 2 * ROW_TILE):
__m128 vi${M}x6024 = _mm_setzero_ps();
$for M in range(3 + 2 * ROW_TILE):
__m128 vi${M}x7135 = _mm_setzero_ps();
$for M in range(3 + 2 * ROW_TILE):
const __m128 vi${M}x89AB = _mm_loadu_ps(i${M});
const __m128 vi${M}xCDEF = _mm_loadu_ps(i${M} + 4);
i${M} += 8;
$for M in range(3 + 2 * ROW_TILE):
__m128 vi${M}x8ACE = _mm_shuffle_ps(vi${M}x89AB, vi${M}xCDEF, _MM_SHUFFLE(2, 0, 2, 0));
__m128 vi${M}x9BDF = _mm_shuffle_ps(vi${M}x89AB, vi${M}xCDEF, _MM_SHUFFLE(3, 1, 3, 1));
size_t w = input_width;
for (; w > 8 * sizeof(float); w -= 8 * sizeof(float)) {
$for K in range(5):
$for M in range(ROW_TILE):
$if K == 0:
__m128 vo${M}p0 = _mm_add_ps(vbias, _mm_mul_ps(vi${2*M+K}x8ACE, vk${K}2));
$elif K < ACCUMULATORS:
__m128 vo${M}p${K} = _mm_mul_ps(vi${2*M+K}x8ACE, vk${K}2);
$else:
vo${M}p${K % ACCUMULATORS} = _mm_add_ps(vo${M}p${K % ACCUMULATORS}, _mm_mul_ps(vi${2*M+K}x8ACE, vk${K}2));
$for M in range(3 + 2 * ROW_TILE):
const __m128 vi${M}xE8AC = _mm_shuffle_ps(vi${M}x8ACE, vi${M}x8ACE, _MM_SHUFFLE(2, 1, 0, 3));
$for K in range(5):
$for M in range(ROW_TILE):
vo${M}p${(K+5) % ACCUMULATORS} = _mm_add_ps(vo${M}p${(K+5) % ACCUMULATORS}, _mm_mul_ps(vi${2*M+K}x9BDF, vk${K}3));
$for M in range(3 + 2 * ROW_TILE):
const __m128 vi${M}x68AC = _mm_move_ss(vi${M}xE8AC, vi${M}x6024);
vi${M}x6024 = vi${M}xE8AC;
$for M in range(3 + 2 * ROW_TILE):
const __m128 vi${M}xF9BD = _mm_shuffle_ps(vi${M}x9BDF, vi${M}x9BDF, _MM_SHUFFLE(2, 1, 0, 3));
$for K in range(5):
$for M in range(ROW_TILE):
vo${M}p${(K+10) % ACCUMULATORS} = _mm_add_ps(vo${M}p${(K+10) % ACCUMULATORS}, _mm_mul_ps(vi${2*M+K}x68AC, vk${K}0));
$for M in range(3 + 2 * ROW_TILE):
const __m128 vi${M}xGHIJ = _mm_loadu_ps(i${M});
const __m128 vi${M}xKLMN = _mm_loadu_ps(i${M} + 4);
i${M} += 8;
$for M in range(3 + 2 * ROW_TILE):
const __m128 vi${M}x79BD = _mm_move_ss(vi${M}xF9BD, vi${M}x7135);
vi${M}x7135 = vi${M}xF9BD;
$for M in range(3 + 2 * ROW_TILE):
const __m128 vi${M}xGIKM = _mm_shuffle_ps(vi${M}xGHIJ, vi${M}xKLMN, _MM_SHUFFLE(2, 0, 2, 0));
const __m128 vi${M}xHJLN = _mm_shuffle_ps(vi${M}xGHIJ, vi${M}xKLMN, _MM_SHUFFLE(3, 1, 3, 1));
vi${M}x9BDF = vi${M}xHJLN;
$for K in range(5):
$for M in range(ROW_TILE):
vo${M}p${(K+15) % ACCUMULATORS} = _mm_add_ps(vo${M}p${(K+15) % ACCUMULATORS}, _mm_mul_ps(vi${2*M+K}x79BD, vk${K}1));
$for M in range(3 + 2 * ROW_TILE):
const __m128 vi${M}xGACE = _mm_move_ss(vi${M}x8ACE, vi${M}xGIKM);
vi${M}x8ACE = vi${M}xGIKM;
$for M in range(3 + 2 * ROW_TILE):
const __m128 vi${M}xACEG = _mm_shuffle_ps(vi${M}xGACE, vi${M}xGACE, _MM_SHUFFLE(0, 3, 2, 1));
$for K in range(5):
$for M in range(ROW_TILE):
vo${M}p${(K+20) % ACCUMULATORS} = _mm_add_ps(vo${M}p${(K+20) % ACCUMULATORS}, _mm_mul_ps(vi${2*M+K}xACEG, vk${K}4));
$if ACCUMULATORS > 1:
$ACC_SLICE = 1
$while ACC_SLICE < ACCUMULATORS:
$for A in range(0, ACCUMULATORS, ACC_SLICE * 2):
$if A + ACC_SLICE < ACCUMULATORS:
$for M in range(ROW_TILE):
vo${M}p${A} = _mm_add_ps(vo${M}p${A}, vo${M}p${A + ACC_SLICE});
$ACC_SLICE *= 2
$for M in range(ROW_TILE):
__m128 vo${M} = _mm_max_ps(vo${M}p0, vmin);
$for M in range(ROW_TILE):
vo${M} = _mm_min_ps(vo${M}, vmax);
$for M in reversed(range(ROW_TILE)):
_mm_storeu_ps(o${M}, vo${M});
o${M} += 4;
}
// Last block has 1-8 pixels to process.
assert(w <= 8 * sizeof(float));
assert(w >= 1 * sizeof(float));
{
$for M in range(3 + 2 * ROW_TILE):
vi${M}x8ACE = _mm_and_ps(vi${M}x8ACE, vmask_even);
vi${M}x9BDF = _mm_and_ps(vi${M}x9BDF, vmask_odd);
$for K in range(5):
$for M in range(ROW_TILE):
$if K == 0:
__m128 vo${M}p0 = _mm_add_ps(vbias, _mm_mul_ps(vi${2*M+K}x8ACE, vk${K}2));
$elif K < ACCUMULATORS:
__m128 vo${M}p${K} = _mm_mul_ps(vi${2*M+K}x8ACE, vk${K}2);
$else:
vo${M}p${K % ACCUMULATORS} = _mm_add_ps(vo${M}p${K % ACCUMULATORS}, _mm_mul_ps(vi${2*M+K}x8ACE, vk${K}2));
$for M in range(3 + 2 * ROW_TILE):
const __m128 vi${M}xE8AC = _mm_shuffle_ps(vi${M}x8ACE, vi${M}x8ACE, _MM_SHUFFLE(2, 1, 0, 3));
$for K in range(5):
$for M in range(ROW_TILE):
vo${M}p${(K+5) % ACCUMULATORS} = _mm_add_ps(vo${M}p${(K+5) % ACCUMULATORS}, _mm_mul_ps(vi${2*M+K}x9BDF, vk${K}3));
$for M in range(3 + 2 * ROW_TILE):
const __m128 vi${M}x68AC = _mm_move_ss(vi${M}xE8AC, vi${M}x6024);
$for M in range(3 + 2 * ROW_TILE):
const __m128 vi${M}xF9BD = _mm_shuffle_ps(vi${M}x9BDF, vi${M}x9BDF, _MM_SHUFFLE(2, 1, 0, 3));
$for K in range(5):
$for M in range(ROW_TILE):
vo${M}p${(K+10) % ACCUMULATORS} = _mm_add_ps(vo${M}p${(K+10) % ACCUMULATORS}, _mm_mul_ps(vi${2*M+K}x68AC, vk${K}0));
$for M in range(3 + 2 * ROW_TILE):
const __m128 vi${M}x79BD = _mm_move_ss(vi${M}xF9BD, vi${M}x7135);
$for K in range(5):
$for M in range(ROW_TILE):
vo${M}p${(K+15) % ACCUMULATORS} = _mm_add_ps(vo${M}p${(K+15) % ACCUMULATORS}, _mm_mul_ps(vi${2*M+K}x79BD, vk${K}1));
const __m128 vzero = _mm_setzero_ps();
$for M in range(3 + 2 * ROW_TILE):
const __m128 vi${M}xGACE = _mm_move_ss(vi${M}x8ACE, vzero);
$for M in range(3 + 2 * ROW_TILE):
const __m128 vi${M}xACEG = _mm_shuffle_ps(vi${M}xGACE, vi${M}xGACE, _MM_SHUFFLE(0, 3, 2, 1));
$for K in range(5):
$for M in range(ROW_TILE):
vo${M}p${(K+20) % ACCUMULATORS} = _mm_add_ps(vo${M}p${(K+20) % ACCUMULATORS}, _mm_mul_ps(vi${2*M+K}xACEG, vk${K}4));
$if ACCUMULATORS > 1:
$ACC_SLICE = 1
$while ACC_SLICE < ACCUMULATORS:
$for A in range(0, ACCUMULATORS, ACC_SLICE * 2):
$if A + ACC_SLICE < ACCUMULATORS:
$for M in range(ROW_TILE):
vo${M}p${A} = _mm_add_ps(vo${M}p${A}, vo${M}p${A + ACC_SLICE});
$ACC_SLICE *= 2
$for M in range(ROW_TILE):
__m128 vo${M} = _mm_max_ps(vo${M}p0, vmin);
$for M in range(ROW_TILE):
vo${M} = _mm_min_ps(vo${M}, vmax);
size_t w_tmp = (w + 1 * sizeof(float)) / (2 * sizeof(float));
if XNN_LIKELY(w_tmp >= 4) {
$for M in reversed(range(ROW_TILE)):
_mm_storeu_ps(o${M}, vo${M});
o${M} += 4;
} else {
if (w_tmp & 2) {
$for M in reversed(range(ROW_TILE)):
_mm_storel_pi((__m64*) o${M}, vo${M});
o${M} += 2;
$for M in range(ROW_TILE):
vo${M} = _mm_movehl_ps(vo${M}, vo${M});
}
if (w_tmp & 1) {
$for M in reversed(range(ROW_TILE)):
_mm_store_ss(o${M}, vo${M});
o${M} += 1;
}
}
}
i0 = (const float*) ((uintptr_t) i${2 * ROW_TILE} - input_decrement);
i1 = (const float*) ((uintptr_t) i${2 * ROW_TILE + 1} - input_decrement);
i2 = (const float*) ((uintptr_t) i${2 * ROW_TILE + 2} - input_decrement);
$for M in range(3, 3 + 2 * ROW_TILE):
i${M} = (const float*) ((uintptr_t) i${M-1} + input_width);
$if ROW_TILE > 1:
o0 = o${ROW_TILE - 1};
$for M in range(1, ROW_TILE):
o${M} = (float*) ((uintptr_t) o${M-1} + output_width);
$if ROW_TILE > 1:
output_height = doz(output_height, ${ROW_TILE});
padded_input_height = doz(padded_input_height, ${ROW_TILE * 2});
$else:
output_height -= 1;
padded_input_height -= 2;
} while (output_height != 0);
}
|