| 1 | namespace dnnl { |
|---|---|
| 2 | namespace impl { |
| 3 | namespace gpu { |
| 4 | namespace ocl { |
| 5 | const char *custom_reorder_kernel = R"==(/******************************************************************************* )==" "\n" |
| 6 | R"==(* Copyright 2019-2022 Intel Corporation )==" "\n" |
| 7 | R"==(* )==" "\n" |
| 8 | R"==(* Licensed under the Apache License, Version 2.0 (the "License"); )==" "\n" |
| 9 | R"==(* you may not use this file except in compliance with the License. )==" "\n" |
| 10 | R"==(* You may obtain a copy of the License at )==" "\n" |
| 11 | R"==(* )==" "\n" |
| 12 | R"==(* http: )==" "\n" |
| 13 | R"==(* )==" "\n" |
| 14 | R"==(* Unless required by applicable law or agreed to in writing, software )==" "\n" |
| 15 | R"==(* distributed under the License is distributed on an "AS IS" BASIS, )==" "\n" |
| 16 | R"==(* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. )==" "\n" |
| 17 | R"==(* See the License for the specific language governing permissions and )==" "\n" |
| 18 | R"==(* limitations under the License. )==" "\n" |
| 19 | R"==(*******************************************************************************/ )==" "\n" |
| 20 | R"==(#include "gpu/ocl/reorder_common.h" )==" "\n" |
| 21 | R"==(KERNEL_ATTR )==" "\n" |
| 22 | R"==(__kernel void custom_reorder(__global SRC_DATA_T *restrict src, )==" "\n" |
| 23 | R"==(__global DST_DATA_T *restrict dst, __global float *restrict src_scales, )==" "\n" |
| 24 | R"==(__global int *restrict src_zps, __global float *restrict dst_scales, )==" "\n" |
| 25 | R"==(__global int *restrict dst_zps, float sum_scale, int sum_zp) { )==" "\n" |
| 26 | R"==(const int src_zp = GET_SRC_ZP(src_zps); )==" "\n" |
| 27 | R"==(const int dst_zp = GET_DST_ZP(dst_zps); )==" "\n" |
| 28 | R"==(#if WITH_SRC_SCALE && SRC_NUM_SCALES == 1 )==" "\n" |
| 29 | R"==(float src_scale = src_scales[0]; )==" "\n" |
| 30 | R"==(#else )==" "\n" |
| 31 | R"==(float src_scale = 1.0f; )==" "\n" |
| 32 | R"==(#endif )==" "\n" |
| 33 | R"==(#if WITH_DST_SCALE && DST_NUM_SCALES == 1 )==" "\n" |
| 34 | R"==(float dst_scale = dst_scales[0]; )==" "\n" |
| 35 | R"==(#else )==" "\n" |
| 36 | R"==(float dst_scale = 1.0f; )==" "\n" |
| 37 | R"==(#endif )==" "\n" |
| 38 | R"==(src += SRC_OFFSET0; )==" "\n" |
| 39 | R"==(dst += DST_OFFSET0; )==" "\n" |
| 40 | R"==(#if UNALIGNED )==" "\n" |
| 41 | R"==(const int d0 = GWS_GET_D0(); )==" "\n" |
| 42 | R"==(const int d1_blk_start = GWS_GET_D1(); )==" "\n" |
| 43 | R"==(const int d2_blk_start = GWS_GET_D2(); )==" "\n" |
| 44 | R"==(const int d3_blk_start = GWS_GET_D3(); )==" "\n" |
| 45 | R"==(const int d4_blk_start = GWS_GET_D4(); )==" "\n" |
| 46 | R"==(const int d5_blk_start = GWS_GET_D5(); )==" "\n" |
| 47 | R"==(const int d1_blk_end = d1_blk_start + GWS_GET_D1_BLOCK(); )==" "\n" |
| 48 | R"==(const int d2_blk_end = d2_blk_start + GWS_GET_D2_BLOCK(); )==" "\n" |
| 49 | R"==(const int d3_blk_end = d3_blk_start + GWS_GET_D3_BLOCK(); )==" "\n" |
| 50 | R"==(const int d4_blk_end = d4_blk_start + GWS_GET_D4_BLOCK(); )==" "\n" |
| 51 | R"==(const int d5_blk_end = d5_blk_start + GWS_GET_D5_BLOCK(); )==" "\n" |
| 52 | R"==(for_(int d1 = d1_blk_start; d1 < d1_blk_end; ++d1) )==" "\n" |
| 53 | R"==(for_(int d2 = d2_blk_start; d2 < d2_blk_end; ++d2) )==" "\n" |
| 54 | R"==(for_(int d3 = d3_blk_start; d3 < d3_blk_end; ++d3) )==" "\n" |
| 55 | R"==(for_(int d4 = d4_blk_start; d4 < d4_blk_end; ++d4) )==" "\n" |
| 56 | R"==(for (int d5 = d5_blk_start; d5 < d5_blk_end; ++d5) { )==" "\n" |
| 57 | R"==(const int src_off = SRC_OFF(d0, d1, d2, d3, d4, d5); )==" "\n" |
| 58 | R"==(const int dst_off = DST_OFF(d0, d1, d2, d3, d4, d5); )==" "\n" |
| 59 | R"==(#if PAD_FILL_ZERO == 1 )==" "\n" |
| 60 | R"==(int pad_d0 = d0 >= SRC_D0; )==" "\n" |
| 61 | R"==(int pad_d1 = NDIMS > 1 && d1 >= SRC_D1; )==" "\n" |
| 62 | R"==(int pad_d2 = NDIMS > 2 && d2 >= SRC_D2; )==" "\n" |
| 63 | R"==(int pad_d3 = NDIMS > 3 && d3 >= SRC_D3; )==" "\n" |
| 64 | R"==(int pad_d4 = NDIMS > 4 && d4 >= SRC_D4; )==" "\n" |
| 65 | R"==(int pad_d5 = NDIMS > 5 && d5 >= SRC_D5; )==" "\n" |
| 66 | R"==(if (pad_d0 || pad_d1 || pad_d2 || pad_d3 || pad_d4 || pad_d5) { )==" "\n" |
| 67 | R"==(dst[dst_off] = 0; )==" "\n" |
| 68 | R"==(continue; )==" "\n" |
| 69 | R"==(} )==" "\n" |
| 70 | R"==(#endif )==" "\n" |
| 71 | R"==(#if WITH_SRC_SCALE && SRC_NUM_SCALES > 1 )==" "\n" |
| 72 | R"==(src_scale = src_scales[SCALE_OFF(SRC, d0, d1, d2, d3, d4, d5)]; )==" "\n" |
| 73 | R"==(#endif )==" "\n" |
| 74 | R"==(#if WITH_DST_SCALE && DST_NUM_SCALES > 1 )==" "\n" |
| 75 | R"==(dst_scale = dst_scales[SCALE_OFF(DST, d0, d1, d2, d3, d4, d5)]; )==" "\n" |
| 76 | R"==(#endif )==" "\n" |
| 77 | R"==(REORDER(dst[dst_off], src[src_off], src_scale, dst_scale, sum_scale, )==" "\n" |
| 78 | R"==(src_zp, dst_zp, sum_zp); )==" "\n" |
| 79 | R"==(} )==" "\n" |
| 80 | R"==(#elif ALT_OFFSETS )==" "\n" |
| 81 | R"==(const int d0 = get_global_id(0); )==" "\n" |
| 82 | R"==(const int d1 = get_global_id(1); )==" "\n" |
| 83 | R"==(const int d2 = get_global_id(2); )==" "\n" |
| 84 | R"==(#ifdef LIMIT_MAX_D0 )==" "\n" |
| 85 | R"==(if (d0 >= LIMIT_MAX_D0) { return; } )==" "\n" |
| 86 | R"==(#endif )==" "\n" |
| 87 | R"==(const int src_base = S0 * d0 + S1 * d1 + S2 * d2; )==" "\n" |
| 88 | R"==(const int dst_base = D0 * d0 + D1 * d1 + D2 * d2; )==" "\n" |
| 89 | R"==(for (int db = 0; db < BLK; ++db) { )==" "\n" |
| 90 | R"==(const int src_off = src_base + db * SB; )==" "\n" |
| 91 | R"==(const int dst_off = dst_base + db * DB; )==" "\n" |
| 92 | R"==(REORDER(dst[dst_off], src[src_off], src_scale, dst_scale, sum_scale, )==" "\n" |
| 93 | R"==(src_zp, dst_zp, sum_zp); )==" "\n" |
| 94 | R"==(} )==" "\n" |
| 95 | R"==(#elif PLAIN_xFxE_TO_ABCDEF )==" "\n" |
| 96 | R"==(const int d0 = GWS_GET_D0(); )==" "\n" |
| 97 | R"==(const int d1 = GWS_GET_D1(); )==" "\n" |
| 98 | R"==(const int d2 = GWS_GET_D2(); )==" "\n" |
| 99 | R"==(const int d3 = GWS_GET_D3(); )==" "\n" |
| 100 | R"==(const int d4 = GWS_GET_D4(); )==" "\n" |
| 101 | R"==(const int d5 = GWS_GET_D5(); )==" "\n" |
| 102 | R"==(#if WITH_SUM_SCALE || WITH_SUM_ZPOINT )==" "\n" |
| 103 | R"==(#define SUM_OUTPUT 1 )==" "\n" |
| 104 | R"==(#else )==" "\n" |
| 105 | R"==(#define SUM_OUTPUT 0 )==" "\n" |
| 106 | R"==(#endif )==" "\n" |
| 107 | R"==(const unsigned sglid = get_sub_group_local_id(); )==" "\n" |
| 108 | R"==(#define REORDER_BLOCK(block_size, src_memory, src_swap, src_offset) \ )==" "\n" |
| 109 | R"==({ \ )==" "\n" |
| 110 | R"==(unroll_for(unsigned sidx = 0; sidx < block_size; ++sidx) { \ )==" "\n" |
| 111 | R"==(const unsigned src_off \ )==" "\n" |
| 112 | R"==(= SRC_OFF(d0, d1, d2, d3, d4, sidx + src_offset); \ )==" "\n" |
| 113 | R"==(src_memory[sidx] = SRC_BLOCK_READ(&src[src_off]); \ )==" "\n" |
| 114 | R"==(} \ )==" "\n" |
| 115 | R"==(unroll_for(int j = 0; j < SUB_GROUP_SIZE; j++) \ )==" "\n" |
| 116 | R"==(unroll_for(int i = 0; i < block_size; i++) { \ )==" "\n" |
| 117 | R"==(unsigned x = (i + j * block_size) / SUB_GROUP_SIZE; \ )==" "\n" |
| 118 | R"==(unsigned y = (i + j * block_size) % SUB_GROUP_SIZE; \ )==" "\n" |
| 119 | R"==(unsigned sg_src = (i + j * block_size) / block_size; \ )==" "\n" |
| 120 | R"==(src_swap[x][y] = intel_sub_group_shuffle(src_mem[i], sg_src); \ )==" "\n" |
| 121 | R"==(} \ )==" "\n" |
| 122 | R"==(\ )==" "\n" |
| 123 | R"==(DST_DATA_T dst_tmp; \ )==" "\n" |
| 124 | R"==(unsigned dst_off; \ )==" "\n" |
| 125 | R"==(if (block_size < 16) dst_off = DST_OFF(d0, d1, d2, d3, d4, 0); \ )==" "\n" |
| 126 | R"==(\ )==" "\n" |
| 127 | R"==(unroll_for(unsigned sidx = 0; sidx < block_size; ++sidx) { \ )==" "\n" |
| 128 | R"==(if (block_size >= 16) \ )==" "\n" |
| 129 | R"==(dst_off = DST_OFF(d0, d1, d2, d3, d4 + sidx, src_offset); \ )==" "\n" |
| 130 | R"==(if (SUM_OUTPUT) dst_tmp = DST_BLOCK_READ(&dst[dst_off]); \ )==" "\n" |
| 131 | R"==(REORDER(dst_tmp, src_swap[sidx][sglid], src_scale, dst_scale, \ )==" "\n" |
| 132 | R"==(sum_scale, src_zp, dst_zp, sum_zp); \ )==" "\n" |
| 133 | R"==(DST_BLOCK_WRITE(&dst[dst_off], dst_tmp); \ )==" "\n" |
| 134 | R"==(if (block_size < 16) dst_off += SUB_GROUP_SIZE; \ )==" "\n" |
| 135 | R"==(} \ )==" "\n" |
| 136 | R"==(} )==" "\n" |
| 137 | R"==(#if DST_D5 > 16 )==" "\n" |
| 138 | R"==(unsigned block_size = 16; )==" "\n" |
| 139 | R"==(#else )==" "\n" |
| 140 | R"==(unsigned block_size = DST_D5; )==" "\n" |
| 141 | R"==(#endif )==" "\n" |
| 142 | R"==(SRC_DATA_T src_mem[16]; )==" "\n" |
| 143 | R"==(SRC_DATA_T src_all[16][SUB_GROUP_SIZE]; )==" "\n" |
| 144 | R"==(REORDER_BLOCK(block_size, src_mem, src_all, d5); )==" "\n" |
| 145 | R"==(#elif TRANSPOSE_NXN || LOCAL_NXN )==" "\n" |
| 146 | R"==(#define BATCH_SIZE SUB_GROUP_SIZE )==" "\n" |
| 147 | R"==(int sgId = get_sub_group_local_id(); )==" "\n" |
| 148 | R"==(int sg_off = get_sub_group_id() * SUB_GROUP_SIZE * BATCH_SIZE; )==" "\n" |
| 149 | R"==(const int d0 = GWS_GET_D0(); )==" "\n" |
| 150 | R"==(const int d1 = GWS_GET_D1(); )==" "\n" |
| 151 | R"==(const int d2 = GWS_GET_D2(); )==" "\n" |
| 152 | R"==(const int d3 = GWS_GET_D3(); )==" "\n" |
| 153 | R"==(const int d4 = GWS_GET_D4(); )==" "\n" |
| 154 | R"==(const int d5 = GWS_GET_D5(); )==" "\n" |
| 155 | R"==(const int d0_block = GWS_GET_D0_BLOCK(); )==" "\n" |
| 156 | R"==(const int d1_block = GWS_GET_D1_BLOCK(); )==" "\n" |
| 157 | R"==(const int d2_block = GWS_GET_D2_BLOCK(); )==" "\n" |
| 158 | R"==(const int d3_block = GWS_GET_D3_BLOCK(); )==" "\n" |
| 159 | R"==(const int d4_block = GWS_GET_D4_BLOCK(); )==" "\n" |
| 160 | R"==(const int d5_block = GWS_GET_D5_BLOCK(); )==" "\n" |
| 161 | R"==(#ifdef TRANSPOSE_NXN )==" "\n" |
| 162 | R"==(SRC_DATA_T src_buf[SUB_GROUP_SIZE]; )==" "\n" |
| 163 | R"==(SRC_DATA_T dst_buf[SUB_GROUP_SIZE]; )==" "\n" |
| 164 | R"==(#else )==" "\n" |
| 165 | R"==(__local SRC_DATA_T tmp[SG_PER_WG * BATCH_SIZE]; )==" "\n" |
| 166 | R"==(#endif )==" "\n" |
| 167 | R"==(SRC_DATA_T send_buf; )==" "\n" |
| 168 | R"==(#if PAD_FILL_ZERO == 1 )==" "\n" |
| 169 | R"==(const int pad_d0 = d0 >= SRC_D0; )==" "\n" |
| 170 | R"==(const int pad_d1 = NDIMS > 1 && d1 >= SRC_D1; )==" "\n" |
| 171 | R"==(const int pad_d2 = NDIMS > 2 && d2 >= SRC_D2; )==" "\n" |
| 172 | R"==(const int pad_d3 = NDIMS > 3 && d3 >= SRC_D3; )==" "\n" |
| 173 | R"==(const int pad_d4 = NDIMS > 4 && d4 >= SRC_D4; )==" "\n" |
| 174 | R"==(const int pad_d5 = NDIMS > 5 && d5 >= SRC_D5; )==" "\n" |
| 175 | R"==(const int pad = pad_d0 || pad_d1 || pad_d2 || pad_d3 || pad_d4 || pad_d5; )==" "\n" |
| 176 | R"==(#else )==" "\n" |
| 177 | R"==(const int pad = 0; )==" "\n" |
| 178 | R"==(#endif )==" "\n" |
| 179 | R"==(if (!pad) { )==" "\n" |
| 180 | R"==(for_(int d0i = 0; d0i < d0_block; d0i++) )==" "\n" |
| 181 | R"==(for_(int d1i = 0; d1i < d1_block; d1i++) )==" "\n" |
| 182 | R"==(for_(int d2i = 0; d2i < d2_block; d2i++) )==" "\n" |
| 183 | R"==(for_(int d3i = 0; d3i < d3_block; d3i++) )==" "\n" |
| 184 | R"==(for_(int d4i = 0; d4i < d4_block; d4i++) )==" "\n" |
| 185 | R"==(for (int d5i = 0; d5i < d5_block; d5i++) { )==" "\n" |
| 186 | R"==(const int iter = d0i + d1i + d2i + d3i + d4i + d5i; )==" "\n" |
| 187 | R"==(const int src_off = SRC_OFF( )==" "\n" |
| 188 | R"==(d0 + d0i, d1 + d1i, d2 + d2i, d3 + d3i, d4 + d4i, d5 + d5i); )==" "\n" |
| 189 | R"==(#ifdef TRANSPOSE_NXN )==" "\n" |
| 190 | R"==(src_buf[iter] = SRC_BLOCK_READ(&src[src_off]); )==" "\n" |
| 191 | R"==(#else )==" "\n" |
| 192 | R"==(tmp[sg_off + iter * SUB_GROUP_SIZE + sgId] )==" "\n" |
| 193 | R"==(= SRC_BLOCK_READ(&src[src_off]); )==" "\n" |
| 194 | R"==(#endif )==" "\n" |
| 195 | R"==(} )==" "\n" |
| 196 | R"==(#ifdef TRANSPOSE_NXN )==" "\n" |
| 197 | R"==(dst_buf[sgId] = src_buf[sgId]; )==" "\n" |
| 198 | R"==(for (int i = 1; i < SUB_GROUP_SIZE; i++) { )==" "\n" |
| 199 | R"==(send_buf = src_buf[(i + sgId) % BATCH_SIZE]; )==" "\n" |
| 200 | R"==(dst_buf[(BATCH_SIZE + sgId - i) % BATCH_SIZE] )==" "\n" |
| 201 | R"==(= intel_sub_group_shuffle( )==" "\n" |
| 202 | R"==(send_buf, (BATCH_SIZE + sgId - i) % BATCH_SIZE); )==" "\n" |
| 203 | R"==(} )==" "\n" |
| 204 | R"==(#endif )==" "\n" |
| 205 | R"==(} )==" "\n" |
| 206 | R"==(for_(int d0i = 0; d0i < d0_block; d0i++) )==" "\n" |
| 207 | R"==(for_(int d1i = 0; d1i < d1_block; d1i++) )==" "\n" |
| 208 | R"==(for_(int d2i = 0; d2i < d2_block; d2i++) )==" "\n" |
| 209 | R"==(for_(int d3i = 0; d3i < d3_block; d3i++) )==" "\n" |
| 210 | R"==(for_(int d4i = 0; d4i < d4_block; d4i++) )==" "\n" |
| 211 | R"==(for (int d5i = 0; d5i < d5_block; d5i++) { )==" "\n" |
| 212 | R"==(const int iter = d0i + d1i + d2i + d3i + d4i + d5i; )==" "\n" |
| 213 | R"==(#if DST_BLOCK_DIM == 0 )==" "\n" |
| 214 | R"==(const int dst_off = DST_OFF(d0 + iter, d1, d2, d3, d4, d5); )==" "\n" |
| 215 | R"==(#elif DST_BLOCK_DIM == 1 )==" "\n" |
| 216 | R"==(const int dst_off = DST_OFF(d0, d1 + iter, d2, d3, d4, d5); )==" "\n" |
| 217 | R"==(#elif DST_BLOCK_DIM == 2 )==" "\n" |
| 218 | R"==(const int dst_off = DST_OFF(d0, d1, d2 + iter, d3, d4, d5); )==" "\n" |
| 219 | R"==(#elif DST_BLOCK_DIM == 3 )==" "\n" |
| 220 | R"==(const int dst_off = DST_OFF(d0, d1, d2, d3 + iter, d4, d5); )==" "\n" |
| 221 | R"==(#elif DST_BLOCK_DIM == 4 )==" "\n" |
| 222 | R"==(const int dst_off = DST_OFF(d0, d1, d2, d3, d4 + iter, d5); )==" "\n" |
| 223 | R"==(#elif DST_BLOCK_DIM == 5 )==" "\n" |
| 224 | R"==(const int dst_off = DST_OFF(d0, d1, d2, d3, d4, d5 + iter); )==" "\n" |
| 225 | R"==(#endif )==" "\n" |
| 226 | R"==(DST_DATA_T dst_tmp; )==" "\n" |
| 227 | R"==(if (!pad) { )==" "\n" |
| 228 | R"==(#if WITH_SUM_SCALE || WITH_SUM_ZPOINT )==" "\n" |
| 229 | R"==(dst_tmp = DST_BLOCK_READ(&dst[dst_off]); )==" "\n" |
| 230 | R"==(#endif )==" "\n" |
| 231 | R"==(#if WITH_SRC_SCALE && SRC_NUM_SCALES > 1 )==" "\n" |
| 232 | R"==(src_scale = src_scales[SCALE_OFF(SRC, d0, d1, d2, d3, d4, d5)]; )==" "\n" |
| 233 | R"==(#endif )==" "\n" |
| 234 | R"==(#if WITH_DST_SCALE && DST_NUM_SCALES > 1 )==" "\n" |
| 235 | R"==(dst_scale = dst_scales[SCALE_OFF(DST, d0, d1, d2, d3, d4, d5)]; )==" "\n" |
| 236 | R"==(#endif )==" "\n" |
| 237 | R"==(#ifdef TRANSPOSE_NXN )==" "\n" |
| 238 | R"==(REORDER(dst_tmp, dst_buf[iter], src_scale, dst_scale, sum_scale, )==" "\n" |
| 239 | R"==(src_zp, dst_zp, sum_zp); )==" "\n" |
| 240 | R"==(#else )==" "\n" |
| 241 | R"==(send_buf = tmp[sg_off + sgId * SUB_GROUP_SIZE + iter]; )==" "\n" |
| 242 | R"==(REORDER(dst_tmp, send_buf, src_scale, dst_scale, sum_scale, src_zp, )==" "\n" |
| 243 | R"==(dst_zp, sum_zp); )==" "\n" |
| 244 | R"==(#endif )==" "\n" |
| 245 | R"==(} else { )==" "\n" |
| 246 | R"==(dst_tmp = 0; )==" "\n" |
| 247 | R"==(} )==" "\n" |
| 248 | R"==(DST_BLOCK_WRITE(&dst[dst_off], dst_tmp); )==" "\n" |
| 249 | R"==(} )==" "\n" |
| 250 | R"==(#elif REORDER_NCHW )==" "\n" |
| 251 | R"==(#define BIGGER_THAN_16 (SRC_D1 >= 16) )==" "\n" |
| 252 | R"==(int sgId = get_sub_group_local_id(); )==" "\n" |
| 253 | R"==(const int d0 = GWS_GET_D0(); )==" "\n" |
| 254 | R"==(const int d1 = GWS_GET_D1(); )==" "\n" |
| 255 | R"==(const int d2 = GWS_GET_D2(); )==" "\n" |
| 256 | R"==(const int d3 = GWS_GET_D3(); )==" "\n" |
| 257 | R"==(const int d1_block = GWS_GET_D1_BLOCK(); )==" "\n" |
| 258 | R"==(SRC_DATA_T src_buf[SUB_GROUP_SIZE]; )==" "\n" |
| 259 | R"==(SRC_DATA_T dst_buf[SUB_GROUP_SIZE]; )==" "\n" |
| 260 | R"==(#if BIGGER_THAN_16 )==" "\n" |
| 261 | R"==(SRC_DATA_T send_buf; )==" "\n" |
| 262 | R"==(#else )==" "\n" |
| 263 | R"==(SRC_DATA_T exch_buf[d1_block][SUB_GROUP_SIZE]; )==" "\n" |
| 264 | R"==(#endif )==" "\n" |
| 265 | R"==(#if BIGGER_THAN_16 )==" "\n" |
| 266 | R"==(#define STRIDE_S SRC_D1 )==" "\n" |
| 267 | R"==(#else )==" "\n" |
| 268 | R"==(#define STRIDE_S 16 )==" "\n" |
| 269 | R"==(#endif )==" "\n" |
| 270 | R"==(#define STRIDE_D (SRC_D2 * SRC_D3) )==" "\n" |
| 271 | R"==(for (int i = 0; i < d1_block; i++) { )==" "\n" |
| 272 | R"==(int src_off = SRC_OFF(d0, d1, d2, d3, 0, 0) + STRIDE_S * i; )==" "\n" |
| 273 | R"==(src_buf[i] = SRC_BLOCK_READ(&src[src_off]); )==" "\n" |
| 274 | R"==(} )==" "\n" |
| 275 | R"==(#if BIGGER_THAN_16 )==" "\n" |
| 276 | R"==(for (int i = 0; i < SUB_GROUP_SIZE; i++) { )==" "\n" |
| 277 | R"==(send_buf = src_buf[(i + sgId) % 16]; )==" "\n" |
| 278 | R"==(dst_buf[(16 + sgId - i) % 16] )==" "\n" |
| 279 | R"==(= intel_sub_group_shuffle(send_buf, (16 + sgId - i) % 16); )==" "\n" |
| 280 | R"==(} )==" "\n" |
| 281 | R"==(#else )==" "\n" |
| 282 | R"==(for (int i = 0; i < d1_block; i++) { )==" "\n" |
| 283 | R"==(for (int sg = 0; sg < SUB_GROUP_SIZE; sg++) { )==" "\n" |
| 284 | R"==(exch_buf[i][sg] = intel_sub_group_shuffle(src_buf[i], sg); )==" "\n" |
| 285 | R"==(} )==" "\n" |
| 286 | R"==(} )==" "\n" |
| 287 | R"==(for (int i = 0; i < d1_block; i++) { )==" "\n" |
| 288 | R"==(int ofs = i + sgId * d1_block; )==" "\n" |
| 289 | R"==(dst_buf[i] = exch_buf[ofs / SUB_GROUP_SIZE][ofs % SUB_GROUP_SIZE]; )==" "\n" |
| 290 | R"==(} )==" "\n" |
| 291 | R"==(#endif )==" "\n" |
| 292 | R"==(for (int i = 0; i < d1_block; i++) { )==" "\n" |
| 293 | R"==(int dst_off = DST_OFF(d0, d1, d2, d3, 0, 0) + STRIDE_D * i; )==" "\n" |
| 294 | R"==(DST_DATA_T dst_tmp; )==" "\n" |
| 295 | R"==(#if WITH_SUM_SCALE || WITH_SUM_ZPOINT )==" "\n" |
| 296 | R"==(dst_tmp = DST_BLOCK_READ(&dst[dst_off]); )==" "\n" |
| 297 | R"==(#endif )==" "\n" |
| 298 | R"==(REORDER(dst_tmp, dst_buf[i], src_scale, dst_scale, sum_scale, src_zp, )==" "\n" |
| 299 | R"==(dst_zp, sum_zp); )==" "\n" |
| 300 | R"==(DST_BLOCK_WRITE(&dst[dst_off], dst_tmp); )==" "\n" |
| 301 | R"==(} )==" "\n" |
| 302 | R"==(#elif PLAIN_TO_ABCD84A42B )==" "\n" |
| 303 | R"==(const int sglid = get_sub_group_local_id(); )==" "\n" |
| 304 | R"==(const int sg = get_sub_group_id(); )==" "\n" |
| 305 | R"==(#define SRC_D3_ALIGNED (SRC_D3 % SUB_GROUP_SIZE == 0) )==" "\n" |
| 306 | R"==(#define IS_SRC_ABCD_LAYOUT (SRC_S3_0 == 1) )==" "\n" |
| 307 | R"==(#define IS_SRC_ACDB_LAYOUT (SRC_S3_0 != 1) )==" "\n" |
| 308 | R"==(const int d0 = GWS_GET_D0(); )==" "\n" |
| 309 | R"==(const int d1 = GWS_GET_D1(); )==" "\n" |
| 310 | R"==(const int d2 = GWS_GET_D2(); )==" "\n" |
| 311 | R"==(const int d3 = GWS_GET_D3(); )==" "\n" |
| 312 | R"==(const int d0_block = GWS_GET_D0_BLOCK(); )==" "\n" |
| 313 | R"==(const int d1_block = GWS_GET_D1_BLOCK(); )==" "\n" |
| 314 | R"==(const int d01_block = d0_block * d1_block; )==" "\n" |
| 315 | R"==(__local SRC_DATA_T loc_buf[SG_PER_WG][d01_block][SUB_GROUP_SIZE]; )==" "\n" |
| 316 | R"==(for (int d0_inner = 0; d0_inner < d0_block; d0_inner++) { )==" "\n" |
| 317 | R"==(for (int d1_inner = 0; d1_inner < d1_block; d1_inner++) { )==" "\n" |
| 318 | R"==(if (SRC_D0 % d0_block != 0 && d0 + d0_inner >= SRC_D0) { )==" "\n" |
| 319 | R"==(loc_buf[sg][d0_inner * d1_block + d1_inner][sglid] = 0; )==" "\n" |
| 320 | R"==(continue; )==" "\n" |
| 321 | R"==(} )==" "\n" |
| 322 | R"==(if (SRC_D1 % d1_block != 0 && d1 + d1_inner >= SRC_D1) { )==" "\n" |
| 323 | R"==(loc_buf[sg][d0_inner * d1_block + d1_inner][sglid] = 0; )==" "\n" |
| 324 | R"==(continue; )==" "\n" |
| 325 | R"==(} )==" "\n" |
| 326 | R"==(if (IS_SRC_ABCD_LAYOUT )==" "\n" |
| 327 | R"==(&& (SRC_D3_ALIGNED || d3 + SUB_GROUP_SIZE < SRC_D3)) { )==" "\n" |
| 328 | R"==(const int src_off )==" "\n" |
| 329 | R"==(= SRC_OFF(d0 + d0_inner, d1 + d1_inner, d2, d3, 0, 0); )==" "\n" |
| 330 | R"==(loc_buf[sg][d0_inner * d1_block + d1_inner][sglid] )==" "\n" |
| 331 | R"==(= src[src_off + sglid]; )==" "\n" |
| 332 | R"==(} else { )==" "\n" |
| 333 | R"==(if (SRC_D3_ALIGNED || d3 + sglid < SRC_D3) { )==" "\n" |
| 334 | R"==(const int src_off = SRC_OFF( )==" "\n" |
| 335 | R"==(d0 + d0_inner, d1 + d1_inner, d2, d3 + sglid, 0, 0); )==" "\n" |
| 336 | R"==(loc_buf[sg][d0_inner * d1_block + d1_inner][sglid] )==" "\n" |
| 337 | R"==(= src[src_off]; )==" "\n" |
| 338 | R"==(} )==" "\n" |
| 339 | R"==(} )==" "\n" |
| 340 | R"==(} )==" "\n" |
| 341 | R"==(} )==" "\n" |
| 342 | R"==(if (SRC_D3_ALIGNED || d3 + SUB_GROUP_SIZE < SRC_D3) { )==" "\n" |
| 343 | R"==(for (int d = 0; d < SUB_GROUP_SIZE; d += 8) { )==" "\n" |
| 344 | R"==(SRC_DATA8_T src_tmp; )==" "\n" |
| 345 | R"==(for (int i = 0; i < 8; i++) { )==" "\n" |
| 346 | R"==(src_tmp[i] = loc_buf[sg][sglid][d + i]; )==" "\n" |
| 347 | R"==(} )==" "\n" |
| 348 | R"==(const int dst_off = DST_OFF(d0, d1, d2, d3 + d, 0, 0); )==" "\n" |
| 349 | R"==(DST_DATA8_T dst_tmp; )==" "\n" |
| 350 | R"==(#if WITH_SUM_SCALE || WITH_SUM_ZPOINT )==" "\n" |
| 351 | R"==(for (int n = 0; n < 8; n++) { )==" "\n" |
| 352 | R"==(dst_tmp[n] = dst[dst_off + sglid + n * SUB_GROUP_SIZE]; )==" "\n" |
| 353 | R"==(} )==" "\n" |
| 354 | R"==(#endif )==" "\n" |
| 355 | R"==(REORDER8(dst_tmp, src_tmp, src_scale, dst_scale, sum_scale, src_zp, )==" "\n" |
| 356 | R"==(dst_zp, sum_zp); )==" "\n" |
| 357 | R"==(for (int n = 0; n < 8; n++) { )==" "\n" |
| 358 | R"==(dst[dst_off + sglid + n * SUB_GROUP_SIZE] = dst_tmp[n]; )==" "\n" |
| 359 | R"==(} )==" "\n" |
| 360 | R"==(} )==" "\n" |
| 361 | R"==(} else { )==" "\n" |
| 362 | R"==(for (int d = 0; d < SUB_GROUP_SIZE; d++) { )==" "\n" |
| 363 | R"==(if (d3 + d < SRC_D3) { )==" "\n" |
| 364 | R"==(const SRC_DATA_T src_tmp = loc_buf[sg][sglid][d]; )==" "\n" |
| 365 | R"==(const int dst_off = DST_OFF(d0, d1, d2, d3 + d, 0, 0); )==" "\n" |
| 366 | R"==(DST_DATA_T dst_tmp; )==" "\n" |
| 367 | R"==(#if WITH_SUM_SCALE || WITH_SUM_ZPOINT )==" "\n" |
| 368 | R"==(dst_tmp[n] = dst[dst_off + sglid]; )==" "\n" |
| 369 | R"==(#endif )==" "\n" |
| 370 | R"==(REORDER(dst_tmp, src_tmp, src_scale, dst_scale, sum_scale, )==" "\n" |
| 371 | R"==(src_zp, dst_zp, sum_zp); )==" "\n" |
| 372 | R"==(dst[dst_off + sglid] = dst_tmp; )==" "\n" |
| 373 | R"==(} )==" "\n" |
| 374 | R"==(} )==" "\n" |
| 375 | R"==(} )==" "\n" |
| 376 | R"==(#elif PLAIN_TO_AB_XX_8AYB )==" "\n" |
| 377 | R"==(int sgId = get_sub_group_local_id(); )==" "\n" |
| 378 | R"==(const int d0 = GWS_GET_D0(); )==" "\n" |
| 379 | R"==(const int d1 = GWS_GET_D1(); )==" "\n" |
| 380 | R"==(const int d0b = GWS_GET_D0_BLOCK(); )==" "\n" |
| 381 | R"==(const int d1b = GWS_GET_D1_BLOCK(); )==" "\n" |
| 382 | R"==(SRC_DATA_T src_buf[d0b]; )==" "\n" |
| 383 | R"==(DST_DATA_T dst_buf[d0b]; )==" "\n" |
| 384 | R"==(for (int d0i = 0; d0i < d0b; ++d0i) { )==" "\n" |
| 385 | R"==(const int src_off = SRC_OFF(d0 + d0i, d1, 0, 0, 0, 0); )==" "\n" |
| 386 | R"==(src_buf[d0i] = SRC_BLOCK_READ(&src[src_off]); )==" "\n" |
| 387 | R"==(} )==" "\n" |
| 388 | R"==(SRC_DATA_T tmp_buf[d0b][SUB_GROUP_SIZE]; )==" "\n" |
| 389 | R"==(for (int i = 0; i < d0b; i++) { )==" "\n" |
| 390 | R"==(for (int sg = 0; sg < SUB_GROUP_SIZE; sg++) { )==" "\n" |
| 391 | R"==(tmp_buf[i][sg] = intel_sub_group_shuffle(src_buf[i], sg); )==" "\n" |
| 392 | R"==(} )==" "\n" |
| 393 | R"==(} )==" "\n" |
| 394 | R"==(#if BLK_L == 4 )==" "\n" |
| 395 | R"==(for (int d0i = 0; d0i < d0b; ++d0i) { )==" "\n" |
| 396 | R"==(dst_buf[d0i] = tmp_buf[(d0i % 2 * BLK_L) + sgId / BLK_L] )==" "\n" |
| 397 | R"==([(d0i / 2) * BLK_L + sgId % BLK_L]; )==" "\n" |
| 398 | R"==(} )==" "\n" |
| 399 | R"==(#else )==" "\n" |
| 400 | R"==(for (int d0i = 0; d0i < d0b; ++d0i) { )==" "\n" |
| 401 | R"==(dst_buf[d0i] = tmp_buf[sgId / BLK_L][d0i * BLK_L + sgId % BLK_L]; )==" "\n" |
| 402 | R"==(} )==" "\n" |
| 403 | R"==(#endif )==" "\n" |
| 404 | R"==(for (int d0i = 0; d0i < d0b; ++d0i) { )==" "\n" |
| 405 | R"==(const int dst_off = DST_OFF(d0, d1, 0, 0, 0, 0) + SUB_GROUP_SIZE * d0i; )==" "\n" |
| 406 | R"==(DST_DATA_T dst_tmp; )==" "\n" |
| 407 | R"==(#if WITH_SUM_SCALE || WITH_SUM_ZPOINT )==" "\n" |
| 408 | R"==(dst_tmp = DST_BLOCK_READ(&dst[dst_off]); )==" "\n" |
| 409 | R"==(#endif )==" "\n" |
| 410 | R"==(REORDER(dst_tmp, dst_buf[d0i], src_scale, dst_scale, sum_scale, src_zp, )==" "\n" |
| 411 | R"==(dst_zp, sum_zp); )==" "\n" |
| 412 | R"==(DST_BLOCK_WRITE(&dst[dst_off], dst_tmp); )==" "\n" |
| 413 | R"==(} )==" "\n" |
| 414 | R"==(#elif VECTORIZE_LAST_DIM )==" "\n" |
| 415 | R"==(const int d0 = GWS_GET_D0(); )==" "\n" |
| 416 | R"==(const int d1 = GWS_GET_D1(); )==" "\n" |
| 417 | R"==(const int d2 = GWS_GET_D2(); )==" "\n" |
| 418 | R"==(const int d3 = GWS_GET_D3(); )==" "\n" |
| 419 | R"==(const int d4 = GWS_GET_D4(); )==" "\n" |
| 420 | R"==(const int d5 = GWS_GET_D5(); )==" "\n" |
| 421 | R"==(const int d0_block = GWS_GET_D0_BLOCK(); )==" "\n" |
| 422 | R"==(const int d1_block = GWS_GET_D1_BLOCK(); )==" "\n" |
| 423 | R"==(const int d2_block = GWS_GET_D2_BLOCK(); )==" "\n" |
| 424 | R"==(const int d3_block = GWS_GET_D3_BLOCK(); )==" "\n" |
| 425 | R"==(const int d4_block = GWS_GET_D4_BLOCK(); )==" "\n" |
| 426 | R"==(for_(int d0i = 0; d0i < d0_block; d0i++) )==" "\n" |
| 427 | R"==(for_(int d1i = 0; d1i < d1_block; d1i++) )==" "\n" |
| 428 | R"==(for_(int d2i = 0; d2i < d2_block; d2i++) )==" "\n" |
| 429 | R"==(for_(int d3i = 0; d3i < d3_block; d3i++) )==" "\n" |
| 430 | R"==(for (int d4i = 0; d4i < d4_block; d4i++) { )==" "\n" |
| 431 | R"==(int src_off )==" "\n" |
| 432 | R"==(= SRC_OFF(d0 + d0i, d1 + d1i, d2 + d2i, d3 + d3i, d4 + d4i, d5); )==" "\n" |
| 433 | R"==(SRC_DATA_T src_tmp = SRC_BLOCK_READ(&src[src_off]); )==" "\n" |
| 434 | R"==(int dst_off )==" "\n" |
| 435 | R"==(= DST_OFF(d0 + d0i, d1 + d1i, d2 + d2i, d3 + d3i, d4 + d4i, d5); )==" "\n" |
| 436 | R"==(DST_DATA_T dst_tmp; )==" "\n" |
| 437 | R"==(#if WITH_SUM_SCALE || WITH_SUM_ZPOINT )==" "\n" |
| 438 | R"==(dst_tmp = DST_BLOCK_READ(&dst[dst_off]); )==" "\n" |
| 439 | R"==(#endif )==" "\n" |
| 440 | R"==(REORDER(dst_tmp, src_tmp, src_scale, dst_scale, sum_scale, src_zp, )==" "\n" |
| 441 | R"==(dst_zp, sum_zp); )==" "\n" |
| 442 | R"==(DST_BLOCK_WRITE(&dst[dst_off], dst_tmp); )==" "\n" |
| 443 | R"==(} )==" "\n" |
| 444 | R"==(#elif PAD_INNERMOST )==" "\n" |
| 445 | R"==(const int sgId = get_sub_group_local_id(); )==" "\n" |
| 446 | R"==(int d[6]; )==" "\n" |
| 447 | R"==(int blk[6]; )==" "\n" |
| 448 | R"==(int b[6] = {0, 0, 0, 0, 0, 0}; )==" "\n" |
| 449 | R"==(d[0] = GWS_GET_D0(); )==" "\n" |
| 450 | R"==(d[1] = GWS_GET_D1(); )==" "\n" |
| 451 | R"==(d[2] = GWS_GET_D2(); )==" "\n" |
| 452 | R"==(d[3] = GWS_GET_D3(); )==" "\n" |
| 453 | R"==(d[4] = GWS_GET_D4(); )==" "\n" |
| 454 | R"==(d[5] = GWS_GET_D5(); )==" "\n" |
| 455 | R"==(blk[0] = GWS_GET_D0_BLOCK(); )==" "\n" |
| 456 | R"==(blk[1] = GWS_GET_D1_BLOCK(); )==" "\n" |
| 457 | R"==(blk[2] = GWS_GET_D2_BLOCK(); )==" "\n" |
| 458 | R"==(blk[3] = GWS_GET_D3_BLOCK(); )==" "\n" |
| 459 | R"==(blk[4] = GWS_GET_D4_BLOCK(); )==" "\n" |
| 460 | R"==(blk[5] = GWS_GET_D5_BLOCK(); )==" "\n" |
| 461 | R"==(__local SRC_DATA_T cache[SG_PER_WG * GROUP * GROUP * VECT_SIZE]; )==" "\n" |
| 462 | R"==(const int sg_off = get_sub_group_id() * VECT_SIZE * (GROUP * GROUP); )==" "\n" |
| 463 | R"==(for (int i = 0; i < blk[SRC_LOOP_DIM]; i++) { )==" "\n" |
| 464 | R"==(b[SRC_LOOP_DIM] = i; )==" "\n" |
| 465 | R"==(const int src_off = SRC_OFF(d[0] + b[0], d[1] + b[1], d[2] + b[2], )==" "\n" |
| 466 | R"==(d[3] + b[3], d[4] + b[4], d[5] + b[5]); )==" "\n" |
| 467 | R"==(unroll_for(int j = 0; j < GROUP; j++) { )==" "\n" |
| 468 | R"==(const int coff = sg_off + VECT_SIZE * GROUP * i + VECT_SIZE * j; )==" "\n" |
| 469 | R"==(const int soff = src_off + INNERMOST_SIZE * j; )==" "\n" |
| 470 | R"==(#if NON_INNERMOST_PADDING == 1 )==" "\n" |
| 471 | R"==(const int pad_d0 = d[0] + b[0] >= SRC_D0; )==" "\n" |
| 472 | R"==(const int pad_d1 = NDIMS > 1 && d[1] + b[1] >= SRC_D1; )==" "\n" |
| 473 | R"==(const int pad_d2 = NDIMS > 2 && d[2] + b[2] >= SRC_D2; )==" "\n" |
| 474 | R"==(const int pad_d3 = NDIMS > 3 && d[3] + b[3] >= SRC_D3; )==" "\n" |
| 475 | R"==(const int pad_d4 = NDIMS > 4 && d[4] + b[4] >= SRC_D4; )==" "\n" |
| 476 | R"==(const int pad_d5 = NDIMS > 5 && d[5] + b[5] >= SRC_D5; )==" "\n" |
| 477 | R"==(if (pad_d0 || pad_d1 || pad_d2 || pad_d3 || pad_d4 || pad_d5) { )==" "\n" |
| 478 | R"==(cache[coff + sgId] = 0; )==" "\n" |
| 479 | R"==(continue; )==" "\n" |
| 480 | R"==(} )==" "\n" |
| 481 | R"==(#endif )==" "\n" |
| 482 | R"==(if (sgId < INNERMOST_SIZE) { )==" "\n" |
| 483 | R"==(cache[coff + sgId] = src[soff + sgId]; )==" "\n" |
| 484 | R"==(} else { )==" "\n" |
| 485 | R"==(cache[coff + sgId] = 0; )==" "\n" |
| 486 | R"==(} )==" "\n" |
| 487 | R"==(} )==" "\n" |
| 488 | R"==(} )==" "\n" |
| 489 | R"==(b[SRC_LOOP_DIM] = 0; )==" "\n" |
| 490 | R"==(for (int i = 0; i < blk[DST_LOOP_DIM]; i++) { )==" "\n" |
| 491 | R"==(b[DST_LOOP_DIM] = i; )==" "\n" |
| 492 | R"==(const int dst_off = DST_OFF(d[0] + b[0], d[1] + b[1], d[2] + b[2], )==" "\n" |
| 493 | R"==(d[3] + b[3], d[4] + b[4], d[5] + b[5]); )==" "\n" |
| 494 | R"==(#if WITH_SRC_SCALE && SRC_NUM_SCALES > 1 )==" "\n" |
| 495 | R"==(src_scale = src_scales[SCALE_OFF(SRC, d[0] + b[0], d[1] + b[1], )==" "\n" |
| 496 | R"==(d[2] + b[2], d[3] + b[3], d[4] + b[4], d[5] + b[5])]; )==" "\n" |
| 497 | R"==(#endif )==" "\n" |
| 498 | R"==(#if WITH_DST_SCALE && DST_NUM_SCALES > 1 )==" "\n" |
| 499 | R"==(dst_scale = dst_scales[SCALE_OFF(DST, d[0] + b[0], d[1] + b[1], )==" "\n" |
| 500 | R"==(d[2] + b[2], d[3] + b[3], d[4] + b[4], d[5] + b[5])]; )==" "\n" |
| 501 | R"==(#endif )==" "\n" |
| 502 | R"==(unroll_for(int j = 0; j < GROUP; j++) { )==" "\n" |
| 503 | R"==(const int coff = sg_off + VECT_SIZE * GROUP * j + VECT_SIZE * i; )==" "\n" |
| 504 | R"==(const int doff = dst_off + VECT_SIZE * j; )==" "\n" |
| 505 | R"==(DST_DATA_T dst_tmp; )==" "\n" |
| 506 | R"==(#if WITH_SUM_SCALE || WITH_SUM_ZPOINT )==" "\n" |
| 507 | R"==(dst_tmp = dst[doff + sgId * DST_INNERMOST_STRIDE]; )==" "\n" |
| 508 | R"==(#endif )==" "\n" |
| 509 | R"==(REORDER(dst_tmp, cache[coff + sgId], src_scale, dst_scale, )==" "\n" |
| 510 | R"==(sum_scale, src_zp, dst_zp, sum_zp); )==" "\n" |
| 511 | R"==(dst[doff + sgId * DST_INNERMOST_STRIDE] = dst_tmp; )==" "\n" |
| 512 | R"==(} )==" "\n" |
| 513 | R"==(} )==" "\n" |
| 514 | R"==(#elif VECTORIZE_GROUPS )==" "\n" |
| 515 | R"==(int d[6]; )==" "\n" |
| 516 | R"==(int blk[6]; )==" "\n" |
| 517 | R"==(int b[6] = {0, 0, 0, 0, 0, 0}; )==" "\n" |
| 518 | R"==(d[0] = GWS_GET_D0(); )==" "\n" |
| 519 | R"==(d[1] = GWS_GET_D1(); )==" "\n" |
| 520 | R"==(d[2] = GWS_GET_D2(); )==" "\n" |
| 521 | R"==(d[3] = GWS_GET_D3(); )==" "\n" |
| 522 | R"==(d[4] = GWS_GET_D4(); )==" "\n" |
| 523 | R"==(d[5] = GWS_GET_D5(); )==" "\n" |
| 524 | R"==(blk[0] = GWS_GET_D0_BLOCK(); )==" "\n" |
| 525 | R"==(blk[1] = GWS_GET_D1_BLOCK(); )==" "\n" |
| 526 | R"==(blk[2] = GWS_GET_D2_BLOCK(); )==" "\n" |
| 527 | R"==(blk[3] = GWS_GET_D3_BLOCK(); )==" "\n" |
| 528 | R"==(blk[4] = GWS_GET_D4_BLOCK(); )==" "\n" |
| 529 | R"==(blk[5] = GWS_GET_D5_BLOCK(); )==" "\n" |
| 530 | R"==(SRC_DATA_T cache[GROUP * GROUP]; )==" "\n" |
| 531 | R"==(int coeff = (VECT_DIM == SRC_LOOP_DIM ? get_sub_group_size() : 1); )==" "\n" |
| 532 | R"==(for (int i = 0; i < blk[SRC_LOOP_DIM]; i++) { )==" "\n" |
| 533 | R"==(b[SRC_LOOP_DIM] = coeff * i; )==" "\n" |
| 534 | R"==(const int src_off = SRC_OFF(d[0] + b[0], d[1] + b[1], d[2] + b[2], )==" "\n" |
| 535 | R"==(d[3] + b[3], d[4] + b[4], d[5] + b[5]); )==" "\n" |
| 536 | R"==(unroll_for(int j = 0; j < GROUP; j++) { )==" "\n" |
| 537 | R"==(int cidx = GROUP * i + j; )==" "\n" |
| 538 | R"==(int sidx = src_off + get_sub_group_size() * j; )==" "\n" |
| 539 | R"==(#if PAD_FILL_ZERO == 1 )==" "\n" |
| 540 | R"==(int pad_d0 = d[0] + b[0] >= SRC_D0; )==" "\n" |
| 541 | R"==(int pad_d1 = NDIMS > 1 && d[1] + b[1] >= SRC_D1; )==" "\n" |
| 542 | R"==(int pad_d2 = NDIMS > 2 && d[2] + b[2] >= SRC_D2; )==" "\n" |
| 543 | R"==(int pad_d3 = NDIMS > 3 && d[3] + b[3] >= SRC_D3; )==" "\n" |
| 544 | R"==(int pad_d4 = NDIMS > 4 && d[4] + b[4] >= SRC_D4; )==" "\n" |
| 545 | R"==(int pad_d5 = NDIMS > 5 && d[5] + b[5] >= SRC_D5; )==" "\n" |
| 546 | R"==(if (pad_d0 || pad_d1 || pad_d2 || pad_d3 || pad_d4 || pad_d5) { )==" "\n" |
| 547 | R"==(cache[cidx] = 0; )==" "\n" |
| 548 | R"==(continue; )==" "\n" |
| 549 | R"==(} )==" "\n" |
| 550 | R"==(#endif )==" "\n" |
| 551 | R"==(cache[cidx] = SRC_BLOCK_READ(&src[sidx]); )==" "\n" |
| 552 | R"==(} )==" "\n" |
| 553 | R"==(} )==" "\n" |
| 554 | R"==(b[SRC_LOOP_DIM] = 0; )==" "\n" |
| 555 | R"==(coeff = (VECT_DIM == DST_LOOP_DIM ? get_sub_group_size() : 1); )==" "\n" |
| 556 | R"==(for (int i = 0; i < blk[DST_LOOP_DIM]; i++) { )==" "\n" |
| 557 | R"==(b[DST_LOOP_DIM] = coeff * i; )==" "\n" |
| 558 | R"==(const int dst_off = DST_OFF(d[0] + b[0], d[1] + b[1], d[2] + b[2], )==" "\n" |
| 559 | R"==(d[3] + b[3], d[4] + b[4], d[5] + b[5]); )==" "\n" |
| 560 | R"==(int pad_d0 = d[0] + b[0] >= DST_PD0; )==" "\n" |
| 561 | R"==(int pad_d1 = NDIMS > 1 && d[1] + b[1] >= DST_PD1; )==" "\n" |
| 562 | R"==(int pad_d2 = NDIMS > 2 && d[2] + b[2] >= DST_PD2; )==" "\n" |
| 563 | R"==(int pad_d3 = NDIMS > 3 && d[3] + b[3] >= DST_PD3; )==" "\n" |
| 564 | R"==(int pad_d4 = NDIMS > 4 && d[4] + b[4] >= DST_PD4; )==" "\n" |
| 565 | R"==(int pad_d5 = NDIMS > 5 && d[5] + b[5] >= DST_PD5; )==" "\n" |
| 566 | R"==(if (pad_d0 || pad_d1 || pad_d2 || pad_d3 || pad_d4 || pad_d5) { )==" "\n" |
| 567 | R"==(continue; )==" "\n" |
| 568 | R"==(} )==" "\n" |
| 569 | R"==(unroll_for(int j = 0; j < GROUP; j++) { )==" "\n" |
| 570 | R"==(int cidx = i + j * GROUP; )==" "\n" |
| 571 | R"==(int didx = dst_off + get_sub_group_size() * j; )==" "\n" |
| 572 | R"==(DST_DATA_T dst_tmp; )==" "\n" |
| 573 | R"==(#if WITH_SUM_SCALE || WITH_SUM_ZPOINT )==" "\n" |
| 574 | R"==(dst_tmp = DST_BLOCK_READ(&dst[didx]); )==" "\n" |
| 575 | R"==(#endif )==" "\n" |
| 576 | R"==(#if WITH_SRC_SCALE && SRC_NUM_SCALES > 1 )==" "\n" |
| 577 | R"==(src_scale = src_scales[SCALE_OFF(SRC, d[0] + b[0], d[1] + b[1], )==" "\n" |
| 578 | R"==(d[2] + b[2], d[3] + b[3], d[4] + b[4], d[5] + b[5])]; )==" "\n" |
| 579 | R"==(#endif )==" "\n" |
| 580 | R"==(#if WITH_DST_SCALE && DST_NUM_SCALES > 1 )==" "\n" |
| 581 | R"==(dst_scale = dst_scales[SCALE_OFF(DST, d[0] + b[0], d[1] + b[1], )==" "\n" |
| 582 | R"==(d[2] + b[2], d[3] + b[3], d[4] + b[4], d[5] + b[5])]; )==" "\n" |
| 583 | R"==(#endif )==" "\n" |
| 584 | R"==(REORDER(dst_tmp, cache[cidx], src_scale, dst_scale, sum_scale, )==" "\n" |
| 585 | R"==(src_zp, dst_zp, sum_zp); )==" "\n" |
| 586 | R"==(DST_BLOCK_WRITE(&dst[didx], dst_tmp); )==" "\n" |
| 587 | R"==(} )==" "\n" |
| 588 | R"==(} )==" "\n" |
| 589 | R"==(#elif USE_DENSE_VECT )==" "\n" |
| 590 | R"==(const int d0_blk_start = GWS_GET_D0(); )==" "\n" |
| 591 | R"==(const int d0_blk_end = d0_blk_start + (GWS_GET_D0_BLOCK() * 16); )==" "\n" |
| 592 | R"==(for (int d0 = d0_blk_start; d0 < d0_blk_end; d0 += 128) { )==" "\n" |
| 593 | R"==(SRC_DATA8_T src_tmp = SRC_BLOCK_READ8(&src[d0]); )==" "\n" |
| 594 | R"==(DST_DATA8_T dst_tmp; )==" "\n" |
| 595 | R"==(#if WITH_SUM_SCALE || WITH_SUM_ZPOINT )==" "\n" |
| 596 | R"==(dst_tmp = DST_BLOCK_READ8(&dst[d0]); )==" "\n" |
| 597 | R"==(#endif )==" "\n" |
| 598 | R"==(REORDER8(dst_tmp, src_tmp, src_scale, dst_scale, sum_scale, src_zp, )==" "\n" |
| 599 | R"==(dst_zp, sum_zp); )==" "\n" |
| 600 | R"==(DST_BLOCK_WRITE8(&dst[d0], dst_tmp); )==" "\n" |
| 601 | R"==(} )==" "\n" |
| 602 | R"==(#elif XAB_XBA )==" "\n" |
| 603 | R"==(const int sgId = get_sub_group_local_id(); )==" "\n" |
| 604 | R"==(const int sgNr = get_sub_group_id(); )==" "\n" |
| 605 | R"==(const int wgId )==" "\n" |
| 606 | R"==(= get_group_id(0) + 10 * get_group_id(1) + 100 * get_group_id(2); )==" "\n" |
| 607 | R"==(int d[6]; )==" "\n" |
| 608 | R"==(int blk[6]; )==" "\n" |
| 609 | R"==(int b[6] = {0, 0, 0, 0, 0, 0}; )==" "\n" |
| 610 | R"==(d[0] = GWS_GET_D0(); )==" "\n" |
| 611 | R"==(d[1] = GWS_GET_D1(); )==" "\n" |
| 612 | R"==(d[2] = GWS_GET_D2(); )==" "\n" |
| 613 | R"==(d[3] = GWS_GET_D3(); )==" "\n" |
| 614 | R"==(d[4] = GWS_GET_D4(); )==" "\n" |
| 615 | R"==(d[5] = GWS_GET_D5(); )==" "\n" |
| 616 | R"==(blk[0] = GWS_GET_D0_BLOCK(); )==" "\n" |
| 617 | R"==(blk[1] = GWS_GET_D1_BLOCK(); )==" "\n" |
| 618 | R"==(blk[2] = GWS_GET_D2_BLOCK(); )==" "\n" |
| 619 | R"==(blk[3] = GWS_GET_D3_BLOCK(); )==" "\n" |
| 620 | R"==(blk[4] = GWS_GET_D4_BLOCK(); )==" "\n" |
| 621 | R"==(blk[5] = GWS_GET_D5_BLOCK(); )==" "\n" |
| 622 | R"==(__local SRC_DATA_T tmp[SG_PER_WG * SUB_GROUP_SIZE * BLOCK_SIZE]; )==" "\n" |
| 623 | R"==(SRC_DATA_T data; )==" "\n" |
| 624 | R"==(const int sgLs = SUB_GROUP_SIZE * BLOCK_SIZE; )==" "\n" |
| 625 | R"==(const int wg_off = sgNr * sgLs; )==" "\n" |
| 626 | R"==(for (int i = 0; i < blk[SRC_BLK_DIM]; i++) { )==" "\n" |
| 627 | R"==(b[SRC_BLK_DIM] = SRC_OFF_COEFF * i; )==" "\n" |
| 628 | R"==(const int src_off = SRC_OFF(d[0] + b[0], d[1] + b[1], d[2] + b[2], )==" "\n" |
| 629 | R"==(d[3] + b[3], d[4] + b[4], d[5] + b[5]); )==" "\n" |
| 630 | R"==(#if PAD_FILL_ZERO == 1 )==" "\n" |
| 631 | R"==(const int pad_d0 = d[0] + b[0] >= SRC_D0; )==" "\n" |
| 632 | R"==(const int pad_d1 = NDIMS > 1 && d[1] + b[1] >= SRC_D1; )==" "\n" |
| 633 | R"==(const int pad_d2 = NDIMS > 2 && d[2] + b[2] >= SRC_D2; )==" "\n" |
| 634 | R"==(const int pad_d3 = NDIMS > 3 && d[3] + b[3] >= SRC_D3; )==" "\n" |
| 635 | R"==(const int pad_d4 = NDIMS > 4 && d[4] + b[4] >= SRC_D4; )==" "\n" |
| 636 | R"==(const int pad_d5 = NDIMS > 5 && d[5] + b[5] >= SRC_D5; )==" "\n" |
| 637 | R"==(if (pad_d0 || pad_d1 || pad_d2 || pad_d3 || pad_d4 || pad_d5) { )==" "\n" |
| 638 | R"==(tmp[wg_off + SUB_GROUP_SIZE * i + sgId] = 0; )==" "\n" |
| 639 | R"==(continue; )==" "\n" |
| 640 | R"==(} )==" "\n" |
| 641 | R"==(#endif )==" "\n" |
| 642 | R"==(data = src[src_off + sgId]; )==" "\n" |
| 643 | R"==(tmp[wg_off + SUB_GROUP_SIZE * i + sgId] = data; )==" "\n" |
| 644 | R"==(} )==" "\n" |
| 645 | R"==(b[SRC_BLK_DIM] = 0; )==" "\n" |
| 646 | R"==(for (int i = 0; i < blk[SRC_BLK_DIM]; i++) { )==" "\n" |
| 647 | R"==(b[DST_BLK_DIM] = DST_OFF_COEFF * i; )==" "\n" |
| 648 | R"==(const int dst_off = DST_OFF(d[0] + b[0], d[1] + b[1], d[2] + b[2], )==" "\n" |
| 649 | R"==(d[3] + b[3], d[4] + b[4], d[5] + b[5]); )==" "\n" |
| 650 | R"==(#if XB_TO_XAB )==" "\n" |
| 651 | R"==(SRC_DATA_T data = tmp[wg_off + (sgId % (SUB_GROUP_SIZE / BLOCK_SIZE)) )==" "\n" |
| 652 | R"==(+ (sgId / (SUB_GROUP_SIZE / BLOCK_SIZE)) * SUB_GROUP_SIZE )==" "\n" |
| 653 | R"==(+ (SUB_GROUP_SIZE / BLOCK_SIZE) * i]; )==" "\n" |
| 654 | R"==(#else )==" "\n" |
| 655 | R"==(SRC_DATA_T data = tmp[wg_off + (sgId % BLOCK_SIZE) * SUB_GROUP_SIZE )==" "\n" |
| 656 | R"==(+ sgId / BLOCK_SIZE + (SUB_GROUP_SIZE / BLOCK_SIZE) * i]; )==" "\n" |
| 657 | R"==(#endif )==" "\n" |
| 658 | R"==(#if WITH_SRC_SCALE && SRC_NUM_SCALES > 1 )==" "\n" |
| 659 | R"==(src_scale = src_scales[SCALE_OFF(SRC, d[0] + b[0], d[1] + b[1], )==" "\n" |
| 660 | R"==(d[2] + b[2], d[3] + b[3], d[4] + b[4], d[5] + b[5])]; )==" "\n" |
| 661 | R"==(#endif )==" "\n" |
| 662 | R"==(#if WITH_DST_SCALE && DST_NUM_SCALES > 1 )==" "\n" |
| 663 | R"==(dst_scale = dst_scales[SCALE_OFF(DST, d[0] + b[0], d[1] + b[1], )==" "\n" |
| 664 | R"==(d[2] + b[2], d[3] + b[3], d[4] + b[4], d[5] + b[5])]; )==" "\n" |
| 665 | R"==(#endif )==" "\n" |
| 666 | R"==(DST_DATA_T dst_tmp; )==" "\n" |
| 667 | R"==(#if WITH_SUM_SCALE || WITH_SUM_ZPOINT )==" "\n" |
| 668 | R"==(dst_tmp = dst[dst_off + sgId]; )==" "\n" |
| 669 | R"==(#endif )==" "\n" |
| 670 | R"==(REORDER(dst_tmp, data, src_scale, dst_scale, sum_scale, src_zp, dst_zp, )==" "\n" |
| 671 | R"==(sum_zp); )==" "\n" |
| 672 | R"==(dst[dst_off + sgId] = dst_tmp; )==" "\n" |
| 673 | R"==(} )==" "\n" |
| 674 | R"==(#else )==" "\n" |
| 675 | R"==(const int d0 = GWS_GET_D0(); )==" "\n" |
| 676 | R"==(const int d1 = GWS_GET_D1(); )==" "\n" |
| 677 | R"==(const int d2 = GWS_GET_D2(); )==" "\n" |
| 678 | R"==(const int d3 = GWS_GET_D3(); )==" "\n" |
| 679 | R"==(const int d4 = GWS_GET_D4(); )==" "\n" |
| 680 | R"==(const int d5 = GWS_GET_D5(); )==" "\n" |
| 681 | R"==(const int local_id = get_sub_group_local_id(); )==" "\n" |
| 682 | R"==(#if SRC_16A16B || DST_16A16B || SRC_16B16A || DST_16B16A )==" "\n" |
| 683 | R"==(src += SRC_OFF(d0, d1, d2, d3, d4, d5); )==" "\n" |
| 684 | R"==(dst += DST_OFF(d0, d1, d2, d3, d4, d5); )==" "\n" |
| 685 | R"==(SRC_DATA8_T in0, in1; )==" "\n" |
| 686 | R"==(#if SRC_16A16B || SRC_16B16A )==" "\n" |
| 687 | R"==(in0 = SRC_BLOCK_READ8(&src[0]); )==" "\n" |
| 688 | R"==(in1 = SRC_BLOCK_READ8(&src[8 * 16]); )==" "\n" |
| 689 | R"==(#else )==" "\n" |
| 690 | R"==(for (int i = 0; i < 8; i++) { )==" "\n" |
| 691 | R"==(#if DST_16B16A )==" "\n" |
| 692 | R"==(in0[i] = src[SRC_OFF(local_id, i, 0, 0, 0, 0)]; )==" "\n" |
| 693 | R"==(in1[i] = src[SRC_OFF(local_id, i + 8, 0, 0, 0, 0)]; )==" "\n" |
| 694 | R"==(#else )==" "\n" |
| 695 | R"==(in0[i] = src[SRC_OFF(i, local_id, 0, 0, 0, 0)]; )==" "\n" |
| 696 | R"==(in1[i] = src[SRC_OFF(i + 8, local_id, 0, 0, 0, 0)]; )==" "\n" |
| 697 | R"==(#endif )==" "\n" |
| 698 | R"==(} )==" "\n" |
| 699 | R"==(#endif )==" "\n" |
| 700 | R"==(DST_DATA8_T dst0, dst1; )==" "\n" |
| 701 | R"==(#if (SRC_16A16B || SRC_16B16A) && (DST_16A16B || DST_16B16A) )==" "\n" |
| 702 | R"==(#if WITH_SUM_SCALE || WITH_SUM_ZPOINT )==" "\n" |
| 703 | R"==(for (int i = 0; i < 8; i++) { )==" "\n" |
| 704 | R"==(#if SRC_16B16A )==" "\n" |
| 705 | R"==(dst0[i] = dst[DST_OFF(local_id, i + 0, 0, 0, 0, 0)]; )==" "\n" |
| 706 | R"==(dst1[i] = dst[DST_OFF(local_id, i + 8, 0, 0, 0, 0)]; )==" "\n" |
| 707 | R"==(#else )==" "\n" |
| 708 | R"==(dst0[i] = dst[DST_OFF(i + 0, local_id, 0, 0, 0, 0)]; )==" "\n" |
| 709 | R"==(dst1[i] = dst[DST_OFF(i + 8, local_id, 0, 0, 0, 0)]; )==" "\n" |
| 710 | R"==(#endif )==" "\n" |
| 711 | R"==(} )==" "\n" |
| 712 | R"==(#endif )==" "\n" |
| 713 | R"==(#elif DST_16A16B || DST_16B16A )==" "\n" |
| 714 | R"==(#if WITH_SUM_SCALE || WITH_SUM_ZPOINT )==" "\n" |
| 715 | R"==(dst0 = DST_BLOCK_READ8(&dst[0]); )==" "\n" |
| 716 | R"==(dst1 = DST_BLOCK_READ8(&dst[8 * 16]); )==" "\n" |
| 717 | R"==(#endif )==" "\n" |
| 718 | R"==(#else )==" "\n" |
| 719 | R"==(#if WITH_SUM_SCALE || WITH_SUM_ZPOINT )==" "\n" |
| 720 | R"==(for (int i = 0; i < 8; i++) { )==" "\n" |
| 721 | R"==(#if SRC_16B16A )==" "\n" |
| 722 | R"==(dst0[i] = dst[DST_OFF(local_id, i + 0, 0, 0, 0, 0)]; )==" "\n" |
| 723 | R"==(dst1[i] = dst[DST_OFF(local_id, i + 8, 0, 0, 0, 0)]; )==" "\n" |
| 724 | R"==(#else )==" "\n" |
| 725 | R"==(dst0[i] = dst[DST_OFF(i + 0, local_id, 0, 0, 0, 0)]; )==" "\n" |
| 726 | R"==(dst1[i] = dst[DST_OFF(i + 8, local_id, 0, 0, 0, 0)]; )==" "\n" |
| 727 | R"==(#endif )==" "\n" |
| 728 | R"==(} )==" "\n" |
| 729 | R"==(#endif )==" "\n" |
| 730 | R"==(#endif )==" "\n" |
| 731 | R"==(REORDER8( )==" "\n" |
| 732 | R"==(dst0, in0, src_scale, dst_scale, sum_scale, src_zp, dst_zp, sum_zp); )==" "\n" |
| 733 | R"==(REORDER8( )==" "\n" |
| 734 | R"==(dst1, in1, src_scale, dst_scale, sum_scale, src_zp, dst_zp, sum_zp); )==" "\n" |
| 735 | R"==(#if (SRC_16A16B || SRC_16B16A) && (DST_16A16B || DST_16B16A) )==" "\n" |
| 736 | R"==(for (int i = 0; i < 8; i++) { )==" "\n" |
| 737 | R"==(#if SRC_16B16A )==" "\n" |
| 738 | R"==(dst[DST_OFF(local_id, i + 0, 0, 0, 0, 0)] = dst0[i]; )==" "\n" |
| 739 | R"==(dst[DST_OFF(local_id, i + 8, 0, 0, 0, 0)] = dst1[i]; )==" "\n" |
| 740 | R"==(#else )==" "\n" |
| 741 | R"==(dst[DST_OFF(i + 0, local_id, 0, 0, 0, 0)] = dst0[i]; )==" "\n" |
| 742 | R"==(dst[DST_OFF(i + 8, local_id, 0, 0, 0, 0)] = dst1[i]; )==" "\n" |
| 743 | R"==(#endif )==" "\n" |
| 744 | R"==(} )==" "\n" |
| 745 | R"==(#elif DST_16A16B || DST_16B16A )==" "\n" |
| 746 | R"==(DST_BLOCK_WRITE8(&dst[0], dst0); )==" "\n" |
| 747 | R"==(DST_BLOCK_WRITE8(&dst[8 * 16], dst1); )==" "\n" |
| 748 | R"==(#else )==" "\n" |
| 749 | R"==(for (int i = 0; i < 8; i++) { )==" "\n" |
| 750 | R"==(#if SRC_16B16A )==" "\n" |
| 751 | R"==(dst[DST_OFF(local_id, i + 0, 0, 0, 0, 0)] = dst0[i]; )==" "\n" |
| 752 | R"==(dst[DST_OFF(local_id, i + 8, 0, 0, 0, 0)] = dst1[i]; )==" "\n" |
| 753 | R"==(#else )==" "\n" |
| 754 | R"==(dst[DST_OFF(i + 0, local_id, 0, 0, 0, 0)] = dst0[i]; )==" "\n" |
| 755 | R"==(dst[DST_OFF(i + 8, local_id, 0, 0, 0, 0)] = dst1[i]; )==" "\n" |
| 756 | R"==(#endif )==" "\n" |
| 757 | R"==(} )==" "\n" |
| 758 | R"==(#endif )==" "\n" |
| 759 | R"==(#elif SRC_16B || DST_16B )==" "\n" |
| 760 | R"==(SRC_DATA_T src_tmp; )==" "\n" |
| 761 | R"==(#if SRC_16B )==" "\n" |
| 762 | R"==(src += SRC_OFF(d0, d1, d2, d3, d4, d5); )==" "\n" |
| 763 | R"==(src_tmp = SRC_BLOCK_READ(&src[0]); )==" "\n" |
| 764 | R"==(#else )==" "\n" |
| 765 | R"==(src += SRC_OFF(d0, d1 + local_id, d2, d3, d4, d5); )==" "\n" |
| 766 | R"==(src_tmp = src[0]; )==" "\n" |
| 767 | R"==(#endif )==" "\n" |
| 768 | R"==(DST_DATA_T dst_tmp; )==" "\n" |
| 769 | R"==(#if DST_16B )==" "\n" |
| 770 | R"==(dst += DST_OFF(d0, d1, d2, d3, d4, d5); )==" "\n" |
| 771 | R"==(#if WITH_SUM_SCALE || WITH_SUM_ZPOINT )==" "\n" |
| 772 | R"==(dst_tmp = DST_BLOCK_READ(&dst[0]); )==" "\n" |
| 773 | R"==(#endif )==" "\n" |
| 774 | R"==(#else )==" "\n" |
| 775 | R"==(dst += DST_OFF(d0, d1 + local_id, d2, d3, d4, d5); )==" "\n" |
| 776 | R"==(#if WITH_SUM_SCALE || WITH_SUM_ZPOINT )==" "\n" |
| 777 | R"==(dst_tmp = dst[0]; )==" "\n" |
| 778 | R"==(#endif )==" "\n" |
| 779 | R"==(#endif )==" "\n" |
| 780 | R"==(REORDER(dst_tmp, src_tmp, src_scale, dst_scale, sum_scale, src_zp, dst_zp, )==" "\n" |
| 781 | R"==(sum_zp); )==" "\n" |
| 782 | R"==(#if DST_16B )==" "\n" |
| 783 | R"==(DST_BLOCK_WRITE(&dst[0], dst_tmp); )==" "\n" |
| 784 | R"==(#else )==" "\n" |
| 785 | R"==(dst[0] = dst_tmp; )==" "\n" |
| 786 | R"==(#endif )==" "\n" |
| 787 | R"==(#elif SRC_16B16C || DST_16B16C || SRC_16C16B || DST_16C16B )==" "\n" |
| 788 | R"==(const int g = d0; )==" "\n" |
| 789 | R"==(SRC_DATA8_T in0, in1; )==" "\n" |
| 790 | R"==(#if SRC_16B16C || SRC_16C16B )==" "\n" |
| 791 | R"==(src += SRC_OFF_G(g, d1, d2, d3, d4, d5); )==" "\n" |
| 792 | R"==(in0 = SRC_BLOCK_READ8(&src[0]); )==" "\n" |
| 793 | R"==(in1 = SRC_BLOCK_READ8(&src[8 * 16]); )==" "\n" |
| 794 | R"==(#else )==" "\n" |
| 795 | R"==(for (int i = 0; i < 8; i++) { )==" "\n" |
| 796 | R"==(#if DST_16C16B )==" "\n" |
| 797 | R"==(in0[i] = src[SRC_OFF_G(g, d1 + local_id, d2 + i + 0, d3, d4, d5)]; )==" "\n" |
| 798 | R"==(in1[i] = src[SRC_OFF_G(g, d1 + local_id, d2 + i + 8, d3, d4, d5)]; )==" "\n" |
| 799 | R"==(#else )==" "\n" |
| 800 | R"==(in0[i] = src[SRC_OFF_G(g, d1 + i + 0, d2 + local_id, d3, d4, d5)]; )==" "\n" |
| 801 | R"==(in1[i] = src[SRC_OFF_G(g, d1 + i + 8, d2 + local_id, d3, d4, d5)]; )==" "\n" |
| 802 | R"==(#endif )==" "\n" |
| 803 | R"==(} )==" "\n" |
| 804 | R"==(#endif )==" "\n" |
| 805 | R"==(DST_DATA8_T dst0, dst1; )==" "\n" |
| 806 | R"==(#if (SRC_16B16C || SRC_16C16B) && (DST_16B16C || DST_16C16B) )==" "\n" |
| 807 | R"==(#if WITH_SUM_SCALE || WITH_SUM_ZPOINT )==" "\n" |
| 808 | R"==(for (int i = 0; i < 8; i++) { )==" "\n" |
| 809 | R"==(#if SRC_16C16B )==" "\n" |
| 810 | R"==(dst0[i] = dst[DST_OFF_G(g, d1 + local_id, d2 + i + 0, d3, d4, d5)]; )==" "\n" |
| 811 | R"==(dst1[i] = dst[DST_OFF_G(g, d1 + local_id, d2 + i + 8, d3, d4, d5)]; )==" "\n" |
| 812 | R"==(#else )==" "\n" |
| 813 | R"==(dst0[i] = dst[DST_OFF_G(g, d1 + i + 0, d2 + local_id, d3, d4, d5)]; )==" "\n" |
| 814 | R"==(dst1[i] = dst[DST_OFF_G(g, d1 + i + 8, d2 + local_id, d3, d4, d5)]; )==" "\n" |
| 815 | R"==(#endif )==" "\n" |
| 816 | R"==(} )==" "\n" |
| 817 | R"==(#endif )==" "\n" |
| 818 | R"==(#elif DST_16B16C || DST_16C16B )==" "\n" |
| 819 | R"==(dst += DST_OFF_G(g, d1, d2, d3, d4, d5); )==" "\n" |
| 820 | R"==(#if WITH_SUM_SCALE || WITH_SUM_ZPOINT )==" "\n" |
| 821 | R"==(dst0 = DST_BLOCK_READ8(&dst[0]); )==" "\n" |
| 822 | R"==(dst1 = DST_BLOCK_READ8(&dst[8 * 16]); )==" "\n" |
| 823 | R"==(#endif )==" "\n" |
| 824 | R"==(#else )==" "\n" |
| 825 | R"==(#if WITH_SUM_SCALE || WITH_SUM_ZPOINT )==" "\n" |
| 826 | R"==(for (int i = 0; i < 8; i++) { )==" "\n" |
| 827 | R"==(#if SRC_16C16B )==" "\n" |
| 828 | R"==(dst0[i] = dst[DST_OFF_G(g, d1 + local_id, d2 + i + 0, d3, d4, d5)]; )==" "\n" |
| 829 | R"==(dst1[i] = dst[DST_OFF_G(g, d1 + local_id, d2 + i + 8, d3, d4, d5)]; )==" "\n" |
| 830 | R"==(#else )==" "\n" |
| 831 | R"==(dst0[i] = dst[DST_OFF_G(g, d1 + i + 0, d2 + local_id, d3, d4, d5)]; )==" "\n" |
| 832 | R"==(dst1[i] = dst[DST_OFF_G(g, d1 + i + 8, d2 + local_id, d3, d4, d5)]; )==" "\n" |
| 833 | R"==(#endif )==" "\n" |
| 834 | R"==(} )==" "\n" |
| 835 | R"==(#endif )==" "\n" |
| 836 | R"==(#endif )==" "\n" |
| 837 | R"==(REORDER8( )==" "\n" |
| 838 | R"==(dst0, in0, src_scale, dst_scale, sum_scale, src_zp, dst_zp, sum_zp); )==" "\n" |
| 839 | R"==(REORDER8( )==" "\n" |
| 840 | R"==(dst1, in1, src_scale, dst_scale, sum_scale, src_zp, dst_zp, sum_zp); )==" "\n" |
| 841 | R"==(#if (SRC_16B16C || SRC_16C16B) && (DST_16B16C || DST_16C16B) )==" "\n" |
| 842 | R"==(for (int i = 0; i < 8; i++) { )==" "\n" |
| 843 | R"==(#if SRC_16C16B )==" "\n" |
| 844 | R"==(dst[DST_OFF_G(g, d1 + local_id, d2 + i + 0, d3, d4, d5)] = dst0[i]; )==" "\n" |
| 845 | R"==(dst[DST_OFF_G(g, d1 + local_id, d2 + i + 8, d3, d4, d5)] = dst1[i]; )==" "\n" |
| 846 | R"==(#else )==" "\n" |
| 847 | R"==(dst[DST_OFF_G(g, d1 + i + 0, d2 + local_id, d3, d4, d5)] = dst0[i]; )==" "\n" |
| 848 | R"==(dst[DST_OFF_G(g, d1 + i + 8, d2 + local_id, d3, d4, d5)] = dst1[i]; )==" "\n" |
| 849 | R"==(#endif )==" "\n" |
| 850 | R"==(} )==" "\n" |
| 851 | R"==(#elif DST_16B16C || DST_16C16B )==" "\n" |
| 852 | R"==(DST_BLOCK_WRITE8(&dst[0], dst0); )==" "\n" |
| 853 | R"==(DST_BLOCK_WRITE8(&dst[8 * 16], dst1); )==" "\n" |
| 854 | R"==(#else )==" "\n" |
| 855 | R"==(for (int i = 0; i < 8; i++) { )==" "\n" |
| 856 | R"==(#if SRC_16C16B )==" "\n" |
| 857 | R"==(dst[DST_OFF_G(g, d1 + local_id, d2 + i + 0, d3, d4, d5)] = dst0[i]; )==" "\n" |
| 858 | R"==(dst[DST_OFF_G(g, d1 + local_id, d2 + i + 8, d3, d4, d5)] = dst1[i]; )==" "\n" |
| 859 | R"==(#else )==" "\n" |
| 860 | R"==(dst[DST_OFF_G(g, d1 + i + 0, d2 + local_id, d3, d4, d5)] = dst0[i]; )==" "\n" |
| 861 | R"==(dst[DST_OFF_G(g, d1 + i + 8, d2 + local_id, d3, d4, d5)] = dst1[i]; )==" "\n" |
| 862 | R"==(#endif )==" "\n" |
| 863 | R"==(} )==" "\n" |
| 864 | R"==(#endif )==" "\n" |
| 865 | R"==(#endif )==" "\n" |
| 866 | R"==(#endif )==" "\n" |
| 867 | R"==(} )==" "\n" |
| 868 | R"==()=="; |
| 869 | } |
| 870 | } |
| 871 | } |
| 872 | } |
Generated on 2022-Nov-30 from project oneDNN revision 84bc949a1
Powered by 
Code Browser 2.1
Generator usage only permitted with license.