| 1 | namespace dnnl { |
|---|---|
| 2 | namespace impl { |
| 3 | namespace gpu { |
| 4 | namespace ocl { |
| 5 | const char *xe_lp_gemm_nocopy_x8x8s32_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/ocl_math_utils.h" )==" "\n" |
| 21 | R"==(#include "gpu/ocl/gemm/ocl_gemm_attrs.h" )==" "\n" |
| 22 | R"==(#include "gpu/ocl/ocl_post_ops.h" )==" "\n" |
| 23 | R"==(#include "gpu/ocl/ocl_types.h" )==" "\n" |
| 24 | R"==(#if defined(S8S8) )==" "\n" |
| 25 | R"==(#define A_TYPE char )==" "\n" |
| 26 | R"==(#define A_TYPE2 char2 )==" "\n" |
| 27 | R"==(#define A_TYPE4 char4 )==" "\n" |
| 28 | R"==(#define B_TYPE char )==" "\n" |
| 29 | R"==(#define B_TYPE4 char4 )==" "\n" |
| 30 | R"==(#define AS_A_TYPE as_char )==" "\n" |
| 31 | R"==(#define AS_A_TYPE2 as_char2 )==" "\n" |
| 32 | R"==(#define AS_A_TYPE4 as_char4 )==" "\n" |
| 33 | R"==(#define AS_B_TYPE as_char )==" "\n" |
| 34 | R"==(#define AS_B_TYPE2 as_char2 )==" "\n" |
| 35 | R"==(#define AS_B_TYPE4 as_char4 )==" "\n" |
| 36 | R"==(#endif )==" "\n" |
| 37 | R"==(#if defined(U8S8) )==" "\n" |
| 38 | R"==(#define A_TYPE uchar )==" "\n" |
| 39 | R"==(#define A_TYPE2 uchar2 )==" "\n" |
| 40 | R"==(#define A_TYPE4 uchar4 )==" "\n" |
| 41 | R"==(#define B_TYPE char )==" "\n" |
| 42 | R"==(#define B_TYPE4 char4 )==" "\n" |
| 43 | R"==(#define AS_A_TYPE as_uchar )==" "\n" |
| 44 | R"==(#define AS_A_TYPE2 as_uchar2 )==" "\n" |
| 45 | R"==(#define AS_A_TYPE4 as_uchar4 )==" "\n" |
| 46 | R"==(#define AS_B_TYPE as_char )==" "\n" |
| 47 | R"==(#define AS_B_TYPE2 as_char2 )==" "\n" |
| 48 | R"==(#define AS_B_TYPE4 as_char4 )==" "\n" |
| 49 | R"==(#endif )==" "\n" |
| 50 | R"==(#if defined(S8U8) )==" "\n" |
| 51 | R"==(#define A_TYPE char )==" "\n" |
| 52 | R"==(#define A_TYPE2 char2 )==" "\n" |
| 53 | R"==(#define A_TYPE4 char4 )==" "\n" |
| 54 | R"==(#define B_TYPE uchar )==" "\n" |
| 55 | R"==(#define B_TYPE4 uchar4 )==" "\n" |
| 56 | R"==(#define AS_A_TYPE as_char )==" "\n" |
| 57 | R"==(#define AS_A_TYPE2 as_char2 )==" "\n" |
| 58 | R"==(#define AS_A_TYPE4 as_char4 )==" "\n" |
| 59 | R"==(#define AS_B_TYPE as_uchar )==" "\n" |
| 60 | R"==(#define AS_B_TYPE2 as_uchar2 )==" "\n" |
| 61 | R"==(#define AS_B_TYPE4 as_uchar4 )==" "\n" |
| 62 | R"==(#endif )==" "\n" |
| 63 | R"==(#if defined(U8U8) )==" "\n" |
| 64 | R"==(#define A_TYPE uchar )==" "\n" |
| 65 | R"==(#define A_TYPE2 uchar2 )==" "\n" |
| 66 | R"==(#define A_TYPE4 uchar4 )==" "\n" |
| 67 | R"==(#define B_TYPE uchar )==" "\n" |
| 68 | R"==(#define B_TYPE4 uchar4 )==" "\n" |
| 69 | R"==(#define AS_A_TYPE as_uchar )==" "\n" |
| 70 | R"==(#define AS_A_TYPE2 as_uchar2 )==" "\n" |
| 71 | R"==(#define AS_A_TYPE4 as_uchar4 )==" "\n" |
| 72 | R"==(#define AS_B_TYPE as_uchar )==" "\n" |
| 73 | R"==(#define AS_B_TYPE2 as_uchar2 )==" "\n" |
| 74 | R"==(#define AS_B_TYPE4 as_uchar4 )==" "\n" |
| 75 | R"==(#endif )==" "\n" |
| 76 | R"==(#if defined(TN) )==" "\n" |
| 77 | R"==(#define DO_FMA DO_FMA_TN )==" "\n" |
| 78 | R"==(#endif )==" "\n" |
| 79 | R"==(#if defined(NN) )==" "\n" |
| 80 | R"==(#define DO_FMA DO_FMA_NN )==" "\n" |
| 81 | R"==(#endif )==" "\n" |
| 82 | R"==(#if defined(NT) )==" "\n" |
| 83 | R"==(#define DO_FMA DO_FMA_NT )==" "\n" |
| 84 | R"==(#endif )==" "\n" |
| 85 | R"==(#if defined(TT) )==" "\n" |
| 86 | R"==(#define DO_FMA DO_FMA_TT )==" "\n" |
| 87 | R"==(#endif )==" "\n" |
| 88 | R"==(#define ADD_ROW_A(z) \ )==" "\n" |
| 89 | R"==(do { \ )==" "\n" |
| 90 | R"==(sumRowA[z] = ai[z].s0 + ai[z].s1 + ai[z].s2 + ai[z].s3; \ )==" "\n" |
| 91 | R"==(} while (0) )==" "\n" |
| 92 | R"==(#define ADD_ROW_AT() \ )==" "\n" |
| 93 | R"==(do { \ )==" "\n" |
| 94 | R"==(sumRowA[0] = ai[0].s0 + ai[1].s0 + ai[2].s0 + ai[3].s0; \ )==" "\n" |
| 95 | R"==(sumRowA[1] = ai[0].s1 + ai[1].s1 + ai[2].s1 + ai[3].s1; \ )==" "\n" |
| 96 | R"==(} while (0) )==" "\n" |
| 97 | R"==(#define ADD_COL_B() \ )==" "\n" |
| 98 | R"==(do { \ )==" "\n" |
| 99 | R"==(sumColB = bi.s0 + bi.s1 + bi.s2 + bi.s3; \ )==" "\n" |
| 100 | R"==(} while (0) )==" "\n" |
| 101 | R"==(#define ADD_COL_BT() \ )==" "\n" |
| 102 | R"==(do { \ )==" "\n" |
| 103 | R"==(sumColB = bi[0] + bi[1] + bi[2] + bi[3]; \ )==" "\n" |
| 104 | R"==(} while (0) )==" "\n" |
| 105 | R"==(#ifdef ALIGNED )==" "\n" |
| 106 | R"==(#define VLOAD4_A(z, p) \ )==" "\n" |
| 107 | R"==(do { \ )==" "\n" |
| 108 | R"==(ai[z] = *((global A_TYPE4 *)p); \ )==" "\n" |
| 109 | R"==(} while (0) )==" "\n" |
| 110 | R"==(#else )==" "\n" |
| 111 | R"==(#define VLOAD4_A(z, p) \ )==" "\n" |
| 112 | R"==(do { \ )==" "\n" |
| 113 | R"==(ai[z].s0 = *(p + 0); \ )==" "\n" |
| 114 | R"==(ai[z].s1 = *(p + 1); \ )==" "\n" |
| 115 | R"==(ai[z].s2 = *(p + 2); \ )==" "\n" |
| 116 | R"==(ai[z].s3 = *(p + 3); \ )==" "\n" |
| 117 | R"==(} while (0) )==" "\n" |
| 118 | R"==(#endif )==" "\n" |
| 119 | R"==(#ifdef ALIGNED )==" "\n" |
| 120 | R"==(#define BLOCK_READ_A(h, hh) \ )==" "\n" |
| 121 | R"==(do { \ )==" "\n" |
| 122 | R"==(ai[hh] = AS_A_TYPE2(intel_sub_group_block_read_uc2( \ )==" "\n" |
| 123 | R"==((global uchar *)(a_ptrs[hh] + h * lda))); \ )==" "\n" |
| 124 | R"==(} while (0) )==" "\n" |
| 125 | R"==(#else )==" "\n" |
| 126 | R"==(#define BLOCK_READ_A(h, hh) \ )==" "\n" |
| 127 | R"==(do { \ )==" "\n" |
| 128 | R"==(ai[hh].s0 = *((a_ptrs[hh] + h * lda) + 0); \ )==" "\n" |
| 129 | R"==(ai[hh].s1 = *((a_ptrs[hh] + h * lda) + 16); \ )==" "\n" |
| 130 | R"==(} while (0) )==" "\n" |
| 131 | R"==(#endif )==" "\n" |
| 132 | R"==(#ifdef ALIGNED )==" "\n" |
| 133 | R"==(#define BLOCK_READ_B(h, hh) \ )==" "\n" |
| 134 | R"==(do { \ )==" "\n" |
| 135 | R"==(bi[hh] = AS_B_TYPE(intel_sub_group_block_read_uc( \ )==" "\n" |
| 136 | R"==((global uchar *)(b_ptrs[hh] + h * ldb))); \ )==" "\n" |
| 137 | R"==(} while (0) )==" "\n" |
| 138 | R"==(#else )==" "\n" |
| 139 | R"==(#define BLOCK_READ_B(h, hh) \ )==" "\n" |
| 140 | R"==(do { \ )==" "\n" |
| 141 | R"==(bi[hh] = *(b_ptrs[hh] + h * ldb); \ )==" "\n" |
| 142 | R"==(} while (0) )==" "\n" |
| 143 | R"==(#endif )==" "\n" |
| 144 | R"==(#ifdef ALIGNED )==" "\n" |
| 145 | R"==(#define VLOAD4_B(p) \ )==" "\n" |
| 146 | R"==(do { \ )==" "\n" |
| 147 | R"==(bi = *((global B_TYPE4 *)p); \ )==" "\n" |
| 148 | R"==(} while (0) )==" "\n" |
| 149 | R"==(#else )==" "\n" |
| 150 | R"==(#define VLOAD4_B(p) \ )==" "\n" |
| 151 | R"==(do { \ )==" "\n" |
| 152 | R"==(bi.s0 = *(p + 0); \ )==" "\n" |
| 153 | R"==(bi.s1 = *(p + 1); \ )==" "\n" |
| 154 | R"==(bi.s2 = *(p + 2); \ )==" "\n" |
| 155 | R"==(bi.s3 = *(p + 3); \ )==" "\n" |
| 156 | R"==(} while (0) )==" "\n" |
| 157 | R"==(#endif )==" "\n" |
| 158 | R"==(#define LOADA_REM(z, p) \ )==" "\n" |
| 159 | R"==(do { \ )==" "\n" |
| 160 | R"==(if (krem == 3) { \ )==" "\n" |
| 161 | R"==(ai[z].s0 = *(p + 0); \ )==" "\n" |
| 162 | R"==(ai[z].s1 = *(p + 1); \ )==" "\n" |
| 163 | R"==(ai[z].s2 = *(p + 2); \ )==" "\n" |
| 164 | R"==(} \ )==" "\n" |
| 165 | R"==(if (krem == 2) { \ )==" "\n" |
| 166 | R"==(ai[z].s0 = *(p + 0); \ )==" "\n" |
| 167 | R"==(ai[z].s1 = *(p + 1); \ )==" "\n" |
| 168 | R"==(} \ )==" "\n" |
| 169 | R"==(if (krem == 1) { ai[z].s0 = *(p + 0); } \ )==" "\n" |
| 170 | R"==(} while (0) )==" "\n" |
| 171 | R"==(#define LOADB_REM(p) \ )==" "\n" |
| 172 | R"==(do { \ )==" "\n" |
| 173 | R"==(if (krem == 3) { \ )==" "\n" |
| 174 | R"==(bi.s0 = *(p + 0); \ )==" "\n" |
| 175 | R"==(bi.s1 = *(p + 1); \ )==" "\n" |
| 176 | R"==(bi.s2 = *(p + 2); \ )==" "\n" |
| 177 | R"==(} \ )==" "\n" |
| 178 | R"==(if (krem == 2) { \ )==" "\n" |
| 179 | R"==(bi.s0 = *(p + 0); \ )==" "\n" |
| 180 | R"==(bi.s1 = *(p + 1); \ )==" "\n" |
| 181 | R"==(} \ )==" "\n" |
| 182 | R"==(if (krem == 1) { bi.s0 = *(p + 0); } \ )==" "\n" |
| 183 | R"==(} while (0) )==" "\n" |
| 184 | R"==(#define COPYA() \ )==" "\n" |
| 185 | R"==(do { \ )==" "\n" |
| 186 | R"==(ait[0].s0 = ai[0].s0; \ )==" "\n" |
| 187 | R"==(ait[0].s1 = ai[1].s0; \ )==" "\n" |
| 188 | R"==(ait[0].s2 = ai[2].s0; \ )==" "\n" |
| 189 | R"==(ait[0].s3 = ai[3].s0; \ )==" "\n" |
| 190 | R"==(ait[1].s0 = ai[0].s1; \ )==" "\n" |
| 191 | R"==(ait[1].s1 = ai[1].s1; \ )==" "\n" |
| 192 | R"==(ait[1].s2 = ai[2].s1; \ )==" "\n" |
| 193 | R"==(ait[1].s3 = ai[3].s1; \ )==" "\n" |
| 194 | R"==(} while (0) )==" "\n" |
| 195 | R"==(#define COPYB() \ )==" "\n" |
| 196 | R"==(do { \ )==" "\n" |
| 197 | R"==(biit.s0 = bi[0]; \ )==" "\n" |
| 198 | R"==(biit.s1 = bi[1]; \ )==" "\n" |
| 199 | R"==(biit.s2 = bi[2]; \ )==" "\n" |
| 200 | R"==(biit.s3 = bi[3]; \ )==" "\n" |
| 201 | R"==(} while (0) )==" "\n" |
| 202 | R"==(#define DO_FMA_TN(h, i) \ )==" "\n" |
| 203 | R"==(do { \ )==" "\n" |
| 204 | R"==(ci[0][i] = idot4(AS_B_TYPE4(sub_group_broadcast(as_int(bi), i)), \ )==" "\n" |
| 205 | R"==(AS_A_TYPE4(ai[0]), ci[0][i]); \ )==" "\n" |
| 206 | R"==(ci[1][i] = idot4(AS_B_TYPE4(sub_group_broadcast(as_int(bi), i)), \ )==" "\n" |
| 207 | R"==(AS_A_TYPE4(ai[1]), ci[1][i]); \ )==" "\n" |
| 208 | R"==(} while (0) )==" "\n" |
| 209 | R"==(#define DO_FMA_NN(h, i) \ )==" "\n" |
| 210 | R"==(do { \ )==" "\n" |
| 211 | R"==(ci[0][i] = idot4(AS_B_TYPE4(sub_group_broadcast(as_int(bi), i)), \ )==" "\n" |
| 212 | R"==(AS_A_TYPE4(ait[0]), ci[0][i]); \ )==" "\n" |
| 213 | R"==(ci[1][i] = idot4(AS_B_TYPE4(sub_group_broadcast(as_int(bi), i)), \ )==" "\n" |
| 214 | R"==(AS_A_TYPE4(ait[1]), ci[1][i]); \ )==" "\n" |
| 215 | R"==(} while (0) )==" "\n" |
| 216 | R"==(#define DO_FMA_NT(h, i) \ )==" "\n" |
| 217 | R"==(do { \ )==" "\n" |
| 218 | R"==(ci[0][i] = idot4(AS_B_TYPE4(sub_group_broadcast(as_int(biit), i)), \ )==" "\n" |
| 219 | R"==(AS_A_TYPE4(ait[0]), ci[0][i]); \ )==" "\n" |
| 220 | R"==(ci[1][i] = idot4(AS_B_TYPE4(sub_group_broadcast(as_int(biit), i)), \ )==" "\n" |
| 221 | R"==(AS_A_TYPE4(ait[1]), ci[1][i]); \ )==" "\n" |
| 222 | R"==(} while (0) )==" "\n" |
| 223 | R"==(#define DO_FMA_TT(h, i) \ )==" "\n" |
| 224 | R"==(do { \ )==" "\n" |
| 225 | R"==(ci[0][i] = idot4(AS_B_TYPE4(sub_group_broadcast(as_int(biit), i)), \ )==" "\n" |
| 226 | R"==(AS_A_TYPE4(ai[0]), ci[0][i]); \ )==" "\n" |
| 227 | R"==(ci[1][i] = idot4(AS_B_TYPE4(sub_group_broadcast(as_int(biit), i)), \ )==" "\n" |
| 228 | R"==(AS_A_TYPE4(ai[1]), ci[1][i]); \ )==" "\n" |
| 229 | R"==(} while (0) )==" "\n" |
| 230 | R"==(#if WITH_ELTWISE == 1 )==" "\n" |
| 231 | R"==(#define POST_OP(val) \ )==" "\n" |
| 232 | R"==(do { \ )==" "\n" |
| 233 | R"==(if (apply_eltwise) \ )==" "\n" |
| 234 | R"==(val = fwd_eltwise( \ )==" "\n" |
| 235 | R"==(val, eltwise_alpha, eltwise_beta, eltwise_scale); \ )==" "\n" |
| 236 | R"==(} while (0) )==" "\n" |
| 237 | R"==(#else )==" "\n" |
| 238 | R"==(#define POST_OP(val) )==" "\n" |
| 239 | R"==(#endif )==" "\n" |
| 240 | R"==(#define FMA_I_LOOP(h) \ )==" "\n" |
| 241 | R"==(do { \ )==" "\n" |
| 242 | R"==(DO_FMA(h, 0); \ )==" "\n" |
| 243 | R"==(DO_FMA(h, 1); \ )==" "\n" |
| 244 | R"==(DO_FMA(h, 2); \ )==" "\n" |
| 245 | R"==(DO_FMA(h, 3); \ )==" "\n" |
| 246 | R"==(DO_FMA(h, 4); \ )==" "\n" |
| 247 | R"==(DO_FMA(h, 5); \ )==" "\n" |
| 248 | R"==(DO_FMA(h, 6); \ )==" "\n" |
| 249 | R"==(DO_FMA(h, 7); \ )==" "\n" |
| 250 | R"==(DO_FMA(h, 8); \ )==" "\n" |
| 251 | R"==(DO_FMA(h, 9); \ )==" "\n" |
| 252 | R"==(DO_FMA(h, 10); \ )==" "\n" |
| 253 | R"==(DO_FMA(h, 11); \ )==" "\n" |
| 254 | R"==(DO_FMA(h, 12); \ )==" "\n" |
| 255 | R"==(DO_FMA(h, 13); \ )==" "\n" |
| 256 | R"==(DO_FMA(h, 14); \ )==" "\n" |
| 257 | R"==(DO_FMA(h, 15); \ )==" "\n" |
| 258 | R"==(} while (0) )==" "\n" |
| 259 | R"==(#define ADD_BOFF(i) \ )==" "\n" |
| 260 | R"==(do { \ )==" "\n" |
| 261 | R"==(ci[0][i] -= ATTR_B0 * sumRowA[0]; \ )==" "\n" |
| 262 | R"==(ci[1][i] -= ATTR_B0 * sumRowA[1]; \ )==" "\n" |
| 263 | R"==(} while (0) )==" "\n" |
| 264 | R"==(#define ADD_BOFF_LOOP() \ )==" "\n" |
| 265 | R"==(do { \ )==" "\n" |
| 266 | R"==(ADD_BOFF(0); \ )==" "\n" |
| 267 | R"==(ADD_BOFF(1); \ )==" "\n" |
| 268 | R"==(ADD_BOFF(2); \ )==" "\n" |
| 269 | R"==(ADD_BOFF(3); \ )==" "\n" |
| 270 | R"==(ADD_BOFF(4); \ )==" "\n" |
| 271 | R"==(ADD_BOFF(5); \ )==" "\n" |
| 272 | R"==(ADD_BOFF(6); \ )==" "\n" |
| 273 | R"==(ADD_BOFF(7); \ )==" "\n" |
| 274 | R"==(ADD_BOFF(8); \ )==" "\n" |
| 275 | R"==(ADD_BOFF(9); \ )==" "\n" |
| 276 | R"==(ADD_BOFF(10); \ )==" "\n" |
| 277 | R"==(ADD_BOFF(11); \ )==" "\n" |
| 278 | R"==(ADD_BOFF(12); \ )==" "\n" |
| 279 | R"==(ADD_BOFF(13); \ )==" "\n" |
| 280 | R"==(ADD_BOFF(14); \ )==" "\n" |
| 281 | R"==(ADD_BOFF(15); \ )==" "\n" |
| 282 | R"==(} while (0) )==" "\n" |
| 283 | R"==(#define ADD_AOFF(h, i) \ )==" "\n" |
| 284 | R"==(do { \ )==" "\n" |
| 285 | R"==(ci[0][i] -= (ATTR_A0 * sub_group_broadcast(as_int(sumColB), i)) \ )==" "\n" |
| 286 | R"==(- (h * ATTR_A0 * ATTR_B0); \ )==" "\n" |
| 287 | R"==(ci[1][i] -= (ATTR_A0 * sub_group_broadcast(as_int(sumColB), i)) \ )==" "\n" |
| 288 | R"==(- (h * ATTR_A0 * ATTR_B0); \ )==" "\n" |
| 289 | R"==(} while (0) )==" "\n" |
| 290 | R"==(#define ADD_AOFF_LOOP(h) \ )==" "\n" |
| 291 | R"==(do { \ )==" "\n" |
| 292 | R"==(ADD_AOFF(h, 0); \ )==" "\n" |
| 293 | R"==(ADD_AOFF(h, 1); \ )==" "\n" |
| 294 | R"==(ADD_AOFF(h, 2); \ )==" "\n" |
| 295 | R"==(ADD_AOFF(h, 3); \ )==" "\n" |
| 296 | R"==(ADD_AOFF(h, 4); \ )==" "\n" |
| 297 | R"==(ADD_AOFF(h, 5); \ )==" "\n" |
| 298 | R"==(ADD_AOFF(h, 6); \ )==" "\n" |
| 299 | R"==(ADD_AOFF(h, 7); \ )==" "\n" |
| 300 | R"==(ADD_AOFF(h, 8); \ )==" "\n" |
| 301 | R"==(ADD_AOFF(h, 9); \ )==" "\n" |
| 302 | R"==(ADD_AOFF(h, 10); \ )==" "\n" |
| 303 | R"==(ADD_AOFF(h, 11); \ )==" "\n" |
| 304 | R"==(ADD_AOFF(h, 12); \ )==" "\n" |
| 305 | R"==(ADD_AOFF(h, 13); \ )==" "\n" |
| 306 | R"==(ADD_AOFF(h, 14); \ )==" "\n" |
| 307 | R"==(ADD_AOFF(h, 15); \ )==" "\n" |
| 308 | R"==(} while (0) )==" "\n" |
| 309 | R"==(#define UPDATE_C_COL(i, betaZero) \ )==" "\n" |
| 310 | R"==(do { \ )==" "\n" |
| 311 | R"==(if (jrem > i) { \ )==" "\n" |
| 312 | R"==(if (irem > 0) { \ )==" "\n" |
| 313 | R"==(if (c_offset_type == 0) { \ )==" "\n" |
| 314 | R"==(float val = ((betaZero) ? 0 : *c) + ci[0][i]; \ )==" "\n" |
| 315 | R"==(POST_OP(val); \ )==" "\n" |
| 316 | R"==(*c = convert_int_sat_rte(val + ((!apply_co) ? 0 : co[0])); \ )==" "\n" |
| 317 | R"==(} \ )==" "\n" |
| 318 | R"==(if (c_offset_type == 1) { \ )==" "\n" |
| 319 | R"==(float val = ((betaZero) ? 0 : *c) + ci[0][i]; \ )==" "\n" |
| 320 | R"==(POST_OP(val); \ )==" "\n" |
| 321 | R"==(*c = convert_int_sat_rte(val + ((!apply_co) ? 0 : co[0])); \ )==" "\n" |
| 322 | R"==(} \ )==" "\n" |
| 323 | R"==(if (c_offset_type == 2) { \ )==" "\n" |
| 324 | R"==(float val = ((betaZero) ? 0 : *c) + ci[0][i]; \ )==" "\n" |
| 325 | R"==(POST_OP(val); \ )==" "\n" |
| 326 | R"==(*c = convert_int_sat_rte(val + ((!apply_co) ? 0 : co[i])); \ )==" "\n" |
| 327 | R"==(} \ )==" "\n" |
| 328 | R"==(} \ )==" "\n" |
| 329 | R"==(if (irem > 16) { \ )==" "\n" |
| 330 | R"==(if (c_offset_type == 0) { \ )==" "\n" |
| 331 | R"==(float val = ((betaZero) ? 0 : *c2) + ci[1][i]; \ )==" "\n" |
| 332 | R"==(POST_OP(val); \ )==" "\n" |
| 333 | R"==(*c2 = convert_int_sat_rte( \ )==" "\n" |
| 334 | R"==(val + ((!apply_co) ? 0 : co[0])); \ )==" "\n" |
| 335 | R"==(} \ )==" "\n" |
| 336 | R"==(if (c_offset_type == 1) { \ )==" "\n" |
| 337 | R"==(float val = ((betaZero) ? 0 : *c2) + ci[1][i]; \ )==" "\n" |
| 338 | R"==(POST_OP(val); \ )==" "\n" |
| 339 | R"==(*c2 = convert_int_sat_rte( \ )==" "\n" |
| 340 | R"==(val + ((!apply_co) ? 0 : co[16])); \ )==" "\n" |
| 341 | R"==(} \ )==" "\n" |
| 342 | R"==(if (c_offset_type == 2) { \ )==" "\n" |
| 343 | R"==(float val = ((betaZero) ? 0 : *c2) + ci[1][i]; \ )==" "\n" |
| 344 | R"==(POST_OP(val); \ )==" "\n" |
| 345 | R"==(*c2 = convert_int_sat_rte( \ )==" "\n" |
| 346 | R"==(val + ((!apply_co) ? 0 : co[i])); \ )==" "\n" |
| 347 | R"==(} \ )==" "\n" |
| 348 | R"==(} \ )==" "\n" |
| 349 | R"==(} \ )==" "\n" |
| 350 | R"==(c = c + ldc; \ )==" "\n" |
| 351 | R"==(c2 = c2 + ldc; \ )==" "\n" |
| 352 | R"==(} while (0) )==" "\n" |
| 353 | R"==(#define UPDATE_C(betaZero) \ )==" "\n" |
| 354 | R"==(do { \ )==" "\n" |
| 355 | R"==(UPDATE_C_COL(0, betaZero); \ )==" "\n" |
| 356 | R"==(UPDATE_C_COL(1, betaZero); \ )==" "\n" |
| 357 | R"==(UPDATE_C_COL(2, betaZero); \ )==" "\n" |
| 358 | R"==(UPDATE_C_COL(3, betaZero); \ )==" "\n" |
| 359 | R"==(UPDATE_C_COL(4, betaZero); \ )==" "\n" |
| 360 | R"==(UPDATE_C_COL(5, betaZero); \ )==" "\n" |
| 361 | R"==(UPDATE_C_COL(6, betaZero); \ )==" "\n" |
| 362 | R"==(UPDATE_C_COL(7, betaZero); \ )==" "\n" |
| 363 | R"==(UPDATE_C_COL(8, betaZero); \ )==" "\n" |
| 364 | R"==(UPDATE_C_COL(9, betaZero); \ )==" "\n" |
| 365 | R"==(UPDATE_C_COL(10, betaZero); \ )==" "\n" |
| 366 | R"==(UPDATE_C_COL(11, betaZero); \ )==" "\n" |
| 367 | R"==(UPDATE_C_COL(12, betaZero); \ )==" "\n" |
| 368 | R"==(UPDATE_C_COL(13, betaZero); \ )==" "\n" |
| 369 | R"==(UPDATE_C_COL(14, betaZero); \ )==" "\n" |
| 370 | R"==(UPDATE_C_COL(15, betaZero); \ )==" "\n" |
| 371 | R"==(} while (0) )==" "\n" |
| 372 | R"==(#ifdef TN )==" "\n" |
| 373 | R"==(__attribute__((intel_reqd_sub_group_size(16))) kernel void )==" "\n" |
| 374 | R"==(xe_lp_gemm_compute_x8x8s32(global A_TYPE *a, global B_TYPE *b, global int *c, )==" "\n" |
| 375 | R"==(int offsetA, int offsetB, int offsetC, int lda, int ldb, int ldc, int m, )==" "\n" |
| 376 | R"==(int n, int k, int beta, global int *ao, global int *bo, global int *co, )==" "\n" |
| 377 | R"==(int offsetCO, int apply_co, local A_TYPE *sa, local B_TYPE *sb, )==" "\n" |
| 378 | R"==(int apply_eltwise, float eltwise_alpha, float eltwise_beta, )==" "\n" |
| 379 | R"==(float eltwise_scale) { )==" "\n" |
| 380 | R"==(A_TYPE4 ai[2]; )==" "\n" |
| 381 | R"==(B_TYPE4 bi; )==" "\n" |
| 382 | R"==(int ci[2][16]; )==" "\n" |
| 383 | R"==(int sumRowA[2] = {0, 0}; )==" "\n" |
| 384 | R"==(int sumColB = 0; )==" "\n" |
| 385 | R"==(int idM = get_group_id(0); )==" "\n" |
| 386 | R"==(int idN = get_group_id(1); )==" "\n" |
| 387 | R"==(int idlM = get_local_id(0); )==" "\n" |
| 388 | R"==(int idlN = get_local_id(1); )==" "\n" |
| 389 | R"==(int lid = get_sub_group_local_id(); )==" "\n" |
| 390 | R"==(int lsm = get_enqueued_local_size(0); )==" "\n" |
| 391 | R"==(int lsn = 8; )==" "\n" |
| 392 | R"==(int i0 = (idM * lsm / 16) * 32 + (get_local_id(0) / 16) * 32; )==" "\n" |
| 393 | R"==(int j0 = idlN * 16 + (idN * lsn * 16); )==" "\n" |
| 394 | R"==(int irem = m - i0 - lid; )==" "\n" |
| 395 | R"==(int jrem = n - j0; )==" "\n" |
| 396 | R"==(int irem2 = m - i0; )==" "\n" |
| 397 | R"==(int jrem2 = n - j0; )==" "\n" |
| 398 | R"==(if (irem < 0) irem = 0; )==" "\n" |
| 399 | R"==(if (jrem < 0) jrem = 0; )==" "\n" |
| 400 | R"==(a += offsetA + (i0 * lda) + (lid * lda); )==" "\n" |
| 401 | R"==(b += offsetB + (j0 * ldb) + (lid * ldb); )==" "\n" |
| 402 | R"==(c += offsetC + (i0) + (j0 * ldc) + lid; )==" "\n" |
| 403 | R"==(int c_offset_type = 0; )==" "\n" |
| 404 | R"==(#ifdef FF )==" "\n" |
| 405 | R"==(co += offsetCO; )==" "\n" |
| 406 | R"==(c_offset_type = 0; )==" "\n" |
| 407 | R"==(#endif )==" "\n" |
| 408 | R"==(#ifdef RR )==" "\n" |
| 409 | R"==(co += offsetCO + i0 + lid; )==" "\n" |
| 410 | R"==(c_offset_type = 1; )==" "\n" |
| 411 | R"==(#endif )==" "\n" |
| 412 | R"==(#ifdef CC )==" "\n" |
| 413 | R"==(co += offsetCO + (j0); )==" "\n" |
| 414 | R"==(c_offset_type = 2; )==" "\n" |
| 415 | R"==(#endif )==" "\n" |
| 416 | R"==(global A_TYPE *a_ptrs[2] = {a, a + 16 * lda}; )==" "\n" |
| 417 | R"==(global B_TYPE *b_ptrs = {b}; )==" "\n" |
| 418 | R"==(for (int y = 0; y < 16; y++) { )==" "\n" |
| 419 | R"==(for (int z = 0; z < 2; z++) { )==" "\n" |
| 420 | R"==(ci[z][y] = 0; )==" "\n" |
| 421 | R"==(} )==" "\n" |
| 422 | R"==(} )==" "\n" |
| 423 | R"==(int k_align = k & ~3; )==" "\n" |
| 424 | R"==(#ifndef ALLOW_READ_OVERRUNS )==" "\n" |
| 425 | R"==(if (irem2 >= 32 && jrem2 >= 16) { )==" "\n" |
| 426 | R"==(#endif )==" "\n" |
| 427 | R"==(for (int h = 0; h < k_align; h += 4) { )==" "\n" |
| 428 | R"==(for (int z = 0; z < 2; z++) { )==" "\n" |
| 429 | R"==(VLOAD4_A(z, (a_ptrs[z] + h)); )==" "\n" |
| 430 | R"==(#ifdef BOFFNONZERO )==" "\n" |
| 431 | R"==(ADD_ROW_A(z); )==" "\n" |
| 432 | R"==(#endif )==" "\n" |
| 433 | R"==(} )==" "\n" |
| 434 | R"==(VLOAD4_B((b_ptrs + h)); )==" "\n" |
| 435 | R"==(#ifdef AOFFNONZERO )==" "\n" |
| 436 | R"==(ADD_COL_B(); )==" "\n" |
| 437 | R"==(#endif )==" "\n" |
| 438 | R"==(FMA_I_LOOP(0); )==" "\n" |
| 439 | R"==(#ifdef BOFFNONZERO )==" "\n" |
| 440 | R"==(ADD_BOFF_LOOP(); )==" "\n" |
| 441 | R"==(#endif )==" "\n" |
| 442 | R"==(#ifdef AOFFNONZERO )==" "\n" |
| 443 | R"==(ADD_AOFF_LOOP(4); )==" "\n" |
| 444 | R"==(#endif )==" "\n" |
| 445 | R"==(} )==" "\n" |
| 446 | R"==(int krem = k & 3; )==" "\n" |
| 447 | R"==(if (krem > 0) { )==" "\n" |
| 448 | R"==(ai[0] = 0; )==" "\n" |
| 449 | R"==(ai[1] = 0; )==" "\n" |
| 450 | R"==(bi = 0; )==" "\n" |
| 451 | R"==(for (int z = 0; z < 2; z++) { )==" "\n" |
| 452 | R"==(LOADA_REM(z, (a_ptrs[z] + k_align)); )==" "\n" |
| 453 | R"==(#ifdef BOFFNONZERO )==" "\n" |
| 454 | R"==(ADD_ROW_A(z); )==" "\n" |
| 455 | R"==(#endif )==" "\n" |
| 456 | R"==(} )==" "\n" |
| 457 | R"==(LOADB_REM((b_ptrs + k_align)); )==" "\n" |
| 458 | R"==(#ifdef AOFFNONZERO )==" "\n" |
| 459 | R"==(ADD_COL_B(); )==" "\n" |
| 460 | R"==(#endif )==" "\n" |
| 461 | R"==(FMA_I_LOOP(0); )==" "\n" |
| 462 | R"==(#ifdef BOFFNONZERO )==" "\n" |
| 463 | R"==(ADD_BOFF_LOOP(); )==" "\n" |
| 464 | R"==(#endif )==" "\n" |
| 465 | R"==(#ifdef AOFFNONZERO )==" "\n" |
| 466 | R"==(ADD_AOFF_LOOP(krem); )==" "\n" |
| 467 | R"==(#endif )==" "\n" |
| 468 | R"==(} )==" "\n" |
| 469 | R"==(#ifndef ALLOW_READ_OVERRUNS )==" "\n" |
| 470 | R"==(} else { )==" "\n" |
| 471 | R"==(for (int h = 0; h < k_align; h += 4) { )==" "\n" |
| 472 | R"==(for (int z = 0; z < 2; z++) { )==" "\n" |
| 473 | R"==(if (irem > z * 16) { )==" "\n" |
| 474 | R"==(VLOAD4_A(z, (a_ptrs[z] + h)); )==" "\n" |
| 475 | R"==(#ifdef BOFFNONZERO )==" "\n" |
| 476 | R"==(ADD_ROW_A(z); )==" "\n" |
| 477 | R"==(#endif )==" "\n" |
| 478 | R"==(} )==" "\n" |
| 479 | R"==(} )==" "\n" |
| 480 | R"==(if (jrem > lid) { )==" "\n" |
| 481 | R"==(VLOAD4_B((b_ptrs + h)); )==" "\n" |
| 482 | R"==(#ifdef AOFFNONZERO )==" "\n" |
| 483 | R"==(ADD_COL_B(); )==" "\n" |
| 484 | R"==(#endif )==" "\n" |
| 485 | R"==(} )==" "\n" |
| 486 | R"==(FMA_I_LOOP(0); )==" "\n" |
| 487 | R"==(#ifdef BOFFNONZERO )==" "\n" |
| 488 | R"==(ADD_BOFF_LOOP(); )==" "\n" |
| 489 | R"==(#endif )==" "\n" |
| 490 | R"==(#ifdef AOFFNONZERO )==" "\n" |
| 491 | R"==(ADD_AOFF_LOOP(4); )==" "\n" |
| 492 | R"==(#endif )==" "\n" |
| 493 | R"==(} )==" "\n" |
| 494 | R"==(int krem = k & 3; )==" "\n" |
| 495 | R"==(if (krem > 0) { )==" "\n" |
| 496 | R"==(ai[0] = 0; )==" "\n" |
| 497 | R"==(ai[1] = 0; )==" "\n" |
| 498 | R"==(bi = 0; )==" "\n" |
| 499 | R"==(for (int z = 0; z < 2; z++) { )==" "\n" |
| 500 | R"==(if (irem > z * 16) { )==" "\n" |
| 501 | R"==(LOADA_REM(z, (a_ptrs[z] + k_align)); )==" "\n" |
| 502 | R"==(#ifdef BOFFNONZERO )==" "\n" |
| 503 | R"==(ADD_ROW_A(z); )==" "\n" |
| 504 | R"==(#endif )==" "\n" |
| 505 | R"==(} )==" "\n" |
| 506 | R"==(} )==" "\n" |
| 507 | R"==(if (jrem > lid) { )==" "\n" |
| 508 | R"==(LOADB_REM((b_ptrs + k_align)); )==" "\n" |
| 509 | R"==(#ifdef AOFFNONZERO )==" "\n" |
| 510 | R"==(ADD_COL_B(); )==" "\n" |
| 511 | R"==(#endif )==" "\n" |
| 512 | R"==(} )==" "\n" |
| 513 | R"==(FMA_I_LOOP(0); )==" "\n" |
| 514 | R"==(#ifdef BOFFNONZERO )==" "\n" |
| 515 | R"==(ADD_BOFF_LOOP(); )==" "\n" |
| 516 | R"==(#endif )==" "\n" |
| 517 | R"==(#ifdef AOFFNONZERO )==" "\n" |
| 518 | R"==(ADD_AOFF_LOOP(krem); )==" "\n" |
| 519 | R"==(#endif )==" "\n" |
| 520 | R"==(} )==" "\n" |
| 521 | R"==(} )==" "\n" |
| 522 | R"==(#endif /* ALLOW_READ_OVERHEAD */ )==" "\n" |
| 523 | R"==(global int *c2 = c + 16; )==" "\n" |
| 524 | R"==(if (beta == 0) )==" "\n" |
| 525 | R"==(UPDATE_C(1); )==" "\n" |
| 526 | R"==(else )==" "\n" |
| 527 | R"==(UPDATE_C(0); )==" "\n" |
| 528 | R"==(} )==" "\n" |
| 529 | R"==(#endif )==" "\n" |
| 530 | R"==(#ifdef NN )==" "\n" |
| 531 | R"==(__attribute__((intel_reqd_sub_group_size(16))) kernel void )==" "\n" |
| 532 | R"==(xe_lp_gemm_compute_x8x8s32(global A_TYPE *a, global B_TYPE *b, global int *c, )==" "\n" |
| 533 | R"==(int offsetA, int offsetB, int offsetC, int lda, int ldb, int ldc, int m, )==" "\n" |
| 534 | R"==(int n, int k, int beta, global int *ao, global int *bo, global int *co, )==" "\n" |
| 535 | R"==(int offsetCO, int apply_co, local A_TYPE *sa, local B_TYPE *sb, )==" "\n" |
| 536 | R"==(int apply_eltwise, float eltwise_alpha, float eltwise_beta, )==" "\n" |
| 537 | R"==(float eltwise_scale) { )==" "\n" |
| 538 | R"==(A_TYPE2 ai[4]; )==" "\n" |
| 539 | R"==(B_TYPE4 bi; )==" "\n" |
| 540 | R"==(int ci[2][16]; )==" "\n" |
| 541 | R"==(int sumRowA[2] = {0, 0}; )==" "\n" |
| 542 | R"==(int sumColB = 0; )==" "\n" |
| 543 | R"==(A_TYPE4 ait[2]; )==" "\n" |
| 544 | R"==(int idM = get_group_id(0); )==" "\n" |
| 545 | R"==(int idN = get_group_id(1); )==" "\n" |
| 546 | R"==(int idlM = get_local_id(0); )==" "\n" |
| 547 | R"==(int idlN = get_local_id(1); )==" "\n" |
| 548 | R"==(int lid = get_sub_group_local_id(); )==" "\n" |
| 549 | R"==(int lsm = get_enqueued_local_size(0); )==" "\n" |
| 550 | R"==(int lsn = 8; )==" "\n" |
| 551 | R"==(int i0 = (idM * lsm / 16) * 32 + (get_local_id(0) / 16) * 32; )==" "\n" |
| 552 | R"==(int j0 = idlN * 16 + (idN * lsn * 16); )==" "\n" |
| 553 | R"==(int irem = m - i0 - lid; )==" "\n" |
| 554 | R"==(int jrem = n - j0; )==" "\n" |
| 555 | R"==(int irem2 = m - i0; )==" "\n" |
| 556 | R"==(int jrem2 = n - j0; )==" "\n" |
| 557 | R"==(if (irem < 0) irem = 0; )==" "\n" |
| 558 | R"==(if (jrem < 0) jrem = 0; )==" "\n" |
| 559 | R"==(#ifdef ALIGNED )==" "\n" |
| 560 | R"==(a += offsetA + i0; )==" "\n" |
| 561 | R"==(#else )==" "\n" |
| 562 | R"==(a += offsetA + i0 + lid; )==" "\n" |
| 563 | R"==(#endif )==" "\n" |
| 564 | R"==(b += offsetB + (j0 * ldb) + (lid * ldb); )==" "\n" |
| 565 | R"==(c += offsetC + (i0) + (j0 * ldc) + lid; )==" "\n" |
| 566 | R"==(int c_offset_type = 0; )==" "\n" |
| 567 | R"==(#ifdef FF )==" "\n" |
| 568 | R"==(co += offsetCO; )==" "\n" |
| 569 | R"==(c_offset_type = 0; )==" "\n" |
| 570 | R"==(#endif )==" "\n" |
| 571 | R"==(#ifdef RR )==" "\n" |
| 572 | R"==(co += offsetCO + i0 + lid; )==" "\n" |
| 573 | R"==(c_offset_type = 1; )==" "\n" |
| 574 | R"==(#endif )==" "\n" |
| 575 | R"==(#ifdef CC )==" "\n" |
| 576 | R"==(co += offsetCO + (j0); )==" "\n" |
| 577 | R"==(c_offset_type = 2; )==" "\n" |
| 578 | R"==(#endif )==" "\n" |
| 579 | R"==(global A_TYPE *a_ptrs[4] = {a, a + 1 * lda, a + 2 * lda, a + 3 * lda}; )==" "\n" |
| 580 | R"==(global B_TYPE *b_ptrs = {b}; )==" "\n" |
| 581 | R"==(for (int y = 0; y < 16; y++) { )==" "\n" |
| 582 | R"==(for (int z = 0; z < 2; z++) { )==" "\n" |
| 583 | R"==(ci[z][y] = 0; )==" "\n" |
| 584 | R"==(} )==" "\n" |
| 585 | R"==(} )==" "\n" |
| 586 | R"==(int k_align = k & ~3; )==" "\n" |
| 587 | R"==(#ifndef ALLOW_READ_OVERRUNS )==" "\n" |
| 588 | R"==(if (irem2 >= 32 && jrem2 >= 16) { )==" "\n" |
| 589 | R"==(#endif )==" "\n" |
| 590 | R"==(for (int h = 0; h < k_align; h += 4) { )==" "\n" |
| 591 | R"==(for (int hh = 0; hh < 4; hh++) { )==" "\n" |
| 592 | R"==(BLOCK_READ_A(h, hh); )==" "\n" |
| 593 | R"==(#ifdef BOFFNONZERO )==" "\n" |
| 594 | R"==(ADD_ROW_AT(); )==" "\n" |
| 595 | R"==(#endif )==" "\n" |
| 596 | R"==(} )==" "\n" |
| 597 | R"==(COPYA(); )==" "\n" |
| 598 | R"==(VLOAD4_B((b_ptrs + h)); )==" "\n" |
| 599 | R"==(#ifdef AOFFNONZERO )==" "\n" |
| 600 | R"==(ADD_COL_B(); )==" "\n" |
| 601 | R"==(#endif )==" "\n" |
| 602 | R"==(FMA_I_LOOP(0); )==" "\n" |
| 603 | R"==(#ifdef BOFFNONZERO )==" "\n" |
| 604 | R"==(ADD_BOFF_LOOP(); )==" "\n" |
| 605 | R"==(#endif )==" "\n" |
| 606 | R"==(#ifdef AOFFNONZERO )==" "\n" |
| 607 | R"==(ADD_AOFF_LOOP(4); )==" "\n" |
| 608 | R"==(#endif )==" "\n" |
| 609 | R"==(} )==" "\n" |
| 610 | R"==(int krem = k & 3; )==" "\n" |
| 611 | R"==(if (krem > 0) { )==" "\n" |
| 612 | R"==(ai[0] = 0; )==" "\n" |
| 613 | R"==(ai[1] = 0; )==" "\n" |
| 614 | R"==(ai[2] = 0; )==" "\n" |
| 615 | R"==(ai[3] = 0; )==" "\n" |
| 616 | R"==(bi = 0; )==" "\n" |
| 617 | R"==(for (int hh = 0; hh < krem; hh++) { )==" "\n" |
| 618 | R"==(BLOCK_READ_A(k_align, hh); )==" "\n" |
| 619 | R"==(#ifdef BOFFNONZERO )==" "\n" |
| 620 | R"==(ADD_ROW_AT(); )==" "\n" |
| 621 | R"==(#endif )==" "\n" |
| 622 | R"==(} )==" "\n" |
| 623 | R"==(COPYA(); )==" "\n" |
| 624 | R"==(LOADB_REM((b_ptrs + k_align)); )==" "\n" |
| 625 | R"==(#ifdef AOFFNONZERO )==" "\n" |
| 626 | R"==(ADD_COL_B(); )==" "\n" |
| 627 | R"==(#endif )==" "\n" |
| 628 | R"==(FMA_I_LOOP(0); )==" "\n" |
| 629 | R"==(#ifdef BOFFNONZERO )==" "\n" |
| 630 | R"==(ADD_BOFF_LOOP(); )==" "\n" |
| 631 | R"==(#endif )==" "\n" |
| 632 | R"==(#ifdef AOFFNONZERO )==" "\n" |
| 633 | R"==(ADD_AOFF_LOOP(krem); )==" "\n" |
| 634 | R"==(#endif )==" "\n" |
| 635 | R"==(} )==" "\n" |
| 636 | R"==(#ifndef ALLOW_READ_OVERRUNS )==" "\n" |
| 637 | R"==(} else { )==" "\n" |
| 638 | R"==(for (int h = 0; h < k_align; h += 4) { )==" "\n" |
| 639 | R"==(for (int hh = 0; hh < 4; hh++) { )==" "\n" |
| 640 | R"==(if (irem2 > lid) { )==" "\n" |
| 641 | R"==(#ifdef ALIGNED )==" "\n" |
| 642 | R"==(ai[hh].s0 = *((a_ptrs[hh] + h * lda) + 0 + lid); )==" "\n" |
| 643 | R"==(ai[hh].s1 = *((a_ptrs[hh] + h * lda) + 16 + lid); )==" "\n" |
| 644 | R"==(#else )==" "\n" |
| 645 | R"==(ai[hh].s0 = *((a_ptrs[hh] + h * lda) + 0); )==" "\n" |
| 646 | R"==(ai[hh].s1 = *((a_ptrs[hh] + h * lda) + 16); )==" "\n" |
| 647 | R"==(#endif )==" "\n" |
| 648 | R"==(#ifdef BOFFNONZERO )==" "\n" |
| 649 | R"==(ADD_ROW_AT(); )==" "\n" |
| 650 | R"==(#endif )==" "\n" |
| 651 | R"==(} )==" "\n" |
| 652 | R"==(} )==" "\n" |
| 653 | R"==(COPYA(); )==" "\n" |
| 654 | R"==(if (jrem > lid) { )==" "\n" |
| 655 | R"==(VLOAD4_B((b_ptrs + h)); )==" "\n" |
| 656 | R"==(#ifdef AOFFNONZERO )==" "\n" |
| 657 | R"==(ADD_COL_B(); )==" "\n" |
| 658 | R"==(#endif )==" "\n" |
| 659 | R"==(} )==" "\n" |
| 660 | R"==(FMA_I_LOOP(0); )==" "\n" |
| 661 | R"==(#ifdef BOFFNONZERO )==" "\n" |
| 662 | R"==(ADD_BOFF_LOOP(); )==" "\n" |
| 663 | R"==(#endif )==" "\n" |
| 664 | R"==(#ifdef AOFFNONZERO )==" "\n" |
| 665 | R"==(ADD_AOFF_LOOP(4); )==" "\n" |
| 666 | R"==(#endif )==" "\n" |
| 667 | R"==(} )==" "\n" |
| 668 | R"==(int krem = k & 3; )==" "\n" |
| 669 | R"==(if (krem > 0) { )==" "\n" |
| 670 | R"==(ai[0] = 0; )==" "\n" |
| 671 | R"==(ai[1] = 0; )==" "\n" |
| 672 | R"==(ai[2] = 0; )==" "\n" |
| 673 | R"==(ai[3] = 0; )==" "\n" |
| 674 | R"==(bi = 0; )==" "\n" |
| 675 | R"==(for (int hh = 0; hh < krem; hh++) { )==" "\n" |
| 676 | R"==(if (irem2 > lid) { )==" "\n" |
| 677 | R"==(#ifdef ALIGNED )==" "\n" |
| 678 | R"==(ai[hh].s0 = *((a_ptrs[hh] + k_align * lda) + 0 + lid); )==" "\n" |
| 679 | R"==(ai[hh].s1 = *((a_ptrs[hh] + k_align * lda) + 16 + lid); )==" "\n" |
| 680 | R"==(#else )==" "\n" |
| 681 | R"==(ai[hh].s0 = *((a_ptrs[hh] + k_align * lda) + 0); )==" "\n" |
| 682 | R"==(ai[hh].s1 = *((a_ptrs[hh] + k_align * lda) + 16); )==" "\n" |
| 683 | R"==(#endif )==" "\n" |
| 684 | R"==(#ifdef BOFFNONZERO )==" "\n" |
| 685 | R"==(ADD_ROW_AT(); )==" "\n" |
| 686 | R"==(#endif )==" "\n" |
| 687 | R"==(} )==" "\n" |
| 688 | R"==(} )==" "\n" |
| 689 | R"==(COPYA(); )==" "\n" |
| 690 | R"==(if (jrem > lid) { )==" "\n" |
| 691 | R"==(LOADB_REM((b_ptrs + k_align)); )==" "\n" |
| 692 | R"==(#ifdef AOFFNONZERO )==" "\n" |
| 693 | R"==(ADD_COL_B(); )==" "\n" |
| 694 | R"==(#endif )==" "\n" |
| 695 | R"==(} )==" "\n" |
| 696 | R"==(FMA_I_LOOP(0); )==" "\n" |
| 697 | R"==(#ifdef BOFFNONZERO )==" "\n" |
| 698 | R"==(ADD_BOFF_LOOP(); )==" "\n" |
| 699 | R"==(#endif )==" "\n" |
| 700 | R"==(#ifdef AOFFNONZERO )==" "\n" |
| 701 | R"==(ADD_AOFF_LOOP(krem); )==" "\n" |
| 702 | R"==(#endif )==" "\n" |
| 703 | R"==(} )==" "\n" |
| 704 | R"==(} )==" "\n" |
| 705 | R"==(#endif /* ALLOW_READ_OVERHEAD */ )==" "\n" |
| 706 | R"==(global int *c2 = c + 16; )==" "\n" |
| 707 | R"==(if (beta == 0) )==" "\n" |
| 708 | R"==(UPDATE_C(1); )==" "\n" |
| 709 | R"==(else )==" "\n" |
| 710 | R"==(UPDATE_C(0); )==" "\n" |
| 711 | R"==(} )==" "\n" |
| 712 | R"==(#endif )==" "\n" |
| 713 | R"==(#ifdef NT )==" "\n" |
| 714 | R"==(__attribute__((intel_reqd_sub_group_size(16))) kernel void )==" "\n" |
| 715 | R"==(xe_lp_gemm_compute_x8x8s32(global A_TYPE *a, global B_TYPE *b, global int *c, )==" "\n" |
| 716 | R"==(int offsetA, int offsetB, int offsetC, int lda, int ldb, int ldc, int m, )==" "\n" |
| 717 | R"==(int n, int k, int beta, global int *ao, global int *bo, global int *co, )==" "\n" |
| 718 | R"==(int offsetCO, int apply_co, local A_TYPE *sa, local B_TYPE *sb, )==" "\n" |
| 719 | R"==(int apply_eltwise, float eltwise_alpha, float eltwise_beta, )==" "\n" |
| 720 | R"==(float eltwise_scale) { )==" "\n" |
| 721 | R"==(A_TYPE2 ai[4]; )==" "\n" |
| 722 | R"==(B_TYPE bi[4]; )==" "\n" |
| 723 | R"==(int ci[2][16]; )==" "\n" |
| 724 | R"==(int sumRowA[2] = {0, 0}; )==" "\n" |
| 725 | R"==(int sumColB = 0; )==" "\n" |
| 726 | R"==(A_TYPE4 ait[2]; )==" "\n" |
| 727 | R"==(A_TYPE4 biit; )==" "\n" |
| 728 | R"==(int idM = get_group_id(0); )==" "\n" |
| 729 | R"==(int idN = get_group_id(1); )==" "\n" |
| 730 | R"==(int idlM = get_local_id(0); )==" "\n" |
| 731 | R"==(int idlN = get_local_id(1); )==" "\n" |
| 732 | R"==(int lid = get_sub_group_local_id(); )==" "\n" |
| 733 | R"==(int lsm = get_enqueued_local_size(0); )==" "\n" |
| 734 | R"==(int lsn = 8; )==" "\n" |
| 735 | R"==(int i0 = (idM * lsm / 16) * 32 + (get_local_id(0) / 16) * 32; )==" "\n" |
| 736 | R"==(int j0 = idlN * 16 + (idN * lsn * 16); )==" "\n" |
| 737 | R"==(int irem = m - i0 - lid; )==" "\n" |
| 738 | R"==(int jrem = n - j0; )==" "\n" |
| 739 | R"==(int irem2 = m - i0; )==" "\n" |
| 740 | R"==(int jrem2 = n - j0; )==" "\n" |
| 741 | R"==(if (irem < 0) irem = 0; )==" "\n" |
| 742 | R"==(if (jrem < 0) jrem = 0; )==" "\n" |
| 743 | R"==(#ifdef ALIGNED )==" "\n" |
| 744 | R"==(a += offsetA + i0; )==" "\n" |
| 745 | R"==(#else )==" "\n" |
| 746 | R"==(a += offsetA + i0 + lid; )==" "\n" |
| 747 | R"==(#endif )==" "\n" |
| 748 | R"==(#ifdef ALIGNED )==" "\n" |
| 749 | R"==(b += offsetB + j0; )==" "\n" |
| 750 | R"==(#else )==" "\n" |
| 751 | R"==(b += offsetB + j0 + lid; )==" "\n" |
| 752 | R"==(#endif )==" "\n" |
| 753 | R"==(c += offsetC + (i0) + (j0 * ldc) + lid; )==" "\n" |
| 754 | R"==(int c_offset_type = 0; )==" "\n" |
| 755 | R"==(#ifdef FF )==" "\n" |
| 756 | R"==(co += offsetCO; )==" "\n" |
| 757 | R"==(c_offset_type = 0; )==" "\n" |
| 758 | R"==(#endif )==" "\n" |
| 759 | R"==(#ifdef RR )==" "\n" |
| 760 | R"==(co += offsetCO + i0 + lid; )==" "\n" |
| 761 | R"==(c_offset_type = 1; )==" "\n" |
| 762 | R"==(#endif )==" "\n" |
| 763 | R"==(#ifdef CC )==" "\n" |
| 764 | R"==(co += offsetCO + (j0); )==" "\n" |
| 765 | R"==(c_offset_type = 2; )==" "\n" |
| 766 | R"==(#endif )==" "\n" |
| 767 | R"==(global A_TYPE *a_ptrs[4] = {a, a + 1 * lda, a + 2 * lda, a + 3 * lda}; )==" "\n" |
| 768 | R"==(global B_TYPE *b_ptrs[4] = {b, b + 1 * ldb, b + 2 * ldb, b + 3 * ldb}; )==" "\n" |
| 769 | R"==(for (int y = 0; y < 16; y++) { )==" "\n" |
| 770 | R"==(for (int z = 0; z < 2; z++) { )==" "\n" |
| 771 | R"==(ci[z][y] = 0; )==" "\n" |
| 772 | R"==(} )==" "\n" |
| 773 | R"==(} )==" "\n" |
| 774 | R"==(int insidea1 = 5; )==" "\n" |
| 775 | R"==(int insidea2 = 5; )==" "\n" |
| 776 | R"==(int insideb = 5; )==" "\n" |
| 777 | R"==(int k_align = k & ~3; )==" "\n" |
| 778 | R"==(#ifndef ALLOW_READ_OVERRUNS )==" "\n" |
| 779 | R"==(if (irem2 >= 32 && jrem2 >= 16) { )==" "\n" |
| 780 | R"==(#endif )==" "\n" |
| 781 | R"==(for (int h = 0; h < k_align; h += 4) { )==" "\n" |
| 782 | R"==(for (int hh = 0; hh < 4; hh++) { )==" "\n" |
| 783 | R"==(BLOCK_READ_A(h, hh); )==" "\n" |
| 784 | R"==(#ifdef BOFFNONZERO )==" "\n" |
| 785 | R"==(ADD_ROW_AT(); )==" "\n" |
| 786 | R"==(#endif )==" "\n" |
| 787 | R"==(} )==" "\n" |
| 788 | R"==(COPYA(); )==" "\n" |
| 789 | R"==(for (int hh = 0; hh < 4; hh++) { )==" "\n" |
| 790 | R"==(BLOCK_READ_B(h, hh); )==" "\n" |
| 791 | R"==(#ifdef AOFFNONZERO )==" "\n" |
| 792 | R"==(ADD_COL_BT(); )==" "\n" |
| 793 | R"==(#endif )==" "\n" |
| 794 | R"==(} )==" "\n" |
| 795 | R"==(COPYB(); )==" "\n" |
| 796 | R"==(FMA_I_LOOP(0); )==" "\n" |
| 797 | R"==(#ifdef BOFFNONZERO )==" "\n" |
| 798 | R"==(ADD_BOFF_LOOP(); )==" "\n" |
| 799 | R"==(#endif )==" "\n" |
| 800 | R"==(#ifdef AOFFNONZERO )==" "\n" |
| 801 | R"==(ADD_AOFF_LOOP(4); )==" "\n" |
| 802 | R"==(#endif )==" "\n" |
| 803 | R"==(} )==" "\n" |
| 804 | R"==(int krem = k & 3; )==" "\n" |
| 805 | R"==(if (krem > 0) { )==" "\n" |
| 806 | R"==(ai[0] = 0; )==" "\n" |
| 807 | R"==(ai[1] = 0; )==" "\n" |
| 808 | R"==(ai[2] = 0; )==" "\n" |
| 809 | R"==(ai[3] = 0; )==" "\n" |
| 810 | R"==(bi[0] = 0; )==" "\n" |
| 811 | R"==(bi[1] = 0; )==" "\n" |
| 812 | R"==(bi[2] = 0; )==" "\n" |
| 813 | R"==(bi[3] = 0; )==" "\n" |
| 814 | R"==(for (int hh = 0; hh < krem; hh++) { )==" "\n" |
| 815 | R"==(BLOCK_READ_A(k_align, hh); )==" "\n" |
| 816 | R"==(#ifdef BOFFNONZERO )==" "\n" |
| 817 | R"==(ADD_ROW_AT(); )==" "\n" |
| 818 | R"==(#endif )==" "\n" |
| 819 | R"==(} )==" "\n" |
| 820 | R"==(COPYA(); )==" "\n" |
| 821 | R"==(for (int hh = 0; hh < krem; hh++) { )==" "\n" |
| 822 | R"==(BLOCK_READ_B(k_align, hh); )==" "\n" |
| 823 | R"==(#ifdef AOFFNONZERO )==" "\n" |
| 824 | R"==(ADD_COL_BT(); )==" "\n" |
| 825 | R"==(#endif )==" "\n" |
| 826 | R"==(} )==" "\n" |
| 827 | R"==(COPYB(); )==" "\n" |
| 828 | R"==(FMA_I_LOOP(0); )==" "\n" |
| 829 | R"==(#ifdef BOFFNONZERO )==" "\n" |
| 830 | R"==(ADD_BOFF_LOOP(); )==" "\n" |
| 831 | R"==(#endif )==" "\n" |
| 832 | R"==(#ifdef AOFFNONZERO )==" "\n" |
| 833 | R"==(ADD_AOFF_LOOP(krem); )==" "\n" |
| 834 | R"==(#endif )==" "\n" |
| 835 | R"==(} )==" "\n" |
| 836 | R"==(#ifndef ALLOW_READ_OVERRUNS )==" "\n" |
| 837 | R"==(} else { )==" "\n" |
| 838 | R"==(for (int h = 0; h < k_align; h += 4) { )==" "\n" |
| 839 | R"==(for (int hh = 0; hh < 4; hh++) { )==" "\n" |
| 840 | R"==(if (irem2 > lid) { )==" "\n" |
| 841 | R"==(#ifdef ALIGNED )==" "\n" |
| 842 | R"==(ai[hh].s0 = *((a_ptrs[hh] + h * lda) + 0 + lid); )==" "\n" |
| 843 | R"==(ai[hh].s1 = *((a_ptrs[hh] + h * lda) + 16 + lid); )==" "\n" |
| 844 | R"==(#else )==" "\n" |
| 845 | R"==(ai[hh].s0 = *((a_ptrs[hh] + h * lda) + 0); )==" "\n" |
| 846 | R"==(ai[hh].s1 = *((a_ptrs[hh] + h * lda) + 16); )==" "\n" |
| 847 | R"==(#endif )==" "\n" |
| 848 | R"==(#ifdef BOFFNONZERO )==" "\n" |
| 849 | R"==(ADD_ROW_AT(); )==" "\n" |
| 850 | R"==(#endif )==" "\n" |
| 851 | R"==(} )==" "\n" |
| 852 | R"==(} )==" "\n" |
| 853 | R"==(COPYA(); )==" "\n" |
| 854 | R"==(for (int hh = 0; hh < 4; hh++) { )==" "\n" |
| 855 | R"==(if (jrem > lid) { )==" "\n" |
| 856 | R"==(#ifdef ALIGNED )==" "\n" |
| 857 | R"==(bi[hh] = *((b_ptrs[hh] + h * ldb) + lid); )==" "\n" |
| 858 | R"==(#else )==" "\n" |
| 859 | R"==(bi[hh] = *(b_ptrs[hh] + h * ldb); )==" "\n" |
| 860 | R"==(#endif )==" "\n" |
| 861 | R"==(#ifdef AOFFNONZERO )==" "\n" |
| 862 | R"==(ADD_COL_BT(); )==" "\n" |
| 863 | R"==(#endif )==" "\n" |
| 864 | R"==(} )==" "\n" |
| 865 | R"==(} )==" "\n" |
| 866 | R"==(COPYB(); )==" "\n" |
| 867 | R"==(FMA_I_LOOP(0); )==" "\n" |
| 868 | R"==(#ifdef BOFFNONZERO )==" "\n" |
| 869 | R"==(ADD_BOFF_LOOP(); )==" "\n" |
| 870 | R"==(#endif )==" "\n" |
| 871 | R"==(#ifdef AOFFNONZERO )==" "\n" |
| 872 | R"==(ADD_AOFF_LOOP(4); )==" "\n" |
| 873 | R"==(#endif )==" "\n" |
| 874 | R"==(} )==" "\n" |
| 875 | R"==(int krem = k & 3; )==" "\n" |
| 876 | R"==(if (krem > 0) { )==" "\n" |
| 877 | R"==(ai[0] = 0; )==" "\n" |
| 878 | R"==(ai[1] = 0; )==" "\n" |
| 879 | R"==(ai[2] = 0; )==" "\n" |
| 880 | R"==(ai[3] = 0; )==" "\n" |
| 881 | R"==(bi[0] = 0; )==" "\n" |
| 882 | R"==(bi[1] = 0; )==" "\n" |
| 883 | R"==(bi[2] = 0; )==" "\n" |
| 884 | R"==(bi[3] = 0; )==" "\n" |
| 885 | R"==(for (int hh = 0; hh < krem; hh++) { )==" "\n" |
| 886 | R"==(if (irem2 > lid) { )==" "\n" |
| 887 | R"==(#ifdef ALIGNED )==" "\n" |
| 888 | R"==(ai[hh].s0 = *((a_ptrs[hh] + k_align * lda) + 0 + lid); )==" "\n" |
| 889 | R"==(ai[hh].s1 = *((a_ptrs[hh] + k_align * lda) + 16 + lid); )==" "\n" |
| 890 | R"==(#else )==" "\n" |
| 891 | R"==(ai[hh].s0 = *((a_ptrs[hh] + k_align * lda) + 0); )==" "\n" |
| 892 | R"==(ai[hh].s1 = *((a_ptrs[hh] + k_align * lda) + 16); )==" "\n" |
| 893 | R"==(#endif )==" "\n" |
| 894 | R"==(#ifdef BOFFNONZERO )==" "\n" |
| 895 | R"==(ADD_ROW_AT(); )==" "\n" |
| 896 | R"==(#endif )==" "\n" |
| 897 | R"==(} )==" "\n" |
| 898 | R"==(} )==" "\n" |
| 899 | R"==(COPYA(); )==" "\n" |
| 900 | R"==(for (int hh = 0; hh < krem; hh++) { )==" "\n" |
| 901 | R"==(if (jrem > lid) { )==" "\n" |
| 902 | R"==(#ifdef ALIGNED )==" "\n" |
| 903 | R"==(bi[hh] = *((b_ptrs[hh] + k_align * ldb) + lid); )==" "\n" |
| 904 | R"==(#else )==" "\n" |
| 905 | R"==(bi[hh] = *(b_ptrs[hh] + k_align * ldb); )==" "\n" |
| 906 | R"==(#endif )==" "\n" |
| 907 | R"==(#ifdef AOFFNONZERO )==" "\n" |
| 908 | R"==(ADD_COL_BT(); )==" "\n" |
| 909 | R"==(#endif )==" "\n" |
| 910 | R"==(} )==" "\n" |
| 911 | R"==(} )==" "\n" |
| 912 | R"==(COPYB(); )==" "\n" |
| 913 | R"==(FMA_I_LOOP(0); )==" "\n" |
| 914 | R"==(#ifdef BOFFNONZERO )==" "\n" |
| 915 | R"==(ADD_BOFF_LOOP(); )==" "\n" |
| 916 | R"==(#endif )==" "\n" |
| 917 | R"==(#ifdef AOFFNONZERO )==" "\n" |
| 918 | R"==(ADD_AOFF_LOOP(krem); )==" "\n" |
| 919 | R"==(#endif )==" "\n" |
| 920 | R"==(} )==" "\n" |
| 921 | R"==(} )==" "\n" |
| 922 | R"==(#endif /* ALLOW_READ_OVERHEAD */ )==" "\n" |
| 923 | R"==(global int *c2 = c + 16; )==" "\n" |
| 924 | R"==(if (beta == 0) )==" "\n" |
| 925 | R"==(UPDATE_C(1); )==" "\n" |
| 926 | R"==(else )==" "\n" |
| 927 | R"==(UPDATE_C(0); )==" "\n" |
| 928 | R"==(} )==" "\n" |
| 929 | R"==(#endif )==" "\n" |
| 930 | R"==(#ifdef TT )==" "\n" |
| 931 | R"==(__attribute__((intel_reqd_sub_group_size(16))) kernel void )==" "\n" |
| 932 | R"==(xe_lp_gemm_compute_x8x8s32(global A_TYPE *a, global B_TYPE *b, global int *c, )==" "\n" |
| 933 | R"==(int offsetA, int offsetB, int offsetC, int lda, int ldb, int ldc, int m, )==" "\n" |
| 934 | R"==(int n, int k, int beta, global int *ao, global int *bo, global int *co, )==" "\n" |
| 935 | R"==(int offsetCO, int apply_co, local A_TYPE *sa, local B_TYPE *sb, )==" "\n" |
| 936 | R"==(int apply_eltwise, float eltwise_alpha, float eltwise_beta, )==" "\n" |
| 937 | R"==(float eltwise_scale) { )==" "\n" |
| 938 | R"==(A_TYPE4 ai[2]; )==" "\n" |
| 939 | R"==(B_TYPE bi[4]; )==" "\n" |
| 940 | R"==(int ci[2][16]; )==" "\n" |
| 941 | R"==(int sumRowA[2] = {0, 0}; )==" "\n" |
| 942 | R"==(int sumColB = 0; )==" "\n" |
| 943 | R"==(A_TYPE4 biit; )==" "\n" |
| 944 | R"==(int idM = get_group_id(0); )==" "\n" |
| 945 | R"==(int idN = get_group_id(1); )==" "\n" |
| 946 | R"==(int idlM = get_local_id(0); )==" "\n" |
| 947 | R"==(int idlN = get_local_id(1); )==" "\n" |
| 948 | R"==(int lid = get_sub_group_local_id(); )==" "\n" |
| 949 | R"==(int lsm = get_enqueued_local_size(0); )==" "\n" |
| 950 | R"==(int lsn = 8; )==" "\n" |
| 951 | R"==(int i0 = (idM * lsm / 16) * 32 + (get_local_id(0) / 16) * 32; )==" "\n" |
| 952 | R"==(int j0 = idlN * 16 + (idN * lsn * 16); )==" "\n" |
| 953 | R"==(int irem = m - i0 - lid; )==" "\n" |
| 954 | R"==(int jrem = n - j0; )==" "\n" |
| 955 | R"==(if (irem < 0) irem = 0; )==" "\n" |
| 956 | R"==(if (jrem < 0) jrem = 0; )==" "\n" |
| 957 | R"==(int irem2 = m - i0; )==" "\n" |
| 958 | R"==(int jrem2 = n - j0; )==" "\n" |
| 959 | R"==(a += offsetA + (i0 * lda) + (lid * lda); )==" "\n" |
| 960 | R"==(#ifdef ALIGNED )==" "\n" |
| 961 | R"==(b += offsetB + j0; )==" "\n" |
| 962 | R"==(#else )==" "\n" |
| 963 | R"==(b += offsetB + j0 + lid; )==" "\n" |
| 964 | R"==(#endif )==" "\n" |
| 965 | R"==(c += offsetC + (i0) + (j0 * ldc) + lid; )==" "\n" |
| 966 | R"==(int c_offset_type = 0; )==" "\n" |
| 967 | R"==(#ifdef FF )==" "\n" |
| 968 | R"==(co += offsetCO; )==" "\n" |
| 969 | R"==(c_offset_type = 0; )==" "\n" |
| 970 | R"==(#endif )==" "\n" |
| 971 | R"==(#ifdef RR )==" "\n" |
| 972 | R"==(co += offsetCO + i0 + lid; )==" "\n" |
| 973 | R"==(c_offset_type = 1; )==" "\n" |
| 974 | R"==(#endif )==" "\n" |
| 975 | R"==(#ifdef CC )==" "\n" |
| 976 | R"==(co += offsetCO + (j0); )==" "\n" |
| 977 | R"==(c_offset_type = 2; )==" "\n" |
| 978 | R"==(#endif )==" "\n" |
| 979 | R"==(global A_TYPE *a_ptrs[2] = {a, a + 16 * lda}; )==" "\n" |
| 980 | R"==(global B_TYPE *b_ptrs[4] = {b, b + 1 * ldb, b + 2 * ldb, b + 3 * ldb}; )==" "\n" |
| 981 | R"==(for (int y = 0; y < 16; y++) { )==" "\n" |
| 982 | R"==(for (int z = 0; z < 2; z++) { )==" "\n" |
| 983 | R"==(ci[z][y] = 0; )==" "\n" |
| 984 | R"==(} )==" "\n" |
| 985 | R"==(} )==" "\n" |
| 986 | R"==(int k_align = k & ~3; )==" "\n" |
| 987 | R"==(#ifndef ALLOW_READ_OVERRUNS )==" "\n" |
| 988 | R"==(if (irem2 >= 32 && jrem2 >= 16) { )==" "\n" |
| 989 | R"==(#endif )==" "\n" |
| 990 | R"==(for (int h = 0; h < k_align; h += 4) { )==" "\n" |
| 991 | R"==(for (int z = 0; z < 2; z++) { )==" "\n" |
| 992 | R"==(VLOAD4_A(z, ((a_ptrs[z]) + h)); )==" "\n" |
| 993 | R"==(#ifdef BOFFNONZERO )==" "\n" |
| 994 | R"==(ADD_ROW_A(z); )==" "\n" |
| 995 | R"==(#endif )==" "\n" |
| 996 | R"==(} )==" "\n" |
| 997 | R"==(for (int hh = 0; hh < 4; hh++) { )==" "\n" |
| 998 | R"==(BLOCK_READ_B(h, hh); )==" "\n" |
| 999 | R"==(#ifdef AOFFNONZERO )==" "\n" |
| 1000 | R"==(ADD_COL_BT(); )==" "\n" |
| 1001 | R"==(#endif )==" "\n" |
| 1002 | R"==(} )==" "\n" |
| 1003 | R"==(COPYB(); )==" "\n" |
| 1004 | R"==(FMA_I_LOOP(0); )==" "\n" |
| 1005 | R"==(#ifdef BOFFNONZERO )==" "\n" |
| 1006 | R"==(ADD_BOFF_LOOP(); )==" "\n" |
| 1007 | R"==(#endif )==" "\n" |
| 1008 | R"==(#ifdef AOFFNONZERO )==" "\n" |
| 1009 | R"==(ADD_AOFF_LOOP(4); )==" "\n" |
| 1010 | R"==(#endif )==" "\n" |
| 1011 | R"==(} )==" "\n" |
| 1012 | R"==(int krem = k & 3; )==" "\n" |
| 1013 | R"==(if (krem > 0) { )==" "\n" |
| 1014 | R"==(ai[0] = 0; )==" "\n" |
| 1015 | R"==(ai[1] = 0; )==" "\n" |
| 1016 | R"==(bi[0] = 0; )==" "\n" |
| 1017 | R"==(bi[1] = 0; )==" "\n" |
| 1018 | R"==(bi[2] = 0; )==" "\n" |
| 1019 | R"==(bi[3] = 0; )==" "\n" |
| 1020 | R"==(for (int z = 0; z < 2; z++) { )==" "\n" |
| 1021 | R"==(LOADA_REM(z, ((a_ptrs[z]) + k_align)); )==" "\n" |
| 1022 | R"==(#ifdef BOFFNONZERO )==" "\n" |
| 1023 | R"==(ADD_ROW_A(z); )==" "\n" |
| 1024 | R"==(#endif )==" "\n" |
| 1025 | R"==(} )==" "\n" |
| 1026 | R"==(for (int hh = 0; hh < krem; hh++) { )==" "\n" |
| 1027 | R"==(BLOCK_READ_B(k_align, hh); )==" "\n" |
| 1028 | R"==(#ifdef AOFFNONZERO )==" "\n" |
| 1029 | R"==(ADD_COL_BT(); )==" "\n" |
| 1030 | R"==(#endif )==" "\n" |
| 1031 | R"==(} )==" "\n" |
| 1032 | R"==(COPYB(); )==" "\n" |
| 1033 | R"==(FMA_I_LOOP(0); )==" "\n" |
| 1034 | R"==(#ifdef BOFFNONZERO )==" "\n" |
| 1035 | R"==(ADD_BOFF_LOOP(); )==" "\n" |
| 1036 | R"==(#endif )==" "\n" |
| 1037 | R"==(#ifdef AOFFNONZERO )==" "\n" |
| 1038 | R"==(ADD_AOFF_LOOP(krem); )==" "\n" |
| 1039 | R"==(#endif )==" "\n" |
| 1040 | R"==(} )==" "\n" |
| 1041 | R"==(#ifndef ALLOW_READ_OVERRUNS )==" "\n" |
| 1042 | R"==(} else { )==" "\n" |
| 1043 | R"==(for (int h = 0; h < k_align; h += 4) { )==" "\n" |
| 1044 | R"==(for (int z = 0; z < 2; z++) { )==" "\n" |
| 1045 | R"==(if (irem > z * 16) { )==" "\n" |
| 1046 | R"==(VLOAD4_A(z, (a_ptrs[z] + h)); )==" "\n" |
| 1047 | R"==(#ifdef BOFFNONZERO )==" "\n" |
| 1048 | R"==(ADD_ROW_A(z); )==" "\n" |
| 1049 | R"==(#endif )==" "\n" |
| 1050 | R"==(} )==" "\n" |
| 1051 | R"==(} )==" "\n" |
| 1052 | R"==(for (int hh = 0; hh < 4; hh++) { )==" "\n" |
| 1053 | R"==(if (jrem > lid) { )==" "\n" |
| 1054 | R"==(#ifdef ALIGNED )==" "\n" |
| 1055 | R"==(bi[hh] = *((b_ptrs[hh] + h * ldb) + lid); )==" "\n" |
| 1056 | R"==(#else )==" "\n" |
| 1057 | R"==(bi[hh] = *(b_ptrs[hh] + h * ldb); )==" "\n" |
| 1058 | R"==(#endif )==" "\n" |
| 1059 | R"==(#ifdef AOFFNONZERO )==" "\n" |
| 1060 | R"==(ADD_COL_BT(); )==" "\n" |
| 1061 | R"==(#endif )==" "\n" |
| 1062 | R"==(} )==" "\n" |
| 1063 | R"==(} )==" "\n" |
| 1064 | R"==(COPYB(); )==" "\n" |
| 1065 | R"==(FMA_I_LOOP(0); )==" "\n" |
| 1066 | R"==(#ifdef BOFFNONZERO )==" "\n" |
| 1067 | R"==(ADD_BOFF_LOOP(); )==" "\n" |
| 1068 | R"==(#endif )==" "\n" |
| 1069 | R"==(#ifdef AOFFNONZERO )==" "\n" |
| 1070 | R"==(ADD_AOFF_LOOP(4); )==" "\n" |
| 1071 | R"==(#endif )==" "\n" |
| 1072 | R"==(} )==" "\n" |
| 1073 | R"==(int krem = k & 3; )==" "\n" |
| 1074 | R"==(if (krem > 0) { )==" "\n" |
| 1075 | R"==(ai[0] = 0; )==" "\n" |
| 1076 | R"==(ai[1] = 0; )==" "\n" |
| 1077 | R"==(bi[0] = 0; )==" "\n" |
| 1078 | R"==(bi[1] = 0; )==" "\n" |
| 1079 | R"==(bi[2] = 0; )==" "\n" |
| 1080 | R"==(bi[3] = 0; )==" "\n" |
| 1081 | R"==(for (int z = 0; z < 2; z++) { )==" "\n" |
| 1082 | R"==(if (irem > z * 16) { )==" "\n" |
| 1083 | R"==(LOADA_REM(z, (a_ptrs[z] + k_align)); )==" "\n" |
| 1084 | R"==(#ifdef BOFFNONZERO )==" "\n" |
| 1085 | R"==(ADD_ROW_A(z); )==" "\n" |
| 1086 | R"==(#endif )==" "\n" |
| 1087 | R"==(} )==" "\n" |
| 1088 | R"==(} )==" "\n" |
| 1089 | R"==(for (int hh = 0; hh < krem; hh++) { )==" "\n" |
| 1090 | R"==(if (jrem > lid) { )==" "\n" |
| 1091 | R"==(#ifdef ALIGNED )==" "\n" |
| 1092 | R"==(bi[hh] = *((b_ptrs[hh] + k_align * ldb) + lid); )==" "\n" |
| 1093 | R"==(#else )==" "\n" |
| 1094 | R"==(bi[hh] = *(b_ptrs[hh] + k_align * ldb); )==" "\n" |
| 1095 | R"==(#endif )==" "\n" |
| 1096 | R"==(#ifdef AOFFNONZERO )==" "\n" |
| 1097 | R"==(ADD_COL_BT(); )==" "\n" |
| 1098 | R"==(#endif )==" "\n" |
| 1099 | R"==(} )==" "\n" |
| 1100 | R"==(} )==" "\n" |
| 1101 | R"==(COPYB(); )==" "\n" |
| 1102 | R"==(FMA_I_LOOP(0); )==" "\n" |
| 1103 | R"==(#ifdef BOFFNONZERO )==" "\n" |
| 1104 | R"==(ADD_BOFF_LOOP(); )==" "\n" |
| 1105 | R"==(#endif )==" "\n" |
| 1106 | R"==(#ifdef AOFFNONZERO )==" "\n" |
| 1107 | R"==(ADD_AOFF_LOOP(krem); )==" "\n" |
| 1108 | R"==(#endif )==" "\n" |
| 1109 | R"==(} )==" "\n" |
| 1110 | R"==(} )==" "\n" |
| 1111 | R"==(#endif /* ALLOW_READ_OVERHEAD */ )==" "\n" |
| 1112 | R"==(global int *c2 = c + 16; )==" "\n" |
| 1113 | R"==(if (beta == 0) )==" "\n" |
| 1114 | R"==(UPDATE_C(1); )==" "\n" |
| 1115 | R"==(else )==" "\n" |
| 1116 | R"==(UPDATE_C(0); )==" "\n" |
| 1117 | R"==(} )==" "\n" |
| 1118 | R"==(#endif )==" "\n" |
| 1119 | R"==()=="; |
| 1120 | } |
| 1121 | } |
| 1122 | } |
| 1123 | } |
Generated on 2022-Nov-30 from project oneDNN revision 84bc949a1
Powered by 
Code Browser 2.1
Generator usage only permitted with license.