| 1 | namespace dnnl { |
|---|---|
| 2 | namespace impl { |
| 3 | namespace gpu { |
| 4 | namespace ocl { |
| 5 | const char *gen9_gemm_nocopy_f16_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/gemm/ocl_gemm_attrs.h" )==" "\n" |
| 21 | R"==(#include "gpu/ocl/ocl_post_ops.h" )==" "\n" |
| 22 | R"==(#include "gpu/ocl/ocl_types.h" )==" "\n" |
| 23 | R"==(#pragma OPENCL EXTENSION cl_khr_fp16 : enable )==" "\n" |
| 24 | R"==(#define VLOAD4_ALIGNED(o, p) *((global half4 *)p + o) )==" "\n" |
| 25 | R"==(#if !defined(TRANS_A) )==" "\n" |
| 26 | R"==(#if !defined(TRANS_B) )==" "\n" |
| 27 | R"==(#define NN )==" "\n" |
| 28 | R"==(#define DO_FMA DO_FMA_NN )==" "\n" |
| 29 | R"==(#else )==" "\n" |
| 30 | R"==(#define NT )==" "\n" |
| 31 | R"==(#define DO_FMA DO_FMA_NT )==" "\n" |
| 32 | R"==(#endif )==" "\n" |
| 33 | R"==(#else )==" "\n" |
| 34 | R"==(#if !defined(TRANS_B) )==" "\n" |
| 35 | R"==(#define TN )==" "\n" |
| 36 | R"==(#define DO_FMA DO_FMA_TN )==" "\n" |
| 37 | R"==(#else )==" "\n" |
| 38 | R"==(#define TT )==" "\n" |
| 39 | R"==(#define DO_FMA DO_FMA_TT )==" "\n" |
| 40 | R"==(#endif )==" "\n" |
| 41 | R"==(#endif )==" "\n" |
| 42 | R"==(#define DO_FMA_NN(h, i, i_mod_2, i_div_2, i_mod_16, i_div_16) \ )==" "\n" |
| 43 | R"==(do { \ )==" "\n" |
| 44 | R"==(c[i][0] = mad(sub_group_broadcast(a[h].s##i_mod_2, i_div_2), \ )==" "\n" |
| 45 | R"==(b[0].s##h, c[i][0]); \ )==" "\n" |
| 46 | R"==(c[i][1] = mad(sub_group_broadcast(a[h].s##i_mod_2, i_div_2), \ )==" "\n" |
| 47 | R"==(b[1].s##h, c[i][1]); \ )==" "\n" |
| 48 | R"==(} while (0) )==" "\n" |
| 49 | R"==(#define DO_FMA_NT(h, i, i_mod_2, i_div_2, i_mod_16, i_div_16) \ )==" "\n" |
| 50 | R"==(do { \ )==" "\n" |
| 51 | R"==(c[i][0] = mad(sub_group_broadcast(a[h].s##i_mod_2, i_div_2), b[h].s0, \ )==" "\n" |
| 52 | R"==(c[i][0]); \ )==" "\n" |
| 53 | R"==(c[i][1] = mad(sub_group_broadcast(a[h].s##i_mod_2, i_div_2), b[h].s1, \ )==" "\n" |
| 54 | R"==(c[i][1]); \ )==" "\n" |
| 55 | R"==(} while (0) )==" "\n" |
| 56 | R"==(#define DO_FMA_TN(h, i, i_mod_2, i_div_2, i_mod_16, i_div_16) \ )==" "\n" |
| 57 | R"==(do { \ )==" "\n" |
| 58 | R"==(c[i][0] = mad(sub_group_broadcast(a[i_div_16].s##h, i_mod_16), \ )==" "\n" |
| 59 | R"==(b[0].s##h, c[i][0]); \ )==" "\n" |
| 60 | R"==(c[i][1] = mad(sub_group_broadcast(a[i_div_16].s##h, i_mod_16), \ )==" "\n" |
| 61 | R"==(b[1].s##h, c[i][1]); \ )==" "\n" |
| 62 | R"==(} while (0) )==" "\n" |
| 63 | R"==(#define DO_FMA_TT(h, i, i_mod_2, i_div_2, i_mod_16, i_div_16) \ )==" "\n" |
| 64 | R"==(do { \ )==" "\n" |
| 65 | R"==(c[i][0] = mad(sub_group_broadcast(a[i_div_16].s##h, i_mod_16), \ )==" "\n" |
| 66 | R"==(b[h].s0, c[i][0]); \ )==" "\n" |
| 67 | R"==(c[i][1] = mad(sub_group_broadcast(a[i_div_16].s##h, i_mod_16), \ )==" "\n" |
| 68 | R"==(b[h].s1, c[i][1]); \ )==" "\n" |
| 69 | R"==(} while (0) )==" "\n" |
| 70 | R"==(#if WITH_ELTWISE == 1 )==" "\n" |
| 71 | R"==(#define POST_OP(val) \ )==" "\n" |
| 72 | R"==(do { \ )==" "\n" |
| 73 | R"==(if (last_k_block && last_k_unroll) \ )==" "\n" |
| 74 | R"==(val = fwd_eltwise( \ )==" "\n" |
| 75 | R"==(val, eltwise_alpha, eltwise_beta, eltwise_scale); \ )==" "\n" |
| 76 | R"==(} while (0) )==" "\n" |
| 77 | R"==(#else )==" "\n" |
| 78 | R"==(#define POST_OP(val) )==" "\n" |
| 79 | R"==(#endif )==" "\n" |
| 80 | R"==(#define FMA_I_LOOP(h) \ )==" "\n" |
| 81 | R"==(do { \ )==" "\n" |
| 82 | R"==(DO_FMA(h, 0, 0, 0, 0, 0); \ )==" "\n" |
| 83 | R"==(DO_FMA(h, 1, 1, 0, 1, 0); \ )==" "\n" |
| 84 | R"==(DO_FMA(h, 2, 0, 1, 2, 0); \ )==" "\n" |
| 85 | R"==(DO_FMA(h, 3, 1, 1, 3, 0); \ )==" "\n" |
| 86 | R"==(DO_FMA(h, 4, 0, 2, 4, 0); \ )==" "\n" |
| 87 | R"==(DO_FMA(h, 5, 1, 2, 5, 0); \ )==" "\n" |
| 88 | R"==(DO_FMA(h, 6, 0, 3, 6, 0); \ )==" "\n" |
| 89 | R"==(DO_FMA(h, 7, 1, 3, 7, 0); \ )==" "\n" |
| 90 | R"==(DO_FMA(h, 8, 0, 4, 8, 0); \ )==" "\n" |
| 91 | R"==(DO_FMA(h, 9, 1, 4, 9, 0); \ )==" "\n" |
| 92 | R"==(DO_FMA(h, 10, 0, 5, 10, 0); \ )==" "\n" |
| 93 | R"==(DO_FMA(h, 11, 1, 5, 11, 0); \ )==" "\n" |
| 94 | R"==(DO_FMA(h, 12, 0, 6, 12, 0); \ )==" "\n" |
| 95 | R"==(DO_FMA(h, 13, 1, 6, 13, 0); \ )==" "\n" |
| 96 | R"==(DO_FMA(h, 14, 0, 7, 14, 0); \ )==" "\n" |
| 97 | R"==(DO_FMA(h, 15, 1, 7, 15, 0); \ )==" "\n" |
| 98 | R"==(DO_FMA(h, 16, 0, 8, 0, 1); \ )==" "\n" |
| 99 | R"==(DO_FMA(h, 17, 1, 8, 1, 1); \ )==" "\n" |
| 100 | R"==(DO_FMA(h, 18, 0, 9, 2, 1); \ )==" "\n" |
| 101 | R"==(DO_FMA(h, 19, 1, 9, 3, 1); \ )==" "\n" |
| 102 | R"==(DO_FMA(h, 20, 0, 10, 4, 1); \ )==" "\n" |
| 103 | R"==(DO_FMA(h, 21, 1, 10, 5, 1); \ )==" "\n" |
| 104 | R"==(DO_FMA(h, 22, 0, 11, 6, 1); \ )==" "\n" |
| 105 | R"==(DO_FMA(h, 23, 1, 11, 7, 1); \ )==" "\n" |
| 106 | R"==(DO_FMA(h, 24, 0, 12, 8, 1); \ )==" "\n" |
| 107 | R"==(DO_FMA(h, 25, 1, 12, 9, 1); \ )==" "\n" |
| 108 | R"==(DO_FMA(h, 26, 0, 13, 10, 1); \ )==" "\n" |
| 109 | R"==(DO_FMA(h, 27, 1, 13, 11, 1); \ )==" "\n" |
| 110 | R"==(DO_FMA(h, 28, 0, 14, 12, 1); \ )==" "\n" |
| 111 | R"==(DO_FMA(h, 29, 1, 14, 13, 1); \ )==" "\n" |
| 112 | R"==(DO_FMA(h, 30, 0, 15, 14, 1); \ )==" "\n" |
| 113 | R"==(DO_FMA(h, 31, 1, 15, 15, 1); \ )==" "\n" |
| 114 | R"==(} while (0) )==" "\n" |
| 115 | R"==(#define UPDATE_C_ROW(i, betaZero) \ )==" "\n" |
| 116 | R"==(do { \ )==" "\n" |
| 117 | R"==(if (irem > i) { \ )==" "\n" |
| 118 | R"==(if (jrem > 0) { \ )==" "\n" |
| 119 | R"==(half val \ )==" "\n" |
| 120 | R"==(= ATTR_ALPHA * c[i][0] + ((betaZero) ? 0 : beta * *C); \ )==" "\n" |
| 121 | R"==(POST_OP(val); \ )==" "\n" |
| 122 | R"==(*C = val; \ )==" "\n" |
| 123 | R"==(} \ )==" "\n" |
| 124 | R"==(if (jrem > 16) { \ )==" "\n" |
| 125 | R"==(half val = ATTR_ALPHA * c[i][1] \ )==" "\n" |
| 126 | R"==(+ ((betaZero) ? 0 : beta * *C2); \ )==" "\n" |
| 127 | R"==(POST_OP(val); \ )==" "\n" |
| 128 | R"==(*C2 = val; \ )==" "\n" |
| 129 | R"==(} \ )==" "\n" |
| 130 | R"==(} \ )==" "\n" |
| 131 | R"==(C++; \ )==" "\n" |
| 132 | R"==(C2++; \ )==" "\n" |
| 133 | R"==(} while (0) )==" "\n" |
| 134 | R"==(#define UPDATE_C(betaZero) \ )==" "\n" |
| 135 | R"==(do { \ )==" "\n" |
| 136 | R"==(UPDATE_C_ROW(0, betaZero); \ )==" "\n" |
| 137 | R"==(UPDATE_C_ROW(1, betaZero); \ )==" "\n" |
| 138 | R"==(UPDATE_C_ROW(2, betaZero); \ )==" "\n" |
| 139 | R"==(UPDATE_C_ROW(3, betaZero); \ )==" "\n" |
| 140 | R"==(UPDATE_C_ROW(4, betaZero); \ )==" "\n" |
| 141 | R"==(UPDATE_C_ROW(5, betaZero); \ )==" "\n" |
| 142 | R"==(UPDATE_C_ROW(6, betaZero); \ )==" "\n" |
| 143 | R"==(UPDATE_C_ROW(7, betaZero); \ )==" "\n" |
| 144 | R"==(UPDATE_C_ROW(8, betaZero); \ )==" "\n" |
| 145 | R"==(UPDATE_C_ROW(9, betaZero); \ )==" "\n" |
| 146 | R"==(UPDATE_C_ROW(10, betaZero); \ )==" "\n" |
| 147 | R"==(UPDATE_C_ROW(11, betaZero); \ )==" "\n" |
| 148 | R"==(UPDATE_C_ROW(12, betaZero); \ )==" "\n" |
| 149 | R"==(UPDATE_C_ROW(13, betaZero); \ )==" "\n" |
| 150 | R"==(UPDATE_C_ROW(14, betaZero); \ )==" "\n" |
| 151 | R"==(UPDATE_C_ROW(15, betaZero); \ )==" "\n" |
| 152 | R"==(UPDATE_C_ROW(16, betaZero); \ )==" "\n" |
| 153 | R"==(UPDATE_C_ROW(17, betaZero); \ )==" "\n" |
| 154 | R"==(UPDATE_C_ROW(18, betaZero); \ )==" "\n" |
| 155 | R"==(UPDATE_C_ROW(19, betaZero); \ )==" "\n" |
| 156 | R"==(UPDATE_C_ROW(20, betaZero); \ )==" "\n" |
| 157 | R"==(UPDATE_C_ROW(21, betaZero); \ )==" "\n" |
| 158 | R"==(UPDATE_C_ROW(22, betaZero); \ )==" "\n" |
| 159 | R"==(UPDATE_C_ROW(23, betaZero); \ )==" "\n" |
| 160 | R"==(UPDATE_C_ROW(24, betaZero); \ )==" "\n" |
| 161 | R"==(UPDATE_C_ROW(25, betaZero); \ )==" "\n" |
| 162 | R"==(UPDATE_C_ROW(26, betaZero); \ )==" "\n" |
| 163 | R"==(UPDATE_C_ROW(27, betaZero); \ )==" "\n" |
| 164 | R"==(UPDATE_C_ROW(28, betaZero); \ )==" "\n" |
| 165 | R"==(UPDATE_C_ROW(29, betaZero); \ )==" "\n" |
| 166 | R"==(UPDATE_C_ROW(30, betaZero); \ )==" "\n" |
| 167 | R"==(UPDATE_C_ROW(31, betaZero); \ )==" "\n" |
| 168 | R"==(} while (0) )==" "\n" |
| 169 | R"==(#ifdef NN )==" "\n" |
| 170 | R"==(__attribute__((intel_reqd_sub_group_size(16))) )==" "\n" |
| 171 | R"==(kernel void )==" "\n" |
| 172 | R"==(gen9_gemm_nocopy_f16(global half *A, global half *B, global half *C, )==" "\n" |
| 173 | R"==(long offset_a, long offset_b, long offset_c, int lda, int ldb, int ldc, )==" "\n" |
| 174 | R"==(int m, int n, int k, global float *alpha, float beta, int last_k_block, )==" "\n" |
| 175 | R"==(float eltwise_alpha, float eltwise_beta, float eltwise_scale )==" "\n" |
| 176 | R"==(#ifdef WITH_K_UNROLL )==" "\n" |
| 177 | R"==(, )==" "\n" |
| 178 | R"==(volatile global int *flag, long offset_f) { )==" "\n" |
| 179 | R"==(#else )==" "\n" |
| 180 | R"==() { )==" "\n" |
| 181 | R"==(#endif )==" "\n" |
| 182 | R"==(half2 a[4]; )==" "\n" |
| 183 | R"==(half4 b[2]; )==" "\n" |
| 184 | R"==(half c[32][2]; )==" "\n" |
| 185 | R"==(int idM = get_global_id(1); )==" "\n" |
| 186 | R"==(int idN = get_global_id(0); )==" "\n" |
| 187 | R"==(int lid = get_sub_group_local_id(); )==" "\n" |
| 188 | R"==(int idK = get_global_id(2); )==" "\n" |
| 189 | R"==(int nku = get_global_size(2); )==" "\n" |
| 190 | R"==(int i0 = idM * 32; )==" "\n" |
| 191 | R"==(int j0 = sub_group_broadcast(idN, 0) * 2 + lid; )==" "\n" |
| 192 | R"==(int irem = m - i0; )==" "\n" |
| 193 | R"==(int jrem = n - j0; )==" "\n" |
| 194 | R"==(if (irem < 0) irem = 0; )==" "\n" |
| 195 | R"==(if (jrem < 0) jrem = 0; )==" "\n" |
| 196 | R"==(int irem2 = (irem + 1) >> 1; )==" "\n" |
| 197 | R"==(int last_k_unroll = (idK == nku - 1); )==" "\n" |
| 198 | R"==(#ifdef WITH_K_UNROLL )==" "\n" |
| 199 | R"==(int k0 = idK * UNROLL_K; )==" "\n" |
| 200 | R"==(int kt = k - k0; )==" "\n" |
| 201 | R"==(if (kt < 0) kt = 0; )==" "\n" |
| 202 | R"==(if (kt > UNROLL_K) kt = UNROLL_K; )==" "\n" |
| 203 | R"==(A += offset_a + i0 + k0 * lda; )==" "\n" |
| 204 | R"==(B += offset_b + j0 * ldb + k0; )==" "\n" |
| 205 | R"==(k = kt; )==" "\n" |
| 206 | R"==(#else )==" "\n" |
| 207 | R"==(A += offset_a + i0; )==" "\n" |
| 208 | R"==(B += offset_b + j0 * ldb; )==" "\n" |
| 209 | R"==(#endif )==" "\n" |
| 210 | R"==(C += offset_c + i0 + j0 * ldc; )==" "\n" |
| 211 | R"==(global half *A_ptrs[4] = {A, A + lda, A + 2 * lda, A + 3 * lda}; )==" "\n" |
| 212 | R"==(global half *B_ptrs[2] = {B, B + 16 * ldb}; )==" "\n" |
| 213 | R"==(for (int y = 0; y < 32; y++) )==" "\n" |
| 214 | R"==(for (int z = 0; z < 2; z++) )==" "\n" |
| 215 | R"==(c[y][z] = 0; )==" "\n" |
| 216 | R"==(#ifdef WITH_K_UNROLL )==" "\n" |
| 217 | R"==(flag += offset_f + idM; )==" "\n" |
| 218 | R"==(if (idK == 0 && lid == 0) *flag = 0; )==" "\n" |
| 219 | R"==(#endif )==" "\n" |
| 220 | R"==(int k_align = k & ~3; )==" "\n" |
| 221 | R"==(#ifndef ALLOW_READ_OVERRUNS )==" "\n" |
| 222 | R"==(if (irem >= 32 && sub_group_broadcast(jrem, 0) >= 32) { )==" "\n" |
| 223 | R"==(#endif )==" "\n" |
| 224 | R"==(for (int h = 0; h < k_align; h += 4) { )==" "\n" |
| 225 | R"==(for (int hh = 0; hh < 4; hh++) )==" "\n" |
| 226 | R"==(a[hh] = as_half2(intel_sub_group_block_read( )==" "\n" |
| 227 | R"==((global uint *)(A_ptrs[hh] + h * lda))); )==" "\n" |
| 228 | R"==(for (int j = 0; j < 2; j++) )==" "\n" |
| 229 | R"==(b[j] = VLOAD4_ALIGNED(0, (B_ptrs[j] + h)); )==" "\n" |
| 230 | R"==(FMA_I_LOOP(0); )==" "\n" |
| 231 | R"==(FMA_I_LOOP(1); )==" "\n" |
| 232 | R"==(FMA_I_LOOP(2); )==" "\n" |
| 233 | R"==(FMA_I_LOOP(3); )==" "\n" |
| 234 | R"==(} )==" "\n" |
| 235 | R"==(for (int h = k_align; h < k; h++) { )==" "\n" |
| 236 | R"==(a[0] = as_half2(intel_sub_group_block_read( )==" "\n" |
| 237 | R"==((global uint *)(A_ptrs[0] + h * lda))); )==" "\n" |
| 238 | R"==(for (int j = 0; j < 2; j++) )==" "\n" |
| 239 | R"==(b[j] = B_ptrs[j][h]; )==" "\n" |
| 240 | R"==(FMA_I_LOOP(0); )==" "\n" |
| 241 | R"==(} )==" "\n" |
| 242 | R"==(#ifndef ALLOW_READ_OVERRUNS )==" "\n" |
| 243 | R"==(} else { )==" "\n" |
| 244 | R"==(for (int h = 0; h < k_align; h += 4) { )==" "\n" |
| 245 | R"==(for (int hh = 0; hh < 4; hh++) { )==" "\n" |
| 246 | R"==(if (irem2 > lid) )==" "\n" |
| 247 | R"==(a[hh] = as_half2( )==" "\n" |
| 248 | R"==(*((global uint *)(A_ptrs[hh] + h * lda) + lid)); )==" "\n" |
| 249 | R"==(} )==" "\n" |
| 250 | R"==(for (int j = 0; j < 2; j++) )==" "\n" |
| 251 | R"==(if (jrem > j * 16) b[j] = VLOAD4_ALIGNED(0, (B_ptrs[j] + h)); )==" "\n" |
| 252 | R"==(FMA_I_LOOP(0); )==" "\n" |
| 253 | R"==(FMA_I_LOOP(1); )==" "\n" |
| 254 | R"==(FMA_I_LOOP(2); )==" "\n" |
| 255 | R"==(FMA_I_LOOP(3); )==" "\n" |
| 256 | R"==(} )==" "\n" |
| 257 | R"==(for (int h = k_align; h < k; h++) { )==" "\n" |
| 258 | R"==(if (irem2 > lid) )==" "\n" |
| 259 | R"==(a[0] = as_half2(*((global uint *)(A_ptrs[0] + h * lda) + lid)); )==" "\n" |
| 260 | R"==(for (int j = 0; j < 2; j++) )==" "\n" |
| 261 | R"==(if (jrem > j * 16) b[j] = B_ptrs[j][h]; )==" "\n" |
| 262 | R"==(FMA_I_LOOP(0); )==" "\n" |
| 263 | R"==(} )==" "\n" |
| 264 | R"==(} )==" "\n" |
| 265 | R"==(#endif )==" "\n" |
| 266 | R"==(global half *C2 = C + 16 * ldc; )==" "\n" |
| 267 | R"==(#ifdef WITH_K_UNROLL )==" "\n" |
| 268 | R"==(do { )==" "\n" |
| 269 | R"==(read_mem_fence(CLK_GLOBAL_MEM_FENCE); )==" "\n" |
| 270 | R"==(} while (*flag != idK); )==" "\n" |
| 271 | R"==(if (idK == 0) { )==" "\n" |
| 272 | R"==(if (beta == 0) )==" "\n" |
| 273 | R"==(UPDATE_C(1); )==" "\n" |
| 274 | R"==(else )==" "\n" |
| 275 | R"==(UPDATE_C(0); )==" "\n" |
| 276 | R"==(} else { )==" "\n" |
| 277 | R"==(beta = 1.0; )==" "\n" |
| 278 | R"==(UPDATE_C(0); )==" "\n" |
| 279 | R"==(} )==" "\n" |
| 280 | R"==(if (lid == 0) *flag = idK + 1; )==" "\n" |
| 281 | R"==(#else )==" "\n" |
| 282 | R"==(if (beta == 0) )==" "\n" |
| 283 | R"==(UPDATE_C(1); )==" "\n" |
| 284 | R"==(else )==" "\n" |
| 285 | R"==(UPDATE_C(0); )==" "\n" |
| 286 | R"==(#endif )==" "\n" |
| 287 | R"==(} )==" "\n" |
| 288 | R"==(#endif )==" "\n" |
| 289 | R"==(#ifdef NT )==" "\n" |
| 290 | R"==(__attribute__((intel_reqd_sub_group_size(16))) )==" "\n" |
| 291 | R"==(kernel void )==" "\n" |
| 292 | R"==(gen9_gemm_nocopy_f16(global half *A, global half *B, global half *C, )==" "\n" |
| 293 | R"==(long offset_a, long offset_b, long offset_c, int lda, int ldb, int ldc, )==" "\n" |
| 294 | R"==(int m, int n, int k, global float *alpha, float beta, int last_k_block, )==" "\n" |
| 295 | R"==(float eltwise_alpha, float eltwise_beta, float eltwise_scale )==" "\n" |
| 296 | R"==(#ifdef WITH_K_UNROLL )==" "\n" |
| 297 | R"==(, )==" "\n" |
| 298 | R"==(volatile global int *flag, long offset_f) { )==" "\n" |
| 299 | R"==(#else )==" "\n" |
| 300 | R"==() { )==" "\n" |
| 301 | R"==(#endif )==" "\n" |
| 302 | R"==(half2 a[4]; )==" "\n" |
| 303 | R"==(half2 b[4]; )==" "\n" |
| 304 | R"==(half c[32][2]; )==" "\n" |
| 305 | R"==(int idM = get_global_id(1); )==" "\n" |
| 306 | R"==(int idN = get_global_id(0); )==" "\n" |
| 307 | R"==(int lid = get_sub_group_local_id(); )==" "\n" |
| 308 | R"==(int idK = get_global_id(2); )==" "\n" |
| 309 | R"==(int nku = get_global_size(2); )==" "\n" |
| 310 | R"==(int i0 = idM * 32; )==" "\n" |
| 311 | R"==(int j00 = sub_group_broadcast(idN, 0) * 2; )==" "\n" |
| 312 | R"==(int j0 = j00 + lid; )==" "\n" |
| 313 | R"==(int irem = m - i0; )==" "\n" |
| 314 | R"==(int jrem = n - j0; )==" "\n" |
| 315 | R"==(if (irem < 0) irem = 0; )==" "\n" |
| 316 | R"==(if (jrem < 0) jrem = 0; )==" "\n" |
| 317 | R"==(int irem2 = (irem + 1) >> 1; )==" "\n" |
| 318 | R"==(int last_k_unroll = (idK == nku - 1); )==" "\n" |
| 319 | R"==(#ifdef WITH_K_UNROLL )==" "\n" |
| 320 | R"==(int k0 = idK * UNROLL_K; )==" "\n" |
| 321 | R"==(int kt = k - k0; )==" "\n" |
| 322 | R"==(if (kt < 0) kt = 0; )==" "\n" |
| 323 | R"==(if (kt > UNROLL_K) kt = UNROLL_K; )==" "\n" |
| 324 | R"==(A += offset_a + i0 + k0 * lda; )==" "\n" |
| 325 | R"==(B += offset_b + j00 + k0 * ldb; )==" "\n" |
| 326 | R"==(k = kt; )==" "\n" |
| 327 | R"==(#else )==" "\n" |
| 328 | R"==(A += offset_a + i0; )==" "\n" |
| 329 | R"==(B += offset_b + j00; )==" "\n" |
| 330 | R"==(#endif )==" "\n" |
| 331 | R"==(C += offset_c + i0 + j0 * ldc; )==" "\n" |
| 332 | R"==(global half *A_ptrs[4] = {A, A + lda, A + 2 * lda, A + 3 * lda}; )==" "\n" |
| 333 | R"==(global half *B_ptrs[4] = {B, B + ldb, B + 2 * ldb, B + 3 * ldb}; )==" "\n" |
| 334 | R"==(for (int y = 0; y < 32; y++) )==" "\n" |
| 335 | R"==(for (int z = 0; z < 2; z++) )==" "\n" |
| 336 | R"==(c[y][z] = 0; )==" "\n" |
| 337 | R"==(#ifdef WITH_K_UNROLL )==" "\n" |
| 338 | R"==(flag += offset_f + idM; )==" "\n" |
| 339 | R"==(if (idK == 0 && lid == 0) *flag = 0; )==" "\n" |
| 340 | R"==(#endif )==" "\n" |
| 341 | R"==(int k_align = k & ~3; )==" "\n" |
| 342 | R"==(#ifndef ALLOW_READ_OVERRUNS )==" "\n" |
| 343 | R"==(if (irem >= 32 && sub_group_broadcast(jrem, 0) >= 32) { )==" "\n" |
| 344 | R"==(#endif )==" "\n" |
| 345 | R"==(for (int h = 0; h < k_align; h += 4) { )==" "\n" |
| 346 | R"==(for (int hh = 0; hh < 4; hh++) { )==" "\n" |
| 347 | R"==(a[hh] = as_half2(intel_sub_group_block_read( )==" "\n" |
| 348 | R"==((global uint *)(A_ptrs[hh] + h * lda))); )==" "\n" |
| 349 | R"==(b[hh] = as_half2(intel_sub_group_block_read_us2( )==" "\n" |
| 350 | R"==((global ushort *)(B_ptrs[hh] + h * ldb))); )==" "\n" |
| 351 | R"==(} )==" "\n" |
| 352 | R"==(FMA_I_LOOP(0); )==" "\n" |
| 353 | R"==(FMA_I_LOOP(1); )==" "\n" |
| 354 | R"==(FMA_I_LOOP(2); )==" "\n" |
| 355 | R"==(FMA_I_LOOP(3); )==" "\n" |
| 356 | R"==(} )==" "\n" |
| 357 | R"==(for (int h = k_align; h < k; h++) { )==" "\n" |
| 358 | R"==(a[0] = as_half2(intel_sub_group_block_read( )==" "\n" |
| 359 | R"==((global uint *)(A_ptrs[0] + h * lda))); )==" "\n" |
| 360 | R"==(b[0] = as_half2(intel_sub_group_block_read_us2( )==" "\n" |
| 361 | R"==((global ushort *)(B_ptrs[0] + h * ldb))); )==" "\n" |
| 362 | R"==(FMA_I_LOOP(0); )==" "\n" |
| 363 | R"==(} )==" "\n" |
| 364 | R"==(#ifndef ALLOW_READ_OVERRUNS )==" "\n" |
| 365 | R"==(} else { )==" "\n" |
| 366 | R"==(for (int h = 0; h < k_align; h += 4) { )==" "\n" |
| 367 | R"==(for (int hh = 0; hh < 4; hh++) { )==" "\n" |
| 368 | R"==(if (irem2 > lid) )==" "\n" |
| 369 | R"==(a[hh] = as_half2( )==" "\n" |
| 370 | R"==(*((global uint *)(A_ptrs[hh] + h * lda) + lid)); )==" "\n" |
| 371 | R"==(if (jrem > 0) b[hh].s0 = B_ptrs[hh][h * ldb + lid]; )==" "\n" |
| 372 | R"==(if (jrem > 16) b[hh].s1 = B_ptrs[hh][h * ldb + lid + 16]; )==" "\n" |
| 373 | R"==(} )==" "\n" |
| 374 | R"==(FMA_I_LOOP(0); )==" "\n" |
| 375 | R"==(FMA_I_LOOP(1); )==" "\n" |
| 376 | R"==(FMA_I_LOOP(2); )==" "\n" |
| 377 | R"==(FMA_I_LOOP(3); )==" "\n" |
| 378 | R"==(} )==" "\n" |
| 379 | R"==(for (int h = k_align; h < k; h++) { )==" "\n" |
| 380 | R"==(if (irem2 > lid) )==" "\n" |
| 381 | R"==(a[0] = as_half2(*((global uint *)(A_ptrs[0] + h * lda) + lid)); )==" "\n" |
| 382 | R"==(if (jrem > 0) b[0].s0 = B_ptrs[0][h * ldb + lid]; )==" "\n" |
| 383 | R"==(if (jrem > 16) b[0].s1 = B_ptrs[0][h * ldb + lid + 16]; )==" "\n" |
| 384 | R"==(FMA_I_LOOP(0); )==" "\n" |
| 385 | R"==(} )==" "\n" |
| 386 | R"==(} )==" "\n" |
| 387 | R"==(#endif )==" "\n" |
| 388 | R"==(global half *C2 = C + 16 * ldc; )==" "\n" |
| 389 | R"==(#ifdef WITH_K_UNROLL )==" "\n" |
| 390 | R"==(do { )==" "\n" |
| 391 | R"==(read_mem_fence(CLK_GLOBAL_MEM_FENCE); )==" "\n" |
| 392 | R"==(} while (*flag != idK); )==" "\n" |
| 393 | R"==(if (idK == 0) { )==" "\n" |
| 394 | R"==(if (beta == 0) )==" "\n" |
| 395 | R"==(UPDATE_C(1); )==" "\n" |
| 396 | R"==(else )==" "\n" |
| 397 | R"==(UPDATE_C(0); )==" "\n" |
| 398 | R"==(} else { )==" "\n" |
| 399 | R"==(beta = 1.0; )==" "\n" |
| 400 | R"==(UPDATE_C(0); )==" "\n" |
| 401 | R"==(} )==" "\n" |
| 402 | R"==(if (lid == 0) *flag = idK + 1; )==" "\n" |
| 403 | R"==(#else )==" "\n" |
| 404 | R"==(if (beta == 0) )==" "\n" |
| 405 | R"==(UPDATE_C(1); )==" "\n" |
| 406 | R"==(else )==" "\n" |
| 407 | R"==(UPDATE_C(0); )==" "\n" |
| 408 | R"==(#endif )==" "\n" |
| 409 | R"==(} )==" "\n" |
| 410 | R"==(#endif )==" "\n" |
| 411 | R"==(#ifdef TN )==" "\n" |
| 412 | R"==(__attribute__((intel_reqd_sub_group_size(16))) )==" "\n" |
| 413 | R"==(kernel void )==" "\n" |
| 414 | R"==(gen9_gemm_nocopy_f16(global half *A, global half *B, global half *C, )==" "\n" |
| 415 | R"==(long offset_a, long offset_b, long offset_c, int lda, int ldb, int ldc, )==" "\n" |
| 416 | R"==(int m, int n, int k, global float *alpha, float beta, int last_k_block, )==" "\n" |
| 417 | R"==(float eltwise_alpha, float eltwise_beta, float eltwise_scale )==" "\n" |
| 418 | R"==(#ifdef WITH_K_UNROLL )==" "\n" |
| 419 | R"==(, )==" "\n" |
| 420 | R"==(volatile global int *flag, long offset_f) { )==" "\n" |
| 421 | R"==(#else )==" "\n" |
| 422 | R"==() { )==" "\n" |
| 423 | R"==(#endif )==" "\n" |
| 424 | R"==(half4 a[2]; )==" "\n" |
| 425 | R"==(half4 b[2]; )==" "\n" |
| 426 | R"==(half c[32][2]; )==" "\n" |
| 427 | R"==(int idM = get_global_id(1); )==" "\n" |
| 428 | R"==(int idN = get_global_id(0); )==" "\n" |
| 429 | R"==(int lid = get_sub_group_local_id(); )==" "\n" |
| 430 | R"==(int idK = get_global_id(2); )==" "\n" |
| 431 | R"==(int nku = get_global_size(2); )==" "\n" |
| 432 | R"==(int i0 = idM * 32; )==" "\n" |
| 433 | R"==(int j0 = sub_group_broadcast(idN, 0) * 2 + lid; )==" "\n" |
| 434 | R"==(int irem = m - i0; )==" "\n" |
| 435 | R"==(int jrem = n - j0; )==" "\n" |
| 436 | R"==(if (irem < 0) irem = 0; )==" "\n" |
| 437 | R"==(if (jrem < 0) jrem = 0; )==" "\n" |
| 438 | R"==(int last_k_unroll = (idK == nku - 1); )==" "\n" |
| 439 | R"==(#ifdef WITH_K_UNROLL )==" "\n" |
| 440 | R"==(int k0 = idK * UNROLL_K; )==" "\n" |
| 441 | R"==(int kt = k - k0; )==" "\n" |
| 442 | R"==(if (kt < 0) kt = 0; )==" "\n" |
| 443 | R"==(if (kt > UNROLL_K) kt = UNROLL_K; )==" "\n" |
| 444 | R"==(A += offset_a + (i0 + lid) * lda + k0; )==" "\n" |
| 445 | R"==(B += offset_b + j0 * ldb + k0; )==" "\n" |
| 446 | R"==(k = kt; )==" "\n" |
| 447 | R"==(#else )==" "\n" |
| 448 | R"==(A += offset_a + (i0 + lid) * lda; )==" "\n" |
| 449 | R"==(B += offset_b + j0 * ldb; )==" "\n" |
| 450 | R"==(#endif )==" "\n" |
| 451 | R"==(C += offset_c + i0 + j0 * ldc; )==" "\n" |
| 452 | R"==(global half *A_ptrs[2] = {A, A + 16 * lda}; )==" "\n" |
| 453 | R"==(global half *B_ptrs[2] = {B, B + 16 * ldb}; )==" "\n" |
| 454 | R"==(for (int y = 0; y < 32; y++) )==" "\n" |
| 455 | R"==(for (int z = 0; z < 2; z++) )==" "\n" |
| 456 | R"==(c[y][z] = 0; )==" "\n" |
| 457 | R"==(#ifdef WITH_K_UNROLL )==" "\n" |
| 458 | R"==(flag += offset_f + idM; )==" "\n" |
| 459 | R"==(if (idK == 0 && lid == 0) *flag = 0; )==" "\n" |
| 460 | R"==(#endif )==" "\n" |
| 461 | R"==(int k_align = k & ~3; )==" "\n" |
| 462 | R"==(#ifndef ALLOW_READ_OVERRUNS )==" "\n" |
| 463 | R"==(if (irem >= 32 && sub_group_broadcast(jrem, 0) >= 32) { )==" "\n" |
| 464 | R"==(#endif )==" "\n" |
| 465 | R"==(for (int h = 0; h < k_align; h += 4) { )==" "\n" |
| 466 | R"==(for (int z = 0; z < 2; z++) { )==" "\n" |
| 467 | R"==(a[z] = VLOAD4_ALIGNED(0, (A_ptrs[z] + h)); )==" "\n" |
| 468 | R"==(b[z] = VLOAD4_ALIGNED(0, (B_ptrs[z] + h)); )==" "\n" |
| 469 | R"==(} )==" "\n" |
| 470 | R"==(FMA_I_LOOP(0); )==" "\n" |
| 471 | R"==(FMA_I_LOOP(1); )==" "\n" |
| 472 | R"==(FMA_I_LOOP(2); )==" "\n" |
| 473 | R"==(FMA_I_LOOP(3); )==" "\n" |
| 474 | R"==(} )==" "\n" |
| 475 | R"==(for (int h = k_align; h < k; h++) { )==" "\n" |
| 476 | R"==(for (int z = 0; z < 2; z++) { )==" "\n" |
| 477 | R"==(a[z] = A_ptrs[z][h]; )==" "\n" |
| 478 | R"==(b[z] = B_ptrs[z][h]; )==" "\n" |
| 479 | R"==(} )==" "\n" |
| 480 | R"==(FMA_I_LOOP(0); )==" "\n" |
| 481 | R"==(} )==" "\n" |
| 482 | R"==(#ifndef ALLOW_READ_OVERRUNS )==" "\n" |
| 483 | R"==(} else { )==" "\n" |
| 484 | R"==(for (int h = 0; h < k_align; h += 4) { )==" "\n" |
| 485 | R"==(for (int z = 0; z < 2; z++) { )==" "\n" |
| 486 | R"==(if (irem > (lid + z * 16)) )==" "\n" |
| 487 | R"==(a[z] = VLOAD4_ALIGNED(0, (A_ptrs[z] + h)); )==" "\n" |
| 488 | R"==(if (jrem > z * 16) b[z] = VLOAD4_ALIGNED(0, (B_ptrs[z] + h)); )==" "\n" |
| 489 | R"==(} )==" "\n" |
| 490 | R"==(FMA_I_LOOP(0); )==" "\n" |
| 491 | R"==(FMA_I_LOOP(1); )==" "\n" |
| 492 | R"==(FMA_I_LOOP(2); )==" "\n" |
| 493 | R"==(FMA_I_LOOP(3); )==" "\n" |
| 494 | R"==(} )==" "\n" |
| 495 | R"==(for (int h = k_align; h < k; h++) { )==" "\n" |
| 496 | R"==(for (int z = 0; z < 2; z++) { )==" "\n" |
| 497 | R"==(if (irem > (lid + z * 16)) a[z] = A_ptrs[z][h]; )==" "\n" |
| 498 | R"==(if (jrem > z * 16) b[z] = B_ptrs[z][h]; )==" "\n" |
| 499 | R"==(} )==" "\n" |
| 500 | R"==(FMA_I_LOOP(0); )==" "\n" |
| 501 | R"==(} )==" "\n" |
| 502 | R"==(} )==" "\n" |
| 503 | R"==(#endif )==" "\n" |
| 504 | R"==(global half *C2 = C + 16 * ldc; )==" "\n" |
| 505 | R"==(#ifdef WITH_K_UNROLL )==" "\n" |
| 506 | R"==(do { )==" "\n" |
| 507 | R"==(read_mem_fence(CLK_GLOBAL_MEM_FENCE); )==" "\n" |
| 508 | R"==(} while (*flag != idK); )==" "\n" |
| 509 | R"==(if (idK == 0) { )==" "\n" |
| 510 | R"==(if (beta == 0) )==" "\n" |
| 511 | R"==(UPDATE_C(1); )==" "\n" |
| 512 | R"==(else )==" "\n" |
| 513 | R"==(UPDATE_C(0); )==" "\n" |
| 514 | R"==(} else { )==" "\n" |
| 515 | R"==(beta = 1.0; )==" "\n" |
| 516 | R"==(UPDATE_C(0); )==" "\n" |
| 517 | R"==(} )==" "\n" |
| 518 | R"==(if (lid == 0) *flag = idK + 1; )==" "\n" |
| 519 | R"==(#else )==" "\n" |
| 520 | R"==(if (beta == 0) )==" "\n" |
| 521 | R"==(UPDATE_C(1); )==" "\n" |
| 522 | R"==(else )==" "\n" |
| 523 | R"==(UPDATE_C(0); )==" "\n" |
| 524 | R"==(#endif )==" "\n" |
| 525 | R"==(} )==" "\n" |
| 526 | R"==(#endif )==" "\n" |
| 527 | R"==(#ifdef TT )==" "\n" |
| 528 | R"==(__attribute__((intel_reqd_sub_group_size(16))) )==" "\n" |
| 529 | R"==(kernel void )==" "\n" |
| 530 | R"==(gen9_gemm_nocopy_f16(global half *A, global half *B, global half *C, )==" "\n" |
| 531 | R"==(long offset_a, long offset_b, long offset_c, int lda, int ldb, int ldc, )==" "\n" |
| 532 | R"==(int m, int n, int k, global float *alpha, float beta, int last_k_block, )==" "\n" |
| 533 | R"==(float eltwise_alpha, float eltwise_beta, float eltwise_scale )==" "\n" |
| 534 | R"==(#ifdef WITH_K_UNROLL )==" "\n" |
| 535 | R"==(, )==" "\n" |
| 536 | R"==(volatile global int *flag, long offset_f) { )==" "\n" |
| 537 | R"==(#else )==" "\n" |
| 538 | R"==() { )==" "\n" |
| 539 | R"==(#endif )==" "\n" |
| 540 | R"==(half4 a[2]; )==" "\n" |
| 541 | R"==(half2 b[4]; )==" "\n" |
| 542 | R"==(half c[32][2]; )==" "\n" |
| 543 | R"==(int idM = get_global_id(1); )==" "\n" |
| 544 | R"==(int idN = get_global_id(0); )==" "\n" |
| 545 | R"==(int lid = get_sub_group_local_id(); )==" "\n" |
| 546 | R"==(int idK = get_global_id(2); )==" "\n" |
| 547 | R"==(int nku = get_global_size(2); )==" "\n" |
| 548 | R"==(int i0 = idM * 32; )==" "\n" |
| 549 | R"==(int j00 = sub_group_broadcast(idN, 0) * 2; )==" "\n" |
| 550 | R"==(int j0 = j00 + lid; )==" "\n" |
| 551 | R"==(int irem = m - i0; )==" "\n" |
| 552 | R"==(int jrem = n - j0; )==" "\n" |
| 553 | R"==(if (irem < 0) irem = 0; )==" "\n" |
| 554 | R"==(if (jrem < 0) jrem = 0; )==" "\n" |
| 555 | R"==(int last_k_unroll = (idK == nku - 1); )==" "\n" |
| 556 | R"==(#ifdef WITH_K_UNROLL )==" "\n" |
| 557 | R"==(int k0 = idK * UNROLL_K; )==" "\n" |
| 558 | R"==(int kt = k - k0; )==" "\n" |
| 559 | R"==(if (kt < 0) kt = 0; )==" "\n" |
| 560 | R"==(if (kt > UNROLL_K) kt = UNROLL_K; )==" "\n" |
| 561 | R"==(A += offset_a + (i0 + lid) * lda + k0; )==" "\n" |
| 562 | R"==(B += offset_b + j00 + k0 * ldb; )==" "\n" |
| 563 | R"==(k = kt; )==" "\n" |
| 564 | R"==(#else )==" "\n" |
| 565 | R"==(A += offset_a + (i0 + lid) * lda; )==" "\n" |
| 566 | R"==(B += offset_b + j00; )==" "\n" |
| 567 | R"==(#endif )==" "\n" |
| 568 | R"==(C += offset_c + i0 + j0 * ldc; )==" "\n" |
| 569 | R"==(global half *A_ptrs[2] = {A, A + 16 * lda}; )==" "\n" |
| 570 | R"==(global half *B_ptrs[4] = {B, B + ldb, B + 2 * ldb, B + 3 * ldb}; )==" "\n" |
| 571 | R"==(for (int y = 0; y < 32; y++) )==" "\n" |
| 572 | R"==(for (int z = 0; z < 2; z++) )==" "\n" |
| 573 | R"==(c[y][z] = 0; )==" "\n" |
| 574 | R"==(#ifdef WITH_K_UNROLL )==" "\n" |
| 575 | R"==(flag += offset_f + idM; )==" "\n" |
| 576 | R"==(if (idK == 0 && lid == 0) *flag = 0; )==" "\n" |
| 577 | R"==(#endif )==" "\n" |
| 578 | R"==(int k_align = k & ~3; )==" "\n" |
| 579 | R"==(#ifndef ALLOW_READ_OVERRUNS )==" "\n" |
| 580 | R"==(if (irem >= 32 && sub_group_broadcast(jrem, 0) >= 32) { )==" "\n" |
| 581 | R"==(#endif )==" "\n" |
| 582 | R"==(for (int h = 0; h < k_align; h += 4) { )==" "\n" |
| 583 | R"==(for (int z = 0; z < 2; z++) )==" "\n" |
| 584 | R"==(a[z] = VLOAD4_ALIGNED(0, (A_ptrs[z] + h)); )==" "\n" |
| 585 | R"==(for (int hh = 0; hh < 4; hh++) )==" "\n" |
| 586 | R"==(b[hh] = as_half2(intel_sub_group_block_read_us2( )==" "\n" |
| 587 | R"==((global ushort *)(B_ptrs[hh] + h * ldb))); )==" "\n" |
| 588 | R"==(FMA_I_LOOP(0); )==" "\n" |
| 589 | R"==(FMA_I_LOOP(1); )==" "\n" |
| 590 | R"==(FMA_I_LOOP(2); )==" "\n" |
| 591 | R"==(FMA_I_LOOP(3); )==" "\n" |
| 592 | R"==(} )==" "\n" |
| 593 | R"==(for (int h = k_align; h < k; h++) { )==" "\n" |
| 594 | R"==(for (int z = 0; z < 2; z++) )==" "\n" |
| 595 | R"==(a[z] = A_ptrs[z][h]; )==" "\n" |
| 596 | R"==(b[0] = as_half2(intel_sub_group_block_read_us2( )==" "\n" |
| 597 | R"==((global ushort *)(B_ptrs[0] + h * ldb))); )==" "\n" |
| 598 | R"==(FMA_I_LOOP(0); )==" "\n" |
| 599 | R"==(} )==" "\n" |
| 600 | R"==(#ifndef ALLOW_READ_OVERRUNS )==" "\n" |
| 601 | R"==(} else { )==" "\n" |
| 602 | R"==(for (int h = 0; h < k_align; h += 4) { )==" "\n" |
| 603 | R"==(for (int z = 0; z < 2; z++) )==" "\n" |
| 604 | R"==(if (irem > (lid + z * 16)) )==" "\n" |
| 605 | R"==(a[z] = VLOAD4_ALIGNED(0, (A_ptrs[z] + h)); )==" "\n" |
| 606 | R"==(for (int hh = 0; hh < 4; hh++) { )==" "\n" |
| 607 | R"==(if (jrem > 0) b[hh].s0 = B_ptrs[hh][h * ldb + lid]; )==" "\n" |
| 608 | R"==(if (jrem > 16) b[hh].s1 = B_ptrs[hh][h * ldb + lid + 16]; )==" "\n" |
| 609 | R"==(} )==" "\n" |
| 610 | R"==(FMA_I_LOOP(0); )==" "\n" |
| 611 | R"==(FMA_I_LOOP(1); )==" "\n" |
| 612 | R"==(FMA_I_LOOP(2); )==" "\n" |
| 613 | R"==(FMA_I_LOOP(3); )==" "\n" |
| 614 | R"==(} )==" "\n" |
| 615 | R"==(for (int h = k_align; h < k; h++) { )==" "\n" |
| 616 | R"==(for (int z = 0; z < 2; z++) )==" "\n" |
| 617 | R"==(a[z] = A_ptrs[z][h]; )==" "\n" |
| 618 | R"==(if (jrem > 0) b[0].s0 = B_ptrs[0][h * ldb + lid]; )==" "\n" |
| 619 | R"==(if (jrem > 16) b[0].s1 = B_ptrs[0][h * ldb + lid + 16]; )==" "\n" |
| 620 | R"==(FMA_I_LOOP(0); )==" "\n" |
| 621 | R"==(} )==" "\n" |
| 622 | R"==(} )==" "\n" |
| 623 | R"==(#endif )==" "\n" |
| 624 | R"==(global half *C2 = C + 16 * ldc; )==" "\n" |
| 625 | R"==(#ifdef WITH_K_UNROLL )==" "\n" |
| 626 | R"==(do { )==" "\n" |
| 627 | R"==(read_mem_fence(CLK_GLOBAL_MEM_FENCE); )==" "\n" |
| 628 | R"==(} while (*flag != idK); )==" "\n" |
| 629 | R"==(if (idK == 0) { )==" "\n" |
| 630 | R"==(if (beta == 0) )==" "\n" |
| 631 | R"==(UPDATE_C(1); )==" "\n" |
| 632 | R"==(else )==" "\n" |
| 633 | R"==(UPDATE_C(0); )==" "\n" |
| 634 | R"==(} else { )==" "\n" |
| 635 | R"==(beta = 1.0; )==" "\n" |
| 636 | R"==(UPDATE_C(0); )==" "\n" |
| 637 | R"==(} )==" "\n" |
| 638 | R"==(if (lid == 0) *flag = idK + 1; )==" "\n" |
| 639 | R"==(#else )==" "\n" |
| 640 | R"==(if (beta == 0) )==" "\n" |
| 641 | R"==(UPDATE_C(1); )==" "\n" |
| 642 | R"==(else )==" "\n" |
| 643 | R"==(UPDATE_C(0); )==" "\n" |
| 644 | R"==(#endif )==" "\n" |
| 645 | R"==(} )==" "\n" |
| 646 | R"==(#endif )==" "\n" |
| 647 | R"==()=="; |
| 648 | } |
| 649 | } |
| 650 | } |
| 651 | } |
Generated on 2022-Nov-30 from project oneDNN revision 84bc949a1
Powered by 
Code Browser 2.1
Generator usage only permitted with license.