1 | namespace dnnl { |
2 | namespace impl { |
3 | namespace gpu { |
4 | namespace ocl { |
5 | const char *gen9_gemm_nocopy_superkernel_f32_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"==(#if DT_F32 != 1 )==" "\n" |
24 | R"==(#error "Only f32 implemented." )==" "\n" |
25 | R"==(#endif )==" "\n" |
26 | R"==(#define DO_FMA_NN(hh, i_mod_16, i_div_16, i_mod_4, i_div_4) \ )==" "\n" |
27 | R"==(do { \ )==" "\n" |
28 | R"==(c[i_div_4].s##i_mod_4 \ )==" "\n" |
29 | R"==(= mad(sub_group_broadcast(a[hh].s##i_div_16, i_mod_16), \ )==" "\n" |
30 | R"==(b.s##hh, c[i_div_4].s##i_mod_4); \ )==" "\n" |
31 | R"==(} while (0) )==" "\n" |
32 | R"==(#define DO_FMA_NT(hh, i_mod_16, i_div_16, i_mod_4, i_div_4) \ )==" "\n" |
33 | R"==(do { \ )==" "\n" |
34 | R"==(c[i_div_4].s##i_mod_4 \ )==" "\n" |
35 | R"==(= mad(sub_group_broadcast(a[hh].s##i_div_16, i_mod_16), b[hh], \ )==" "\n" |
36 | R"==(c[i_div_4].s##i_mod_4); \ )==" "\n" |
37 | R"==(} while (0) )==" "\n" |
38 | R"==(#if !defined(TRANS_A) )==" "\n" |
39 | R"==(#if !defined(TRANS_B) )==" "\n" |
40 | R"==(#define NN )==" "\n" |
41 | R"==(#define DO_FMA DO_FMA_NN )==" "\n" |
42 | R"==(#else )==" "\n" |
43 | R"==(#define NT )==" "\n" |
44 | R"==(#define DO_FMA DO_FMA_NT )==" "\n" |
45 | R"==(#endif )==" "\n" |
46 | R"==(#else )==" "\n" |
47 | R"==(#error "No superkernel implementation." )==" "\n" |
48 | R"==(#endif )==" "\n" |
49 | R"==(#define FMA_I_LOOP_32_ROW(hh) \ )==" "\n" |
50 | R"==(do { \ )==" "\n" |
51 | R"==(DO_FMA(hh, 0, 0, 0, 0); \ )==" "\n" |
52 | R"==(DO_FMA(hh, 1, 0, 1, 0); \ )==" "\n" |
53 | R"==(DO_FMA(hh, 2, 0, 2, 0); \ )==" "\n" |
54 | R"==(DO_FMA(hh, 3, 0, 3, 0); \ )==" "\n" |
55 | R"==(DO_FMA(hh, 4, 0, 0, 1); \ )==" "\n" |
56 | R"==(DO_FMA(hh, 5, 0, 1, 1); \ )==" "\n" |
57 | R"==(DO_FMA(hh, 6, 0, 2, 1); \ )==" "\n" |
58 | R"==(DO_FMA(hh, 7, 0, 3, 1); \ )==" "\n" |
59 | R"==(DO_FMA(hh, 8, 0, 0, 2); \ )==" "\n" |
60 | R"==(DO_FMA(hh, 9, 0, 1, 2); \ )==" "\n" |
61 | R"==(DO_FMA(hh, 10, 0, 2, 2); \ )==" "\n" |
62 | R"==(DO_FMA(hh, 11, 0, 3, 2); \ )==" "\n" |
63 | R"==(DO_FMA(hh, 12, 0, 0, 3); \ )==" "\n" |
64 | R"==(DO_FMA(hh, 13, 0, 1, 3); \ )==" "\n" |
65 | R"==(DO_FMA(hh, 14, 0, 2, 3); \ )==" "\n" |
66 | R"==(DO_FMA(hh, 15, 0, 3, 3); \ )==" "\n" |
67 | R"==(DO_FMA(hh, 16, 1, 0, 4); \ )==" "\n" |
68 | R"==(DO_FMA(hh, 17, 1, 1, 4); \ )==" "\n" |
69 | R"==(DO_FMA(hh, 18, 1, 2, 4); \ )==" "\n" |
70 | R"==(DO_FMA(hh, 19, 1, 3, 4); \ )==" "\n" |
71 | R"==(DO_FMA(hh, 20, 1, 0, 5); \ )==" "\n" |
72 | R"==(DO_FMA(hh, 21, 1, 1, 5); \ )==" "\n" |
73 | R"==(DO_FMA(hh, 22, 1, 2, 5); \ )==" "\n" |
74 | R"==(DO_FMA(hh, 23, 1, 3, 5); \ )==" "\n" |
75 | R"==(DO_FMA(hh, 24, 1, 0, 6); \ )==" "\n" |
76 | R"==(DO_FMA(hh, 25, 1, 1, 6); \ )==" "\n" |
77 | R"==(DO_FMA(hh, 26, 1, 2, 6); \ )==" "\n" |
78 | R"==(DO_FMA(hh, 27, 1, 3, 6); \ )==" "\n" |
79 | R"==(DO_FMA(hh, 28, 1, 0, 7); \ )==" "\n" |
80 | R"==(DO_FMA(hh, 29, 1, 1, 7); \ )==" "\n" |
81 | R"==(DO_FMA(hh, 30, 1, 2, 7); \ )==" "\n" |
82 | R"==(DO_FMA(hh, 31, 1, 3, 7); \ )==" "\n" |
83 | R"==(} while (0) )==" "\n" |
84 | R"==(#define FMA_I_LOOP_16_ROW(hh) \ )==" "\n" |
85 | R"==(do { \ )==" "\n" |
86 | R"==(DO_FMA(hh, 0, 0, 0, 0); \ )==" "\n" |
87 | R"==(DO_FMA(hh, 1, 0, 1, 0); \ )==" "\n" |
88 | R"==(DO_FMA(hh, 2, 0, 2, 0); \ )==" "\n" |
89 | R"==(DO_FMA(hh, 3, 0, 3, 0); \ )==" "\n" |
90 | R"==(DO_FMA(hh, 4, 0, 0, 1); \ )==" "\n" |
91 | R"==(DO_FMA(hh, 5, 0, 1, 1); \ )==" "\n" |
92 | R"==(DO_FMA(hh, 6, 0, 2, 1); \ )==" "\n" |
93 | R"==(DO_FMA(hh, 7, 0, 3, 1); \ )==" "\n" |
94 | R"==(DO_FMA(hh, 8, 0, 0, 2); \ )==" "\n" |
95 | R"==(DO_FMA(hh, 9, 0, 1, 2); \ )==" "\n" |
96 | R"==(DO_FMA(hh, 10, 0, 2, 2); \ )==" "\n" |
97 | R"==(DO_FMA(hh, 11, 0, 3, 2); \ )==" "\n" |
98 | R"==(DO_FMA(hh, 12, 0, 0, 3); \ )==" "\n" |
99 | R"==(DO_FMA(hh, 13, 0, 1, 3); \ )==" "\n" |
100 | R"==(DO_FMA(hh, 14, 0, 2, 3); \ )==" "\n" |
101 | R"==(DO_FMA(hh, 15, 0, 3, 3); \ )==" "\n" |
102 | R"==(} while (0) )==" "\n" |
103 | R"==(#if WITH_ELTWISE == 1 )==" "\n" |
104 | R"==(#define POST_OP(val) \ )==" "\n" |
105 | R"==(do { \ )==" "\n" |
106 | R"==(if (last_k_block) \ )==" "\n" |
107 | R"==(val = fwd_eltwise( \ )==" "\n" |
108 | R"==(val, eltwise_alpha, eltwise_beta, eltwise_scale); \ )==" "\n" |
109 | R"==(} while (0) )==" "\n" |
110 | R"==(#else )==" "\n" |
111 | R"==(#define POST_OP(val) )==" "\n" |
112 | R"==(#endif )==" "\n" |
113 | R"==(#define UPDATE_C_ROW(i, ii, betaZero) \ )==" "\n" |
114 | R"==(do { \ )==" "\n" |
115 | R"==(if (jrem > 0) \ )==" "\n" |
116 | R"==(if (irem > i) { \ )==" "\n" |
117 | R"==(float val = ATTR_ALPHA * c[i / 4].s##ii \ )==" "\n" |
118 | R"==(+ ((betaZero) ? 0 : beta * *C); \ )==" "\n" |
119 | R"==(POST_OP(val); \ )==" "\n" |
120 | R"==(*C = val; \ )==" "\n" |
121 | R"==(} \ )==" "\n" |
122 | R"==(C++; \ )==" "\n" |
123 | R"==(} while (0) )==" "\n" |
124 | R"==(#define UPDATE_C_32_ROW(betaZero) \ )==" "\n" |
125 | R"==(do { \ )==" "\n" |
126 | R"==(UPDATE_C_ROW(0, 0, betaZero); \ )==" "\n" |
127 | R"==(UPDATE_C_ROW(1, 1, betaZero); \ )==" "\n" |
128 | R"==(UPDATE_C_ROW(2, 2, betaZero); \ )==" "\n" |
129 | R"==(UPDATE_C_ROW(3, 3, betaZero); \ )==" "\n" |
130 | R"==(UPDATE_C_ROW(4, 0, betaZero); \ )==" "\n" |
131 | R"==(UPDATE_C_ROW(5, 1, betaZero); \ )==" "\n" |
132 | R"==(UPDATE_C_ROW(6, 2, betaZero); \ )==" "\n" |
133 | R"==(UPDATE_C_ROW(7, 3, betaZero); \ )==" "\n" |
134 | R"==(UPDATE_C_ROW(8, 0, betaZero); \ )==" "\n" |
135 | R"==(UPDATE_C_ROW(9, 1, betaZero); \ )==" "\n" |
136 | R"==(UPDATE_C_ROW(10, 2, betaZero); \ )==" "\n" |
137 | R"==(UPDATE_C_ROW(11, 3, betaZero); \ )==" "\n" |
138 | R"==(UPDATE_C_ROW(12, 0, betaZero); \ )==" "\n" |
139 | R"==(UPDATE_C_ROW(13, 1, betaZero); \ )==" "\n" |
140 | R"==(UPDATE_C_ROW(14, 2, betaZero); \ )==" "\n" |
141 | R"==(UPDATE_C_ROW(15, 3, betaZero); \ )==" "\n" |
142 | R"==(UPDATE_C_ROW(16, 0, betaZero); \ )==" "\n" |
143 | R"==(UPDATE_C_ROW(17, 1, betaZero); \ )==" "\n" |
144 | R"==(UPDATE_C_ROW(18, 2, betaZero); \ )==" "\n" |
145 | R"==(UPDATE_C_ROW(19, 3, betaZero); \ )==" "\n" |
146 | R"==(UPDATE_C_ROW(20, 0, betaZero); \ )==" "\n" |
147 | R"==(UPDATE_C_ROW(21, 1, betaZero); \ )==" "\n" |
148 | R"==(UPDATE_C_ROW(22, 2, betaZero); \ )==" "\n" |
149 | R"==(UPDATE_C_ROW(23, 3, betaZero); \ )==" "\n" |
150 | R"==(UPDATE_C_ROW(24, 0, betaZero); \ )==" "\n" |
151 | R"==(UPDATE_C_ROW(25, 1, betaZero); \ )==" "\n" |
152 | R"==(UPDATE_C_ROW(26, 2, betaZero); \ )==" "\n" |
153 | R"==(UPDATE_C_ROW(27, 3, betaZero); \ )==" "\n" |
154 | R"==(UPDATE_C_ROW(28, 0, betaZero); \ )==" "\n" |
155 | R"==(UPDATE_C_ROW(29, 1, betaZero); \ )==" "\n" |
156 | R"==(UPDATE_C_ROW(30, 2, betaZero); \ )==" "\n" |
157 | R"==(UPDATE_C_ROW(31, 3, betaZero); \ )==" "\n" |
158 | R"==(} while (0) )==" "\n" |
159 | R"==(#define SUPERKERNEL_PROLOGUE \ )==" "\n" |
160 | R"==(global volatile int *p = plan; \ )==" "\n" |
161 | R"==(int id = get_group_id(0); \ )==" "\n" |
162 | R"==(\ )==" "\n" |
163 | R"==(A0 += offsetA; \ )==" "\n" |
164 | R"==(B0 += offsetB; \ )==" "\n" |
165 | R"==(C0 += offsetC; \ )==" "\n" |
166 | R"==(\ )==" "\n" |
167 | R"==(while (id < threads) { \ )==" "\n" |
168 | R"==(uint i0, j0; \ )==" "\n" |
169 | R"==(uint kid0, kid1; \ )==" "\n" |
170 | R"==(\ )==" "\n" |
171 | R"==(i0 = plan[2 * id + 2]; \ )==" "\n" |
172 | R"==(j0 = plan[2 * id + 3]; \ )==" "\n" |
173 | R"==(kid0 = (i0 >> 31); \ )==" "\n" |
174 | R"==(kid1 = (j0 >> 31); \ )==" "\n" |
175 | R"==(i0 &= ~(1 << 31); \ )==" "\n" |
176 | R"==(j0 &= ~(1 << 31); \ )==" "\n" |
177 | R"==(j0 += get_local_id(0); )==" "\n" |
178 | R"==(#define SUPERKERNEL_EPILOGUE \ )==" "\n" |
179 | R"==(if (get_sub_group_local_id() == 0) id = atomic_inc(plan); \ )==" "\n" |
180 | R"==(\ )==" "\n" |
181 | R"==(sub_group_barrier(0); \ )==" "\n" |
182 | R"==(id = sub_group_broadcast(id, 0); \ )==" "\n" |
183 | R"==(} \ )==" "\n" |
184 | R"==(if (get_sub_group_local_id() == 0) { \ )==" "\n" |
185 | R"==(if (atomic_inc(plan + 1) == (get_num_groups(0) - 1)) { \ )==" "\n" |
186 | R"==(mem_fence(CLK_GLOBAL_MEM_FENCE); \ )==" "\n" |
187 | R"==(plan[0] = get_num_groups(0); \ )==" "\n" |
188 | R"==(plan[1] = 0; \ )==" "\n" |
189 | R"==(} \ )==" "\n" |
190 | R"==(} )==" "\n" |
191 | R"==(#ifdef NN )==" "\n" |
192 | R"==(__attribute__((intel_reqd_sub_group_size(16))) )==" "\n" |
193 | R"==(kernel void )==" "\n" |
194 | R"==(gen9_gemm_nocopy_superkernel_f32(global int *plan, int threads, )==" "\n" |
195 | R"==(global float *A0, global float *B0, global float *C0, long offsetA, )==" "\n" |
196 | R"==(long offsetB, long offsetC, int lda, int ldb, int ldc, int m, int n, )==" "\n" |
197 | R"==(int k, global float *alpha, float beta, int last_k_block, )==" "\n" |
198 | R"==(float eltwise_alpha, float eltwise_beta, float eltwise_scale) { )==" "\n" |
199 | R"==(SUPERKERNEL_PROLOGUE )==" "\n" |
200 | R"==(float2 a[4]; )==" "\n" |
201 | R"==(float4 b; )==" "\n" |
202 | R"==(float4 c[8]; )==" "\n" |
203 | R"==(int irem = m - i0; )==" "\n" |
204 | R"==(int jrem = n - j0; )==" "\n" |
205 | R"==(if (irem < 0) irem = 0; )==" "\n" |
206 | R"==(if (jrem < 0) jrem = 0; )==" "\n" |
207 | R"==(global float *A = A0 + i0; )==" "\n" |
208 | R"==(global float *B = B0 + j0 * ldb; )==" "\n" |
209 | R"==(global float *C = C0 + i0 + j0 * ldc; )==" "\n" |
210 | R"==(global float *A_cols[4] = {A, A + lda, A + 2 * lda, A + 3 * lda}; )==" "\n" |
211 | R"==(int ldax4 = lda << 2; )==" "\n" |
212 | R"==(int ldbx4 = ldb << 2; )==" "\n" |
213 | R"==(if (kid0 == 0) { )==" "\n" |
214 | R"==(for (int z = 0; z < 8; z++) )==" "\n" |
215 | R"==(c[z] = 0.f; )==" "\n" |
216 | R"==(for (int h = 0; h < (k >> 2); h++) { )==" "\n" |
217 | R"==(for (int j = 0; j < 4; j++) { )==" "\n" |
218 | R"==(a[j] = as_float2( )==" "\n" |
219 | R"==(intel_sub_group_block_read2((global uint *)A_cols[j])); )==" "\n" |
220 | R"==(A_cols[j] += ldax4; )==" "\n" |
221 | R"==(} )==" "\n" |
222 | R"==(b = vload4(0, B); )==" "\n" |
223 | R"==(B += 4; )==" "\n" |
224 | R"==(FMA_I_LOOP_32_ROW(0); )==" "\n" |
225 | R"==(FMA_I_LOOP_32_ROW(1); )==" "\n" |
226 | R"==(FMA_I_LOOP_32_ROW(2); )==" "\n" |
227 | R"==(FMA_I_LOOP_32_ROW(3); )==" "\n" |
228 | R"==(} )==" "\n" |
229 | R"==(int krem = k & 3; )==" "\n" |
230 | R"==(if (krem > 0) { )==" "\n" |
231 | R"==(for (int j = 0; j < 4; j++) )==" "\n" |
232 | R"==(a[j] = as_float2( )==" "\n" |
233 | R"==(intel_sub_group_block_read2((global uint *)A_cols[j])); )==" "\n" |
234 | R"==(b = vload4(0, B); )==" "\n" |
235 | R"==(FMA_I_LOOP_32_ROW(0); )==" "\n" |
236 | R"==(if (krem > 1) FMA_I_LOOP_32_ROW(1); )==" "\n" |
237 | R"==(if (krem > 2) FMA_I_LOOP_32_ROW(2); )==" "\n" |
238 | R"==(} )==" "\n" |
239 | R"==(} else { )==" "\n" |
240 | R"==(if (irem > 16) irem = 16; )==" "\n" |
241 | R"==(for (int z = 0; z < 4; z++) )==" "\n" |
242 | R"==(c[z] = 0.f; )==" "\n" |
243 | R"==(for (int h = 0; h < (k >> 2); h++) { )==" "\n" |
244 | R"==(for (int j = 0; j < 4; j++) { )==" "\n" |
245 | R"==(a[j].s0 = as_float( )==" "\n" |
246 | R"==(intel_sub_group_block_read((global uint *)A_cols[j])); )==" "\n" |
247 | R"==(A_cols[j] += ldax4; )==" "\n" |
248 | R"==(} )==" "\n" |
249 | R"==(b = vload4(0, B); )==" "\n" |
250 | R"==(B += 4; )==" "\n" |
251 | R"==(FMA_I_LOOP_16_ROW(0); )==" "\n" |
252 | R"==(FMA_I_LOOP_16_ROW(1); )==" "\n" |
253 | R"==(FMA_I_LOOP_16_ROW(2); )==" "\n" |
254 | R"==(FMA_I_LOOP_16_ROW(3); )==" "\n" |
255 | R"==(} )==" "\n" |
256 | R"==(int krem = k & 3; )==" "\n" |
257 | R"==(if (krem > 0) { )==" "\n" |
258 | R"==(for (int j = 0; j < 4; j++) )==" "\n" |
259 | R"==(a[j].s0 = as_float( )==" "\n" |
260 | R"==(intel_sub_group_block_read((global uint *)A_cols[j])); )==" "\n" |
261 | R"==(b = vload4(0, B); )==" "\n" |
262 | R"==(FMA_I_LOOP_16_ROW(0); )==" "\n" |
263 | R"==(if (krem > 1) FMA_I_LOOP_16_ROW(1); )==" "\n" |
264 | R"==(if (krem > 2) FMA_I_LOOP_16_ROW(2); )==" "\n" |
265 | R"==(} )==" "\n" |
266 | R"==(} )==" "\n" |
267 | R"==(if (beta == 0) )==" "\n" |
268 | R"==(UPDATE_C_32_ROW(1); )==" "\n" |
269 | R"==(else )==" "\n" |
270 | R"==(UPDATE_C_32_ROW(0); )==" "\n" |
271 | R"==(SUPERKERNEL_EPILOGUE )==" "\n" |
272 | R"==(} )==" "\n" |
273 | R"==(#endif )==" "\n" |
274 | R"==(#ifdef NT )==" "\n" |
275 | R"==(__attribute__((intel_reqd_sub_group_size(16))) )==" "\n" |
276 | R"==(kernel void )==" "\n" |
277 | R"==(gen9_gemm_nocopy_superkernel_f32(global int *plan, int threads, )==" "\n" |
278 | R"==(global float *A0, global float *B0, global float *C0, long offsetA, )==" "\n" |
279 | R"==(long offsetB, long offsetC, int lda, int ldb, int ldc, int m, int n, )==" "\n" |
280 | R"==(int k, global float *alpha, float beta, int last_k_block, )==" "\n" |
281 | R"==(float eltwise_alpha, float eltwise_beta, float eltwise_scale) { )==" "\n" |
282 | R"==(SUPERKERNEL_PROLOGUE )==" "\n" |
283 | R"==(float2 a[2]; )==" "\n" |
284 | R"==(float b[2]; )==" "\n" |
285 | R"==(float4 c[8]; )==" "\n" |
286 | R"==(int irem = m - i0; )==" "\n" |
287 | R"==(int jrem = n - j0; )==" "\n" |
288 | R"==(if (irem < 0) irem = 0; )==" "\n" |
289 | R"==(if (jrem < 0) jrem = 0; )==" "\n" |
290 | R"==(global float *A = A0 + i0; )==" "\n" |
291 | R"==(global float *B = B0 + j0; )==" "\n" |
292 | R"==(global float *C = C0 + i0 + j0 * ldc; )==" "\n" |
293 | R"==(global float *A_cols[2] = {A, A + lda}; )==" "\n" |
294 | R"==(global float *B_rows[2] = {B, B + ldb}; )==" "\n" |
295 | R"==(int ldax2 = lda << 1; )==" "\n" |
296 | R"==(int ldbx2 = ldb << 1; )==" "\n" |
297 | R"==(if (kid0 == 0) { )==" "\n" |
298 | R"==(for (int z = 0; z < 8; z++) )==" "\n" |
299 | R"==(c[z] = 0.f; )==" "\n" |
300 | R"==(for (int h = 0; h < (k >> 1); h++) { )==" "\n" |
301 | R"==(for (int j = 0; j < 2; j++) { )==" "\n" |
302 | R"==(a[j] = as_float2( )==" "\n" |
303 | R"==(intel_sub_group_block_read2((global uint *)A_cols[j])); )==" "\n" |
304 | R"==(A_cols[j] += ldax2; )==" "\n" |
305 | R"==(} )==" "\n" |
306 | R"==(for (int i = 0; i < 2; i++) { )==" "\n" |
307 | R"==(b[i] = as_float( )==" "\n" |
308 | R"==(intel_sub_group_block_read((global uint *)B_rows[i])); )==" "\n" |
309 | R"==(B_rows[i] += ldbx2; )==" "\n" |
310 | R"==(} )==" "\n" |
311 | R"==(FMA_I_LOOP_32_ROW(0); )==" "\n" |
312 | R"==(FMA_I_LOOP_32_ROW(1); )==" "\n" |
313 | R"==(} )==" "\n" |
314 | R"==(int krem = k & 1; )==" "\n" |
315 | R"==(if (krem > 0) { )==" "\n" |
316 | R"==(a[0] = as_float2( )==" "\n" |
317 | R"==(intel_sub_group_block_read2((global uint *)A_cols[0])); )==" "\n" |
318 | R"==(b[0] = as_float( )==" "\n" |
319 | R"==(intel_sub_group_block_read((global uint *)B_rows[0])); )==" "\n" |
320 | R"==(FMA_I_LOOP_32_ROW(0); )==" "\n" |
321 | R"==(} )==" "\n" |
322 | R"==(} else { )==" "\n" |
323 | R"==(if (irem > 16) irem = 16; )==" "\n" |
324 | R"==(for (int z = 0; z < 4; z++) )==" "\n" |
325 | R"==(c[z] = 0.f; )==" "\n" |
326 | R"==(for (int h = 0; h < (k >> 1); h++) { )==" "\n" |
327 | R"==(for (int j = 0; j < 2; j++) { )==" "\n" |
328 | R"==(a[j].s0 = as_float( )==" "\n" |
329 | R"==(intel_sub_group_block_read((global uint *)A_cols[j])); )==" "\n" |
330 | R"==(A_cols[j] += ldax2; )==" "\n" |
331 | R"==(} )==" "\n" |
332 | R"==(for (int i = 0; i < 2; i++) { )==" "\n" |
333 | R"==(b[i] = as_float( )==" "\n" |
334 | R"==(intel_sub_group_block_read((global uint *)B_rows[i])); )==" "\n" |
335 | R"==(B_rows[i] += ldbx2; )==" "\n" |
336 | R"==(} )==" "\n" |
337 | R"==(FMA_I_LOOP_16_ROW(0); )==" "\n" |
338 | R"==(FMA_I_LOOP_16_ROW(1); )==" "\n" |
339 | R"==(} )==" "\n" |
340 | R"==(int krem = k & 1; )==" "\n" |
341 | R"==(if (krem > 0) { )==" "\n" |
342 | R"==(a[0].s0 = as_float( )==" "\n" |
343 | R"==(intel_sub_group_block_read((global uint *)A_cols[0])); )==" "\n" |
344 | R"==(b[0] = as_float( )==" "\n" |
345 | R"==(intel_sub_group_block_read((global uint *)B_rows[0])); )==" "\n" |
346 | R"==(FMA_I_LOOP_16_ROW(0); )==" "\n" |
347 | R"==(} )==" "\n" |
348 | R"==(} )==" "\n" |
349 | R"==(if (beta == 0) )==" "\n" |
350 | R"==(UPDATE_C_32_ROW(1); )==" "\n" |
351 | R"==(else )==" "\n" |
352 | R"==(UPDATE_C_32_ROW(0); )==" "\n" |
353 | R"==(SUPERKERNEL_EPILOGUE )==" "\n" |
354 | R"==(} )==" "\n" |
355 | R"==(#endif )==" "\n" |
356 | R"==()==" ; |
357 | } |
358 | } |
359 | } |
360 | } |