1 | namespace dnnl { |
2 | namespace impl { |
3 | namespace gpu { |
4 | namespace ocl { |
5 | const char *gen9_gemm_nocopy_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 "Incorrect datatype." )==" "\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"==(#define DO_FMA_TN(hh, i, i_mod_4, i_div_4) \ )==" "\n" |
39 | R"==(do { \ )==" "\n" |
40 | R"==(c[i_div_4][0].s##i_mod_4 = mad(sub_group_broadcast(a.s##hh, i), \ )==" "\n" |
41 | R"==(b[0].s##hh, c[i_div_4][0].s##i_mod_4); \ )==" "\n" |
42 | R"==(c[i_div_4][1].s##i_mod_4 = mad(sub_group_broadcast(a.s##hh, i), \ )==" "\n" |
43 | R"==(b[1].s##hh, c[i_div_4][1].s##i_mod_4); \ )==" "\n" |
44 | R"==(} while (0) )==" "\n" |
45 | R"==(#define DO_FMA_TT(hh, i, i_mod_4, i_div_4) \ )==" "\n" |
46 | R"==(do { \ )==" "\n" |
47 | R"==(c[i_div_4][0].s##i_mod_4 = mad(sub_group_broadcast(a.s##hh, i), \ )==" "\n" |
48 | R"==(b[hh].s0, c[i_div_4][0].s##i_mod_4); \ )==" "\n" |
49 | R"==(c[i_div_4][1].s##i_mod_4 = mad(sub_group_broadcast(a.s##hh, i), \ )==" "\n" |
50 | R"==(b[hh].s1, c[i_div_4][1].s##i_mod_4); \ )==" "\n" |
51 | R"==(} while (0) )==" "\n" |
52 | R"==(#if !defined(TRANS_A) )==" "\n" |
53 | R"==(#if !defined(TRANS_B) )==" "\n" |
54 | R"==(#define NN )==" "\n" |
55 | R"==(#define DO_FMA DO_FMA_NN )==" "\n" |
56 | R"==(#else )==" "\n" |
57 | R"==(#define NT )==" "\n" |
58 | R"==(#define DO_FMA DO_FMA_NT )==" "\n" |
59 | R"==(#endif )==" "\n" |
60 | R"==(#else )==" "\n" |
61 | R"==(#if !defined(TRANS_B) )==" "\n" |
62 | R"==(#define TN )==" "\n" |
63 | R"==(#define DO_FMA DO_FMA_TN )==" "\n" |
64 | R"==(#else )==" "\n" |
65 | R"==(#define TT )==" "\n" |
66 | R"==(#define DO_FMA DO_FMA_TT )==" "\n" |
67 | R"==(#endif )==" "\n" |
68 | R"==(#endif )==" "\n" |
69 | R"==(#if WITH_ELTWISE == 1 )==" "\n" |
70 | R"==(#define POST_OP(val) \ )==" "\n" |
71 | R"==(do { \ )==" "\n" |
72 | R"==(if (last_k_block && last_k_unroll) \ )==" "\n" |
73 | R"==(val = fwd_eltwise( \ )==" "\n" |
74 | R"==(val, eltwise_alpha, eltwise_beta, eltwise_scale); \ )==" "\n" |
75 | R"==(} while (0) )==" "\n" |
76 | R"==(#else )==" "\n" |
77 | R"==(#define POST_OP(val) )==" "\n" |
78 | R"==(#endif )==" "\n" |
79 | R"==(#define FMA_I_LOOP_32_ROW(hh) \ )==" "\n" |
80 | R"==(do { \ )==" "\n" |
81 | R"==(DO_FMA(hh, 0, 0, 0, 0); \ )==" "\n" |
82 | R"==(DO_FMA(hh, 1, 0, 1, 0); \ )==" "\n" |
83 | R"==(DO_FMA(hh, 2, 0, 2, 0); \ )==" "\n" |
84 | R"==(DO_FMA(hh, 3, 0, 3, 0); \ )==" "\n" |
85 | R"==(DO_FMA(hh, 4, 0, 0, 1); \ )==" "\n" |
86 | R"==(DO_FMA(hh, 5, 0, 1, 1); \ )==" "\n" |
87 | R"==(DO_FMA(hh, 6, 0, 2, 1); \ )==" "\n" |
88 | R"==(DO_FMA(hh, 7, 0, 3, 1); \ )==" "\n" |
89 | R"==(DO_FMA(hh, 8, 0, 0, 2); \ )==" "\n" |
90 | R"==(DO_FMA(hh, 9, 0, 1, 2); \ )==" "\n" |
91 | R"==(DO_FMA(hh, 10, 0, 2, 2); \ )==" "\n" |
92 | R"==(DO_FMA(hh, 11, 0, 3, 2); \ )==" "\n" |
93 | R"==(DO_FMA(hh, 12, 0, 0, 3); \ )==" "\n" |
94 | R"==(DO_FMA(hh, 13, 0, 1, 3); \ )==" "\n" |
95 | R"==(DO_FMA(hh, 14, 0, 2, 3); \ )==" "\n" |
96 | R"==(DO_FMA(hh, 15, 0, 3, 3); \ )==" "\n" |
97 | R"==(DO_FMA(hh, 16, 1, 0, 4); \ )==" "\n" |
98 | R"==(DO_FMA(hh, 17, 1, 1, 4); \ )==" "\n" |
99 | R"==(DO_FMA(hh, 18, 1, 2, 4); \ )==" "\n" |
100 | R"==(DO_FMA(hh, 19, 1, 3, 4); \ )==" "\n" |
101 | R"==(DO_FMA(hh, 20, 1, 0, 5); \ )==" "\n" |
102 | R"==(DO_FMA(hh, 21, 1, 1, 5); \ )==" "\n" |
103 | R"==(DO_FMA(hh, 22, 1, 2, 5); \ )==" "\n" |
104 | R"==(DO_FMA(hh, 23, 1, 3, 5); \ )==" "\n" |
105 | R"==(DO_FMA(hh, 24, 1, 0, 6); \ )==" "\n" |
106 | R"==(DO_FMA(hh, 25, 1, 1, 6); \ )==" "\n" |
107 | R"==(DO_FMA(hh, 26, 1, 2, 6); \ )==" "\n" |
108 | R"==(DO_FMA(hh, 27, 1, 3, 6); \ )==" "\n" |
109 | R"==(DO_FMA(hh, 28, 1, 0, 7); \ )==" "\n" |
110 | R"==(DO_FMA(hh, 29, 1, 1, 7); \ )==" "\n" |
111 | R"==(DO_FMA(hh, 30, 1, 2, 7); \ )==" "\n" |
112 | R"==(DO_FMA(hh, 31, 1, 3, 7); \ )==" "\n" |
113 | R"==(} while (0) )==" "\n" |
114 | R"==(#define FMA_I_LOOP_16_ROW(hh) \ )==" "\n" |
115 | R"==(do { \ )==" "\n" |
116 | R"==(DO_FMA(hh, 0, 0, 0); \ )==" "\n" |
117 | R"==(DO_FMA(hh, 1, 1, 0); \ )==" "\n" |
118 | R"==(DO_FMA(hh, 2, 2, 0); \ )==" "\n" |
119 | R"==(DO_FMA(hh, 3, 3, 0); \ )==" "\n" |
120 | R"==(DO_FMA(hh, 4, 0, 1); \ )==" "\n" |
121 | R"==(DO_FMA(hh, 5, 1, 1); \ )==" "\n" |
122 | R"==(DO_FMA(hh, 6, 2, 1); \ )==" "\n" |
123 | R"==(DO_FMA(hh, 7, 3, 1); \ )==" "\n" |
124 | R"==(DO_FMA(hh, 8, 0, 2); \ )==" "\n" |
125 | R"==(DO_FMA(hh, 9, 1, 2); \ )==" "\n" |
126 | R"==(DO_FMA(hh, 10, 2, 2); \ )==" "\n" |
127 | R"==(DO_FMA(hh, 11, 3, 2); \ )==" "\n" |
128 | R"==(DO_FMA(hh, 12, 0, 3); \ )==" "\n" |
129 | R"==(DO_FMA(hh, 13, 1, 3); \ )==" "\n" |
130 | R"==(DO_FMA(hh, 14, 2, 3); \ )==" "\n" |
131 | R"==(DO_FMA(hh, 15, 3, 3); \ )==" "\n" |
132 | R"==(} while (0) )==" "\n" |
133 | R"==(#define UPDATE_C_ROW(i, ii, betaZero) \ )==" "\n" |
134 | R"==(do { \ )==" "\n" |
135 | R"==(if (jrem > 0) \ )==" "\n" |
136 | R"==(if (irem > i) { \ )==" "\n" |
137 | R"==(float val = ATTR_ALPHA * c[i / 4].s##ii \ )==" "\n" |
138 | R"==(+ ((betaZero) ? 0 : beta * *C); \ )==" "\n" |
139 | R"==(POST_OP(val); \ )==" "\n" |
140 | R"==(*C = val; \ )==" "\n" |
141 | R"==(} \ )==" "\n" |
142 | R"==(C++; \ )==" "\n" |
143 | R"==(} while (0) )==" "\n" |
144 | R"==(#define UPDATE_C_ROW_2X(i, ii, betaZero) \ )==" "\n" |
145 | R"==(do { \ )==" "\n" |
146 | R"==(if (irem > i) { \ )==" "\n" |
147 | R"==(if (jrem > 0) { \ )==" "\n" |
148 | R"==(float val = ATTR_ALPHA * c[i / 4][0].s##ii \ )==" "\n" |
149 | R"==(+ ((betaZero) ? 0 : beta * *(C_ptrs[0])); \ )==" "\n" |
150 | R"==(POST_OP(val); \ )==" "\n" |
151 | R"==(*(C_ptrs[0]) = val; \ )==" "\n" |
152 | R"==(} \ )==" "\n" |
153 | R"==(if (jrem > 16) { \ )==" "\n" |
154 | R"==(float val = ATTR_ALPHA * c[i / 4][1].s##ii \ )==" "\n" |
155 | R"==(+ ((betaZero) ? 0 : beta * *(C_ptrs[1])); \ )==" "\n" |
156 | R"==(POST_OP(val); \ )==" "\n" |
157 | R"==(*(C_ptrs[1]) = val; \ )==" "\n" |
158 | R"==(} \ )==" "\n" |
159 | R"==(} \ )==" "\n" |
160 | R"==(C_ptrs[0]++; \ )==" "\n" |
161 | R"==(C_ptrs[1]++; \ )==" "\n" |
162 | R"==(} while (0) )==" "\n" |
163 | R"==(#define UPDATE_C_32_ROW(betaZero) \ )==" "\n" |
164 | R"==(do { \ )==" "\n" |
165 | R"==(UPDATE_C_ROW(0, 0, betaZero); \ )==" "\n" |
166 | R"==(UPDATE_C_ROW(1, 1, betaZero); \ )==" "\n" |
167 | R"==(UPDATE_C_ROW(2, 2, betaZero); \ )==" "\n" |
168 | R"==(UPDATE_C_ROW(3, 3, betaZero); \ )==" "\n" |
169 | R"==(UPDATE_C_ROW(4, 0, betaZero); \ )==" "\n" |
170 | R"==(UPDATE_C_ROW(5, 1, betaZero); \ )==" "\n" |
171 | R"==(UPDATE_C_ROW(6, 2, betaZero); \ )==" "\n" |
172 | R"==(UPDATE_C_ROW(7, 3, betaZero); \ )==" "\n" |
173 | R"==(UPDATE_C_ROW(8, 0, betaZero); \ )==" "\n" |
174 | R"==(UPDATE_C_ROW(9, 1, betaZero); \ )==" "\n" |
175 | R"==(UPDATE_C_ROW(10, 2, betaZero); \ )==" "\n" |
176 | R"==(UPDATE_C_ROW(11, 3, betaZero); \ )==" "\n" |
177 | R"==(UPDATE_C_ROW(12, 0, betaZero); \ )==" "\n" |
178 | R"==(UPDATE_C_ROW(13, 1, betaZero); \ )==" "\n" |
179 | R"==(UPDATE_C_ROW(14, 2, betaZero); \ )==" "\n" |
180 | R"==(UPDATE_C_ROW(15, 3, betaZero); \ )==" "\n" |
181 | R"==(UPDATE_C_ROW(16, 0, betaZero); \ )==" "\n" |
182 | R"==(UPDATE_C_ROW(17, 1, betaZero); \ )==" "\n" |
183 | R"==(UPDATE_C_ROW(18, 2, betaZero); \ )==" "\n" |
184 | R"==(UPDATE_C_ROW(19, 3, betaZero); \ )==" "\n" |
185 | R"==(UPDATE_C_ROW(20, 0, betaZero); \ )==" "\n" |
186 | R"==(UPDATE_C_ROW(21, 1, betaZero); \ )==" "\n" |
187 | R"==(UPDATE_C_ROW(22, 2, betaZero); \ )==" "\n" |
188 | R"==(UPDATE_C_ROW(23, 3, betaZero); \ )==" "\n" |
189 | R"==(UPDATE_C_ROW(24, 0, betaZero); \ )==" "\n" |
190 | R"==(UPDATE_C_ROW(25, 1, betaZero); \ )==" "\n" |
191 | R"==(UPDATE_C_ROW(26, 2, betaZero); \ )==" "\n" |
192 | R"==(UPDATE_C_ROW(27, 3, betaZero); \ )==" "\n" |
193 | R"==(UPDATE_C_ROW(28, 0, betaZero); \ )==" "\n" |
194 | R"==(UPDATE_C_ROW(29, 1, betaZero); \ )==" "\n" |
195 | R"==(UPDATE_C_ROW(30, 2, betaZero); \ )==" "\n" |
196 | R"==(UPDATE_C_ROW(31, 3, betaZero); \ )==" "\n" |
197 | R"==(} while (0) )==" "\n" |
198 | R"==(#define UPDATE_C_16_ROW(betaZero) \ )==" "\n" |
199 | R"==(do { \ )==" "\n" |
200 | R"==(UPDATE_C_ROW_2X(0, 0, betaZero); \ )==" "\n" |
201 | R"==(UPDATE_C_ROW_2X(1, 1, betaZero); \ )==" "\n" |
202 | R"==(UPDATE_C_ROW_2X(2, 2, betaZero); \ )==" "\n" |
203 | R"==(UPDATE_C_ROW_2X(3, 3, betaZero); \ )==" "\n" |
204 | R"==(UPDATE_C_ROW_2X(4, 0, betaZero); \ )==" "\n" |
205 | R"==(UPDATE_C_ROW_2X(5, 1, betaZero); \ )==" "\n" |
206 | R"==(UPDATE_C_ROW_2X(6, 2, betaZero); \ )==" "\n" |
207 | R"==(UPDATE_C_ROW_2X(7, 3, betaZero); \ )==" "\n" |
208 | R"==(UPDATE_C_ROW_2X(8, 0, betaZero); \ )==" "\n" |
209 | R"==(UPDATE_C_ROW_2X(9, 1, betaZero); \ )==" "\n" |
210 | R"==(UPDATE_C_ROW_2X(10, 2, betaZero); \ )==" "\n" |
211 | R"==(UPDATE_C_ROW_2X(11, 3, betaZero); \ )==" "\n" |
212 | R"==(UPDATE_C_ROW_2X(12, 0, betaZero); \ )==" "\n" |
213 | R"==(UPDATE_C_ROW_2X(13, 1, betaZero); \ )==" "\n" |
214 | R"==(UPDATE_C_ROW_2X(14, 2, betaZero); \ )==" "\n" |
215 | R"==(UPDATE_C_ROW_2X(15, 3, betaZero); \ )==" "\n" |
216 | R"==(} while (0) )==" "\n" |
217 | R"==(#ifdef NN )==" "\n" |
218 | R"==(__attribute__((intel_reqd_sub_group_size(16))) )==" "\n" |
219 | R"==(kernel void )==" "\n" |
220 | R"==(gen9_gemm_nocopy_f32(global float *A, global float *B, global float *C, )==" "\n" |
221 | R"==(long offset_a, long offset_b, long offset_c, int lda, int ldb, int ldc, )==" "\n" |
222 | R"==(int m, int n, int k, global float *alpha, float beta, int last_k_block, )==" "\n" |
223 | R"==(float eltwise_alpha, float eltwise_beta, float eltwise_scale )==" "\n" |
224 | R"==(#ifdef WITH_K_UNROLL )==" "\n" |
225 | R"==(, )==" "\n" |
226 | R"==(volatile global int *flag, long offset_f) { )==" "\n" |
227 | R"==(#else )==" "\n" |
228 | R"==() { )==" "\n" |
229 | R"==(#endif )==" "\n" |
230 | R"==(float2 a[4]; )==" "\n" |
231 | R"==(float4 b; )==" "\n" |
232 | R"==(float4 c[8]; )==" "\n" |
233 | R"==(int idM = get_global_id(1); )==" "\n" |
234 | R"==(int idN = get_global_id(0); )==" "\n" |
235 | R"==(int lid = get_sub_group_local_id(); )==" "\n" |
236 | R"==(int idK = get_global_id(2); )==" "\n" |
237 | R"==(int nku = get_global_size(2); )==" "\n" |
238 | R"==(int i0 = idM * 32; )==" "\n" |
239 | R"==(int j0 = idN; )==" "\n" |
240 | R"==(int irem = m - i0; )==" "\n" |
241 | R"==(int jrem = n - j0; )==" "\n" |
242 | R"==(if (irem < 0) irem = 0; )==" "\n" |
243 | R"==(if (jrem < 0) jrem = 0; )==" "\n" |
244 | R"==(int last_k_unroll = (idK == nku - 1); )==" "\n" |
245 | R"==(#ifdef WITH_K_UNROLL )==" "\n" |
246 | R"==(int k0 = idK * UNROLL_K; )==" "\n" |
247 | R"==(int kt = k - k0; )==" "\n" |
248 | R"==(if (kt < 0) kt = 0; )==" "\n" |
249 | R"==(if (kt > UNROLL_K) kt = UNROLL_K; )==" "\n" |
250 | R"==(A += offset_a + i0 + k0 * lda; )==" "\n" |
251 | R"==(B += offset_b + j0 * ldb + k0; )==" "\n" |
252 | R"==(k = kt; )==" "\n" |
253 | R"==(#else )==" "\n" |
254 | R"==(A += offset_a + i0; )==" "\n" |
255 | R"==(B += offset_b + j0 * ldb; )==" "\n" |
256 | R"==(#endif )==" "\n" |
257 | R"==(C += offset_c + i0 + j0 * ldc; )==" "\n" |
258 | R"==(global float *A_cols[4] = {A, A + lda, A + 2 * lda, A + 3 * lda}; )==" "\n" |
259 | R"==(int ldax4 = lda << 2; )==" "\n" |
260 | R"==(int ldbx4 = ldb << 2; )==" "\n" |
261 | R"==(for (int z = 0; z < 8; z++) )==" "\n" |
262 | R"==(c[z] = 0.f; )==" "\n" |
263 | R"==(#ifdef WITH_K_UNROLL )==" "\n" |
264 | R"==(flag += offset_f + idM; )==" "\n" |
265 | R"==(if (idK == 0 && lid == 0) *flag = 0; )==" "\n" |
266 | R"==(#endif )==" "\n" |
267 | R"==(#ifndef ALLOW_READ_OVERRUNS )==" "\n" |
268 | R"==(if (irem >= 32 && sub_group_broadcast(jrem, 0) >= 16) { )==" "\n" |
269 | R"==(#endif )==" "\n" |
270 | R"==(for (int h = 0; h < (k >> 2); h++) { )==" "\n" |
271 | R"==(for (int hh = 0; hh < 4; hh++) { )==" "\n" |
272 | R"==(a[hh] = as_float2( )==" "\n" |
273 | R"==(intel_sub_group_block_read2((global uint *)A_cols[hh])); )==" "\n" |
274 | R"==(A_cols[hh] += ldax4; )==" "\n" |
275 | R"==(} )==" "\n" |
276 | R"==(b = vload4(0, B); )==" "\n" |
277 | R"==(B += 4; )==" "\n" |
278 | R"==(FMA_I_LOOP_32_ROW(0); )==" "\n" |
279 | R"==(FMA_I_LOOP_32_ROW(1); )==" "\n" |
280 | R"==(FMA_I_LOOP_32_ROW(2); )==" "\n" |
281 | R"==(FMA_I_LOOP_32_ROW(3); )==" "\n" |
282 | R"==(} )==" "\n" |
283 | R"==(int krem = k & 3; )==" "\n" |
284 | R"==(for (int h = 0; h < krem; h++) { )==" "\n" |
285 | R"==(a[0] = as_float2( )==" "\n" |
286 | R"==(intel_sub_group_block_read2((global uint *)A_cols[0])); )==" "\n" |
287 | R"==(A_cols[0] += lda; )==" "\n" |
288 | R"==(b = *B++; )==" "\n" |
289 | R"==(FMA_I_LOOP_32_ROW(0); )==" "\n" |
290 | R"==(} )==" "\n" |
291 | R"==(#ifndef ALLOW_READ_OVERRUNS )==" "\n" |
292 | R"==(} else { )==" "\n" |
293 | R"==(for (int h = 0; h < (k >> 1); h++) { )==" "\n" |
294 | R"==(for (int hh = 0; hh < 2; hh++) { )==" "\n" |
295 | R"==(if (irem > lid) a[hh].s0 = A_cols[hh][lid]; )==" "\n" |
296 | R"==(if (irem > (lid + 16)) a[hh].s1 = A_cols[hh][lid + 16]; )==" "\n" |
297 | R"==(A_cols[hh] += (lda << 1); )==" "\n" |
298 | R"==(} )==" "\n" |
299 | R"==(if (jrem > 0) b.s01 = vload2(0, B); )==" "\n" |
300 | R"==(B += 2; )==" "\n" |
301 | R"==(FMA_I_LOOP_32_ROW(0); )==" "\n" |
302 | R"==(FMA_I_LOOP_32_ROW(1); )==" "\n" |
303 | R"==(} )==" "\n" |
304 | R"==(if (k & 1) { )==" "\n" |
305 | R"==(if (irem > lid) a[0].s0 = A_cols[0][lid]; )==" "\n" |
306 | R"==(if (irem > (lid + 16)) a[0].s1 = A_cols[0][lid + 16]; )==" "\n" |
307 | R"==(if (jrem > 0) b = *B; )==" "\n" |
308 | R"==(FMA_I_LOOP_32_ROW(0); )==" "\n" |
309 | R"==(} )==" "\n" |
310 | R"==(} )==" "\n" |
311 | R"==(#endif )==" "\n" |
312 | R"==(#ifdef WITH_K_UNROLL )==" "\n" |
313 | R"==(do { )==" "\n" |
314 | R"==(read_mem_fence(CLK_GLOBAL_MEM_FENCE); )==" "\n" |
315 | R"==(} while (*flag != idK); )==" "\n" |
316 | R"==(if (idK == 0) { )==" "\n" |
317 | R"==(if (beta == 0) )==" "\n" |
318 | R"==(UPDATE_C_32_ROW(1); )==" "\n" |
319 | R"==(else )==" "\n" |
320 | R"==(UPDATE_C_32_ROW(0); )==" "\n" |
321 | R"==(} else { )==" "\n" |
322 | R"==(beta = 1.0; )==" "\n" |
323 | R"==(UPDATE_C_32_ROW(0); )==" "\n" |
324 | R"==(} )==" "\n" |
325 | R"==(if (lid == 0) *flag = idK + 1; )==" "\n" |
326 | R"==(#else )==" "\n" |
327 | R"==(if (beta == 0) )==" "\n" |
328 | R"==(UPDATE_C_32_ROW(1); )==" "\n" |
329 | R"==(else )==" "\n" |
330 | R"==(UPDATE_C_32_ROW(0); )==" "\n" |
331 | R"==(#endif )==" "\n" |
332 | R"==(} )==" "\n" |
333 | R"==(#endif )==" "\n" |
334 | R"==(#ifdef NT )==" "\n" |
335 | R"==(__attribute__((intel_reqd_sub_group_size(16))) )==" "\n" |
336 | R"==(kernel void )==" "\n" |
337 | R"==(gen9_gemm_nocopy_f32(global float *A, global float *B, global float *C, )==" "\n" |
338 | R"==(long offset_a, long offset_b, long offset_c, int lda, int ldb, int ldc, )==" "\n" |
339 | R"==(int m, int n, int k, global float *alpha, float beta, int last_k_block, )==" "\n" |
340 | R"==(float eltwise_alpha, float eltwise_beta, float eltwise_scale )==" "\n" |
341 | R"==(#ifdef WITH_K_UNROLL )==" "\n" |
342 | R"==(, )==" "\n" |
343 | R"==(volatile global int *flag, long offset_f) { )==" "\n" |
344 | R"==(#else )==" "\n" |
345 | R"==() { )==" "\n" |
346 | R"==(#endif )==" "\n" |
347 | R"==(float2 a[2]; )==" "\n" |
348 | R"==(float b[2]; )==" "\n" |
349 | R"==(float4 c[8]; )==" "\n" |
350 | R"==(int idM = get_global_id(1); )==" "\n" |
351 | R"==(int idN = get_global_id(0); )==" "\n" |
352 | R"==(int lid = get_sub_group_local_id(); )==" "\n" |
353 | R"==(int idK = get_global_id(2); )==" "\n" |
354 | R"==(int nku = get_global_size(2); )==" "\n" |
355 | R"==(int i0 = idM * 32; )==" "\n" |
356 | R"==(int j0 = idN; )==" "\n" |
357 | R"==(int irem = m - i0; )==" "\n" |
358 | R"==(int jrem = n - j0; )==" "\n" |
359 | R"==(if (irem < 0) irem = 0; )==" "\n" |
360 | R"==(if (jrem < 0) jrem = 0; )==" "\n" |
361 | R"==(int last_k_unroll = (idK == nku - 1); )==" "\n" |
362 | R"==(#ifdef WITH_K_UNROLL )==" "\n" |
363 | R"==(int k0 = idK * UNROLL_K; )==" "\n" |
364 | R"==(int kt = k - k0; )==" "\n" |
365 | R"==(if (kt < 0) kt = 0; )==" "\n" |
366 | R"==(if (kt > UNROLL_K) kt = UNROLL_K; )==" "\n" |
367 | R"==(A += offset_a + i0 + k0 * lda; )==" "\n" |
368 | R"==(B += offset_b + sub_group_broadcast(j0, 0) + k0 * ldb; )==" "\n" |
369 | R"==(k = kt; )==" "\n" |
370 | R"==(#else )==" "\n" |
371 | R"==(A += offset_a + i0; )==" "\n" |
372 | R"==(B += offset_b + sub_group_broadcast(j0, 0); )==" "\n" |
373 | R"==(#endif )==" "\n" |
374 | R"==(C += offset_c + i0 + j0 * ldc; )==" "\n" |
375 | R"==(global float *A_cols[2] = {A, A + lda}; )==" "\n" |
376 | R"==(global float *B_rows[2] = {B, B + ldb}; )==" "\n" |
377 | R"==(int ldax2 = lda << 1; )==" "\n" |
378 | R"==(int ldbx2 = ldb << 1; )==" "\n" |
379 | R"==(for (int z = 0; z < 8; z++) )==" "\n" |
380 | R"==(c[z] = 0.f; )==" "\n" |
381 | R"==(#ifdef WITH_K_UNROLL )==" "\n" |
382 | R"==(flag += offset_f + idM; )==" "\n" |
383 | R"==(if (idK == 0 && lid == 0) *flag = 0; )==" "\n" |
384 | R"==(#endif )==" "\n" |
385 | R"==(#ifndef ALLOW_READ_OVERRUNS )==" "\n" |
386 | R"==(if (irem >= 32 && sub_group_broadcast(jrem, 0) >= 16) { )==" "\n" |
387 | R"==(#endif )==" "\n" |
388 | R"==(for (int h = 0; h < (k >> 1); h++) { )==" "\n" |
389 | R"==(for (int hh = 0; hh < 2; hh++) { )==" "\n" |
390 | R"==(a[hh] = as_float2( )==" "\n" |
391 | R"==(intel_sub_group_block_read2((global uint *)A_cols[hh])); )==" "\n" |
392 | R"==(A_cols[hh] += ldax2; )==" "\n" |
393 | R"==(} )==" "\n" |
394 | R"==(for (int hh = 0; hh < 2; hh++) { )==" "\n" |
395 | R"==(b[hh] = as_float( )==" "\n" |
396 | R"==(intel_sub_group_block_read((global uint *)B_rows[hh])); )==" "\n" |
397 | R"==(B_rows[hh] += ldbx2; )==" "\n" |
398 | R"==(} )==" "\n" |
399 | R"==(FMA_I_LOOP_32_ROW(0); )==" "\n" |
400 | R"==(FMA_I_LOOP_32_ROW(1); )==" "\n" |
401 | R"==(} )==" "\n" |
402 | R"==(int krem = k & 1; )==" "\n" |
403 | R"==(if (krem > 0) { )==" "\n" |
404 | R"==(a[0] = as_float2( )==" "\n" |
405 | R"==(intel_sub_group_block_read2((global uint *)A_cols[0])); )==" "\n" |
406 | R"==(b[0] = as_float( )==" "\n" |
407 | R"==(intel_sub_group_block_read((global uint *)B_rows[0])); )==" "\n" |
408 | R"==(FMA_I_LOOP_32_ROW(0); )==" "\n" |
409 | R"==(} )==" "\n" |
410 | R"==(#ifndef ALLOW_READ_OVERRUNS )==" "\n" |
411 | R"==(} else { )==" "\n" |
412 | R"==(for (int h = 0; h < (k >> 1); h++) { )==" "\n" |
413 | R"==(for (int hh = 0; hh < 2; hh++) { )==" "\n" |
414 | R"==(if (irem > lid) a[hh].s0 = A_cols[hh][lid]; )==" "\n" |
415 | R"==(if (irem > (lid + 16)) a[hh].s1 = A_cols[hh][lid + 16]; )==" "\n" |
416 | R"==(A_cols[hh] += ldax2; )==" "\n" |
417 | R"==(} )==" "\n" |
418 | R"==(for (int hh = 0; hh < 2; hh++) { )==" "\n" |
419 | R"==(if (jrem > 0) b[hh] = B_rows[hh][lid]; )==" "\n" |
420 | R"==(B_rows[hh] += ldbx2; )==" "\n" |
421 | R"==(} )==" "\n" |
422 | R"==(FMA_I_LOOP_32_ROW(0); )==" "\n" |
423 | R"==(FMA_I_LOOP_32_ROW(1); )==" "\n" |
424 | R"==(} )==" "\n" |
425 | R"==(int krem = k & 1; )==" "\n" |
426 | R"==(if (krem > 0) { )==" "\n" |
427 | R"==(if (irem > lid) a[0].s0 = A_cols[0][lid]; )==" "\n" |
428 | R"==(if (irem > (lid + 16)) a[0].s1 = A_cols[0][lid + 16]; )==" "\n" |
429 | R"==(if (jrem > 0) b[0] = B_rows[0][lid]; )==" "\n" |
430 | R"==(FMA_I_LOOP_32_ROW(0); )==" "\n" |
431 | R"==(} )==" "\n" |
432 | R"==(} )==" "\n" |
433 | R"==(#endif )==" "\n" |
434 | R"==(#ifdef WITH_K_UNROLL )==" "\n" |
435 | R"==(do { )==" "\n" |
436 | R"==(read_mem_fence(CLK_GLOBAL_MEM_FENCE); )==" "\n" |
437 | R"==(} while (*flag != idK); )==" "\n" |
438 | R"==(if (idK == 0) { )==" "\n" |
439 | R"==(if (beta == 0) )==" "\n" |
440 | R"==(UPDATE_C_32_ROW(1); )==" "\n" |
441 | R"==(else )==" "\n" |
442 | R"==(UPDATE_C_32_ROW(0); )==" "\n" |
443 | R"==(} else { )==" "\n" |
444 | R"==(beta = 1.0; )==" "\n" |
445 | R"==(UPDATE_C_32_ROW(0); )==" "\n" |
446 | R"==(} )==" "\n" |
447 | R"==(if (lid == 0) *flag = idK + 1; )==" "\n" |
448 | R"==(#else )==" "\n" |
449 | R"==(if (beta == 0) )==" "\n" |
450 | R"==(UPDATE_C_32_ROW(1); )==" "\n" |
451 | R"==(else )==" "\n" |
452 | R"==(UPDATE_C_32_ROW(0); )==" "\n" |
453 | R"==(#endif )==" "\n" |
454 | R"==(} )==" "\n" |
455 | R"==(#endif )==" "\n" |
456 | R"==(#ifdef TN )==" "\n" |
457 | R"==(__attribute__((intel_reqd_sub_group_size(16))) )==" "\n" |
458 | R"==(kernel void )==" "\n" |
459 | R"==(gen9_gemm_nocopy_f32(global float *A, global float *B, global float *C, )==" "\n" |
460 | R"==(long offset_a, long offset_b, long offset_c, int lda, int ldb, int ldc, )==" "\n" |
461 | R"==(int m, int n, int k, global float *alpha, float beta, int last_k_block, )==" "\n" |
462 | R"==(float eltwise_alpha, float eltwise_beta, float eltwise_scale )==" "\n" |
463 | R"==(#ifdef WITH_K_UNROLL )==" "\n" |
464 | R"==(, )==" "\n" |
465 | R"==(volatile global int *flag, long offset_f) { )==" "\n" |
466 | R"==(#else )==" "\n" |
467 | R"==() { )==" "\n" |
468 | R"==(#endif )==" "\n" |
469 | R"==(float4 a; )==" "\n" |
470 | R"==(float4 b[2]; )==" "\n" |
471 | R"==(float4 c[4][2]; )==" "\n" |
472 | R"==(int idM = get_global_id(1); )==" "\n" |
473 | R"==(int idN = get_global_id(0); )==" "\n" |
474 | R"==(int lid = get_sub_group_local_id(); )==" "\n" |
475 | R"==(int idK = get_global_id(2); )==" "\n" |
476 | R"==(int nku = get_global_size(2); )==" "\n" |
477 | R"==(int i0 = idM * 16; )==" "\n" |
478 | R"==(int j0 = sub_group_broadcast(idN, 0) * 2 + lid; )==" "\n" |
479 | R"==(int irem = m - i0; )==" "\n" |
480 | R"==(int jrem = n - j0; )==" "\n" |
481 | R"==(if (irem < 0) irem = 0; )==" "\n" |
482 | R"==(if (jrem < 0) jrem = 0; )==" "\n" |
483 | R"==(int last_k_unroll = (idK == nku - 1); )==" "\n" |
484 | R"==(#ifdef WITH_K_UNROLL )==" "\n" |
485 | R"==(int k0 = idK * UNROLL_K; )==" "\n" |
486 | R"==(int kt = k - k0; )==" "\n" |
487 | R"==(if (kt < 0) kt = 0; )==" "\n" |
488 | R"==(if (kt > UNROLL_K) kt = UNROLL_K; )==" "\n" |
489 | R"==(A += offset_a + (i0 + lid) * lda + k0; )==" "\n" |
490 | R"==(B += offset_b + j0 * ldb + k0; )==" "\n" |
491 | R"==(k = kt; )==" "\n" |
492 | R"==(#else )==" "\n" |
493 | R"==(A += offset_a + (i0 + lid) * lda; )==" "\n" |
494 | R"==(B += offset_b + j0 * ldb; )==" "\n" |
495 | R"==(#endif )==" "\n" |
496 | R"==(C += offset_c + i0 + j0 * ldc; )==" "\n" |
497 | R"==(global float *B_ptrs[2] = {B, B + 16 * ldb}; )==" "\n" |
498 | R"==(for (int ii = 0; ii < 4; ii++) )==" "\n" |
499 | R"==(for (int jj = 0; jj < 2; jj++) )==" "\n" |
500 | R"==(c[ii][jj] = 0.f; )==" "\n" |
501 | R"==(#ifdef WITH_K_UNROLL )==" "\n" |
502 | R"==(flag += offset_f + idM; )==" "\n" |
503 | R"==(if (idK == 0 && lid == 0) *flag = 0; )==" "\n" |
504 | R"==(#endif )==" "\n" |
505 | R"==(for (int h = 0; h < (k >> 2); h++) { )==" "\n" |
506 | R"==(if (irem > lid) a = vload4(0, A); )==" "\n" |
507 | R"==(A += 4; )==" "\n" |
508 | R"==(for (int hh = 0; hh < 2; hh++) { )==" "\n" |
509 | R"==(if (jrem > hh * 16) b[hh] = vload4(0, B_ptrs[hh]); )==" "\n" |
510 | R"==(B_ptrs[hh] += 4; )==" "\n" |
511 | R"==(} )==" "\n" |
512 | R"==(FMA_I_LOOP_16_ROW(0); )==" "\n" |
513 | R"==(FMA_I_LOOP_16_ROW(1); )==" "\n" |
514 | R"==(FMA_I_LOOP_16_ROW(2); )==" "\n" |
515 | R"==(FMA_I_LOOP_16_ROW(3); )==" "\n" |
516 | R"==(} )==" "\n" |
517 | R"==(int krem = k & 3; )==" "\n" |
518 | R"==(for (int h = 0; h < krem; h++) { )==" "\n" |
519 | R"==(if (irem > lid) a = *A++; )==" "\n" |
520 | R"==(for (int hh = 0; hh < 2; hh++) { )==" "\n" |
521 | R"==(if (jrem > hh * 16) b[hh] = *B_ptrs[hh]; )==" "\n" |
522 | R"==(B_ptrs[hh]++; )==" "\n" |
523 | R"==(} )==" "\n" |
524 | R"==(FMA_I_LOOP_16_ROW(0); )==" "\n" |
525 | R"==(} )==" "\n" |
526 | R"==(global float *C_ptrs[2] = {C, C + 16 * ldc}; )==" "\n" |
527 | R"==(#ifdef WITH_K_UNROLL )==" "\n" |
528 | R"==(do { )==" "\n" |
529 | R"==(read_mem_fence(CLK_GLOBAL_MEM_FENCE); )==" "\n" |
530 | R"==(} while (*flag != idK); )==" "\n" |
531 | R"==(if (idK == 0) { )==" "\n" |
532 | R"==(if (beta == 0) )==" "\n" |
533 | R"==(UPDATE_C_16_ROW(1); )==" "\n" |
534 | R"==(else )==" "\n" |
535 | R"==(UPDATE_C_16_ROW(0); )==" "\n" |
536 | R"==(} else { )==" "\n" |
537 | R"==(beta = 1.0; )==" "\n" |
538 | R"==(UPDATE_C_16_ROW(0); )==" "\n" |
539 | R"==(} )==" "\n" |
540 | R"==(if (lid == 0) *flag = idK + 1; )==" "\n" |
541 | R"==(#else )==" "\n" |
542 | R"==(if (beta == 0) )==" "\n" |
543 | R"==(UPDATE_C_16_ROW(1); )==" "\n" |
544 | R"==(else )==" "\n" |
545 | R"==(UPDATE_C_16_ROW(0); )==" "\n" |
546 | R"==(#endif )==" "\n" |
547 | R"==(} )==" "\n" |
548 | R"==(#endif )==" "\n" |
549 | R"==(#ifdef TT )==" "\n" |
550 | R"==(__attribute__((intel_reqd_sub_group_size(16))) )==" "\n" |
551 | R"==(kernel void )==" "\n" |
552 | R"==(gen9_gemm_nocopy_f32(global float *A, global float *B, global float *C, )==" "\n" |
553 | R"==(long offset_a, long offset_b, long offset_c, int lda, int ldb, int ldc, )==" "\n" |
554 | R"==(int m, int n, int k, global float *alpha, float beta, int last_k_block, )==" "\n" |
555 | R"==(float eltwise_alpha, float eltwise_beta, float eltwise_scale )==" "\n" |
556 | R"==(#ifdef WITH_K_UNROLL )==" "\n" |
557 | R"==(, )==" "\n" |
558 | R"==(volatile global int *flag, long offset_f) { )==" "\n" |
559 | R"==(#else )==" "\n" |
560 | R"==() { )==" "\n" |
561 | R"==(#endif )==" "\n" |
562 | R"==(float4 a; )==" "\n" |
563 | R"==(float2 b[4]; )==" "\n" |
564 | R"==(float4 c[4][2]; )==" "\n" |
565 | R"==(int idM = get_global_id(1); )==" "\n" |
566 | R"==(int idN = get_global_id(0); )==" "\n" |
567 | R"==(int lid = get_sub_group_local_id(); )==" "\n" |
568 | R"==(int idK = get_global_id(2); )==" "\n" |
569 | R"==(int nku = get_global_size(2); )==" "\n" |
570 | R"==(int i0 = idM * 16; )==" "\n" |
571 | R"==(int j0 = sub_group_broadcast(idN, 0) * 2 + lid; )==" "\n" |
572 | R"==(int irem = m - i0; )==" "\n" |
573 | R"==(int jrem = n - j0; )==" "\n" |
574 | R"==(if (irem < 0) irem = 0; )==" "\n" |
575 | R"==(if (jrem < 0) jrem = 0; )==" "\n" |
576 | R"==(int last_k_unroll = (idK == nku - 1); )==" "\n" |
577 | R"==(#ifdef WITH_K_UNROLL )==" "\n" |
578 | R"==(int k0 = idK * UNROLL_K; )==" "\n" |
579 | R"==(int kt = k - k0; )==" "\n" |
580 | R"==(if (kt < 0) kt = 0; )==" "\n" |
581 | R"==(if (kt > UNROLL_K) kt = UNROLL_K; )==" "\n" |
582 | R"==(A += offset_a + (i0 + lid) * lda + k0; )==" "\n" |
583 | R"==(B += offset_b + sub_group_broadcast(j0, 0) + k0 * ldb; )==" "\n" |
584 | R"==(k = kt; )==" "\n" |
585 | R"==(#else )==" "\n" |
586 | R"==(A += offset_a + (i0 + lid) * lda; )==" "\n" |
587 | R"==(B += offset_b + sub_group_broadcast(j0, 0); )==" "\n" |
588 | R"==(#endif )==" "\n" |
589 | R"==(C += offset_c + i0 + j0 * ldc; )==" "\n" |
590 | R"==(global float *B_rows[4] = {B, B + ldb, B + 2 * ldb, B + 3 * ldb}; )==" "\n" |
591 | R"==(int ldbx4 = ldb << 2; )==" "\n" |
592 | R"==(for (int ii = 0; ii < 4; ii++) )==" "\n" |
593 | R"==(for (int jj = 0; jj < 2; jj++) )==" "\n" |
594 | R"==(c[ii][jj] = 0.f; )==" "\n" |
595 | R"==(#ifdef WITH_K_UNROLL )==" "\n" |
596 | R"==(flag += offset_f + idM; )==" "\n" |
597 | R"==(if (idK == 0 && lid == 0) *flag = 0; )==" "\n" |
598 | R"==(#endif )==" "\n" |
599 | R"==(#ifndef ALLOW_READ_OVERRUNS )==" "\n" |
600 | R"==(if (irem >= 16 && sub_group_broadcast(jrem, 0) >= 32) { )==" "\n" |
601 | R"==(#endif )==" "\n" |
602 | R"==(for (int h = 0; h < (k >> 2); h++) { )==" "\n" |
603 | R"==(a = vload4(0, A); )==" "\n" |
604 | R"==(A += 4; )==" "\n" |
605 | R"==(for (int hh = 0; hh < 4; hh++) { )==" "\n" |
606 | R"==(b[hh] = as_float2( )==" "\n" |
607 | R"==(intel_sub_group_block_read2((global uint *)B_rows[hh])); )==" "\n" |
608 | R"==(B_rows[hh] += ldbx4; )==" "\n" |
609 | R"==(} )==" "\n" |
610 | R"==(FMA_I_LOOP_16_ROW(0); )==" "\n" |
611 | R"==(FMA_I_LOOP_16_ROW(1); )==" "\n" |
612 | R"==(FMA_I_LOOP_16_ROW(2); )==" "\n" |
613 | R"==(FMA_I_LOOP_16_ROW(3); )==" "\n" |
614 | R"==(} )==" "\n" |
615 | R"==(int krem = k & 3; )==" "\n" |
616 | R"==(for (int h = 0; h < krem; h++) { )==" "\n" |
617 | R"==(a = *A++; )==" "\n" |
618 | R"==(b[0] = as_float2( )==" "\n" |
619 | R"==(intel_sub_group_block_read2((global uint *)B_rows[0])); )==" "\n" |
620 | R"==(B_rows[0] += ldb; )==" "\n" |
621 | R"==(FMA_I_LOOP_16_ROW(0); )==" "\n" |
622 | R"==(} )==" "\n" |
623 | R"==(#ifndef ALLOW_READ_OVERRUNS )==" "\n" |
624 | R"==(} else { )==" "\n" |
625 | R"==(for (int h = 0; h < (k >> 2); h++) { )==" "\n" |
626 | R"==(if (irem > lid) a = vload4(0, A); )==" "\n" |
627 | R"==(A += 4; )==" "\n" |
628 | R"==(for (int hh = 0; hh < 4; hh++) { )==" "\n" |
629 | R"==(if (jrem > 0) b[hh].s0 = B_rows[hh][lid]; )==" "\n" |
630 | R"==(if (jrem > 16) b[hh].s1 = B_rows[hh][lid + 16]; )==" "\n" |
631 | R"==(B_rows[hh] += ldbx4; )==" "\n" |
632 | R"==(} )==" "\n" |
633 | R"==(FMA_I_LOOP_16_ROW(0); )==" "\n" |
634 | R"==(FMA_I_LOOP_16_ROW(1); )==" "\n" |
635 | R"==(FMA_I_LOOP_16_ROW(2); )==" "\n" |
636 | R"==(FMA_I_LOOP_16_ROW(3); )==" "\n" |
637 | R"==(} )==" "\n" |
638 | R"==(int krem = k & 3; )==" "\n" |
639 | R"==(for (int h = 0; h < krem; h++) { )==" "\n" |
640 | R"==(if (irem > lid) a = *A++; )==" "\n" |
641 | R"==(if (jrem > 0) b[0].s0 = B_rows[0][lid]; )==" "\n" |
642 | R"==(if (jrem > 16) b[0].s1 = B_rows[0][lid + 16]; )==" "\n" |
643 | R"==(B_rows[0] += ldb; )==" "\n" |
644 | R"==(FMA_I_LOOP_16_ROW(0); )==" "\n" |
645 | R"==(} )==" "\n" |
646 | R"==(} )==" "\n" |
647 | R"==(#endif )==" "\n" |
648 | R"==(global float *C_ptrs[2] = {C, C + 16 * ldc}; )==" "\n" |
649 | R"==(#ifdef WITH_K_UNROLL )==" "\n" |
650 | R"==(do { )==" "\n" |
651 | R"==(read_mem_fence(CLK_GLOBAL_MEM_FENCE); )==" "\n" |
652 | R"==(} while (*flag != idK); )==" "\n" |
653 | R"==(if (idK == 0) { )==" "\n" |
654 | R"==(if (beta == 0) )==" "\n" |
655 | R"==(UPDATE_C_16_ROW(1); )==" "\n" |
656 | R"==(else )==" "\n" |
657 | R"==(UPDATE_C_16_ROW(0); )==" "\n" |
658 | R"==(} else { )==" "\n" |
659 | R"==(beta = 1.0; )==" "\n" |
660 | R"==(UPDATE_C_16_ROW(0); )==" "\n" |
661 | R"==(} )==" "\n" |
662 | R"==(if (lid == 0) *flag = idK + 1; )==" "\n" |
663 | R"==(#else )==" "\n" |
664 | R"==(if (beta == 0) )==" "\n" |
665 | R"==(UPDATE_C_16_ROW(1); )==" "\n" |
666 | R"==(else )==" "\n" |
667 | R"==(UPDATE_C_16_ROW(0); )==" "\n" |
668 | R"==(#endif )==" "\n" |
669 | R"==(} )==" "\n" |
670 | R"==(#endif )==" "\n" |
671 | R"==()==" ; |
672 | } |
673 | } |
674 | } |
675 | } |