1 | namespace dnnl { |
2 | namespace impl { |
3 | namespace gpu { |
4 | namespace ocl { |
5 | const char *gen9_reduction_kernel = R"==(/******************************************************************************* )==" "\n" |
6 | R"==(* Copyright 2021-2022 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_post_ops.h" )==" "\n" |
21 | R"==(#include "gpu/ocl/ocl_types.h" )==" "\n" |
22 | R"==(#if defined(IS_MAX) )==" "\n" |
23 | R"==(#define INIT_ACC TO_DEF_ACC_DATA_T(DATA_MIN) )==" "\n" |
24 | R"==(#elif defined(IS_MIN) )==" "\n" |
25 | R"==(#define INIT_ACC TO_DEF_ACC_DATA_T(DATA_MAX) )==" "\n" |
26 | R"==(#elif defined(IS_MUL) )==" "\n" |
27 | R"==(#define INIT_ACC TO_DEF_ACC_DATA_T(DATA_ONE) )==" "\n" |
28 | R"==(#else )==" "\n" |
29 | R"==(#define INIT_ACC TO_DEF_ACC_DATA_T(DATA_ZERO) )==" "\n" |
30 | R"==(#endif )==" "\n" |
31 | R"==(#if defined(IS_MAX) )==" "\n" |
32 | R"==(#if defined(SRC_DT_S8) || defined(SRC_DT_U8) )==" "\n" |
33 | R"==(#define ACCUMULATE(x, y) max(x, y) )==" "\n" |
34 | R"==(#else )==" "\n" |
35 | R"==(#define ACCUMULATE(x, y) fmax(x, y) )==" "\n" |
36 | R"==(#endif )==" "\n" |
37 | R"==(#elif defined(IS_MIN) )==" "\n" |
38 | R"==(#if defined(SRC_DT_S8) || defined(SRC_DT_U8) )==" "\n" |
39 | R"==(#define ACCUMULATE(x, y) min(x, y) )==" "\n" |
40 | R"==(#else )==" "\n" |
41 | R"==(#define ACCUMULATE(x, y) fmin(x, y) )==" "\n" |
42 | R"==(#endif )==" "\n" |
43 | R"==(#elif defined(IS_MEAN) || defined(IS_SUM) )==" "\n" |
44 | R"==(#define ACCUMULATE(x, y) (x + y) )==" "\n" |
45 | R"==(#elif defined(IS_MUL) )==" "\n" |
46 | R"==(#define ACCUMULATE(x, y) (x * y) )==" "\n" |
47 | R"==(#else )==" "\n" |
48 | R"==(#define ACCUMULATE(x, y) (x + pow(fabs(y), POWER)) )==" "\n" |
49 | R"==(#endif )==" "\n" |
50 | R"==(#if defined(IS_MAX) || defined(IS_MIN) || defined(IS_MUL) )==" "\n" |
51 | R"==(#define ACCUMULATE_AGAIN(x, y) ACCUMULATE(x, y) )==" "\n" |
52 | R"==(#else )==" "\n" |
53 | R"==(#define ACCUMULATE_AGAIN(x, y) (x + y) )==" "\n" |
54 | R"==(#endif )==" "\n" |
55 | R"==(#if defined(IS_MEAN) )==" "\n" |
56 | R"==(#define FINALIZE_REDUCTION(x) (x / REDUCTION_SIZE) )==" "\n" |
57 | R"==(#elif defined(IS_LP_MAX) )==" "\n" |
58 | R"==(#define FINALIZE_REDUCTION(x) rootn(fmax(x, EPS), POWER) )==" "\n" |
59 | R"==(#elif defined(IS_LP_SUM) )==" "\n" |
60 | R"==(#define FINALIZE_REDUCTION(x) rootn(x + EPS, POWER) )==" "\n" |
61 | R"==(#elif defined(IS_P_MAX) )==" "\n" |
62 | R"==(#define FINALIZE_REDUCTION(x) fmax(x, EPS) )==" "\n" |
63 | R"==(#elif defined(IS_P_SUM) )==" "\n" |
64 | R"==(#define FINALIZE_REDUCTION(x) (x + EPS) )==" "\n" |
65 | R"==(#else )==" "\n" |
66 | R"==(#define FINALIZE_REDUCTION(x) (x) )==" "\n" |
67 | R"==(#endif )==" "\n" |
68 | R"==(#if WITH_SUM )==" "\n" |
69 | R"==(#define INIT_SUM(sum_data) DST_TO_REF(sum_data) )==" "\n" |
70 | R"==(#else )==" "\n" |
71 | R"==(#define INIT_SUM(sum_data) 0.0f )==" "\n" |
72 | R"==(#endif )==" "\n" |
73 | R"==(#if defined(IS_MAX) )==" "\n" |
74 | R"==(#define SUB_GROUP_REDUCE(x, c_block) sub_group_reduce_max(x) )==" "\n" |
75 | R"==(#elif defined(IS_MIN) )==" "\n" |
76 | R"==(#define SUB_GROUP_REDUCE(x, c_block) sub_group_reduce_min(x) )==" "\n" |
77 | R"==(#elif defined(IS_MUL) )==" "\n" |
78 | R"==(#define SUB_GROUP_REDUCE(x, c_block) \ )==" "\n" |
79 | R"==(({ \ )==" "\n" |
80 | R"==(int cid_end \ )==" "\n" |
81 | R"==(= (INITIAL_C % SUB_GROUP_SIZE == 0 ? SUB_GROUP_SIZE \ )==" "\n" |
82 | R"==(: (INITIAL_C - c_block)); \ )==" "\n" |
83 | R"==(DEF_ACC_DATA_T sub_group_acc = 1.0; \ )==" "\n" |
84 | R"==(for (int channel_id = 0; channel_id < cid_end; channel_id++) { \ )==" "\n" |
85 | R"==(sub_group_acc *= intel_sub_group_shuffle(c_acc, channel_id); \ )==" "\n" |
86 | R"==(} \ )==" "\n" |
87 | R"==(sub_group_acc; \ )==" "\n" |
88 | R"==(}) )==" "\n" |
89 | R"==(#else )==" "\n" |
90 | R"==(#define SUB_GROUP_REDUCE(x, c_block) sub_group_reduce_add(x) )==" "\n" |
91 | R"==(#endif )==" "\n" |
92 | R"==(#if INITIAL_C_CHUNKS == 1 )==" "\n" |
93 | R"==(#define C_BLOCK_READ BLOCK_READ )==" "\n" |
94 | R"==(#define AS_C_BLOCK_DATA_T AS_DATA_T )==" "\n" |
95 | R"==(#define CONVERT_C_BLOCK_T TO_DEF_ACC_DATA_T )==" "\n" |
96 | R"==(#define C_BLOCK_T DEF_ACC_DATA_T )==" "\n" |
97 | R"==(#elif INITIAL_C_CHUNKS == 2 )==" "\n" |
98 | R"==(#define C_BLOCK_READ BLOCK_READ2 )==" "\n" |
99 | R"==(#define AS_C_BLOCK_DATA_T AS_DATA2_T )==" "\n" |
100 | R"==(#define CONVERT_C_BLOCK_T TO_DEF_ACC_DATA2_T )==" "\n" |
101 | R"==(#define C_BLOCK_T DEF_ACC_DATA2_T )==" "\n" |
102 | R"==(#elif INITIAL_C_CHUNKS == 4 )==" "\n" |
103 | R"==(#define C_BLOCK_READ BLOCK_READ4 )==" "\n" |
104 | R"==(#define AS_C_BLOCK_DATA_T AS_DATA4_T )==" "\n" |
105 | R"==(#define CONVERT_C_BLOCK_T TO_DEF_ACC_DATA4_T )==" "\n" |
106 | R"==(#define C_BLOCK_T DEF_ACC_DATA4_T )==" "\n" |
107 | R"==(#elif INITIAL_C_CHUNKS == 8 )==" "\n" |
108 | R"==(#define C_BLOCK_READ BLOCK_READ8 )==" "\n" |
109 | R"==(#define AS_C_BLOCK_DATA_T AS_DATA8_T )==" "\n" |
110 | R"==(#define CONVERT_C_BLOCK_T TO_DEF_ACC_DATA8_T )==" "\n" |
111 | R"==(#define C_BLOCK_T DEF_ACC_DATA8_T )==" "\n" |
112 | R"==(#endif )==" "\n" |
113 | R"==(#define ROUND_DOWN(a, b) ((a) - ((a) % (b))) )==" "\n" |
114 | R"==(#undef ROUND_UP )==" "\n" |
115 | R"==(#define ROUND_UP(a, b) ROUND_DOWN((a + b - 1), (b)) )==" "\n" |
116 | R"==(#define INITIAL_SRC_OFFSET(n, c, hwd) \ )==" "\n" |
117 | R"==((((n) / N_BLOCK_SIZE) * INITIAL_HWD_DIM * N_BLOCK_SIZE * ROUND_UP(INITIAL_C, C_BLOCK_SIZE) + \ )==" "\n" |
118 | R"==(((c) / C_BLOCK_SIZE) * INITIAL_HWD_DIM * N_BLOCK_SIZE * C_BLOCK_SIZE + \ )==" "\n" |
119 | R"==((hwd) * N_BLOCK_SIZE * C_BLOCK_SIZE + \ )==" "\n" |
120 | R"==(((n) % N_BLOCK_SIZE) * C_BLOCK_SIZE + \ )==" "\n" |
121 | R"==(((c) % C_BLOCK_SIZE)) )==" "\n" |
122 | R"==(#define INITIAL_DST_OFFSET(n, c, hwd) \ )==" "\n" |
123 | R"==(((n / N_BLOCK_SIZE) * FINAL_HWD_DIM * N_BLOCK_SIZE * ROUND_UP(FINAL_C_DIM, C_BLOCK_SIZE) + \ )==" "\n" |
124 | R"==(((c) / C_BLOCK_SIZE) * FINAL_HWD_DIM * N_BLOCK_SIZE * C_BLOCK_SIZE + \ )==" "\n" |
125 | R"==((hwd) * N_BLOCK_SIZE * C_BLOCK_SIZE + \ )==" "\n" |
126 | R"==(((n) % N_BLOCK_SIZE) * C_BLOCK_SIZE + \ )==" "\n" |
127 | R"==(((c) % C_BLOCK_SIZE)) )==" "\n" |
128 | R"==(#define FINAL_SRC_OFFSET(n, c, hwd) INITIAL_DST_OFFSET(n, c, hwd) )==" "\n" |
129 | R"==(#define FINAL_DST_OFFSET(n, c, hwd) \ )==" "\n" |
130 | R"==(((n) / N_BLOCK_SIZE) * (FINAL_HWD_DIM / FINAL_HWD_CHUNK_SIZE) * N_BLOCK_SIZE * ROUND_UP(FINAL_C_DIM / FINAL_C_CHUNK_SIZE, C_BLOCK_SIZE) + \ )==" "\n" |
131 | R"==(((c) / C_BLOCK_SIZE) * (FINAL_HWD_DIM / FINAL_HWD_CHUNK_SIZE) * N_BLOCK_SIZE * C_BLOCK_SIZE + \ )==" "\n" |
132 | R"==((hwd) * N_BLOCK_SIZE * C_BLOCK_SIZE + \ )==" "\n" |
133 | R"==(((n) % N_BLOCK_SIZE) * C_BLOCK_SIZE + \ )==" "\n" |
134 | R"==(((c) % C_BLOCK_SIZE) )==" "\n" |
135 | R"==(#if WITH_POST_OP )==" "\n" |
136 | R"==(#define APPLY_POST_OPS(sum_data, data, n_idx, c_idx, hwd_idx) \ )==" "\n" |
137 | R"==({ \ )==" "\n" |
138 | R"==(float sum_init_val = INIT_SUM(sum_data); \ )==" "\n" |
139 | R"==(const int D = hwd_start / (DST_H_DIM * DST_W_DIM); \ )==" "\n" |
140 | R"==(const int H = (hwd_start % (DST_H_DIM * DST_W_DIM)) / DST_W_DIM; \ )==" "\n" |
141 | R"==(const int W = hwd_start % DST_W_DIM; \ )==" "\n" |
142 | R"==(APPLY_POST_OPS_SERIAL(data, float, sum_init_val, float, n_idx, 1, \ )==" "\n" |
143 | R"==(c_idx, 1, D, 1, H, 1, W, 1, 0, 1); \ )==" "\n" |
144 | R"==(} )==" "\n" |
145 | R"==(#else )==" "\n" |
146 | R"==(#define APPLY_POST_OPS(sum_data, data, n_idx, c_idx, hwd_idx) \ )==" "\n" |
147 | R"==({} )==" "\n" |
148 | R"==(#endif )==" "\n" |
149 | R"==(#define WRITE_FINAL_RESULT(dst_elem, acc, n_start, c_start, hwd_start) \ )==" "\n" |
150 | R"==({ \ )==" "\n" |
151 | R"==(float acc_float = FINALIZE_REDUCTION(convert_float(acc)); \ )==" "\n" |
152 | R"==(APPLY_POST_OPS(dst_elem, acc_float, n_start, c_start, hwd_start); \ )==" "\n" |
153 | R"==(dst_elem = TO_DST(acc_float); \ )==" "\n" |
154 | R"==(} )==" "\n" |
155 | R"==(#if SKIP_FINAL_PHASE )==" "\n" |
156 | R"==(#define WRITE_INITIAL_RESULT WRITE_FINAL_RESULT )==" "\n" |
157 | R"==(#define INITIAL_DST_DTYPE DST_DATA_T )==" "\n" |
158 | R"==(#else )==" "\n" |
159 | R"==(#define WRITE_INITIAL_RESULT(dst_elem, data, n_start, c_start, hwd_start) \ )==" "\n" |
160 | R"==({ dst_elem = data; } )==" "\n" |
161 | R"==(#define INITIAL_DST_DTYPE DEF_ACC_DATA_T )==" "\n" |
162 | R"==(#endif )==" "\n" |
163 | R"==(NAMED_KERNEL_ATTR(INITIAL) )==" "\n" |
164 | R"==(__kernel void gen9_initial_reduce(__global SRC_DATA_T *src, )==" "\n" |
165 | R"==(__global INITIAL_DST_DTYPE *dst )==" "\n" |
166 | R"==(#if SKIP_FINAL_PHASE )==" "\n" |
167 | R"==(POST_OP_ARGS )==" "\n" |
168 | R"==(#endif )==" "\n" |
169 | R"==() { )==" "\n" |
170 | R"==(const int n_chunk_idx = GWS_GET_INITIAL_N(); )==" "\n" |
171 | R"==(const int c = GWS_GET_INITIAL_C(); )==" "\n" |
172 | R"==(const int c_block_idx = c / C_BLOCK_SIZE; )==" "\n" |
173 | R"==(const int hwd_chunk_idx = GWS_GET_INITIAL_HWD_CHUNK_ID(); )==" "\n" |
174 | R"==(const int hwd_start = hwd_chunk_idx * INITIAL_HWD_CHUNK_SIZE; )==" "\n" |
175 | R"==(const int current_hwd_chunk = min(INITIAL_HWD_CHUNK_SIZE, )==" "\n" |
176 | R"==(INITIAL_HWD_DIM - hwd_chunk_idx * INITIAL_HWD_CHUNK_SIZE); )==" "\n" |
177 | R"==(const int aligned_hwd_chunk = ROUND_DOWN(current_hwd_chunk, VECT_DT_N); )==" "\n" |
178 | R"==(const int n_start = n_chunk_idx * INITIAL_N_CHUNK_SIZE; )==" "\n" |
179 | R"==(const int n_end = min(n_start + INITIAL_N_CHUNK_SIZE, INITIAL_N); )==" "\n" |
180 | R"==(#if SKIP_FINAL_PHASE )==" "\n" |
181 | R"==(for (int n_idx = n_start; n_idx < n_start + INITIAL_N_CHUNK_SIZE; n_idx++) { )==" "\n" |
182 | R"==(for (int c_idx = c; c_idx < c + INITIAL_C_CHUNKS * SUB_GROUP_SIZE; )==" "\n" |
183 | R"==(c_idx++) { )==" "\n" |
184 | R"==(if (n_idx >= DST_N && n_idx < DST_N_PADDED )==" "\n" |
185 | R"==(|| c_idx >= DST_C && c_idx < DST_C_PADDED) { )==" "\n" |
186 | R"==(for (int hwd_idx = hwd_start; )==" "\n" |
187 | R"==(#if IS_HWD_REDUCED )==" "\n" |
188 | R"==(hwd_idx < hwd_start + FINAL_HWD_CHUNK_SIZE; )==" "\n" |
189 | R"==(#else )==" "\n" |
190 | R"==(hwd_idx < hwd_start + INITIAL_HWD_CHUNK_SIZE; )==" "\n" |
191 | R"==(#endif )==" "\n" |
192 | R"==(hwd_idx++) { )==" "\n" |
193 | R"==(int n = (!IS_C_REDUCED && IS_N_REDUCED && NDIMS == 3 )==" "\n" |
194 | R"==(&& DST_N_PADDED == 1) )==" "\n" |
195 | R"==(? 0 )==" "\n" |
196 | R"==(: n_idx; )==" "\n" |
197 | R"==(const int dst_off = FINAL_DST_OFFSET(n, c_idx, hwd_idx); )==" "\n" |
198 | R"==(dst[dst_off] = TO_DST(0.0f); )==" "\n" |
199 | R"==(} )==" "\n" |
200 | R"==(} )==" "\n" |
201 | R"==(} )==" "\n" |
202 | R"==(} )==" "\n" |
203 | R"==(#endif )==" "\n" |
204 | R"==(int channel_id = c + get_sub_group_local_id(); )==" "\n" |
205 | R"==(if (channel_id >= INITIAL_C || n_start >= INITIAL_N) { return; } )==" "\n" |
206 | R"==(VECT_DEF_ACC_DATA_T vector_acc = INIT_ACC; )==" "\n" |
207 | R"==(for (int n = n_start; n < n_end; n++) { )==" "\n" |
208 | R"==(for (int hwd_id = 0; hwd_id < aligned_hwd_chunk; hwd_id += VECT_DT_N) { )==" "\n" |
209 | R"==(for (int c_chunk = 0; c_chunk < INITIAL_C_CHUNKS; c_chunk++) { )==" "\n" |
210 | R"==(const int off = INITIAL_SRC_OFFSET(n, c, )==" "\n" |
211 | R"==(hwd_start + hwd_id )==" "\n" |
212 | R"==(+ c_chunk * VECT_DT_N / INITIAL_C_CHUNKS); )==" "\n" |
213 | R"==(VECT_DEF_ACC_DATA_T data )==" "\n" |
214 | R"==(= AS_VECT_DEF_ACC_DATA_T(AS_VECT_DATA_T(VECT_BLOCK_READ( )==" "\n" |
215 | R"==((const __global BLOCK_DATA_T *)&src[off]))); )==" "\n" |
216 | R"==(vector_acc = ACCUMULATE(vector_acc, data); )==" "\n" |
217 | R"==(} )==" "\n" |
218 | R"==(} )==" "\n" |
219 | R"==(for (int hwd_id = aligned_hwd_chunk; hwd_id < current_hwd_chunk; )==" "\n" |
220 | R"==(hwd_id++) { )==" "\n" |
221 | R"==(const int off = INITIAL_SRC_OFFSET(n, c, hwd_start + hwd_id); )==" "\n" |
222 | R"==(C_BLOCK_T data = CONVERT_C_BLOCK_T(AS_C_BLOCK_DATA_T( )==" "\n" |
223 | R"==(C_BLOCK_READ((const __global BLOCK_DATA_T *)&src[off]))); )==" "\n" |
224 | R"==(#if VECT_DT_N == 1 )==" "\n" |
225 | R"==(vector_acc = ACCUMULATE(vector_acc, data); )==" "\n" |
226 | R"==(#else )==" "\n" |
227 | R"==(#if INITIAL_C_CHUNKS == 1 )==" "\n" |
228 | R"==(vector_acc[0] = ACCUMULATE(vector_acc[0], data); )==" "\n" |
229 | R"==(#else )==" "\n" |
230 | R"==(for (int i = 0; i < INITIAL_C_CHUNKS; i++) { )==" "\n" |
231 | R"==(vector_acc[i] = ACCUMULATE(vector_acc[i], data[i]); )==" "\n" |
232 | R"==(} )==" "\n" |
233 | R"==(#endif )==" "\n" |
234 | R"==(#endif )==" "\n" |
235 | R"==(} )==" "\n" |
236 | R"==(} )==" "\n" |
237 | R"==(#if VECT_DT_N == 1 )==" "\n" |
238 | R"==(VECT_DEF_ACC_DATA_T acc = vector_acc; )==" "\n" |
239 | R"==(#else )==" "\n" |
240 | R"==(const int elems_to_accumulate = aligned_hwd_chunk > 0 ? VECT_DT_N : 1; )==" "\n" |
241 | R"==(#if INITIAL_C_CHUNKS == 1 )==" "\n" |
242 | R"==(DEF_ACC_DATA_T acc = INIT_ACC; )==" "\n" |
243 | R"==(for (int i = 0; i < elems_to_accumulate; i++) { )==" "\n" |
244 | R"==(acc = ACCUMULATE_AGAIN(acc, vector_acc[i]); )==" "\n" |
245 | R"==(} )==" "\n" |
246 | R"==(#else )==" "\n" |
247 | R"==(C_BLOCK_T acc = INIT_ACC; )==" "\n" |
248 | R"==(for (int i = 0; i < elems_to_accumulate; i += INITIAL_C_CHUNKS) { )==" "\n" |
249 | R"==(unroll_for(int j = 0; j < INITIAL_C_CHUNKS; j++) { )==" "\n" |
250 | R"==(acc[j] = ACCUMULATE_AGAIN(acc[j], vector_acc[i + j]); )==" "\n" |
251 | R"==(} )==" "\n" |
252 | R"==(} )==" "\n" |
253 | R"==(#endif )==" "\n" |
254 | R"==(#endif )==" "\n" |
255 | R"==(const int local_id = get_sub_group_local_id(); )==" "\n" |
256 | R"==(#if IS_C_REDUCED )==" "\n" |
257 | R"==(#if INITIAL_C_CHUNKS == 1 )==" "\n" |
258 | R"==(DEF_ACC_DATA_T c_acc = acc; )==" "\n" |
259 | R"==(#else )==" "\n" |
260 | R"==(DEF_ACC_DATA_T c_acc = acc[0]; )==" "\n" |
261 | R"==(for (int i = 1; i < INITIAL_C_CHUNKS; i++) { )==" "\n" |
262 | R"==(c_acc += acc[i]; )==" "\n" |
263 | R"==(} )==" "\n" |
264 | R"==(#endif )==" "\n" |
265 | R"==(const int dst_off )==" "\n" |
266 | R"==(= INITIAL_DST_OFFSET(n_chunk_idx, c_block_idx, hwd_chunk_idx); )==" "\n" |
267 | R"==(c_acc = SUB_GROUP_REDUCE(c_acc, c); )==" "\n" |
268 | R"==(if (local_id == 0) { )==" "\n" |
269 | R"==(WRITE_INITIAL_RESULT(dst[dst_off], c_acc, n_start, c, hwd_start); )==" "\n" |
270 | R"==(} )==" "\n" |
271 | R"==(#else )==" "\n" |
272 | R"==(const int dst_c = c + local_id; )==" "\n" |
273 | R"==(#if INITIAL_C_CHUNKS == 1 )==" "\n" |
274 | R"==(WRITE_INITIAL_RESULT( )==" "\n" |
275 | R"==(dst[INITIAL_DST_OFFSET(n_chunk_idx, dst_c, hwd_chunk_idx)], acc, )==" "\n" |
276 | R"==(n_start, dst_c, hwd_start); )==" "\n" |
277 | R"==(#else )==" "\n" |
278 | R"==(for (int i = 0; i < INITIAL_C_CHUNKS; i++) { )==" "\n" |
279 | R"==(int c_off = i * SUB_GROUP_SIZE; )==" "\n" |
280 | R"==(WRITE_INITIAL_RESULT(dst[INITIAL_DST_OFFSET(n_chunk_idx, dst_c + c_off, )==" "\n" |
281 | R"==(hwd_chunk_idx)], )==" "\n" |
282 | R"==(acc[i], n_start, c, hwd_start); )==" "\n" |
283 | R"==(} )==" "\n" |
284 | R"==(#endif )==" "\n" |
285 | R"==(#endif )==" "\n" |
286 | R"==(} )==" "\n" |
287 | R"==(#if !SKIP_FINAL_PHASE )==" "\n" |
288 | R"==(NAMED_KERNEL_ATTR(FINAL) )==" "\n" |
289 | R"==(__kernel void gen9_final_reduce( )==" "\n" |
290 | R"==(__global DEF_ACC_DATA_T *src, __global DST_DATA_T *dst POST_OP_ARGS) { )==" "\n" |
291 | R"==(const int n_start = GWS_GET_FINAL_N() * FINAL_N_CHUNK_SIZE; )==" "\n" |
292 | R"==(const int c_start = GWS_GET_FINAL_C() * FINAL_C_CHUNK_SIZE; )==" "\n" |
293 | R"==(const int hwd_start = GWS_GET_FINAL_HWD() * FINAL_HWD_CHUNK_SIZE; )==" "\n" |
294 | R"==(DEF_ACC_DATA_T acc = INIT_ACC; )==" "\n" |
295 | R"==(const int max_n = max(DST_N_PADDED, FINAL_N_DIM); )==" "\n" |
296 | R"==(const int max_c = max(DST_C_PADDED, FINAL_C_DIM); )==" "\n" |
297 | R"==(const int n_end = min(max_n, n_start + FINAL_N_CHUNK_SIZE); )==" "\n" |
298 | R"==(const int c_end = min(max_c, c_start + FINAL_C_CHUNK_SIZE); )==" "\n" |
299 | R"==(const int hwd_end = min(FINAL_HWD_DIM, hwd_start + FINAL_HWD_CHUNK_SIZE); )==" "\n" |
300 | R"==(for (int n = n_start; n < n_end; n++) { )==" "\n" |
301 | R"==(for (int c = c_start; c < c_end; c++) { )==" "\n" |
302 | R"==(for (int hwd = hwd_start; hwd < hwd_end; hwd++) { )==" "\n" |
303 | R"==(if ((n >= DST_N && n < DST_N_PADDED) )==" "\n" |
304 | R"==(|| (c >= DST_C && c < DST_C_PADDED)) { )==" "\n" |
305 | R"==(#if NDIMS == 2 && DST_N_PADDED == 1 )==" "\n" |
306 | R"==(const int dst_off = FINAL_DST_OFFSET(0, c, hwd); )==" "\n" |
307 | R"==(#elif NDIMS >= 3 && DST_N_PADDED == 1 && IS_HWD_REDUCED )==" "\n" |
308 | R"==(const int dst_off = FINAL_DST_OFFSET(0, c, 0); )==" "\n" |
309 | R"==(#elif IS_HWD_REDUCED )==" "\n" |
310 | R"==(const int dst_off = FINAL_DST_OFFSET(n, c, 0); )==" "\n" |
311 | R"==(#else )==" "\n" |
312 | R"==(const int dst_off = FINAL_DST_OFFSET(n, c, hwd); )==" "\n" |
313 | R"==(#endif )==" "\n" |
314 | R"==(dst[dst_off] = TO_DST(0.0f); )==" "\n" |
315 | R"==(} )==" "\n" |
316 | R"==(if (n < FINAL_N_DIM && c < FINAL_C_DIM) { )==" "\n" |
317 | R"==(const int off = FINAL_SRC_OFFSET(n, c, hwd); )==" "\n" |
318 | R"==(const DEF_ACC_DATA_T data = src[off]; )==" "\n" |
319 | R"==(acc = ACCUMULATE_AGAIN(acc, data); )==" "\n" |
320 | R"==(} )==" "\n" |
321 | R"==(} )==" "\n" |
322 | R"==(} )==" "\n" |
323 | R"==(} )==" "\n" |
324 | R"==(if (n_start < DST_N && c_start < DST_C) { )==" "\n" |
325 | R"==(const int off = FINAL_DST_OFFSET(n_start, c_start, hwd_start); )==" "\n" |
326 | R"==(WRITE_FINAL_RESULT(dst[off], acc, n_start, c_start, hwd_start); )==" "\n" |
327 | R"==(} )==" "\n" |
328 | R"==(} )==" "\n" |
329 | R"==(#endif )==" "\n" |
330 | R"==()==" ; |
331 | } |
332 | } |
333 | } |
334 | } |