1 | namespace dnnl { |
2 | namespace impl { |
3 | namespace gpu { |
4 | namespace ocl { |
5 | const char *gen9_conv_dw_fwd_data_kernel = R"==(/******************************************************************************* )==" "\n" |
6 | R"==(* Copyright 2019-2021 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 IS_DW != 1 )==" "\n" |
23 | R"==(#error "Kernel supports depth-wise convolutions only" )==" "\n" |
24 | R"==(#endif )==" "\n" |
25 | R"==(#ifdef DST_DT_S8 )==" "\n" |
26 | R"==(#if VER_32MB16C )==" "\n" |
27 | R"==(#define DST_MB_BLOCK MB_BLOCK )==" "\n" |
28 | R"==(#else )==" "\n" |
29 | R"==(#define DST_MB_BLOCK (MB_BLOCK * 2) )==" "\n" |
30 | R"==(#endif )==" "\n" |
31 | R"==(#define DST_OC_BLOCK (OC_BLOCK * 2) )==" "\n" |
32 | R"==(#endif )==" "\n" |
33 | R"==(#define APPLY_POST_OPS_COMMON(nelems, accumulator, dest_data, mb_shift) \ )==" "\n" |
34 | R"==({ \ )==" "\n" |
35 | R"==(const int po_mb = mb_shift + mb; \ )==" "\n" |
36 | R"==(const int po_oc = g; \ )==" "\n" |
37 | R"==(int po_mb_count; \ )==" "\n" |
38 | R"==(if (VER_16MB16C == 1) { \ )==" "\n" |
39 | R"==(po_mb_count = nelems; \ )==" "\n" |
40 | R"==(} else { \ )==" "\n" |
41 | R"==(po_mb_count = 1; \ )==" "\n" |
42 | R"==(} \ )==" "\n" |
43 | R"==(APPLY_POST_OPS_TRY_BURST(accumulator, DATA_T, dest_data, DATA_T, \ )==" "\n" |
44 | R"==(po_mb, po_mb_count, po_oc, SUB_GROUP_SIZE, \ )==" "\n" |
45 | R"==(get_sub_group_local_id()); \ )==" "\n" |
46 | R"==(} )==" "\n" |
47 | R"==(__attribute__((reqd_work_group_size(LWS_0, LWS_1, LWS_2))) )==" "\n" |
48 | R"==(#if SUB_GROUP_SIZE != 1 )==" "\n" |
49 | R"==(__attribute__((intel_reqd_sub_group_size(SUB_GROUP_SIZE))) )==" "\n" |
50 | R"==(#endif )==" "\n" |
51 | R"==(__kernel void )==" "\n" |
52 | R"==(gen9_conv_dw_fwd(const __global DATA_T *src, const __global DATA_T *wei, )==" "\n" |
53 | R"==(const __global DATA_T *bias, __global DST_DATA_T *dst POST_OP_ARGS) { )==" "\n" |
54 | R"==(MAYBE_SKIP_NON_UNIFORM_WG(); )==" "\n" |
55 | R"==(#if VER_8OW16C )==" "\n" |
56 | R"==(const int osp = get_global_id(1); )==" "\n" |
57 | R"==(const int od = osp / (OWB * OH); )==" "\n" |
58 | R"==(const int ohw = osp % (OWB * OH); )==" "\n" |
59 | R"==(const int ow = (ohw % OWB) * OW_BLOCK; )==" "\n" |
60 | R"==(const int oh = ohw / OWB; )==" "\n" |
61 | R"==(const int g )==" "\n" |
62 | R"==(= (get_group_id(0) * (LWS_0 / SUB_GROUP_SIZE) + get_sub_group_id()) )==" "\n" |
63 | R"==(* OC_BLOCK; )==" "\n" |
64 | R"==(const int mb = get_global_id(2) * MB_BLOCK; )==" "\n" |
65 | R"==(const int id = od * SD - PD; )==" "\n" |
66 | R"==(const int ih = oh * SH - PH; )==" "\n" |
67 | R"==(const int iw = ow * SW - PW; )==" "\n" |
68 | R"==(#ifdef DST_DT_S8 )==" "\n" |
69 | R"==(const int G_32block = G % 32 ? (32 + G - (G % 32)) : G; )==" "\n" |
70 | R"==(dst += mb * G_32block * OD * OH * OW )==" "\n" |
71 | R"==(+ (g / 32 * 32) * OD * OH * OW * MB_BLOCK )==" "\n" |
72 | R"==(+ (od * OH * OW + oh * OW + ow) * MB_BLOCK * (DST_OC_BLOCK) )==" "\n" |
73 | R"==(+ (g % 32); )==" "\n" |
74 | R"==(#else )==" "\n" |
75 | R"==(dst += mb * G * OD * OH * OW + g * OD * OH * OW * MB_BLOCK )==" "\n" |
76 | R"==(+ (od * OH * OW + oh * OW + ow) * MB_BLOCK * OC_BLOCK; )==" "\n" |
77 | R"==(#endif )==" "\n" |
78 | R"==(src += mb )==" "\n" |
79 | R"==(* ((G_WO_PADDING / IC_BLOCK) )==" "\n" |
80 | R"==(+ (G_WO_PADDING % IC_BLOCK > 0 ? 1 : 0)) )==" "\n" |
81 | R"==(* IC_BLOCK * ID * IH * IW )==" "\n" |
82 | R"==(+ g * ID * IH * IW * MB_BLOCK )==" "\n" |
83 | R"==(+ (id * IH * IW + ih * IW + iw) * MB_BLOCK * IC_BLOCK; )==" "\n" |
84 | R"==(wei += g * KD * KH * KW; )==" "\n" |
85 | R"==(DATA_T S00[OW_BLOCK] = {DATA_ZERO}; )==" "\n" |
86 | R"==(if (WITH_BIAS) { )==" "\n" |
87 | R"==(const int bg_off = g + get_sub_group_local_id(); )==" "\n" |
88 | R"==(DATA_T b = (G_WO_PADDING % OC_BLOCK == 0 || bg_off < G_WO_PADDING) )==" "\n" |
89 | R"==(? bias[bg_off] )==" "\n" |
90 | R"==(: DATA_ZERO; )==" "\n" |
91 | R"==(unroll_for(int k = 0; k < OW_BLOCK; k++) { S00[k] = b; } )==" "\n" |
92 | R"==(} )==" "\n" |
93 | R"==(#if KH != 1 || KW != 1 || KD != 1 )==" "\n" |
94 | R"==(for (int kd = 0; kd < KD; kd++) )==" "\n" |
95 | R"==(for (int kh = 0; kh < KH; kh++) { )==" "\n" |
96 | R"==(if (id + kd * (1 + DD) < 0 || id + kd * (1 + DD) >= ID) continue; )==" "\n" |
97 | R"==(if (ih + kh * (1 + DH) < 0 || ih + kh * (1 + DH) >= IH) continue; )==" "\n" |
98 | R"==(const __global DATA_T *src1 = src )==" "\n" |
99 | R"==(+ (kd * (1 + DD) * IH + kh * (1 + DH)) * IW * MB_BLOCK )==" "\n" |
100 | R"==(* IC_BLOCK; )==" "\n" |
101 | R"==(DATA_T tempA[SW * OW_BLOCK + KW * (1 + DW)] = {0}; )==" "\n" |
102 | R"==(__attribute__((opencl_unroll_hint( )==" "\n" |
103 | R"==(SW * OW_BLOCK + KW * (1 + DW)))) )==" "\n" |
104 | R"==(for (int i = 0; i < SW * OW_BLOCK + KW * (1 + DW); i++) { )==" "\n" |
105 | R"==(if ((i + iw) >= 0 && (i + iw) < IW) { )==" "\n" |
106 | R"==(tempA[i] = AS_DATA_T(BLOCK_READ((const __global BLOCK_DATA_T )==" "\n" |
107 | R"==(*)(&src1[i * IC_BLOCK]))); )==" "\n" |
108 | R"==(} )==" "\n" |
109 | R"==(} )==" "\n" |
110 | R"==(for (int kw = 0; kw < KW; kw++) { )==" "\n" |
111 | R"==(const __global DATA_T *wei1 )==" "\n" |
112 | R"==(= wei + (kd * KH * KW + kh * KW + kw) * OC_BLOCK; )==" "\n" |
113 | R"==(#else )==" "\n" |
114 | R"==(const int kw = 0; )==" "\n" |
115 | R"==(const __global DATA_T *wei1 = wei; )==" "\n" |
116 | R"==(const __global DATA_T *src1 = src; )==" "\n" |
117 | R"==(#endif )==" "\n" |
118 | R"==(DATA_T B0 = AS_DATA_T( )==" "\n" |
119 | R"==(BLOCK_READ((const __global BLOCK_DATA_T *)(wei1))); )==" "\n" |
120 | R"==(DATA_T A0; )==" "\n" |
121 | R"==(__attribute__((opencl_unroll_hint(OW_BLOCK))) )==" "\n" |
122 | R"==(for (int k = 0; k < OW_BLOCK; k++) { )==" "\n" |
123 | R"==(if (G != G_WO_PADDING && g >= G_WO_PADDING) { )==" "\n" |
124 | R"==(S00[k] = DATA_ZERO; )==" "\n" |
125 | R"==(continue; )==" "\n" |
126 | R"==(} )==" "\n" |
127 | R"==(#if KH != 1 || KW != 1 || KD != 1 )==" "\n" |
128 | R"==(A0 = tempA[k * SW + kw * (1 + DW)]; )==" "\n" |
129 | R"==(#else )==" "\n" |
130 | R"==(if (iw + kw * (1 + DW) + k * SW < 0 )==" "\n" |
131 | R"==(|| iw + kw * (1 + DW) + k * SW >= IW) )==" "\n" |
132 | R"==(A0 = DATA_ZERO; )==" "\n" |
133 | R"==(else )==" "\n" |
134 | R"==(A0 = AS_DATA_T(BLOCK_READ( )==" "\n" |
135 | R"==((const __global BLOCK_DATA_T *)(&src1[k * SW * IC_BLOCK]))); )==" "\n" |
136 | R"==(#endif )==" "\n" |
137 | R"==(S00[k] = fma(A0, (DATA_T)B0, S00[k]); )==" "\n" |
138 | R"==(} )==" "\n" |
139 | R"==(#if KH != 1 || KW != 1 || KD != 1 )==" "\n" |
140 | R"==(} )==" "\n" |
141 | R"==(} )==" "\n" |
142 | R"==(#endif )==" "\n" |
143 | R"==(DATA_T D00[OW_BLOCK] = {0}; )==" "\n" |
144 | R"==(#if WITH_SUM )==" "\n" |
145 | R"==(#ifdef DST_DT_S8 )==" "\n" |
146 | R"==(__attribute__((opencl_unroll_hint(OW_BLOCK))) )==" "\n" |
147 | R"==(for (int k = 0; k < OW_BLOCK; k++) { )==" "\n" |
148 | R"==(D00[k] = CONVERT_DATA_T(BLOCK_READ_DST( )==" "\n" |
149 | R"==((const __global DST_DATA_T *)&dst[k * DST_OC_BLOCK])); )==" "\n" |
150 | R"==(} )==" "\n" |
151 | R"==(#else )==" "\n" |
152 | R"==(__attribute__((opencl_unroll_hint(OW_BLOCK))) )==" "\n" |
153 | R"==(for (int k = 0; k < OW_BLOCK; k++) { )==" "\n" |
154 | R"==(D00[k] = AS_DATA_T( )==" "\n" |
155 | R"==(BLOCK_READ((const __global BLOCK_DATA_T *)&dst[k * OC_BLOCK])); )==" "\n" |
156 | R"==(} )==" "\n" |
157 | R"==(#endif )==" "\n" |
158 | R"==(#endif )==" "\n" |
159 | R"==(APPLY_POST_OPS_COMMON(OW_BLOCK, S00, D00, 0); )==" "\n" |
160 | R"==(if (OW % OW_BLOCK == 0 || ow + OW_BLOCK <= OW) { )==" "\n" |
161 | R"==(__attribute__((opencl_unroll_hint)) )==" "\n" |
162 | R"==(for (int k = 0; k < OW_BLOCK; k++) { )==" "\n" |
163 | R"==(#ifdef DST_DT_S8 )==" "\n" |
164 | R"==(BLOCK_WRITE_DST((__global DST_DATA_T *)&dst[k * DST_OC_BLOCK], )==" "\n" |
165 | R"==(CONVERT_DST_DATA_T(S00[k])); )==" "\n" |
166 | R"==(#else )==" "\n" |
167 | R"==(BLOCK_WRITE((__global BLOCK_DATA_T *)&dst[k * OC_BLOCK], )==" "\n" |
168 | R"==(AS_UINT_T(S00[k])); )==" "\n" |
169 | R"==(#endif )==" "\n" |
170 | R"==(} )==" "\n" |
171 | R"==(} else { )==" "\n" |
172 | R"==(__attribute__((opencl_unroll_hint)) )==" "\n" |
173 | R"==(for (int k = 0; k < OW % OW_BLOCK; k++) { )==" "\n" |
174 | R"==(#ifdef DST_DT_S8 )==" "\n" |
175 | R"==(BLOCK_WRITE_DST((__global DST_DATA_T *)&dst[k * DST_OC_BLOCK], )==" "\n" |
176 | R"==(CONVERT_DST_DATA_T(S00[k])); )==" "\n" |
177 | R"==(#else )==" "\n" |
178 | R"==(BLOCK_WRITE((__global BLOCK_DATA_T *)&dst[k * OC_BLOCK], )==" "\n" |
179 | R"==(AS_UINT_T(S00[k])); )==" "\n" |
180 | R"==(#endif )==" "\n" |
181 | R"==(} )==" "\n" |
182 | R"==(} )==" "\n" |
183 | R"==(#endif )==" "\n" |
184 | R"==(#if VER_16MB16C || VER_32MB16C )==" "\n" |
185 | R"==(const int osp = get_global_id(1); )==" "\n" |
186 | R"==(const int od = osp / (OWB * OH); )==" "\n" |
187 | R"==(const int ohw = osp % (OWB * OH); )==" "\n" |
188 | R"==(const int ow = (ohw % OWB) * OW_BLOCK; )==" "\n" |
189 | R"==(const int oh = ohw / OWB; )==" "\n" |
190 | R"==(const int g )==" "\n" |
191 | R"==(= (get_group_id(0) * (LWS_0 / SUB_GROUP_SIZE) + get_sub_group_id()) )==" "\n" |
192 | R"==(* OC_BLOCK; )==" "\n" |
193 | R"==(const int mb = get_global_id(2) * MB_BLOCK; )==" "\n" |
194 | R"==(const int id = od * SD - PD; )==" "\n" |
195 | R"==(const int ih = oh * SH - PH; )==" "\n" |
196 | R"==(const int iw = ow * SW - PW; )==" "\n" |
197 | R"==(#ifdef DST_DT_S8 )==" "\n" |
198 | R"==(const int G_32block = G % 32 ? (32 + G - (G % 32)) : G; )==" "\n" |
199 | R"==(dst += (mb / DST_MB_BLOCK) * G_32block * OD * OH * OW * DST_MB_BLOCK )==" "\n" |
200 | R"==(+ (mb % DST_MB_BLOCK) * DST_OC_BLOCK )==" "\n" |
201 | R"==(+ (g / DST_OC_BLOCK) * OD * OH * OW * DST_MB_BLOCK * DST_OC_BLOCK )==" "\n" |
202 | R"==(+ (od * OH * OW + oh * OW + ow) * DST_MB_BLOCK * DST_OC_BLOCK )==" "\n" |
203 | R"==(+ (g % DST_OC_BLOCK); )==" "\n" |
204 | R"==(#else )==" "\n" |
205 | R"==(dst += mb * G * OD * OH * OW + g * OD * OH * OW * MB_BLOCK )==" "\n" |
206 | R"==(+ (od * OH * OW + oh * OW + ow) * MB_BLOCK * OC_BLOCK; )==" "\n" |
207 | R"==(#endif )==" "\n" |
208 | R"==(src += mb )==" "\n" |
209 | R"==(* ((G_WO_PADDING / IC_BLOCK) )==" "\n" |
210 | R"==(+ (G_WO_PADDING % IC_BLOCK > 0 ? 1 : 0)) )==" "\n" |
211 | R"==(* IC_BLOCK * ID * IH * IW )==" "\n" |
212 | R"==(+ g * ID * IH * IW * MB_BLOCK )==" "\n" |
213 | R"==(+ (id * IH * IW + ih * IW + iw) * MB_BLOCK * IC_BLOCK; )==" "\n" |
214 | R"==(wei += g * KD * KH * KW; )==" "\n" |
215 | R"==(DATA8_T S00 = DATA_ZERO; )==" "\n" |
216 | R"==(DATA8_T S01 = DATA_ZERO; )==" "\n" |
217 | R"==(#if VER_32MB16C )==" "\n" |
218 | R"==(DATA8_T S02 = DATA_ZERO; )==" "\n" |
219 | R"==(DATA8_T S03 = DATA_ZERO; )==" "\n" |
220 | R"==(#endif )==" "\n" |
221 | R"==(if (WITH_BIAS) { )==" "\n" |
222 | R"==(const int bg_off = g + get_sub_group_local_id(); )==" "\n" |
223 | R"==(DATA_T b = (G_WO_PADDING % OC_BLOCK == 0 || bg_off < G_WO_PADDING) )==" "\n" |
224 | R"==(? bias[bg_off] )==" "\n" |
225 | R"==(: DATA_ZERO; )==" "\n" |
226 | R"==(unroll_for(int k = 0; k < 8; k++) { )==" "\n" |
227 | R"==(S00[k] = b; )==" "\n" |
228 | R"==(S01[k] = b; )==" "\n" |
229 | R"==(#if VER_32MB16C )==" "\n" |
230 | R"==(S02[k] = b; )==" "\n" |
231 | R"==(S03[k] = b; )==" "\n" |
232 | R"==(#endif )==" "\n" |
233 | R"==(} )==" "\n" |
234 | R"==(} )==" "\n" |
235 | R"==(#if KH != 1 || KW != 1 || KD != 1 )==" "\n" |
236 | R"==(for (int kd = 0; kd < KD; kd++) )==" "\n" |
237 | R"==(for (int kh = 0; kh < KH; kh++) )==" "\n" |
238 | R"==(for (int kw = 0; kw < KW; kw++) { )==" "\n" |
239 | R"==(if (id + kd * (1 + DD) < 0 || id + kd * (1 + DD) >= ID) )==" "\n" |
240 | R"==(continue; )==" "\n" |
241 | R"==(if (ih + kh * (1 + DH) < 0 || ih + kh * (1 + DH) >= IH) )==" "\n" |
242 | R"==(continue; )==" "\n" |
243 | R"==(if (iw + kw * (1 + DW) < 0 || iw + kw * (1 + DW) >= IW) )==" "\n" |
244 | R"==(continue; )==" "\n" |
245 | R"==(const __global DATA_T *wei1 )==" "\n" |
246 | R"==(= wei + (kd * KH * KW + kh * KW + kw) * OC_BLOCK; )==" "\n" |
247 | R"==(const __global DATA_T *src1 = src )==" "\n" |
248 | R"==(+ (kd * (1 + DD) * IH * IW + kh * (1 + DH) * IW )==" "\n" |
249 | R"==(+ kw * (1 + DW)) )==" "\n" |
250 | R"==(* MB_BLOCK * IC_BLOCK; )==" "\n" |
251 | R"==(#else )==" "\n" |
252 | R"==(const __global DATA_T *wei1 = wei; )==" "\n" |
253 | R"==(const __global DATA_T *src1 = src; )==" "\n" |
254 | R"==(#endif )==" "\n" |
255 | R"==(if (G != G_WO_PADDING && g >= G_WO_PADDING) { )==" "\n" |
256 | R"==(S00 = DATA_ZERO; )==" "\n" |
257 | R"==(S01 = DATA_ZERO; )==" "\n" |
258 | R"==(#if VER_32MB16C )==" "\n" |
259 | R"==(S02 = DATA_ZERO; )==" "\n" |
260 | R"==(S03 = DATA_ZERO; )==" "\n" |
261 | R"==(#endif )==" "\n" |
262 | R"==(continue; )==" "\n" |
263 | R"==(} )==" "\n" |
264 | R"==(DATA8_T A0 = AS_DATA8_T( )==" "\n" |
265 | R"==(BLOCK_READ8((const __global BLOCK_DATA_T *)(src1))); )==" "\n" |
266 | R"==(DATA8_T A1 = AS_DATA8_T(BLOCK_READ8( )==" "\n" |
267 | R"==((const __global BLOCK_DATA_T *)&src1[8 * IC_BLOCK])); )==" "\n" |
268 | R"==(#if VER_32MB16C )==" "\n" |
269 | R"==(DATA8_T A2 = AS_DATA8_T(BLOCK_READ8( )==" "\n" |
270 | R"==((const __global BLOCK_DATA_T *)&src1[16 * IC_BLOCK])); )==" "\n" |
271 | R"==(DATA8_T A3 = AS_DATA8_T(BLOCK_READ8( )==" "\n" |
272 | R"==((const __global BLOCK_DATA_T *)&src1[24 * IC_BLOCK])); )==" "\n" |
273 | R"==(#endif )==" "\n" |
274 | R"==(DATA_T B0 = AS_DATA_T( )==" "\n" |
275 | R"==(BLOCK_READ((const __global BLOCK_DATA_T *)(wei1))); )==" "\n" |
276 | R"==(S00 = fma(A0, (DATA8_T)B0, S00); )==" "\n" |
277 | R"==(S01 = fma(A1, (DATA8_T)B0, S01); )==" "\n" |
278 | R"==(#if VER_32MB16C )==" "\n" |
279 | R"==(S02 = fma(A2, (DATA8_T)B0, S02); )==" "\n" |
280 | R"==(S03 = fma(A3, (DATA8_T)B0, S03); )==" "\n" |
281 | R"==(#endif )==" "\n" |
282 | R"==(#if KH != 1 || KW != 1 || KD != 1 )==" "\n" |
283 | R"==(} )==" "\n" |
284 | R"==(#endif )==" "\n" |
285 | R"==(DATA8_T D00; )==" "\n" |
286 | R"==(DATA8_T D01; )==" "\n" |
287 | R"==(#if VER_32MB16C )==" "\n" |
288 | R"==(DATA8_T D02; )==" "\n" |
289 | R"==(DATA8_T D03; )==" "\n" |
290 | R"==(#endif )==" "\n" |
291 | R"==(#if WITH_SUM )==" "\n" |
292 | R"==(#ifdef DST_DT_S8 )==" "\n" |
293 | R"==(for (int i = 0; i < 8; ++i) { )==" "\n" |
294 | R"==(D00[i] = CONVERT_DATA_T( )==" "\n" |
295 | R"==(BLOCK_READ_DST((__global DST_DATA_T *)&dst[i * 32])); )==" "\n" |
296 | R"==(D01[i] = CONVERT_DATA_T( )==" "\n" |
297 | R"==(BLOCK_READ_DST((__global DST_DATA_T *)&dst[(i * 32) + 256])); )==" "\n" |
298 | R"==(#if VER_32MB16C )==" "\n" |
299 | R"==(D02[i] = CONVERT_DATA_T( )==" "\n" |
300 | R"==(BLOCK_READ_DST((__global DST_DATA_T *)&dst[i * 32] + 512)); )==" "\n" |
301 | R"==(D03[i] = CONVERT_DATA_T( )==" "\n" |
302 | R"==(BLOCK_READ_DST((__global DST_DATA_T *)&dst[(i * 32) + 768])); )==" "\n" |
303 | R"==(#endif )==" "\n" |
304 | R"==(} )==" "\n" |
305 | R"==(#else )==" "\n" |
306 | R"==(D00 = AS_DATA8_T(BLOCK_READ8((const __global BLOCK_DATA_T *)dst)); )==" "\n" |
307 | R"==(D01 = AS_DATA8_T( )==" "\n" |
308 | R"==(BLOCK_READ8((const __global BLOCK_DATA_T *)&dst[8 * OC_BLOCK])); )==" "\n" |
309 | R"==(#if VER_32MB16C )==" "\n" |
310 | R"==(D02 = AS_DATA8_T( )==" "\n" |
311 | R"==(BLOCK_READ8((const __global BLOCK_DATA_T *)&dst[16 * OC_BLOCK])); )==" "\n" |
312 | R"==(D03 = AS_DATA8_T( )==" "\n" |
313 | R"==(BLOCK_READ8((const __global BLOCK_DATA_T *)&dst[24 * OC_BLOCK])); )==" "\n" |
314 | R"==(#endif )==" "\n" |
315 | R"==(#endif )==" "\n" |
316 | R"==(#endif )==" "\n" |
317 | R"==(APPLY_POST_OPS_COMMON(8, S00, D00, 0); )==" "\n" |
318 | R"==(APPLY_POST_OPS_COMMON(8, S01, D01, 8); )==" "\n" |
319 | R"==(#if VER_32MB16C )==" "\n" |
320 | R"==(APPLY_POST_OPS_COMMON(8, S02, D02, 16); )==" "\n" |
321 | R"==(APPLY_POST_OPS_COMMON(8, S03, D03, 24); )==" "\n" |
322 | R"==(#endif )==" "\n" |
323 | R"==(#ifdef DST_DT_S8 )==" "\n" |
324 | R"==(for (int i = 0; i < 8; ++i) { )==" "\n" |
325 | R"==(BLOCK_WRITE_DST((__global DST_DATA_T *)&dst[i * DST_OC_BLOCK], )==" "\n" |
326 | R"==(CONVERT_DST_DATA_T(S00[i])); )==" "\n" |
327 | R"==(BLOCK_WRITE_DST((__global DST_DATA_T *)&dst[(i + 8) * DST_OC_BLOCK], )==" "\n" |
328 | R"==(CONVERT_DST_DATA_T(S01[i])); )==" "\n" |
329 | R"==(#if VER_32MB16C )==" "\n" |
330 | R"==(BLOCK_WRITE_DST((__global DST_DATA_T *)&dst[(i + 16) * DST_OC_BLOCK], )==" "\n" |
331 | R"==(CONVERT_DST_DATA_T(S02[i])); )==" "\n" |
332 | R"==(BLOCK_WRITE_DST((__global DST_DATA_T *)&dst[(i + 24) * DST_OC_BLOCK], )==" "\n" |
333 | R"==(CONVERT_DST_DATA_T(S03[i])); )==" "\n" |
334 | R"==(#endif )==" "\n" |
335 | R"==(} )==" "\n" |
336 | R"==(#else )==" "\n" |
337 | R"==(BLOCK_WRITE8((__global BLOCK_DATA_T *)&dst[0], AS_UINT8_T(S00)); )==" "\n" |
338 | R"==(BLOCK_WRITE8((__global BLOCK_DATA_T *)&dst[8 * OC_BLOCK], AS_UINT8_T(S01)); )==" "\n" |
339 | R"==(#if VER_32MB16C )==" "\n" |
340 | R"==(BLOCK_WRITE8((__global BLOCK_DATA_T *)&dst[16 * OC_BLOCK], AS_UINT8_T(S02)); )==" "\n" |
341 | R"==(BLOCK_WRITE8((__global BLOCK_DATA_T *)&dst[24 * OC_BLOCK], AS_UINT8_T(S03)); )==" "\n" |
342 | R"==(#endif )==" "\n" |
343 | R"==(#endif )==" "\n" |
344 | R"==(#endif )==" "\n" |
345 | R"==(return; )==" "\n" |
346 | R"==(} )==" "\n" |
347 | R"==()==" ; |
348 | } |
349 | } |
350 | } |
351 | } |