1namespace dnnl {
2namespace impl {
3namespace gpu {
4namespace ocl {
5const char *xe_lp_conv_fwd_data_mb_block_x8s8x_kernel = R"==(/******************************************************************************* )==""\n"
6R"==(* Copyright 2019-2022 Intel Corporation )==""\n"
7R"==(* )==""\n"
8R"==(* Licensed under the Apache License, Version 2.0 (the "License"); )==""\n"
9R"==(* you may not use this file except in compliance with the License. )==""\n"
10R"==(* You may obtain a copy of the License at )==""\n"
11R"==(* )==""\n"
12R"==(* http: )==""\n"
13R"==(* )==""\n"
14R"==(* Unless required by applicable law or agreed to in writing, software )==""\n"
15R"==(* distributed under the License is distributed on an "AS IS" BASIS, )==""\n"
16R"==(* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. )==""\n"
17R"==(* See the License for the specific language governing permissions and )==""\n"
18R"==(* limitations under the License. )==""\n"
19R"==(*******************************************************************************/ )==""\n"
20R"==(#include "gpu/ocl/ocl_math_utils.h" )==""\n"
21R"==(#include "gpu/ocl/ocl_post_ops.h" )==""\n"
22R"==(#include "gpu/ocl/ocl_types.h" )==""\n"
23R"==(#include "gpu/ocl/ocl_zero_points.h" )==""\n"
24R"==(#define SRC_DATA_BLOCK_T SRC_MMAD_DATA8_T )==""\n"
25R"==(#define AS_SRC_DATA_BLOCK_T AS_SRC_MMAD_DATA8_T )==""\n"
26R"==(#define BLOCK_READ_SRC(data, idx) \ )==""\n"
27R"==(data = AS_SRC_DATA_BLOCK_T( \ )==""\n"
28R"==(intel_sub_group_block_read8((__global uint *)&src[idx])); )==""\n"
29R"==(#define BLOCK_READ_WHT(data, idx) \ )==""\n"
30R"==(data = as_int8(intel_sub_group_block_read8((__global uint *)&wei[idx])); )==""\n"
31R"==(#if OC % OC_BLOCK == 0 )==""\n"
32R"==(#define BLOCK_READ_BIA(data, idx) \ )==""\n"
33R"==(data = as_float4(intel_sub_group_block_read4((__global uint *)&bias[idx])); )==""\n"
34R"==(#else )==""\n"
35R"==(#define BLOCK_READ_BIA(data, idx) \ )==""\n"
36R"==(data = (float4)0; \ )==""\n"
37R"==(int i; \ )==""\n"
38R"==(for (i = idx; i < idx + OC_BLOCK && i < OC - (OC % SUB_GROUP_SIZE); \ )==""\n"
39R"==(i += SUB_GROUP_SIZE) { \ )==""\n"
40R"==(data[(i - idx) / SUB_GROUP_SIZE] = as_float( \ )==""\n"
41R"==(intel_sub_group_block_read((__global uint *)&bias[i])); \ )==""\n"
42R"==(} \ )==""\n"
43R"==(if ((get_sub_group_local_id() < OC % SUB_GROUP_SIZE) && i < OC \ )==""\n"
44R"==(&& (i - idx) / SUB_GROUP_SIZE < 4) { \ )==""\n"
45R"==(data[(i - idx) / SUB_GROUP_SIZE] \ )==""\n"
46R"==(= as_float(bias[i + get_sub_group_local_id()]); \ )==""\n"
47R"==(} )==""\n"
48R"==(#endif )==""\n"
49R"==(#define BLOCK_READ_SCALES(data, idx) \ )==""\n"
50R"==(data = as_float4(intel_sub_group_block_read4( \ )==""\n"
51R"==((__global uint *)&runtime_scales[idx])); )==""\n"
52R"==(#if SCALES_PER_OC )==""\n"
53R"==(#define SCALE scales )==""\n"
54R"==(#elif SCALES_COMMON )==""\n"
55R"==(#define SCALE runtime_scales[0] )==""\n"
56R"==(#else )==""\n"
57R"==(#define SCALE 1 )==""\n"
58R"==(#endif )==""\n"
59R"==(__attribute__((intel_reqd_sub_group_size(SUB_GROUP_SIZE))) )==""\n"
60R"==(__attribute__((reqd_work_group_size(LWS_0, LWS_1, LWS_2))) __kernel void )==""\n"
61R"==(conv_fwd_mb_block_x8s8x(const __global uchar *src, const __global char *wei, )==""\n"
62R"==(const __global float *bias, __global DST_DATA_T *dst POST_OP_ARGS, )==""\n"
63R"==(const __global float *runtime_scales, )==""\n"
64R"==(const __global int *src_compensation, const __global int *src_zpoints, )==""\n"
65R"==(const __global int *dst_compensation) { )==""\n"
66R"==(#ifdef MB_FULL_BLOCK )==""\n"
67R"==(const int mb_blocks = 1; )==""\n"
68R"==(#else )==""\n"
69R"==(const int mb_blocks = 2; )==""\n"
70R"==(#endif )==""\n"
71R"==(const int group_oc = get_group_id(0) * OC_GROUP; )==""\n"
72R"==(const int group_mb = get_group_id(2) * MB_GROUP / mb_blocks; )==""\n"
73R"==(const int group_sp = get_group_id(1) * SP_GROUP; )==""\n"
74R"==(const int mb = get_group_id(2) % mb_blocks; )==""\n"
75R"==(const int sub_group_id = get_sub_group_id(); )==""\n"
76R"==(const int ocl_local_id = get_local_id(0); )==""\n"
77R"==(const int oc = (sub_group_id % OC_GROUP); )==""\n"
78R"==(const int sp = (sub_group_id / OC_GROUP); )==""\n"
79R"==(const int g = (group_oc + oc) / OC_NCHUNK; )==""\n"
80R"==(const int group_ic = IC_NCHUNK * g; )==""\n"
81R"==(const int god = group_sp / (OW_PADDED * OH); )==""\n"
82R"==(const int gohw = group_sp % (OW_PADDED * OH); )==""\n"
83R"==(const int goh = gohw / OW_PADDED; )==""\n"
84R"==(const int gow = gohw % OW_PADDED; )==""\n"
85R"==(const int gid = god * SD; )==""\n"
86R"==(const int gih = goh * SH; )==""\n"
87R"==(const int giw = gow * SW; )==""\n"
88R"==(const int local_oh = sp / OW_PADDED; )==""\n"
89R"==(const int local_ow = sp % OW_PADDED; )==""\n"
90R"==(const int local_ih = local_oh * SH; )==""\n"
91R"==(const int local_iw = local_ow * SW; )==""\n"
92R"==(const int od = god; )==""\n"
93R"==(const int ow = gow + local_ow; )==""\n"
94R"==(const int oh = goh + local_oh; )==""\n"
95R"==(const int id = gid - PD; )==""\n"
96R"==(const int iw = giw + local_iw - PW; )==""\n"
97R"==(const int ih = gih + local_ih - PH; )==""\n"
98R"==(if (ow >= OW) return; )==""\n"
99R"==(dst += OC_BLOCK * OD * OH * OW * MB_BLOCK * (group_oc + oc); )==""\n"
100R"==(dst += OC_BLOCK * OD * OH * OW * OC_NCHUNK * G * MB_BLOCK * group_mb; )==""\n"
101R"==(dst += OC_BLOCK * MB_BLOCK / mb_blocks * mb; )==""\n"
102R"==(dst += OC_BLOCK * MB_BLOCK * (OW * OH * od + OW * oh + ow); )==""\n"
103R"==(src += IC_BLOCK * ID * IH * IW * MB_BLOCK * group_ic; )==""\n"
104R"==(src += IC_BLOCK * ID * IH * IW * IC_NCHUNK * G * MB_BLOCK * group_mb; )==""\n"
105R"==(src += IC_BLOCK * MB_BLOCK / mb_blocks * mb; )==""\n"
106R"==(src += IC_BLOCK * MB_BLOCK * (IW * IH * id + IW * ih + iw); )==""\n"
107R"==(wei += IC_BLOCK * KD * KH * KW * OC_BLOCK * (group_oc + oc) * IC_NCHUNK; )==""\n"
108R"==(int8 C00 = 0, C01 = 0, C02 = 0, C03 = 0; )==""\n"
109R"==(int8 C10 = 0, C11 = 0, C12 = 0, C13 = 0; )==""\n"
110R"==(int8 C20 = 0, C21 = 0, C22 = 0, C23 = 0; )==""\n"
111R"==(int8 C30 = 0, C31 = 0, C32 = 0, C33 = 0; )==""\n"
112R"==(for (int ic_chunk = 0; ic_chunk < IC_NCHUNK; ic_chunk++) { )==""\n"
113R"==(if (MB % MB_BLOCK != 0 )==""\n"
114R"==(&& (group_mb * MB_BLOCK + mb * MB_BLOCK / mb_blocks) >= MB) { )==""\n"
115R"==(break; )==""\n"
116R"==(} )==""\n"
117R"==(SRC_DATA_BLOCK_T S0, S1, S2, S3; )==""\n"
118R"==(int8 W0, W1, W2, W3; )==""\n"
119R"==(for (int kd = 0; kd < KD; kd++) { )==""\n"
120R"==(#if WITH_SRC_ZPOINTS )==""\n"
121R"==(const int is_pad_d )==""\n"
122R"==(= kd * (1 + DD) + id < 0 || kd * (1 + DD) + id >= ID; )==""\n"
123R"==(#else )==""\n"
124R"==(if (kd * (1 + DD) + id < 0 || kd * (1 + DD) + id >= ID) { )==""\n"
125R"==(src += IC_BLOCK * MB_BLOCK * IH * IW * (1 + DD); )==""\n"
126R"==(wei += IC_BLOCK * OC_BLOCK * KH * KW; )==""\n"
127R"==(continue; )==""\n"
128R"==(} )==""\n"
129R"==(#endif )==""\n"
130R"==(for (int kh = 0; kh < KH; kh++) { )==""\n"
131R"==(#if WITH_SRC_ZPOINTS )==""\n"
132R"==(const int is_pad_h )==""\n"
133R"==(= kh * (1 + DH) + ih < 0 || kh * (1 + DH) + ih >= IH; )==""\n"
134R"==(#else )==""\n"
135R"==(if (kh * (1 + DH) + ih < 0 || kh * (1 + DH) + ih >= IH) { )==""\n"
136R"==(src += IC_BLOCK * MB_BLOCK * IW * (1 + DH); )==""\n"
137R"==(wei += IC_BLOCK * OC_BLOCK * KW; )==""\n"
138R"==(continue; )==""\n"
139R"==(} )==""\n"
140R"==(#endif )==""\n"
141R"==(for (int kw = 0; kw < KW; kw++) { )==""\n"
142R"==(#if WITH_SRC_ZPOINTS )==""\n"
143R"==(const int is_pad_w = kw * (1 + DW) + iw < 0 )==""\n"
144R"==(|| kw * (1 + DW) + iw >= IW; )==""\n"
145R"==(if (is_pad_w || is_pad_h || is_pad_d) { )==""\n"
146R"==(#if WITH_SRC_ZPOINTS_PER_IC )==""\n"
147R"==(const int4 z = read_src_zero_points_32c( )==""\n"
148R"==(src_zpoints, (group_ic + ic_chunk) * IC_BLOCK); )==""\n"
149R"==(#else )==""\n"
150R"==(const int z = read_src_zero_point(src_zpoints); )==""\n"
151R"==(#endif )==""\n"
152R"==(BLOCK_READ_WHT(W0, 0); )==""\n"
153R"==(BLOCK_READ_WHT(W1, 8 * IC_BLOCK); )==""\n"
154R"==(BLOCK_READ_WHT(W2, 16 * IC_BLOCK); )==""\n"
155R"==(BLOCK_READ_WHT(W3, 24 * IC_BLOCK); )==""\n"
156R"==(int4 acc = 0; )==""\n"
157R"==(#if WITH_SRC_ZPOINTS_PER_IC )==""\n"
158R"==(acc.s0 += calc_src_compensation_x32(z, W0); )==""\n"
159R"==(acc.s1 += calc_src_compensation_x32(z, W1); )==""\n"
160R"==(acc.s2 += calc_src_compensation_x32(z, W2); )==""\n"
161R"==(acc.s3 += calc_src_compensation_x32(z, W3); )==""\n"
162R"==(#else )==""\n"
163R"==(unroll_for(uint i = 0; i < 8; ++i) { )==""\n"
164R"==(acc.s0 = idot4(0x01010101, W0[i], acc.s0); )==""\n"
165R"==(acc.s1 = idot4(0x01010101, W1[i], acc.s1); )==""\n"
166R"==(acc.s2 = idot4(0x01010101, W2[i], acc.s2); )==""\n"
167R"==(acc.s3 = idot4(0x01010101, W3[i], acc.s3); )==""\n"
168R"==(} )==""\n"
169R"==(acc = z * acc; )==""\n"
170R"==(#endif )==""\n"
171R"==(C00 += acc.s0; )==""\n"
172R"==(C01 += acc.s1; )==""\n"
173R"==(C02 += acc.s2; )==""\n"
174R"==(C03 += acc.s3; )==""\n"
175R"==(#if MB > 8 )==""\n"
176R"==(C10 += acc.s0; )==""\n"
177R"==(C11 += acc.s1; )==""\n"
178R"==(C12 += acc.s2; )==""\n"
179R"==(C13 += acc.s3; )==""\n"
180R"==(#ifdef MB_FULL_BLOCK )==""\n"
181R"==(C20 += acc.s0; )==""\n"
182R"==(C21 += acc.s1; )==""\n"
183R"==(C22 += acc.s2; )==""\n"
184R"==(C23 += acc.s3; )==""\n"
185R"==(C30 += acc.s0; )==""\n"
186R"==(C31 += acc.s1; )==""\n"
187R"==(C32 += acc.s2; )==""\n"
188R"==(C33 += acc.s3; )==""\n"
189R"==(#endif )==""\n"
190R"==(#endif )==""\n"
191R"==(} else )==""\n"
192R"==(#else )==""\n"
193R"==(if (kw * (1 + DW) + iw >= 0 && kw * (1 + DW) + iw < IW) )==""\n"
194R"==(#endif )==""\n"
195R"==({ )==""\n"
196R"==(BLOCK_READ_SRC(S0, 0); )==""\n"
197R"==(#if MB > 8 )==""\n"
198R"==(BLOCK_READ_SRC(S1, 8 * IC_BLOCK); )==""\n"
199R"==(#ifdef MB_FULL_BLOCK )==""\n"
200R"==(BLOCK_READ_SRC(S2, 16 * IC_BLOCK); )==""\n"
201R"==(BLOCK_READ_SRC(S3, 24 * IC_BLOCK); )==""\n"
202R"==(#endif )==""\n"
203R"==(#endif )==""\n"
204R"==(BLOCK_READ_WHT(W0, 0); )==""\n"
205R"==(BLOCK_READ_WHT(W1, 8 * IC_BLOCK); )==""\n"
206R"==(BLOCK_READ_WHT(W2, 16 * IC_BLOCK); )==""\n"
207R"==(BLOCK_READ_WHT(W3, 24 * IC_BLOCK); )==""\n"
208R"==(C00 = mmad8x8(S0, W0, C00); )==""\n"
209R"==(C01 = mmad8x8(S0, W1, C01); )==""\n"
210R"==(C02 = mmad8x8(S0, W2, C02); )==""\n"
211R"==(C03 = mmad8x8(S0, W3, C03); )==""\n"
212R"==(#if MB > 8 )==""\n"
213R"==(C10 = mmad8x8(S1, W0, C10); )==""\n"
214R"==(C11 = mmad8x8(S1, W1, C11); )==""\n"
215R"==(C12 = mmad8x8(S1, W2, C12); )==""\n"
216R"==(C13 = mmad8x8(S1, W3, C13); )==""\n"
217R"==(#ifdef MB_FULL_BLOCK )==""\n"
218R"==(C20 = mmad8x8(S2, W0, C20); )==""\n"
219R"==(C21 = mmad8x8(S2, W1, C21); )==""\n"
220R"==(C22 = mmad8x8(S2, W2, C22); )==""\n"
221R"==(C23 = mmad8x8(S2, W3, C23); )==""\n"
222R"==(C30 = mmad8x8(S3, W0, C30); )==""\n"
223R"==(C31 = mmad8x8(S3, W1, C31); )==""\n"
224R"==(C32 = mmad8x8(S3, W2, C32); )==""\n"
225R"==(C33 = mmad8x8(S3, W3, C33); )==""\n"
226R"==(#endif )==""\n"
227R"==(#endif )==""\n"
228R"==(} )==""\n"
229R"==(src += IC_BLOCK * MB_BLOCK * (1 + DW); )==""\n"
230R"==(wei += IC_BLOCK * OC_BLOCK; )==""\n"
231R"==(} )==""\n"
232R"==(src += IC_BLOCK * MB_BLOCK * (IW * (1 + DH) - KW * (1 + DW)); )==""\n"
233R"==(} )==""\n"
234R"==(src += IC_BLOCK * MB_BLOCK * (IH * (1 + DD) - KH * (1 + DH)) * IW; )==""\n"
235R"==(} )==""\n"
236R"==(src += IC_BLOCK * MB_BLOCK * (ID - KD * (1 + DD)) * IH * IW; )==""\n"
237R"==(} )==""\n"
238R"==(#if WITH_SRC_ZPOINTS )==""\n"
239R"==(int4 src_comp = as_int4(intel_sub_group_block_read4( )==""\n"
240R"==((__global uint *)(&src_compensation[(group_oc + oc) * OC_BLOCK]))); )==""\n"
241R"==(C00 -= src_comp.s0; )==""\n"
242R"==(C01 -= src_comp.s1; )==""\n"
243R"==(C02 -= src_comp.s2; )==""\n"
244R"==(C03 -= src_comp.s3; )==""\n"
245R"==(#if MB > 8 )==""\n"
246R"==(C10 -= src_comp.s0; )==""\n"
247R"==(C11 -= src_comp.s1; )==""\n"
248R"==(C12 -= src_comp.s2; )==""\n"
249R"==(C13 -= src_comp.s3; )==""\n"
250R"==(#ifdef MB_FULL_BLOCK )==""\n"
251R"==(C20 -= src_comp.s0; )==""\n"
252R"==(C21 -= src_comp.s1; )==""\n"
253R"==(C22 -= src_comp.s2; )==""\n"
254R"==(C23 -= src_comp.s3; )==""\n"
255R"==(C30 -= src_comp.s0; )==""\n"
256R"==(C31 -= src_comp.s1; )==""\n"
257R"==(C32 -= src_comp.s2; )==""\n"
258R"==(C33 -= src_comp.s3; )==""\n"
259R"==(#endif )==""\n"
260R"==(#endif )==""\n"
261R"==(#endif )==""\n"
262R"==(float4 tmp; )==""\n"
263R"==(DST_DATA4_T dst_pack[8]; )==""\n"
264R"==(DST_DATA4_T D0[8]; )==""\n"
265R"==(DST_DATA4_T D1[8]; )==""\n"
266R"==(DST_DATA4_T D2[8]; )==""\n"
267R"==(DST_DATA4_T D3[8]; )==""\n"
268R"==(#if SCALES_PER_OC )==""\n"
269R"==(float4 scales; )==""\n"
270R"==(BLOCK_READ_SCALES(scales, (group_oc + oc) * OC_BLOCK); )==""\n"
271R"==(#endif )==""\n"
272R"==(#if WITH_BIAS )==""\n"
273R"==(float4 bia; )==""\n"
274R"==(BLOCK_READ_BIA(bia, (group_oc + oc) * OC_BLOCK); )==""\n"
275R"==(#define QUANTIZE_ADD_BIAS() tmp = SCALE * fma(tmp, (float4)1, bia); )==""\n"
276R"==(#else )==""\n"
277R"==(#define QUANTIZE_ADD_BIAS() tmp *= SCALE; )==""\n"
278R"==(#endif )==""\n"
279R"==(#if WITH_SUM )==""\n"
280R"==(*(DST_DATA16_T *)(D0 + 0) = BLOCK_READ_DST16(dst); )==""\n"
281R"==(*(DST_DATA16_T *)(D0 + 4) = BLOCK_READ_DST16(dst + 4 * OC_BLOCK); )==""\n"
282R"==(#if MB > 8 )==""\n"
283R"==(*(DST_DATA16_T *)(D1 + 0) = BLOCK_READ_DST16(dst + 8 * OC_BLOCK); )==""\n"
284R"==(*(DST_DATA16_T *)(D1 + 4) = BLOCK_READ_DST16(dst + 12 * OC_BLOCK); )==""\n"
285R"==(#ifdef MB_FULL_BLOCK )==""\n"
286R"==(*(DST_DATA16_T *)(D2 + 0) = BLOCK_READ_DST16(dst + 16 * OC_BLOCK); )==""\n"
287R"==(*(DST_DATA16_T *)(D2 + 4) = BLOCK_READ_DST16(dst + 20 * OC_BLOCK); )==""\n"
288R"==(*(DST_DATA16_T *)(D3 + 0) = BLOCK_READ_DST16(dst + 24 * OC_BLOCK); )==""\n"
289R"==(*(DST_DATA16_T *)(D3 + 4) = BLOCK_READ_DST16(dst + 28 * OC_BLOCK); )==""\n"
290R"==(#endif )==""\n"
291R"==(#endif )==""\n"
292R"==(#endif )==""\n"
293R"==(#if WITH_DST_ZPOINTS )==""\n"
294R"==(int4 dst_zp = read_dst_zero_points_32c( )==""\n"
295R"==(dst_compensation, (group_oc + oc) * OC_BLOCK); )==""\n"
296R"==(#define ADD_DST_COMPENSATION() tmp += convert_float4(dst_zp); )==""\n"
297R"==(#else )==""\n"
298R"==(#define ADD_DST_COMPENSATION() )==""\n"
299R"==(#endif )==""\n"
300R"==(#if WITH_SRC_ZPOINTS )==""\n"
301R"==(#define ZERO_PAD_DST() tmp = zero_pad_dst_32c(tmp, (group_oc + oc) * OC_BLOCK); )==""\n"
302R"==(#else )==""\n"
303R"==(#define ZERO_PAD_DST() )==""\n"
304R"==(#endif )==""\n"
305R"==(#define PACK(C0, C1, C2, C3, idx) \ )==""\n"
306R"==(do { \ )==""\n"
307R"==(tmp[0] = C0[idx]; \ )==""\n"
308R"==(tmp[1] = C1[idx]; \ )==""\n"
309R"==(tmp[2] = C2[idx]; \ )==""\n"
310R"==(tmp[3] = C3[idx]; \ )==""\n"
311R"==(} while (0) )==""\n"
312R"==(#define CONVERT_PACK(idx) \ )==""\n"
313R"==(do { \ )==""\n"
314R"==(dst_pack[idx] = CONVERT_DST_DATA4_T(tmp); \ )==""\n"
315R"==(} while (0) )==""\n"
316R"==(#define STORE_DST(C0, C1, C2, C3, D, mb_stride) \ )==""\n"
317R"==(do { \ )==""\n"
318R"==(for (int n_i = 0; n_i < 8; n_i++) { \ )==""\n"
319R"==(PACK(C0, C1, C2, C3, n_i); \ )==""\n"
320R"==(const int po_mb = (group_mb * MB_BLOCK + mb * MB_BLOCK / mb_blocks \ )==""\n"
321R"==(+ mb_stride + n_i); \ )==""\n"
322R"==(if (MB % MB_BLOCK == 0 || po_mb < MB) { \ )==""\n"
323R"==(QUANTIZE_ADD_BIAS(); \ )==""\n"
324R"==(const int po_oc = (group_oc + oc) * OC_BLOCK; \ )==""\n"
325R"==(float4 dni \ )==""\n"
326R"==(= convert_float4(SUM_TO_REF(AS_SUM_DATA4_T(D[n_i]))); \ )==""\n"
327R"==(APPLY_POST_OPS_TRY_BURST(tmp, float, dni, float, po_mb, 1, \ )==""\n"
328R"==(po_oc, 4 * SUB_GROUP_SIZE, ocl_local_id); \ )==""\n"
329R"==(ADD_DST_COMPENSATION(); \ )==""\n"
330R"==(ZERO_PAD_DST(); \ )==""\n"
331R"==(} \ )==""\n"
332R"==(CONVERT_PACK(n_i); \ )==""\n"
333R"==(} \ )==""\n"
334R"==(BLOCK_WRITE_DST16( \ )==""\n"
335R"==(&dst[mb_stride * OC_BLOCK], *(DST_DATA16_T *)dst_pack); \ )==""\n"
336R"==(BLOCK_WRITE_DST16(&dst[mb_stride * OC_BLOCK + 16 * 8], \ )==""\n"
337R"==(*(DST_DATA16_T *)(dst_pack + 4)); \ )==""\n"
338R"==(} while (0) )==""\n"
339R"==(STORE_DST(C00, C01, C02, C03, D0, 0); )==""\n"
340R"==(STORE_DST(C10, C11, C12, C13, D1, 8); )==""\n"
341R"==(#ifdef MB_FULL_BLOCK )==""\n"
342R"==(STORE_DST(C20, C21, C22, C23, D2, 16); )==""\n"
343R"==(STORE_DST(C30, C31, C32, C33, D3, 24); )==""\n"
344R"==(#endif )==""\n"
345R"==(} )==""\n"
346R"==()==";
347}
348}
349}
350}