1namespace dnnl {
2namespace impl {
3namespace gpu {
4namespace ocl {
5const char *gen9_binary_kernel = R"==(/******************************************************************************* )==""\n"
6R"==(* Copyright 2020-2022 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/binary_types.h" )==""\n"
21R"==(#if IS_PLAIN_LAYOUT )==""\n"
22R"==(KERNEL_ATTR )==""\n"
23R"==(__kernel void gen9_binary(__global SRC0_DATA_T *src0, )==""\n"
24R"==(__global SRC1_DATA_T *src1, __global DST_DATA_T *dst POST_OP_ARGS, )==""\n"
25R"==(__global float *src0_scale, __global float *src1_scale) { )==""\n"
26R"==(int dims0[6] = {0}; )==""\n"
27R"==(unsigned mid_dim = GWS_GET_MIXED_DIM(); )==""\n"
28R"==(dims0[5] = mid_dim % DST_D5; )==""\n"
29R"==(mid_dim /= DST_D5; )==""\n"
30R"==(dims0[4] = mid_dim % DST_D4; )==""\n"
31R"==(mid_dim /= DST_D4; )==""\n"
32R"==(dims0[3] = mid_dim % DST_D3; )==""\n"
33R"==(mid_dim /= DST_D3; )==""\n"
34R"==(dims0[2] = mid_dim % DST_D2; )==""\n"
35R"==(mid_dim /= DST_D2; )==""\n"
36R"==(dims0[1] = mid_dim % DST_D1; )==""\n"
37R"==(mid_dim /= DST_D1; )==""\n"
38R"==(dims0[0] = mid_dim; )==""\n"
39R"==(int src0_off = SRC0_OFF( )==""\n"
40R"==(dims0[0], dims0[1], dims0[2], dims0[3], dims0[4], dims0[5]); )==""\n"
41R"==(src0 += src0_off; )==""\n"
42R"==(int src1_off = SRC1_OFF(dims0[0] * (!BCAST_DIM0), dims0[1] * (!BCAST_DIM1), )==""\n"
43R"==(dims0[2] * (!BCAST_DIM2), dims0[3] * (!BCAST_DIM3), )==""\n"
44R"==(dims0[4] * (!BCAST_DIM4), dims0[5] * (!BCAST_DIM5)); )==""\n"
45R"==(src1 += src1_off; )==""\n"
46R"==(int dst_off = DST_OFF( )==""\n"
47R"==(dims0[0], dims0[1], dims0[2], dims0[3], dims0[4], dims0[5]); )==""\n"
48R"==(dst += dst_off; )==""\n"
49R"==(#if WITH_SRC0_SCALE )==""\n"
50R"==(#define src0_scale_val src0_scale[0] )==""\n"
51R"==(#else )==""\n"
52R"==(#define src0_scale_val 1 )==""\n"
53R"==(#endif )==""\n"
54R"==(#if WITH_SRC1_SCALE )==""\n"
55R"==(#define src1_scale_val src1_scale[0] )==""\n"
56R"==(#else )==""\n"
57R"==(#define src1_scale_val 1 )==""\n"
58R"==(#endif )==""\n"
59R"==(float tmp_src0[NVECT]; )==""\n"
60R"==(READ_DATA(NVECT, SRC0, (&src0[0]), (&tmp_src0[0]), src0_scale_val); )==""\n"
61R"==(#if BCAST_AT_INNERMOST_DIM )==""\n"
62R"==(float tmp_src1[1]; )==""\n"
63R"==(tmp_src1[0] = src1_scale_val * CONVERT_FLOAT_T(src1[0]); )==""\n"
64R"==(#define SRC1_IDX_MASK 0 )==""\n"
65R"==(#else )==""\n"
66R"==(float tmp_src1[NVECT]; )==""\n"
67R"==(READ_DATA(NVECT, SRC1, (&src1[0]), (&tmp_src1[0]), src1_scale_val); )==""\n"
68R"==(#define SRC1_IDX_MASK 1 )==""\n"
69R"==(#endif )==""\n"
70R"==(float tmp[NVECT]; )==""\n"
71R"==(unroll_for(unsigned idx = 0; idx < NVECT; ++idx) { )==""\n"
72R"==(tmp[idx] = get_eltwise_op(tmp_src0[idx], tmp_src1[idx * SRC1_IDX_MASK]); )==""\n"
73R"==(} )==""\n"
74R"==(float dst_data[NVECT]; )==""\n"
75R"==(#if WITH_SUM )==""\n"
76R"==(READ_DATA(NVECT, DST, (&dst[0]), (&dst_data[0]), 1); )==""\n"
77R"==(#endif )==""\n"
78R"==(dims0[NDIMS - 1] += get_sub_group_local_id(); )==""\n"
79R"==(unroll_for(unsigned idx = 0; idx < NVECT; ++idx) { )==""\n"
80R"==(float d_i = tmp[idx]; )==""\n"
81R"==(float dst_i = dst_data[idx]; )==""\n"
82R"==(APPLY_POST_OPS_SERIAL(d_i, float, dst_i, float, dims0[0], 1, dims0[1], )==""\n"
83R"==(1, dims0[2], 1, dims0[3], 1, dims0[4], 1, dims0[5], 1); )==""\n"
84R"==(tmp[idx] = d_i; )==""\n"
85R"==(dims0[NDIMS - 1] += 16; )==""\n"
86R"==(} )==""\n"
87R"==(WRITE_DATA(NVECT, DST, (&tmp[0]), (&dst[0])); )==""\n"
88R"==(} )==""\n"
89R"==(#elif PLAIN_TO_ABCD4AXB )==""\n"
90R"==(KERNEL_ATTR )==""\n"
91R"==(__kernel void gen9_binary(__global SRC0_DATA_T *src0, )==""\n"
92R"==(__global SRC1_DATA_T *src1, __global DST_DATA_T *dst POST_OP_ARGS, )==""\n"
93R"==(__global float *src0_scale, __global float *src1_scale) { )==""\n"
94R"==(src0 += SRC0_OFFSET0; )==""\n"
95R"==(src1 += SRC1_OFFSET0; )==""\n"
96R"==(dst += DST_OFFSET0; )==""\n"
97R"==(int sglid = get_sub_group_local_id(); )==""\n"
98R"==(const int d0 = GWS_GET_D0(); )==""\n"
99R"==(const int d1 = GWS_GET_D1(); )==""\n"
100R"==(const int d2 = GWS_GET_D2(); )==""\n"
101R"==(const int d3 = GWS_GET_D3(); )==""\n"
102R"==(const int d4 = GWS_GET_D3(); )==""\n"
103R"==(const int d5 = GWS_GET_D3(); )==""\n"
104R"==(const int d0_block = GWS_GET_D0_BLOCK(); )==""\n"
105R"==(const int d1_block = GWS_GET_D1_BLOCK(); )==""\n"
106R"==(const int d01_block = d0_block * d1_block; )==""\n"
107R"==(SRC0_DATA_T tmp_buf0[d01_block] = {0}; )==""\n"
108R"==(SRC1_DATA_T tmp_buf1[d01_block] = {0}; )==""\n"
109R"==(DST_DATA_T res_buf[d01_block] = {0}; )==""\n"
110R"==(const int d0_inner_block = min(d0_block, SRC0_D0); )==""\n"
111R"==(const int d1_inner_block = min(d1_block, SRC0_D1); )==""\n"
112R"==(for (int d0_inner = 0; d0_inner < d0_inner_block; d0_inner++) { )==""\n"
113R"==(for (int d1_inner = 0; d1_inner < d1_inner_block; d1_inner++) { )==""\n"
114R"==(if (SRC0_D0 % d0_inner_block != 0 && d0 + d0_inner >= SRC0_D0) )==""\n"
115R"==(continue; )==""\n"
116R"==(if (SRC0_D1 % d1_inner_block != 0 && d1 + d1_inner >= SRC0_D1) )==""\n"
117R"==(continue; )==""\n"
118R"==(int src0_off; )==""\n"
119R"==(int src1_off; )==""\n"
120R"==(if (SRC0_S3_0 == 1) { )==""\n"
121R"==(src0_off = SRC0_OFF(d0 + d0_inner, d1 + d1_inner, d2, d3, 0, 0); )==""\n"
122R"==(tmp_buf0[d0_inner * d1_block + d1_inner] )==""\n"
123R"==(= SRC0_BLOCK_READ(&src0[src0_off]); )==""\n"
124R"==(src1_off = SRC1_OFF((d0 + d0_inner) * (!BCAST_DIM0), )==""\n"
125R"==((d1 + d1_inner) * (!BCAST_DIM1), d2 * (!BCAST_DIM2), )==""\n"
126R"==(d3 * (!BCAST_DIM3), 0, 0); )==""\n"
127R"==(} else { )==""\n"
128R"==(src0_off = SRC0_OFF( )==""\n"
129R"==(d0 + d0_inner, d1 + d1_inner, d2, d3 + sglid, 0, 0); )==""\n"
130R"==(tmp_buf0[d0_inner * d1_block + d1_inner] = src0[src0_off]; )==""\n"
131R"==(src1_off = SRC1_OFF((d0 + d0_inner) * (!BCAST_DIM0), )==""\n"
132R"==((d1 + d1_inner) * (!BCAST_DIM1), d2 * (!BCAST_DIM2), )==""\n"
133R"==((d3 + sglid) * (!BCAST_DIM3), 0, 0); )==""\n"
134R"==(} )==""\n"
135R"==(#if BCAST_AT_INNERMOST_DIM == 1 )==""\n"
136R"==(tmp_buf1[d0_inner * d1_block + d1_inner] = src1[src1_off]; )==""\n"
137R"==(#else )==""\n"
138R"==(tmp_buf1[d0_inner * d1_block + d1_inner] )==""\n"
139R"==(= SRC1_BLOCK_READ(&src1[src1_off]); )==""\n"
140R"==(#endif )==""\n"
141R"==(} )==""\n"
142R"==(} )==""\n"
143R"==(int i = 0; )==""\n"
144R"==(for (int d0_i = 0; d0_i < d0_block; d0_i++) { )==""\n"
145R"==(for (int d1_i = 0; d1_i < d1_block; d1_i++) { )==""\n"
146R"==(float tmp_src0 = CONVERT_FLOAT_T(tmp_buf0[i]); )==""\n"
147R"==(float tmp_src1 = CONVERT_FLOAT_T(tmp_buf1[i]); )==""\n"
148R"==(float res; )==""\n"
149R"==(float dst_data; )==""\n"
150R"==(#if WITH_SRC0_SCALE )==""\n"
151R"==(tmp_src0 = tmp_src0 * src0_scale[0]; )==""\n"
152R"==(#endif )==""\n"
153R"==(#if WITH_SRC1_SCALE )==""\n"
154R"==(tmp_src1 = tmp_src1 * src1_scale[0]; )==""\n"
155R"==(#endif )==""\n"
156R"==(res = get_eltwise_op(tmp_src0, tmp_src1); )==""\n"
157R"==(APPLY_POST_OPS_SERIAL(res, float, dst_data, float, d0 + d0_i, 1, )==""\n"
158R"==(d1 + d1_i, 1, d2, 1, d3 + sglid, 1, d4, 1, d5, 1); )==""\n"
159R"==(res_buf[i] = TO_DST(res); )==""\n"
160R"==(++i; )==""\n"
161R"==(} )==""\n"
162R"==(} )==""\n"
163R"==(DST_DATA_T res_all[d01_block][SUB_GROUP_SIZE]; )==""\n"
164R"==(for (int i = 0; i < d01_block; i++) )==""\n"
165R"==(for (int j = 0; j < SUB_GROUP_SIZE; j++) )==""\n"
166R"==(res_all[i][j] = intel_sub_group_shuffle(res_buf[i], j); )==""\n"
167R"==(for (int d = 0; d < SUB_GROUP_SIZE; d += 8) { )==""\n"
168R"==(DST_DATA8_T res_tmp; )==""\n"
169R"==(for (int i = 0; i < 8; i++) )==""\n"
170R"==(res_tmp[i] = res_all[sglid][d + i]; )==""\n"
171R"==(int dst_off = DST_OFF(d0, d1, d2, d3 + d, 0, 0); )==""\n"
172R"==(DST_BLOCK_WRITE8(&dst[dst_off], res_tmp); )==""\n"
173R"==(} )==""\n"
174R"==(} )==""\n"
175R"==(#elif IS_XA16B )==""\n"
176R"==(KERNEL_ATTR )==""\n"
177R"==(__kernel void gen9_binary(__global SRC0_DATA_T *src0, )==""\n"
178R"==(__global SRC1_DATA_T *src1, __global DST_DATA_T *dst POST_OP_ARGS, )==""\n"
179R"==(__global float *src0_scale, __global float *src1_scale) { )==""\n"
180R"==(int dims0[6] = {0}; )==""\n"
181R"==(dims0[0] = GWS_GET_D0(); )==""\n"
182R"==(dims0[1] = GWS_GET_D1(); )==""\n"
183R"==(dims0[2] = GWS_GET_D2(); )==""\n"
184R"==(dims0[3] = GWS_GET_D3(); )==""\n"
185R"==(dims0[4] = GWS_GET_D4(); )==""\n"
186R"==(dims0[5] = GWS_GET_D5(); )==""\n"
187R"==(int src0_off = SRC0_OFF( )==""\n"
188R"==(dims0[0], dims0[1], dims0[2], dims0[3], dims0[4], dims0[5]); )==""\n"
189R"==(int dst_off = DST_OFF( )==""\n"
190R"==(dims0[0], dims0[1], dims0[2], dims0[3], dims0[4], dims0[5]); )==""\n"
191R"==(int src1_off = SRC1_OFF( )==""\n"
192R"==(dims0[0], dims0[1], dims0[2], dims0[3], dims0[4], dims0[5]); )==""\n"
193R"==(int sub_grp_id = get_sub_group_local_id(); )==""\n"
194R"==(for (int channels = 0; channels < SRC0_PD1; channels += GWS_LWS0_DEFAULT) { )==""\n"
195R"==(float8 d = 0; )==""\n"
196R"==(float8 dst_data; )==""\n"
197R"==(__global SRC1_DATA_T *t_src1 = src1 + src1_off; )==""\n"
198R"==(__global DST_DATA_T *t_dst = dst + dst_off; )==""\n"
199R"==(__global SRC0_DATA_T *t_src0 = src0 + src0_off; )==""\n"
200R"==(if ((SRC0_D1 % SUB_GROUP_SIZE != 0) )==""\n"
201R"==(&& (dims0[1] + sub_grp_id) >= SRC0_D1) { )==""\n"
202R"==(d = 0; )==""\n"
203R"==(} else { )==""\n"
204R"==(float8 tmp_src0 = CONVERT_FLOAT8_T(SRC0_BLOCK_READ8(&t_src0[0])); )==""\n"
205R"==(float8 tmp_src1 = CONVERT_FLOAT8_T(SRC1_BLOCK_READ8(&t_src1[0])); )==""\n"
206R"==(#if WITH_SRC0_SCALE )==""\n"
207R"==(tmp_src0 = tmp_src0 * src0_scale[0]; )==""\n"
208R"==(#endif )==""\n"
209R"==(#if WITH_SRC1_SCALE )==""\n"
210R"==(tmp_src1 = tmp_src1 * src1_scale[0]; )==""\n"
211R"==(#endif )==""\n"
212R"==(d = get_eltwise_op(tmp_src0, tmp_src1); )==""\n"
213R"==(#if WITH_SUM )==""\n"
214R"==(dst_data = CONVERT_FLOAT8_T(DST_BLOCK_READ8(&t_dst[0])); )==""\n"
215R"==(#endif )==""\n"
216R"==(const int po_mb = dims0[0]; )==""\n"
217R"==(const int po_oc = dims0[1] + sub_grp_id; )==""\n"
218R"==(for (int lcl_mb = 0; lcl_mb < NVECT; ++lcl_mb) { )==""\n"
219R"==(if (po_mb + lcl_mb >= SRC0_D0) { )==""\n"
220R"==(d[lcl_mb] = 0; )==""\n"
221R"==(} else { )==""\n"
222R"==(float d_i = d[lcl_mb]; )==""\n"
223R"==(float dst_i = dst_data[lcl_mb]; )==""\n"
224R"==(APPLY_POST_OPS_SERIAL(d_i, float, dst_i, float, )==""\n"
225R"==(po_mb + lcl_mb, 1, po_oc, 1, dims0[2], 1, dims0[3], )==""\n"
226R"==(1, dims0[4], 1, dims0[5], 1); )==""\n"
227R"==(d[lcl_mb] = d_i; )==""\n"
228R"==(} )==""\n"
229R"==(} )==""\n"
230R"==(} )==""\n"
231R"==(DST_BLOCK_WRITE8(&t_dst[0], TO_DST8(d)); )==""\n"
232R"==(src0_off += MB_BLOCK * SUB_GROUP_SIZE * SRC0_PD2 * SRC0_PD3 * SRC0_PD4 )==""\n"
233R"==(* SRC0_PD5; )==""\n"
234R"==(src1_off += MB_BLOCK * SUB_GROUP_SIZE * SRC1_PD2 * SRC1_PD3 * SRC1_PD4 )==""\n"
235R"==(* SRC1_PD5; )==""\n"
236R"==(dst_off += MB_BLOCK * SUB_GROUP_SIZE * DST_PD2 * DST_PD3 * DST_PD4 )==""\n"
237R"==(* DST_PD5; )==""\n"
238R"==(} )==""\n"
239R"==(} )==""\n"
240R"==(#else )==""\n"
241R"==(KERNEL_ATTR )==""\n"
242R"==(__kernel void gen9_binary(__global SRC0_DATA_T *src0, )==""\n"
243R"==(__global SRC1_DATA_T *src1, __global DST_DATA_T *dst POST_OP_ARGS, )==""\n"
244R"==(__global float *src0_scale, __global float *src1_scale) { )==""\n"
245R"==(int dims0[6] = {0}; )==""\n"
246R"==(dims0[0] = GWS_GET_D0(); )==""\n"
247R"==(dims0[1] = GWS_GET_D1(); )==""\n"
248R"==(dims0[2] = GWS_GET_D2(); )==""\n"
249R"==(dims0[3] = GWS_GET_D3(); )==""\n"
250R"==(dims0[4] = GWS_GET_D4(); )==""\n"
251R"==(dims0[5] = GWS_GET_D5(); )==""\n"
252R"==(int src0_off = SRC0_OFF( )==""\n"
253R"==(dims0[0], dims0[1], dims0[2], dims0[3], dims0[4], dims0[5]); )==""\n"
254R"==(src0 += src0_off; )==""\n"
255R"==(int dst_off = DST_OFF( )==""\n"
256R"==(dims0[0], dims0[1], dims0[2], dims0[3], dims0[4], dims0[5]); )==""\n"
257R"==(dst += dst_off; )==""\n"
258R"==(int src1_off = SRC1_OFF(dims0[0] * (!BCAST_DIM0), dims0[1] * (!BCAST_DIM1), )==""\n"
259R"==(dims0[2] * (!BCAST_DIM2), dims0[3] * (!BCAST_DIM3), )==""\n"
260R"==(dims0[4] * (!BCAST_DIM4), dims0[5] * (!BCAST_DIM5)); )==""\n"
261R"==(src1 += src1_off; )==""\n"
262R"==(#if NVECT == 1 )==""\n"
263R"==(float d = 0; )==""\n"
264R"==(float dst_data; )==""\n"
265R"==(float tmp_src0 = CONVERT_FLOAT_T(SRC0_BLOCK_READ(&src0[0])); )==""\n"
266R"==(#elif NVECT == 2 )==""\n"
267R"==(float2 d = 0; )==""\n"
268R"==(float2 dst_data; )==""\n"
269R"==(float2 tmp_src0 = CONVERT_FLOAT2_T(SRC0_BLOCK_READ2(&src0[0])); )==""\n"
270R"==(#elif NVECT == 4 )==""\n"
271R"==(float4 d = 0; )==""\n"
272R"==(float4 dst_data; )==""\n"
273R"==(float4 tmp_src0 = CONVERT_FLOAT4_T(SRC0_BLOCK_READ4(&src0[0])); )==""\n"
274R"==(#elif NVECT == 8 )==""\n"
275R"==(float8 d = 0; )==""\n"
276R"==(float8 dst_data; )==""\n"
277R"==(float8 tmp_src0 = CONVERT_FLOAT8_T(SRC0_BLOCK_READ8(&src0[0])); )==""\n"
278R"==(#endif )==""\n"
279R"==(#if BCAST_DIM1 )==""\n"
280R"==(float tmp_src1 = CONVERT_FLOAT_T(src1[0]); )==""\n"
281R"==(#else )==""\n"
282R"==(#if BCAST_AT_INNERMOST_DIM == 1 || NVECT == 1 )==""\n"
283R"==(float tmp_src1 = CONVERT_FLOAT_T(SRC1_BLOCK_READ(&src1[0])); )==""\n"
284R"==(#elif NVECT == 2 )==""\n"
285R"==(float2 tmp_src1 = CONVERT_FLOAT2_T(SRC1_BLOCK_READ2(&src1[0])); )==""\n"
286R"==(#elif NVECT == 4 )==""\n"
287R"==(float4 tmp_src1 = CONVERT_FLOAT4_T(SRC1_BLOCK_READ4(&src1[0])); )==""\n"
288R"==(#elif NVECT == 8 )==""\n"
289R"==(float8 tmp_src1 = CONVERT_FLOAT8_T(SRC1_BLOCK_READ8(&src1[0])); )==""\n"
290R"==(#endif )==""\n"
291R"==(#endif )==""\n"
292R"==(#if WITH_SRC0_SCALE )==""\n"
293R"==(tmp_src0 = tmp_src0 * src0_scale[0]; )==""\n"
294R"==(#endif )==""\n"
295R"==(#if WITH_SRC1_SCALE )==""\n"
296R"==(tmp_src1 = tmp_src1 * src1_scale[0]; )==""\n"
297R"==(#endif )==""\n"
298R"==(d = get_eltwise_op(tmp_src0, tmp_src1); )==""\n"
299R"==(#if WITH_SUM )==""\n"
300R"==(#if NVECT == 1 )==""\n"
301R"==(dst_data = CONVERT_FLOAT_T(DST_BLOCK_READ(&dst[0])); )==""\n"
302R"==(#elif NVECT == 2 )==""\n"
303R"==(dst_data = CONVERT_FLOAT2_T(DST_BLOCK_READ2(&dst[0])); )==""\n"
304R"==(#elif NVECT == 4 )==""\n"
305R"==(dst_data = CONVERT_FLOAT4_T(DST_BLOCK_READ4(&dst[0])); )==""\n"
306R"==(#elif NVECT == 8 )==""\n"
307R"==(dst_data = CONVERT_FLOAT8_T(DST_BLOCK_READ8(&dst[0])); )==""\n"
308R"==(#endif )==""\n"
309R"==(#endif )==""\n"
310R"==(const int po_mb = dims0[0]; )==""\n"
311R"==(const int po_oc = dims0[1] + get_sub_group_local_id(); )==""\n"
312R"==(#if NVECT == 1 )==""\n"
313R"==(APPLY_POST_OPS_SERIAL(d, float, dst_data, float, po_mb, 1, po_oc, 1, )==""\n"
314R"==(dims0[2], 1, dims0[3], 1, dims0[4], 1, dims0[5], 1); )==""\n"
315R"==(#else )==""\n"
316R"==(for (int vidx = 0; vidx < NVECT; ++vidx) { )==""\n"
317R"==(float d_i = d[vidx]; )==""\n"
318R"==(float dst_i = dst_data[vidx]; )==""\n"
319R"==(APPLY_POST_OPS_SERIAL(d_i, float, dst_i, float, po_mb, 1, po_oc, 1, )==""\n"
320R"==(dims0[2], 1, dims0[3], 1, dims0[4], 1, dims0[5], 1); )==""\n"
321R"==(d[vidx] = d_i; )==""\n"
322R"==(++dims0[NDIMS - 1]; )==""\n"
323R"==(} )==""\n"
324R"==(#endif )==""\n"
325R"==(#if NVECT == 1 )==""\n"
326R"==(DST_BLOCK_WRITE(&dst[0], TO_DST(d)); )==""\n"
327R"==(#elif NVECT == 2 )==""\n"
328R"==(DST_BLOCK_WRITE2(&dst[0], TO_DST2(d)); )==""\n"
329R"==(#elif NVECT == 4 )==""\n"
330R"==(DST_BLOCK_WRITE4(&dst[0], TO_DST4(d)); )==""\n"
331R"==(#elif NVECT == 8 )==""\n"
332R"==(DST_BLOCK_WRITE8(&dst[0], TO_DST8(d)); )==""\n"
333R"==(#endif )==""\n"
334R"==(} )==""\n"
335R"==(#endif )==""\n"
336R"==()==";
337}
338}
339}
340}