1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
|
// =================================================================================================
// This file is part of the CLBlast project. The project is licensed under Apache Version 2.0. This
// project loosely follows the Google C++ styleguide and uses a tab-size of two spaces and a max-
// width of 100 characters per line.
//
// Author(s):
// Cedric Nugteren <www.cedricnugteren.nl>
//
// This is part 2 of 3 of the GEMM kernel. See part 1 for more information.
//
// =================================================================================================
// Enables loading of this file using the C++ pre-processor's #include (C++11 standard raw string
// literal). Comment-out this line for syntax-highlighting when developing.
R"(
// =================================================================================================
// Caches global off-chip memory into local (shared) memory on-chip. This function is specific for
// caching the A input matrix.
inline void GlobalToLocalDirectA(const __global realMD* restrict agm, __local real* alm,
const int a_ld, const int a_offset, const int kwg,
const int a_transpose, const int a_conjugate) {
#if MDIMCD == MDIMAD
const int la0 = get_local_id(0);
const int la1 = get_local_id(1);
#else
const int tid = get_local_id(0) + MDIMCD*get_local_id(1);
const int la0 = tid % MDIMAD;
const int la1 = tid / MDIMAD;
#endif
#pragma unroll
for (int mia=0; mia<MWAD/VWMD; ++mia) {
#pragma unroll
for (int kia=0; kia<KWAD; ++kia) {
// Computes the indices for the global memory
int mg = mia + la0*(MWAD/VWMD);
int kg = kia + la1*KWAD;
int idm = (a_transpose) ? mg + kwg/VWMD : mg + GetGroupID0()*(WGD/VWMD);
int idk = (a_transpose) ? kg + GetGroupID0()*WGD : kg + kwg;
// Loads the data from global memory into the local memory
const realMD avec = agm[idk*(a_ld/VWMD) + idm + (a_offset/VWMD)];
#if VWMD == 1
alm[kg*(WGD + PADA) + mg] = avec;
#elif VWMD == 2
alm[kg*(WGD + PADA) + mg*VWMD + 0] = avec.x;
alm[kg*(WGD + PADA) + mg*VWMD + 1] = avec.y;
#elif VWMD == 4
alm[kg*(WGD + PADA) + mg*VWMD + 0] = avec.x;
alm[kg*(WGD + PADA) + mg*VWMD + 1] = avec.y;
alm[kg*(WGD + PADA) + mg*VWMD + 2] = avec.z;
alm[kg*(WGD + PADA) + mg*VWMD + 3] = avec.w;
#elif VWMD == 8
alm[kg*(WGD + PADA) + mg*VWMD + 0] = avec.s0;
alm[kg*(WGD + PADA) + mg*VWMD + 1] = avec.s1;
alm[kg*(WGD + PADA) + mg*VWMD + 2] = avec.s2;
alm[kg*(WGD + PADA) + mg*VWMD + 3] = avec.s3;
alm[kg*(WGD + PADA) + mg*VWMD + 4] = avec.s4;
alm[kg*(WGD + PADA) + mg*VWMD + 5] = avec.s5;
alm[kg*(WGD + PADA) + mg*VWMD + 6] = avec.s6;
alm[kg*(WGD + PADA) + mg*VWMD + 7] = avec.s7;
#elif VWMD == 16
alm[kg*(WGD + PADA) + mg*VWMD + 0] = avec.s0;
alm[kg*(WGD + PADA) + mg*VWMD + 1] = avec.s1;
alm[kg*(WGD + PADA) + mg*VWMD + 2] = avec.s2;
alm[kg*(WGD + PADA) + mg*VWMD + 3] = avec.s3;
alm[kg*(WGD + PADA) + mg*VWMD + 4] = avec.s4;
alm[kg*(WGD + PADA) + mg*VWMD + 5] = avec.s5;
alm[kg*(WGD + PADA) + mg*VWMD + 6] = avec.s6;
alm[kg*(WGD + PADA) + mg*VWMD + 7] = avec.s7;
alm[kg*(WGD + PADA) + mg*VWMD + 8] = avec.s8;
alm[kg*(WGD + PADA) + mg*VWMD + 9] = avec.s9;
alm[kg*(WGD + PADA) + mg*VWMD + 10] = avec.sA;
alm[kg*(WGD + PADA) + mg*VWMD + 11] = avec.sB;
alm[kg*(WGD + PADA) + mg*VWMD + 12] = avec.sC;
alm[kg*(WGD + PADA) + mg*VWMD + 13] = avec.sD;
alm[kg*(WGD + PADA) + mg*VWMD + 14] = avec.sE;
alm[kg*(WGD + PADA) + mg*VWMD + 15] = avec.sF;
#endif
if (a_conjugate) {
for (int vm=0; vm<VWMD; ++vm) {
COMPLEX_CONJUGATE(alm[kg*(WGD + PADA) + mg*VWMD + vm]);
}
}
}
}
}
// Same as above, but now for the B input matrix
inline void GlobalToLocalDirectB(const __global realND* restrict bgm, __local real* blm,
const int b_ld, const int b_offset, const int kwg,
const int b_transpose, const int b_conjugate) {
#if MDIMCD == NDIMBD
const int lb0 = get_local_id(0);
const int lb1 = get_local_id(1);
#else
const int tid = get_local_id(0) + MDIMCD*get_local_id(1);
const int lb0 = tid % NDIMBD;
const int lb1 = tid / NDIMBD;
#endif
#pragma unroll
for (int kib=0; kib<KWBD; ++kib) {
#pragma unroll
for (int nib=0; nib<NWBD/VWND; ++nib) {
// Computes the indices for the global memory
int ng = nib + lb0*(NWBD/VWND);
int kg = kib + lb1*KWBD;
int idn = (b_transpose) ? ng + kwg/VWND : ng + GetGroupID1()*(WGD/VWND);
int idk = (b_transpose) ? kg + GetGroupID1()*WGD : kg + kwg;
// Loads the data from global memory into the local memory
const realND bvec = bgm[idk*(b_ld/VWND) + idn + (b_offset/VWND)];
#if VWND == 1
blm[kg*(WGD + PADB) + ng] = bvec;
#elif VWND == 2
blm[kg*(WGD + PADB) + ng*VWND + 0] = bvec.x;
blm[kg*(WGD + PADB) + ng*VWND + 1] = bvec.y;
#elif VWND == 4
blm[kg*(WGD + PADB) + ng*VWND + 0] = bvec.x;
blm[kg*(WGD + PADB) + ng*VWND + 1] = bvec.y;
blm[kg*(WGD + PADB) + ng*VWND + 2] = bvec.z;
blm[kg*(WGD + PADB) + ng*VWND + 3] = bvec.w;
#elif VWND == 8
blm[kg*(WGD + PADB) + ng*VWND + 0] = bvec.s0;
blm[kg*(WGD + PADB) + ng*VWND + 1] = bvec.s1;
blm[kg*(WGD + PADB) + ng*VWND + 2] = bvec.s2;
blm[kg*(WGD + PADB) + ng*VWND + 3] = bvec.s3;
blm[kg*(WGD + PADB) + ng*VWND + 4] = bvec.s4;
blm[kg*(WGD + PADB) + ng*VWND + 5] = bvec.s5;
blm[kg*(WGD + PADB) + ng*VWND + 6] = bvec.s6;
blm[kg*(WGD + PADB) + ng*VWND + 7] = bvec.s7;
#elif VWND == 16
blm[kg*(WGD + PADB) + ng*VWND + 0] = bvec.s0;
blm[kg*(WGD + PADB) + ng*VWND + 1] = bvec.s1;
blm[kg*(WGD + PADB) + ng*VWND + 2] = bvec.s2;
blm[kg*(WGD + PADB) + ng*VWND + 3] = bvec.s3;
blm[kg*(WGD + PADB) + ng*VWND + 4] = bvec.s4;
blm[kg*(WGD + PADB) + ng*VWND + 5] = bvec.s5;
blm[kg*(WGD + PADB) + ng*VWND + 6] = bvec.s6;
blm[kg*(WGD + PADB) + ng*VWND + 7] = bvec.s7;
blm[kg*(WGD + PADB) + ng*VWND + 8] = bvec.s8;
blm[kg*(WGD + PADB) + ng*VWND + 9] = bvec.s9;
blm[kg*(WGD + PADB) + ng*VWND + 10] = bvec.sA;
blm[kg*(WGD + PADB) + ng*VWND + 11] = bvec.sB;
blm[kg*(WGD + PADB) + ng*VWND + 12] = bvec.sC;
blm[kg*(WGD + PADB) + ng*VWND + 13] = bvec.sD;
blm[kg*(WGD + PADB) + ng*VWND + 14] = bvec.sE;
blm[kg*(WGD + PADB) + ng*VWND + 15] = bvec.sF;
#endif
if (b_conjugate) {
for (int vn=0; vn<VWND; ++vn) {
COMPLEX_CONJUGATE(blm[kg*(WGD + PADB) + ng*VWND + vn]);
}
}
}
}
}
// =================================================================================================
// Caches global off-chip memory into local (shared) memory on-chip. This function is specific for
// caching the A input matrix. In contrast to the functions above, this function performs doesn't
// use the vector data-types.
inline void GlobalToLocalScalarA(const __global real* restrict agms, __local real* alm,
const int a_ld, const int a_offset, const int kwg,
const int a_transpose, const int a_conjugate) {
#if MDIMCD == MDIMAD
const int la0 = get_local_id(0);
const int la1 = get_local_id(1);
#else
const int tid = get_local_id(0) + MDIMCD*get_local_id(1);
const int la0 = tid % MDIMAD;
const int la1 = tid / MDIMAD;
#endif
#pragma unroll
for (int mia=0; mia<MWAD; ++mia) {
#pragma unroll
for (int kia=0; kia<KWAD; ++kia) {
// Computes the indices for the global memory
int mg = mia + la0*MWAD;
int kg = kia + la1*KWAD;
int idm = (a_transpose) ? mg + kwg : mg + GetGroupID0()*WGD;
int idk = (a_transpose) ? kg + GetGroupID0()*WGD : kg + kwg;
// Loads the data from global memory into the local memory
real result = agms[idk*a_ld + idm + a_offset];
if (a_conjugate) { COMPLEX_CONJUGATE(result); }
alm[kg*(WGD + PADA) + mg] = result;
}
}
}
// Same as above, but now for the B input matrix
inline void GlobalToLocalScalarB(const __global real* restrict bgms, __local real* blm,
const int b_ld, const int b_offset, const int kwg,
const int b_transpose, const int b_conjugate) {
#if MDIMCD == NDIMBD
const int lb0 = get_local_id(0);
const int lb1 = get_local_id(1);
#else
const int tid = get_local_id(0) + MDIMCD*get_local_id(1);
const int lb0 = tid % NDIMBD;
const int lb1 = tid / NDIMBD;
#endif
#pragma unroll
for (int kib=0; kib<KWBD; ++kib) {
#pragma unroll
for (int nib=0; nib<NWBD; ++nib) {
// Computes the indices for the global memory
int ng = nib + lb0*NWBD;
int kg = kib + lb1*KWBD;
int idn = (b_transpose) ? ng + kwg : ng + GetGroupID1()*WGD;
int idk = (b_transpose) ? kg + GetGroupID1()*WGD : kg + kwg;
// Loads the data from global memory into the local memory
real result = bgms[idk*b_ld + idn + b_offset];
if (b_conjugate) { COMPLEX_CONJUGATE(result); }
blm[kg*(WGD + PADB) + ng] = result;
}
}
}
// =================================================================================================
// Caches global off-chip memory into local (shared) memory on-chip. This function is specific for
// caching the A input matrix. In contrast to the functions above, this function performs bounds
// checks and doesn't use the vector data-types.
inline void GlobalToLocalCheckedA(const __global real* restrict agms, __local real* alm,
const int a_ld, const int a_offset, const int kwg,
const int a_transpose, const int a_conjugate,
const int kSizeM, const int kSizeK) {
#if MDIMCD == MDIMAD
const int la0 = get_local_id(0);
const int la1 = get_local_id(1);
#else
const int tid = get_local_id(0) + MDIMCD*get_local_id(1);
const int la0 = tid % MDIMAD;
const int la1 = tid / MDIMAD;
#endif
#pragma unroll
for (int mia=0; mia<MWAD; ++mia) {
#pragma unroll
for (int kia=0; kia<KWAD; ++kia) {
// Computes the indices for the global memory
int mg = mia + la0*MWAD;
int kg = kia + la1*KWAD;
int idm = (a_transpose) ? mg + kwg : mg + GetGroupID0()*WGD;
int idk = (a_transpose) ? kg + GetGroupID0()*WGD : kg + kwg;
// Loads the data from global memory into the local memory
int condition = (a_transpose) ? (idm < kSizeK) && (idk < kSizeM) :
(idm < kSizeM) && (idk < kSizeK);
if (condition) {
real result = agms[idk*a_ld + idm + a_offset];
if (a_conjugate) { COMPLEX_CONJUGATE(result); }
alm[kg*(WGD + PADA) + mg] = result;
}
else {
SetToZero(alm[kg*(WGD + PADA) + mg]);
}
}
}
}
// Same as above, but now for the B input matrix
inline void GlobalToLocalCheckedB(const __global real* restrict bgms, __local real* blm,
const int b_ld, const int b_offset, const int kwg,
const int b_transpose, const int b_conjugate,
const int kSizeN, const int kSizeK) {
#if MDIMCD == NDIMBD
const int lb0 = get_local_id(0);
const int lb1 = get_local_id(1);
#else
const int tid = get_local_id(0) + MDIMCD*get_local_id(1);
const int lb0 = tid % NDIMBD;
const int lb1 = tid / NDIMBD;
#endif
#pragma unroll
for (int kib=0; kib<KWBD; ++kib) {
#pragma unroll
for (int nib=0; nib<NWBD; ++nib) {
// Computes the indices for the global memory
int ng = nib + lb0*NWBD;
int kg = kib + lb1*KWBD;
int idn = (b_transpose) ? ng + kwg : ng + GetGroupID1()*WGD;
int idk = (b_transpose) ? kg + GetGroupID1()*WGD : kg + kwg;
// Loads the data from global memory into the local memory
int condition = (b_transpose) ? (idn < kSizeK) && (idk < kSizeN) :
(idn < kSizeN) && (idk < kSizeK);
if (condition) {
real result = bgms[idk*b_ld + idn + b_offset];
if (b_conjugate) { COMPLEX_CONJUGATE(result); }
blm[kg*(WGD + PADB) + ng] = result;
}
else {
SetToZero(blm[kg*(WGD + PADB) + ng]);
}
}
}
}
// =================================================================================================
// End of the C++11 raw string literal
)"
// =================================================================================================
|
