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
|
// =================================================================================================
// 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 3 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"(
// =================================================================================================
// Main body of the kernel. This is the direct version without pre/post processing and restrictions.
INLINE_FUNC void XgemmDirect(const int kSizeM, const int kSizeN, const int kSizeK,
const real_arg arg_alpha,
const real_arg arg_beta,
const __global realMD* restrict agm, const int a_offset, const int a_ld,
const __global realND* restrict bgm, const int b_offset, const int b_ld,
__global real* cgm, const int c_offset, const int c_ld,
LOCAL_PTR real* alm, LOCAL_PTR real* blm,
const int a_transpose, const int b_transpose, const int c_transpose,
const int a_conjugate, const int b_conjugate) {
const real alpha = GetRealArg(arg_alpha);
const real beta = GetRealArg(arg_beta);
// Extra pointers to scalar versions of global memory
const __global real* restrict agms = (const __global real* restrict) agm;
const __global real* restrict bgms = (const __global real* restrict) bgm;
// Allocates workitem-private memory (registers)
#pragma promote_to_registers
real apd[MWID];
#pragma promote_to_registers
real bpd[NWID];
#pragma promote_to_registers
real cpd[NWID * MWID];
// Initializes the accumulation registers
#pragma unroll
for (int _mi = 0; _mi < MWID; _mi += 1) {
#pragma unroll
for (int _ni = 0; _ni < NWID; _ni += 1) {
SetToZero(cpd[_ni * MWID + _mi]);
}
}
// The faster version of GEMM is not allowed on the (incomplete) borders. Therefore, this section
// processes only the main parts: output blocks of WGD by WGD.
const int idm = get_local_id(0) * MWID + GetGroupID0() * WGD;
const int idn = get_local_id(1) * NWID + GetGroupID1() * WGD;
if ((idm < (kSizeM/WGD)*WGD) && (idn < (kSizeN/WGD)*WGD)) {
// Loops over all complete workgroup tiles (K-dimension)
int kwg = 0;
for (; kwg < (kSizeK/WGD) * WGD; kwg += WGD) {
// Loads data: off-chip --> local (matrix A and B)
if (a_ld % VWMD == 0 && a_offset % VWMD == 0) {
GlobalToLocalDirectA(agm, alm, a_ld, a_offset, kwg, a_transpose, a_conjugate);
}
else {
GlobalToLocalScalarA(agms, alm, a_ld, a_offset, kwg, a_transpose, a_conjugate);
}
if (b_ld % VWND == 0 && b_offset % VWND == 0) {
GlobalToLocalDirectB(bgm, blm, b_ld, b_offset, kwg, b_transpose, b_conjugate);
}
else {
GlobalToLocalScalarB(bgms, blm, b_ld, b_offset, kwg, b_transpose, b_conjugate);
}
barrier(CLK_LOCAL_MEM_FENCE);
// Loops over all workitem tiles, unrolled by a factor KWID
for (int pwi = 0; pwi < WGD; pwi += KWID) {
#pragma unroll
for (int _pit = 0; _pit < KWID; _pit += 1) {
int kg = pwi + _pit;
// Loads data: local --> private (matrix A and B)
#pragma unroll
for (int _mi = 0; _mi < MWID; _mi += 1) {
apd[_mi] = LocalToPrivateDirectA(alm, _mi, kg, a_transpose);
}
#pragma unroll
for (int _ni = 0; _ni < NWID; _ni += 1) {
bpd[_ni] = LocalToPrivateDirectB(blm, _ni, kg, b_transpose);
}
// Performs the accumulation (Cpmd += Apmd * Bpmd)
#pragma unroll
for (int _ni = 0; _ni < NWID; _ni += 1) {
#pragma unroll
for (int _mi = 0; _mi < MWID; _mi += 1) {
MultiplyAdd(cpd[_ni * MWID + _mi], apd[_mi], bpd[_ni]);
}
}
}
}
barrier(CLK_LOCAL_MEM_FENCE);
}
// Loop over the remaining part (incomplete tile in K-dimension)
for (; kwg < kSizeK; ++kwg) {
// Loads data: off-chip --> private (matrix A and B)
#pragma unroll
for (int _mi = 0; _mi < MWID; _mi += 1) {
apd[_mi] = GlobalToPrivateDirectA(agms, _mi, a_ld, a_offset, idm, kwg, a_transpose, a_conjugate);
}
#pragma unroll
for (int _ni = 0; _ni < NWID; _ni += 1) {
bpd[_ni] = GlobalToPrivateDirectB(bgms, _ni, b_ld, b_offset, idn, kwg, b_transpose, b_conjugate);
}
// Performs the accumulation (Cpmd += Apmd * Bpmd)
#pragma unroll
for (int _ni = 0; _ni < NWID; _ni += 1) {
#pragma unroll
for (int _mi = 0; _mi < MWID; _mi += 1) {
MultiplyAdd(cpd[_ni * MWID + _mi], apd[_mi], bpd[_ni]);
}
}
}
// Stores a tile of results and performs the multiplication with alpha and beta
#pragma unroll
for (int _ni = 0; _ni < NWID; _ni += 1) {
#pragma unroll
for (int _mi = 0; _mi < MWID; _mi += 1) {
StoreResultsDirect(cgm, cpd[_ni * MWID + _mi], _mi, _ni, idm, idn,
alpha, beta, c_ld, c_offset, c_transpose);
}
}
}
// Simple but slower version for the parts on the edge (incomplete tiles in M and N-dimensions)
else {
// Loops over all complete workgroup tiles (K-dimension)
int kwg = 0;
for (; kwg < (kSizeK/WGD) * WGD; kwg+=WGD) {
// Loads data: off-chip --> local (matrix A and B)
GlobalToLocalCheckedA(agms, alm, a_ld, a_offset, kwg, a_transpose, a_conjugate, kSizeM, kSizeK);
GlobalToLocalCheckedB(bgms, blm, b_ld, b_offset, kwg, b_transpose, b_conjugate, kSizeN, kSizeK);
barrier(CLK_LOCAL_MEM_FENCE);
// Loops over all workitem tiles, unrolled by a factor KWID
for (int pwi = 0; pwi < WGD; pwi += KWID) {
#pragma unroll
for (int _pit = 0; _pit < KWID; _pit += 1) {
int kg = pwi + _pit;
// Loads data: local --> private (matrix A and B)
#pragma unroll
for (int _mi = 0; _mi < MWID; _mi += 1) {
apd[_mi] = LocalToPrivateDirectA(alm, _mi, kg, a_transpose);
}
#pragma unroll
for (int _ni = 0; _ni < NWID; _ni += 1) {
bpd[_ni] = LocalToPrivateDirectB(blm, _ni, kg, b_transpose);
}
// Performs the accumulation (C += A * B)
#pragma unroll
for (int _ni = 0; _ni < NWID; _ni += 1) {
#pragma unroll
for (int _mi = 0; _mi < MWID; _mi += 1) {
MultiplyAdd(cpd[_ni * MWID + _mi], apd[_mi], bpd[_ni]);
}
}
}
}
barrier(CLK_LOCAL_MEM_FENCE);
}
// Loop over the remaining part (incomplete tile in K-dimension)
for (; kwg < kSizeK; ++kwg) {
// Loads data: off-chip --> private (matrix A and B)
#pragma unroll
for (int _mi = 0; _mi < MWID; _mi += 1) {
apd[_mi] = GlobalToPrivateCheckedA(agms, _mi, a_ld, a_offset, idm, kwg, a_transpose, a_conjugate, kSizeM);
}
#pragma unroll
for (int _ni = 0; _ni < NWID; _ni += 1) {
bpd[_ni] = GlobalToPrivateCheckedB(bgms, _ni, b_ld, b_offset, idn, kwg, b_transpose, b_conjugate, kSizeN);
}
// Performs the accumulation (C += A * B)
#pragma unroll
for (int _ni = 0; _ni < NWID; _ni += 1) {
#pragma unroll
for (int _mi = 0; _mi < MWID; _mi += 1) {
MultiplyAdd(cpd[_ni * MWID + _mi], apd[_mi], bpd[_ni]);
}
}
}
// Stores a tile of results and performs the multiplication with alpha and beta
#pragma unroll
for (int _ni = 0; _ni < NWID; _ni += 1) {
#pragma unroll
for (int _mi = 0; _mi < MWID; _mi += 1) {
StoreResultsChecked(cgm, cpd[_ni * MWID + _mi], _mi, _ni, idm, idn, kSizeM, kSizeN,
alpha, beta, c_ld, c_offset, c_transpose);
}
}
}
}
// =================================================================================================
// Direct version of the GEMM kernel with [A, B] = [non-transposed, non-transposed]
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#else
__kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1)))
#endif
void XgemmDirectNN(const int kSizeM, const int kSizeN, const int kSizeK,
const real_arg arg_alpha, const real_arg arg_beta,
const __global realMD* restrict agm, const int a_offset, const int a_ld,
const __global realND* restrict bgm, const int b_offset, const int b_ld,
__global real* cgm, const int c_offset, const int c_ld,
const int c_transpose, const int a_conjugate, const int b_conjugate) {
__local real alm[WGD * (WGD + PADA)];
__local real blm[WGD * (WGD + PADB)];
XgemmDirect(kSizeM, kSizeN, kSizeK, arg_alpha, arg_beta,
agm, a_offset, a_ld, bgm, b_offset, b_ld, cgm, c_offset, c_ld,
alm, blm, 0, 0, c_transpose, a_conjugate, b_conjugate);
}
// Direct version of the GEMM kernel with [A, B] = [non-transposed, transposed]
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#else
__kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1)))
#endif
void XgemmDirectNT(const int kSizeM, const int kSizeN, const int kSizeK,
const real_arg arg_alpha, const real_arg arg_beta,
const __global realMD* restrict agm, const int a_offset, const int a_ld,
const __global realND* restrict bgm, const int b_offset, const int b_ld,
__global real* cgm, const int c_offset, const int c_ld,
const int c_transpose, const int a_conjugate, const int b_conjugate) {
__local real alm[WGD * (WGD + PADA)];
__local real blm[WGD * (WGD + PADB)];
XgemmDirect(kSizeM, kSizeN, kSizeK, arg_alpha, arg_beta,
agm, a_offset, a_ld, bgm, b_offset, b_ld, cgm, c_offset, c_ld,
alm, blm, 0, 1, c_transpose, a_conjugate, b_conjugate);
}
// Direct version of the GEMM kernel with [A, B] = [transposed, non-transposed]
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#else
__kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1)))
#endif
void XgemmDirectTN(const int kSizeM, const int kSizeN, const int kSizeK,
const real_arg arg_alpha, const real_arg arg_beta,
const __global realMD* restrict agm, const int a_offset, const int a_ld,
const __global realND* restrict bgm, const int b_offset, const int b_ld,
__global real* cgm, const int c_offset, const int c_ld,
const int c_transpose, const int a_conjugate, const int b_conjugate) {
__local real alm[WGD * (WGD + PADA)];
__local real blm[WGD * (WGD + PADB)];
XgemmDirect(kSizeM, kSizeN, kSizeK, arg_alpha, arg_beta,
agm, a_offset, a_ld, bgm, b_offset, b_ld, cgm, c_offset, c_ld,
alm, blm, 1, 0, c_transpose, a_conjugate, b_conjugate);
}
// Direct version of the GEMM kernel with [A, B] = [transposed, transposed]
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#else
__kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1)))
#endif
void XgemmDirectTT(const int kSizeM, const int kSizeN, const int kSizeK,
const real_arg arg_alpha, const real_arg arg_beta,
const __global realMD* restrict agm, const int a_offset, const int a_ld,
const __global realND* restrict bgm, const int b_offset, const int b_ld,
__global real* cgm, const int c_offset, const int c_ld,
const int c_transpose, const int a_conjugate, const int b_conjugate) {
__local real alm[WGD * (WGD + PADA)];
__local real blm[WGD * (WGD + PADB)];
XgemmDirect(kSizeM, kSizeN, kSizeK, arg_alpha, arg_beta,
agm, a_offset, a_ld, bgm, b_offset, b_ld, cgm, c_offset, c_ld,
alm, blm, 1, 1, c_transpose, a_conjugate, b_conjugate);
}
// =================================================================================================
// End of the C++11 raw string literal
)"
// =================================================================================================
|
