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
|
// =================================================================================================
// 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 file contains the batched version of the direct GEMM kernels. See part 1 for information
// about the non-batched version of the kernel.
//
// =================================================================================================
// 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"(
// =================================================================================================
#if defined(ROUTINE_GEMMBATCHED)
// Direct version of the batched GEMM kernel with [A, B] = [non-transposed, non-transposed]
__kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1)))
void XgemmDirectBatchedNN(const int kSizeM, const int kSizeN, const int kSizeK,
const __constant real_arg* arg_alphas, const __constant real_arg* arg_betas,
const __global realMD* restrict agm, const __constant int* a_offsets, const int a_ld,
const __global realND* restrict bgm, const __constant int* b_offsets, const int b_ld,
__global real* cgm, const __constant int* c_offsets, const int c_ld,
const int c_transpose, const int a_conjugate, const int b_conjugate) {
const int batch = get_group_id(2);
const real_arg arg_alpha = arg_alphas[batch];
const real_arg arg_beta = arg_betas[batch];
const int a_offset = a_offsets[batch];
const int b_offset = b_offsets[batch];
const int c_offset = c_offsets[batch];
__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 batched GEMM kernel with [A, B] = [non-transposed, transposed]
__kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1)))
void XgemmDirectBatchedNT(const int kSizeM, const int kSizeN, const int kSizeK,
const __constant real_arg* arg_alphas, const __constant real_arg* arg_betas,
const __global realMD* restrict agm, const __constant int* a_offsets, const int a_ld,
const __global realND* restrict bgm, const __constant int* b_offsets, const int b_ld,
__global real* cgm, const __constant int* c_offsets, const int c_ld,
const int c_transpose, const int a_conjugate, const int b_conjugate) {
const int batch = get_group_id(2);
const real_arg arg_alpha = arg_alphas[batch];
const real_arg arg_beta = arg_betas[batch];
const int a_offset = a_offsets[batch];
const int b_offset = b_offsets[batch];
const int c_offset = c_offsets[batch];
__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 batched GEMM kernel with [A, B] = [transposed, non-transposed]
__kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1)))
void XgemmDirectBatchedTN(const int kSizeM, const int kSizeN, const int kSizeK,
const __constant real_arg* arg_alphas, const __constant real_arg* arg_betas,
const __global realMD* restrict agm, const __constant int* a_offsets, const int a_ld,
const __global realND* restrict bgm, const __constant int* b_offsets, const int b_ld,
__global real* cgm, const __constant int* c_offsets, const int c_ld,
const int c_transpose, const int a_conjugate, const int b_conjugate) {
const int batch = get_group_id(2);
const real_arg arg_alpha = arg_alphas[batch];
const real_arg arg_beta = arg_betas[batch];
const int a_offset = a_offsets[batch];
const int b_offset = b_offsets[batch];
const int c_offset = c_offsets[batch];
__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 batched GEMM kernel with [A, B] = [transposed, transposed]
__kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1)))
void XgemmDirectBatchedTT(const int kSizeM, const int kSizeN, const int kSizeK,
const __constant real_arg* arg_alphas, const __constant real_arg* arg_betas,
const __global realMD* restrict agm, const __constant int* a_offsets, const int a_ld,
const __global realND* restrict bgm, const __constant int* b_offsets, const int b_ld,
__global real* cgm, const __constant int* c_offsets, const int c_ld,
const int c_transpose, const int a_conjugate, const int b_conjugate) {
const int batch = get_group_id(2);
const real_arg arg_alpha = arg_alphas[batch];
const real_arg arg_beta = arg_betas[batch];
const int a_offset = a_offsets[batch];
const int b_offset = b_offsets[batch];
const int c_offset = c_offsets[batch];
__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);
}
#endif
// =================================================================================================
#if defined(ROUTINE_GEMMSTRIDEDBATCHED)
// Direct version of the strided-batched GEMM kernel with [A, B] = [non-transposed, non-transposed]
__kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1)))
void XgemmDirectStridedBatchedNN(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 int a_stride,
const __global realND* restrict bgm, const int b_offset, const int b_ld, const int b_stride,
__global real* cgm, const int c_offset, const int c_ld, const int c_stride,
const int c_transpose, const int a_conjugate, const int b_conjugate) {
const int batch = get_group_id(2);
const int a_offset_batch = a_offset + a_stride * batch;
const int b_offset_batch = b_offset + b_stride * batch;
const int c_offset_batch = c_offset + c_stride * batch;
__local real alm[WGD * (WGD + PADA)];
__local real blm[WGD * (WGD + PADB)];
XgemmDirect(kSizeM, kSizeN, kSizeK, arg_alpha, arg_beta,
agm, a_offset_batch, a_ld, bgm, b_offset_batch, b_ld, cgm, c_offset_batch, c_ld,
alm, blm, 0, 0, c_transpose, a_conjugate, b_conjugate);
}
// Direct version of the strided-batched GEMM kernel with [A, B] = [non-transposed, transposed]
__kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1)))
void XgemmDirectStridedBatchedNT(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 int a_stride,
const __global realND* restrict bgm, const int b_offset, const int b_ld, const int b_stride,
__global real* cgm, const int c_offset, const int c_ld, const int c_stride,
const int c_transpose, const int a_conjugate, const int b_conjugate) {
const int batch = get_group_id(2);
const int a_offset_batch = a_offset + a_stride * batch;
const int b_offset_batch = b_offset + b_stride * batch;
const int c_offset_batch = c_offset + c_stride * batch;
__local real alm[WGD * (WGD + PADA)];
__local real blm[WGD * (WGD + PADB)];
XgemmDirect(kSizeM, kSizeN, kSizeK, arg_alpha, arg_beta,
agm, a_offset_batch, a_ld, bgm, b_offset_batch, b_ld, cgm, c_offset_batch, c_ld,
alm, blm, 0, 1, c_transpose, a_conjugate, b_conjugate);
}
// Direct version of the strided-batched GEMM kernel with [A, B] = [transposed, non-transposed]
__kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1)))
void XgemmDirectStridedBatchedTN(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 int a_stride,
const __global realND* restrict bgm, const int b_offset, const int b_ld, const int b_stride,
__global real* cgm, const int c_offset, const int c_ld, const int c_stride,
const int c_transpose, const int a_conjugate, const int b_conjugate) {
const int batch = get_group_id(2);
const int a_offset_batch = a_offset + a_stride * batch;
const int b_offset_batch = b_offset + b_stride * batch;
const int c_offset_batch = c_offset + c_stride * batch;
__local real alm[WGD * (WGD + PADA)];
__local real blm[WGD * (WGD + PADB)];
XgemmDirect(kSizeM, kSizeN, kSizeK, arg_alpha, arg_beta,
agm, a_offset_batch, a_ld, bgm, b_offset_batch, b_ld, cgm, c_offset_batch, c_ld,
alm, blm, 1, 0, c_transpose, a_conjugate, b_conjugate);
}
// Direct version of the strided-batched GEMM kernel with [A, B] = [transposed, transposed]
__kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1)))
void XgemmDirectStridedBatchedTT(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 int a_stride,
const __global realND* restrict bgm, const int b_offset, const int b_ld, const int b_stride,
__global real* cgm, const int c_offset, const int c_ld, const int c_stride,
const int c_transpose, const int a_conjugate, const int b_conjugate) {
const int batch = get_group_id(2);
const int a_offset_batch = a_offset + a_stride * batch;
const int b_offset_batch = b_offset + b_stride * batch;
const int c_offset_batch = c_offset + c_stride * batch;
__local real alm[WGD * (WGD + PADA)];
__local real blm[WGD * (WGD + PADB)];
XgemmDirect(kSizeM, kSizeN, kSizeK, arg_alpha, arg_beta,
agm, a_offset_batch, a_ld, bgm, b_offset_batch, b_ld, cgm, c_offset_batch, c_ld,
alm, blm, 1, 1, c_transpose, a_conjugate, b_conjugate);
}
#endif
// =================================================================================================
// End of the C++11 raw string literal
)"
// =================================================================================================
|
