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
|
// =================================================================================================
// 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 an implementation of 3D convolution on a 4D image using GEMM kernels. It
// uses parameters from the direct GEMM 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_CONVGEMM)
// ConvGEMM kernel
__kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1)))
void Xconvgemm(const int num_patches, const int num_kernels, const int patch_size,
const __global realMD* restrict colgm, const int col_offset, const int col_stride,
const __global realND* restrict kernelgm, const int kernel_offset,
__global real* resultgm, const int result_offset, const int result_stride) {
// Batch offsets
const int batch = get_group_id(2);
const int col_offset_batch = col_offset + col_stride * batch;
const int result_offset_batch = result_offset + result_stride * batch;
__local real alm[WGD * (WGD + PADA)];
__local real blm[WGD * (WGD + PADB)];
// Extra pointers to scalar versions of global memory
const __global real* restrict colgms = (const __global real* restrict) colgm;
const __global real* restrict kernelgms = (const __global real* restrict) kernelgm;
// 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 < (num_patches/WGD)*WGD) && (idn < (num_kernels/WGD)*WGD)) {
// Loops over all complete workgroup tiles (K-dimension)
int kwg = 0;
for (; kwg < (patch_size/WGD) * WGD; kwg += WGD) {
// Loads data: off-chip --> local (matrix A and B)
if (num_patches % VWMD == 0 && col_offset_batch % VWMD == 0) {
GlobalToLocalDirectA(colgm, alm, num_patches, col_offset_batch, kwg, false, false);
}
else {
GlobalToLocalScalarA(colgms, alm, num_patches, col_offset_batch, kwg, false, false);
}
if (patch_size % VWND == 0 && kernel_offset % VWND == 0) {
GlobalToLocalDirectB(kernelgm, blm, patch_size, kernel_offset, kwg, true, false);
}
else {
GlobalToLocalScalarB(kernelgms, blm, patch_size, kernel_offset, kwg, true, false);
}
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, false);
}
#pragma unroll
for (int _ni = 0; _ni < NWID; _ni += 1) {
bpd[_ni] = LocalToPrivateDirectB(blm, _ni, kg, true);
}
// 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 < patch_size; ++kwg) {
// Loads data: off-chip --> private (matrix A and B)
#pragma unroll
for (int _mi = 0; _mi < MWID; _mi += 1) {
apd[_mi] = GlobalToPrivateDirectA(colgms, _mi, num_patches, col_offset_batch, idm, kwg, false, false);
}
#pragma unroll
for (int _ni = 0; _ni < NWID; _ni += 1) {
bpd[_ni] = GlobalToPrivateDirectB(kernelgms, _ni, patch_size, kernel_offset, idn, kwg, true, false);
}
// 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
#pragma unroll
for (int _ni = 0; _ni < NWID; _ni += 1) {
#pragma unroll
for (int _mi = 0; _mi < MWID; _mi += 1) {
StoreResultsDirect(resultgm, cpd[_ni * MWID + _mi], _mi, _ni, idm, idn,
ONE, ZERO, num_patches, result_offset_batch, false);
}
}
}
// 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 < (patch_size/WGD) * WGD; kwg+=WGD) {
// Loads data: off-chip --> local (matrix A and B)
GlobalToLocalCheckedA(colgms, alm, num_patches, col_offset_batch, kwg, false, false, num_patches, patch_size);
GlobalToLocalCheckedB(kernelgms, blm, patch_size, kernel_offset, kwg, true, false, num_kernels, patch_size);
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, false);
}
#pragma unroll
for (int _ni = 0; _ni < NWID; _ni += 1) {
bpd[_ni] = LocalToPrivateDirectB(blm, _ni, kg, true);
}
// 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 < patch_size; ++kwg) {
// Loads data: off-chip --> private (matrix A and B)
#pragma unroll
for (int _mi = 0; _mi < MWID; _mi += 1) {
apd[_mi] = GlobalToPrivateCheckedA(colgms, _mi, num_patches, col_offset_batch, idm, kwg, false, false, num_patches);
}
#pragma unroll
for (int _ni = 0; _ni < NWID; _ni += 1) {
bpd[_ni] = GlobalToPrivateCheckedB(kernelgms, _ni, patch_size, kernel_offset, idn, kwg, true, false, num_kernels);
}
// 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
#pragma unroll
for (int _ni = 0; _ni < NWID; _ni += 1) {
#pragma unroll
for (int _mi = 0; _mi < MWID; _mi += 1) {
StoreResultsChecked(resultgm, cpd[_ni * MWID + _mi], _mi, _ni, idm, idn, num_patches, num_kernels,
ONE, ZERO, num_patches, result_offset_batch, false);
}
}
}
}
#endif
// =================================================================================================
// End of the C++11 raw string literal
)"
// =================================================================================================
|
