-
Notifications
You must be signed in to change notification settings - Fork 4
/
cuda_matmul.h
366 lines (327 loc) · 12.4 KB
/
cuda_matmul.h
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
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
/*
* cuda_matmul.h
*
* Created on: 21.5.2013
* Author: Teemu Rantalaiho ([email protected])
*
*
* Copyright 2013 Teemu Rantalaiho
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*
*
*
*/
#ifndef CUDA_MATMUL_H_
#define CUDA_MATMUL_H_
#include <cuda_runtime_api.h>
/*
* Signatures:
*
* TRANSFORMFUNTYPE:
* struct <name>{
* OUTPUTTYPE operator()( INPUTTYPE input, INDEXTYPE x, INDEXTYPE y, SRCTYPE src, INDEXTYPE srcx);
* };
* // NOTE: when srcx = (transpose==true) ? y : x;
* SUMFUNTYPE:
* struct <name>{
* OUTPUTTYPE operator()( OUTPUTTYPE x, OUTPUTTYPE y);
* };
* STOREFUNTYPE:
* struct <name> {
* void operator()( DSTTYPE result, INDEXTYPE i, OUTPUTTYPE var){
* result[i] = var;
* }
* };
*
*/
template <typename OUTPUTTYPE, typename TRANSFORMFUNTYPE, typename INDEXTYPE, typename INPUTTYPE, typename SUMFUNTYPE, typename STOREFUNTYPE, typename SRCTYPE, typename DSTTYPE>
static inline
cudaError_t callFullMatMul(
INPUTTYPE input, TRANSFORMFUNTYPE xformFunctor, SUMFUNTYPE sumFunctor, STOREFUNTYPE storeDstFun,
INDEXTYPE sizex, INDEXTYPE sizey, SRCTYPE src, DSTTYPE result,
bool transpose = false, cudaStream_t stream = 0, bool outInDev = true);
template <typename INDEXTYPE, typename RADIXTYPE>
static inline
cudaError_t callFloatMatMul(
const RADIXTYPE* mat, INDEXTYPE sizex, INDEXTYPE sizey, INDEXTYPE stride,
const RADIXTYPE* src, RADIXTYPE* result, bool transpose = false,
cudaStream_t stream = 0, bool outInDev = true);
#ifndef __global__
#define __global__
#endif
#ifndef __shared__
#define __shared__
#endif
#ifndef __device__
#define __device__
#endif
#define MM_BLOCKSIZE_LOG2 8
#define MM_BLOCKSIZE (1<<MM_BLOCKSIZE_LOG2)
template <typename TRANSFORMFUNTYPE, typename INDEXTYPE, typename INPUTTYPE, typename SUMFUNTYPE, typename STOREFUNTYPE, typename OUTPUTTYPE, typename SRCTYPE, typename DSTTYPE>
__global__
void callFullMatmulKernel(
INPUTTYPE input, TRANSFORMFUNTYPE xformFunctor, SUMFUNTYPE sumFunctor, STOREFUNTYPE storeDstFun,
INDEXTYPE sizex, INDEXTYPE sizey, SRCTYPE src, DSTTYPE result,
int nstepsX, INDEXTYPE starty = 0, OUTPUTTYPE* tmpOut = NULL)
{
int tid = threadIdx.x;
INDEXTYPE y = (INDEXTYPE)(blockIdx.y + starty);
INDEXTYPE x = (INDEXTYPE)(tid + blockIdx.x * blockDim.x);
OUTPUTTYPE myRes;
int stride = gridDim.x << MM_BLOCKSIZE_LOG2;
if (x < sizex && y < sizey){
myRes = xformFunctor(input, x, y, src, x);
x += stride;
}
//#ifndef UNROLL_NLOG2_CUDA_STEPS
#define NUNROLL_LOG2 2
//#else
//#define NUNROLL_LOG2 UNROLL_NLOG2_CUDA_STEPS
//#endif
#define NUNROLL (1 << NUNROLL_LOG2)
int nFullSteps = (nstepsX - 1) >> NUNROLL_LOG2;
for (int fstep = 0; fstep < nFullSteps; fstep++){
#pragma unroll
for (int substep = 0; substep < NUNROLL; substep++){
OUTPUTTYPE tmpres = xformFunctor(input, x, y, src, x);
myRes = sumFunctor(myRes, tmpres);
x += stride;
}
}
while (x < sizex){
OUTPUTTYPE tmpres = xformFunctor(input, x, y, src, x);
myRes = sumFunctor(myRes, tmpres);
x += stride;
}
{
__shared__ OUTPUTTYPE tmparr[MM_BLOCKSIZE];
tmparr[tid] = myRes;
__syncthreads();
if (sizex < MM_BLOCKSIZE){
if (tid == 0){
for (int i = 1; i < sizex; i++)
myRes = sumFunctor(myRes, tmparr[i]);
}
} else {
if (tid < 32){
#pragma unroll
for (int i = 1; i < (MM_BLOCKSIZE >> 5); i++)
myRes = sumFunctor(myRes, tmparr[tid + (i<<5)]);
tmparr[tid] = myRes;
__threadfence_block();
tmparr[tid] = sumFunctor(tmparr[tid], tmparr[tid+16]);
__threadfence_block();
tmparr[tid] = sumFunctor(tmparr[tid], tmparr[tid+8]);
__threadfence_block();
tmparr[tid] = sumFunctor(tmparr[tid], tmparr[tid+4]);
__threadfence_block();
tmparr[tid] = sumFunctor(tmparr[tid], tmparr[tid+2]);
__threadfence_block();
myRes = sumFunctor(tmparr[tid], tmparr[tid+1]);
}
}
}
if (threadIdx.x == 0){
if (tmpOut)
tmpOut[blockIdx.x + blockIdx.y * gridDim.x] = myRes;
else
storeDstFun(result, y, myRes);
}
}
#define FINALSUMT 64
template <typename INDEXTYPE, typename SUMFUNTYPE, typename STOREFUNTYPE, typename OUTPUTTYPE, typename DSTTYPE>
__global__
void FinalSumsKernel(
SUMFUNTYPE sumFunctor, STOREFUNTYPE storeDstFun,
INDEXTYPE sizex, OUTPUTTYPE* tmpptr, DSTTYPE result, INDEXTYPE startY)
{
__shared__ OUTPUTTYPE tmparr[FINALSUMT];
OUTPUTTYPE myRes;
int tid = threadIdx.x;
if (tid < sizex)
tmparr[tid] = tmpptr[tid + blockIdx.x * sizex];
__syncthreads();
if (tid >= FINALSUMT/2)
return;
if (sizex == FINALSUMT){
#if FINALSUMT == 64
__threadfence_block();
tmparr[tid] = sumFunctor(tmparr[tid], tmparr[tid+32]);
#endif
__threadfence_block();
tmparr[tid] = sumFunctor(tmparr[tid], tmparr[tid+16]);
__threadfence_block();
tmparr[tid] = sumFunctor(tmparr[tid], tmparr[tid+8]);
__threadfence_block();
tmparr[tid] = sumFunctor(tmparr[tid], tmparr[tid+4]);
__threadfence_block();
tmparr[tid] = sumFunctor(tmparr[tid], tmparr[tid+2]);
__threadfence_block();
myRes = sumFunctor(tmparr[tid], tmparr[tid+1]);
} else {
__threadfence_block();
myRes = tmparr[0];
for (int i = 1; i < sizex; i++)
myRes = sumFunctor(myRes, tmparr[i]);
}
if (tid == 0)
storeDstFun(result, blockIdx.x + startY, myRes);
}
static inline int divLog2RoundUp(int size, int divlog2)
{
int div = 1 << divlog2;
int paddedSize = (size + div - 1) & (~(div - 1));
int res = paddedSize >> divlog2;
return res;
}
#include <stdio.h>
template <typename TRANSFORMFUNTYPE, typename INDEXTYPE, typename INPUTTYPE, typename SUMFUNTYPE, typename STOREFUNTYPE, typename OUTPUTTYPE, typename SRCTYPE, typename DSTTYPE>
static
cudaError_t callFullMatMulImpl(
INPUTTYPE input, TRANSFORMFUNTYPE xformFunctor, SUMFUNTYPE sumFunctor, STOREFUNTYPE storeDstFun,
INDEXTYPE sizex, INDEXTYPE sizey, SRCTYPE src, DSTTYPE result,
cudaStream_t stream, bool outInDev)
{
//int stepsX = divLog2RoundUp(sizex, MM_BLOCKSIZE_LOG2);
int stepsX = sizex >> MM_BLOCKSIZE_LOG2;
INDEXTYPE blocksX = 1;
OUTPUTTYPE* tmpptr = NULL;
dim3 block = MM_BLOCKSIZE;
if (sizey < 32 && stepsX > 64 * NUNROLL){
blocksX = FINALSUMT;
if (blocksX > stepsX)
blocksX = stepsX;
stepsX = (sizex/blocksX) >> MM_BLOCKSIZE_LOG2;
}
dim3 grid(blocksX, sizey, 1);
if (!outInDev){
printf("Sorry - no support yet for CPU-output buffers...\n");
return cudaSuccess;
}
if (sizex <= 0 || sizey <= 0)
return cudaSuccess;
if (blocksX > 1){
size_t needed = sizeof(OUTPUTTYPE) * blocksX * (sizey < 32768 ? sizey : 32768);
cudaMalloc(&tmpptr, needed);
}
/* printf("block = (%d, %d, %d), grid = (%d,%d)\n", block.x, block.y, block.z, grid.x, grid.y);*/
if (sizey > 32768){
grid.y = 32768;
int startY = 0;
while (startY < sizey){
if (startY + 32768 > sizey)
grid.y = sizey - startY;
callFullMatmulKernel
<TRANSFORMFUNTYPE, INDEXTYPE, INPUTTYPE, SUMFUNTYPE, STOREFUNTYPE, OUTPUTTYPE, SRCTYPE, DSTTYPE>
<<<grid, block,0,stream>>>
(input, xformFunctor, sumFunctor, storeDstFun,
sizex, sizey, src, result, stepsX, startY, tmpptr);
if (blocksX > 1){
dim3 fgrid = 32768;
dim3 fblock = FINALSUMT;
FinalSumsKernel<INDEXTYPE, SUMFUNTYPE, STOREFUNTYPE, OUTPUTTYPE, DSTTYPE>
<<<fgrid, fblock,0,stream>>>(sumFunctor, storeDstFun, blocksX, tmpptr, result, startY);
}
startY += 32768;
}
}
else
{
callFullMatmulKernel
<TRANSFORMFUNTYPE, INDEXTYPE, INPUTTYPE, SUMFUNTYPE, STOREFUNTYPE, OUTPUTTYPE, SRCTYPE, DSTTYPE>
<<<grid, block,0,stream>>>
(input, xformFunctor, sumFunctor, storeDstFun, sizex, sizey, src, result, stepsX, 0, tmpptr);
if (blocksX > 1){
dim3 fgrid = sizey;
dim3 fblock = FINALSUMT;
FinalSumsKernel<INDEXTYPE, SUMFUNTYPE, STOREFUNTYPE, OUTPUTTYPE, DSTTYPE>
<<<fgrid, fblock,0,stream>>>(sumFunctor, storeDstFun, blocksX, tmpptr, result, 0);
}
}
return cudaGetLastError();
}
template <typename INDEXTYPE, typename OUTPUTTYPE, typename TRANSFORMFUNTYPE, typename INPUTTYPE, typename SRCTYPE>
struct FullMatMulTransposeWrapper {
TRANSFORMFUNTYPE userFunctor;
inline __device__
OUTPUTTYPE operator()( INPUTTYPE input, INDEXTYPE x, INDEXTYPE y, SRCTYPE src, INDEXTYPE srcx){
return userFunctor(input, y, x, src, x);
}
};
template <typename OUTPUTTYPE, typename TRANSFORMFUNTYPE, typename INDEXTYPE, typename INPUTTYPE, typename SUMFUNTYPE, typename STOREFUNTYPE, typename SRCTYPE, typename DSTTYPE>
cudaError_t callFullMatMul(
INPUTTYPE input, TRANSFORMFUNTYPE xformFunctor, SUMFUNTYPE sumFunctor, STOREFUNTYPE storeDstFun,
INDEXTYPE sizex, INDEXTYPE sizey, SRCTYPE src, DSTTYPE result,
bool transpose, cudaStream_t stream, bool outInDev)
{
cudaError_t err;
if (transpose){
struct FullMatMulTransposeWrapper<INDEXTYPE, OUTPUTTYPE, TRANSFORMFUNTYPE,INPUTTYPE,SRCTYPE> wrapperFun;
wrapperFun.userFunctor = xformFunctor;
err = callFullMatMulImpl
<FullMatMulTransposeWrapper<INDEXTYPE, OUTPUTTYPE, TRANSFORMFUNTYPE,INPUTTYPE,SRCTYPE>,
INDEXTYPE, INPUTTYPE, SUMFUNTYPE, STOREFUNTYPE, OUTPUTTYPE, SRCTYPE, DSTTYPE>
(input, wrapperFun, sumFunctor, storeDstFun, sizey, sizex, src, result, stream, outInDev);
}
else
{
err = callFullMatMulImpl
<TRANSFORMFUNTYPE, INDEXTYPE, INPUTTYPE, SUMFUNTYPE, STOREFUNTYPE, OUTPUTTYPE, SRCTYPE, DSTTYPE>
(input, xformFunctor, sumFunctor, storeDstFun, sizex, sizey, src, result, stream, outInDev);
}
return err;
}
template <typename INDEXTYPE, typename RADIXTYPE>
struct floatMatMulFun {
INDEXTYPE stride;
__device__
RADIXTYPE operator()( const RADIXTYPE* data, INDEXTYPE x, INDEXTYPE y, const RADIXTYPE* src, INDEXTYPE srcx){
RADIXTYPE m_ij = data[y*stride + x];
RADIXTYPE res = m_ij*src[srcx];
return res;
}
};
template <typename RADIXTYPE>
struct radixSumFun {
__device__
RADIXTYPE operator()( RADIXTYPE a, RADIXTYPE b){
RADIXTYPE res = a + b;
return res;
}
};
template <typename INDEXTYPE, typename RADIXTYPE>
struct storefloatFun {
__device__
void operator()( RADIXTYPE* result, INDEXTYPE i, RADIXTYPE var){
result[i] = var;
}
};
template <typename INDEXTYPE, typename RADIXTYPE>
cudaError_t callFloatMatMul(
const RADIXTYPE* mat, INDEXTYPE sizex, INDEXTYPE sizey, INDEXTYPE stride,
const RADIXTYPE* src, RADIXTYPE* result, bool transpose,
cudaStream_t stream, bool outInDev)
{
struct floatMatMulFun<INDEXTYPE, RADIXTYPE> mulfun;
struct radixSumFun<RADIXTYPE> sumfun;
struct storefloatFun<INDEXTYPE, RADIXTYPE> storeDstfun;
mulfun.stride = stride;
cudaError_t err = callFullMatMul<RADIXTYPE, floatMatMulFun<INDEXTYPE, RADIXTYPE>, INDEXTYPE, const RADIXTYPE*, radixSumFun<RADIXTYPE>,
storefloatFun<INDEXTYPE, RADIXTYPE>, const RADIXTYPE*, RADIXTYPE*>
(mat, mulfun, sumfun, storeDstfun, sizex, sizey, src, result, transpose, stream, outInDev);
if (err != cudaSuccess)
printf("Error in callFullMatMul! err = %s\n", cudaGetErrorString(err));
return err;
}
#endif /* CUDA_MATMUL_H_ */