-
Notifications
You must be signed in to change notification settings - Fork 0
/
transpose.cu
298 lines (255 loc) · 9.8 KB
/
transpose.cu
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
// Copyright 2012 NVIDIA Corporation
//
// 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.
#include <stdio.h>
#include <assert.h>
// Convenience function for checking CUDA runtime API results
// can be wrapped around any runtime API call. No-op in release builds.
inline
cudaError_t checkCuda(cudaError_t result)
{
#if defined(DEBUG) || defined(_DEBUG)
if (result != cudaSuccess) {
fprintf(stderr, "CUDA Runtime Error: %s\n", cudaGetErrorString(result));
assert(result == cudaSuccess);
}
#endif
return result;
}
const int TILE_DIM = 8;
const int BLOCK_ROWS = 4;
const int NUM_REPS = 100;
// Check errors and print GB/s
void postprocess(const float *ref, const float *res, int n, float ms)
{
bool passed = true;
for (int i = 0; i < n; i++)
if (res[i] != ref[i]) {
printf("%d %f %f\n", i, res[i], ref[i]);
printf("%25s\n", "*** FAILED ***");
passed = false;
break;
}
if (passed)
printf("%20.2f\n", 2 * n * sizeof(float) * 1e-6 * NUM_REPS / ms );
}
// simple copy kernel
// Used as reference case representing best effective bandwidth.
__global__ void copy(float *odata, const float *idata)
{
int x = blockIdx.x * TILE_DIM + threadIdx.x;
int y = blockIdx.y * TILE_DIM + threadIdx.y;
int width = gridDim.x * TILE_DIM;
for (int j = 0; j < TILE_DIM; j+= BLOCK_ROWS)
odata[(y+j)*width + x] = idata[(y+j)*width + x];
}
// copy kernel using shared memory
// Also used as reference case, demonstrating effect of using shared memory.
__global__ void copySharedMem(float *odata, const float *idata)
{
__shared__ float tile[TILE_DIM * TILE_DIM];
int x = blockIdx.x * TILE_DIM + threadIdx.x;
int y = blockIdx.y * TILE_DIM + threadIdx.y;
int width = gridDim.x * TILE_DIM;
for (int j = 0; j < TILE_DIM; j += BLOCK_ROWS)
tile[(threadIdx.y+j)*TILE_DIM + threadIdx.x] = idata[(y+j)*width + x];
__syncthreads();
for (int j = 0; j < TILE_DIM; j += BLOCK_ROWS)
odata[(y+j)*width + x] = tile[(threadIdx.y+j)*TILE_DIM + threadIdx.x];
}
// naive transpose
// Simplest transpose; doesn't use shared memory.
// Global memory reads are coalesced but writes are not.
__global__ void transposeNaive(float *odata, const float *idata)
{
int x = blockIdx.x * TILE_DIM + threadIdx.x;
int y = blockIdx.y * TILE_DIM + threadIdx.y;
int width = gridDim.x * TILE_DIM;
for (int j = 0; j < TILE_DIM; j+= BLOCK_ROWS)
odata[x*width + (y+j)] = idata[(y+j)*width + x];
}
// coalesced transpose
// Uses shared memory to achieve coalesing in both reads and writes
// Tile width == #banks causes shared memory bank conflicts.
__global__ void transposeCoalesced(float *odata, const float *idata)
{
__shared__ float tile[TILE_DIM][TILE_DIM];
int x = blockIdx.x * TILE_DIM + threadIdx.x;
int y = blockIdx.y * TILE_DIM + threadIdx.y;
int width = gridDim.x * TILE_DIM;
for (int j = 0; j < TILE_DIM; j += BLOCK_ROWS)
tile[threadIdx.y+j][threadIdx.x] = idata[(y+j)*width + x];
__syncthreads();
x = blockIdx.y * TILE_DIM + threadIdx.x; // transpose block offset
y = blockIdx.x * TILE_DIM + threadIdx.y;
for (int j = 0; j < TILE_DIM; j += BLOCK_ROWS)
odata[(y+j)*width + x] = tile[threadIdx.x][threadIdx.y + j];
}
// No bank-conflict transpose
// Same as transposeCoalesced except the first tile dimension is padded
// to avoid shared memory bank conflicts.
__global__ void transposeNoBankConflicts(float *odata, const float *idata)
{
__shared__ float tile[TILE_DIM][TILE_DIM+1];
int x = blockIdx.x * TILE_DIM + threadIdx.x;
int y = blockIdx.y * TILE_DIM + threadIdx.y;
int width = gridDim.x * TILE_DIM;
for (int j = 0; j < TILE_DIM; j += BLOCK_ROWS)
tile[threadIdx.y+j][threadIdx.x] = idata[(y+j)*width + x];
__syncthreads();
x = blockIdx.y * TILE_DIM + threadIdx.x; // transpose block offset
y = blockIdx.x * TILE_DIM + threadIdx.y;
for (int j = 0; j < TILE_DIM; j += BLOCK_ROWS)
odata[(y+j)*width + x] = tile[threadIdx.x][threadIdx.y + j];
}
int main(int argc, char **argv)
{
const int nx = 1024;
const int ny = 1024;
const int mem_size = nx*ny*sizeof(float);
dim3 dimGrid(nx/TILE_DIM, ny/TILE_DIM, 1);
dim3 dimBlock(TILE_DIM, BLOCK_ROWS, 1);
int devId = 0;
if (argc > 1) devId = atoi(argv[1]);
cudaDeviceProp prop;
checkCuda( cudaGetDeviceProperties(&prop, devId));
printf("\nDevice : %s\n", prop.name);
printf("Matrix size: %d %d, Block size: %d %d, Tile size: %d %d\n",
nx, ny, TILE_DIM, BLOCK_ROWS, TILE_DIM, TILE_DIM);
printf("dimGrid: %d %d %d. dimBlock: %d %d %d\n",
dimGrid.x, dimGrid.y, dimGrid.z, dimBlock.x, dimBlock.y, dimBlock.z);
checkCuda( cudaSetDevice(devId) );
float *h_idata = (float*)malloc(mem_size);
float *h_cdata = (float*)malloc(mem_size);
float *h_tdata = (float*)malloc(mem_size);
float *gold = (float*)malloc(mem_size);
float *d_idata, *d_cdata, *d_tdata;
checkCuda( cudaMalloc(&d_idata, mem_size) );
checkCuda( cudaMalloc(&d_cdata, mem_size) );
checkCuda( cudaMalloc(&d_tdata, mem_size) );
// check parameters and calculate execution configuration
if (nx % TILE_DIM || ny % TILE_DIM) {
printf("nx and ny must be a multiple of TILE_DIM\n");
goto error_exit;
}
/*if (TILE_DIM % BLOCK_ROWS) {
printf("TILE_DIM must be a multiple of BLOCK_ROWS\n");
goto error_exit;
}*/
// host
for (int j = 0; j < ny; j++)
for (int i = 0; i < nx; i++)
h_idata[j*nx + i] = j*nx + i;
// correct result for error checking
for (int j = 0; j < ny; j++)
for (int i = 0; i < nx; i++)
gold[j*nx + i] = h_idata[i*nx + j];
// device
checkCuda( cudaMemcpy(d_idata, h_idata, mem_size, cudaMemcpyHostToDevice) );
// events for timing
cudaEvent_t startEvent, stopEvent;
checkCuda( cudaEventCreate(&startEvent) );
checkCuda( cudaEventCreate(&stopEvent) );
float ms;
// ------------
// time kernels
// ------------
printf("%25s%25s\n", "Routine", "Bandwidth (GB/s)");
/* // ----
// copy
// ----
printf("%25s", "copy");
checkCuda( cudaMemset(d_cdata, 0, mem_size) );
// warm up
copy<<<dimGrid, dimBlock>>>(d_cdata, d_idata);
checkCuda( cudaEventRecord(startEvent, 0) );
for (int i = 0; i < NUM_REPS; i++)
copy<<<dimGrid, dimBlock>>>(d_cdata, d_idata);
checkCuda( cudaEventRecord(stopEvent, 0) );
checkCuda( cudaEventSynchronize(stopEvent) );
checkCuda( cudaEventElapsedTime(&ms, startEvent, stopEvent) );
checkCuda( cudaMemcpy(h_cdata, d_cdata, mem_size, cudaMemcpyDeviceToHost) );
postprocess(h_idata, h_cdata, nx*ny, ms);
// -------------
// copySharedMem
// -------------
printf("%25s", "shared memory copy");
checkCuda( cudaMemset(d_cdata, 0, mem_size) );
// warm up
copySharedMem<<<dimGrid, dimBlock>>>(d_cdata, d_idata);
checkCuda( cudaEventRecord(startEvent, 0) );
for (int i = 0; i < NUM_REPS; i++)
copySharedMem<<<dimGrid, dimBlock>>>(d_cdata, d_idata);
checkCuda( cudaEventRecord(stopEvent, 0) );
checkCuda( cudaEventSynchronize(stopEvent) );
checkCuda( cudaEventElapsedTime(&ms, startEvent, stopEvent) );
checkCuda( cudaMemcpy(h_cdata, d_cdata, mem_size, cudaMemcpyDeviceToHost) );
postprocess(h_idata, h_cdata, nx * ny, ms);
*/
// --------------
// transposeNaive
// --------------
printf("%25s", "naive transpose");
checkCuda( cudaMemset(d_tdata, 0, mem_size) );
// warmup
transposeNaive<<<dimGrid, dimBlock>>>(d_tdata, d_idata);
checkCuda( cudaEventRecord(startEvent, 0) );
for (int i = 0; i < NUM_REPS; i++)
transposeNaive<<<dimGrid, dimBlock>>>(d_tdata, d_idata);
checkCuda( cudaEventRecord(stopEvent, 0) );
checkCuda( cudaEventSynchronize(stopEvent) );
checkCuda( cudaEventElapsedTime(&ms, startEvent, stopEvent) );
checkCuda( cudaMemcpy(h_tdata, d_tdata, mem_size, cudaMemcpyDeviceToHost) );
postprocess(gold, h_tdata, nx * ny, ms);
/* // ------------------
// transposeCoalesced
// ------------------
printf("%25s", "coalesced transpose");
checkCuda( cudaMemset(d_tdata, 0, mem_size) );
// warmup
transposeCoalesced<<<dimGrid, dimBlock>>>(d_tdata, d_idata);
checkCuda( cudaEventRecord(startEvent, 0) );
for (int i = 0; i < NUM_REPS; i++)
transposeCoalesced<<<dimGrid, dimBlock>>>(d_tdata, d_idata);
checkCuda( cudaEventRecord(stopEvent, 0) );
checkCuda( cudaEventSynchronize(stopEvent) );
checkCuda( cudaEventElapsedTime(&ms, startEvent, stopEvent) );
checkCuda( cudaMemcpy(h_tdata, d_tdata, mem_size, cudaMemcpyDeviceToHost) );
postprocess(gold, h_tdata, nx * ny, ms);
// ------------------------
// transposeNoBankConflicts
// ------------------------
printf("%25s", "conflict-free transpose");
checkCuda( cudaMemset(d_tdata, 0, mem_size) );
// warmup
transposeNoBankConflicts<<<dimGrid, dimBlock>>>(d_tdata, d_idata);
checkCuda( cudaEventRecord(startEvent, 0) );
for (int i = 0; i < NUM_REPS; i++)
transposeNoBankConflicts<<<dimGrid, dimBlock>>>(d_tdata, d_idata);
checkCuda( cudaEventRecord(stopEvent, 0) );
checkCuda( cudaEventSynchronize(stopEvent) );
checkCuda( cudaEventElapsedTime(&ms, startEvent, stopEvent) );
checkCuda( cudaMemcpy(h_tdata, d_tdata, mem_size, cudaMemcpyDeviceToHost) );
postprocess(gold, h_tdata, nx * ny, ms);*/
error_exit:
// cleanup
checkCuda( cudaEventDestroy(startEvent) );
checkCuda( cudaEventDestroy(stopEvent) );
checkCuda( cudaFree(d_tdata) );
checkCuda( cudaFree(d_cdata) );
checkCuda( cudaFree(d_idata) );
free(h_idata);
free(h_tdata);
free(h_cdata);
free(gold);
}