-
Notifications
You must be signed in to change notification settings - Fork 0
/
solutionWithSharedMemory.cu
303 lines (235 loc) · 8.79 KB
/
solutionWithSharedMemory.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
299
300
301
302
303
// System includes
#include <stdio.h>
#include <stdlib.h>
#include <assert.h>
// CUDA runtime
#include <cuda_runtime.h>
// Helper functions and utilities to work with CUDA
#include <helper_functions.h>
#include "device_functions.h"
#include <helper_cuda.h>
#define MAX_PIXEL_INTENSITY 63
#define IMAGE_MATRIX_DIMENSION 15
#define RANDOM_SEED 5
#define WINDOW_DIMENSION 5
#define WINDOW_COUNT_PER_FRAME 11
#define FRAME_DIMENSION_WIDTH 15
#define FRAME_DIMENSION_HEIGHT 5
#define FRAME_COUNT 11
#define THREADS_IN_BLOCK_WIDTH WINDOW_COUNT_PER_FRAME
#define THREADS_IN_BLOCK_HEIGHT 1
#define BLOCKS_IN_GRID_WIDTH 1
#define BLOCKS_IN_GRID_HEIGHT 1
// Matrices are stored in row-major order:
// M(row, col) = *(M.elements + row * M.stride + col)
typedef struct {
int width;
int height;
int stride;
unsigned int* elements;
} Matrix;
// To measure time
cudaEvent_t start, stop;
float elapsed_time_ms;
__device__ unsigned int GetElement(const Matrix* A, int row, int col)
{
return A->elements[row * A->stride + col];
}
__device__ void SetElement(Matrix* A, int row, int col, unsigned int value)
{
A->elements[row * A->stride + col] = value;
}
// Get the WINDOW_DIMENSIONxWINDOW_DIMENSION sub-matrix Asub of A that is
// located col sub-matrices to the right and row sub-matrices down
// from the upper-left corner of A
__device__ void GetWindowFromGlobal(Matrix globalMatrix, Matrix *matrixInWindow, int frameNumber, int windowNumber)
{
unsigned int element;
int startIndex = globalMatrix.stride * frameNumber + windowNumber;
for (int i = 0; i < matrixInWindow->height; i++)
{
for (int j = 0; j < matrixInWindow->width; j++)
{
element =
globalMatrix.elements[startIndex + i * globalMatrix.stride + j];
matrixInWindow->elements[i * matrixInWindow->stride + j] = element;
}
}
}
__device__ void GetWindowFromShared(Matrix sharedMatrix, Matrix* matrixInWindow, int windowNumber)
{
unsigned int element;
int startIndex = windowNumber;
for (int i = 0; i < matrixInWindow->height; i++)
{
for (int j = 0; j < matrixInWindow->width; j++)
{
element =
sharedMatrix.elements[startIndex + i * sharedMatrix.stride + j];
matrixInWindow->elements[i * matrixInWindow->stride + j] = element;
}
}
}
__device__ void SetWindowToGlobal(Matrix* matrixInWindow, Matrix* globalMatrix,
int frameNumber, int windowNumber)
{
unsigned int value;
int startIndex = globalMatrix->stride * frameNumber + windowNumber;
for (int i = 0; i < matrixInWindow->height; i++)
{
for (int j = 0; j < matrixInWindow->width; j++)
{
value = GetElement(matrixInWindow, i, j);
atomicAdd(&(globalMatrix->elements[startIndex + i * globalMatrix->stride + j]), value);
}
}
}
__device__ void SetWindowToShared(Matrix* matrixInWindow, Matrix* sharedMatrix, int windowNumber)
{
unsigned int value;
int startIndex = windowNumber;
for (int i = 0; i < matrixInWindow->height; i++)
{
for (int j = 0; j < matrixInWindow->width; j++)
{
value = GetElement(matrixInWindow, i, j);
sharedMatrix->elements[startIndex + i * sharedMatrix->stride + j] += value;
}
}
}
__global__ void MatMulKernel(Matrix A, Matrix C, Matrix matrixInWindow, Matrix resultMatrixInWindow, int frameIndex)
{
int frameNumber = frameIndex;
int windowNumber = threadIdx.x;
__shared__ Matrix A_shared;
A_shared.width = FRAME_DIMENSION_WIDTH;
A_shared.height = FRAME_DIMENSION_HEIGHT;
A_shared.stride = A_shared.width;
A_shared.elements =
(unsigned int*)malloc(FRAME_DIMENSION_WIDTH *
FRAME_DIMENSION_HEIGHT *
sizeof(unsigned int));
if (windowNumber % WINDOW_DIMENSION == 0)
{
GetWindowFromGlobal(A, &matrixInWindow, frameNumber, windowNumber);
SetWindowToShared(&matrixInWindow, &A_shared, windowNumber);
}
__syncthreads();
GetWindowFromShared(A_shared, &matrixInWindow, windowNumber);
GetWindowFromGlobal(C, &resultMatrixInWindow, frameNumber, windowNumber);
unsigned int accumulator;
for (int i = 0; i < matrixInWindow.height; i++)
{
for (int j = 0; j < matrixInWindow.width; j++)
{
accumulator = 0;
for (int k = 0; k < WINDOW_DIMENSION; k++)
{
accumulator += GetElement(&matrixInWindow, i, k) *
GetElement(&matrixInWindow, k, j);
}
SetElement(&resultMatrixInWindow, i, j, accumulator);
}
}
__syncthreads();
SetWindowToGlobal(&resultMatrixInWindow, &C, frameNumber, windowNumber);
}
static void generateMatrixElements(Matrix* A)
{
size_t size = A->width * A->height;
srand(RANDOM_SEED);
for (int i = 0; i < size; i++)
{
A->elements[i] = rand() % MAX_PIXEL_INTENSITY;
}
}
static void printMatrixOnConsole(Matrix* matrix, char* matrixName)
{
printf(" %s\n", matrixName);
for (int row = 0; row < matrix->height; row++)
{
for (int column = 0; column < matrix->width; column++)
{
printf(" %d\t", matrix->elements[row * matrix->width + column]);
}
printf("\n");
}
printf("\n\n");
}
int main(int argc, char** argv)
{
Matrix A;
Matrix C;
int elementCount;
size_t size;
cudaEventCreate(&start);
cudaEventCreate(&stop);
A.width = C.width = IMAGE_MATRIX_DIMENSION;
A.height = C.height = IMAGE_MATRIX_DIMENSION;
A.stride = C.stride = IMAGE_MATRIX_DIMENSION;
elementCount = IMAGE_MATRIX_DIMENSION * IMAGE_MATRIX_DIMENSION;
size = elementCount * sizeof(unsigned int);
A.elements = (unsigned int*)malloc(size);
C.elements = (unsigned int*)calloc(elementCount, sizeof(unsigned int));
generateMatrixElements(&A);
printMatrixOnConsole(&A, "Matrix A - Initial");
printMatrixOnConsole(&C, "Matrix C - Initial");
for (int frameIndex = 0; frameIndex < FRAME_COUNT; frameIndex++)
{
Matrix frameMatrix;
frameMatrix.height = FRAME_DIMENSION_HEIGHT;
frameMatrix.width = FRAME_DIMENSION_WIDTH;
frameMatrix.stride = frameMatrix.width;
// Load A to device memory
Matrix d_A;
d_A.width = A.width;
d_A.height = A.height;
d_A.stride = A.stride;
cudaMalloc(&d_A.elements, size);
cudaMemcpy(d_A.elements, A.elements, size, cudaMemcpyHostToDevice);
// Allocate C in device memory
Matrix d_C;
d_C.width = C.width;
d_C.height = C.height;
d_C.stride = C.stride;
cudaMalloc(&d_C.elements, size);
cudaMemcpy(d_C.elements, C.elements, size, cudaMemcpyHostToDevice);
// Allocate matrix for computations is device
// Note: It makes easier to debug with NSight when allocating it from the host
Matrix matrixInWindow;
matrixInWindow.width = WINDOW_DIMENSION;
matrixInWindow.height = WINDOW_DIMENSION;
matrixInWindow.stride = matrixInWindow.width;
cudaMalloc(&matrixInWindow.elements, WINDOW_DIMENSION * WINDOW_DIMENSION * sizeof(unsigned int));
// Allocate matrix for computations is device
// Note: It makes easier to debug with NSight when allocating it from the host
Matrix resultMatrixInWindow;
resultMatrixInWindow.width = WINDOW_DIMENSION;
resultMatrixInWindow.height = WINDOW_DIMENSION;
resultMatrixInWindow.stride = matrixInWindow.width;
cudaMalloc(&resultMatrixInWindow.elements, WINDOW_DIMENSION * WINDOW_DIMENSION * sizeof(unsigned int));
// Start to measure time
cudaEventRecord(start, 0);
// Invoke kernel
dim3 dimBlock(THREADS_IN_BLOCK_WIDTH, THREADS_IN_BLOCK_HEIGHT);
dim3 dimGrid(BLOCKS_IN_GRID_WIDTH, BLOCKS_IN_GRID_HEIGHT);
MatMulKernel <<< dimGrid, dimBlock >>> (d_A, d_C, matrixInWindow, resultMatrixInWindow, frameIndex);
// Wait for GPU to finish before accessing on host
cudaDeviceSynchronize();
// Read C from device memory
cudaMemcpy(C.elements, d_C.elements, size, cudaMemcpyDeviceToHost);
// Stop to measure time
cudaEventRecord(stop, 0);
cudaEventSynchronize(stop);
cudaEventElapsedTime(&elapsed_time_ms, start, stop);
// Free device memory
cudaFree(d_A.elements);
cudaFree(d_C.elements);
cudaFree(matrixInWindow.elements);
cudaFree(resultMatrixInWindow.elements);
printf("After Frame %d Run\n", frameIndex);
printf(" Elapsed Time (ms): %f \n", elapsed_time_ms);
printMatrixOnConsole(&C, "Matrix C");
}
return 0;
}