二維矩陣相乘——cpu&&gpu

1. #include <stdio.h>
2. #include <stdlib.h>
3. #include <time.h>
4. #include <iostream>
5. #include <cmath>
6. using namespace std;
7. #define M 3200
8. #define N 3200
9. #define P 3200
10. #define B_S 32
11. //#define SHOW
12. //A[M][P]*B[P][N]=C[M][N]
13. __global__
14. void mulMatri_gpu(float* A, float* B, float* C, int m, int n, int p)
15. {
16. int index_x = blockIdx.x * blockDim.x + threadIdx.x;
17. int index_y = blockIdx.y * blockDim.y + threadIdx.y;
18. if (index_x >= n || index_y >= m) return;
19. float Pvalue = 0;
20. for (int k = 0; k < p; k++)
21. {
22. Pvalue += A[index_y*p + k] * B[k*n + index_x];
23. }
24. C[index_y*n + index_x] = Pvalue;
25. }
26. void mulMatri_cpu(float* A, float* B, float* C, int m, int n, int p)
27. {
29. int i, j, k;
30. for (i = 0; i<m; i++)
31. for (j = 0; j<n; j++){
32. C[i*n + j] = 0;
33. for (k = 0; k<p; k++)
34. C[i*n + j] += A[i*p + k] * B[k*n + j];
35. }
38. }
39. void compute_gpu(float* A, float *B, float *C)
40. {
42. float *da, *db, *dc;
43. cudaMalloc((void **)&da, sizeof(float)*M*P);
44. cudaMalloc((void **)&db, sizeof(float)*P*N);
45. cudaMalloc((void **)&dc, sizeof(float)*M*N);
46. ///測試時間
47. float elapsedTime = 0.0f;
48. cudaEvent_t start, stop;
49. cudaEventCreate(&start);
50. cudaEventCreate(&stop);
51. cudaEventRecord(start, 0);
53. cudaMemcpy(da, A, sizeof(float)*M*P, cudaMemcpyHostToDevice);
54. cudaMemcpy(db, B, sizeof(float)*P*N, cudaMemcpyHostToDevice);
55. dim3 dimGrid((M + B_S - 1) / B_S, (N + B_S - 1) / B_S);
56. dim3 dimBlock(B_S, B_S);
57. mulMatri_gpu << <dimGrid, dimBlock >> >(da, db, dc, M, N, P);
58. //cudaDeviceSynchronize();
59. cudaMemcpy(C, dc, sizeof(float)*N*M, cudaMemcpyDeviceToHost);
60. ///時間結束
61. cudaEventRecord(stop, 0);
62. cudaEventSynchronize(stop);
63. cudaEventElapsedTime(&elapsedTime, start, stop);
64. printf("the time on gpu is %f ms\n", elapsedTime);
66. cudaFree(da);
67. cudaFree(db);
68. cudaFree(dc);
69. cudaEventDestroy(start);
70. cudaEventDestroy(stop);
71. }
72. void compute_cpu(float* A, float *B, float *C)
73. {
74. clock_t start, finish;
75. start = clock();
76. mulMatri_cpu(A, B, C, M, N, P);
77. finish = clock();
78. printf("the time on cpu is %f ms\n", (double)(finish - start));
79. }
80. void verify(float *C1, float *C2, int m, int n)
81. {
82. for (int i = 0; i < m; i++)
83. for (int j = 0; j < n; j++)
84. {
85. if ((C2[i*n + j] - C1[i*m + j])>1e-5)
86. {
87. printf("error! results are not equel！");
88. break;
89. }
91. }
92. }
93. int main()
94. {
95. float* A = (float*)malloc(M*P*sizeof(float));
96. float* B = (float*)malloc(P*N*sizeof(float));
97. float* C1 = (float*)malloc(M*N*sizeof(float));
98. float* C2 = (float*)malloc(M*N*sizeof(float));
99. for (int i = 0; i<M; i++)
100. for (int j = 0; j < P; j++)
101. A[i*P + j] = rand() % 10;
102. for (int i = 0; i<P; i++)
103. for (int j = 0; j < N; j++)
104. B[i*N + j] = rand() % 10;
105. #ifdef SHOW
106. for (int i = 0; i<M; i++){
107. for (int j = 0; j<P; j++)
108. cout << A[i*P + j] << " ";
109. cout << endl;
111. }
112. for (int i = 0; i<P; i++){
113. for (int j = 0; j<N; j++)
114. cout << B[i*N + j] << " ";
115. cout << endl;
116. }
117. #endif
119. compute_cpu(A, B, C1);
120. #ifdef SHOW
121. for (int i = 0; i<M; i++){
122. for (int j = 0; j<N; j++)
123. cout << C1[i*N + j] << " ";
124. cout << endl;
126. }
127. #endif
128. compute_gpu(A, B, C2);
129. #ifdef SHOW
130. for (int i = 0; i<M; i++){
131. for (int j = 0; j<N; j++)
132. cout << C2[i*N + j] << " ";
133. cout << endl;
135. }
136. #endif
137. verify(C1, C2, M, N);
138. free(A);
139. free(B);
140. free(C1);
141. free(C2);
142. return 0;
143. }

來自爲知筆記(Wiz)