1. template <int BLOCK_SIZE> __global__ void
2. matrixMulCUDA_4(float* C, float* A, float* B, int wA, int wB)
3. {
4.         // Block index
5.         int bx = blockIdx.x;
6.         int by = blockIdx.y;
7.  
8.         // Thread index
9.         int tx = threadIdx.x;
10.         int ty = threadIdx.y;
11.  
12.         // Index of the first sub-matrix of A processed by the block
13.         int aBegin = wA * BLOCK_SIZE * by;
14.  
15.         // Index of the last sub-matrix of A processed by the block
16.         int aEnd = aBegin + wA - 1;
17.  
18.         // Step size used to iterate through the sub-matrices of A
19.         int aStep = BLOCK_SIZE;
20.  
21.         // Index of the first sub-matrix of B processed by the block
22.         int bBegin = BLOCK_SIZE * bx;
23.  
24.         // Step size used to iterate through the sub-matrices of B
25.         int bStep = BLOCK_SIZE * wB;
26.  
27.         // Csub is used to store the element of the block sub-matrix
28.         // that is computed by the thread
29.         float Csub = 0;
30.  
31.         // Declaration of the shared memory array Aa used to
32.         // store the sub-matrix of A
33.         __shared__ float Aa[BLOCK_SIZE][BLOCK_SIZE];
34.  
35.         // Declaration of the shared memory array Ba used to
36.         // store the sub-matrix of B
37.         __shared__ float Ba[BLOCK_SIZE][BLOCK_SIZE];
38.  
39.         // Declaration of the shared memory array Aa used to
40.         // store the sub-matrix of A
41.         __shared__ float Ab[BLOCK_SIZE][BLOCK_SIZE];
42.  
43.         // Declaration of the shared memory array Ba used to
44.         // store the sub-matrix of B
45.         __shared__ float Bb[BLOCK_SIZE][BLOCK_SIZE];
46.  
47.         // Load the matrices from device memory
48.         // to shared memory; each thread loads
49.         // one element of each matrix
50.         Aa[ty][tx] = A[aBegin + wA * ty + tx];
51.         Ba[ty][tx] = B[bBegin + wB * ty + tx];
52.  
53.         // Loop over all the sub-matrices of A and B
54.         // required to compute the block sub-matrix
55.         for (int a = aBegin, b = bBegin;
56.                 a <= aEnd;
57.                 a += aStep, b += bStep)
58.         {
59.                 // Load the matrices from sub-matrix A to sub-matrix B
60.                 Ab[ty][tx] = Aa[ty][tx];
61.                 Bb[ty][tx] = Ba[ty][tx];
62.  
63.                 // Load the matrices from device memory
64.                 // to shared memory A;
65.                 // In last iteration skips this part
66.                 if (a + aStep < aEnd) {
67.                         Aa[ty][tx] = A[(a + aStep) + wA * ty + tx];
68.                         Ba[ty][tx] = B[(b + bStep) + wB * ty + tx];
69.                 }
70.  
71.                 // Synchronize to make sure the matrices are loaded
72.                 __syncthreads();
73.  
74.                 // Multiply the two matrices together;
75.                 // each thread computes one element
76.                 // of the block sub-matrix
77. #pragma unroll
78.  
79.                 for (int k = 0; k < BLOCK_SIZE; ++k)
80.                 {
81.                         Csub += Ab[ty][k] * Bb[k][tx];
82.                 }
83.  
84.                 // Synchronize to make sure that the preceding
85.                 // computation is done before loading two new
86.                 // sub-matrices of A and B in the next iteration
87.                 __syncthreads();
88.         }
89.  
90.         // Write the block sub-matrix to device memory;
91.         // each thread writes one element
92.         int c = wB * BLOCK_SIZE * by + BLOCK_SIZE * bx;
93.         C[c + wB * ty + tx] = Csub;
94. }
95.