1. template <int BLOCK_SIZE> __global__ void
2. matrixMulCUDA_3_2w1w(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 % BLOCK_SIZE;
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.         // Loop over all the sub-matrices of A and B
32.         // required to compute the block sub-matrix
33.         for (int a = aBegin, b = bBegin;
34.                 a <= aEnd;
35.                 a += aStep, b += bStep)
36.         {
37.                 // Declaration of the shared memory array As used to
38.                 // store the sub-matrix of A
39.                 __shared__ float As[BLOCK_SIZE][BLOCK_SIZE];
40.  
41.                 // Declaration of the shared memory array Bs used to
42.                 // store the sub-matrix of B
43.  
44.                 __shared__ float Bs[BLOCK_SIZE][BLOCK_SIZE];
45.                 // Load the matrices from device memory
46.                 // to shared memory; each thread loads
47.                 // one element of each matrix
48.                 As[ty][tx] = A[a + wA * ty + tx];
49.                 Bs[ty][tx] = B[b + wB * ty + tx];
50.  
51.                 // Synchronize to make sure the matrices are loaded
52.                 __syncthreads();
53.  
54.                 // Multiply the two matrices together;
55.                 // each thread computes one element
56.                 // of the block sub-matrix
57.  
58. #pragma unroll
59.                 for (int k = 0; k < BLOCK_SIZE / 2; k++)
60.                 {
61.                         Csub += As[ty][k + (BLOCK_SIZE / 2) * tx] * Bs[k + (BLOCK_SIZE / 2) * tx][tx];
62.                 }
63.  
64.                 // Synchronize to make sure that the preceding
65.                 // computation is done before loading two new
66.                 // sub-matrices of A and B in the next iteration
67.                 __syncthreads();
68.         }
69.  
70.         // Write the block sub-matrix to device memory;
71.         // each thread writes one element
72.         int c = wB * BLOCK_SIZE * by + BLOCK_SIZE * bx;
73.         C[c + wB * ty + tx] += Csub;
74. }
75.