1. template <int BLOCK_SIZE> __global__ void
2. matrixMulCUDA_4_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.         // Declaration of the shared memory array Bs used to
32.         // store the sub-matrix of B
33.         __shared__ float Bs[BLOCK_SIZE][BLOCK_SIZE];
34.         __shared__ float BBs[BLOCK_SIZE][BLOCK_SIZE];
35.         __shared__ float As[BLOCK_SIZE][BLOCK_SIZE];
36.         __shared__ float AAs[BLOCK_SIZE][BLOCK_SIZE];
37.  
38.         // Initial load of data
39.         As[ty][tx] = A[aBegin + wA * ty + tx];
40.         Bs[ty][tx] = B[bBegin + wB * ty + tx];
41.  
42.         // Loop over all the sub-matrices of A and B
43.         // required to compute the block sub-matrix
44.         for (int a = aBegin, b = bBegin;
45.                 a <= aEnd;
46.                 a += aStep, b += bStep)
47.         {
48.  
49.                 // Copy values between matrixes
50.                 AAs[ty][tx] = As[ty][tx];
51.                 BBs[ty][tx] = Bs[ty][tx];
52.  
53.                 // Synchronize to make sure the matrices are loaded
54.                 __syncthreads();
55.  
56.                 // Load the matrices from device memory
57.                 // to shared memory; each thread loads
58.                 // one element of each matrix
59.                 As[ty][tx] = A[a + wA * ty + tx];
60.                 Bs[ty][tx] = B[b + wB * ty + tx];
61. #pragma unroll
62.                 for (int k = 0; k < BLOCK_SIZE / 2; k++)
63.                 {
64.                         Csub += AAs[ty][k + (BLOCK_SIZE / 2) * tx] * BBs[k + (BLOCK_SIZE / 2) * tx][tx];
65.                 }
66.  
67.                 // Synchronize to make sure that the preceding
68.                 // computation is done before loading two new
69.                 // sub-matrices of A and B in the next iteration
70.                 __syncthreads();
71.         }
72.  
73.         // Write the block sub-matrix to device memory;
74.         // each thread writes one element
75.         int c = wB * BLOCK_SIZE * by + BLOCK_SIZE * bx;
76.         C[c + wB * ty + tx] += Csub;
77. }
78.