1. template <int BLOCK_SIZE> __global__ void
2. matrixMulCUDA_4_1w2w(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[2] = { 0,0 };
30.  
31.         // Declaration of the shared memory array As used to
32.         // store the sub-matrix of A
33.         __shared__ float Aa[BLOCK_SIZE][BLOCK_SIZE];
34.         __shared__ float Ab[BLOCK_SIZE][BLOCK_SIZE];
35.  
36.         // Declaration of the shared memory array Bs used to
37.         // store the sub-matrix of B
38.         __shared__ float Bs[BLOCK_SIZE][2 * BLOCK_SIZE];
39.         __shared__ float Bb[BLOCK_SIZE][2 * BLOCK_SIZE];
40.  
41.         // Initial load
42.         Aa[ty][tx] = A[aBegin + wA * ty + tx];
43.         Bs[ty][tx] = B[bBegin + wB * ty + tx];
44.         Bs[ty][tx + BLOCK_SIZE] = B[bBegin + BLOCK_SIZE * wB * ty + tx];
45.  
46.         // Synchronize to make sure that initial matrices are loaded
47.         __syncthreads();
48.  
49.         // Loop over all the sub-matrices of A and B
50.         // required to compute the block sub-matrix
51.         for (int a = aBegin, b = bBegin;
52.                 a <= aEnd;
53.                 a += aStep, b += bStep)
54.         {
55.                 // copy contents betweeen shared matrixes
56.                 Ab[ty][tx] = Aa[ty][tx];
57.                 Bb[ty][tx] = Bs[ty][tx];
58.                 Bb[ty][tx + BLOCK_SIZE] = Bs[ty][tx + BLOCK_SIZE];
59.  
60.                 // Synchronize to make sure the matrices are loaded
61.                 __syncthreads();
62.  
63.                 // Load the matrices from device memory
64.                 // to shared memory; each thread loads
65.                 // two elements of each matrix
66.                 if (a + aStep < aEnd) {
67.                         Aa[ty][tx] = A[a + wA * ty + tx];
68.                         Bs[ty][tx] = B[b + wB * ty + tx];
69.                         Bs[ty][tx + BLOCK_SIZE] = B[b + BLOCK_SIZE * wB * ty + tx];
70.                 }
71.  
72.                 // Multiply the two matrices together;
73.                 // each thread computes one element
74.                 // of the block sub-matrix
75.                 for (int g = 0; g < 2; g++) {
76. #pragma unroll
77.                         for (int k = 0; k < BLOCK_SIZE; ++k)
78.                         {
79.                                 CSub[g] += Ab[ty][k + (g * BLOCK_SIZE)] * Bb[k][tx + (g * BLOCK_SIZE)];
80.                         }
81.                 }
82.                
83.                 // Synchronize to make sure that the preceding
84.                 // computation is done before loading two new
85.                 // sub-matrices of A and B in the next iteration
86.                 __syncthreads();
87.         }
88.        
89.         // Write the block sub-matrix to device memory;
90.         // each thread writes one element
91.         int c = wB * BLOCK_SIZE * by + BLOCK_SIZE * bx;
92.         for (int g = 0; g < 2; g++) {
93.                 C[c + (BLOCK_SIZE * g) + wB * ty + tx] = CSub[g];
94.         }
95. }