template  __global__ void
matrixMulCUDA_3_2w1w(float* matrixMulCUDA_4_2w1w(float* C, float* A, float* B, int wA, int wB)
{
        // Block index
        int bx = blockIdx.x;
        int by = blockIdx.y;
        // Thread index
        int tx = threadIdx.x;
x % BLOCK_SIZE;
        int ty = threadIdx.y;

        for (int x = 1; x <= 2; x++)
        {
                // Index of the first sub-matrix of A processed by the block
                int aBegin = wA * BLOCK_SIZE * by;
                // Index of the last sub-matrix of A processed by the block
                int aEnd = aBegin + wA - 1;
                // Step size used to iterate through the sub-matrices of A
                int aStep = BLOCK_SIZE;
                // Index of the first sub-matrix of B processed by the block
                int bBegin = BLOCK_SIZE * bx;
                // Step size used to iterate through the sub-matrices of B
                int bStep = BLOCK_SIZE * wB;
                // Csub is used to store the element of the block sub-matrix
                // that is computed by the thread
                float Csub = 0;
                // Declaration of the shared memory array Bs used to
                // store the sub-matrix of B
                __shared__ float Bs[BLOCK_SIZE][BLOCK_SIZE];
                __shared__ float BBs[BLOCK_SIZE][BLOCK_SIZE];
                __shared__ float As[BLOCK_SIZE / 2][BLOCK_SIZE];
                __shared__ float AAs[BLOCK_SIZE / 2][BLOCK_SIZE];
                // Initial load of data
                if (x == 1)
                {
                        if (ty < BLOCK_SIZE / 2)
                        {
                                As[ty][tx] = A[aBegin + wA * ty + tx];
                        }
                        Bs[ty][tx] = B[bBegin + wB * ty + tx];
                }
                else if (x == 2)
                {
                        if (ty >= BLOCK_SIZE / 2)
                        {
                                As[ty][tx] = A[aBegin + wA * ty + tx];
                        }
                }
                // Loop over all the sub-matrices of A and B
                // required to compute the block sub-matrix
                for (int a = aBegin, b = bBegin;
                        a <= aEnd;
                        a += aStep, b += bStep)
                {
                        // Declaration of Copy values between matrixes
                        if (x == 1 && ty < BLOCK_SIZE / 2)
                        {
                                AAs[ty][tx] = As[ty][tx];
                        }
                        else if (x == 2 && ty >= BLOCK_SIZE / 2)
                        {
                                AAs[ty][tx] = As[ty][tx];
                        }
                        BBs[ty][tx] = Bs[ty][tx];
                        // Synchronize to make sure the shared memory array As used to
                        // store the sub-matrix of A
                        __shared__ float As[BLOCK_SIZE / 2][BLOCK_SIZE];
                        // Declaration of the shared memory array Bs used to
                        // store the sub-matrix of B
                        __shared__ float Bs[BLOCK_SIZE][BLOCK_SIZE];
matrices are loaded
                        __syncthreads();
                        // Load the matrices from device memory
                        // to shared memory; each thread loads
                        // one element of each matrix
                        if (x == 1)
                        {
                                if (ty < BLOCK_SIZE / 2)
                                {
                                        As[ty][tx] = A[a + wA * ty + tx];
                                }
                                Bs[ty][tx] = B[b + wB * ty + tx];
                        }
                        else if (x == 2)
                        {
                                if (ty >= BLOCK_SIZE / 2)
                                {
                                        As[ty][tx] = A[a + wA * ty + tx];
                                }
                        }
                        // Synchronize to make sure the matrices are loaded
                        __syncthreads();
                        // Multiply the two matrices together;
                        // 2 thread computes one element
                        // of the block sub-matrix
#pragma unroll
                        for (int k = 0; k < BLOCK_SIZE / 2; k++)
                        {
                                if (ty < BLOCK_SIZE / 2)
                                {
                                        Csub += As[ty][k] * Bs[k][tx];
                                }
                                else
                                {
                                        Csub += As[ty][k + BLOCK_SIZE / 2] * Bs[k + BLOCK_SIZE / 2][tx];
                                }
                        }
                        // Synchronize to make sure that the preceding
                        // computation is done before loading two new
                        // sub-matrices of A and B in the next iteration
                        __syncthreads();
                }
                // Write the block sub-matrix to device memory;
                // each thread writes one element
                int c = wB * BLOCK_SIZE * by + BLOCK_SIZE * bx;
                if (ty < BLOCK_SIZE / 2 && x == 1)
                {
                        C[c + wB * ty + tx] += Csub;
                }
                else if (ty >= BLOCK_SIZE / 2 && x == 2)
                {
                        C[c + wB * ty + tx] += Csub;
                }
        }
}