1
__global__ void matrixMul(float *A, float *B, float *C, int width) {
2
    int row = blockIdx.y * blockDim.y + threadIdx.y;
3
    int col = blockIdx.x * blockDim.x + threadIdx.x;
4
    
5
    float sum = 0.0f;
6
    for (int i = 0; i < width; i++) {
7
        sum += A[row * width + i] * B[i * width + col];
8
    }
9
    
10
    C[row * width + col] = sum;
11
}

1
__global__ void matrixMulOptimized(float *A, float *B, float *C, int width) {
2
    __shared__ float sharedA[TILE_SIZE][TILE_SIZE];
3
    __shared__ float sharedB[TILE_SIZE][TILE_SIZE];
4
    
5
    int bx = blockIdx.x, by = blockIdx.y;
6
    int tx = threadIdx.x, ty = threadIdx.y;
7
    
8
    int row = by * TILE_SIZE + ty;
9
    int col = bx * TILE_SIZE + tx;
10
    
11
    float sum = 0.0f;
12
    for (int tile = 0; tile < width/TILE_SIZE; tile++) {
13
        sharedA[ty][tx] = A[row * width + (tile * TILE_SIZE + tx)];
14
        sharedB[ty][tx] = B[(tile * TILE_SIZE + ty) * width + col];
15
        __syncthreads();
16
        
17
        #pragma unroll
18
        for (int i = 0; i < TILE_SIZE; i++) {
19
            sum += sharedA[ty][i] * sharedB[i][tx];
20
        }
21
        __syncthreads();
22
    }
23
    
24
    C[row * width + col] = sum;
25
}