__syncthreads() not syncing the threads, although not in if statement

Accelerated Computing CUDA CUDA Programming and Performance
1 
__global__ void matVecMultCUDAKernel(int* aOnGPU, int* bOnGPU, int* cOnGPU, int matSize) {
	__shared__ int aShared[BLOCK_SIZE][BLOCK_SIZE];
	__shared__ int bShared[BLOCK_SIZE];

	int myRow = blockIdx.x * blockDim.x + threadIdx.x;
	int myRowInBlock = threadIdx.x, myColInBlock = threadIdx.y;
	int rowSum = 0;

	for (int m = 0; m < (matSize + BLOCK_SIZE - 1) / BLOCK_SIZE; m++) {
		aShared[myRowInBlock][myColInBlock] = getValFromMatrix(aOnGPU,myRow,m*BLOCK_SIZE+myColInBlock,matSize);
		if (myColInBlock==0) {bShared[myRowInBlock] = getValFromVector(bOnGPU,m*BLOCK_SIZE+myRowInBlock,matSize);}

		__syncthreads(); // Sync threads to make sure all fields have been written by all threads in the block to cShared and xShared

		if (myColInBlock==0) {
			for (int k=0;k<BLOCK_SIZE;k++) {
//				rowSum += getValFromMatrix(aOnGPU,myRow,m*BLOCK_SIZE+k,matSize) * getValFromVector(bOnGPU,m*BLOCK_SIZE+k,matSize);
				rowSum += aShared[myRowInBlock][k] * bShared[k];
			}
		}
	}

	if (myColInBlock==0 && myRow<matSize) {cOnGPU[myRow] = rowSum;}
}

The above kernel gives incorrent results for matrix-vector multiplication. In debugging, I can see that some warps in the block are executing when m=0 and others when m=1. If i decomment line 17 and comment line 18, the results are correct again. Is there something wrong with how I’m using _syncthreads()?

2

I think i figured it out, i needed to add an additional __syncthreads(); between line 20 and 21, since otherwise threads may update the shared arrays before it has been read fully