Wrong results using CUDA streams and memCpyAsync, become correct adding cudaDeviceSynchronize

Question

I'm working on a CUDA matrix multiplication, but I did some modifications to observe how they affect performances.

I want to observe the behavior and performances of a matrix multiplication kernel, making some changes. I'm measuring the changes in GPU events time, I'm testing it in two speicific different conditions:

• I have an amount of matrices (say matN) for A, B and C, then I transfer (H2D) one matrix for A, one for B and multply them, to transfer back (D2H) one C;

• I have matN for A, B and C, but I transfer >1(say chunk) matrices for A and for B, I compute exactly chunk multiplications, and transfer back chunk result matrices C.

In the first case (chunk = 1) all works as expected, but in the second case (chunk > 1) I get some of Cs are correct, while others are not.

But if I put a cudaDeviceSynchronize() after the cudaMemcpyAsync, I get correct results.

Here's the code doing what I've just described:

/**** main.cpp ****/

    int chunk = matN/iters;    
    #ifdef LOWPAR
        GRIDx= 1;
        GRIDy= 1;
        label="LOW";
    #else
       int sizeX = M;
       int sizeY = N;
       GRIDx = ceil((sizeX)/BLOCK);
       GRIDy = ceil((sizeY)/BLOCK);
       label="";
    #endif

    const int bytesA = M*K*sizeof(float);
    const int bytesB = K*N*sizeof(float);
    const int bytesC = M*N*sizeof(float);

    //device mem allocation
    float *Ad, *Bd, *Cd;
    gpuErrchk( cudaMalloc((void **)&Ad, bytesA*chunk) );
    gpuErrchk( cudaMalloc((void **)&Bd, bytesB*chunk) );
    gpuErrchk( cudaMalloc((void **)&Cd, bytesC*chunk) );
    //host pinned mem allocation
    float *A, *B, *C;
    gpuErrchk( cudaMallocHost((void **)&A, bytesA*matN) );
    gpuErrchk( cudaMallocHost((void **)&B, bytesB*matN) );
    gpuErrchk( cudaMallocHost((void **)&C, bytesC*matN) );

    //host data init
    for(int i=0; i<matN; ++i){
        randomMatrix(M, K, A+(i*M*K));
        randomMatrix(K, N, B+(i*K*N));
    } 

    //event start
    createAndStartEvent(&startEvent, &stopEvent);

    if (square)
    {          
        label += "SQUARE";
        int size = N*N;
        for (int i = 0; i < iters; ++i) { 
            int j = i%nStream;            
            int idx = i*size*chunk;
            newSquareMatMulKer(A+idx, B+idx, C+idx, Ad, Bd, Cd, N, chunk, stream[j]); 
        }
    }
    else {
        ...
    } 

    msTot = endEvent(&startEvent, &stopEvent);
    #ifdef MEASURES          
        printMeasures(square, label, msTot, millis.count(), matN, iters, devId);
    #else
        float *_A, *_B, *_C, *tmpC;
        tmpC = (float *)calloc(1,bytesC*chunk);
        for (int s=0; s<matN; ++s)
        {
            _A = A+(s*M*K);
            _B = B+(s*K*N);
            _C = C+(s*M*N);
            memset(tmpC, 0, bytesC*chunk);

            hostMatMul(_A, _B, tmpC, M, K, N);
            checkMatEquality(_C, tmpC, M, N);
        }   
    #endif


/**** matmul.cu ****/

__global__ void squareMatMulKernel(float* A, float* B, float* C, int N, int chunk) {

    int ROW = blockIdx.x*blockDim.x+threadIdx.x;
    int COL = blockIdx.y*blockDim.y+threadIdx.y;
  

    if (ROW<N && COL<N) {
        int size=N*N;
        int offs = 0;
        float tmpSum=0.0f;
        
        for (int s=0; s<chunk; ++s)
        {
            offs = s*size;
            tmpSum = 0.0f;
        
            for (int i = 0; i < N; ++i) {
                tmpSum += A[offs+(ROW*N)+i] * B[offs+(i*N)+COL];
            }
        
            C[offs+(ROW*N)+COL] = tmpSum;
        }
    }
    return ;
}




void newSquareMatMulKer(float *A, float *B, float *C, float *Ad, float *Bd, float *Cd, 
            int n, int chunk, cudaStream_t strm)
{
    int size = n*n;
    int bytesMat = size*sizeof(float);

    dim3 dimBlock(BLOCK,BLOCK,1);
    dim3 dimGrid(GRIDx, GRIDy,1); 

    gpuErrchk( cudaMemcpyAsync(Ad, A, bytesMat*chunk, cudaMemcpyHostToDevice, strm) );    
    gpuErrchk( cudaMemcpyAsync(Bd, B, bytesMat*chunk, cudaMemcpyHostToDevice, strm) );   

    #ifdef LOWPAR
        squareMatMulGridStrideKer<<<dimGrid, dimBlock, 0, strm>>>(Ad, Bd, Cd, n, chunk);
    #else
        squareMatMulKernel<<<dimGrid, dimBlock, 0, strm>>>(Ad, Bd, Cd, n, chunk);
    #endif
    squareMatMulKernel<<<dimGrid, dimBlock, 0, strm>>>(Ad, Bd, Cd, n, chunk);

    gpuErrchk( cudaMemcpyAsync( C, Cd, bytesMat*chunk, cudaMemcpyDeviceToHost, strm) );

    cudaDeviceSynchronize();
        ^ ^ ^ ^ ^ ^
}

I tried to debug using cuda-gdb but nothing strange showed up, gpuErrchk doesn't throw errors in CUDA API calls. I run the code using memcheck too, both with and without cudaDeviceSynchronize and I got no error.

I think it can be a synchronization issue, but I can't understand the reason behind that. Can someone spot where I'm wrong? Other code advices are really appreciated too.

Answer 1

If you are using multiples streams, you may override Ad and Bd before using them.

Example with iters = 2 and nStream = 2 :

for (int i = 0; i < iters; ++i) { 
  int j = i%nStream;            
  int idx = i*size*chunk;
  newSquareMatMulKer(A+idx, B+idx, C+idx, Ad, Bd, Cd, N, chunk, stream[j]); 
}

From this loop, you will call

newSquareMatMulKer(A, B, C, Ad, Bd, Cd, N, chunk, stream[0]); // call 0
newSquareMatMulKer(A+idx, B+idx, C+idx, Ad, Bd, Cd, N, chunk, stream[1]); // call 1

As you are using the same memory area on device for both call, you may have several synchronizations issues:

• call 1 start to copy A and B on device before call 0:squareMatMulKernel end, so you may use incorrect values of A and/or B to compute your first iteration.

• call 1:squareMatMulKernel start before you retrieve the values of C from call 0, so you may override C with values from call 1.

To fix this problem, I see two approaches:

• Using synchronization as in your example with cudaDeviceSynchronize();.

• You can allocate more memory two device side (one workspace per stream), for example.

''

//device mem allocation
float *Ad, *Bd, *Cd;
gpuErrchk( cudaMalloc((void **)&Ad, bytesA*chunk*nStream) );
gpuErrchk( cudaMalloc((void **)&Bd, bytesB*chunk*nStream) );
gpuErrchk( cudaMalloc((void **)&Cd, bytesC*chunk*nStream) );

/* code here */

for (int i = 0; i < iters; ++i) { 
  int j = i%nStream;            
  int idx = i*size*chunk;
  int offset_stream = j*size*chunk;
  newSquareMatMulKer(A+idx, B+idx, C+idx, 
    Ad + offset_stream , 
    Bd + offset_stream , 
    Cd + offset_stream , N, chunk, stream[j]); 
}

In this case you don't need synchronization before the end of the loop.