vector addition in CUDA using streams

Question

I m a beginner in CUDA parallel programming. I tried a program for vector addition using CUDA streams. When i compile, i get the following error.

Solution is not correct. The solution did not match the expected results at row 0. Expecting (1+0.5=1.5) but got 0.

I checked with cuda by examples book and similar questions online. Couldn't find a solution. Can anyone help me solve this error? Thanks in advance.

#include    <wb.h>
#define wbCheck(stmt) do {                      \
    cudaError_t err = stmt;                     \
    if (err != cudaSuccess) {                   \
        wbLog(ERROR, "Failed to run stmt ", #stmt);         \
        wbLog(ERROR, "Got CUDA error ...  ", cudaGetErrorString(err)); \
        return -1;                          \
    }                               \
} while(0)


__global__ void vecAdd(float * in1, float * in2, float * out, int len) 
{
int i = threadIdx.x + blockDim.x * blockIdx.x;
  if (i < len)
out[i] = in1[i] + in2[i];
}

int main(int argc, char ** argv) 
{
cudaStream_t stream0, stream1,stream2,stream3;
cudaStreamCreate(&stream0);
cudaStreamCreate(&stream1);
cudaStreamCreate(&stream2);
cudaStreamCreate(&stream3);
wbArg_t args;
int inputLength;

float *h_A, *h_B, *h_C;
float *d_A0, *d_B0, *d_C0;
float *d_A1, *d_B1, *d_C1;
float *d_A2, *d_B2, *d_C2;
float *d_A3, *d_B3, *d_C3;

args = wbArg_read(argc, argv);

wbTime_start(Generic, "Importing data and creating memory on host");
h_A = (float *) wbImport(wbArg_getInputFile(args, 0), &inputLength);
h_B = (float *) wbImport(wbArg_getInputFile(args, 1), &inputLength);
h_C = (float *) malloc(inputLength * sizeof(float));
wbTime_stop(Generic, "Importing data and creating memory on host");
wbLog(TRACE, "The input length is ", inputLength);
wbLog(TRACE, "h_A ", *h_A);
wbLog(TRACE, "h_B", *h_B);


int size = inputLength * sizeof(float);
int SegSize = inputLength/4;


wbCheck(cudaMalloc((void **) &d_A0, size));
wbCheck(cudaMalloc((void **) &d_B0, size));
wbCheck(cudaMalloc((void **) &d_C0, size));

wbCheck(cudaMalloc((void **) &d_A1, size));
wbCheck(cudaMalloc((void **) &d_B1, size));
wbCheck(cudaMalloc((void **) &d_C1, size));

wbCheck(cudaMalloc((void **) &d_A2, size));
wbCheck(cudaMalloc((void **) &d_B2, size));
wbCheck(cudaMalloc((void **) &d_C2, size));

wbCheck(cudaMalloc((void **) &d_A3, size));
wbCheck(cudaMalloc((void **) &d_B3, size));
wbCheck(cudaMalloc((void **) &d_C3, size));


cudaHostAlloc((void **) &h_A, size, cudaHostAllocDefault);
cudaHostAlloc((void **) &h_B, size, cudaHostAllocDefault);
cudaHostAlloc((void **) &h_C, size, cudaHostAllocDefault);


dim3 DimGrid((inputLength -1)/256 +1 , 1 , 1);
dim3 DimBlock(256 , 1, 1);


for (int i=0; i<size; i+=inputLength*4) 
{
cudaMemcpyAsync(d_A0, h_A+i, SegSize*sizeof(float),cudaMemcpyHostToDevice, stream0);
cudaMemcpyAsync(d_B0, h_B+i, SegSize*sizeof(float),cudaMemcpyHostToDevice, stream0);

cudaMemcpyAsync(d_A1, h_A+i+SegSize, SegSize*sizeof(float),cudaMemcpyHostToDevice,stream1);
cudaMemcpyAsync(d_B1, h_B+i+SegSize, SegSize*sizeof(float),cudaMemcpyHostToDevice,stream1);

cudaMemcpyAsync(d_A2, h_A+i+SegSize+SegSize, SegSize*sizeof(float),cudaMemcpyHostToDevice, stream2);
cudaMemcpyAsync(d_B2, h_B+i+SegSize+SegSize, SegSize*sizeof(float),cudaMemcpyHostToDevice, stream2);

cudaMemcpyAsync(d_A3, h_A+i+SegSize+SegSize+SegSize, SegSize*sizeof(float),cudaMemcpyHostToDevice, stream3);
cudaMemcpyAsync(d_B3, h_B+i+SegSize+SegSize+SegSize, SegSize*sizeof(float),cudaMemcpyHostToDevice, stream3);

vecAdd<<<DimGrid, DimBlock, 0, stream0>>>(d_A0, d_B0, d_C0,inputLength);
vecAdd<<<DimGrid, DimBlock, 0, stream1>>>(d_A1, d_B1, d_C1,inputLength);
vecAdd<<<DimGrid, DimBlock, 0, stream2>>>(d_A2, d_B2, d_C2,inputLength);
vecAdd<<<DimGrid, DimBlock, 0, stream3>>>(d_A3, d_B3, d_C3,inputLength);


cudaDeviceSynchronize();


cudaMemcpyAsync(h_C+i, d_C0, SegSize*sizeof(float),cudaMemcpyDeviceToHost, stream0);
cudaMemcpyAsync(h_C+i+SegSize, d_C1, SegSize*sizeof(float),cudaMemcpyDeviceToHost,stream1);
cudaMemcpyAsync(h_C+i+SegSize+SegSize, d_C2, SegSize*sizeof(float),cudaMemcpyDeviceToHost,stream2);
cudaMemcpyAsync(h_C+i+SegSize+SegSize+SegSize, d_C3, SegSize*sizeof(float),cudaMemcpyDeviceToHost,stream3);
    wbLog(TRACE, "on addition is ", *h_C);

}

cudaFree(d_A0);
cudaFree(d_B0);
cudaFree(d_C0);

cudaFree(d_A1);
cudaFree(d_B1);
cudaFree(d_C1);

cudaFree(d_A2);
cudaFree(d_B2);
cudaFree(d_C2);

cudaFree(d_A3);
cudaFree(d_B3);
cudaFree(d_C3);

wbSolution(args, h_C, inputLength);
cudaFreeHost(h_A);
cudaFreeHost(h_B);
cudaFreeHost(h_C);

return 0;
}

Answer 1

One problem is how you are handling h_A, h_B, and h_C:

h_A = (float *) wbImport(wbArg_getInputFile(args, 0), &inputLength);
h_B = (float *) wbImport(wbArg_getInputFile(args, 1), &inputLength);

The above lines of code are creating an allocation for h_A and h_B and importing some data (presumably).

These lines of code:

cudaHostAlloc((void **) &h_A, size, cudaHostAllocDefault);
cudaHostAlloc((void **) &h_B, size, cudaHostAllocDefault);
cudaHostAlloc((void **) &h_C, size, cudaHostAllocDefault);

Are not doing what you think. They are creating a new allocation for h_A, h_B and h_C. Whatever data those pointers previously referenced is no longer accessible from those pointers (i.e. for all intents and purposes, it is lost).

CUDA should be able to work just fine with the pointers and allocations being created here:

h_A = (float *) wbImport(wbArg_getInputFile(args, 0), &inputLength);
h_B = (float *) wbImport(wbArg_getInputFile(args, 1), &inputLength);
h_C = (float *) malloc(inputLength * sizeof(float));

So delete these lines of code:

cudaHostAlloc((void **) &h_A, size, cudaHostAllocDefault);
cudaHostAlloc((void **) &h_B, size, cudaHostAllocDefault);
cudaHostAlloc((void **) &h_C, size, cudaHostAllocDefault);

and delete these:

cudaFreeHost(h_A);
cudaFreeHost(h_B);
cudaFreeHost(h_C);

And you should be closer to a solution.