Python 3.10.12 : Few example for CUDA and NVCC - part 044.

NVCC use CUDA C/C++ source code and allows developers to write high-performance GPU-accelerated applications by leveraging the power of NVIDIA GPUs for parallel processing tasks.

Today I test some simple examples with this tool on Google Colab using the nvcc4jupyter python package.

You need to install it with the pip and know how to use the CUDA C/C++ source code, or use the basic example from documentation.

pip install nvcc4jupyter

I change some source code because is need to install this library and I don't have time to learn and test.

But this will allow me to test better, because on my desktop I don't have a good hardware.

This is the source I change and I cut the source code linked on error_handling.h.

This is the changed source code , you can see more on my GitHub repo for Google Colab ! !

#include 
//#include "error_handling.h"

const int DSIZE = 4096;
const int block_size = 256;

// vector add kernel: C = A + B
__global__ void vadd(const float *A, const float *B, float *C, int ds){
    int idx = threadIdx.x + blockIdx.x * blockDim.x;
    if (idx < ds) {
        C[idx] = A[idx] + B[idx];
    }
}

int main(){
    float *h_A, *h_B, *h_C, *d_A, *d_B, *d_C;

    // allocate space for vectors in host memory
    h_A = new float[DSIZE];
    h_B = new float[DSIZE];
    h_C = new float[DSIZE];

    // initialize vectors in host memory to random values (except for the
    // result vector whose values do not matter as they will be overwritten)
    for (int i = 0; i < DSIZE; i++) {
        h_A[i] = rand()/(float)RAND_MAX;
        h_B[i] = rand()/(float)RAND_MAX;
    }

    // allocate space for vectors in device memory
    cudaMalloc(&d_A, DSIZE*sizeof(float));
    cudaMalloc(&d_B, DSIZE*sizeof(float));
    cudaMalloc(&d_C, DSIZE*sizeof(float));
    //cudaCheckErrors("cudaMalloc failure"); // error checking

    // copy vectors A and B from host to device:
    cudaMemcpy(d_A, h_A, DSIZE*sizeof(float), cudaMemcpyHostToDevice);
    cudaMemcpy(d_B, h_B, DSIZE*sizeof(float), cudaMemcpyHostToDevice);
    //cudaCheckErrors("cudaMemcpy H2D failure");

    // launch the vector adding kernel
    vadd<<<(DSIZE+block_size-1)/block_size, block_size>>>(d_A, d_B, d_C, DSIZE);
    //cudaCheckErrors("kernel launch failure");

    // wait for the kernel to finish execution
    cudaDeviceSynchronize();
    //cudaCheckErrors("kernel execution failure");

    cudaMemcpy(h_C, d_C, DSIZE*sizeof(float), cudaMemcpyDeviceToHost);
    //cudaCheckErrors("cudaMemcpy D2H failure");

    printf("A[0] = %f\n", h_A[0]);
    printf("B[0] = %f\n", h_B[0]);
    printf("C[0] = %f\n", h_C[0]);
    return 0;
}

This is result ...

A[0] = 0.840188
B[0] = 0.394383
C[0] = 0.000000