Is CUDA good? The gap between CUDA and CPU in computing

LitchiCheng

Is CUDA good? The gap between CUDA and CPU in computing [Copy link]

邀请：@chenzhufly   @skywalker_lee   @wsdymg   @bigbat   参与回复

Test platform

Jetson Music Nano

Test code

Code repository (please give me a STAR if it is useful to you!)

https://github.com/LitchiCheng/CUDA_Test

The following is some test code:

#include <cstdio>
#include "cuda_runtime.h"

#include <sys/time.h>
#include <time.h>

#include <math.h>

#include "utility/timecost.h"

__global__ void kernelAdd(float * A, float * B, float * C)
{
    int ix=threadIdx.x+blockDim.x*blockIdx.x;
    int iy=threadIdx.y+blockDim.y*blockIdx.y;
    int idx=ix+iy*blockDim.x*gridDim.x;
    C[idx]=cos(A[idx])+sin(B[idx]);
    // if(idx == 2077){
        // printf("idx[%d],ix[%d],iy[%d],bdx[%d],bdy[%d],bix[%d],biy[%d],gdx[%d],gdy[%d],C[%f]\r\n", \
        // idx,ix,iy,blockDim.x,blockDim.y,blockIdx.x,blockIdx.y,gridDim.x,gridDim.y,C[idx]);
    // }
}


int main()
{

for(int cycle=0;cycle < 32;cycle++)
{
    int gridSize2Dx = 16*cycle;
    int gridSize2Dy = 16*cycle;
    //blocksize MAX 1024
    int blockSize2Dx = 32;
    int blockSize2Dy = 32;
    int sum = gridSize2Dx*gridSize2Dy*blockSize2Dx*blockSize2Dy;
    int sum_bytes = sum*sizeof(float);
    printf("size %d \r\n", sum);

    float* A_host=(float*)malloc(sum_bytes);
    for(int i=0;i<sum;i++){
        A_host[i]=(float)i;
    }
  
    float* B_host=(float*)malloc(sum_bytes);
    for(int i=0;i<sum;i++){
        B_host[i]=i;
    }

    float* C_host=(float*)malloc(sum_bytes);

    float *A_dev=NULL;
    float *B_dev=NULL;
    float *C_dev=NULL;
    {
        timecost t1("cuda");
        cudaMalloc((void**)&A_dev,sum_bytes);
        cudaMemcpy(A_dev,A_host,sum_bytes,cudaMemcpyHostToDevice);
        cudaMalloc((void**)&B_dev,sum_bytes);
        cudaMemcpy(B_dev,B_host,sum_bytes,cudaMemcpyHostToDevice);

        cudaMalloc((void**)&C_dev,sum_bytes);
        dim3 gridSize2D(gridSize2Dx, gridSize2Dy);
        dim3 blockSize2D(blockSize2Dx, blockSize2Dy);
        kernelAdd<<<gridSize2D, blockSize2D>>>(A_dev,B_dev,C_dev);
        int ret = 0;
        ret = cudaMemcpy(C_host,C_dev,sum_bytes,cudaMemcpyDeviceToHost);
    }

    for(int i=0;i<sum;i++){
        // printf("C[%d]:%f \n", i, C_host[i]);
        C_host[i] = 0.0f;
    }

    {
        timecost t2("cpu");
        for(int i=0;i<sum;i++){
            C_host[i]=cos(A_host[i])+sin(B_host[i]);
        }
    }

    for(int i=0;i<sum;i++){
        // printf("C[%d]:%f \n", i, C_host[i]);
        C_host[i] = 0.0f;
    }

    cudaFree(A_dev);
    free(A_host);
    cudaFree(B_dev);
    free(B_host);
    cudaFree(C_dev);
    free(C_host);
}
    return 0;
}

Use Python script to count the time consumption (orange CPU time consumption, blue CUDA time consumption, x-axis is the increasing amount of calculation)

in conclusion

After the above tests, a simple comparison between CUDA and CPU shows that the larger the amount of data, the more advantages CUDA has. When the amount of data is very small, it is worse to use CPU. Of course, the writing of CUDA programs is also a major influencing factor, such as memory allocation, decoupling of program logic, etc.

video

wangerxian

CUDA uses graphics card calculations. When the amount of data is large, the advantages of graphics cards are reflected.

wangerxian

However, there may be some problems with using cuda, such as the graphics card not supporting it. I used cuda before, but after updating the graphics card driver, my computer frequently had blue screens.

LitchiCheng

wangerxian posted on 2024-2-29 10:51 However, there may be some problems with using cuda, such as the graphics card not supporting it. I used cuda before, but after updating the graphics card driver, my computer frequently had blue screens

Mine is a Jetson board, which is relatively better, not the graphics card of my computer.

wangerxian

LitchiCheng posted on 2024-2-29 22:52 This is a jetson board, which is relatively better, not the graphics card of my own computer

Did you buy the Jetson board to learn artificial intelligence?

LitchiCheng

wangerxian posted on 2024-3-1 15:41 Did you buy the Jetson board to learn artificial intelligence?

For pre-research testing

nmg

I'm curious, are all engineers' computers equipped with NVIDIA graphics cards?

xao523523

111