gpgpu-sim dose not support CUDA ray tracing programs?

[WilliamWangPeng]

HI dear author,
It's truly a honor to write a letter to you, I'm building your project nowadays and found the error when running a very simple CUDA ray tracing program as following, I wonder if there is a chance that you know the reason :)

Why this programs does not work?

#include <iostream>
#include <time.h>
#include "cuda_profiler_api.h"
// limited version of checkCudaErrors from helper_cuda.h in CUDA examples
#define checkCudaErrors(val) check_cuda( (val), #val, __FILE__, __LINE__ )

void check_cuda(cudaError_t result, char const *const func, const char *const file, int const line) {
    if (result) {
        std::cerr << "CUDA error = " << static_cast<unsigned int>(result) << " at " <<
            file << ":" << line << " '" << func << "' \n";
        // Make sure we call CUDA Device Reset before exiting
        cudaDeviceReset();
        exit(99);
    }
}

__global__ void render(float *fb, int max_x, int max_y) {
    int i = threadIdx.x + blockIdx.x * blockDim.x;
    int j = threadIdx.y + blockIdx.y * blockDim.y;
    if((i >= max_x) || (j >= max_y)) return;
    int pixel_index = j*max_x*3 + i*3;
    fb[pixel_index + 0] = float(i) / max_x;
    fb[pixel_index + 1] = float(j) / max_y;
    fb[pixel_index + 2] = 0.2;
}

int main() {
    int nx = 1200;
    int ny = 600;
    int tx = 8;
    int ty = 8;

    std::cerr << "Rendering a " << nx << "x" << ny << " image ";
    std::cerr << "in " << tx << "x" << ty << " blocks.\n";

    int num_pixels = nx*ny;
    size_t fb_size = 3*num_pixels*sizeof(float);

    float *fb;
    checkCudaErrors(cudaMallocManaged((void **)&fb, fb_size));

    clock_t start, stop;
    start = clock();

    dim3 blocks(nx/tx+1,ny/ty+1);

  dim3 threads(tx,ty);
  cudaProfilerStart();
  render<<<blocks, threads>>>(fb, nx, ny);
  cudaProfilerStop();
  checkCudaErrors(cudaGetLastError());
    checkCudaErrors(cudaDeviceSynchronize());
    stop = clock();
    double timer_seconds = ((double)(stop - start)) / CLOCKS_PER_SEC;
    std::cerr << "took " << timer_seconds << " seconds.\n";

    std::cout << "P3\n" << nx << " " << ny << "\n255\n";
    for (int j = ny-1; j >= 0; j--) {
        for (int i = 0; i < nx; i++) {
            size_t pixel_index = j*3*nx + i*3;
            float r = fb[pixel_index + 0];
            float g = fb[pixel_index + 1];
            float b = fb[pixel_index + 2];
            int ir = int(255.99*r);
            int ig = int(255.99*g);
            int ib = int(255.99*b);
            std::cout << ir << " " << ig << " " << ib << "\n";
        }
    }

    checkCudaErrors(cudaFree(fb));

}

thank you
best regards to you
William

[mattsinc]

Hi William,

This repo is not the right place to ask your question -- this repo is about adding HTCondor support for running GPGPU-Sim.

Having said that, if you look at the GPGPU-Sim code repo: https://github.com/gpgpu-sim/gpgpu-sim_distribution/blob/master/libcuda/cuda_runtime_api.cc#L4742, the function you are asking about is not supported.

Thanks, Matt

[WilliamWangPeng]

Hi William,

This repo is not the right place to ask your question -- this repo is about adding HTCondor support for running GPGPU-Sim.

Having said that, if you look at the GPGPU-Sim code repo: https://github.com/gpgpu-sim/gpgpu-sim_distribution/blob/master/libcuda/cuda_runtime_api.cc#L4742, the function you are asking about is not supported.

Thanks, Matt

thank you so much @mattsinc
Do you know if that is easy to support this API ?

thank you best regards with kindness
William

[mattsinc]

In theory it would require adding a counter or timer of some sort to go from the time you call profiler start to the time you call profiler stop. Not terribly hard, but also something that would require some trial and error to do. Probably better to ask the maintainers either way.

Matt