performance-tests/opencl_add_gpu.c

#include <stdio.h>
#include <stdlib.h>
#include <CL/cl.h>
#include <sys/time.h>

#define ARRAY_SIZE 15000576   // Tiny array
#define LOCAL_SIZE 256    // Small work-group size

const char *kernelSource =
"__kernel void add(__global const float *a, __global const float *b, __global float *c) { "
"    int i = get_global_id(0);                                             "
" if (i < 15000576) {														"
"    c[i] = a[i] + b[i];                                                   "
"}                                                                        "
"}                                                                        ";

double get_time() {
    struct timeval tv;
    gettimeofday(&tv, NULL);
    return tv.tv_sec + tv.tv_usec * 1e-6;
}

int main() {
    float *a = (float*)malloc(sizeof(float) * ARRAY_SIZE);
    float *b = (float*)malloc(sizeof(float) * ARRAY_SIZE);
    float *c = (float*)malloc(sizeof(float) * ARRAY_SIZE);

    // Initialize arrays
    for (int i = 0; i < ARRAY_SIZE; i++) {
        a[i] = i;
        b[i] = i * 2;
    }

    // Get GPU platform and device
    cl_platform_id platform;
    cl_device_id gpuDevice;
    cl_uint platformCount, deviceCount;
    clGetPlatformIDs(1, &platform, &platformCount);
    clGetDeviceIDs(platform, CL_DEVICE_TYPE_GPU, 1, &gpuDevice, &deviceCount);

    // Print GPU device info
    char deviceName[128];
    clGetDeviceInfo(gpuDevice, CL_DEVICE_NAME, 128, deviceName, NULL);
    printf("Using GPU device: %s\n", deviceName);

    // Create context and command queue
    cl_context context = clCreateContext(NULL, 1, &gpuDevice, NULL, NULL, NULL);
    cl_command_queue command_queue = clCreateCommandQueue(context, gpuDevice, 0, NULL);

    // Create buffers
    double start = get_time();
    cl_mem a_mem = clCreateBuffer(context, CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR, ARRAY_SIZE * sizeof(float), (void *)a, NULL);
    cl_mem b_mem = clCreateBuffer(context, CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR, ARRAY_SIZE * sizeof(float), (void *)b, NULL);
    cl_mem c_mem = clCreateBuffer(context, CL_MEM_WRITE_ONLY, ARRAY_SIZE * sizeof(float), NULL, NULL);
    double buffer_create_time = get_time() - start;

    // Time data transfer to device (not needed with CL_MEM_COPY_HOST_PTR)
    double transfer_to_device_time = 0.0;

    // Create and build program
    start = get_time();
    cl_program program = clCreateProgramWithSource(context, 1, &kernelSource, NULL, NULL);
    cl_int ret = clBuildProgram(program, 1, &gpuDevice, NULL, NULL, NULL);
    if (ret != CL_SUCCESS) {
        size_t log_size;
        clGetProgramBuildInfo(program, gpuDevice, CL_PROGRAM_BUILD_LOG, 0, NULL, &log_size);
        char *log = (char *)malloc(log_size);
        clGetProgramBuildInfo(program, gpuDevice, CL_PROGRAM_BUILD_LOG, log_size, log, NULL);
        printf("Kernel compilation error:\n%s\n", log);
        free(log);
        return 1;
    }
    double program_build_time = get_time() - start;

    // Create kernel
    start = get_time();
    cl_kernel kernel = clCreateKernel(program, "add", NULL);
    if (!kernel) {
        printf("Error creating kernel\n");
        return 1;
    }
    double kernel_create_time = get_time() - start;

    // Set kernel arguments
    start = get_time();
    ret = clSetKernelArg(kernel, 0, sizeof(cl_mem), (void *)&a_mem);
    ret |= clSetKernelArg(kernel, 1, sizeof(cl_mem), (void *)&b_mem);
    ret |= clSetKernelArg(kernel, 2, sizeof(cl_mem), (void *)&c_mem);
    if (ret != CL_SUCCESS) {
        printf("Error setting kernel arguments: %d\n", ret);
        return 1;
    }
    double kernel_arg_time = get_time() - start;

    // Execute kernel
    start = get_time();
    size_t global_size = ARRAY_SIZE;
    size_t local_size = LOCAL_SIZE;
    ret = clEnqueueNDRangeKernel(command_queue, kernel, 1, NULL, &global_size, &local_size, 0, NULL, NULL);
    if (ret != CL_SUCCESS) {
        printf("Error executing kernel: %d\n", ret);
        return 1;
    }
    clFinish(command_queue);
    double kernel_exec_time = get_time() - start;

    // Read result back
    start = get_time();
    ret = clEnqueueReadBuffer(command_queue, c_mem, CL_TRUE, 0, ARRAY_SIZE * sizeof(float), c, 0, NULL, NULL);
    if (ret != CL_SUCCESS) {
        printf("Error reading from memory buffer: %d\n", ret);
        return 1;
    }
    double transfer_to_host_time = get_time() - start;

    // Print timing results
    printf("Buffer creation time:      %.6f ms\n", buffer_create_time * 1000);
    printf("Program build time:        %.6f ms\n", program_build_time * 1000);
    printf("Kernel creation time:      %.6f ms\n", kernel_create_time * 1000);
    printf("Kernel argument setup time: %.6f ms\n", kernel_arg_time * 1000);
    printf("Kernel execution time:     %.6f ms\n", kernel_exec_time * 1000);
    printf("Data transfer to host:     %.6f ms\n", transfer_to_host_time * 1000);
    printf("Total time:                %.6f ms\n",
           (buffer_create_time + program_build_time + kernel_create_time +
            kernel_arg_time + kernel_exec_time + transfer_to_host_time) * 1000);

    // Print the result
    for (int i = 0; i < 8; i++) {
        printf("c[%d] = %f\n", i, c[i]);
    }

    // Cleanup
    clReleaseMemObject(a_mem);
    clReleaseMemObject(b_mem);
    clReleaseMemObject(c_mem);
    clReleaseKernel(kernel);
    clReleaseProgram(program);
    clReleaseCommandQueue(command_queue);
    clReleaseContext(context);

    return 0;
}