#include "mat_mul.h"
#include <stdio.h>
#include <CL/cl.h>
#include <fstream>
#include <string.h>

#define CHECK_ERROR(err) \
    if (err != CL_SUCCESS) { \
        printf("[%s:%d] OpenCL error %d\n", __FILE__, __LINE__, err); \
        exit(EXIT_FAILURE); \
    }
#define TS 32
#define WPT 8
#define DEV_NUM 4

static cl_int err;
static cl_platform_id platform;
static cl_device_id device[DEV_NUM];
static cl_context context;
static cl_command_queue queue[DEV_NUM];
static cl_program program;
static cl_kernel kernel[DEV_NUM];
static cl_mem a_d[DEV_NUM], b_d[DEV_NUM], c_d[DEV_NUM];

static float *A, *B, *C;
static int M, N, K;
static int size[DEV_NUM];
static int ndev;

static std::ofstream fout;

void mat_mul(float *_A, float *_B, float *_C, int _M, int _N, int _K) {
    A = _A, B = _B, C = _C;
    M = _M, N = _N, K = _K;

#ifdef _DEBUG
    fout.open("dump.out");
#endif

    for (int i=0; i<ndev; i++) {
        int rows = M / ndev;
        int start = i * rows;
        int end = start + rows;

        if (i == ndev - 1) {
            end = M;
        }
        size[i] = end - start;

        // Setup kernel arguments
        err = clSetKernelArg(kernel[i], 0, sizeof(cl_mem), &a_d[i]);
        CHECK_ERROR(err);
        err = clSetKernelArg(kernel[i], 1, sizeof(cl_mem), &b_d[i]);
        CHECK_ERROR(err);
        err = clSetKernelArg(kernel[i], 2, sizeof(cl_mem), &c_d[i]);
        CHECK_ERROR(err);
        err = clSetKernelArg(kernel[i], 3, sizeof(int), &size[i]);
        CHECK_ERROR(err);
        err = clSetKernelArg(kernel[i], 4, sizeof(int), &N);
        CHECK_ERROR(err);
        err = clSetKernelArg(kernel[i], 5, sizeof(int), &K);
        CHECK_ERROR(err);

        // Setup global work size and local work size
        size_t gws[2] = {(size_t)(size[i] + WPT - 1)/WPT, (size_t)N}, lws[2] = {TS/WPT, TS};
        for (int i = 0; i < 2; ++i) {
            // By OpenCL spec, global work size should be MULTIPLE of local work size
            // Formula below achieve it
            // e.g., gws = 25, lws = 16, then (25 + 16 - 1) / 16 * 16 = 40 / 16 * 16 = 2 * 16 = 32
            gws[i] = (gws[i] + lws[i] - 1) / lws[i] * lws[i];
        }
#ifdef _DEBUG
        fout << "[GPU" << i << "] size=" << size[i] << " gws={" << gws[0] << "," << gws[1] << "}, lws={" << lws[0] << "," << lws[1] << "}" << std::endl;
#endif

        // Run kernel
        err = clEnqueueNDRangeKernel(queue[i], kernel[i], 2, NULL, gws, lws, 0, NULL, NULL);
        CHECK_ERROR(err);
    }

    for (int i=0; i<ndev; i++) {
        // DO NOT REMOVE; NEEDED FOR TIME MEASURE
        err = clFinish(queue[i]);
        CHECK_ERROR(err);
    }
}

static void print_platform_info(cl_platform_id platform) {
    size_t sz;
    char *buf;
    CHECK_ERROR(clGetPlatformInfo(platform, CL_PLATFORM_NAME, 0, NULL, &sz));
    buf = (char*)malloc(sz);
    CHECK_ERROR(clGetPlatformInfo(platform, CL_PLATFORM_NAME, sz, buf, NULL));
    printf("Detected OpenCL platform: %s\n", buf);
    free(buf);
}

static void print_device_info(cl_device_id device) {
    size_t sz;
    char *buf;
    CHECK_ERROR(clGetDeviceInfo(device, CL_DEVICE_NAME, 0, NULL, &sz));
    buf = (char*)malloc(sz);
    CHECK_ERROR(clGetDeviceInfo(device, CL_DEVICE_NAME, sz, buf, NULL));
    printf("Detected OpenCL device: %s\n", buf);
    free(buf);
}

static cl_program create_and_build_program_with_source(cl_context context, cl_device_id device, const char *file_name) {
    FILE *file = fopen(file_name, "rb");
    if (file == NULL) {
        printf("Failed to open %s\n", file_name);
        exit(EXIT_FAILURE);
    }
    fseek(file, 0, SEEK_END);
    size_t source_size = ftell(file);
    rewind(file);
    char *source_code = (char*)malloc(source_size + 1);
    size_t ntotal = 0;
    while (ntotal < source_size) {
        int nread = fread(source_code, sizeof(char), source_size, file);
        ntotal += nread;
    }
    source_code[source_size] = '\0';
    fclose(file);
    cl_program program = clCreateProgramWithSource(context, 1, (const char **)&source_code, &source_size, &err);
    CHECK_ERROR(err);
    free(source_code);
    err = clBuildProgram(program, 1, &device, "", NULL, NULL);
    if (err == CL_BUILD_PROGRAM_FAILURE) {
        size_t log_size;
        CHECK_ERROR(clGetProgramBuildInfo(program, device, CL_PROGRAM_BUILD_LOG, 0, NULL, &log_size));
        char *log = (char*)malloc(log_size + 1);
        CHECK_ERROR(clGetProgramBuildInfo(program, device, CL_PROGRAM_BUILD_LOG, log_size, log, NULL));
        log[log_size] = 0;
        printf("Compile error:\n%s\n", log);
        free(log);
    }
    CHECK_ERROR(err);
    return program;
}

void mat_mul_init(float *A, float *B, float *C, int M, int N, int K) {
    // Get OpenCL platform
    err = clGetPlatformIDs(1, &platform, NULL);
    CHECK_ERROR(err);
    print_platform_info(platform);

    // Get OpenCL device
    err = clGetDeviceIDs(platform, CL_DEVICE_TYPE_GPU, DEV_NUM, device, (unsigned int*) &ndev);
    CHECK_ERROR(err);
    for (int i=0; i<DEV_NUM; i++)
        print_device_info(device[i]);

    // Create OpenCL context
    context = clCreateContext(NULL, ndev, device, NULL, NULL, &err);
    CHECK_ERROR(err);

    for (int i=0; i<ndev; i++) {
        // Create OpenCL command queue
        queue[i] = clCreateCommandQueue(context, device[i], 0, &err);
        CHECK_ERROR(err);

        // Compile program from "kernel.cl"
        program = create_and_build_program_with_source(context, device[i], "kernel.cl");

        // Extract kernel from compiled program
        kernel[i] = clCreateKernel(program, "sgemm", &err); 
        CHECK_ERROR(err);
    }

    for (int i=0; i<ndev; i++) {
        // Create GPU buffers, namely GPU memory like malloc
        int rows = M / ndev;
        int start = i * rows;
        int end = start + rows;

        if (i == ndev - 1) {
            end = M;
        }

        rows = end - start;
        a_d[i] = clCreateBuffer(context, CL_MEM_READ_WRITE, rows * K * sizeof(float), NULL, &err);
        CHECK_ERROR(err);
        b_d[i] = clCreateBuffer(context, CL_MEM_READ_WRITE, K * N * sizeof(float), NULL, &err);
        CHECK_ERROR(err);
        c_d[i] = clCreateBuffer(context, CL_MEM_READ_WRITE, rows * N * sizeof(float), NULL, &err);
        CHECK_ERROR(err);

        // Write to GPU; A (cpu) -> a_d (gpu), B (cpu) -> b_d (gpu)
        err = clEnqueueWriteBuffer(queue[i], a_d[i], CL_TRUE, 0, rows * K * sizeof(float), &A[start * K], 0, NULL, NULL);
        CHECK_ERROR(err);
        err = clEnqueueWriteBuffer(queue[i], b_d[i], CL_TRUE, 0, K * N * sizeof(float), B, 0, NULL, NULL);
        CHECK_ERROR(err);

        // DO NOT REMOVE; NEEDED FOR TIME MEASURE
        err = clFinish(queue[i]);
        CHECK_ERROR(err);
    }
}

void mat_mul_final(float *A, float *B, float *C, int M, int N, int K) {
//    memset(C, -1, M * N * sizeof(float));
    for (int i=0; i<ndev; i++) {
        int rows = M / ndev;
        int start = i * rows;
        int end = start + rows;

        if (i == ndev - 1) {
            end = M;
        }
        rows = end - start;
        // Read from GPU; c_d (gpu) -> C (cpu)
        err = clEnqueueReadBuffer(queue[i], c_d[i], CL_TRUE, 0, rows * N * sizeof(float), &C[start * N], 0, NULL, NULL);
        CHECK_ERROR(err);
    }

    for (int i=0; i<ndev; i++) {
        // DO NOT REMOVE; NEEDED FOR TIME MEASURE
        err = clFinish(queue[i]);
        CHECK_ERROR(err);
    }
}