#include "mat_mul.h"

#include <stdio.h>
#include <CL/cl.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 MAX_NUM_DEVICES 4

static cl_int err;
static cl_platform_id platform;
//static cl_device_id device;
static cl_device_id devices[MAX_NUM_DEVICES];
static cl_context context;
//static cl_command_queue queue;
static cl_command_queue queues[MAX_NUM_DEVICES];
static cl_program program;
//static cl_kernel kernel;
static cl_kernel kernels[MAX_NUM_DEVICES];
static cl_mem a_d[MAX_NUM_DEVICES], b_d[MAX_NUM_DEVICES], c_d[MAX_NUM_DEVICES];

static float *A, *B, *C;
static int M, N, K;

static int seperate_M[MAX_NUM_DEVICES];

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;

  // Setup kernel arguments
  for (int i=0; i<MAX_NUM_DEVICES; ++i){
    err = clSetKernelArg(kernels[i], 0, sizeof(cl_mem), &a_d[i]);
    CHECK_ERROR(err);
    err = clSetKernelArg(kernels[i], 1, sizeof(cl_mem), &b_d[i]);
    CHECK_ERROR(err);
    err = clSetKernelArg(kernels[i], 2, sizeof(cl_mem), &c_d[i]);
    CHECK_ERROR(err);
    err = clSetKernelArg(kernels[i], 3, sizeof(int), &seperate_M[i]);
    CHECK_ERROR(err);
    err = clSetKernelArg(kernels[i], 4, sizeof(int), &N);
    CHECK_ERROR(err);
    err = clSetKernelArg(kernels[i], 5, sizeof(int), &K);
    CHECK_ERROR(err);
  }

  for (int i=0; i<MAX_NUM_DEVICES; ++i){
    // Setup global work size and local work size
    //size_t gws[2] = {(size_t)M, (size_t)N}, lws[2] = {1, 1};
    size_t gws[2] = {(size_t)(seperate_M[i] + WPT - 1)/WPT, (size_t)N}, lws[2] = {TS/WPT, TS};
    for (int j = 0; j < 2; ++j) {
      // 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[j] = (gws[j] + lws[j] - 1) / lws[j] * lws[j];
    }

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

  // DO NOT REMOVE; NEEDED FOR TIME MEASURE
  for (int i=0; i<MAX_NUM_DEVICES; ++i){
    err = clFinish(queues[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, MAX_NUM_DEVICES, device, "", NULL, NULL);
  if (err == CL_BUILD_PROGRAM_FAILURE) {
    size_t log_size;
    CHECK_ERROR(clGetProgramBuildInfo(program, device[0], CL_PROGRAM_BUILD_LOG, 0, NULL, &log_size));
    char *log = (char*)malloc(log_size + 1);
    CHECK_ERROR(clGetProgramBuildInfo(program, device[0], 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, MAX_NUM_DEVICES, devices, NULL);
  CHECK_ERROR(err);
  for(int i=0; i<MAX_NUM_DEVICES; ++i)
    print_device_info(devices[i]);
  context = clCreateContext(NULL, MAX_NUM_DEVICES, devices, NULL, NULL, &err);
  CHECK_ERROR(err);

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

  // Compile program from "kernel.cl"
  program = create_and_build_program_with_source(context, devices, "kernel.cl");

  // Extract kernel from compiled program
  for (int i=0; i<MAX_NUM_DEVICES; ++i){
    kernels[i] = clCreateKernel(program, "sgemm", &err); 
    CHECK_ERROR(err);
  }

  // Create GPU buffers
  int tmp_seperate = M / MAX_NUM_DEVICES;
  for (int i=0; i<MAX_NUM_DEVICES; ++i){
    seperate_M[i] = tmp_seperate;
    if (i == MAX_NUM_DEVICES-1) seperate_M[i] += M % MAX_NUM_DEVICES;
  }

  for (int i=0; i<MAX_NUM_DEVICES; ++i){
    a_d[i] = clCreateBuffer(context, CL_MEM_READ_WRITE, seperate_M[i] * 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, seperate_M[i] * N * sizeof(float), NULL, &err);
    CHECK_ERROR(err);
  }

  // Write to GPU; A (cpu) -> a_d (gpu), B (cpu) -> b_d (gpu)
  int offset = 0;
  for (int i=0; i<MAX_NUM_DEVICES; ++i){
    err = clEnqueueWriteBuffer(queues[i], a_d[i], CL_TRUE, 0, seperate_M[i] * K * sizeof(float), &A[offset], 0, NULL, NULL);
    CHECK_ERROR(err);
    offset += seperate_M[i] * K;
    err = clEnqueueWriteBuffer(queues[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
  for (int i=0; i<MAX_NUM_DEVICES; ++i){
    err = clFinish(queues[i]);
    CHECK_ERROR(err);
  }
}

void mat_mul_final(float *A, float *B, float *C, int M, int N, int K) {
  // Read from GPU; c_d (gpu) -> C (cpu)
  int offset = 0;
  for (int i=0; i<MAX_NUM_DEVICES; ++i) {
    err = clEnqueueReadBuffer(queues[i], c_d[i], CL_TRUE, 0, seperate_M[i] * N * sizeof(float), &C[offset], 0, NULL, NULL);
    CHECK_ERROR(err);
    offset += seperate_M[i] * N;
  }

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