#include "mat_mul.h"

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

#define NUM_GPUS        4
#define TILE_SIZE       32
#define WORK_PER_THREAD 8
#define VECTOR_WIDTH    8
#define REG_TILE_SIZE   (TILE_SIZE / WORK_PER_THREAD)
#define min(A, B)       (((A)>(B))?(B):(A))

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

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

static float *A, *B, *C;
static int M, N, K;
static int row_start[NUM_GPUS];
static int row_end[NUM_GPUS];
static int num_rows[NUM_GPUS];

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
  #pragma omp parallel for
  for(int i = 0; i < NUM_GPUS; i++)
  {
    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), &num_rows[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
    // For kernel.cl
    size_t gws[2] = { (size_t)((num_rows[i] + WORK_PER_THREAD - 1) / WORK_PER_THREAD), (size_t)N };
    size_t lws[2] = { (size_t)(TILE_SIZE / WORK_PER_THREAD), (size_t)(TILE_SIZE) };
    // For kernel2.cl
    // size_t gws[2] = { (size_t)(num_rows[i]), (size_t)(N / VECTOR_WIDTH) };
    // size_t lws[2] = { (size_t)(TILE_SIZE), (size_t)(TILE_SIZE / VECTOR_WIDTH) };
    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];
    }

    // printf("#%d gws = (%d, %d), lws = (%d, %d)\n", i, (int)gws[0], (int)gws[1], (int)lws[0], (int)lws[1]);

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

  // DO NOT REMOVE; NEEDED FOR TIME MEASURE
  #pragma omp parallel for
  for(int i = 0; i < NUM_GPUS; i++)
  {
    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, NUM_GPUS, device, "", NULL, NULL);
  CHECK_ERROR(err);

  if (err == CL_BUILD_PROGRAM_FAILURE)
  {
    size_t log_size;
    for(int i = 0; i < NUM_GPUS; i++)
    {
      CHECK_ERROR(clGetProgramBuildInfo(program, device[i], CL_PROGRAM_BUILD_LOG, 0, NULL, &log_size));
      char *log = (char *)malloc(log_size + 1);
      CHECK_ERROR(clGetProgramBuildInfo(program, device[i], CL_PROGRAM_BUILD_LOG, log_size, log, NULL));
      log[log_size] = 0;
      printf("Compile error:\n%s\n", log);
      free(log);
    }
  }

  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, NUM_GPUS, device, &device_cnt);
  CHECK_ERROR(err);
  // for(int i = 0; i < NUM_GPUS; i++)
  // {
  //   print_device_info(device[i]);  
  // }

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

  // Compile program from "kernel.cl"
  program = create_and_build_program_with_source(context, device, "kernel.cl");
  // program = create_and_build_program_with_source(context, device, "kernel2.cl");

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

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

    // Calculate Distribution Unit
    row_start[i] = M / NUM_GPUS * i + min(i, M % NUM_GPUS);
    row_end[i] = M / NUM_GPUS * (i + 1) + min(i + 1, M % NUM_GPUS);
    num_rows[i] = row_end[i] - row_start[i];

    // Create GPU buffers
    a_d[i] = clCreateBuffer(context, CL_MEM_READ_WRITE, num_rows[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, num_rows[i] * N * sizeof(float), NULL, &err);
    CHECK_ERROR(err);

    err = clEnqueueWriteBuffer(queue[i], a_d[i], CL_FALSE, 0, num_rows[i] * K * sizeof(float), &A[row_start[i] * K], 0, NULL, NULL);
    CHECK_ERROR(err);
    err = clEnqueueWriteBuffer(queue[i], b_d[i], CL_FALSE, 0, K * N * sizeof(float), B, 0, NULL, NULL);
    CHECK_ERROR(err);
  }

  // DO NOT REMOVE; NEEDED FOR TIME MEASURE
  #pragma omp parallel for
  for(int i = 0; i < NUM_GPUS; i++)
  {
    err = clFinish(queue[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)
  #pragma omp parallel for
  for(int i = 0; i < NUM_GPUS; i++)
  {
    err = clEnqueueReadBuffer(queue[i], c_d[i], CL_FALSE, 0, num_rows[i] * N * sizeof(float), &C[row_start[i] * N], 0, NULL, NULL);
    CHECK_ERROR(err);
  }

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