#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 64
#define WPT 32

static cl_int err;
static cl_platform_id platform;
static cl_device_id device[4];
static cl_context context;
static cl_command_queue queue[4];
static cl_program program[4];
static cl_kernel kernel[4];
static cl_mem a_d[4], b_d[4], c_d[4];

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

static int num_device;
//	static int M_splitted[4];

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 n = 0 ; n < num_device ; n++) {
	  int M_splitted = (n == num_device-1) ? (M - (M/num_device * n)) : M / num_device;
	  err = clSetKernelArg(kernel[n], 0, sizeof(cl_mem), &a_d[n]);
	  CHECK_ERROR(err);
	  err = clSetKernelArg(kernel[n], 1, sizeof(cl_mem), &b_d[n]);
	  CHECK_ERROR(err);
	  err = clSetKernelArg(kernel[n], 2, sizeof(cl_mem), &c_d[n]);
	  CHECK_ERROR(err);
	  err = clSetKernelArg(kernel[n], 3, sizeof(int), &M_splitted);
	  CHECK_ERROR(err);
	  err = clSetKernelArg(kernel[n], 4, sizeof(int), &N);
	  CHECK_ERROR(err);
	  err = clSetKernelArg(kernel[n], 5, sizeof(int), &K);
	  CHECK_ERROR(err);

	  // Setup global work size and local work size
	  size_t gws[2] = {(size_t)(M_splitted + WPT - 1)/WPT, (size_t)N};
	  size_t 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];
	  }
	  // Run kernel
	  err = clEnqueueNDRangeKernel(queue[n], kernel[n], 2, NULL, gws, lws, 0, NULL, NULL);
	  CHECK_ERROR(err);
  }

  // DO NOT REMOVE; NEEDED FOR TIME MEASURE
  //
  for (int n = 0 ; n < num_device ; n++) {
	  err = clFinish(queue[n]);
	  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);

  int num_device_temp;
  // Get OpenCL device
  err = clGetDeviceIDs(platform, CL_DEVICE_TYPE_GPU, 0, NULL, (cl_uint*)&num_device_temp);

  if (M >= num_device_temp) num_device = num_device_temp;
  else num_device = M;

  err = clGetDeviceIDs(platform, CL_DEVICE_TYPE_GPU, num_device, device, NULL);
  CHECK_ERROR(err);
  for (int n = 0 ; n < num_device ; n++) {
//	  M_splitted[n] = (n == num_device-1) ? (M - (n * M/num_device)) : M/num_device;
	  print_device_info(device[n]);
  }

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

  for (int n = 0 ; n < num_device ; n++) {
	  int M_splitted = (n == num_device-1) ? (M - (M/num_device * n)) : M / num_device;
	  // Create OpenCL command queue
	  queue[n] = clCreateCommandQueue(context, device[n], 0, &err);
	  CHECK_ERROR(err);

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

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

	 // Create GPU buffers
	 a_d[n] = clCreateBuffer(context, CL_MEM_READ_WRITE, M_splitted * K * sizeof(float), NULL, &err);
	 CHECK_ERROR(err);
	 b_d[n] = clCreateBuffer(context, CL_MEM_READ_WRITE, K * N * sizeof(float), NULL, &err);
	 CHECK_ERROR(err);
	 c_d[n] = clCreateBuffer(context, CL_MEM_READ_WRITE, M_splitted * N * sizeof(float), NULL, &err);
	 CHECK_ERROR(err);

  // Write to GPU; A (cpu) -> a_d (gpu), B (cpu) -> b_d (gpu)
	  err = clEnqueueWriteBuffer(queue[n], a_d[n], CL_TRUE, 0, M_splitted * K * sizeof(float), &A[M/num_device * K * n], 0, NULL, NULL);
	  CHECK_ERROR(err);
	  err = clEnqueueWriteBuffer(queue[n], b_d[n], CL_TRUE, 0, K * N * sizeof(float), B, 0, NULL, NULL);
	  CHECK_ERROR(err);
  }
  // DO NOT REMOVE; NEEDED FOR TIME MEASURE
  for (int n = 0 ; n < num_device ; n++) {
	  err = clFinish(queue[n]);
	  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)
  for (int n = 0 ; n < num_device ; n++) {
	  int M_splitted = (n == num_device-1) ? (M - (M/num_device * n)) : M / num_device;
	  err = clEnqueueReadBuffer(queue[n], c_d[n], CL_TRUE, 0, M_splitted * N * sizeof(float), &C[M/num_device * N * n], 0, NULL, NULL);
	  CHECK_ERROR(err);
  }

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