#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); \
  }

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

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

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
  //err = clSetKernelArg(kernel, 0, sizeof(cl_mem), &a_d);
  //CHECK_ERROR(err);
  //err = clSetKernelArg(kernel, 1, sizeof(cl_mem), &b_d);
  //CHECK_ERROR(err);
  //err = clSetKernelArg(kernel, 2, sizeof(cl_mem), &c_d);
  //CHECK_ERROR(err);
  //err = clSetKernelArg(kernel, 3, sizeof(int), &M);
  //CHECK_ERROR(err);
  //err = clSetKernelArg(kernel, 4, sizeof(int), &N);
  //CHECK_ERROR(err);
  //err = clSetKernelArg(kernel, 5, sizeof(int), &K);
  //CHECK_ERROR(err);

  // Setup global work size and local work size
  // 1st method
  //size_t gws[2] = {(size_t)M, (size_t)N}, lws[2] = {32,32}; // 72  GFlops
  //size_t gws[2] = {(size_t)M, (size_t)N}, lws[2] = {16,16}; // 156
  //size_t gws[2] = {(size_t)M, (size_t)N}, lws[2] = {8,8};   // 457
  //size_t gws[2] = {(size_t)M, (size_t)N}, lws[2] = {4,4};   // 121
  //
  // 2nd method
  //size_t gws[2] = {(size_t)M, (size_t)N}, lws[2] = {32,32}; // 680
  //
  // 3rd method
//#define TS 32
//#define WPT 8
  //size_t gws[2] = {(size_t)M, (size_t)N/WPT}, lws[2] = {TS,TS/WPT}; // 1433
  //
  // 4th method
//#define TS 32
//#define WIDTH 4
  //size_t gws[2] = {(size_t)M, (size_t)N/WIDTH}, lws[2] = {TS,TS/WIDTH}; // 2675
  //
  // 5th method
//#define TS 32
//#define WIDTH 8
  //size_t gws[2] = {(size_t)M, (size_t)N/WIDTH}, lws[2] = {TS,TS/WIDTH}; // 3638
  //
  // 6th method - Tile-unaligned
//#define TS 32
//#define WIDTH 8
  //size_t gws[2] = {(size_t)M, (size_t)(N+WIDTH-1)/WIDTH}, lws[2] = {TS,TS/WIDTH}; // ...
  //
  // 7th method - 3rd + unaligned
//#define TS 32
//#define WPT 8
//  printf("\n");
//  size_t gws[2] = {(size_t)(M+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];
  //  printf("gws%d:%d, lws%d:%d\n", i, gws[i], i, lws[i]);
  //}

  for(int d=0; d<DEVICE_NUM; d++) {
    //printf("GPU%d M_start:%d\n", d, M_start[d]);
    //printf("GPU%d M_size :%d\n", d, M_size[d]);

    #define TS 32
    #define WPT 8
    //printf("\n");
    size_t gws[2] = {(size_t)(M_size[d]+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];
      //printf("gws%d:%ld, lws%d:%ld\n", i, gws[i], i, lws[i]);
    }

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

    // Run kernel
    //printf("RUN Q%d\n", d);
    err = clEnqueueNDRangeKernel(queue[d], kernel, 2, NULL, gws, lws, 0, NULL, NULL);
    CHECK_ERROR(err);
  }

  for(int d=0; d<DEVICE_NUM; d++) {
    // DO NOT REMOVE; NEEDED FOR TIME MEASURE
    err = clFinish(queue[d]);
    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_list) {
  size_t sz;
  char *buf;
  for(int d=0; d<DEVICE_NUM; d++) {
      CHECK_ERROR(clGetDeviceInfo(device[d], CL_DEVICE_NAME, 0, NULL, &sz));
      buf = (char*)malloc(sz);
      CHECK_ERROR(clGetDeviceInfo(device[d], CL_DEVICE_NAME, sz, buf, NULL));
      printf("Detected OpenCL device(%d): %s\n", d, buf);
      free(buf);
  }
}

static cl_program create_and_build_program_with_source(cl_context context, cl_device_id *device_list, 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, DEVICE_NUM, device_list, "", NULL, NULL);
  if (err == CL_BUILD_PROGRAM_FAILURE) {
    size_t log_size;
    for(int d=0; d<DEVICE_NUM; d++) {
      CHECK_ERROR(clGetProgramBuildInfo(program, device_list[d], CL_PROGRAM_BUILD_LOG, 0, NULL, &log_size));
      char *log = (char*)malloc(log_size + 1);
      CHECK_ERROR(clGetProgramBuildInfo(program, device_list[d], 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, 1, &device, NULL);
  CHECK_ERROR(err);
  print_device_info(device);

  // Create OpenCL context
  context = clCreateContext(NULL, 1, &device, NULL, NULL, &err);
  CHECK_ERROR(err);

  // Create OpenCL command queue
  queue = clCreateCommandQueue(context, device, 0, &err);
  CHECK_ERROR(err);

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

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

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

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

  // DO NOT REMOVE; NEEDED FOR TIME MEASURE
  err = clFinish(queue);
  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)
  err = clEnqueueReadBuffer(queue, c_d, CL_TRUE, 0, M * N * sizeof(float), C, 0, NULL, NULL);
  CHECK_ERROR(err);

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

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 4 OpenCL devices
  err = clGetDeviceIDs(platform, CL_DEVICE_TYPE_GPU, DEVICE_NUM, device, NULL);
  CHECK_ERROR(err);
  print_device_info(device);

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

  // Create OpenCL command queue
  for(int i=0; i<DEVICE_NUM; i++) {
    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, "kernel.cl");

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

  // Divide A and B matrix equally.
  for(int d=0; d<DEVICE_NUM; d++) {
	  M_start[d] = M/DEVICE_NUM*d;
	  M_size[d] = (d == DEVICE_NUM-1) ? M - M/DEVICE_NUM*(DEVICE_NUM-1)
                                          : M/DEVICE_NUM;
  }

  // Create GPU buffers
  for(int d=0; d<DEVICE_NUM; d++) {
    a_d[d] = clCreateBuffer(context, CL_MEM_READ_WRITE, M_size[d] * K * sizeof(float), NULL, &err);
    CHECK_ERROR(err);
  }
  for(int d=0; d<DEVICE_NUM; d++) {
    b_d[d] = clCreateBuffer(context, CL_MEM_READ_WRITE, K * N * sizeof(float), NULL, &err);
    CHECK_ERROR(err);
  }
  for(int d=0; d<DEVICE_NUM; d++) {
    c_d[d] = clCreateBuffer(context, CL_MEM_READ_WRITE, M_size[d] * N * sizeof(float), NULL, &err);
    CHECK_ERROR(err);
  }

  // Need to use all queue ???
  // Write to GPU; A (cpu) -> a_d (gpu), B (cpu) -> b_d (gpu)
  for(int d=0; d<DEVICE_NUM; d++) {
    //printf("WriteBuffer\n");
    err = clEnqueueWriteBuffer(queue[d], a_d[d], CL_TRUE, 0, M_size[d] * K * sizeof(float), A + M_start[d] * K, 0, NULL, NULL);
    CHECK_ERROR(err);
  }
  for(int d=0; d<DEVICE_NUM; d++) {
    err = clEnqueueWriteBuffer(queue[d], b_d[d], CL_TRUE, 0, K * N * sizeof(float), B, 0, NULL, NULL);
    CHECK_ERROR(err);
  }

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

void mat_mul_final(float *A, float *B, float *C, int M, int N, int K) {

  for(int d=0; d<DEVICE_NUM; d++) {
    // Read from GPU; c_d (gpu) -> C (cpu)
    err = clEnqueueReadBuffer(queue[d], c_d[d], CL_TRUE, 0, M_size[d] * N * sizeof(float), C + M_start[d] * N, 0, NULL, NULL);
    CHECK_ERROR(err);
  }

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