#include "mat_mul.h"
#include <omp.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 RTS TS/WPT

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;
static cl_kernel kernel[4];
static cl_mem a_d, b_d, c_d, a_d1, b_d1, c_d1, a_d2, b_d2, c_d2, a_d3, b_d3, c_d3;
static float *A, *B, *C;
//static float *A_new, *B_new, *C_new;
static int M, N, K;
static int ndev;
static int M_div[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;
  
  err = clSetKernelArg(kernel[0], 0, sizeof(cl_mem), &a_d);
  CHECK_ERROR(err);
  err = clSetKernelArg(kernel[0], 1, sizeof(cl_mem), &b_d);
  CHECK_ERROR(err);
  err = clSetKernelArg(kernel[0], 2, sizeof(cl_mem), &c_d);
  CHECK_ERROR(err);
  err = clSetKernelArg(kernel[0], 3, sizeof(int), &M_div[0]);
  CHECK_ERROR(err);
  err = clSetKernelArg(kernel[0], 4, sizeof(int), &N);
  CHECK_ERROR(err);
  err = clSetKernelArg(kernel[0], 5, sizeof(int), &K);
  CHECK_ERROR(err);

  err = clSetKernelArg(kernel[1], 0, sizeof(cl_mem), &a_d1);
  CHECK_ERROR(err);
  err = clSetKernelArg(kernel[1], 1, sizeof(cl_mem), &b_d1);
  CHECK_ERROR(err);
  err = clSetKernelArg(kernel[1], 2, sizeof(cl_mem), &c_d1);
  CHECK_ERROR(err);
  err = clSetKernelArg(kernel[1], 3, sizeof(int), &M_div[1]);
  CHECK_ERROR(err);
  err = clSetKernelArg(kernel[1], 4, sizeof(int), &N);
  CHECK_ERROR(err);
  err = clSetKernelArg(kernel[1], 5, sizeof(int), &K);
  CHECK_ERROR(err);

  err = clSetKernelArg(kernel[2], 0, sizeof(cl_mem), &a_d2);
  CHECK_ERROR(err);
  err = clSetKernelArg(kernel[2], 1, sizeof(cl_mem), &b_d2);
  CHECK_ERROR(err);
  err = clSetKernelArg(kernel[2], 2, sizeof(cl_mem), &c_d2);
  CHECK_ERROR(err);
  err = clSetKernelArg(kernel[2], 3, sizeof(int), &M_div[2]);
  CHECK_ERROR(err);
  err = clSetKernelArg(kernel[2], 4, sizeof(int), &N);
  CHECK_ERROR(err);
  err = clSetKernelArg(kernel[2], 5, sizeof(int), &K);
  CHECK_ERROR(err);

  err = clSetKernelArg(kernel[3], 0, sizeof(cl_mem), &a_d3);
  CHECK_ERROR(err);
  err = clSetKernelArg(kernel[3], 1, sizeof(cl_mem), &b_d3);
  CHECK_ERROR(err);
  err = clSetKernelArg(kernel[3], 2, sizeof(cl_mem), &c_d3);
  CHECK_ERROR(err);
  err = clSetKernelArg(kernel[3], 3, sizeof(int), &M_div[3]);
  CHECK_ERROR(err);
  err = clSetKernelArg(kernel[3], 4, sizeof(int), &N);
  CHECK_ERROR(err);
  err = clSetKernelArg(kernel[3], 5, sizeof(int), &K);
  CHECK_ERROR(err);

  for(int i=0; i<ndev; i++) {
     // Setup global work size and local work size
     size_t gws[2] = {(size_t)(M_div[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];
     }
   
     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, ndev, 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, 0, NULL, (unsigned int*) &ndev);
  CHECK_ERROR(err);
  err = clGetDeviceIDs(platform, CL_DEVICE_TYPE_GPU, ndev, device, NULL);
  CHECK_ERROR(err);
  for(int i=0; i<ndev; i++)
     print_device_info(device[i]);

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

  // Create OpenCL command queue
  for(int i=0; i< ndev; 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");

  for(int i=0; i< ndev; i++){
     kernel[i] = clCreateKernel(program, "sgemm", &err); 
     CHECK_ERROR(err);
  }

  int M_temp = M/ndev;
  for(int i=0; i<ndev; ++i){
     M_div[i]= M_temp;
     if(i == ndev-1)
       M_div[i] = M_div[i]+ M%4;
  }
  // Create GPU buffers
  a_d = clCreateBuffer(context, CL_MEM_READ_WRITE, M/4 * 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/4 * N * sizeof(float), NULL, &err);
  CHECK_ERROR(err);

  a_d1 = clCreateBuffer(context, CL_MEM_READ_WRITE, M/4 * K * sizeof(float), NULL, &err);
  CHECK_ERROR(err);
  b_d1 = clCreateBuffer(context, CL_MEM_READ_WRITE, K * N * sizeof(float), NULL, &err);
  CHECK_ERROR(err);
  c_d1 = clCreateBuffer(context, CL_MEM_READ_WRITE, M/4 * N * sizeof(float), NULL, &err);
  CHECK_ERROR(err);

  a_d2 = clCreateBuffer(context, CL_MEM_READ_WRITE, M/4 * K * sizeof(float), NULL, &err);
  CHECK_ERROR(err);
  b_d2 = clCreateBuffer(context, CL_MEM_READ_WRITE, K * N * sizeof(float), NULL, &err);
  CHECK_ERROR(err);
  c_d2 = clCreateBuffer(context, CL_MEM_READ_WRITE, M/4 * N * sizeof(float), NULL, &err);
  CHECK_ERROR(err);

  a_d3 = clCreateBuffer(context, CL_MEM_READ_WRITE, (M/4 + M%4) * K * sizeof(float), NULL, &err);
  CHECK_ERROR(err);
  b_d3 = clCreateBuffer(context, CL_MEM_READ_WRITE, K * N * sizeof(float), NULL, &err);
  CHECK_ERROR(err);
  c_d3 = clCreateBuffer(context, CL_MEM_READ_WRITE, (M/4 + M%4) * N * sizeof(float), NULL, &err);
  CHECK_ERROR(err);


  // Write to GPU; A (cpu) -> a_d (gpu), B (cpu) -> b_d (gpu)
  err = clEnqueueWriteBuffer(queue[0], a_d, CL_TRUE, 0, M/4 * K * sizeof(float), (void*)((size_t)A), 0, NULL, NULL);
  CHECK_ERROR(err);
  err = clEnqueueWriteBuffer(queue[0], b_d, CL_TRUE, 0, K * N * sizeof(float), B, 0, NULL, NULL);
  CHECK_ERROR(err);

  err = clEnqueueWriteBuffer(queue[1], a_d1, CL_TRUE, 0, M/4 * K * sizeof(float), (void*)((size_t)A+(M/4*K*sizeof(float))), 0, NULL, NULL);
  CHECK_ERROR(err);
  err = clEnqueueWriteBuffer(queue[1], b_d1, CL_TRUE, 0, K * N * sizeof(float), B, 0, NULL, NULL);
  CHECK_ERROR(err);

  err = clEnqueueWriteBuffer(queue[2], a_d2, CL_TRUE, 0, M/4 * K * sizeof(float), (void*)((size_t)A+(M/4*K*sizeof(float)*2)), 0, NULL, NULL);
  CHECK_ERROR(err);
  err = clEnqueueWriteBuffer(queue[2], b_d2, CL_TRUE, 0, K * N * sizeof(float), B, 0, NULL, NULL);
  CHECK_ERROR(err);

  err = clEnqueueWriteBuffer(queue[3], a_d3, CL_TRUE, 0, (M/4+M%4) * K * sizeof(float), (void*)((size_t)A+(M/4*K*sizeof(float)*3)), 0, NULL, NULL);
  CHECK_ERROR(err);
  err = clEnqueueWriteBuffer(queue[3], b_d3, 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<ndev; 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)
  err = clEnqueueReadBuffer(queue[0], c_d, CL_TRUE, 0, M/4 * N * sizeof(float), (void*)((size_t)C), 0, NULL, NULL);
  CHECK_ERROR(err);
  err = clEnqueueReadBuffer(queue[1], c_d1, CL_TRUE, 0, M/4 * N * sizeof(float), (void*)((size_t)C+(M/4*N*sizeof(float))), 0, NULL, NULL);
  CHECK_ERROR(err);
  err = clEnqueueReadBuffer(queue[2], c_d2, CL_TRUE, 0, M/4 * N * sizeof(float), (void*)((size_t)C+(M/4*N*sizeof(float)*2)), 0, NULL, NULL);
  CHECK_ERROR(err);
  err = clEnqueueReadBuffer(queue[3], c_d3, CL_TRUE, 0, (M/4+M%4) * N * sizeof(float), (void*)((size_t)C+(M/4*N*sizeof(float)*3)), 0, NULL, NULL);
  CHECK_ERROR(err);


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