#include "mat_mul.h"

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

#define MAT_COPY (0)

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

#define ALIGN_UP(_X, _Y) (((_X) + (_Y) - 1) & ~((_Y) - 1))
#define MAX_DEV (4)

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

static float *A, *B, *C;
static int M, N, K;
static int NON_OPTIMAL;
static int nDevCnt;
#if (MAT_COPY)
static float *A_backup, *B_backup, *C_backup;
static int M_backup, N_backup, K_backup;
#endif // MAT_COPY
#define NUM_WORK_ITEM (32)
#define VECTOR_WIDTH (16)
#define USING_NON_VECTOR (1)

#if (MAT_COPY)
static void mat_copy(float* __restrict pfDst, float* __restrict pfSrc, int nXDsize, int nYDsize, int nXSsize, int nYSsize, int nAddPadding)
{
  if (nAddPadding)
  {
#pragma omp parallel for
    for (int i = 0; i < nXDsize; i++)
    {
      for (int j = 0; j < nYDsize; j++)
      {
        if (i >= nXSsize || j >= nYSsize)
        {
          *(pfDst + i * nXDsize + j) = 0.0f;
        }
        else
        {
          *(pfDst + i * nXDsize + j) =*(pfSrc + i * nXSsize + j);
        }
      }
    }
  }
  else
  {    
#pragma omp parallel for
    for (int i = 0; i < nXDsize; i++)
    {
      for (int j = 0; j < nYDsize; j++)
      {
        *(pfDst + i * nXSsize + j) =*(pfSrc + i * nXDsize + j);
      }
    }
  }
}
#endif // MAT_COPY

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;


  if (_M % (NUM_WORK_ITEM * nDevCnt) != 0
    || _N % NUM_WORK_ITEM != 0
    || _K % NUM_WORK_ITEM != 0)
    {
      NON_OPTIMAL = 1; 
    }
    else
    {
      NON_OPTIMAL = 0;
    }

  // Setup kernel arguments
  for (int i = 0; i < nDevCnt; i++)
  {
    const int nLocalM = (i == (nDevCnt - 1)) ? M - ((M / nDevCnt) * (nDevCnt - 1)) : (M / nDevCnt);

    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), &nLocalM);
    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);
    err = clSetKernelArg(kernel[i], 6, sizeof(int), &NON_OPTIMAL);
    CHECK_ERROR(err);


    // Setup global work size and local work size
    size_t gws[2] , lws[2];
#if (!USING_NON_VECTOR)
    if (NON_OPTIMAL == 0)
    {
      gws[0] = (size_t)nLocalM;
      gws[1] = (size_t)N / VECTOR_WIDTH;
      lws[0] = NUM_WORK_ITEM;
      lws[1] = NUM_WORK_ITEM / VECTOR_WIDTH;    
    }
    else
#endif
    {
      gws[0] = (size_t)ALIGN_UP(nLocalM, VECTOR_WIDTH) / VECTOR_WIDTH;
      gws[1] = (size_t)ALIGN_UP(N, NUM_WORK_ITEM);
      lws[0] = NUM_WORK_ITEM / VECTOR_WIDTH;
      lws[1] = NUM_WORK_ITEM; 
    }

    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",gws[0],gws[1],lws[0],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
  for (int i = 0; i < nDevCnt; 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, 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);

  // Get OpenCL device
  err = clGetDeviceIDs(platform, CL_DEVICE_TYPE_GPU, 0, NULL, (cl_uint*)&nDevCnt);
  err = clGetDeviceIDs(platform, CL_DEVICE_TYPE_GPU, (cl_uint)nDevCnt, device, NULL);
  CHECK_ERROR(err);
  // for (int i = 0; i < nDevCnt; i++)
  {
    print_device_info(device[0]);
  }

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

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

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

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

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

    // Write to GPU; A (cpu) -> a_d (gpu), B (cpu) -> b_d (gpu)
    err = clEnqueueWriteBuffer(queue[i], a_d[i], CL_TRUE, 0, nLocalM * K * sizeof(float), &A[i * (M / nDevCnt) * K], 0, NULL, NULL);
    CHECK_ERROR(err);
    err = clEnqueueWriteBuffer(queue[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 < nDevCnt; 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)
  for (int i = 0; i < nDevCnt; i++)
  {
    int nLocalM = (i == (nDevCnt - 1)) ? M - ((M / nDevCnt) * (nDevCnt - 1)) : (M / nDevCnt);
    err = clEnqueueReadBuffer(queue[i], c_d[i], CL_TRUE, 0, nLocalM * N * sizeof(float), &C[i * (M / nDevCnt) * N], 0, NULL, NULL);
    CHECK_ERROR(err);
  }

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