#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;
static cl_device_id device[4];
static cl_context context;
static cl_command_queue queue[4];
static cl_program program1;
static cl_program program2;
static cl_program program3;
static cl_program program4;
static cl_kernel kernel1;
static cl_kernel kernel2;
static cl_kernel kernel3;
static cl_kernel kernel4;
static cl_mem a1_d, b1_d, c1_d;
static cl_mem a2_d, b2_d, c2_d;
static cl_mem a3_d, b3_d, c3_d;
static cl_mem a4_d, b4_d, c4_d;

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(kernel1, 0, sizeof(cl_mem), &a1_d);
  CHECK_ERROR(err);
  err = clSetKernelArg(kernel1, 1, sizeof(cl_mem), &b1_d);
  CHECK_ERROR(err);
  err = clSetKernelArg(kernel1, 2, sizeof(cl_mem), &c1_d);
  CHECK_ERROR(err);
  err = clSetKernelArg(kernel1, 3, sizeof(int), &M);
  CHECK_ERROR(err);
  err = clSetKernelArg(kernel1, 4, sizeof(int), &N);
  CHECK_ERROR(err);
  err = clSetKernelArg(kernel1, 5, sizeof(int), &K);
  CHECK_ERROR(err);

  err = clSetKernelArg(kernel2, 0, sizeof(cl_mem), &a2_d);
  CHECK_ERROR(err);
  err = clSetKernelArg(kernel2, 1, sizeof(cl_mem), &b2_d);
  CHECK_ERROR(err);
  err = clSetKernelArg(kernel2, 2, sizeof(cl_mem), &c2_d);
  CHECK_ERROR(err);
  err = clSetKernelArg(kernel2, 3, sizeof(int), &M);
  CHECK_ERROR(err);
  err = clSetKernelArg(kernel2, 4, sizeof(int), &N);
  CHECK_ERROR(err);
  err = clSetKernelArg(kernel2, 5, sizeof(int), &K);
  CHECK_ERROR(err);

  err = clSetKernelArg(kernel3, 0, sizeof(cl_mem), &a3_d);
  CHECK_ERROR(err);
  err = clSetKernelArg(kernel3, 1, sizeof(cl_mem), &b3_d);
  CHECK_ERROR(err);
  err = clSetKernelArg(kernel3, 2, sizeof(cl_mem), &c3_d);
  CHECK_ERROR(err);
  err = clSetKernelArg(kernel3, 3, sizeof(int), &M);
  CHECK_ERROR(err);
  err = clSetKernelArg(kernel3, 4, sizeof(int), &N);
  CHECK_ERROR(err);
  err = clSetKernelArg(kernel3, 5, sizeof(int), &K);
  CHECK_ERROR(err);

  err = clSetKernelArg(kernel4, 0, sizeof(cl_mem), &a4_d);
  CHECK_ERROR(err);
  err = clSetKernelArg(kernel4, 1, sizeof(cl_mem), &b4_d);
  CHECK_ERROR(err);
  err = clSetKernelArg(kernel4, 2, sizeof(cl_mem), &c4_d);
  CHECK_ERROR(err);
  err = clSetKernelArg(kernel4, 3, sizeof(int), &M);
  CHECK_ERROR(err);
  err = clSetKernelArg(kernel4, 4, sizeof(int), &N);
  CHECK_ERROR(err);
  err = clSetKernelArg(kernel4, 5, sizeof(int), &K);
  CHECK_ERROR(err);

int m1,m2,m3,m4;
int m = (M/32)*32;
m1 = m/4;
m2 = m1*2;
m3 = m1*3;
m4 = m1*4 + m%4;

size_t M1,M2,M3,M4;
M1 = m1;
M2 = m2 - m1;
M3 = m3 - m2;
M4 = m4 - m3;

size_t TS = 32;
size_t WPT = 8;
  // Setup global work size and local work size
//  size_t gws1[2] = {M/WPT, N}, lws1[2] = {TS/WPT, TS};
   size_t gws1[2] = {(size_t)M1/WPT, (size_t)N}, lws1[2] = {TS/WPT, TS};
   size_t gws2[2] = {(size_t)M2/WPT, (size_t)N}, lws2[2] = {TS/WPT, TS};
   size_t gws3[2] = {(size_t)M3/WPT, (size_t)N}, lws3[2] = {TS/WPT, TS};
   size_t gws4[2] = {(size_t)M4/WPT, (size_t)N}, lws4[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

    gws1[i] = (gws1[i] + lws1[i] - 1) / lws1[i] * lws1[i];
    gws2[i] = (gws2[i] + lws2[i] - 1) / lws2[i] * lws2[i];
    gws3[i] = (gws3[i] + lws3[i] - 1) / lws3[i] * lws3[i];
    gws4[i] = (gws4[i] + lws3[i] - 1) / lws4[i] * lws4[i];

  }
      
  // Run kernel
  err = clEnqueueNDRangeKernel(queue[0], kernel1, 2, NULL, gws1, lws1, 0, NULL, NULL);
  CHECK_ERROR(err);

  err = clEnqueueNDRangeKernel(queue[1], kernel2, 2, NULL, gws2, lws2, 0, NULL, NULL);
  CHECK_ERROR(err);

  err = clEnqueueNDRangeKernel(queue[2], kernel3, 2, NULL, gws3, lws3, 0, NULL, NULL);
  CHECK_ERROR(err);

  err = clEnqueueNDRangeKernel(queue[3], kernel4, 2, NULL, gws4, lws4, 0, NULL, NULL);
  CHECK_ERROR(err);
  // DO NOT REMOVE; NEEDED FOR TIME MEASURE
  err = clFinish(queue[0]);
  CHECK_ERROR(err);
  err = clFinish(queue[1]);
  CHECK_ERROR(err);
  err = clFinish(queue[2]);
  CHECK_ERROR(err);
  err = clFinish(queue[3]);
  CHECK_ERROR(err);

  for(int i = m4; i < M; i++){
    for(int k=0; k < K; k++){
            float aik = A[i*K + k];
      for(int j = 0; j < N; j++){
          C[i*N + j] += aik * B[k*N + j];
      }
    }
  }

}

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));
  printf("Detected OpenCL device11: %d\n", CL_DEVICE_NAME);
  buf = (char*)malloc(sz);
  CHECK_ERROR(clGetDeviceInfo(device, CL_DEVICE_NAME, sz, buf, NULL));
  printf("Detected OpenCL device: %s\n", buf);
  printf("Detected OpenCL device22: %d\n", CL_DEVICE_NAME);
  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, 4, device, NULL);
  CHECK_ERROR(err);
  print_device_info(device[0]);
  print_device_info(device[1]);
  print_device_info(device[2]);
  print_device_info(device[3]);

  // Create OpenCL context
  context = clCreateContext(NULL, 4, device, NULL, NULL, &err);
  CHECK_ERROR(err);

  // Create OpenCL command queue
  queue[0] = clCreateCommandQueue(context, device[0], 0, &err);
  CHECK_ERROR(err);
  queue[1] = clCreateCommandQueue(context, device[1], 0, &err);
  CHECK_ERROR(err);
  queue[2] = clCreateCommandQueue(context, device[2], 0, &err);
  CHECK_ERROR(err);
  queue[3] = clCreateCommandQueue(context, device[3], 0, &err);
  CHECK_ERROR(err);

  // Compile program from "kernel.cl"
  program1 = create_and_build_program_with_source(context, device[0], "kernel.cl");
  program2 = create_and_build_program_with_source(context, device[1], "kernel.cl");
  program3 = create_and_build_program_with_source(context, device[2], "kernel.cl");
  program4 = create_and_build_program_with_source(context, device[3], "kernel.cl");

  // Extract kernel from compiled program
  kernel1 = clCreateKernel(program1, "sgemm", &err); 
  CHECK_ERROR(err);

  kernel2 = clCreateKernel(program2, "sgemm", &err); 
  CHECK_ERROR(err);

  kernel3 = clCreateKernel(program3, "sgemm", &err); 
  CHECK_ERROR(err);

  kernel4 = clCreateKernel(program4, "sgemm", &err); 
  CHECK_ERROR(err);

int m1,m2,m3,m4;
int m = (M/32)*32;
m1 = m/4;
m2 = m1*2;
m3 = m1*3;
m4 = m1*4 + m%4;

  // Create GPU buffers
  a1_d = clCreateBuffer(context, CL_MEM_READ_WRITE, m1 * K * sizeof(float), NULL, &err);
  CHECK_ERROR(err);
  b1_d = clCreateBuffer(context, CL_MEM_READ_WRITE, K * N * sizeof(float), NULL, &err);
  CHECK_ERROR(err);
  c1_d = clCreateBuffer(context, CL_MEM_READ_WRITE, m1 * N * sizeof(float), NULL, &err);
  CHECK_ERROR(err);


  a2_d = clCreateBuffer(context, CL_MEM_READ_WRITE, (m2-m1) * K * sizeof(float), NULL, &err);
  CHECK_ERROR(err);
  b2_d = clCreateBuffer(context, CL_MEM_READ_WRITE, K * N * sizeof(float), NULL, &err);
  CHECK_ERROR(err);
  c2_d = clCreateBuffer(context, CL_MEM_READ_WRITE, (m2-m1) * N * sizeof(float), NULL, &err);
  CHECK_ERROR(err);


  a3_d = clCreateBuffer(context, CL_MEM_READ_WRITE, (m3-m2) * K* sizeof(float), NULL, &err);
  CHECK_ERROR(err);
  b3_d = clCreateBuffer(context, CL_MEM_READ_WRITE, K * N * sizeof(float), NULL, &err);
  CHECK_ERROR(err);
  c3_d = clCreateBuffer(context, CL_MEM_READ_WRITE, (m3-m2) * N * sizeof(float), NULL, &err);
  CHECK_ERROR(err);


  a4_d = clCreateBuffer(context, CL_MEM_READ_WRITE, (m4-m3) * K * sizeof(float), NULL, &err);
  CHECK_ERROR(err);
  b4_d = clCreateBuffer(context, CL_MEM_READ_WRITE, K * N * sizeof(float), NULL, &err);
  CHECK_ERROR(err);
  c4_d = clCreateBuffer(context, CL_MEM_READ_WRITE, (m4-m3) * 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], a1_d, CL_TRUE, 0, m1 * K * sizeof(float), &A[0], 0, NULL, NULL);
  CHECK_ERROR(err);
  err = clEnqueueWriteBuffer(queue[0], b1_d, CL_TRUE, 0, K * N * sizeof(float), B, 0, NULL, NULL);
  CHECK_ERROR(err);


  err = clEnqueueWriteBuffer(queue[1], a2_d, CL_TRUE, 0, (m2-m1) * K * sizeof(float), &A[m1*K], 0, NULL, NULL);
  CHECK_ERROR(err);
  err = clEnqueueWriteBuffer(queue[1], b2_d, CL_TRUE, 0, K * N * sizeof(float), B, 0, NULL, NULL);
  CHECK_ERROR(err);


  err = clEnqueueWriteBuffer(queue[2], a3_d, CL_TRUE, 0, (m3-m2) * K * sizeof(float), &A[m2*K], 0, NULL, NULL);
  CHECK_ERROR(err);
  err = clEnqueueWriteBuffer(queue[2], b3_d, CL_TRUE, 0, K * N * sizeof(float), B, 0, NULL, NULL);
  CHECK_ERROR(err);


  err = clEnqueueWriteBuffer(queue[3], a4_d, CL_TRUE, 0, (m4-m3) * K * sizeof(float), &A[m3*K], 0, NULL, NULL);
  CHECK_ERROR(err);
  err = clEnqueueWriteBuffer(queue[3], b4_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[0]);
  CHECK_ERROR(err);

  err = clFinish(queue[1]);
  CHECK_ERROR(err);

  err = clFinish(queue[2]);
  CHECK_ERROR(err);

  err = clFinish(queue[3]);
  CHECK_ERROR(err);
}

void mat_mul_final(float *A, float *B, float *C, int M, int N, int K) {
int m1,m2,m3,m4;
int m = (M/32)*32;
m1 = m/4;
m2 = m1*2;
m3 = m1*3;
m4 = m1*4 + m%4;
  // Read from GPU; c_d (gpu) -> C (cpu)
  err = clEnqueueReadBuffer(queue[0], c1_d, CL_TRUE, 0, m1 * N * sizeof(float), &C[0], 0, NULL, NULL);
  CHECK_ERROR(err);


  err = clEnqueueReadBuffer(queue[1], c2_d, CL_TRUE, 0, (m2-m1) * N * sizeof(float), &C[m1*N], 0, NULL, NULL);
  CHECK_ERROR(err);


  err = clEnqueueReadBuffer(queue[2], c3_d, CL_TRUE, 0, (m3-m2) * N * sizeof(float), &C[m2*N], 0, NULL, NULL);
  CHECK_ERROR(err);


  err = clEnqueueReadBuffer(queue[3], c4_d, CL_TRUE, 0, (m4-m3) * N * sizeof(float), &C[m3*N], 0, NULL, NULL);
  CHECK_ERROR(err);

  // DO NOT REMOVE; NEEDED FOR TIME MEASURE
  err = clFinish(queue[0]);
  CHECK_ERROR(err);
  err = clFinish(queue[1]);
  CHECK_ERROR(err);
  err = clFinish(queue[2]);
  CHECK_ERROR(err);
  err = clFinish(queue[3]);
  CHECK_ERROR(err);
}