#include "mat_mul.h"
#include "util.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 MAX_DEV 4
#define WPT 8

#define sizeA  (M * K * sizeof(float))
#define sizeB  (K * N * sizeof(float))
#define sizeC  (M * N * sizeof(float))
#define sizeEA (extra * K * sizeof(float))
#define sizeEC (extra * N * sizeof(float))

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;
static cl_kernel kernel[MAX_DEV];
static int ndev;
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 extra;

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 i = 0; i < ndev; i++) {
		M = (i == ndev - 1 && extra)? M/ndev + extra : M/ndev;

		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), &M);
		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);

  	// Setup global work size and local work size
  	size_t gws[2] = {(size_t)(M + WPT - 1)/WPT, (size_t)N}, lws[2] = {32/WPT, 32};
  	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("GPU(%d), M(%d), gws(%d, %d)\n", i + 1, M, (int)gws[0], (int)gws[1]);

  	// Run kernel
		err = clEnqueueNDRangeKernel(queue[i], kernel[i], 2, NULL, gws, lws, 0, NULL, NULL);
		CHECK_ERROR(err);

		M = _M;
	}

  // DO NOT REMOVE; NEEDED FOR TIME MEASURE
	for (int i = 0; i < ndev; 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;
	for (int i = 0; i < ndev; i++) {
		CHECK_ERROR(clGetDeviceInfo(device[i], CL_DEVICE_NAME, 0, NULL, &sz));
		buf = (char*)malloc(sz);
		CHECK_ERROR(clGetDeviceInfo(device[i], 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, MAX_DEV, device, (unsigned int*)&ndev);
  CHECK_ERROR(err);
  print_device_info(device);

	// Use one GPU only if M is smaller than MAX_DEV
	if (M < MAX_DEV) ndev = 1;

	//int tempM = M;
	extra = M % ndev;
	M /= ndev;	

  // 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");

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

  // Create GPU buffers
	for (int i = 0; i < ndev; i++) {
		if (i == ndev - 1 && extra) {
			//printf("%d extra rows are allocated to (%d)th GPU\n", extra, i + 1);
			a_d[i] = clCreateBuffer(context, CL_MEM_READ_WRITE, sizeA + sizeEA, NULL, &err);
			CHECK_ERROR(err);
			c_d[i] = clCreateBuffer(context, CL_MEM_READ_WRITE, sizeC + sizeEC, NULL, &err);
			CHECK_ERROR(err);
		}
		else {
			a_d[i] = clCreateBuffer(context, CL_MEM_READ_WRITE, sizeA, NULL, &err);
			CHECK_ERROR(err);
			c_d[i] = clCreateBuffer(context, CL_MEM_READ_WRITE, sizeC, NULL, &err);
			CHECK_ERROR(err);
		}

		b_d[i] = clCreateBuffer(context, CL_MEM_READ_WRITE, sizeB, NULL, &err);
		CHECK_ERROR(err);
	}

  // Write to GPU; A (cpu) -> a_d (gpu), B (cpu) -> b_d (gpu)
	for (int i = 0; i < ndev; i++) {
		int size = (i == ndev - 1 && extra)? sizeA + sizeEA : sizeA;
		err = clEnqueueWriteBuffer(queue[i], a_d[i], CL_TRUE, 0, size, (void*)((size_t)A + sizeA * i), 0, NULL, NULL);
		CHECK_ERROR(err);
		err = clEnqueueWriteBuffer(queue[i], b_d[i], CL_TRUE, 0, sizeB, B, 0, NULL, NULL);
		CHECK_ERROR(err);
		
		//float * dA[4];
		//alloc_mat(&dA[i], size / (K*sizeof(float)), K);
	  //err = clEnqueueReadBuffer(queue[i], a_d[i], CL_TRUE, 0, size, (void*)dA[i], 0, NULL, NULL);
		//CHECK_ERROR(err);
		//printf("A matrix of (%d) GPU, size(%d), M(%d), N(%d), sizeA(%d), sizeEA(%d), extra(%d)\n", i+1, (int)size, M, N, (int)sizeA, (int)sizeEA, extra);
		//print_mat(dA[i], size / (K*sizeof(float)), K);
	}

	//printf("Original A matrix\n");
	//print_mat(A, tempM, K);
	

  // 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) {
	M /= ndev;

  // Read from GPU; c_d (gpu) -> C (cpu)
	for (int i = 0; i < ndev; i++) {
		if (i == ndev - 1 && extra) {
			err = clEnqueueReadBuffer(queue[i], c_d[i], CL_TRUE, 0, sizeC + sizeEC, (void*)((size_t)C + sizeC * i), 0, NULL, NULL);
		}
		else {
			err = clEnqueueReadBuffer(queue[i], c_d[i], CL_TRUE, 0, sizeC, (void*)((size_t)C + sizeC * i), 0, NULL, NULL);
		}
		//printf("M(%d), N(%d), sizeC(%d), sizeEC(%d)\n", M, N, (int)sizeC, (int)sizeEC);
		CHECK_ERROR(err);
	}

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