#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); \
  }

#define NUM_GPUS 4
#define BLOCK_SIZE 16

static cl_int err;
static cl_platform_id platform;
static cl_device_id device[NUM_GPUS];
static cl_context context;
static cl_command_queue queue[NUM_GPUS];
static cl_program program[NUM_GPUS];
static cl_kernel kernel[NUM_GPUS];
static cl_mem a_d[NUM_GPUS], b_d[NUM_GPUS], c_d[NUM_GPUS];
static cl_uint deviceCount;
static int workOffset[NUM_GPUS];
static int workSize[NUM_GPUS];

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 global work size and local work size
  size_t gws[2] = {(size_t)N, (size_t)workSize[0]}, lws[2] = {BLOCK_SIZE, BLOCK_SIZE};

  // Run kernel
  for (unsigned int i = 0; i < deviceCount; i++) {
    gws[0] = ((size_t)N + lws[0] - 1) / lws[0] * lws[0];
    gws[1] = ((size_t)workSize[i] + lws[1] - 1) / lws[1] * lws[1];
    err = clEnqueueNDRangeKernel(queue[i], kernel[i], 2, NULL, gws, lws, 0, NULL, NULL);
    CHECK_ERROR(err);
  }

  // DO NOT REMOVE; NEEDED FOR TIME MEASURE
  for (unsigned int i = 0; i < deviceCount; 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, &deviceCount);
  CHECK_ERROR(err);
  err = clGetDeviceIDs(platform, CL_DEVICE_TYPE_GPU, deviceCount, device, NULL);
  CHECK_ERROR(err);
  for (unsigned int i = 0; i < deviceCount; i++) {
    print_device_info(device[i]);
  }

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

  // Create OpenCL command queue
  for (unsigned int i = 0; i < deviceCount; i++) {
    queue[i] = clCreateCommandQueue(context, device[i], 0, &err);
    CHECK_ERROR(err);
  }

  // Compile program from "kernel.cl"
  for (unsigned int i = 0; i < deviceCount; i++) {
    program[i] = create_and_build_program_with_source(context, device[i], "kernel.cl");
  }

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

  // Create GPU buffers
  int rounded_N = (N + BLOCK_SIZE - 1) / BLOCK_SIZE * BLOCK_SIZE;
  int sizePerGPU = M / deviceCount;
  cl_uint init_mem = 0;
  workOffset[0] = 0;
  for (unsigned int i = 0; i < deviceCount; i++) {
    // input buffer
    workSize[i] = (i != (deviceCount - 1)) ? sizePerGPU : (M - workOffset[i]);
    int rounded_M = (workSize[i] + BLOCK_SIZE - 1) / BLOCK_SIZE * BLOCK_SIZE;

    a_d[i] = clCreateBuffer(context, CL_MEM_READ_ONLY, rounded_M * K * sizeof(float), NULL, &err);
    CHECK_ERROR(err);
    b_d[i] = clCreateBuffer(context, CL_MEM_READ_ONLY, K * rounded_N * sizeof(float), NULL, &err);
    CHECK_ERROR(err);
    c_d[i] = clCreateBuffer(context, CL_MEM_WRITE_ONLY, rounded_M * rounded_N * sizeof(float), NULL, &err);
    CHECK_ERROR(err);
    err = clEnqueueFillBuffer(queue[i], a_d[i], &init_mem, sizeof(cl_uint), 0, rounded_M * K * sizeof(float), 0, NULL, NULL);
    CHECK_ERROR(err);
    err = clEnqueueFillBuffer(queue[i], b_d[i], &init_mem, sizeof(cl_uint), 0, K * rounded_N * sizeof(float), 0, NULL, NULL);
    CHECK_ERROR(err);
    err = clEnqueueFillBuffer(queue[i], c_d[i], &init_mem, sizeof(cl_uint), 0, rounded_M * rounded_N * sizeof(float), 0, NULL, NULL);
    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, workSize[i] * K * sizeof(float), &A[workOffset[i] * K], 0, NULL, NULL);
    CHECK_ERROR(err);
    for (int j = 0; j < K; j++) {
      err = clEnqueueWriteBuffer(queue[i], b_d[i], CL_TRUE, j * rounded_N * sizeof(float), N * sizeof(float), &B[j * N], 0, NULL, NULL);
      CHECK_ERROR(err);
    }

    if (i < deviceCount - 1) {
      workOffset[i + 1] = workOffset[i] + workSize[i];
    }
  }

  // Setup kernel arguments
  for (unsigned int i = 0; i < deviceCount; i++) {
    int n = 0;
    err = clSetKernelArg(kernel[i], n++, sizeof(cl_mem), &a_d[i]);
    CHECK_ERROR(err);
    err = clSetKernelArg(kernel[i], n++, sizeof(cl_mem), &b_d[i]);
    CHECK_ERROR(err);
    err = clSetKernelArg(kernel[i], n++, sizeof(cl_mem), &c_d[i]);
    CHECK_ERROR(err);
    err = clSetKernelArg(kernel[i], n++, sizeof(float) * BLOCK_SIZE * BLOCK_SIZE, 0);
    CHECK_ERROR(err);
    err = clSetKernelArg(kernel[i], n++, sizeof(float) * BLOCK_SIZE * BLOCK_SIZE, 0);
    CHECK_ERROR(err);
    err = clSetKernelArg(kernel[i], n++, sizeof(int), &workSize[i]);
    CHECK_ERROR(err);
    err = clSetKernelArg(kernel[i], n++, sizeof(int), &rounded_N);
    CHECK_ERROR(err);
    err = clSetKernelArg(kernel[i], n++, sizeof(int), &K);
    CHECK_ERROR(err);
  }

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

void mat_mul_final(float *A, float *B, float *C, int M, int N, int K) {
  int rounded_N = (N + BLOCK_SIZE - 1) / BLOCK_SIZE * BLOCK_SIZE;

  // Read from GPU; c_d (gpu) -> C (cpu)
  for (unsigned int i = 0; i < deviceCount; i++) {
    for (int j = 0; j < workSize[i]; j++) {
      err = clEnqueueReadBuffer(queue[i], c_d[i], CL_TRUE, j * rounded_N * sizeof(float), N * sizeof(float), &C[workOffset[i] * N + j * N], 0, NULL, NULL);
      CHECK_ERROR(err);
    }
  }

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