#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 MAX_DEVICES 4
#define TS 32
#define WPT 32
#define TSM TS
#define TSN TS
#define TSK TS
#define CEIL_DIV(x,y) (((x) + (y) - 1) / (y))
#define PADDINGX TS
#define PADDINGY TS

static cl_int err;
static cl_platform_id platform;
static cl_device_id device[MAX_DEVICES];
static cl_context context;
static cl_command_queue queue[MAX_DEVICES];
static cl_program program[MAX_DEVICES];
static cl_kernel padakernel[MAX_DEVICES], padbkernel[MAX_DEVICES], kernel[MAX_DEVICES], removekernel[MAX_DEVICES];
static cl_mem a_d[MAX_DEVICES], b_d[MAX_DEVICES], c_d[MAX_DEVICES];
static cl_mem a_ad[MAX_DEVICES], b_ad[MAX_DEVICES], c_ad[MAX_DEVICES];

static float *A, *B, *C;
static int M, N, K;
static int K_XL;
static int M_XL[MAX_DEVICES];
static int N_XL;
static int startM[MAX_DEVICES], endM[MAX_DEVICES];

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;
  int m_size;
  size_t gws[2] = {(size_t)(M + WPT - 1)/ WPT, (size_t)N}, lws[2] = {TS/WPT, TS};

  for (int k = 0; k < MAX_DEVICES; k++) {
    m_size = endM[k] - startM[k];

    // Setup kernel arguments for adding zero padding of A
    err = clSetKernelArg(padakernel[k], 0, sizeof(int), &m_size);
    CHECK_ERROR(err);
    err = clSetKernelArg(padakernel[k], 1, sizeof(int), &K);
    CHECK_ERROR(err);
    err = clSetKernelArg(padakernel[k], 2, sizeof(cl_mem), &a_d[k]);
    CHECK_ERROR(err);
    err = clSetKernelArg(padakernel[k], 3, sizeof(int), &M_XL[k]);
    CHECK_ERROR(err);
    err = clSetKernelArg(padakernel[k], 4, sizeof(int), &K_XL);
    CHECK_ERROR(err);
    err = clSetKernelArg(padakernel[k], 5, sizeof(cl_mem), &a_ad[k]);
    CHECK_ERROR(err);

    // Setup kernel arguments for adding zero padding of B
    err = clSetKernelArg(padbkernel[k], 0, sizeof(int), &K);
    CHECK_ERROR(err);
    err = clSetKernelArg(padbkernel[k], 1, sizeof(int), &N);
    CHECK_ERROR(err);
    err = clSetKernelArg(padbkernel[k], 2, sizeof(cl_mem), &b_d[k]);
    CHECK_ERROR(err);
    err = clSetKernelArg(padbkernel[k], 3, sizeof(int), &K_XL);
    CHECK_ERROR(err);
    err = clSetKernelArg(padbkernel[k], 4, sizeof(int), &N_XL);
    CHECK_ERROR(err);
    err = clSetKernelArg(padbkernel[k], 5, sizeof(cl_mem), &b_ad[k]);
    CHECK_ERROR(err);

    // Setup kernel arguments for calculating matrix multiplication
    err = clSetKernelArg(kernel[k], 0, sizeof(cl_mem), &a_ad[k]);
    CHECK_ERROR(err);
    err = clSetKernelArg(kernel[k], 1, sizeof(cl_mem), &b_ad[k]);
    CHECK_ERROR(err);
    err = clSetKernelArg(kernel[k], 2, sizeof(cl_mem), &c_ad[k]);
    CHECK_ERROR(err);
    err = clSetKernelArg(kernel[k], 3, sizeof(int), &M_XL[k]);
    CHECK_ERROR(err);
    err = clSetKernelArg(kernel[k], 4, sizeof(int), &N_XL);
    CHECK_ERROR(err);
    err = clSetKernelArg(kernel[k], 5, sizeof(int), &K_XL);
    CHECK_ERROR(err);

    // Setup kernel arguments for removing zero padding of C
    err = clSetKernelArg(removekernel[k], 0, sizeof(int), &M_XL[k]);
    CHECK_ERROR(err);
    err = clSetKernelArg(removekernel[k], 1, sizeof(int), &N_XL);
    CHECK_ERROR(err);
    err = clSetKernelArg(removekernel[k], 2, sizeof(cl_mem), &c_ad[k]);
    CHECK_ERROR(err);
    err = clSetKernelArg(removekernel[k], 3, sizeof(int), &m_size);
    CHECK_ERROR(err);
    err = clSetKernelArg(removekernel[k], 4, sizeof(int), &N);
    CHECK_ERROR(err);
    err = clSetKernelArg(removekernel[k], 5, sizeof(cl_mem), &c_d[k]);
    CHECK_ERROR(err);

    size_t agws[2] = {(size_t)M_XL[k], (size_t)K_XL}, alws[2] = {PADDINGX, PADDINGY};
    // Run kernel for adding zero padding of A
    err = clEnqueueNDRangeKernel(queue[k], padakernel[k], 2, NULL, agws, alws, 0, NULL, NULL);
    CHECK_ERROR(err);

    size_t bgws[2] = {(size_t)K_XL, (size_t)N_XL}, blws[2] = {PADDINGX, PADDINGY};
    // Run kernel for adding zero padding of B
    err = clEnqueueNDRangeKernel(queue[k], padbkernel[k], 2, NULL, bgws, blws, 0, NULL, NULL);
    CHECK_ERROR(err);

    // Setup global work size and local work size
    gws[0] = (size_t)(m_size + WPT - 1)/ WPT;
 
    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];
    }

    // Run kernel for calculating matrix multiplication
    err = clEnqueueNDRangeKernel(queue[k], kernel[k], 2, NULL, gws, lws, 0, NULL, NULL);
    CHECK_ERROR(err);

    size_t cgws[2] = {(size_t)M_XL[k], (size_t)N_XL}, clws[2] = {PADDINGX, PADDINGY};
    // Run kernel for removing zero padding of C
    err = clEnqueueNDRangeKernel(queue[k], removekernel[k], 2, NULL, cgws, clws, 0, NULL, NULL);
    CHECK_ERROR(err);

    // DO NOT REMOVE; NEEDED FOR TIME MEASURE
    err = clFinish(queue[k]);
    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;
}

/*
static cl_program create_and_build_n_program_with_source(cl_context context, cl_device_id device, int ndev, 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) {

  int slice = M / MAX_DEVICES, m_size;
  K_XL = CEIL_DIV(K, TSK) * TSK;
  N_XL = CEIL_DIV(N, TSN) * TSN;

  // 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_DEVICES, device, NULL);
  CHECK_ERROR(err);

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

  for (int i = 0; i < MAX_DEVICES; i++) {
    print_device_info(device[i]);

    // Create OpenCL command queue
    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");

    padakernel[i] = clCreateKernel(program[i], "paddingAddZeroes", &err);
    CHECK_ERROR(err);

    padbkernel[i] = clCreateKernel(program[i], "paddingAddZeroes", &err);
    CHECK_ERROR(err);

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

    removekernel[i] = clCreateKernel(program[i], "paddingRemoveZeroes", &err);
    CHECK_ERROR(err);

    startM[i] = i * slice;
    endM[i] = (i == MAX_DEVICES - 1) ? M : (i + 1) * slice;
    m_size = endM[i] - startM[i];
	M_XL[i] = CEIL_DIV(m_size, TSM) * TSM;

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

    a_ad[i] = clCreateBuffer(context, CL_MEM_READ_ONLY, M_XL[i] * K_XL * sizeof(float), NULL, &err);
    CHECK_ERROR(err);
    b_ad[i] = clCreateBuffer(context, CL_MEM_READ_ONLY, K_XL * N_XL * sizeof(float), NULL, &err);
    CHECK_ERROR(err);
    c_ad[i] = clCreateBuffer(context, CL_MEM_READ_WRITE, M_XL[i] * N_XL * 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_FALSE, 0, m_size * K * sizeof(float), &A[startM[i] * K], 0, NULL, NULL);
    CHECK_ERROR(err);
    err = clEnqueueWriteBuffer(queue[i], b_d[i], CL_FALSE, 0, K * N * sizeof(float), B, 0, NULL, NULL);
    CHECK_ERROR(err);
  }
  
  for (int i = 0; i < MAX_DEVICES; i++) {
    // DO NOT REMOVE; NEEDED FOR TIME MEASURE
    err = clFinish(queue[i]);
    CHECK_ERROR(err);
  }
}

void mat_mul_final(float *A, float *B, float *C, int M, int N, int K) {
  int m_size;
  for (int i = 0; i < MAX_DEVICES; i++) {
    m_size = endM[i] - startM[i];

    // Read from GPU; c_d (gpu) -> C (cpu)
    err = clEnqueueReadBuffer(queue[i], c_d[i], CL_FALSE, 0, m_size * N * sizeof(float), &C[startM[i] * N], 0, NULL, NULL);
    CHECK_ERROR(err);
  }
  for (int i = 0; i < MAX_DEVICES; i++) {
    // DO NOT REMOVE; NEEDED FOR TIME MEASURE
    err = clFinish(queue[i]);
    CHECK_ERROR(err);
  }
}