/*
 * A cuBLAS implementation of the Matrix-Vector multiplication
 * 
 * Author: Petros Anastasiadis(panastas@cslab.ece.ntua.gr) 
 *
 * For more info about cuBLAS see http://docs.nvidia.com/cuda/cublas/index.html 
 *
 */

#include <errno.h>
#include <stdlib.h>
#include <stdio.h>
#include <string.h>
#include <cuda_runtime.h>
#include <cublas_v2.h>
#include <cusparse_v2.h>
/* Need to include External_Functions for these */
#include "matrix_op.h"
#include "util.h"
#include "input.h"
#include "gpu_util.h"


int main(int argc, char **argv)
{
	/* Initializations */
	int i, j, n, m;
	double timer;

	if (argc < 3) error("Usage: ./Program N M");
	else if ( argc == 3) { /*./Program N M */
		n = atoi(argv[1]);
		m = atoi(argv[2]);		
	}
	else error("Too many Arguments");

	/* Allocate space */
	double *x 			= (double *) malloc(m * sizeof(*x));
	double *M 			= (double *) malloc(n * m * sizeof(*M));
	if( !x || !M ) error("memory allocation failed");

	/* Initialize matrices */
	ser_matrix_init_rand(M,n,m,1.0); /* Normal matrices generated randomly */

	/* Initialize vectors */
	vec_init_rand(x, m, 1.0);

	/* Initialize cuda/cublas variables */
	int device_num=0;
	cudaGetDeviceCount(&device_num);
	if (!device_num) {
		printf("No available Cuda Devices...terminating");
		return 0;
	}
	double alf=1.0; /* Y=a*A*x+b */
	double beta=0.0;
	cublasHandle_t handle;
	double *A, * y, *x_c;
	
	printf("Single GPU cuBLAS Version(N=%d, M=%d): ", n, m);

	/* Initialize Unified memmory visible and accesible from both CPU and GPU */
	cudaMallocManaged(&A, m*n * sizeof(double));
	cudaMallocManaged(&y, n * sizeof(double));
	cudaMallocManaged(&x_c, m * sizeof(double));
	cudaDeviceSynchronize();
	cudaCheckErrors("Unified Alloc failed");
	if ( !A || !y || !x_c) error("unified alloc failed");
	for (i = 0; i < m; i++) x_c[i] = x[i];
	matrix_col_major(M, A, n, m); /* We transpose the matrix because cuBLAS works with column-major format */
	cublasCreate(&handle);

	/*GPU Warmup */
	cublasDgemv(handle, CUBLAS_OP_N, n, m, &alf, A , n, x_c, 1, &beta, y, 1);
	cudaDeviceSynchronize();

	timer=csecond();
	for (j = 0; j < NR_ITER; ++j) {	
			cublasDgemv(handle, CUBLAS_OP_N, n, m, &alf, A , n, x_c, 1, &beta, y, 1);
			cudaDeviceSynchronize();
	}
	timer = csecond() - timer;	
	cudaCheckErrors("cublasDgemv failed");
	
#ifdef _DEBUG_ 
	/* Output y vector to a file for debugging */
	FILE * fp;
	char filename[] = "cuBLAS.debug" ; /* Common directory for all implementations, change if needed */
	if(( fp = fopen( filename, "w")) == NULL)  error("Output file creation failed\n");
	for (i = 0; i < n; ++i) fprintf(fp, "%lf ", y[i]) ;
	fclose(fp) ;
#endif
	report_results(timer);

	return 0;
}