cublas problem: some blas 1 functions extremely slow!

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi, all

I found that after calling “cublasSgemv”, calling function “cublasSdot” or “cublasSnrm2” will be extremely slow !!!

my test function is very simple, after setting values for matrix dev_A and vector dev_x, firstly compute

dev_v= dev_A * dev_x;

then compute dev_v dot product dev_v;

my result is

GPU sgemv time:0.038000

GPU sdot time:11.852000

here are my codes:

#include <cuda.h>

#include <stdio.h>

#include <cutil.h>

#include <cublas.h>

#include <cuda_runtime_api.h>

#define size 10000

// Global variables

  float A;

  float x;

// random generator

  void matrixGen(float* A, int m, int n)

  {

	for (int i=0; i<m; i++){

	  for (int j=0; j<n; j++)

		{

		  *(A+i*n+j)=(float)rand()/RAND_MAX;

		}

	}

  }

main(){

float * dev_A;

  float * dev_x;

  float * dev_v;

  float beta;

  int N=size;

  cublasStatus stat0;

matrixGen(A, N, N);

  matrixGen(x, N, 1);

  cublasInit();

// allocate memory and set variable

stat0=cublasAlloc(N*N, sizeof(float), (void**)&dev_A);

  stat0=cublasAlloc(N, sizeof(float), (void**)&dev_v);

  stat0=cublasAlloc(N, sizeof(float), (void**)&dev_x);

  stat0=cublasSetMatrix(N, N, sizeof(float), A, N, dev_A, N);

  stat0=cublasSetVector(N, sizeof(float), x, 1, dev_x, 1);

unsigned int  timer=0;

 CUT_SAFE_CALL(cutCreateTimer(&timer));

 CUT_SAFE_CALL(cutStartTimer(timer));

 cublasSgemv('T', N, N, 1.0, dev_A, N, dev_x, 1, 0, dev_v, 1);

CUT_SAFE_CALL(cutStopTimer(timer));

 printf("\n\n GPU sgemv time:%f\n", cutGetTimerValue(timer));

 CUT_SAFE_CALL(cutDeleteTimer(timer));

timer=0;

 CUT_SAFE_CALL(cutCreateTimer(&timer));

 CUT_SAFE_CALL(cutStartTimer(timer));

beta=cublasSdot(N, dev_v, 1, dev_v, 1);

CUT_SAFE_CALL(cutStopTimer(timer));

 printf("\n\n GPU sdot time:%f\n", cutGetTimerValue(timer));

 CUT_SAFE_CALL(cutDeleteTimer(timer));

}

does any expert have any idea why is that?

many thanks in advance!

2

Your timing is wrong. CUBLAS runs asynchronous to the host, just like regular CUDA kernels. To time your SGEMV call correctly, do this:

CUT_SAFE_CALL(cutCreateTimer(&timer));

CUT_SAFE_CALL(cutStartTimer(timer));

cublasSgemv('T', N, N, 1.0, dev_A, N, dev_x, 1, 0, dev_v, 1);

cudaThreadSynchronize();

CUT_SAFE_CALL(cutStopTimer(timer));

What you think is the running time of SSDOT is mostly the running time of SGEMV.

3

oh, I see, many thanks for pointing out that!