topic cblas_sgemm speed is abnormal in Intel® oneAPI Math Kernel Library

cblas_sgemm speed is abnormal

Hou_y_1 — Wed, 03 Jan 2018 02:00:16 GMT

	int a = 169*64;
	int b = 64*1024;
        const int c = 5;
	float* A = new float[169*64];
	float* B = new float[64*1024];
	float* C = new float[169*1024];
	srand(time(NULL));
	for (int i=0;i<a;i++)
	{
		A = rand()%1000/100.0;

		if (i%c==0)
		{
			A = -4.204e-045;
		}
	}
	for (int j=0;j<b;j++)
	{
		B = rand()%10000/1000.0;
	}
	while (true)
	{
		double t0 = cvGetTickCount();
		cblas_sgemm(CblasRowMajor, CblasNoTrans, CblasNoTrans, 169, 1024, 64, 1.0, A, 64, B, 1024, .0, C, 1024);
		double t1 = cvGetTickCount()-t0;
		cout<<"consume time:"<<t1/cvGetTickFrequency()/1000.0<<endl;
	}

excute code above, change constant c, the consume time is different. I guess the running time will be slower when the metrix contains denorimalized value. why?

Floating-point operations on

Gennady_F_Intel — Thu, 04 Jan 2018 04:27:44 GMT

Floating-point operations on denormals are slower than on normalized operands because denormal operands and results are usually handled through a software assist mechanism rather than directly in hardware. This software processing causes Intel MKL functions that consume denormals to run slower than with normalized floating-point numbers.

Hi,

Zhen_Z_Intel — Fri, 05 Jan 2018 03:04:00 GMT

Hi,

Denormal number calculation will be slow. You may can use Intel C/C++ compiler with /Qftz option flush to zero, and the perf of MKL sgemm would be improved. Or you can modify your source code to process all denormal to a normal number, such as numeric_limits<float>::min().

Best regards,
Fiona