pow function not compiling in CUDA kernel

Accelerated Computing CUDA CUDA Programming and Performance
1

hello, I use the pow() function twice throughout a global function, and the first causes a problem but it doesn’t complain about the second.

Just by taking a look, can anyone see what might be wrong?

works:
t2 = 24.0pow(10.0, 6.953-4.09*log10(envreal_.zwtemp[*iz][*iy][*ix]+26.0));

doesn’t work:
xweight[*i] = pow(xlength[*i]/kwt2lena, 1.0/kwt2lenb);

I get the message:
cuda.cu(166): error: calling a host function from a device/global function is only allowed in device emulation mode

thank you

2

are you compiling with support for double precision? if not, it’s probably picking up the host’s pow(double, double) instead of the device’s pow(float, float)

3

Hi,

I had a similar problem, and I do compile with “nvcc -arch=sm_13”.

It seems that CUDA simply doesn’t have declarations of pow(double,float) and pow(float,double), while it DOES HAVE pow(float,float) and pow(double, double). It can be easily seen here:

[codebox]

    double dA=2., dB=3.;

float	fA=2., fB=3.;

    double rdd,rff,rdf,rfd;

rdd=pow(dA,dB);

rff=pow(fA,fB);

rfd=pow(fA,dB);

rdf=pow(dA,fB);[/codebox]

Here the lines with rdd and rff compile ok, but the rfd and rdf give out the error " error: calling a host function from a device/global function is only allowed in device emulation mode".

A simple fix is to explicitly cast both arguments either to float or to double.

Cheers!