Executing CUDA or PTX Code on the GPU

Hello,

I have a question regarding to CUDA code on the GPU.

I use Matlab 2011a parallel tool kit trial version, VS 2008 and CUDA 3.2.

I just tried to compile the sample CUDA code from parallel tool kit tutorial. (Following is the CUDA kernel code)

File name: test.cu global void add1( double * pi, double c ) { *pi += c; }

I just tried to compile by using following command as described in tutorial.

nvcc -ptx test.cu

but it showed me following error message.

nvcc fatal : Visual Studio configuration file 'vsvars32.bat' could not be foun d for installation at './../../..'

I could fix this error message by using following command

nvcc -ccbin "C:\Program Files (x86)\Mic rosoft Visual Studio 9.0\VC\bin" -I"C:\Program Files (x86)\NVIDIA GPU Computing Toolkit\CUDA\v3.2\include" -ptx test.cu

Then, I could get the test.ptx file and

I just tried to excute this file by using following command.

 k=parallel.gpu.CUDAKernel('test.ptx', 'test.cu');
But, it is not runnable. I see this error message.

Could you tell me what should I do to run this code?

??? Error using ==> iCheckPTXEntryAgainstCProto at 384 The number of inputs to the PTX code (2) is NOT the same as the number of inputs in the C prototype (0)

Error in ==> C:\Program Files\MATLAB\R2011a\toolbox\distcomp\gpu\+parallel\+internal\+gpu\handleKernelArgs.p>handleKernelArgs at 81

Answer 1

[Edric Ellis]

I'm not quite sure why that's failing, but you may be able to work around that by constructing the kernel like so:

k = parallel.gpu.CUDAKernel( 'test.ptx', 'double *, double' )

(The formatting of your "test.cu" is hasn't come out right above - we attempt to parse the prototype from the .cu code, and perhaps something is confusing our parsing...)

[Darryl]

I tried to attemp as you said, but it shows following error message.

??? Error using ==> iCheckPTXEntryAgainstCProto at 411

Expected pointer type but found .u32

Error in ==> C:\Program

Files\MATLAB\R2011a\toolbox\distcomp\gpu\+parallel\+internal\+gpu\handleKernelArgs.p>handleKernelArgs at 81

Do you think this is because I am using a trial version?

[Edric Ellis]

It should have nothing to do with a trial version. Also, if you're running on a 64-bit platform, you might need to compile the PTX in 64-bit mode by adding "-m 64" to your nvcc command-line

[Darryl]

Thanks!!! -m 64 was the right method to solve this program!

Answer 2

[Gaszton]

"??? Error using ==> iCheckPTXEntryAgainstCProto at 384 The number of inputs to the PTX code (2) is NOT the same as the number of inputs in the C prototype (0)"

i got this error when my .ptx file was not recomplied after modifying the .cu source file

Other tip: I suggest you, to modify the nvcc profile file, located here: ...\NVIDIA GPU Computing Toolkit\CUDA\v3.2\bin\

put this extras in it:

   #
   # Location of Microsoft Visual Studio compiler
   #

   compiler-bindir = d:\Programs\Microsoft Visual Studio 9.0\VC\bin

   include-path = "d:\Programs\Microsoft Visual Studio 9.0\VC\include"

(Copy paste the path of your VStudio folder)

I dont know why, but the "include-path" has no effect, so you still have to specify the VS include directory to nvcc.

I you want to avoid typing it in the cmd, create a .bat batch file:

   @ECHO OFF
   cmd /k nvcc -I "D:\programs\Microsoft Visual Studio 9.0\VC\include" -gencode=arch=compute_20,code=\"sm_21,compute_20\" -ptx yourcudacode.cu

[Darryl]

I checked the compiling. I recompiled if there is any modification in the code. I tried as what you said.

I don't know why, but it stil keep showing the error message.

??? Error using ==> iCheckPTXEntryAgainstCProto at 384

The number of inputs to the PTX code (2) is NOT the same as the number of inputs in the C prototype (0)

Error in ==> C:\Program

Files\MATLAB\R2011a\toolbox\distcomp\gpu\+parallel\+internal\+gpu\handleKernelArgs.p>handleKernelArgs at 81

Answer 3

[King Fisher]

I have exactly the same problem :(