Bugs/limitations revealed using Nvidia OCL SDK samples!

[rtfss1gmail_com]

Hi,

I was testing Intel OCL using Nvidia OCL SDK samples and have

found some failures I want to report so they can be fixed for next beta..

Anyway i think plus to gaining full khronos standard compliance

you should also test Nvidia OCL SDK samples as they can reveal

also potential bugs in your implementation..

as nvidia samples create Gpu context first I have done is convert

to cpu contexts using CL_DEVICE_TYPE_CPU..

From now I'm interested in oclMarchingcubes and oclNbody sample..

My findings are:

First oclMarchingcubes failed at creating a 2d image

(clCreateImage2D) in allocateTextures function (really it

allocates 1d texture as wants a 2d texture of dimensions 256x1..

so seems that creating 2d images with height=1 shows some bug..

Also at first I thought it was related due to desired image format

not supported and using clGetSupportedImageFormats I remember to

get an empty list so could you fix that function to report all

supported image formats..

After changing the texture to a 1x256 and fixing indexing in

kernels acordingly I found next bug..

some kernels in some samples use inline keyword for device

functions is also present in DCT8x8 sample also..

this seems to be recognized (and possibly ignored) in both AMD

(for both CPU and GPU backends) and Nvidia..

So I have to remove this for kernels to compile..

After all this changes I check sample still works in Nvidia and

AMD backends but still doesn't work in your backend..

Seems problem is in scan kernels and remark AMD CPU backend works..

It fails in openclScan call from computeIsosurface..

Unhandled exception at 0x774904be in oclMarchingCubes.exe: 0xC0000029: Se ha detectado un destino de desenredo no vlido durante una operacin de desenredo.

call stack shows llvm_backend.dll..

With nbody sample similar I get exactly same problem

Unhandled exception at 0x774904be in oclNbody.exe: 0xC0000029: Se ha detectado un destino de desenredo no vlido durante una operacin de desenredo.

Bugs/limitations revealed using Nvidia OCL SDK samples!
Hi,I was testing Intel OCL using Nvidia OCL SDK samples and have
found some failures I want to report so they can be fixedAnyway i think plus to gaining full khronos standard compliance
you should also test Nvidia OCL SDK samples as they can reveal
also potential bugs in your implementation..as nvidia samples create Gpu context first I have done is convert
to CL_DEVICE_TYPE_CPU..From now I'm interested in oclMarchingcubes and oclNbody sample..My findings are:First oclMarchingcubes failed at creating a 2d image
(clCreateImage2D) in allocateTextures function (really it
allocates 1d texture as wants a 2d texture of dimensions 256x1..so seems that creating 2d images with height=1 shows some bug..Also at first I thought it was related due to desired image format
not supported and using clGetSupportedImageFormats I remember to
get an empty list so could you fix that function to report all
supported image formats..After changing the texture to a 1x256 and fixing indexing inkernels acordingly I found next bug..some kernels in some samples use inline keyword for device
functions is also present in DCT8x8 sample also..this seems to be recognized (and possibly ignored) in both AMD
(for both CPU and GPU backends) and Nvidia..So I have to remove this for kernels to compile..After all this changes I check sample still works in Nvidia and
AMD backends but still doesn't work in your backend..Seems problem is in scan kernels and remark AMD CPU backend works..It fails in openclScan call from computeIsosurface..
Unhandled exception at 0x774904be in oclMarchingCubes.exe: 0xC0000029: Se ha detectado un destino de desenredo no vlido durante una operacin de desenredo.call stack shows llvm_backend.dll..
With nbody sample similar I get exactly same problemUnhandled exception at 0x774904be in oclNbody.exe: 0xC0000029: Se ha detectado un destino de desenredo no vlido durante una operacin de desenredo.

[Shiri_M_Intel]

Hi
Thanks for the detailed report and bug hunting!!

It is very useful.

We will fix these issues in future releases.

Thanks, Shiri