program will crash because of line 1473 in nvmatrix/src/nvmatrix.cu about cuda-convnet2

[GoogleCodeExporter]

This confuses me. cudaTextureObject_t is an unsigned long long, so the 
comparison with zero should be fine. I'll need more details to reproduce this. 
I've never seen it myself. 

Original comment by [email protected] on 4 Aug 2014 at 6:40

[GoogleCodeExporter]

I can't upload snapshot, so list related code as fellow :



1458 cudaTextureObject_t NVMatrix::getTextureObject() {
1459    if (_texObj == 0) {
1460        assert(isContiguous());
 1461        //size_t memFree, memTotal;
1462 
1463        struct cudaResourceDesc resDesc;
1464        memset(&resDesc, 0, sizeof(resDesc));
1465        resDesc.resType = cudaResourceTypeLinear;
1466        resDesc.res.linear.devPtr = getDevData();
 1467        resDesc.res.linear.sizeInBytes = getNumDataBytes();
1468        resDesc.res.linear.desc = cudaCreateChannelDesc(32, 0, 0, 0, 
cudaChannelFormatKindFloat);
1469        struct cudaTextureDesc texDesc;
1470        memset(&texDesc, 0, sizeof(texDesc));
 1471        checkCudaErrors(cudaCreateTextureObject(&_texObj, &resDesc, &texDesc, NULL));
1472    }
1473    assert(_texObj != 0);
1474    return _texObj;
1475 }



_texObj returned by line 1471 is ok if it is zero, but that will make line 1473 
fail. 

Original comment by [email protected] on 5 Aug 2014 at 1:41

[GoogleCodeExporter]

Oh, so you're saying that 0 is a valid value for _texObj that might be set by 
cudaCreateTextureObject. I didn't realize this. I'll have to work around that 
somehow then. Thanks.

Original comment by [email protected] on 11 Aug 2014 at 6:30

[GoogleCodeExporter]

I  think why you using cudaTextureObject_t is because you want to 
utilize readonly cache in GK110. Another way to use the readonly cache 
is using const __restrict__ pointer, such as const float* __restrict__ 
images. that will solve this bug, and resolve the memory amount 
limitation problem of texture and makes code looks better

hope that information will hope.

BTW, I have worked at Baidu Company for six months, my boss is Ren Wu, 
he say you are his friend:).

于 2014/8/12 星期二 2:30, [email protected] 写道:

Original comment by [email protected] on 12 Aug 2014 at 12:48

[GoogleCodeExporter]

Texture memory is (for mysterious reasons) still pretty noticeably faster than 
__restrict__ pointers in the cases where I use it, but I'll keep this in mind, 
thanks. 

Original comment by [email protected] on 12 Aug 2014 at 6:27