cuDNN: NHWC format not supported unless manually specified?

Accelerated Computing GPU-Accelerated Libraries
1

I noticed that calling cudnnSetTensor4dDescriptor with the CUDNN_TENSOR_NHWC format returns 9 (“not supported”). However,

int ns = c*w*h;
int cs = 1;
int hs = c*w;
int ws = c;

status = cudnnSetTensor4dDescriptorEx(d, dataType, n, c, h, w, ns, cs, hs, ws);

which should be equivalent, returns “success”. Why the discrepancy, and is the NHWC format supported at all?

2

if you try to use this tensor to do a convolution, the convolution routine will return that it is not supported as well.
WE plan to support this format in the next release

3

Cannot you tell, when is planned the next release?

The support of OpenCV-formatted (NHWC) images may be critically important.

4

Bumping this topic up, because I am currently having the problem that my data is in NHWC and according to the preliminary documentation of v3 (next version) says that NHWC will not be supported for backprop.

5

May I also ask: Is there any library function that I can use to convert between NHWC and NCHW on the GPU? Many thanks in advance.

~t3l