ZDNN_MISALIGNED_TENSOR when invoking NNPA on zOS when using a constants binary file.

[christopherlmunoz]

This is on zOS using the --store-constants-binary-file. This issue is specifically on a zdnn_conv2d call. I can provide a full output if needed.

bash-5.2$ client/bin/modelzoo --lib mnist-8-with-binary.so --file mnist-8.tests --validate -m VERBOSE --fc-parms 0.01,0.158752,9,10

ERROR: invoke_nnpa() (./zdnn.c:288): ZDNN_MISALIGNED_TENSOR: Tensor address is not on 4K-byte boundary

With ZDNN_ENABLE_PRECHECK enabled, the Kernel and Bias tensors are non 4K aligned addresses.

Contents of zdnn_ztensor: kernel
Pre-transformed Descriptor:
Outermost Innermost
Dimensions: 5 5 1 8
Layout: ZDNN_HWCK Format: ZDNN_FORMAT_4DFEATURE
Type: FP16
Transformed Descriptor:
Outermost Innermost
Dimensions: 5 5 1 8
Layout: ZDNN_HWCK Format: ZDNN_FORMAT_4DKERNEL
Type: ZDNN_DLFLOAT16
Buffer Addr: b4df2ffd363083d7 Size: 102400
Transformed: True

Thanks! FYI @tungld

[tungld]

Created a PR #3002 to make debugging easier.

[tehbone]

Can you please paste the mmap code that is used?

[tehbone]

Now that I look at it, the buffer address looks to be out of sorts, leading to the following questions:

1. Is __MAP_64 used? If not, the buffer address is incredibly wrong.
2. I suspect that some offset is used to calculate the tensor's buffer address. Is that correct?

[tungld]

The starting address of the buffer was wrong and it is fixed by this #3002.