[BUG] cutlass conv2d kernel failed when N*H*W*C*2 larger than 0x100000000 [4G]

[leiwen83]

Describe the bug

I try tune the conv2d kernel whose input feature larger than 4G, and then get failure.
Seems to me there seems need some change for this support?

Steps/Code to reproduce bug
Follow this guide http://matthewrocklin.com/blog/work/2018/02/28/minimal-bug-reports to craft a minimal bug report. This helps us reproduce the issue you're having and resolve the issue more quickly.

./tools/profiler/cutlass_profiler --n=1 --h=5120 --w=8192 --c=32 --k=8 --r=3 --s=3 --pad_h=1 --pad_w=1 --stride_h=1 --stride_w=1 --dilation_h=1 --dilation_w=1

[hwu36]

we don't support that. to support it, you need to change all the strides to 64bits.

[leiwen83]

we don't support that. to support it, you need to change all the strides to 64bits.

I see...
For all strides to 64bits, do you have any suggested file/struture need to be modified?
Currently Stride is using Index as int32_t in include/cutlass/tensor_ref.h, so if make Index as int64_t, could this issue be fixed?

[hwu36]

you need to start from https://github.com/NVIDIA/cutlass/blob/main/include/cutlass/layout/tensor.h#L67. change this one to using Stride = Coord<kStrideRank, LongIndex >;, delete this one https://github.com/NVIDIA/cutlass/blob/main/include/cutlass/layout/tensor.h#L116 . Then fix any warnings/errors complaining int64->int32 conversion.

[leiwen83]

Got it, thx