daisytuner / docc / 25506496266

push

github

Target specific expand (#681)

* expands conv node patch-wise

* Fuse paddings into convolutions

- Added new convolution node that has a padding attribute
- Added a MLIR pass that fuses a padding node into a convolution by creating the new convolution node and setting the padding attribute
- Fixed the expand method of the SDFG convolution node to correctly support padding
- Added partial resnet18 nets to the resnet18 benchmark

* adds tiled h-dimensions for conv node

* adds is<> utility method to blocks

* adds tensor elimination

* adds local buffer reuse pass

* Enable fusion of padding into poolings

- Added custom linalg pooling node that supports padding
- Added MLIR pass that fuses tensor.pad into linalg.pooling_nchw_max / linalg.pooling_nchw_sum by creating the custom linalg pooling node
- Added translation to SDFG for custom linalg pooling node
- Fixed expand method of SDFG pooling node to correctly support padding
- Added integration tests that test padding for max-pooling, avg-pooling, and sum-pooling

* Removed memset from convolution expand

* removes obsolete header

* Makes expander structure compile

* Fix formatting

* Add container for new expanded maps

* User analysis fix on ROCm

* Added im2row expansion for ROCm

* Add CUDA batchnorm expansion also for ROCm

---------

Co-authored-by: Lukas Truemper <lukas.truemper@outlook.de>
Co-authored-by: Moritz Timmer <25349452+Moehre2@users.noreply.github.com>

58 of 71 new or added lines in 5 files covered. (81.69%)

32178 of 49015 relevant lines covered (65.65%)

12931.86 hits per line