* Switch to CUDA10.2
* Update win-gpu-tensorrt-ci-pipeline.yml
* Update win-gpu-tensorrt-ci-pipeline.yml
* remove dynamic_shape
* update onnx-tensorrt submodule
* check if input shape is specified for TensorRT subgraph input and enable some TensorRT unit tests
* fix format issue
* add shape inference instruction for TensorRT
* update according to the reviews
* Update win-gpu-tensorrt-ci-pipeline.yml
* WIP: Re-enable x86 .NET testing in Release pipelines
Enabling x86 testing will make sure that ORT packages doesn’t break x86 projects of customers
* Remove setting some env variables
* Comment out a test failing on x86 builds
* More changes
* Minor fix
* More changes
* More changes
* s
* s
* s
* Revert minor change
* More changes
* More changes
* More changes 2
* explicitly set platform target
* Delete bin and obj folders
* Clean output dirs
* Add back TargetFramwork
* Disable x86 .net framework tests
* Skip x86 tests in MKLML pipeline
* Don't create a copy of model proto when checking to see if there is fp16 input
* PRcomments about making functions const
* Loop through nodeargs in graph object to see if there are fp16 datatypes
* Rename check to checking only inputs
* port the mimalloc allocator
* hook mimalloc opt into common.h and reduction ops
* repurpose USE_MIMALLOC to only denote subbing in of default allocator with mimalloc and some refactoring
* fix unintended cherry pick diffs
* polish alloctor_mimalloc
* explicitly disable mimalloc where it already had been disabled
* update mimalloc to pull in stl allocator
* switch mimalloc stl allocator to use mimalloc library version
* turn mimalloc on by default (only the stl changes are enabled, the python interacting ones are off already and shall remain so)
* move FastAllocVector into cpu specific code
* separate out defines into arena and stl changes
* the rest of the define renames
* bfc arena allocator
* some typos and rename the bfc arena allocator to fit existing class naming conventions
* adjustments in response to comments
* different template instantiations are friends
1. Add LTCG back. It was set to default OFF in my previous PR to speed up Windows build. It is only needed in release pipelines.
2. Remove --use_featurizers from all the packaging pipelines
3. Make sure all the packages have openmp
Use CUDA 10.1 for Linux build
(Windows change is already in)
Please note, cublas 10.2.1.243 is for CUDA SDK 10.1.243, not CUDA 10.2.x. CUDA 10.2.89 need cublas 10.2.2.89. They match on the last part of the digits.
libcublas10-10.1.0.105 won't work!!!
The cuda docker image by viswamy is already using 10.1, no need to change.
* GPT2 Gelu Fusion & Test
* change header path
* Refine code & add missing test onnx file
* Fix builds & refine float/double/fp16 compare.
* Fix builds
* Add Bias Check and UTs
* Fix build and uts
* Fuse with second formula & test
* minor change
* disable FastGelu to see whether the builds can pass
* Verify where is wrong
* disable for debugging
* Revert "disable for debugging"
This reverts commit 535c0817fb36fb95a75773a7f00c8b969dd5362c.
* Revert "Verify where is wrong"
This reverts commit ffc43ec1d136636ba2cee30df49f563a75e84676.
* disable the transformer for inference currently
* Enable FastGeluFusion and fix segement fault when run bertsquad10.onnx test
* Add more Unit tests convering Gelu subgraph use graph input/output
(cherry picked from commit 0739ab985240c6d9acdb8f0afd40c5fb316166af)
* Mode Bias Fusion in BiasGelu.cc
Co-authored-by: Changming Sun <chasun@microsoft.com>
Add support to fuse ReorderOutput+Transpose(NHWC). Converting from NCHWc to NHWC tensors is a trivial copy of data and avoids the cost of a transpose node.
This fixes a customer reported issue where the NCHWc optimizer was dropping graph outputs when an edge was used as both a graph output and an input to another NCHWc node.
