* Make CDist faster via Eigen squaredNorma and GEMM.
* Add call to abs() as the GEMM output may differ slightly due to floating point accuracy and result in a negative distance which returns NaN if sqrt() is applied to it.
* Update math::Gemm to use the type for alpha and beta instead of hardcoding to float. Matches the GemmEx definition.
* Provide Eigen based replication of the GEMM call on x86 if T=double.
* Make test model data deterministic.
* Do the GEMM first so we can avoid potentially subtracting two numbers that are very close to each other.
* update C# API to optimize inference latency
* rename PinnedOnnxValue to fixedBufferOnnxValue and fix build break
* add more test cases
* add conditions on string tensors for pre-allocated outputs
* change to random inputs
* fix word spell
* resolve comments
* resolve comments
* remove FixedBufferOnnxValueTests.cs
* fix trivial typos in doc
* wangye/pivot (#3432)
* check in
* work version
* add ForecastingPivot kernel
* fix mac os and linux build error
* update FeaturizerLibrary Version
* resolve comments
* remove changes
* Add Kernel for LagLeadOperator & RollingWindowFeaturizer (#3434)
* update
* update todo
* resolve comments
* relax eps for TruncatedSVD transformer
* mute TruncatedSVD_transformer due to undeterministic test result
* resolve comments
* update
* test
* update
* fix
* Make the Slice implementation based on type sizes and reduce templatized code to a minimum.
* Remove using 'dynamic' as a template param to Slice as well.
The commit 06fc9506fd which refactored cpu Pool class broke ACL EP build.
Also worked on the commit a4fe60c4d3 as it also affects the new class.
Move the declaration of the new MaxPoolV8 cpu class in the header file. Implement MaxPool 8-11 in ACL EP.
Co-authored-by: Andrei-Alexandru <andrei-alexandru.avram@nxp.com>
* Add CPU implementation for FastGelu operator
* Update optimization script to fuse Gelu or FastGelu according to Elf or Tanh is used in graph.
* Merge BiasGelu and FastGelu into one class
* Enable FastGelu Fusion optimizer for CPU Execution Provider.
Add opt_level option for graph optimization level in bert perf test.
Support BERT models that output each layer, where SkipLayerNormalization has more than 4 children.
Check weight and bias are 1D for layer norm fusion.
Add a dummy class Gpt2OnnxModel for further changes of GPT2 model.
An ExternOp's input needs buffers, so we cannot add compute_inline
schedule on it even if it's a scalar tensor. Instead, we need to
schedule it as compute_root.
Made some fixes to enable loss scale to be wired up to ORT from the Python frontend. In particular, now addition of loss scaling is done unconditionally if mixed precision is enabled. The generated loss scale input name is passed back to the frontend.
Also fixed how inputs were added during the training graph configuration. Graph::SetInputs() was causing some issues - it seems to not be working correctly.
Also added some mixed precision Python frontend tests.
Co-authored-by: liqun <liqun@OrtTrainingDev4.af05slrtruoetgaxwwjv5nsq5e.px.internal.cloudapp.net>
