* Dump subgraph ID and fused graph ID
Dump subgraph ID and fused graph ID for better debugging
* Remove local static fused_count
added a field global_fused_count_ to NupharExecutionProvider class
* Correctly handle implicit inputs for fused nodes
Previously, nuphar's partitioning function didn't include
node's implicit inputs into the inputs list of MetaDef, and hence
a crash was triggered in the onnx graph checker.
This commit fixed the issue. Furthermore, it also fixed a related
issue where we didn't add implicit inputs into
graph_inputs_excluding_initializers_ in Graph::SetGraphInputsOutputs.
the issue was that graph_inputs_including_initializers_ populated by
SetInputs (e.g. called by FunctionImpl::FunctionImpl) may contain
implicit inputs which were not of any node's initializers in the graph.
Because they were not part of any initializers, these implicit inputs
couldn't be visited by going through all nodes' inputs.
Consequently, they would *not* be added into graph_inputs_excluding_initializers_.
We fixed the issue by first copying the populated graph_inputs_including_initializers_
into graph_inputs_excluding_initalizers_, which then had both initializers and
non-initializers as its initial content. Later, we erase initializers from the
list. In this way, we can ensure all implicit inputs to remain in
graph_inputs_excluding_initializers_.
* refined comments and fixed duplicates
Address CR by revisiting comments in terms of implicit inputs
Also fixed an issue by skipping duplicates while copying inputs
from graph_inputs_including_initializers_.
* address CR
explain why we need to collect nodes' implicit inputs
* don't rely on pointer values for iterating std::set
Previously, openvino relied on iterating a set of NodeArg pointers
to construct inputs and outputs for a fused graph. It could cause
non-determinism. The reason was that although iterating std::set by
itself is stable, pointer values of NodeArgs may vary. Consequently,
we could end up visiting the set's elements in different orders for
different runs for the same test, which resulted in constructing
inputs (and outputs) with different orders to the fused graph.
For example, for the same test, we may have inputs [A, B] in some
runs but inputs[B, A] in others.
Let's use std::string as the key type to avoid such nondeterminism.
This commit also added implicit inputs into meta->inputs while returning
the capability from the openvino provider.
* Fixed another latent issue in openvino's GetCapability function
The issue was that we couldn't simply erase fused_inputs and fused_outputs
while iterating the nodes. For example, an output NodeArg may have multiple
uses, and it's wrong if we erase it from fused_outputs when we encounter only
one of its uses as input.
* [Nuphar EP] performance improvements
1. Add new ops: Shape, Expand
2. Add support for steps in Slice
3. Simplify Gather
4. Always inline alias nodes
5. Transpose nodes with inner loop being symbolic falls back to CPU provider when vectorization is not possible
6. Add opt_inproj option to model_editor to extract MatMuls inside Scan for input projection to outside
* Added GatherElements to Nuphar
This change added GatherElements (op_ver 11) to the Nuphar provider.
* address CR feedback
* create a utilify function for accessing index safely
* address more CR
* SafeIndex -> ClampIndex
* Added Scatter and ScatterElements to Nuphar
Implemented Scatter (op_ver 9 - 10) and ScatterElements (op_ver 11)
nuphar.
Because TVM's compute is output-oriented, our current implementation
uses extern calls for simplicity.
* fixed build issue after rebase
* remove dead code
* Address CR
* removed dead code
* use GetAttrOrDefault
* Address more CR feedback
* add GetStrides to codegen/common/utils.h
* added a unit test for Bool input data
Remove gsl subodule and replace with a local copy of gsl-lite
Refactor for onnxruntime::make_unique
gsl::span size and index are now size_t
Remove lambda auto argument type detection.
Remove constexpr from fail_fast in gsl due to Linux not being happy.
Comment out std::stream support due to MacOS std lib broken.
Move make_unique into include/core/common so it is accessible for server builds.
Relax requirements for onnxruntime/test/providers/cpu/ml/write_scores_test.cc
due to x86 build.
Add ONNXRUNTIME_ROOT to Server Lib includes so gsl is recognized
Enhance proto3 compatibility.
Replace has_*() method to corresponding enum handling so we can deal with
proto3 generated stream from proto2 code.
Add utility wrappers for remaining has_*() methods so we can
easily deal with them if/when we switch to proto3.
* Implement Nuphar execution provider
Nuphar execution provider is a TVM-based compilation provider. It has shown great speedups for RNN models using Scan.
This PR is mainly for a preview of the shared codegen library for other TVM-based providers.
* Fix submodules
* Fix TVM submodule
* Update Nuphar to latest and resolve confliction
* Remove stale files caused by merge -X theirs
* Revert heap buffer change to not introduce onnxruntime_framework into onnxruntime_perf_test
* Fix bad merge
* Merge from Nuphar
* Fix warning treated as error, revert some unnecessary changes
* Revert some more test changes
* Some more test revert or comments to make review easier
New tests could be added later
* One more revert of unnecessary changes
* More change revert. Test could be added back later.
Description:
This change adds the common part of TVM based codegen library. It includes following parts:
* Microsoft TVM Inventory (MTI): a set of TVM ops for neural networks, similar to TOPI
* Compiler pass for traversing ONNX graph and generate TVM ops
* Compiler pass for traversing generated graph and specify TVM schedule
* Compiler pass for handling weight layout
* Utils for debugging
Motivation and Context:
TVM is an open deep learning compiler stack for cpu, gpu and specialized accelerators. To leverage it in ONNX, we built an execution provider named Nuphar. Currently, Nuphar gets good performance on CPUs with AVX2 on quantized LSTM models.
This codegen library was part of Nuphar execution provider. It is split out for sharing with other execution providers, as we'd like to reuse TVM in more devices.
