Adds QNN EP HTP shared memory allocator.
The HTP shared memory allocator (`HtpSharedMemoryAllocator`) calls the
rpcmem shared library (libcdsprpc.so/dll) to allocate and free memory
that can be shared between HTP and CPU.
The allocator can be enabled by setting QNN EP option
`enable_htp_shared_memory_allocator` to `1`.
`QNNExecutionProvider::CreatePreferredAllocators()` will then return an
instance of `HtpSharedMemoryAllocator`.
For each QNN context, we also need to register and unregister memory
handles in order to use the HTP shared memory. This memory handle
management is added to `QnnBackendManager`, which also manages the QNN
context handles.
For more information about using HTP shared memory with QNN, see:
https://docs.qualcomm.com/bundle/publicresource/topics/80-63442-50/htp_shared_buffer_tutorial.html#shared-buffer-tutorial
Limitations:
- HTP shared memory usage is only supported for graph inputs and
outputs. Intermediate values are not supported.
- An allocation is assigned to a single shared memory buffer. The
allocator is not smart enough to have multiple allocations share a
single shared memory buffer.
Co-authored-by: Baiju Meswani <bmeswani@microsoft.com>
### Description
part of https://github.com/microsoft/onnxruntime/issues/21448
This change is intend to save CPU memory during model load for
inference.
Added session option save_prepacked_constant_initializers, with
save_prepacked_constant_initializers turn on:
1. optimize model with inference session, prepacked external initializer
will be saved into data file.
2. load optimized model and external data file with prepacked
initializer, no prepack is needed
3. run inference with optimized model and data file
Tested with model Phi-3-mini-instruct-onnx,
with ORT 1.12.0:
with this change:
Peak memory usage dropped from **5.438 GB to 2.726GB**.
This change takes advantage of ORT loads external initializer with mmap
on CPU. Prepack will use extra memory on heap, omit prepack process can
save this part of memory (roughly same size as external initializers).
next step:
Change all the kernels on CPU with PrePack method implemented and test
properly. Will do in next PR.
### Motivation and Context
<!-- - Why is this change required? What problem does it solve?
- If it fixes an open issue, please link to the issue here. -->
### Description
Updates the ROCm EP opsets to match the current CUDA EP opsets. Also
enable the test CApiTest.basic_cuda_graph_with_annotation.
Note that some changes are whitespace-only. These changes were made to
improve the comparison of corresponding ROCm and CUDA EP source files
when using a side by side diff tool.
### Motivation and Context
The ROCm EP derives from the CUDA EP. Many source files are shared
between the EPs and "hipified" during the ROCm EP build, however quite a
few files within the ROCm EP are under source control after their
initial hipification. Over time these ROCm EP files get stale relative
to their CUDA EP counterparts. It becomes necessary to re-hipify these
otherwise static files in order to pick up important changes such as
opset differences.
### Description
<!-- Describe your changes. -->
### Motivation and Context
<!-- - Why is this change required? What problem does it solve?
- If it fixes an open issue, please link to the issue here. -->
### Description
It might be easier if we just directly include the original gsl headers.
"core/common/gsl.h" is an indirection that doesn't provide extra help.
### Description
- 4-bit QuantizeLinear(21). **Blocked quantization still missing (i.e.,
do not support the new `block_size` attribute)**
- 4-bit DequantizeLinear(21). **Blocked dequantization still missing
(i.e., do not support the new `block_size` attribute)**
- 4-bit Transpose(21).
- Update quantization tool with int4 types.
- Disable QDQ fusions for 4-bit types. See:
https://github.com/microsoft/onnxruntime/blob/main/onnxruntime/core/optimizer/qdq_transformer/selectors_actions/qdq_selector_action_transformer.cc
- MLAS 4-bit quantization kernels for intel, neon, powerpc.
##### Notes
To calculate a tensor's storage size, we normally get the number of
elements from the shape (i.e., `tensor_shape.Size()`) and multiply by
the size of a single element. This does not directly work for sub-byte
elements like int4 as each element in a `Tensor<Int4x2>` stores **two**
packed int4 elements in a byte. The `Tensor::
CalculateTensorStorageSize` should be called to perform the correct
calculation for any tensor element type.
### Motivation and Context
ONNX 1.16 added the int4 and uint4 types. This initial PR adds the int4
type to ORT and adds int4 implementations for the Quant, Dequant, and
Transpose ops on CPU EP. We still need to add int4 support for many ops
and execution providers. See the ONNX 1.16 release notes:
https://github.com/onnx/onnx/releases.
The order of defines for these test have to be in the same order. If we
check for TRT -> CUDA ->DML wen cannot reverse that order in later
defines as we might want to build for multiple EPs.
+@PatriceVignola
This adds a new "Graph Capture" option to the DML ep, similar to the
cuda graph functionality. Here's how graph capture works:
- A user can enable graph capture in the session options by setting
`ep.dml.enable_graph_capture` to `true`
- When they want to capture a run, they set `gpu_graph_id` in their
`RunOptions` to a number bigger than 0 (0 is reserved for internal use
according to the cuda graph documentation).
- Then, when they start the inference, the graph will be captured and
stored in the DML EP for future use
- When they execute the run for a second time with the same id, the
`ReplayGraph` function in the DML EP will be called instead of executing
the kernels, resulting in very low overhead and avoiding kernel
recompilation.
This feature can give up-to-par or even better performance than
specifying the static dimensions at session creation time, but is also
much more flexible.
### Description
<!-- Describe your changes. -->
Add API function GetAliasMap and ReleaseAliasMap in OrtCustomOp
### Motivation and Context
<!-- - Why is this change required? What problem does it solve?
- If it fixes an open issue, please link to the issue here. -->
Add API function GetAliasMap and ReleaseAliasMap in OrtCustomOp
### Description
<!-- Describe your changes. -->
use OrtCustomOp's new API GetMayInplace in CreateKernelCreateInfo to
hook the inplace map of custom ops
### Motivation and Context
<!-- - Why is this change required? What problem does it solve?
- If it fixes an open issue, please link to the issue here. -->
This PR is to use OrtCustomOp's new API GetMayInplace in
CreateKernelCreateInfo to hook the inplace map of custom ops
### Description
Expose Reserve() in OrtAllocator to allow custom allocators to work when
session.use_device_allocator_for_initializers is specified.
Update: this change has been verified by Bing Ads and brings a
significant benefit in terms of memory utilization: 30GB less memory and
also better CPU utilization.
### Motivation and Context
https://microsoft-my.sharepoint.com/:w:/p/prs/Eeidf5YNtWtKrPVkfuTDsuABak1oL4QRpuBGuhqRbLKoJg?e=Zl3bah
### Description
<!-- Describe your changes. -->
This change addresses the following issues with the current CustomOP
Output Type inference
- The function does not take into account optional inputs. When input is
absent the inference is silently aborted, and no output type is inferred
(P1 customer issue)
- Inferring output type based on the input type for multi-kernel custom
ops is done based on the latest in sequence kernel definition. There is
not an attempt made to match the kernel based on the input type.
- Inference is aborted when variadic inputs/outputs are detected when
the generated input/output names fail to obtain type constraints. This
is not immediately clear from the code, because custom op schema is not
available within the inference function.
- No error reporting.
### Motivation and Context
<!-- - Why is this change required? What problem does it solve?
- If it fixes an open issue, please link to the issue here. -->
Most of CustomOPs lack their own type and shape inference function as it
was recently introduced. For that reason, it is important to fix this.
This change is inspired by a customer issue.
This is a follow up on:
- https://github.com/microsoft/onnxruntime/pull/15184
- https://github.com/cbourjau/ort-custom-op/pull/11
- https://github.com/microsoft/onnxruntime-extensions/issues/451
### Description
<!-- Describe your changes. -->
Add 2 C API for ORT extension:
- KernelInfo_GetAllocator
- OrtCustomOp::GetMayInplace
### Motivation and Context
<!-- - Why is this change required? What problem does it solve?
- If it fixes an open issue, please link to the issue here. -->
Add 2 C API for ORT extension project, which will leverage these 2 APIs
for GroupQueryAttention custom op.
### Description
<!-- Describe your changes. -->
1. add a config key in run_options to control cuda graph in runtime.
2. enhance cuda graph class to support mutiple graph saving and
retrieving in one ORT session
3. provide model modification/inference example on Phi2
4. benchmark shows an average of 13% latency reduction in token
generation.
limitation: TRT ep and ROCM ep hasn't applied this feature. we can
revisit this in the future.
### Motivation and Context
<!-- - Why is this change required? What problem does it solve?
- If it fixes an open issue, please link to the issue here. -->
### Description
1. Add two build jobs for enabling Address Sanitizer in CI. One for
Windows CPU, One for Linux CPU.
2. Set default compiler flags/linker flags in build.py for normal
Windows/Linux/MacOS build. This can help control compiler flags in a
more centralized way.
3. All Windows binaries in our official packages will be built with
"/PROFILE" flag. Symbols of onnxruntime.dll can be found at [Microsoft
public symbol
server](https://learn.microsoft.com/en-us/windows-hardware/drivers/debugger/microsoft-public-symbols).
Limitations:
1. On Linux Address Sanitizer ignores RPATH settings in ELF binaries.
Therefore once Address Sanitizer is enabled, before running tests we
need to manually set LD_LIBRARY_PATH properly otherwise
libonnxruntime.so may not be able to find custom ops and shared EPs.
4. On Linux we also need to set LD_PRELOAD before running some tests(if
the main executable, like python, is not built with address sanitizer.
On Windows we do not need to.
5. On Windows before running python tests we should manually copy
address sanitizer DLL to the onnxruntime/capi directory, because python
3.8 and above has enabled "Safe DLL Search Mode" that wouldn't use the
information provided by PATH env.
6. On Linux Address Sanitizer found a lot of memory leaks from our
python binding code. Therefore right now we cannot enable Address
Sanitizer when building ONNX Runtime with python binding.
7. Address Sanitizer itself uses a lot of memory address space and
delays memory deallocations, which is easy to cause OOM issues in 32-bit
applications. We cannot run all the tests in onnxruntime_test_all in
32-bit mode with Address Sanitizer due to this reason. However, we still
can run individual tests in such a way. We just cannot run all of them
in one process.
### Motivation and Context
To catch memory issues.
### Description
onnxruntime may raise an error "type inference failed" but when a custom
operator sets IsHomogeneous to false in its schema. This change make
sure that TypeInferenceFunction and schema type constraints are aligned
to prevent that from happening.
---------
Co-authored-by: Xavier Dupre <xadupre@microsoft.com@orttrainingdev9.d32nl1ml4oruzj4qz3bqlggovf.px.internal.cloudapp.net>
Co-authored-by: Scott McKay <Scott.McKay@microsoft.com>
When the model has "shape tensor" as one of the inputs and user provides
explicit profile shapes for it, TRT EP doesn't correctly set the "shape
tensor" input.
Also, there is a bug for applying explicit profile shapes for the shape
tensor input.
Note: It seems the model has shape tensor input is a rare case. Most of
the cases, the inputs are all execution tensor.
### Description
Validate outputs type and shapes. Make sure sparse initializers are
taken into account.
### Motivation and Context
ORT currently does not validate output types or shapes. Further, neither
inputs or outputs take into account sparse initializers that are
converted from dense.
It is currently possible to pre-allocate a wrong type/shape buffer for
output.
Cc: @Craigacp
Add a generic `UpdateCUDAProviderOptionsWithValue()` C API to update
CUDA EP provider options where its data type is pointer that can't be
represented by string.
Note: Please see some comments for the similar [PR
](https://github.com/microsoft/onnxruntime/pull/16965)for TRT EP.
Add a generic `UpdateTensorRTProviderOptionsWithValue()` C API to update
TensorRT provider options where its data type is pointer that can't be
represented by string.
### Description
Introduce `Float16/BFloat16` support for C# and C++ APIs.
User should be able to perform conversions from `float` to/from
`Float16/BFloat16`, compare values and tests for `NaN, Inifnity, and
whether the number is denormalized.`
### Motivation and Context
User filed issues such as:
https://github.com/microsoft/onnxruntime/issues/14303
### Description
C API for custom ops does not support float 8 types. This PR changes
that.
### Motivation and Context
The list of operators supporting float 8 is very limited. It should be
extended to custom ops to let developpers add customized operators for
these specific types.
CUDA EP already supports [CUDA
graph](https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#cuda-graphs),
also we observed some models can benefit from using CUDA graph with
`trtexec`. Therefore, this PR enables the CUDA graph support for TRT EP.
The implementation is based on
https://github.com/microsoft/onnxruntime/pull/9978 with the same
[constraints](https://github.com/microsoft/onnxruntime/pull/9978) as
below:
- Models with control-flow ops (i.e. If, Loop and Scan ops) are not
supported.
- Usage of CUDA Graphs is limited to models where-in all the model ops
(graph nodes) can be partitioned to the TRT EP.
- The input/output types of models need to be tensors.
- Shapes of inputs/outputs cannot change across inference calls.
- IObinding is required.
### Description
This PR is to refactor ExecutionProvider API for memory management,
which is to move allocators from EP level to SessionState level and
indexed by OrtDevice
### Motivation and Context
<!-- - Why is this change required? What problem does it solve?
- If it fixes an open issue, please link to the issue here. -->
This PR is to refactor ExecutionProvider API for memory management,
which is to move allocators from EP level to SessionState level and
indexed by OrtDevice. By this change, EP level will shift the burden of
maintaining allocators, which will be user friendly for EP developers
---------
Co-authored-by: Lei Cao <leca@microsoft.com@orttrainingdev8.d32nl1ml4oruzj4qz3bqlggovf.px.internal.cloudapp.net>
Add a configuration `max_power_of_two_extend_bytes ` to limit the arena extension size.
### Motivation and Context
<!-- - Why is this change required? What problem does it solve?
- If it fixes an open issue, please link to the issue here. -->
In our real scenario, we observe that if the model is big enough the
BfcArena will extend uncontrollable.
As showed by the following figures, if a model uses more than 16GB
memory, the BfcArena will totally apply for 32GB memory according to the
`kNextPowerOfTwo` strategy. With the new strategy, the extension is
limited. The default maximum extension size is 1GB.
#### Without the new configuration
After loading the model, ORT uses 32G GPU memory.
#### With the new configuration
After loading the model, ORT uses 23G GPU memory.
Co-authored-by: Yuhong Guo <yuhong.gyh@antgroup.com>
### Description
Fix an error in test/shared_lib/test_inference.cc. It should use
ASSERT_NEAR to test float values.
### Motivation and Context
Our OpenVino pipeline is failing because of this.
### Description
This PR partially reverts changes introduced in
https://github.com/microsoft/onnxruntime/pull/15643
We make two API return std::string always in UTF-8.
We also move the entry points from OrtApiBase to OrtApi to make them
versioned.
### Motivation and Context
`GetVersionString` always returns x.y.z numbers that are not subject to
internationalization.
`GetBuildInfoString` can hold international chars, but UTF-8 should be
fine to contain those.
We prefix them with u8"" in case the compiler default charset is not
UTF-8.
Furthermore, creating platform dependent APIs is discouraged.
`ORTCHAR_T` is platform dependent and was created for paths only.
On non-unix platforms would still produce `std::string` that can only
contain UTF-8
The API was introduced after the latest release, and can still be
adjusted.
Implement a set of new APIs for lightweight custom ops registration, to
save efforts from schema-composing.
A few highlights:
- Support build-time type inference;
- Support function-as-op for "stateless" ops;
- Support structure-as-op for "stateful" ops;
- Support varied input/output forms such as span, scalar, and tensors,
either optional or non-optional.
---------
Co-authored-by: Randy Shuai <rashuai@microsoft.com>
Implement a set of new APIs for lightweight custom ops registration, to
save efforts on schema-composing.
A few highlights:
1. Support build-time type inference;
2. Support function-as-op for "stateless" ops;
3. Support structure-as-op for "stateful" ops;
4. Support varied input/output forms such as span, scalar, and tensors,
either optional or non-optional.
---------
Co-authored-by: Randy Shuai <rashuai@microsoft.com>
The PR is to allow custom op of different input types to have same op
name in a graph.
The idea to go over all ops of same name and merge their input/output
types into a type-inference function.
With the enhancement, custom op node inside a graph can have same
op-type given that the input/output types are different.
---------
Co-authored-by: Randy Shuai <rashuai@microsoft.com>
### Description
Reduce a number of auxillary objects created to reduce GC pressure.
Eliminate GCHandle type of memory pinning in most of the places.
Improve string marshalling by allocating unmanaged memory that does not
require pinning. Change native methods from `IntPtr` to `byte[]`
(marshalling pinning is more efficient).
Allocate input/output UTF-8 names in unmanaged heap for the lifetime of
InferenceSession. So we do not keep converting them and pinning on every
Run.
Introduce a new native API that allows to allocate and convert/copy
strings directly into a native tensor.
The PR delivers around 50% latency improvements and less GC pauses.
Inspired by: https://github.com/microsoft/onnxruntime/pull/15520
### Motivation and Context
Client experience GC pressure and performance degradation when dealing
with string tensors.
Co-Authored-By: @tannergooding
### Description
Run clang-format in CI. Formatted all c/c++, objective-c/c++ files.
Excluded
```
'onnxruntime/core/mlas/**',
'onnxruntime/contrib_ops/cuda/bert/tensorrt_fused_multihead_attention/**',
```
because they contain assembly or is data heavy
### Motivation and Context
Coding style consistency
