### Description
Remove `Resolve()` on the entire graph as each function is resolved.
We retain `Resolve()` after each inlining iteration.
### Motivation and Context
Poor performance for inlining the model and session initialization.
Original model before Resolve() removal
FunctionTest.Profiling (**65953 ms**)
After Resolve() Removal
FunctionTest.Profiling (**2911 ms**)
RelWithDebInfo pre-inlined model. Presumably because it runs Level1
optimizers
Non-inlined model consists of functions and Level1 optimizers have no
effect.
FunctionTest.Profiling (**9851 ms**)
To avoid a huge cu file and make code more readable:
- Move PrepareQKV to separate cu file (attention_prepare_qkv.cu)
- Move ConcatPastToPresent to attention_concat.cu
- Add default value for AttentionData
- Add a data structure QkvData to track Q, K and V pointers and track
QKV format.
- [x] Optimize SDXL models exported by optimum.
- [x] Enable it to run locally instead of using module.
- [x] Detect external data file in original model, and save with same
format by default.
- [x] Add tests
### Example
```
pip install optimum transformers diffusers onnx onnxruntime-gpu>=1.16
optimum-cli export onnx --model stabilityai/stable-diffusion-xl-base-1.0 --task stable-diffusion-xl ./sd_xl_base_onnx
python -m onnxruntime.transformers.models.stable_diffusion.optimize_pipeline -i ./sd_xl_base_onnx -o ./sd_xl_base_fp16 --float16
```
### Known issues
(1) VAE decoder cannot be converted to float16. Otherwise, there will be
black image in output.
(2) To use the float16 models, need a minor change in optimum to convert
the inputs for VAE decoder from float16 to float32 since we keep VAE
decoder as float32. The change is to append a line like the following
after [this
line](afd2b5a366/optimum/pipelines/diffusers/pipeline_stable_diffusion_xl.py (L483))
```
latents = latents.astype(np.float32)
```
### Description
WebNN CPU backend expects slope of PRelu to be a static value. For now,
we will not support it.
### Motivation and Context
Fallback this case to pass the CI.
### Description
The files should not have the minor version number. The names were added
in #17365 by mistake.
### Motivation and Context
We did not successfully exclude them out.
### Description
<!-- Describe your changes. -->
This adds a missing member initialization.
### 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. -->
It caused an access violation in
`Dml::GraphDescBuilder::BuildGraphDesc`.
### Description
<!-- Describe your changes. -->
In previous implementation, there are two loops to iterate H * W
elements to calculate the `mean` and `squaredNorm` value in one thread,
meanwhile it outputs H * W elements in one thread. That results it's
very very slow when H * W is a large value. And usually, H * W does be a
large value in a model. For example, in the `candy-8` model, the shapes
of [H, W] are [224,224], [112,112], [56,56] for `InstanceNormalization`
op. And in my ADL, `[1,224,224,32]` consumes 17 ms. See below:
```
[profiling] kernel "23848328|[InstanceNormalization] 23848328" input[0]: [1,224,224,32] | float32, input[1]: [32] | float32, input[2]: [32] | float32, output[0]: [1,224,224,32] | float32, execution time: 17007914 ns
```
In this PR, it uses workgroup memory to optimize the original algorithm.
The advantage is that it can parallelly utilize the 64 (workgroupSize)
threads in one workgroup to calculate `mean` and `squaredNorm` value.
Meanwhile, it only outputs `H * W / workgroupSize` outputs for one
thread, which greatly reduces the overhead for one thread. With this
optimization, `[1,224,224,32]` becomes 3 ms and the main overhead is the
extra two `transpose`. The `createInstanceNormProgramInfo` only needs
`0.64` ms. See below:
```
[profiling] kernel "23003600|[InstanceNormalization] 23003600" input[0]: [1,224,224,32] | float32, output[0]: [1,32,224,224] | float32, execution time: 1543792 ns
program-manager.ts:115
[profiling] kernel "23003600|[InstanceNormalization] 23003600" input[0]: [1,32,224,224] | float32, input[1]: [32] | float32, input[2]: [32] | float32, output[0]: [1,32,224,224] | float32, execution time: 642652 ns
program-manager.ts:115
[profiling] kernel "23003600|[InstanceNormalization] 23003600" input[0]: [1,32,224,224] | float32, output[0]: [1,224,224,32] | float32, execution time: 991608 ns
```
This PR currently only applies the new algorithm to NCHW format. For
NHWC format, one way is to transpose the input so that it can use the
new algorithm. But the disadvantage is that 2 extra transpose are added.
@dakenf also gives another way to optimize NHWC. Details see
[here](d45a96616d/js/web/lib/wasm/jsep/webgpu/ops/instance-norm.ts).
I checked @dakenf's method. The perf is similar with transpose +
optimized NCHW. But on different GPUs, one is a little better than
another or vice versa. So I prefer this PR only does the NCHW part.
@dakenf can submit his optimization on NHWC.
### 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
<!-- Describe your changes. -->
As title.
iOS language code uses different syntax for specifying language
code/region code:
https://developer.apple.com/documentation/xcode/choosing-localization-regions-and-scripts
current `default_locale` is not working for iOS.
### 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. -->
Issue:
https://github.com/microsoft/onnxruntime/issues/17017
---------
Co-authored-by: rachguo <rachguo@rachguos-Mini.attlocal.net>
Co-authored-by: Edward Chen <18449977+edgchen1@users.noreply.github.com>
### Description
The new cpuinfo library doesn't use clog on Android. Newer XNNPack
versions have removed the dependency on clog, but the one we use still
has it. So I cherry-pick the XNNPack to our patch file.
### Description
<!-- Describe your changes. -->
For some use case need to create boolean tensor.
I've tested on [this
project](https://github.com/hans00/react-native-transformers-example)
### 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 handle `ONNX_TENSOR_ELEMENT_DATA_TYPE_BOOL`
And it required #15556 (It seems not include in latest release
(v1.15.1))
### Description
<!-- Describe your changes. -->
Make status.h independent from gsl.
### 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 the coming new feature external EP API (see the prototype
https://github.com/microsoft/onnxruntime/pull/16718), we need to expose
stream in the public header, however, stream is dependent on status.h
which is dependent on gsl. We are seeking a way to decouple stream from
gsl.
From Changming's comment offline, prefast is disabled so all
GSL_SUPPRESS are not taking any effect now. He will handle the warnings
when enable prefast in the future
### 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
PR 15470 updated some C/C++ dependencies. The change caused ROCM EP's
nightly build to fail. see issue
https://github.com/ROCm-Developer-Tools/HIP/issues/2082 for a
background. So, the root cause is HIP compiler has a special requirement
that HIP's include dirs must be used before the operating system's
include folder: /usr/include. HIP adds "-isystem" in front of
"/usr/include". gcc or clang will search the folders added with "-I"
first, then the "-isystem" folder. It works fine as long as we do not
add "-I/usr/include" to the compile commands for *.cu files. It would be wrong if
we already have installed an open source library to /usr and want to use the
prebuilt library from there instead of the current build dir.
### Motivation and Context
@fdwr This is the part 2 of the pybind work that was started earlier.
This adds the following features to the python IO binding
implementation:
- Use a bucketized allocator in order to reduce the number of resource
allocations
- Implement the following functions: `ortvalue_from_numpy`,
`update_inplace`, `ortvalue_from_shape_and_type` and `numpy`
- Modify the `onnxruntime_test_python_iobinding` tests to also run on
DML
---------
Co-authored-by: Jeff Bloomfield <jeffbloo@microsoft.com>
### Description
supplement of #17417
### 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
update prepack script to use exact version.
the prepack script for onnxruntime-node, onnxruntime-web and
onnxruntime-react-native is used to update their referencing version of
dependency "onnxruntime-common".
Previously "~" (tilde symbol) is used. This may cause NPM choose an
older version (if the old version matches the version requirement and
was previously installed already so hit the cache). see also
https://semver.npmjs.com/. [This
build](https://dev.azure.com/onnxruntime/onnxruntime/_build/results?buildId=1134671&view=results)
is caused by this issue.
### 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. -->
This reverts commit bb136f86c8, then
re-implement it in a different way.
I reverted the original change, then added a version constraint to the
find_package args.
If you still found it picks up wrong gtest version after this change,
you may disable `find_package` by setting
'FETCHCONTENT_TRY_FIND_PACKAGE_MODE' to NEVER. For example, the latest
gtest version is 1.14.0. If at a later time Google releases a new
version of gtest and that one is incompatible with the ONNX Runtime
source code you get today and your dev environment already installed the
new version and you do not want to create a new clean build environment
that is without the package, you can add `--cmake_extra_defines
FETCHCONTENT_TRY_FIND_PACKAGE_MODE=NEVER` to your build command to solve
the problem.
### Description
<!-- Describe your changes. -->
The name of nightly ACPT image has been updated to
`ptebic.azurecr.io/internal/aifx/acpt/nightly-ubuntu-cuda-torch-dev`
As the previous image alias had `cu118`, `torch210dev` or `py38`, any
version update will break the training nightly pipeline
### 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. -->
Using constant image alias to avoid pipeline failure.
### Description
Delete all Prefast tasks because the new VS 17.7 version crashes every
time when we run the task on our CI build servers. However, we cannot
reproduce it locally. And this problem blocks us installing security
patches to our CI build machines.
Will use [CodeQL](https://codeql.github.com/) instead.
### Motivation and Context
Address some security alerts.
### Description
release session after use in npm test.
This is one of the prerequisites for supporting IO binding for WebGPU
buffer in onnxruntime-web.
list of prerequisites PRs:
#17465#17469#17470 (this one)
During optimization of SDXL UNet, the prune_graph takes up to 5 minutes.
The cause is to find a node in all nodes is time-consuming. This
optimization will reduce the latency of prune_graph to 2 seconds.
New algorithm will use a hash table (key is first node output, value is
node) to speed up.
The old provisioning profile no longer works. Switched to a temporary one that we can use before a new one is available. The temporary one has a different name.
Revert to the old TRT EP behavior of securing the whole compute_function
by lock_guard.
Current TRT EP which only puts lock_guard around a critical section
(obvious wrong) inside compute_function.
The issue can happen where one thread is updating the engine in
compute_function whereas another thread still accesses the
stale/corrupted engine instance in compute_function, for example, the
code outside the critical section, `int total_bindings =
trt_engine->getNbBindings()`.
So, make the whole compute_function the critical section should be okay.
### Description
This PR proposes a change that should speed up inference for the
TreeEnsemble* kernels. Previously, when traversing a decision tree, the
`TreeNodeElement` pointer would be incremented or decremented to the
appropriate child node - I assume this was because the
`truenode_inc_or_first_weight` and `falsenode_inc_or_n_weights` member
were overloaded for two purposes.
In this PR, we now assign the true branch pointer. We also initialise
`nodes_` in a pre-order traversal which means that the false branch's
position can be resolved statically and does not need to be stored.
I observe the following speed ups. The benchmarks used are derived from
those in https://github.com/siboehm/lleaves/tree/master/benchmarks and
the baseline is the main branch.
NYC Dataset
--------------
| Number of threads | Baseline | Pointer assignment | Pre-ordered
initialisation | Pointer assignment % improvement | Pre-ordered
initialisation % improvement |
|--------------------:|-----------:|---------------------:|-----------------------------:|-----------------------------------:|-------------------------------------------:|
| 1 | 176.539 | 155.709 | 145.119 | 11.7989 | 17.7976 |
| 4 | 59.9015 | 51.9652 | 50.0884 | 13.2488 | 16.382 |
| 8 | 34.5561 | 31.3024 | 28.2535 | 9.41581 | 18.2387 |
Airline Dataset
---------------
| Number of threads | Baseline | Pointer assignment | Pre-ordered
initialisation | Pointer assignment % improvement | Pre-ordered
initialisation % improvement |
|--------------------:|-----------:|---------------------:|-----------------------------:|-----------------------------------:|-------------------------------------------:|
| 1 | 2127.34 | 1389.7 | 920.373 | 34.6745 | 56.736 |
| 4 | 723.307 | 481.634 | 310.618 | 33.4122 | 57.0558 |
| 8 | 420.722 | 278.397 | 185.265 | 33.8286 | 55.9651 |
mtpl2 Dataset
--------------
| Number of threads | Baseline | Pointer assignment | Pre-ordered
initialisation | Pointer assignment % improvement | Pre-ordered
initialisation % improvement |
|--------------------:|-----------:|---------------------:|-----------------------------:|-----------------------------------:|-------------------------------------------:|
| 1 | 1143.62 | 1020.04 | 998.171 | 10.8055 | 13.0988 |
| 4 | 386.153 | 339.905 | 328.061 | 11.9764 | 14.3729 |
| 8 | 225.995 | 200.665 | 199.057 | 11.2084 | 13.4408 |
These were run using an M2 Pro with 16GB of RAM. All times are in
milliseconds and averages over 10 runs with a batch size of 100,000.
### Motivation and Context
Performance improvements.
### Description
Updates the version of QNN SDK used by CI Pipelines. Enables some tests
fixed by 2.14.1, but still need to look into Resize in a separate PR.
### Motivation and Context
Test latest version of QNN SDK.
A recent change was made in
5a83a67f32
to make `ep_type` a reference instead of having it be a copy, presumably
to avoid assigning strings (so `auto& ep_type =
node->GetExecutionProviderType()` instead of `auto ep_type =
node->GetExecutionProviderType()`). The problem with this change is that
calling `node->SetExecutionProviderType(kCpuExecutionProvider)` will
change the value of the reference itself, which means that it's
impossible to revert the node to its previous EP.
This change fixes this bug and adds an optimization over the previous
approach by only assigning a string when we know that we are dealing
with a non-CPU node.
### Description
Added Einsum operator support to JSEP.
### 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
Update the Web CI pipelines:
- remove parameter 'WebTemplate': Since we start to support webgpu, the
linux-web-ci.yml is no longer working and it is already out-of-date.
remove this file and parameter so that we always use win-web-ci.yml
- change flag `RunWebGpuTests` into 2 flags, for release and debug.
Currently for CI we only run webgpu tests on release build. But we want
to have the capability to run webgpu tests on debug build as well.
After this PR is merged, next step is to enable both Debug and Release
webgpu tests in PostMerge pipeline.
This makes it possible to call `optimize_by_onnxruntime` for float32 unet if `--use_external_data_format` is also used.
### Motivation and Context
When using `optimize_pipeline.py` without `--float16`, `optimize_by_onnxruntime` was not called for unet.
### Description
Add new name "WebGPU_Buffer" to OrtMemoryInfo.
This is one of the prerequisites for supporting IO binding for WebGPU
buffer in onnxruntime-web.
list of prerequisites PRs:
#17465#17469 (this one)
### Description
[Successful pipeline
run](https://dev.azure.com/onnxruntime/onnxruntime/_build/results?buildId=1123141&view=results)
Added flag to build the training artifacts & updated the
pull-wasm-artifacts script to pull the training artifacts as well.
Bundled into this PR are minor formatting fixes + naming fixes.
### Motivation and Context
[This PR](https://github.com/microsoft/onnxruntime/pull/16521) extended
the WASM API wrapper to build training WASM artifacts as well.
The ORT training WASM artifacts are required to support ORT training web
bindings.
### Description
This PR contains a few changes in /js/common/ to support a coming PR for
a full implementation of webgpu IO binding.
- allows pass-through if value is already a Tensor instance in return
value of `handler.run()` called by `InferenceSession.run()`
(inference-session-impl.ts). Specifically, onnxruntime-node and
onnxruntime-react-native uses native bindings to generate a Tensor-like
object so we need to create a real Tensor instance here; for
onnxruntime-web the return value is already a Tensor instance.
- adds new types for GPU buffer supported types: `'float32'|'int32'` ->
`'float32'|'float16'|'int32'|'int64'|'uint32'|'bool'`
- exposes types `GpuBufferDataTypes` together with `CpuPinnedDataTypes`
and `TextureDataTypes` as exported
