### 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. -->
`Module.jsepRegisterMLConstant` will be shorten by Closure Compiler in
offical release, this would cause undefined error.
Fix it by using `Module['jsepRegisterMLConstant']`.
### 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 PR make MatMul shaders not depend on inputs broadcasting pattern,
but only depend on input ranks and their shape provided in uniform. This
change fix the issue that currently shaders code are different for
different broadcasting, but have identical cache key and results in
wrong cache hit.
This CL make WebGPU backend support subgroup features and thus allow
using subgroup optimizations in the future.
### Description
With this CL WebGPU backends will create devices with subgroups and
subgroups-f16 features (both are under origin trial in Chrome) or
chromium-experimental-subgroups feature enabled whenever available.
### Motivation and Context
This CL would allow WebGPU operator shaders to use subgroup
optimizations in the future, and might get some significant speedup with
these optimization.
This PR fixes a bug that occurs when searching for compatible `MLTensor`
in the cache. We were missing checking the number of dimensions in the
shape. This would mean that a cached buffer of shape `[1]` could match
for `[1, 1, 256, 256]`.
This PR also adds better handling when attempting to force an `MLTensor`
to a different shape.
In current implementation, all the staging buffers for weights uploading
are destroyed after first batch of kernel execution. It requires a lot
of memory as all the staging buffers couldn't be reused. It also hurts
the startup time (weights uploading only happens in session creation),
as weights uploading is delayed to a very late time.
This PR uses a very aggressive way to submit queue and destroy staging
buffers, so that the related GPU memory could be reused as much as
possible, though the real situation depends on the WebGPU and driver
implementation. The aggressive queue submission also moves GPU
operations to a very early time, which helps the startup time.
Some buffer uploading benchmarks are composed to compare multiple
solutions, regarding to the memory and time consumption. Benchmarks can
be found at
https://github.com/webatintel/webbench/blob/master/webgpu/buffer-upload.html,
while detailed test data can be found at
https://docs.google.com/document/d/1KgygOkb9ZNzkgzQ_tWOGlEI9ScmMBHDjDojjPFLmVXU/edit.
I also tested phi3.5 on 2 machines, first inference time improved from
5141ms to 3579ms and from 4327ms to 2947ms separately.
#22031
For reduce related ops, we should increase workgroupSize to improve
parallelism if only one workgroup is dispatched.
The total ReduceMean time becomes 8.98 ms from 77.79 ms on my iGPUs.
BUG #22031
The total Gemm time in demucs model becomes 181.14 ms from over 1000 ms
on my iGPUs.
### 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. -->
BUG #22031
In the demucs model, there are lots of MatMul ops with shapes like
below:
`input[0]: [3448,1,512] | float32, input[1]: [512,1536] | float32,
output[0]: [3448,1,1536] | float32`
We can see that for this kind of shape, the batch size is a big value,
but M = 1. Our current algorithm is based on [M, N] to partition tiles,
which is not efficient for such kind of shapes. This PR reshapes the
inputs to improve the matmul performance.
Before: [3448,1,512] x [512,1536] = [3448,1,1536]
After: [1, 3448, 512] x [512, 1536] = [1, 3448, 1536] , then the output
can be reshaped to [3448, 1, 1536]
The overall MatMul time in demucs model becomes 1778.45 ms from 4418.17
ms on my iGPUs.
---------
Co-authored-by: Yulong Wang <7679871+fs-eire@users.noreply.github.com>
### Description
This change adds a cache of `MLContext`s keyed by their options to the
`WebNNBackend`. This makes is so that multiple `InferenceSession`s
create with the same options will share the same context.
### Motivation and Context
Since `MLTensor`s are tied `MLContext`s, developer can't easily share
tensors between `InferenceSession` (outside of manually an `MLContext`
and specifying the `context` options). This leads strange behaviors such
as,
```js
const sessionsA = ort.InferenceSession.create(urlA, {
executionProviders: ["webnn"],
preferredOutputLocation: "ml-buffer",
});
const sessionsB = ort.InferenceSession.create(urlB, {
executionProviders: ["webnn"],
});
const temp = await sessionA.run({/* arguments */});
const result = await sessionB.run({"input":temp["output"]}); // ERROR: Failed to execute 'dispatch' on 'MLContext': Invalid inputs: The context of MLGraph doesn't match the context of the MLTensor with name "input".
```
We encountered this behavior when updating the transformers.js version
in the developer preview demos. microsoft/webnn-developer-preview#46
BUG #22031
Optimize below two situations:
1. Increase workgroupSize if only one workgroup is dispatched.
2. Avoid transpose if not necessary.
The overall time of demucs model becomes 106.36 ms from 154.60 ms on my
dGPUs with this PR and PR #22577
### Description
This PR introduces support for registering external data inside WebNN
EP.
### Motivation and Context
- The WebNN EP needs to register the initializers at graph compilation
stage, for initializers from external data, it can't leverage the
general external data loader framework because the graph compilation of
WebNN EP is executed before external data loader called.
- Exposes the `utils::GetExternalDataInfo`, it is useful for WebNN EP to
read the external tensor's infomation.
- Define a new `registerMLConstant` in JSEP to create WebNN constants
from external data in WebNN backend, with the info of tensor as
parameters, as well as the `Module.MountedFiles`, which holds all
preloaded external files.
### Description
This change enables caching `MLTensor`s between inferences runs. This is
done by keeping a reference to `MLTensor`s alive after they have been
released. `MLTensor`s are only destroyed once the sessions goes out of
scope.
### Motivation and Context
Creating and destroying `MTensor`s on every run has a non-trivial
performance penalty. This performance penalty materializes when using
`ort.Tensors`[location=cpu] for inputs/outputs or when using the CPU EP
as a fallback EP for unsupported operators. The former could be
mitigated by developer using `ort.Tensors`[location=ml-tensor]. The
latter cannot be mitigated by developers.
### Description
<!-- Describe your changes. -->
This PR further optimizes matmulnbits specially for iGPUs. The phi3 demo
becomes ~12 tokens/second from ~8 tokens on iGPUs.
Some todos:
1. Make the optimization more general, Remove the blockSize = 32
limitation.
2. Tune the parameter, such as workgroupSize, components size (currently
only support components = 1), to see the performance change.
### Description
<!-- Describe your changes. -->
With this optimization, 96 MultiHeadAttention|Transpose ops in phi3
disappear. Phi3 becomes 113 tokens from 107 tokens on my dGPUs.
The optimization mainly skips the transpose op if one of the transposed
dims is 1. Reshape is enough.
In current implementation, axis in softmax has to be the last, which is
an obvious limitation. This PR removes this limitation and will fix
issues #20710 and #22176.
### Description
Enables using the MLTensor to pass data between models.
### Motivation and Context
Using MLTensor instead of ArrayBuffers reduces the number of copies
between the CPU and devices as well as the renderer and GPU process in
Chromium.
### Description
<!-- Describe your changes. -->
For InstanceNormalization, it has `y = scale * (x - mean) /
sqrt(variance + epsilon) + B` , where mean and variance are computed per
instance per channel. Calculating mean and variance per channel is a
reduce processing, which is NCHW layout friendly since it makes the
adjacent threads can access contiguous data in gpu memory.
This PR optimizes both NHWC and NCHW InstanceNormalization. To
efficiently calculate the mean and variance, we need to make sure the
input is NCHW instead of NHWC. Then use shared memory to do the reduce
operation to get `channel_scale` and `channel_shift`.
With this PR, getting `channel_scale` and `channel_shift` are same for
NHWC and NCHW InstanceNormalization. And the overall performance becomes
very close now.
Below data comes from SD Turbo profiling results.
Before (InstanceNormalization overall time: 140.84 ms)
InstanceNormalization\|InstanceNormComputeMean | 129.70
-- | --
InstanceNormalization\|InstanceNormalizationNHWC | 10.55
InstanceNormalization\|InstanceNormComputeChannelScaleShift | 0.59
After (InstanceNormalization overall time: 59.44 ms)
InstanceNormalization\|InstanceNormComputeChannelScaleShift | 28.57
-- | --
InstanceNormalization\|TransposeShared | 20.19
InstanceNormalization\|InstanceNormalizationNHWC | 10.68
Perf test data(100000 times)
Array: 12.599999997764826ms
String: 1.6000000014901161ms
Perf test case:
```
const permFunctionBodyArray = (rank: number, input: string): string => {
const reverseFunc = [];
reverseFunc.push(`fn perm(i: int) -> int {
var a: int};`);
for (let i = 0; i < rank; ++i) {
reverseFunc.push(input);
}
reverseFunc.push('return a;}');
return reverseFunc.join('\n');
};
const permFunctionBodyString = (rank: number, input: string): string => {
let reverseFunc= `fn perm(i: int}) -> int {
var a: int;`;
for (let i = 0; i < rank; ++i) {
reverseFunc+=input;
}
reverseFunc+='return a;}';
return reverseFunc;//.join('\n');
};
const count = 100000;
let start, end
console.time('array');
start = performance.now();
for(let i =0 ; i < count; i ++) {
permFunctionBodyArray(3, 'input');
}
end = performance.now();
console.timeEnd('array');
console.log("Array: "+ (end-start));
console.time('string');
start = performance.now();
for(let i =0 ; i < count; i ++) {
permFunctionBodyString(3, 'input');
}
end = performance.now();
console.log("String: " +(end-start));
console.timeEnd('string');
```
### 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 is to fix issue #22031 to run model demucs.
For conv-transpose, outputPadding.length could be 1, while spatialRank
is 2. The fix is to append enough 0s to outputPadding. For conv, the
issue is similar. kernelShape.length sometimes could be 1, while
inputs[1].dims.length is 4. The fix is also to append enough 0s to
kernelShape.
### Description
<!-- Describe your changes. -->
#21618
This PR optimizes grouped conv by 1) more sequential memory access in
gpu 2) reusing input's data to reduce global memory access times.
See `Conv|GroupedConv` op in
[Wav2Vec2](https://huggingface.co/facebook/wav2vec2-base-960h) becomes
92 ms from 1058 ms on iGPUs with 32 EU.
For the whole model on my iGPUs with 32 EU,
wav2vec2 model becomes 982ms from 1942 ms.
squeezebert-uncased model becomes 71.86ms from 431.77ms.
### 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. -->
Fix bugs in previous implementation and add more situations to go the
optimized path.
Below situations will go to the optimized path.
1. 2d inputs or squeezed 2d inputs
2. channels last or channels first transpose. For example, channel last
transpose: [1, 256, 512, 512] -> [1, 512, 512, 256]
For this case, the transpose becomes [256, 512x512] -> [512x512, 256]
### 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. -->
For SD Turbo demo, the total transpose time becomes 39.98ms from
122.09ms. And the correspnding percents becomes 3.89% from 11.05% in
this demo.
This PR will also help #21618, the total transpose time in that demo
becomes 17.32 ms from 70.25 ms on my iGPUs.
### 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. -->
### 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. -->
---------
Co-authored-by: Yulong Wang <7679871+fs-eire@users.noreply.github.com>
### Description
<!-- Describe your changes. -->
See 2x speedup for phi3 on the integrated intel gpu with this
optimization.
The optimization is mainly to store input A's data into local variable
instead of loading them from global memory each time when calculate them
with B data.
### 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. -->
Avoid producing presentKey/presentValue outputs if pastKey/pastValue
don't exists.
### 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. -->
### 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. -->
This PR fixes the `AttentionProbsSoftmax` recompilation issue when
executing the phi3 model. With this fix, it will further improve the
phi3 performance.
### 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
Previously, MultiHeadAttention supports relative position bias of shape
[1, N, S, T] or [B, N, S, T], and DecoderMaskedMultiHeadAttention
supports [1, N, S, T]. This will extend the support to allow [1, N, S,
T], [B, N, S, T], [B, 1, S, T] and [1, 1, S, T] for CUDA and CPU EPs.
- [x] Rename the input of "relative position bias" to "attention bias"
because it can also be used for other types of bias, like ALiBi
(Attention with Linear Biases) or attention mask.
- [x] Update unfused kernel to support broadcasting 2nd dimension of
attention bias.
- [x] Update efficient attention to support broadcasting 2nd dimension
of attention bias.
- [x] Update operators (MultiHeadAttention,
DecoderMaskedMultiHeadAttention, Attention, PackedAttention,
PackedMultiHeadAttention) to support broadcast attention bias on CUDA
and CPU EPs.
- [x] Update ROCm, DML and WebGPU naming to be consistent. (Note that
those EPs do not support broadcasting attention_bias for now).
- [x] Add attention bias tests for MultiHeadAttention.
- [x] Update operator documents
- [x] Update benchmark script
Other changes:
* Fix some checks in multihead-attention.ts
* Add helper functions to dump tensors given dimensions.
### Description
See
454996d496
for manual changes (excluded auto-generated formatting changes)
### Why
Because the toolsets for old clang-format is out-of-date. This reduces
the development efficiency.
- The NPM package `clang-format` is already in maintenance mode. not
updated since 2 years ago.
- The VSCode extension for clang-format is not maintained for a while,
and a recent Node.js security update made it not working at all in
Windows.
No one in community seems interested in fixing those.
Choose Prettier as it is the most popular TS/JS formatter.
### How to merge
It's easy to break the build:
- Be careful of any new commits on main not included in this PR.
- Be careful that after this PR is merged, other PRs that already passed
CI can merge.
So, make sure there is no new commits before merging this one, and
invalidate js PRs that already passed CI, force them to merge to latest.
