Vectorize met 2 failed cases in a CI bot with NVIDIA GPU, but we
couldn't repro with all the GPUs at hand, including NVIDIA GPUs. This PR
introduces GPUAdapterInfo and enables this opt on non-NVIDIA GPUs to
make the bots happy.
No obivous perf gain can be seen if we enable vectorize on NVIDIA.
However, it shows big perf improvement on Intel. On my Gen12 Intel GPU,
mobilenetv2-12 perf was improved from 11.14ms to 7.1ms.
### Description
For Concat operation, the zero-size input tensor shape need to be
preserved and, unlike non-zero tensors, the dims are not constrained to
match other input tensors' dims.
### 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. -->
1. Fix Where operator to handle Boolean input less than 4 bytes.
2. Fix JSEP test harness to use tensor names consistently.
### 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 used in sam-h-decoder-f16.
### 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
This is required to make shape uniforms really work.
### Motivation and Context
The bug was unveiled in a model with multiple Split nodes. The later
nodes would try to reuse a previous pipeline cache, while the old shapes
were hardcoded as constants in cache.
### Description
Add MatMulNBits to support MatMul using 4-bit quantized weights
### 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
This PR 1) adds LeakyRelu activation for fusedConv; 2) makes `vec4<f16>`
value work with `float32` uniforms attributes.
For example:
`clamp(value, vec4<f16>(uniforms.clip_min),
vec4<f16>(uniforms.clip_max)` will throw compilation errors since
`uniforms.clip_min` and `uniforms.clip_min` are `f32` not `f16`. So we
need to change it to `clamp(value, vec4<f16>(f16(uniforms.clip_min)),
vec4<f16>(f16(uniforms.clip_max))`
And above problem was introduced when we make activation attributes as
uniforms instead of constant.
BTW, after adding LeakyRelu, `realesrgan-t256` model can pass.
### Description
This PR expands the graph capture capability to JS EP, which is similar
to #16081. But for JS EP, we don't use the CUDA Graph, instead, we
records all gpu commands and replay them, which removes most of the cpu
overhead to avoid the the situation that gpu waiting for cpu.
mobilenetv2-12 becomes 3.7ms from 6ms on NV 3090 and becomes 3.38ms from
4.58ms on Intel A770.
All limitations are similar with CUDA EP:
1. Models with control-flow ops (i.e. If, Loop and Scan ops) are not
supported.
2. Usage of graph capture is limited to models where-in all ops in the
model can be partitioned to the JS EP or CPU EP and no memory copy
between them.
3. Shapes of inputs/outputs cannot change across inference calls.
4. IObinding is required.
The usage is like below:
Method 1: specify outputs buffers explicitly.
```
const sessionOptions = {
executionProviders: [
{
name: "webgpu",
},
],
enableGraphCapture: true,
};
const session = await ort.InferenceSession.create('./models/mobilenetv2-12.onnx', sessionOptions);
// prepare the inputBuffer/outputBuffer
... ...
const feeds = {
'input': ort.Tensor.fromGpuBuffer(inputBuffer, { dataType: 'float32', dims })
};
const fetches = {
'output': ort.Tensor.fromGpuBuffer(outputBuffer, { dataType: 'float32', dims: [1, 1000] })
};
let results = await session.run(feeds, fetches); // The first run will begin to capture the graph.
// update inputBuffer content
... ...
results = = await session.run(feeds, fetches); // The 2ed run and after will directly call replay to execute the graph.
... ...
session.release();
```
Method 2: Don't specify outputs buffers explicitly. Internally, when
graph capture is enabled, it will set all outputs location to
'gpu-buffer'.
```
const sessionOptions = {
executionProviders: [
{
name: "webgpu",
},
],
enableGraphCapture: true,
};
const session = await ort.InferenceSession.create('./models/mobilenetv2-12.onnx', sessionOptions);
// prepare the inputBuffer
... ...
const feeds = {
'input': ort.Tensor.fromGpuBuffer(inputBuffer, { dataType: 'float32', dims })
};
let results = await session.run(feeds); // The first run will begin to capture the graph.
// update inputBuffer content
... ...
results = = await session.run(feeds); // The 2ed run and after will directly call replay to execute the graph.
... ...
session.release();
### 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
```math
\tanh(x)=\frac{e^x-e^{-x}}{e^x+e^{-x}}=
\left\{
\begin{array}{cc}
-\frac{1-e^{-2\cdot(-x)}}{1+e^{-2\cdot(-x)}}, & x<0 \\
0, & x=0 \\
\frac{1-e^{-2x}}{1+e^{-2x}}, & x>0
\end{array}
\right.
```
### Motivation and Context
On some platforms,
$$\tanh(1000)=\frac{e^{1000}-e^{-1000}}{e^{1000}+e^{-1000}}$$ would
produce NaN instead of 0.999... or 1 (imagine $e^{1000}=\infty$ and
$\frac{\infty}{\infty}$ explodes).
### Description
Added Uniforms to SkipLayerNorm
### 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. -->
Improve performance
---------
Co-authored-by: Yulong Wang <7679871+fs-eire@users.noreply.github.com>
### Description
This op is required in mobilenetv3-small-100. With this PR,
mobilenetv3-small-100 model becomes less than 10 ms from over 100 ms on
ADL.
We submit kernels in a batch (a fixed number 16 is used except for the
last batch) for better performance. However, timestamp query support is
at pass level so we disable the batch execution in profiling mode in
previous implementation. Actually we can have multiple passes in a batch
so that we don't have to disable batch execution, which is the first
enhancement of this PR.
Furthermore, WebGPU has an extension to support timestamp query inside
passes, which isn't supported by all the platforms (e.g., Windows
supports it, while macOS doesn't). This is expected to have lower cost
compared with multiple passes solution. So this PR also introduce this
support when available.
This PR also refactors some implementation related to kernelInfo, and
try to unify the related kernel names.
resize for fp16 has 2 issues: scales are always f32 and roi can be f32
or f16.
scales:
this is fixed.
roi
this is fixed for the case where roi is not passed as optional input
with f16. To fix this it requires a much larger change and I did not
want to risk this short before a release. For all practical purpose
passing roi as input with f16 should be rare and we can fix it in the
near future.
### Description
Change `A / sqrt(B)` to `A * inverseSqrt(B)` in BatchNormalization,
InstanceNormalization, LayerNormalization and SkipLayerNormalization.
### Motivation and Context
For the same reason as the existence of the `inverseSqrt` built-in in
WebGPU spec.
Disable createGroupedConvVectorizeProgramInfo path due to bots failures
on below two cases:
[webgpu]Conv - conv - vectorize group - B
[webgpu]Conv - conv - vectorize group - D
### Description
This PR provides a vectorized algorithm for NHWC GroupedConv to improve
performance.
The aggregate time of GroupedConv in mobilenetv2-12 becomes ~1ms from
~4ms on Intel Alder Lake machine. About 20% improvement for the whole
model.
This resolves the below build errors:
```
lib/wasm/jsep/webgpu/op-resolve-rules.ts:19:23 - error TS2724: '"./ops/instance-norm"' has no exported member named 'parseInstanceNormAttributes'. Did you mean 'InstanceNormAttributes'?
19 import {instanceNorm, parseInstanceNormAttributes} from './ops/instance-norm';
~~~~~~~~~~~~~~~~~~~~~~~~~~~
lib/wasm/jsep/webgpu/op-resolve-rules.ts:19:23 - error TS6133: 'parseInstanceNormAttributes' is declared but its value is never read.
19 import {instanceNorm, parseInstanceNormAttributes} from './ops/instance-norm';
~~~~~~~~~~~~~~~~~~~~~~~~~~~
lib/wasm/jsep/webgpu/op-resolve-rules.ts:20:20 - error TS2305: Module '"./ops/layer-norm"' has no exported member 'parseLayerNormAttributes'.
20 import {layerNorm, parseLayerNormAttributes} from './ops/layer-norm';
~~~~~~~~~~~~~~~~~~~~~~~~
lib/wasm/jsep/webgpu/op-resolve-rules.ts:20:20 - error TS6133: 'parseLayerNormAttributes' is declared but its value is never read.
20 import {layerNorm, parseLayerNormAttributes} from './ops/layer-norm';
```
### 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
Also update the op test suite.
### Motivation and Context
Previously the *total* size in case `Expand - last dim is not divisible
by 4` was a multiple of 4, even though the *last dimension* was not, so
the bug has never been caught.
### 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
The patch fixes a floating point accuracy issue in Resize by preferring
integer indices and integer arithmetic where possible.
### Motivation and Context
Model test `test_resize_upsample_sizes_nearest_floor_align_corners` was
observed to be failing on certain platforms. The root cause is the
inaccurate floating point evaluation of 21 / 7 (2.999... vs 3), which
results in the wrong input element to be indexed (floor(2.999...) vs
floor(3)).
### 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
Previously, shape and strides were added unconditionally even they are
not used. This PR fixes this issue and only adds shape and strides when
they are required.
It's useful when some shapes are not used as uniform if the program
depends on type instead of rank.
### Description
Add trilinear interpolation to Resize and changed activation_params attribute as optional for FuseConv.
### 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. -->
Check whether the min/max inputs are provided and use default values if not provided.
### 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
The changes in this PR includes:
1) Fix f16 errors in InstanceNormalization with NCHW format.
2) Use vec to further optimize the original algorithm.
3) (Removed) Don't do layout conversion for InstanceNormalization for
JSEP since InstanceNormalization itself is suitable for NCHW layout and
has better performance in our current implementation.
Tested on sd-vae-decoder-f16.onnx, it becomes 285 ms from 314 ms. The
aggregate gpu profiling data can be found as below (Note the data is
based change 3).):
Before:
<html>
<body>
<!--StartFragment--><span><span class="ui-provider ef bbg bbh bbi bbj
bbk bbl bbm bbn bbo bbp bbq bbr bbs bbt bbu bbv bbw bbx bby bbz bca bcb
bcc bcd bce bcf bcg bch bci bcj bck bcl bcm bcn" dir="ltr">
Kernel | Time (Ms) | Percentage (%)
-- | -- | --
Conv | 201.55 | 69.56
InstanceNormalization | 42.49 | 14.67
Transpose | 28.95 | 9.99
Mul | 5.69 | 1.96
Add | 3.82 | 1.32
MatMul | 3.27 | 1.13
Sigmoid | 2.24 | 0.77
Resize | 1.16 | 0.40
Softmax | 0.34 | 0.12
Cast | 0.24 | 0.08
Sum | 289.75
<br class="Apple-interchange-newline"><!--EndFragment-->
</body>
</html>
After:
<html>
<body>
<!--StartFragment--><span><span class="ui-provider ef bbg bbh bbi bbj
bbk bbl bbm bbn bbo bbp bbq bbr bbs bbt bbu bbv bbw bbx bby bbz bca bcb
bcc bcd bce bcf bcg bch bci bcj bck bcl bcm bcn" dir="ltr">
Kernel | Time (Ms) | Percentage (%)
-- | -- | --
Conv | 205.44 | 79.43
InstanceNormalization | 18.24 | 7.05
Transpose | 17.64 | 6.82
Mul | 5.69 | 2.20
Add | 3.81 | 1.47
MatMul | 3.56 | 1.38
Sigmoid | 2.24 | 0.86
Resize | 1.19 | 0.46
Softmax | 0.59 | 0.23
Cast | 0.24 | 0.09
Sum | 258.65 |
</span></span><!--EndFragment-->
</body>
</html>
From above table, we can see that two ops time are greatly reduced. One
is InstanceNormalization and the other is Transpose. The reason that the
transpose time is reduced is because each InstanceNormalization is
surrounded with two reshape ops in sd-vae-decoder-f16.onnx. Due to JSEP
is prefer NHWC and InstanceNormalization is layout sensitive op, so two
extra transpose ops are inserted dynamically when executing this model.
After this change, those inserted transpose ops are not needed anymore.
So the overall transpose time is reduced.
