mirror of
https://github.com/saymrwulf/onnxruntime.git
synced 2026-05-14 20:48:00 +00:00
4f2096be38
### Description
<!-- Describe your changes. -->
Update XNNPACK to latest version
- adds fp16 kernels and various other improvements
- requires pthreadpool update as well
Most code updates in the XNNPACK EP are to adjust to the new XNNPACK API
- 'setup' is split into 'reshape' and 'setup'
- some ops use a workspace buffer
- copied workspace allocation from XNNPACK unit test code
- some suffixes changed
Added wrapper for XNNPACK caches to base XNNPACK EP kernel
- simplifies usage
- XNNPACK split out the code and weights caches, but the code cache
isn't currently usable via the public API
- we could use the internal types if we think it's required for
performance reasons. non-trivial though as we'd need to propagate ifdef
values from the XNNPACK build up to the ORT build.
- using XNNPACK internals would also mean we would not be able to
support using a pre-build XNNPACK package
- not an issue currently
Fixed opset registration for internal NHWC domain
- was not being tied to the ONNX version, so nodes inserted by layout
transformation had the incorrect opset
- a number of other places needed updating once this issue was fixed
Remove support for NCHW Resize from XNNPACK EP so it's NHWC only
- we only supported NCHW for fp32,
- doing so adds complexity in multiple places (XNNPACK EP kernel
implementation, layout transformation and transpose optimization)
- unclear if that complexity provides any benefit. can add back if
required by production scenario
### 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. -->
We're looking at enabling fp16 support for CoreML and NNAPI. If we do
that we need a good fallback story if the CPU EP will be used. The
XNNPACK fp16 kernels will hopefully provide that.
NOTE: This PR doesn't add fp16 support to the XNNPACK EP kernels. That
can be done as required in separate EPs and should be relatively simple
to do.
4.3 KiB
4.3 KiB
Operators Support Table
The following table shows ONNX
operators and the supported opset domain/versions in WebGPU EP by ONNX Runtime Web. For example,
4-6, 8+ means ONNX Runtime Web currently support opset version 4 to 6, 8 and above.
This file is automatically generated from the def files via this script. Do not modify directly.
| Operator | Opset | Comments |
|---|---|---|
| Abs | ai.onnx(6-12,13+) | |
| Acos | ai.onnx(7+) | |
| Acosh | ai.onnx(9+) | |
| Add | ai.onnx(7-12,13,14+) | |
| ArgMax | ai.onnx(1-10,11-12,13+) | |
| ArgMin | ai.onnx(1-10,11-12,13+) | |
| Asin | ai.onnx(7+) | |
| Asinh | ai.onnx(9+) | |
| Atan | ai.onnx(7+) | |
| Atanh | ai.onnx(9+) | |
| AveragePool | ai.onnx(7-9,10,11+); com.ms.internal.nhwc(7-9,10,11+) | need perf optimization; need implementing activation |
| BiasAdd | com.microsoft(1+) | |
| BiasSplitGelu | com.microsoft(1+) | |
| Cast | ai.onnx(6-8,9-12,13-18,19+) | |
| Ceil | ai.onnx(6-12,13+) | |
| Clip | ai.onnx(6-10,11,12,13+) | |
| Concat | ai.onnx(1-3,4-10,11-12,13+) | |
| Conv | ai.onnx(1-10,11+); com.ms.internal.nhwc(1-10,11+) | need perf optimization; conv3d is not supported; need implementing activation |
| ConvTranspose | ai.onnx(1-10,11+); com.ms.internal.nhwc(1-10,11+) | need perf optimization; ConvTranspose3d is not supported; need implementing activation |
| Cos | ai.onnx(7+) | |
| Cosh | ai.onnx(9+) | |
| Div | ai.onnx(7-12,13,14+) | |
| Einsum | ai.onnx(12+) | |
| Elu | ai.onnx(6+) | |
| Equal | ai.onnx(7-10,11-12,13-18,19+) | |
| Erf | ai.onnx(9-12,13+) | |
| Exp | ai.onnx(6-12,13+) | |
| Expand | ai.onnx(8-12,13+) | |
| Flatten | ai.onnx(1-8,9-10,11-12,13+) | |
| Floor | ai.onnx(6-12,13+) | |
| FusedConv | com.microsoft(1+) | |
| Gather | ai.onnx(1-10,11-12,13+) | |
| GatherElements | ai.onnx(11-12,13+) | |
| Gelu | com.microsoft(1+) | |
| Gemm | ai.onnx(7-8,9-10,11-12,13+) | |
| GlobalAveragePool | ai.onnx(1+); com.ms.internal.nhwc(1+) | |
| GlobalMaxPool | ai.onnx(1+); com.ms.internal.nhwc(1+) | |
| Greater | ai.onnx(7-8,9-12,13+) | |
| GreaterOrEqual | ai.onnx(12-15,16+) | |
| If | ai.onnx(1-10,11-12,13-18,19+) | |
| InstanceNormalization | ai.onnx(6+); com.ms.internal.nhwc(6+) | |
| LayerNormalization | ai.onnx(17+) | |
| LeakyRelu | ai.onnx(6-15,16+) | |
| Less | ai.onnx(7-8,9-12,13+) | |
| LessOrEqual | ai.onnx(12-15,16+) | |
| Log | ai.onnx(6-12,13+) | |
| MatMul | ai.onnx(1-12,13+) | |
| MaxPool | ai.onnx(1-7,8-9,10,11,12+); com.ms.internal.nhwc(1-7,8-9,10,11,12+) | need perf optimization; need implementing activation |
| MemcpyFromHost | ai.onnx(1+) | |
| MemcpyToHost | ai.onnx(1+) | |
| Mul | ai.onnx(7-12,13,14+) | |
| Neg | ai.onnx(6-12,13+) | |
| Not | ai.onnx(1+) | |
| Pad | ai.onnx(2-10,11-12,13-17,18,19+) | |
| Pow | ai.onnx(7-11,12,13-14,15+) | |
| Range | ai.onnx(11+) | |
| Reciprocal | ai.onnx(6-12,13+) | |
| ReduceL1 | ai.onnx(1-10,11-12,13-17,18+) | |
| ReduceL2 | ai.onnx(1-10,11-12,13-17,18+) | |
| ReduceLogSum | ai.onnx(1-10,11-12,13-17,18+) | |
| ReduceLogSumExp | ai.onnx(1-10,11-12,13-17,18+) | |
| ReduceMax | ai.onnx(1-10,11,12,13-17,18+) | |
| ReduceMean | ai.onnx(1-10,11-12,13-17,18+) | |
| ReduceMin | ai.onnx(1-10,11,12,13-17,18+) | |
| ReduceProd | ai.onnx(1-10,11-12,13-17,18+) | |
| ReduceSum | ai.onnx(1-10,11-12,13+) | |
| ReduceSumSquare | ai.onnx(1-10,11-12,13-17,18+) | |
| Relu | ai.onnx(6-12,13,14+) | |
| Reshape | ai.onnx(5-12,13,14+) | no GPU kernel |
| Resize | ai.onnx(10,11-12,13-17,18,19+); com.ms.internal.nhwc(10,11-12,13-17,18,19+) | CoordinateTransformMode align_corners is not supported with downsampling |
| Shape | ai.onnx(1-12,13-14,15+) | no GPU kernel; an ORT warning is generated - need to fix |
| Sigmoid | ai.onnx(6-12,13+) | |
| Sin | ai.onnx(7+) | |
| Sinh | ai.onnx(9+) | |
| SkipLayerNormalization | com.microsoft(1+) | |
| Slice | ai.onnx(1-9,10,11-12,13+) | |
| Softmax | ai.onnx(1-10,11-12,13+) | |
| Split | ai.onnx(1,2-10,11-12,13-17,18+) | |
| Sqrt | ai.onnx(6-12,13+) | |
| Squeeze | ai.onnx(1-10,11-12,13+) | |
| Sub | ai.onnx(7-12,13,14+) | |
| Tan | ai.onnx(7+) | |
| Tanh | ai.onnx(6-12,13+) | |
| ThresholdedRelu | ai.onnx(10+) | |
| Tile | ai.onnx(6-12,13+) | |
| Transpose | ai.onnx(1-12,13+) | need perf optimization |
| Unsqueeze | ai.onnx(1-10,11-12,13+) | |
| Where | ai.onnx(9-15,16+) |