MIOpen
MIOpen copied to clipboard
ROCmSoftwarePlatform
AMD's Machine Intelligence Library
- Add [HingeEmbeddingLoss](https://pytorch.org/docs/stable/generated/torch.nn.HingeEmbeddingLoss.html) API wrapped by MIOPEN_BETA_API, solver, kernel, driver and gtest for fp32, fp16 and bfp16 - Performance speedup compared to ROCm/pytorch: |type |Forward| |--------|-------| |float32 | 2.24| (Note...
* Added AdaptiveAvgPool [1D](https://pytorch.org/docs/stable/generated/torch.nn.AdaptiveAvgPool1d.html) [2D](https://pytorch.org/docs/stable/generated/torch.nn.AdaptiveAvgPool2d.html) [3D](https://pytorch.org/docs/stable/generated/torch.nn.AdaptiveAvgPool3d.html) forward and backward. * Added driver test and gtest for AdaptiveAvgPool . * New API is guarded by MIOPEN_BETA_API macro. * Average over all...
- Add [NormalizeBackward](https://pytorch.org/docs/stable/generated/torch.nn.functional.normalize.html) API wrapped by MIOPEN_BETA_API - Add solver and kernel for cases: contiguous tensor, reduce dimension is last dimension, reduce size % 256 = 0 (256 is chosen...
- cget tool supports -G parameter to specify generator. [Reference](https://cget.readthedocs.io/en/latest/src/commands.html#install) - This code change provides capability to leverage that support and let user pick a generator with install_deps.cmake, limiting to...
This PR removes solvers: - `ConvMPBidirectWinograd` - `ConvMPBidirectWinograd` - `ConvMPBidirectWinograd` - `ConvMPBidirectWinograd` - `ConvMPBidirectWinograd` - `ConvMPBidirectWinograd_xdlops` - `ConvMPBidirectWinograd_xdlops` - `ConvMPBidirectWinograd_xdlops` - `ConvMPBidirectWinograd_xdlops` - `ConvMPBidirectWinograd_xdlops` These solvers were added in https://github.com/ROCm/MIOpen/pull/358...
* Added Interpolate forward and backward with multi-mode operation and kernel. * Added driver test and gtest for [Interpolate ](https://pytorch.org/docs/stable/generated/torch.ao.nn.quantized.functional.interpolate.html). * New API is guarded by MIOPEN_BETA_API macro. * Average...
Add ImageTransform kernels: - `ImageAdjustBrightness` (`ImageAdjustBrightnessFwdContiguous`) - `ImageAdjustHue` (Including `ImageAdjustHue` and `ImageAdjustHuecontiguous`) - `ImageAdjustSaturation` (`RGBToGrayscale`, `RGBToGrayscaleContiguous`, `BlendContiguous`) - `ImageNormalize` (`ImageNormalizeFwdContiguous`) Detailed performance results are [here](https://docs.google.com/spreadsheets/d/1DBncEW_Df_RLpA2gDhEHJna1SCkRc6peRdw4Yh4m4Bs/edit?usp=sharing) Average performance: ### ImageAdjustBrightness ||contiguous|...
- Add [MultilabelSoftMarginLoss](https://pytorch.org/docs/stable/generated/torch.nn.MultiLabelSoftMarginLoss.html) forward operation and kernel. Backward is not better compared to ROCm in general. - Given input tensor is (N,C), MIOpen is better than ROCm if C
This PR adds unit tests for issue https://github.com/ROCm/MIOpen/issues/3540
- Add Softmax operation with forward and backward kernels. - Add driver and gtest for kernels. ### Average improvement over ROCm | type | fwd | bwd | |:--------:|:----:|:----:| |...