Releases: ROCm/MIOpen
Releases · ROCm/MIOpen
MIOpen v1.3.0
Notes:
- Performance improvements for RNNs
- Performance improvements for convolutions using 1x1 filters
- Performance improvement for Batch Normalization
- This release adds preliminary fp16 support for Inference using CNNs
- Bug fixes for various components of MIOpen
Changes:
- Added 2 new API for RNNs: miopenGetRNNLayerParamOffset and miopenGetRNNLayerBiasOffset
- Added support for uninitialized hidden states and nullptr outputs in RNNs
- Added support for Set and Scale operations for strided tensors with dimensions 1 to 5
- Added multi-thread and multi-process support for the performance database
- Improved performance for OpTensor
- Fixed bug in convolutions for backward bias
- Fixed logic issues in get and set layer functions and related w_supertensor test
- Fixed hang in batch norm with batch sizes greater than 256
Known Issues:
- RNNs do not support fp16
- Training with CNNs does not support fp16
MIOpen v1.2.1
Notes:
This release adds support for ROCm 1.7.1.
MIOpen v1.2.0
Notes:
- This release adds the support for recurrent neural networks (RNNs) for three flavors - Vanilla, LSTMs, and GRU
- Users can now themselves update the perf-db file, which hosts the tuning parameters for convolutions, by setting appropriate environment variables
Changes:
- Over 50% improvement in ResNet performance since the last release
- Multiple padding modes like Same and Valid added
- Winograd convolution kernels added for strided bwd-data convolutions
- Tensor Ops allow for beta and alpha scaling values and support up to 5 dimensions with strides and offsets
- Tensor Copy supports up to 5 dimesnional copies with strides and offsets
- Unit-tests for LRN are added
- Several bug fixes for all the layers of the library
Known issues:
- RNNs may give incorrect result due to a known compiler bug; issue may particulary arise during some RNNs configs with GEMM of size power of 4
- Potential issue where OpenCL resources will be exhausted for large RNN
MIOpen v.1.1.4
Merge branch '1.1.x' of github.com:AMDComputeLibraries/MLOpen into 1.1.x
MIOpen v.1.1.1
- Performance improvements for the HIP backend
- Robust error-checking
MIOpen v1.1
Notes:
- The scaling parameter alpha and shift parameter beta for layers kernels are only supported for alpha = 1 and beta = 0. The exceptions to this are for miopenOptTensor, miopenConvolutionForwardBias, and miopenConvolutionBackwardBias.
- Currently, only 32-bit floats are supported in MIOpen.
- MIOpen only supports tensor layout NCHW.
Changes:
- Added persistent cache for compiled GPU kernels
- Performance improvements for batch normalization kernels
- Performance improvements for all types of convolutions for 1x1 filters
- Performance improvements for all types of convolutions with non-unit strides
- Performance improvements for backward-weights convolutions for 3x3 filters
- Performance improvements for the AddTensor operation
- Various bug fixes for Winograd convolutions
1.0.2
Bump version