MIOpen v2.3.0

Notes:

• This release contains new implementations of the implicitGEMM and Winograd algorithms, performance improvements for convolutions, further support for 3D convolutional networks, and various bug fixes.

Changes:

• Added 3D Pooling layers
• Added backwards data algorithm for implicitGEMM
• Added GEMM performance improvements via relaxed constraints in rocBLAS-Tensile
• Added full CO v3 support for all kernels in MIOpen
• Added new Winograd group convolution kernels
• Added an API to query MIOpen's version
• Added parallel compilation in initial convolutional algorithm search; partial solution to #130
• Added SQLite binary program cache
• Improved logging across all layers
• Improved MIOpen's internal design for calling convolutional solvers
• Fixed various bugs for the implicitGEMM algorithm

MIOpen v2.2.1

Notes:

• This release contains bug fixes, documentation updates, and further code object version 3 support

Changes:

• Added support for multiple ROCm installations
• Added additional support for code object v3
• Fixed issue with incorrect LRN calculation #127
• Fixed incorrect performance database documentation
• Fixed issue with incorrect workspace calculation in group convolutions
• Fixed issue with unsupported hardware instructions used with inline assembly

MIOpen v2.2.0

Notes:

• This release contains bug fixes, performance improvements, and expanded applicability for specific convolutional algorithms.
• MIOpen has posted a citable paper on ArXiv here.
• An SQLite database has been added to replace the text-based performance database. While the text file still exists, by default SQLite is used over the text-based performance database; see documentation from more details.

Changes:

• Added per solution algorithm filtering environmental variable for debugging
• Added SQLite3 database and build dependency. The text-based performance database support is deprecated and will be removed in the next release.
• Added citation page to documentation pointing to MIOpen's paper
• Added to the overall documentation
• Fixed fusion compilation check issue
• Fixed fusion group convolution warning
• Improved performance of forward pooling
• Improved performance of convolutions
• Improved performance of spatial training batch normalization for some large batch size input configurations
• Improved applicability of implicit GEMM convolution algorithm
• Improved performance of calls to miopenConvolutionXXXGetWorkSpaceSize() functions
• Improved conformance to code object version 3
• Disabled SCGEMM convolution algorithm by default; this algorithm is deprecated and will be removed in future releases
• Changed "hip_hcc" to "hip-hcc" for the MIOpen package requirements in CMakeLists.txt

MIOpen v2.1.0

Notes:

• This release contains new layers, bug fixes, and a new convolution algorithm.

Changes:

• Added a dropout layer API for training
• Added a new SCGEMM algorithm for convolutions
• Added further support for bfp16 in convolutions
• Added a docker hub link for MIOpen docker images.
• Fixed issue with NaN appearing on batch normalization backwards pass in fp16
• Fixed softmax kernel bug in log mode #112
• Fixed gfx803 support issue #869
• Fixed gfx803 kernel issue #117
• Fixed issue with disabled GEMM #119
• Improved performance of batch normalization fp16 forward training layers
• Improved performance of convolutions layers
• Removed MIOpenGEMM as a requirement for the HIP backend. It is now optional.

MIOpen v2.0.1

Notes:

• This release contains bug fixes and performance improvements.
• Additionally, the convolution algorithm Implicit GEMM is now enabled by default
• Known issues:
  • Backward propagation for batch normalization in fp16 mode may trigger NaN in some cases
  • Softmax Log mode may produce an incorrect result in back propagation

Changes:

• Added Winograd multi-pass convolution kernel
• Fixed issue with hip compiler paths
• Fixed immediate mode behavior with auto-tuning environment variable
• Fixed issue with system find-db in-memory cache, the fix enable the cache by default
• Improved logging
• Improved how symbols are hidden in the library
• Updated default behavior to enable implicit GEMM

MIOpen v2.0.0

Notes:

• This release contains several new features including an immediate mode for selecting convolutions, bfloat16 support, new layers, modes, and algorithms.
• MIOpenDriver, a tool for benchmarking and developing kernels is now shipped with MIOpen.
• BFloat16 now supported in HIP requires an updated rocBLAS as a GEMM backend.
• Immediate mode API now provides the ability to quickly obtain a convolution kernel.
• MIOpen now contains HIP source kernels and implements the ImplicitGEMM kernels. This is a new feature and is currently disabled by default. Use the environmental variable "MIOPEN_DEBUG_CONV_IMPLICIT_GEMM=1" to activation this feature. ImplicitGEMM requires an up to date HIP version of at least 1.5.9211.
• A new "loss" catagory of layers has been added, of which, CTC loss is the first. See the API reference for more details.
• 2.0 is the last release of active support for gfx803 architectures. In future releases, MIOpen will not actively debug and develop new features specifically for gfx803.
• System Find-Db in memory cache is disabled by default. Please see build instructions to enable this feature.

Changes:

• Added support for bfloat16 datatype in convolutions
• Added softmax channel mode and new softmax version 2 API
• Added fast / accurate / log softmax algorithms
• Added new implicit GEMM convolution algorithm for forward and backwards data passes, disabled by default
• Added int32 datatype support for output tensors in int8 convolutions
• Added immediate mode for finding the best convolution kernel for a given configuration
• Added a Find-Db infrastructure which stashes results of find on a user's system
• Added a shipped System Find-Db containing offline run Find() results
• Added an additional, faster batch norm assembly kernel for fp16
• Added CTC loss layer
• Added MIOpenDriver as a default component in MIOpen's build #34
• Fixed C compatability for boolean types in C API #103
• Fixed incorrect calculation in per-activation batch norm backwards pass #104
• Fixed bug #95 with asm batch norm ISA
• Fixed IsApplicable bug in Conv3x3Asm for group convolutions
• Improved performance of 1x1 stride 2 fp32 convolutions in the forward and backwards data passes
• Improved 3-D convolution stability
• Improved applicability of direct convolution backwards weights for 2x2, 5x10, and 5x20 filter sizes
• Improved maintainability in kernels and cpp code
• Updated rocBLAS minimum version to branch master-rocm-2.6

MIOpen v1.8.1

Notes:

• This release contains minor bug fixes and additional performance database improvements.

Changes:

• Fixed accuracy issue with backwards weights
• Fixed issue with name parsing for newer architectures
• Added narrow workaround for 5x10 and 5x20 filter performance regression
• Improved support in performance database for Radeon VII

MIOpen v1.8.0

Notes:

• This release contains full 3-D convolution support and int8 support for inference.
• Additionally, there are major updates in the performance database for major models including those found in Torchvision.
• An assortment of bugs have been resolved in this release.

Changes:

• Fixed various issues in assembly kernels
• Fixed issue #92 and #79 for miopenOpTensor
• Fixed issue #88 for bzip2
• Fixed issue #77 algorithm mismatch
• Added Winograd support for fp32 backwards weights
• Added pooling inclusive mode
• Added tuning for direct group convolution algorithms
• Added additional kernel support for group convolutions
• Added API for 3-D convolutions
• Added support for int8 inference convolutions
• Added integer selection for pooling indexing
• Added minimum dependencies support
• Added RNN fp16 support on the MIOpen-HIP backend
• Added 1x1 convolution + bias + activation fusions
• Added workaround for issue #84 GPU memory access fault
• Added performance tuning for direct backwards weights
• Improved performance database coverage
• Improved internal quality by reducing redunant code
• Improved build instructions in README.md
• Improved performance database coverage for fusions
• Updated Docker components and requirements

Known Issues:

• RNNs do not support fp16 on the MIOpen-OpenCL backend
• OpenCL backend does not support GEMM convolutions in fp16

MIOpen v1.7.1

Notes:

• This release contains minor bug fixes and performance improvements.

Changes:

• Fixed corrupt and obsolete performance database entries
• Fixed issue #70
• Fixed issue #72
• Fixed issue #77
• Removed default dependency of RNNs on rocBLAS
• Added a workaround for softmax fp16 correctness issue
• Added check to only make MIOpen with static boost libraries
• Improved performance database coverage

Known Issues:

• RNNs do not support fp16
• OpenCL backend does not support GEMM convolutions in fp16
• Layer fusions for convolution 1x1 fp16 are not supported
• Layer fusions for large image 1x1 convolutions may cause an exception instead of a warning during compile phase if plan is not supported

MIOpen v1.7.0

Notes:

• This release contains general bug fixes and an updated performance database
• Group convolutions backwards weights performance has been improved
• Logging across the library has been improved
• Performance database has been updated

Changes:

• Fixed logging issues with group convolution and pooling
• Fixed sphinx version issue in document generation
• Fixed issues with corrupt entries in performance database
• Removed external dependency on libSSL and libCrypto
• Added support for large image backwards weights in direct convolution
• Added fp16 support for RNNs on the HIP backend
• Improved performance database coverage

Known Issues:

• RNNs do not support fp16
• OpenCL backend does not support GEMM convolutions in fp16
• Layer fusions for convolution 1x1 fp16 are not supported
• Layer fusions for large image 1x1 convolutions may cause an exception instead of a warning during compile phase if plan is not supported