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