Skip to content
Merged
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
24 changes: 24 additions & 0 deletions projects/miopen/CHANGELOG.md
Original file line number Diff line number Diff line change
Expand Up @@ -6,14 +6,38 @@ Full documentation for MIOpen is available [here](https://rocm.docs.amd.com/proj
## MIOpen 3.5.2 for ROCm 7.14.0
### Added
* Added an informational log when the Composable Kernel (CK) dynamic library is loaded.
* [Conv] Added a gfx950 (MI350X/MI355X) SystemDB, so gfx950 runs use tuned find/perf-database entries instead of falling back to runtime tuning on every invocation.
* [Conv] Added initial gfx1250 architecture support.
* Added JSON-formatted performance logs (`MIOPEN_PERFORMANCE_LOGS`) for convolution driver commands, replacing ad hoc text logs to simplify parsing kernel names and execution times.

### Changed
* [Conv] Load the Composable Kernel (CK) dynamic library during handle creation (`miopenCreate`) instead of on first CK use, so the one-time library-load cost is paid at a predictable setup point rather than as an unpredictable mid-run stall.
* [Conv] Naive convolution solvers are now skipped by default during find when any non-naive solver succeeds across any algorithm. Set `MIOPEN_NAIVE_DISABLE_IF_ALT=0` to restore the previous behavior.
* [Conv] Refreshed the gfx942 SystemDB with updated tuned entries from recent perf-eval runs.
* [Conv] Refined the 2D AI heuristics (Tunanet solver selection and kernel-tuning-network candidate ranking) with feature-engineered inputs for gfx942 and gfx950, improving selection for grouped implicit-GEMM convolutions.

### Resolved Issues
* [RNN] Fix RNN workspace tensor descriptor int overflow
* [Conv] Enabled grouped Composable Kernel (CK) xdlops fwd, bwd, and wrw convolution (2D and 3D) for tensors whose strides exceed the int32 range.
* [Conv] Fixed `miopenStatusInternalError` thrown by Find on depthwise NHWC grouped convolutions under `MIOPEN_FIND_MODE=NORMAL`.
* Fixed a thread-safety issue where concurrent access to the AI-heuristic model caches was not guarded by a mutex.
* [Conv] Fixed Composable Kernel (CK) grouped-convolution solvers not always being registered in host code.
* [Conv][gfx1250] Fixed a legacy CK buffer resource descriptor issue.
* [Conv] Removed stale gfx942 SystemDB entries that incorrectly recorded the naive convolution solver as fastest, allowing the correct (non-naive) solver to be selected.
* [Softmax] Fixed a NaN result when `beta=0`.
* [BatchNorm] Fixed a cross-channel vectorized memory access bug in the backward spatial variant 1 kernel.
* [Conv] Fixed a use-after-free of `kernel_id` in `InitInvokerFactoryNCHW`.
* Fixed a kernarg corruption bug in the SoftMarginLoss forward reduce kernel.
* Fixed a build failure when compiling with `MIOPEN_ENABLE_AI_KERNEL_TUNING=OFF`.
* Fixed the install RPATH for `MIOpenDriver` and Composable Kernel (CK) backend libraries.
* Fixed a dangling lambda capture that could cause a segfault at `MIOPEN_LOG_LEVEL=6`.
* [Windows] Enabled long-path support for `addkernels`.

### Removed
* Removed OpenCL solver `ConvOclDirectFwd1x1`.
* Removed OpenCL convolution solver `conv_ocl_dir2Dfwdgen`.
* Removed OpenCL solvers `ConvOclBwdWrW1x1` and `ConvOclBwdWrW2`.
* Removed OpenCL solvers `OpTensorLeadingOnes` and `Op5dTensorGeneric` (replaced by HIP equivalents).

## MIOpen 3.5.1 for ROCm 7.13.0
### Added
Expand Down
Loading