From b0a8e8da9bf4ef97b6d5c392953c6eee16995df4 Mon Sep 17 00:00:00 2001 From: Jonathan Lichtner <195780826+JonathanLichtnerAMD@users.noreply.github.com> Date: Thu, 30 Jul 2026 20:08:48 -0600 Subject: [PATCH 1/2] docs(miopen): add ROCm 10.0.0 changelog section Add a MIOpen 3.6.0 for ROCm 10.0.0 section documenting the changelog-worthy changes present in release/therock-10.0 that are not in release/therock-7.14. JIRA ID: ALMIOPEN-2364 Co-Authored-By: Claude Opus 4.8 --- projects/miopen/CHANGELOG.md | 16 ++++++++++++++++ 1 file changed, 16 insertions(+) diff --git a/projects/miopen/CHANGELOG.md b/projects/miopen/CHANGELOG.md index de8197154459..240ab0c52f50 100644 --- a/projects/miopen/CHANGELOG.md +++ b/projects/miopen/CHANGELOG.md @@ -3,6 +3,22 @@ Full documentation for MIOpen is available [here](https://rocm.docs.amd.com/projects/MIOpen/en/latest/) +## MIOpen 3.6.0 for ROCm 10.0.0 +### Added +* [Conv] Added gfx950 (MI350X/MI355X) 7x7 depthwise forward and backward-data convolution support (fp16/bf16), fixing a slow fallback-to-naive-kernel regression in ConvNeXt-style depthwise convolutions. + +### Changed +* [Conv] Restored gfx12x support in the Winograd Rage solver, recovering performance that regressed when earlier gfx12 support was reverted. +* [Conv] Refreshed the gfx1100, gfx1102, and gfx1201 (Navi) SystemDBs with updated tuned find/perf-database entries. +* [Conv] Refreshed the gfx950 SystemDB with additional tuned entries to cover more models. + +### Resolved Issues +* [BatchNorm] Fixed an off-by-stride indexing bug in the backward CalcStats mean/variance remainder loop that caused a ~1% systematic bias in NCHW batch normalization backward results. +* [Conv][gfx1250] Fixed an integer overflow in tensor operation kernels for large allocations that could cause memory access faults. + +### Removed +* Removed the OpenCL (OCL) backend; MIOpen now supports the HIP backend only. + ## MIOpen 3.5.2 for ROCm 7.14.0 ### Added * Added an informational log when the Composable Kernel (CK) dynamic library is loaded. From 32dc210b2a1a6b5da3b1cc47977933beced4dc25 Mon Sep 17 00:00:00 2001 From: Jonathan Lichtner <195780826+JonathanLichtnerAMD@users.noreply.github.com> Date: Mon, 10 Aug 2026 11:43:11 -0600 Subject: [PATCH 2/2] docs(miopen): document global-work-size launch-rejection fix for 10.0.0 An additional MIOpen fix (#10188, cherry-picked as #10359) landed on release/therock-10.0 after the 3.6.0 changelog section was first written. It is not present on release/therock-7.14, so it belongs in the 7.14->10.0 delta this section documents. JIRA ID: ALMIOPEN-2364 Co-Authored-By: Claude Sonnet 5 --- projects/miopen/CHANGELOG.md | 1 + 1 file changed, 1 insertion(+) diff --git a/projects/miopen/CHANGELOG.md b/projects/miopen/CHANGELOG.md index 240ab0c52f50..7520c831939c 100644 --- a/projects/miopen/CHANGELOG.md +++ b/projects/miopen/CHANGELOG.md @@ -15,6 +15,7 @@ Full documentation for MIOpen is available [here](https://rocm.docs.amd.com/proj ### Resolved Issues * [BatchNorm] Fixed an off-by-stride indexing bug in the backward CalcStats mean/variance remainder loop that caused a ~1% systematic bias in NCHW batch normalization backward results. * [Conv][gfx1250] Fixed an integer overflow in tensor operation kernels for large allocations that could cause memory access faults. +* [Conv] Fixed a naive convolution solver failure mode where a global work size of 2^32 or more work-items silently failed to launch and left a stale HIP error visible after Find returned success; such launches are now rejected up front. ### Removed * Removed the OpenCL (OCL) backend; MIOpen now supports the HIP backend only.