From c54e0d903357d5b581b5a8c79c3d75575c46aecb Mon Sep 17 00:00:00 2001 From: David Peabody Date: Wed, 18 Mar 2026 11:09:18 -0600 Subject: [PATCH] Update hipDNN roadmap for end of Q1 2026 Reflect actual completion status based on code review of hipDNN and dnn-providers. Q1 items completed ahead of schedule (Fusilli batchnorm, SDPA frontend, LayerNorm/RMSNorm frontend, Python POC, benchmarking tooling) are added with done status. PyTorch integration moved to early Q2 with upstream PR reference. Q2 updated with in-progress markers and current SDPA/normalization/GEMM status. Co-Authored-By: Claude Opus 4.6 --- projects/hipdnn/docs/Roadmap.md | 41 ++++++++++++++++++++------------- 1 file changed, 25 insertions(+), 16 deletions(-) diff --git a/projects/hipdnn/docs/Roadmap.md b/projects/hipdnn/docs/Roadmap.md index 5d74b832bc14..f719044f3ff5 100644 --- a/projects/hipdnn/docs/Roadmap.md +++ b/projects/hipdnn/docs/Roadmap.md @@ -9,9 +9,9 @@ This document outlines the development roadmap for hipDNN, a comprehensive graph > > ⏳ = In progress -## P0 ~ Q1 2026 (Current milestone) +## P0 ~ Q1 2026 -**Focus:** Stable foundation, core operations & initial PyTorch integration +**Focus:** Stable foundation & core operations ### Conv - **Convolution MIOpen plugin support** ✅ @@ -21,36 +21,44 @@ This document outlines the development roadmap for hipDNN, a comprehensive graph ### Normalization - **Batch normalization MIOpen plugin support** ✅ - Including basic fusions ✅ +- **Batch normalization Fusilli plugin support** ✅ +- **LayerNorm & RMSNorm frontend API** ✅ ### GEMM - **Initial frontend GEMM API support** ✅ - Fusilli plugin integration (see note) ✅ -- hipBLASLt plugin initial enablement ⏳ +- hipBLASLt plugin initial enablement ✅ + +### SDPA +- **SDPA frontend API & backend descriptors** ✅ ### Core -- **Stable, robust library to build upon** ⏳ -- Kernel engine settings (Engine knob configurations API + implementation) ⏳ +- **Stable, robust library to build upon** ✅ +- Kernel engine settings (Engine knob configurations API + implementation) ✅ - Ex. Flag for enabling benchmarking mode on MIOpen plugin +- Initial Python bindings POC ✅ +- Initial benchmarking & performance tooling ✅ -### PyTorch -- **PyTorch integration for opt-in hipDNN backend** ⏳ +> **Notes:** +> - Fusilli plugin is opt-in, and not defaulted on yet. +> - PyTorch integration was moved to early Q2 -> **Notes:** Fusilli plugin is opt-in, and not defaulted on yet. +## P1 ~ Q2 2026 (Current milestone) -## P1 ~ Q2 2026 +**Focus:** PyTorch integration, more operations, basic engine selection heuristic & core improvements -**Focus:** More operations, basic engine selection heuristic & core improvements +### PyTorch +- **PyTorch integration for opt-in hipDNN backend** ⏳(Early Q2) ### GEMM -- **hipBLASLt plugin expanded operation & datatype support** +- **hipBLASLt plugin expanded operation & datatype support** ⏳ ### SDPA -- Initial limited SDPA frontend API support -- Initial limited SDPA kernel provider (TBD) +- Limited SDPA provided by Fusilli, CK & ASM ⏳ ### Normalization -- Adding new **HIP kernel provider plugin** to expand normalization support -- Expanded operation API & coverage to support Layernorm & RMS +- Adding new **HIP kernel provider plugin** to expand normalization support ⏳ +- Expanded operation API & coverage to support Layernorm & RMS ⏳ - Expanded layout & datatype coverage for batchnorm ### Heuristics @@ -61,6 +69,7 @@ This document outlines the development roadmap for hipDNN, a comprehensive graph - Plugin SDK utilities to streamline plugin development for new providers ⏳ - Benchmarking & performance python tools ⏳ - Python API wrappers ⏳ +- **Client auto-tuning API** ## P2 ~ Q3 2026 @@ -77,7 +86,7 @@ This document outlines the development roadmap for hipDNN, a comprehensive graph - Kernel engine tagging & filtering - Behavioral & numeric notes for filtering - Client API to enable filtering -- **Client auto-tuning API** + ## P3 ~ Q4 2026 & beyond