docs(hipdnn): update roadmap

projects/hipdnn/docs/Roadmap.md

@@ -48,7 +48,7 @@ This document outlines the development roadmap for hipDNN, a comprehensive graph
 - Forward LSE/softmax-stats output and a GPU forward reference ✅
 - Overridable tensor shapes API (required for variable sequence lengths) ✅
   - Phase 1 override-shape plumbing landed (RFC 0008) ✅
-- SDPA backward pass on gfx942 ⏳
+- SDPA backward pass on gfx942 ✅
   - FP32 gradient accumulation ✅
   - No-mask, causal and sliding-window mask modes ✅
   - BF16 and FP16 at head dim 128 ✅
@@ -89,118 +89,153 @@ This document outlines the development roadmap for hipDNN, a comprehensive graph
 
 **Focus:** SDPA, GEMM with MX low-precision data types, better heuristics & improved kernel provider selection
 
-> [!NOTE]
-> Q3 2026 ended on 2026-09-30. Items that were not complete moved to [Carried over from Q3 2026](#carried-over-from-q3-2026) under Q4 2026. Only completed Q3 items remain in this section.
+### SDPA
+- Forward GPU reference ✅
+- gfx950 causal forward ✅
+- **rocKE SDPA in the HIP kernel provider** — ROCm Kernel Engine SDPA kernels served through hipDNN ✅
+  - gfx950 dense attention engine ✅
+  - gfx1151 (Strix Halo) inference SDPA engine ✅
+
+### GEMM
+- **hipBLASLt plugin expanded operation & datatype support** ✅
+  - FP8 (OCP) dequantize + GEMM ✅
+- MX GEMMs (MX FP4 and MX FP6) through the hipBLASLt provider plugin ✅
+- Documented constraints surfaced for graph builders (alignment, batch, epilogues) ✅
 
 ### Heuristics
-- Heuristic plugin API (RFC 0007) ✅
+- Universal Heuristic Descriptor design (RFC 0019) ✅
+- Benchmarked engine-ranking cache, in-process and persisted to disk ✅
+
+### Normalization
+- Expanded LayerNorm & RMSNorm kernel coverage in the HIP kernel provider ✅
 
 ### Support matrix
 - Whole-graph bundled integration tests ✅
   - Default authoring format for new graph-verification tests ✅
   - Enabled as the default execution path in CI ✅
 
-## Q4 2026 & beyond (Current milestone)
+### Universal descriptors
+- **Data-driven kernel ingestion** via universal kernel, match, heuristic and engine descriptors (RFCs 0017–0020) ✅
+  - Generic kernel ingestor with build-time kernel packaging ✅
 
-**Focus:** Broader operation, layout & datatype coverage; deeper framework integrations; and the graph-execution, tensor and packaging work that moves hipDNN from tech preview toward general availability. We still value community input on what you would like to see prioritized!
+### cuDNN compatibility
+- **cuDNN v9 frontend compatibility shim** behind `HIPDNN_ENABLE_CUDNN_COMPATIBILITY` ✅
+  - Shim core node coverage landed ✅
+  - Node and enum coverage expansion ✅
 
-### Carried over from Q3 2026
+### Core
+- Runtime pass-by-value scalars (RFC 0016) ✅
 
-> [!NOTE]
-> These items were planned for Q3 2026. They were not complete when the quarter ended, so they moved here. Markers show the state at the time of the move.
+## Q4 2026 (Current milestone)
 
-#### SDPA
+**Focus:** Broad SDPA coverage, kernel-ingestor maturity, deeper integration testing, and expanded benchmarking.
+
+### SDPA
+- **Broad SDPA forward and backward coverage** ⏳
+  - rocKE integration ⏳
+  - Ahead-of-time (AOT) compiled FlyDSL kernels for broad coverage ⏳
 - Wider SDPA coverage: FP16 and FP8 forward, head dims up to 256, broader GQA configurations ⏳
-- Production-quality backward pass ⏳ (the forward GPU reference landed in Q2)
-- gfx950 backward and gfx950 causal forward
-- **rocKE SDPA plugin for hipDNN** — ROCm Kernel Engine SDPA provider plugin
-  - ROCm Kernel Engine (rocKE) in-progress, including a JIT SDPA path ⏳
+  - FP16 forward and broader GQA configurations ✅
+  - FP8 forward path wired through the SDPA engine ✅
+  - Head dim 256 on gfx1151 ✅ (CDNA head dim 256 still in progress)
+- Production-quality backward pass, including gfx950 ⏳
 - Ragged tensor support for variable sequence lengths (RFC 0014) ⏳
+  - API, descriptor lowering and GPU reference ✅
 
-#### GEMM
-- **hipBLASLt plugin expanded operation & datatype support** ⏳
-  - FP8 (OCP) dequantize + GEMM path landed behind `HIPBLASLTPROVIDER_ENABLE_MX_GEMM` (off by default) ✅
-- MX GEMMs through the hipBLASLt provider plugin ⏳
-- Documented constraints surfaced for graph builders (alignment, batch, epilogues)
+### Kernel ingestor
+- Performance improvements ✅
+- Execution plan save/load ✅
+- Expression handling improvements ⏳
+- Runtime improvements ⏳
 
-#### MOE (Mixture of Experts)
-- MOE frontend and backend POC (limited coverage) ⏳
-  - Frontend API, schema & backend descriptors landed ✅
-  - Initial provider kernel dispatch ⏳
+### MIOpen provider
+- Close the operation-coverage gap in the hipDNN MIOpen provider ⏳
 
-#### Heuristics
-- Plugin architecture ⏳
-- Phase 1 heuristic plugin: providing heuristic engine selection for limited architectures or team may pivot to support heuristics on rocKE
+### GEMM
+- **MX GEMM in the hipBLASLt provider** ⏳
+  - gfx1250 enablement ⏳
+  - Enabled by default ⏳
 
-#### Normalization
-- Expanded LayerNorm & RMSNorm kernel coverage in the HIP kernel provider ⏳
-- Expanded layout & datatype coverage for batchnorm ⏳
+### Convolution
+- Expanded convolution integration from rocKE ⏳
 
-#### Support matrix
-- Matrix published as a regular CI artifact ⏳
+### Integration testing
+- Per-op review and enhancements ⏳
+- "Gauntlet" mode for deeply proving integrations
 
-#### Testing
-- **Op coverage: detailed per-op integration tests**, in order Attention, then Convolution, then Normalization
-  - A test plan per tier (quick, standard, comprehensive, full) for each op
-  - Pairwise feature coverage per op, so the smallest set flags obvious breaks across providers
-  - Bundles that execute the plans
-  - Golden-reference baseline per op
+### Benchmarking & performance testing
+- Improved op support in [ROCm/dnn-benchmarking](https://github.com/ROCm/dnn-benchmarking), including SDPA backward and other backward ops
+- Graph metadata for improved tracking and filtering of the benchmark corpus
+- Expanded set of benchmarked graphs
 
-#### Universal descriptors
-- Data-driven kernel ingestion via universal kernel, match, dispatch, engine and heuristic descriptors (RFC 0017) ⏳
+### cuDNN compatibility
+- Continued cuDNN v9 frontend compatibility, including a user guide and samples ⏳
+- cuDNN v8 frontend compatibility ⏳
 
-#### Benchmarking & performance testing
-- Benchmarking CI running on AMD GPUs in [ROCm/dnn-benchmarking](https://github.com/ROCm/dnn-benchmarking) ⏳
+### Heuristics
+- Learned engine and kernel ranking via Universal Heuristic Descriptors, with shipped trained models ⏳
+- Heuristic generation pipeline for custom engines, letting custom engines compete with built-in engines in selection ⏳
 
-#### PyTorch
-- **PyTorch integration for opt-in hipDNN backend** ⏳
+### MOE (Mixture of Experts)
+- MOE frontend and backend POC (limited coverage) ⏳
+  - Frontend API, schema & backend descriptors landed ✅
+  - Initial provider kernel dispatch ⏳
 
-#### cuDNN compatibility
-- **cuDNN v9 frontend compatibility shim** behind `HIPDNN_ENABLE_CUDNN_COMPATIBILITY` ⏳
-  - Shim core node coverage landed ✅
-  - Node and enum coverage expansion in progress ⏳
+### Support matrix
+- Matrix published as a regular CI artifact ⏳
+
+### PyTorch
+- **PyTorch integration for opt-in hipDNN backend** ⏳
+  - `hipdnn_torch` injection layer ✅
 
-#### MIOpen integration
-- **MIOpen ↔ hipDNN shim** enabling MIOpen to route through hipDNN, behind `MIOPEN_ENABLE_HIPDNN_WRAPPER` (off by default) ⏳
+### MIOpen integration
+- **Opt-in runtime-feature flag to route MIOpen → hipDNN for convolutions** ⏳
   - MIOpen superbuild integration ✅
+  - Convolution entry points forwarded to hipDNN behind `MIOPEN_ENABLE_HIPDNN_WRAPPER` ✅
 
-#### Core
-- Add **hipRTC & caching support** to plugin SDK (Empowers plugin developers, and standardizes caching of artifacts)
-- Kernel engine tagging & filtering
-  - Numeric notes for filtering
-  - Client API to enable filtering
+### Core
 - Python API wrappers (general availability beyond POC) ⏳
-- Plugin SDK utility expansion to further streamline new-provider development
+  - Engine discovery, auto-tuning and DLPack interoperability ✅
+  - Frontend installable as a wheel ✅
+- Plugin SDK utility expansion to further streamline new-provider development ⏳
 
-### Testing
-- Unified gate for a new integration or kernel source, built on the per-op coverage, giving high confidence in its quality
-- Fuzzy mode for release testing (Q1 2027): exploratory runs within an affordable pre-release budget to find uncovered cases, grow the set, and show where coverage stands
+## Beyond Q4 2026
+
+**Focus:** Broader operation, layout & datatype coverage; deeper framework integrations; and the graph-execution, tensor and packaging work that moves hipDNN from tech preview toward general availability. We still value community input on what you would like to see prioritized!
 
 ### Increase operational support coverage
 - Additional high performance static fusion support for priority use cases
 - Additional JIT graph support for operations
 - Improve general operational support for operations:
   - Additional layout support
   - Additional datatype support
+- Expanded layout & datatype coverage for batchnorm
 
 ### Benchmarking & performance testing
-- Bindings installable as wheels ⏳
+- Benchmarking CI running on AMD GPUs in [ROCm/dnn-benchmarking](https://github.com/ROCm/dnn-benchmarking)
 - App installable as wheel
-- Automated benchmark runs weekly or daily off `develop`, with release-by-release regression tracking (Q1 2027)
 
 ### More framework integrations
 - vLLM backend
+- Deeper PyTorch integration, using hipDNN for attention
 - Currently discussing timelines for various framework integrations. Roadmap will be updated as they are defined.
 
 ### Normalization
 - **Distributed normalization support**
 
+### MIOpen integration
+- Harden the MIOpen → hipDNN convolution reroute until it can be the default-on path
+
 ### Core
+- Add **hipRTC & caching support** to plugin SDK (Empowers plugin developers, and standardizes caching of artifacts)
+- Kernel engine tagging & filtering
+  - Numeric notes for filtering
+  - Client API to enable filtering
 - Fallback (graph splitting) engine
 - Expanded performance and validation suites for hipDNN full install (using real user workloads and benchmarks to drive testing)
 - AOT graph compilation without devices present (Pre-compile graph support)
 - **hipGraph support**
-- Support dynamic linking to backend (enables forwards and backwards compatible client libraries) ⏳
+- Support dynamic linking to backend (enables forwards and backwards compatible client libraries)
   - Save/Load Execution plans (binary serialize/deserialize of graph + execution plan landed in Q2; full save/load across dynamic linking tracked here)
 - Non-standard tensor support (non-packed, vectorized)