upload 2nd version of sdpa backward

.github/ISSUE_TEMPLATE/bug_report.yml

@@ -1,5 +1,5 @@
 name: Bug Report
-description: Create a bug report to help us improve CUTLASS
+description: Create a bug report to help us improve SYCL*TLA
 title: "[BUG] "
 labels: ["? - Needs Triage", "bug"]
 assignees: []
@@ -10,8 +10,9 @@ body:
     attributes:
       label: Which component has the problem?
       options:
-        - CuTe DSL
-        - CUTLASS C++
+        - CuTe APIs
+        - CUTLASS APIs
+        - Python (APIs or Pypi package)
     validations:
       required: true
   - type: textarea

.github/ISSUE_TEMPLATE/documentation_request.md

@@ -1,6 +1,6 @@
 ---
 name: Documentation request
-about: Report incorrect or needed documentation to improve CUTLASS
+about: Report incorrect or needed documentation to improve SYCL*TLA
 title: "[DOC]"
 labels: "? - Needs Triage, documentation"
 assignees: ''

.github/ISSUE_TEMPLATE/feature_request.yml

@@ -1,5 +1,5 @@
 name: Feature Request
-description: Suggest an idea for CUTLASS
+description: Suggest an idea for SYCL*TLA
 title: "[FEA] "
 labels: ["? - Needs Triage", "feature request"]
 assignees: []
@@ -10,8 +10,9 @@ body:
     attributes:
       label: Which component requires the feature?
       options:
-        - CuTe DSL
-        - CUTLASS C++
+        - CuTe APIs
+        - CUTLASS APIs
+        - Python (APIs or Pypi package)
     validations:
       required: true
   - type: textarea
@@ -21,7 +22,7 @@ body:
       description: Please fill out all sections below
       value: |
         **Is your feature request related to a problem? Please describe.**
-        A clear and concise description of what the problem is. Ex. I wish I could use CUTLASS to do [...]
+        A clear and concise description of what the problem is. Ex. I wish I could use SYCL*TLA to do [...]
 
         **Describe the solution you'd like**
         A clear and concise description of what you want to happen.

.github/ISSUE_TEMPLATE/submit_question.md

@@ -1,6 +1,6 @@
 ---
 name: Submit question
-about: Ask a general question about CUTLASS
+about: Ask a general question about SYCL*TLA
 title: "[QST]"
 labels: "? - Needs Triage, question"
 assignees: ''

.github/PULL_REQUEST_TEMPLATE.md

@@ -0,0 +1,19 @@
+## Description
+<!-- What does this PR do? -->
+
+## Type
+- [ ] Bug  - [ ] Feature  - [ ] Performance  - [ ] Refactor
+
+## Testing
+- [ ] Tests pass  - [ ] Xe12  - [ ] Xe20
+
+## Performance
+| Metric | Before | After |
+|--------|--------|-------|
+|        |        |       |
+
+## References
+Fixes #
+
+## Checklist
+- [ ] Copyright  - [ ] Co-pilot Review  - [ ] Deprecated APIs not used

.github/PULL_REQUEST_TEMPLATE/bug_fix.md

@@ -0,0 +1,20 @@
+## Bug
+<!-- What's broken? -->
+
+Severity: <!-- Critical/High/Medium/Low -->
+
+## Root Cause
+<!-- Why? -->
+
+## Fix
+<!-- How? -->
+
+## Verification
+Before: <!-- error -->
+After: <!-- fixed -->
+
+## Testing
+- [ ] Regression/Units test
+- [ ] Tests pass
+
+## Details

.github/PULL_REQUEST_TEMPLATE/feature.md

@@ -0,0 +1,23 @@
+## Feature
+<!-- What capability? -->
+
+## Use Case
+<!-- Why? Who needs it? -->
+
+## API
+```cpp
+// Signature
+```
+
+## Example
+```cpp
+// Usage
+```
+
+## Testing
+- [ ] Tests  - [ ] Example  - [ ] Docs
+
+## ToDo
+- [ ] Implement A
+- [ ] Implement B
+- [ ] Document

.github/PULL_REQUEST_TEMPLATE/performance.md

@@ -0,0 +1,19 @@
+## Optimization
+<!-- What? -->
+
+## Profiling
+Tool: <!-- Benchmark Tests or Profiling Tools -->
+Bottleneck: <!-- What's slow? -->
+
+## Results
+| Case | Before | After | Gain |
+|------|--------|-------|------|
+|      |        |       |      |
+
+## Changes
+<!-- How? -->
+
+## Testing
+- [ ] Tests pass  - [ ] Xe12  - [ ] Xe20
+
+Related: #

.github/PULL_REQUEST_TEMPLATE/refactoring.md

@@ -0,0 +1,23 @@
+## Refactoring
+<!-- What code? -->
+
+## Why
+<!-- Motivation? -->
+
+## Changes
+<!-- Technique? -->
+
+## Preservation
+- [ ] Tests unchanged
+- [ ] Perf unchanged
+
+## Quality
+| Metric | Before | After |
+|--------|--------|-------|
+| LOC    |        |       |
+| Performance |    |       |
+
+## ToDo
+- [ ] Implement A
+- [ ] Implement B
+- [ ] Document

.github/actions/install-intel-graphics/action.yml

@@ -17,20 +17,22 @@ runs:
       run: |
         shopt -s expand_aliases
         which sudo || alias sudo=""
-        if [[ "${{ inputs.GPU }}" == "BMG" ]]; then
+        if [[ "${{ inputs.GPU }}" == "BMG" || "${{ inputs.GPU }}" == "PVC" ]]; then
           sudo add-apt-repository ppa:kobuk-team/intel-graphics
           sudo apt update
         else
-          . /etc/os-release
-          wget https://repositories.intel.com/gpu/ubuntu/dists/${VERSION_CODENAME}/intel-gpu-ubuntu-${VERSION_CODENAME}.run
-          chmod +x intel-gpu-ubuntu-${VERSION_CODENAME}.run
-          sudo ./intel-gpu-ubuntu-${VERSION_CODENAME}.run
-          sudo apt install -y \
-            intel-media-va-driver-non-free libmfx-gen1 libvpl2 \
-            libegl-mesa0 libegl1-mesa-dev libgl1-mesa-dev  \
-            libgles2-mesa-dev libigdgmm12 libxatracker2 mesa-va-drivers \
-            mesa-vdpau-drivers mesa-vulkan-drivers va-driver-all vainfo \
-            libigc-dev intel-igc-cm libigdfcl-dev libigfxcmrt-dev libze-dev hwinfo
+          # LTS PVC drivers
+          # . /etc/os-release
+          # wget https://repositories.intel.com/gpu/ubuntu/dists/${VERSION_CODENAME}/intel-gpu-ubuntu-${VERSION_CODENAME}.run
+          # chmod +x intel-gpu-ubuntu-${VERSION_CODENAME}.run
+          # sudo ./intel-gpu-ubuntu-${VERSION_CODENAME}.run
+          # sudo apt install -y \
+          #   intel-media-va-driver-non-free libmfx-gen1 libvpl2 \
+          #   libegl-mesa0 libegl1-mesa-dev libgl1-mesa-dev  \
+          #   libgles2-mesa-dev libigdgmm12 libxatracker2 mesa-va-drivers \
+          #   mesa-vdpau-drivers mesa-vulkan-drivers va-driver-all vainfo \
+          #   libigc-dev intel-igc-cm libigdfcl-dev libigfxcmrt-dev libze-dev hwinfo
+          exit 1
         fi
         sudo apt-get install -y libze-intel-gpu1 libze-dev intel-metrics-discovery \
           intel-opencl-icd ocl-icd-opencl-dev clinfo intel-gsc intel-ocloc g++

.github/copilot-instructions.md

@@ -0,0 +1,167 @@
+# Copilot Coding Agent Onboarding — SYCL*TLA
+
+Purpose
+-------
+This file is a short, focused onboarding guide so a coding agent (Copilot coding agent) can make correct, CI-safe changes to the SYCL*TLA repository without long exploratory searches. Keep edits conservative: prefer small, well-tested changes and follow the PR checklist described below.
+
+Top-level constraints (read first)
+---------------------------------
+- **Intel copyright headers**: Many files carry dual NVIDIA/Intel copyright headers. Do not remove or alter copyright headers on modified files.
+- **Intel Xe APIs**: The codebase uses new Intel "Xe" APIs (Xe12 for PVC, Xe20 for BMG) and Intel oneAPI toolchain conventions; prefer SYCL-compatible code and avoid adding CUDA-only code paths without explicit gating.
+- **CI Requirements**: Changes must build and pass CI workflows in `.github/workflows/*` (notably `intel_test.yml`, `intel_test_gpp_host.yml`, `sycl_python_test.yml`).
+- **Test Coverage**: Check for test coverage before making changes. C++ tests are in `test/unit/`, Python tests in `test/python/`.
+- **PR Descriptions**: Must include: what changed, why, local build/test steps performed, and expected CI/benchmark impact (see PR templates in `.github/PULL_REQUEST_TEMPLATE/`).
+
+Quick actions to always run locally before creating a PR
+------------------------------------------------------
+1. **ALWAYS source Intel environment first** (required for builds that target Intel compilers; if not available, CMake configure will still catch syntax errors but linking will fail):
+
+```bash
+source /opt/intel/oneapi/setvars.sh
+export CXX=icpx
+export CC=icx
+```
+
+2. **ALWAYS create a clean build directory** and configure for SYCL:
+
+```bash
+rm -rf build && mkdir build && cd build
+cmake .. -G Ninja \
+  -DCUTLASS_ENABLE_SYCL=ON \
+  -DDPCPP_SYCL_TARGET=intel_gpu_bmg_g21 \
+  -DCUTLASS_SYCL_RUNNING_CI=ON
+ninja
+```
+
+**Critical Notes:**
+- `-DDPCPP_SYCL_TARGET` must match your hardware: `intel_gpu_bmg_g21` for BMG (Arc B580), `intel_gpu_pvc` for PVC (Data Center Max). This affects intrinsic availability.
+- Build time: ~10-20 minutes for full build on 8-core machine.
+- If Intel oneAPI is not installed, CMake configure will still catch syntax errors but linking and target-specific checks will fail.
+- **NEVER commit without running a full build locally first**.
+
+Build / Test / Lint summary
+---------------------------
+- **Bootstrap**: No special bootstrap required. Python dependencies in `pyproject.toml` (`networkx`, `numpy`, `pydot`, `scipy`, `treelib`) are needed for Python tests. Install with `pip install -e .` in project root.
+- **Build**: Use CMake 3.22+ and Ninja (see commands above). **ALWAYS** run from clean build directory to avoid stale state.
+- **C++ Unit Tests**: After build, run `cmake --build . --target test_unit` (runs all unit tests in `test/unit/`).
+- **C++ Examples**: `cmake --build . --target test_examples` (builds and validates examples in `examples/`).
+- **Python Tests**: 
+  ```bash
+  cd python
+  python3 -m pytest -q
+  ```
+  CI runs specific test like `test/python/cutlass/gemm/gemm_bf16_pvc.py`. **ALWAYS** set `export CUTLASS_USE_SYCL=1` before running Python tests.
+- **Linting**: No automated linter. Follow existing code style and ensure `-Werror` flag passes (set in CI).
+
+**Environment Variables Required for Runtime:**
+```bash
+export ONEAPI_DEVICE_SELECTOR=level_zero:gpu
+export IGC_ExtraOCLOptions="-cl-intel-256-GRF-per-thread"
+export SYCL_PROGRAM_COMPILE_OPTIONS="-ze-opt-large-register-file -gline-tables-only"
+export IGC_VectorAliasBBThreshold=100000000000
+```
+These are set in CI workflows and should be set locally for accurate testing.
+
+Common failure modes & mitigations
+---------------------------------
+- **Missing Intel environment**: builds fail at linking or with unknown compilers. Mitigation: Source `/opt/intel/oneapi/setvars.sh` or unset `CXX`/`CC` to use system compilers for syntax-only checks.
+- **Wrong SYCL target**: some intrinsics are target-specific (e.g., 2D block prefetch intrinsics). Match the CI target or use conservative code paths.
+- **Layout constraints in Intel Xe epilogues** (ColumnMajor/RowMajor): prefer to reuse existing epilogue code and tests to avoid violating layout constraints. If making changes, run the affected tests locally.
+- **Missing libraries in LD_LIBRARY_PATH** for runtime: set `LD_LIBRARY_PATH` to include `build/tools/library` when running `python` tests that load `.so` wrappers.
+- **CMake cache issues**: If you see unexpected build behavior, **ALWAYS** delete `build/` completely and reconfigure. Stale CMake cache causes many hard-to-debug issues.
+- **Python import errors**: If Python tests fail with import errors, run `pip install -e .` from project root first.
+
+CI and validation pipelines (what will run)
+-------------------------------------------
+- See `.github/workflows/` for exact pipelines. Most important:
+  - `intel_test.yml` — primary CI build for Intel targets
+  - `intel_test_gpp_host.yml` — GPP host builds
+  - `sycl_python_test.yml` — Python test workflow
+  - `nvidia_test.yml` / `cuda_test.yml` — CUDA-targeted tests (keep changes SYCL-first unless explicitly modifying CUDA paths)
+
+How the agent should validate its changes
+-----------------------------------------
+1. Run a local CMake configure and build (fast smoke test):
+
+```bash
+rm -rf build && mkdir build && cd build
+cmake .. -G Ninja -DCUTLASS_ENABLE_SYCL=ON -DDPCPP_SYCL_TARGET=intel_gpu_bmg_g21 -DCUTLASS_SYCL_RUNNING_CI=ON
+ninja -k 0
+```
+
+2. Run the Python test subset that touches modified components (or all Python tests if the change is cross-cutting):
+
+```bash
+cd python
+python3 -m pytest -q
+```
+
+3. For C++ kernel changes, run unit tests: `cmake --build . --target test_unit -j 8`
+
+4. For examples changes, run: `cmake --build . --target test_examples -j 1`
+--------------------------------------------------
+- Short summary of change and the files modified.
+- Build steps executed locally (CMake + Ninja commands, environment variables set).
+- Tests run and their results (include pytest subset names and pass/fail counts).
+- If the change affects performance or kernel selection, include expected performance impact and a short benchmark (size and results).
+- State whether the Intel oneAPI environment was required to fully validate the change.
+
+Project layout (quick map)
+--------------------------
+- Root files: `CMakeLists.txt`, `README.md`, `CHANGELOG-SYCL.md`, `SYCL.cmake`, `pyproject.toml`
+- Major directories:
+  - `include/` — core headers and kernel templates
+  - `python/` — Python wrapper, generator, tests
+  - `examples/` — usage examples, e.g., `11_xe20_cutlass_library`
+  - `test/` — C++ tests and validation kernels
+  - `tools/` — build/test utilities
+  - `media/` — documentation and architecture notes (search `media/docs/cpp/xe_rearchitecture.md`, `media/docs/python/xe_cutlass_library.md`)
+
+Files the agent should inspect when making changes
+--------------------------------------------------
+- `python/cutlass_library/generator.py` — kernel generation and filtering logic
+- `python/cutlass_library/arch_constants.py` — architecture detection and constants
+- `include/cutlass/gemm/kernel/*` — GEMM kernel implementations
+- `.github/workflows/*` — CI steps; ensure changes don't break these workflows
+
+Search tips
+-----------
+- Use `grep -R "HACK\|TODO\|WORKAROUND\|FIXME"` to find fragile areas.
+- Search for `intel` and `Xe` keywords to find Intel-specific code paths.
+
+Testing coverage
+----------------
+- The repo contains Python tests in `python/` and C++ tests under `test/`.
+- Before assuming full coverage, run the test suite locally and include failing tests in your PR notes.
+
+Special rules for changes
+-------------------------
+- Keep changes minimal and well-scoped. If modifying kernel selection or architecture constants, include tests or fallbacks.
+- Preserve Intel copyright headers.
+- Avoid introducing CUDA-only code paths in SYCL code.
+
+When to run a wider search
+--------------------------
+Trust these instructions first. Only perform a broad code search if:
+- The instructions are clearly missing information for the requested change, or
+- A test or build step fails unexpectedly after following these steps.
+
+Where to look for help
+----------------------
+- `README.md` and `media/docs/*` for architecture details
+- `.github/workflows/` for CI expectations
+- Open issues and PR templates in `.github/ISSUE_TEMPLATE` and `.github/PULL_REQUEST_TEMPLATE`
+
+Short checklist before opening a PR
+-----------------------------------
+- [ ] Build configured and compiled locally (or syntax-checked if environment unavailable)
+- [ ] Relevant tests run locally and passed
+- [ ] PR description includes steps and validation results
+- [ ] No removal of Intel copyright headers
+
+If anything here proves inaccurate
+----------------------------------
+Run the minimal searches you need, then update this file with the corrected steps so future agents benefit.
+
+---
+This file is intentionally short (<= 2 pages). For deeper onboarding, consult `README.md`, `media/docs/*`, and the workflows in `.github/workflows/`.