[WIP]

.github/workflows/buildRyzenWheels.yml

@@ -221,14 +221,14 @@ jobs:
 
   build-windows:
     name: Build and upload mlir_aie wheels (Windows)
-  
+
     runs-on: windows-2022
-  
+
     permissions:
       id-token: write
       contents: write
       packages: read
-  
+
     strategy:
       fail-fast: false
       matrix:
@@ -250,13 +250,13 @@ jobs:
 
           - python_version: "3.13"
             ENABLE_RTTI: OFF
-  
+
           - python_version: "3.14"
             ENABLE_RTTI: ON
-  
+
           - python_version: "3.14"
             ENABLE_RTTI: OFF
-  
+
     steps:
       - uses: actions/checkout@v6
         with:
@@ -355,32 +355,32 @@ jobs:
           CMAKE_ARGS: -DAIE_BUILD_CHESS_CLANG=OFF -DOPENSSL_USE_STATIC_LIBS=TRUE -DAIE_ENABLE_XRT_PYTHON_BINDINGS=OFF
         run: |
           set -euo pipefail
-  
+
           # No Vitis on Windows!
           unset VITIS XILINXD_LICENSE_FILE || true
-  
+
           git config --global --add safe.directory "$PWD"
           MLIR_VERSION=$(git rev-parse --short HEAD)
           echo "Building mlir-aie version $MLIR_VERSION"
-  
+
           python -m venv aie-venv
           source aie-venv/Scripts/activate
-  
+
           python -m pip install --upgrade pip
           pip install -r python/requirements_ml.txt
           pip install -r python/requirements_dev.txt
-  
+
           export ENABLE_RTTI="${ENABLE_RTTI}"
-  
+
           NO_RTTI=""  # Set a default value
           NO_RTTI_UNDERSCORE=""  # Set a default value
           if [ x"$ENABLE_RTTI" == x"OFF" ]; then
             NO_RTTI="-no-rtti"
             NO_RTTI_UNDERSCORE="_no_rtti"
           fi
-  
+
           VERSION=$(utils/clone-llvm.sh --get-wheel-version)
-  
+
           # Grab the MLIR distro wheel and extract
           for attempt in 1 2 3; do
             rm -f mlir*.whl
@@ -394,10 +394,10 @@ jobs:
             sleep $((attempt * 5))
           done
           python -m zipfile -e mlir*.whl .
-  
+
           # Linux-style timestamp magic should work fine on Windows.
           find "mlir${NO_RTTI_UNDERSCORE}" -exec touch -a -m -t 201108231405.14 {} \;
-  
+
           # Match Linux wheel version metadata on non-tag builds.
           export DATETIME=$(date +"%Y%m%d%H")
           if [ x"${{ inputs.AIE_COMMIT }}" == x"" ]; then
@@ -416,19 +416,19 @@ jobs:
           export MLIR_AIE_SOURCE_DIR="${ROOT_WIN}"
           export WHEELHOUSE_DIR="${ROOT_WIN}/wheelhouse"
           export CMAKE_MODULE_PATH="${ROOT_WIN}/cmake/modulesXilinx"
-  
+
           mkdir -p "${WHEELHOUSE_DIR}"
           pushd utils/mlir_aie_wheels
-  
+
           pip install wheel importlib_metadata "ninja!=1.13.0"
           CIBW_ARCHS=AMD64 pip wheel . -v -w "${WHEELHOUSE_DIR}" --no-build-isolation
-  
+
           popd
-  
+
           # Try to repair the wheel on Windows using delvewheel (closest auditwheel equivalent).
           pip install delvewheel
           python -m delvewheel repair --ignore-existing --analyze-existing-exes -w "${WHEELHOUSE_DIR}/repaired_wheel" "${WHEELHOUSE_DIR}"/mlir_aie*.whl
-  
+
       - name: Upload mlir_aie
         uses: actions/upload-artifact@v4
         with:

.github/workflows/mlirDistro.yml

@@ -157,7 +157,7 @@ jobs:
     - name: set ENV
       shell: bash
       run: |
-        
+
         PIP_FIND_LINKS_URL="https://github.com/Xilinx/mlir-aie/releases/expanded_assets/mlir-distro"
         if [ x"${{ github.event_name }}" == x"pull_request" ]; then
           PIP_FIND_LINKS_URL="$PIP_FIND_LINKS_URL https://github.com/Xilinx/mlir-aie/releases/expanded_assets/dev-wheels"
@@ -202,15 +202,15 @@ jobs:
       shell: bash
       working-directory: ${{ env.TEMP }}
       run: |
-        
+
         ls "${{ steps.setup_base.outputs.WORKSPACE_ROOT }}"
-        
+
         if [ x"${{ matrix.OS }}" == x"windows-2022" ]; then
           WORKSPACE_ROOT="${{ steps.setup_base.outputs.WORKSPACE_ROOT }}\utils\mlir_wheels"
         else
           WORKSPACE_ROOT="${{ steps.setup_base.outputs.WORKSPACE_ROOT }}/utils/mlir_wheels"
         fi
-        
+
         echo "WORKSPACE_ROOT=$WORKSPACE_ROOT" | tee -a $GITHUB_OUTPUT
 
       # setup llvm
@@ -219,7 +219,7 @@ jobs:
       working-directory: ${{ steps.workspace_root.outputs.WORKSPACE_ROOT }}
       shell: bash
       run: |
-        
+
         curl -s https://codeload.github.com/llvm/llvm-project/zip/${{ needs.get_llvm_project_commit.outputs.LLVM_PROJECT_COMMIT }} -o llvm.zip
         unzip -q llvm.zip
         rm -rf llvm.zip
@@ -233,7 +233,7 @@ jobs:
       shell: bash
       working-directory: ${{ steps.workspace_root.outputs.WORKSPACE_ROOT }}
       run: |
-        
+
         APPLY_PATCHES=${{ inputs.APPLY_PATCHES == '' && 'true' || inputs.APPLY_PATCHES }} \
         CIBW_ARCHS=${{ matrix.ARCH }} \
         CMAKE_GENERATOR=Ninja \
@@ -249,12 +249,12 @@ jobs:
       working-directory: ${{ steps.workspace_root.outputs.WORKSPACE_ROOT }}
       shell: bash
       run: |
-        
+
         export APPLY_PATCHES=${{ inputs.APPLY_PATCHES == '' && 'true' || inputs.APPLY_PATCHES }}
         ./scripts/apply_patches.sh
-        
+
         pip install --upgrade pip
-        
+
         CIBW_ARCHS=${{ matrix.ARCH }} \
         CMAKE_GENERATOR=Ninja \
         DATETIME=${{ needs.get_llvm_project_commit.outputs.DATETIME }} \
@@ -267,15 +267,15 @@ jobs:
       working-directory: ${{ steps.workspace_root.outputs.WORKSPACE_ROOT }}
       shell: bash
       run: |
-        
+
         rm -rf llvm-project
         rm -rf build
 
     - name: Docker prune
       if: contains(inputs.MATRIX_OS, 'ubuntu')
       shell: bash
       run: |
-        
+
         docker system prune -a -f
 
     - name: Get wheel version
@@ -293,7 +293,7 @@ jobs:
       working-directory: ${{ steps.workspace_root.outputs.WORKSPACE_ROOT }}
       shell: bash
       run: |
-        
+
         ccache -s
         HOST_CCACHE_DIR="$(ccache --get-config cache_dir)"
         rm -rf $HOST_CCACHE_DIR
@@ -309,7 +309,7 @@ jobs:
       working-directory: ${{ steps.workspace_root.outputs.WORKSPACE_ROOT }}
       shell: bash
       run: |
-        
+
         ccache --print-stats
         HOST_CCACHE_DIR="$(ccache --get-config cache_dir)"
         # Set the timestamp to the beginning of the current hour.
@@ -334,19 +334,19 @@ jobs:
           fi
           unzip -j wheelhouse/mlir*whl "mlir/bin/$TOOL" -d native_tools/
         done
-        
+
         if [ x"${{ matrix.OS }}" == x"ubuntu-22.04" ]; then
           PLAT="linux"
         elif [ x"${{ matrix.OS }}" == x"windows-2022" ]; then
           PLAT="win"
         fi
-        
+
         PLAT=${PLAT}_$(echo ${{ matrix.ARCH }} | tr '[:upper:]' '[:lower:]')
         pushd native_tools
 
         MLIR_WHEEL_VERSION=${{ steps.get_wheel_version.outputs.MLIR_WHEEL_VERSION }} \
         python setup.py bdist_wheel --dist-dir ../wheelhouse --plat $PLAT
-        
+
         popd
 
     # done

README.md

@@ -140,8 +140,8 @@ xrt-smi examine
 1. Install IRON library by installing the `mlir-aie` wheels:
 
    For installing the `mlir-aie` wheels, there are 3 options. Note that for whichever path you take,
-   it is important to sync the `mlir-aie` wheels version, the github repo commit, and the requirements versions. 
-   If you install from something other than the latest wheels, make sure 
+   it is important to sync the `mlir-aie` wheels version, the github repo commit, and the requirements versions.
+   If you install from something other than the latest wheels, make sure
    you use the repo commit -- and installation instructions -- from that point in time.
 
    1. **Latest:** For the latest wheels (not necessarily a release):

cmake/install_headers.cmake

@@ -13,8 +13,10 @@ function(install_headers SRCPATH BUILDPATH INSTALLPATH HEADERS_NAME)
 
   message("Copying ${HEADERS_NAME} includes from ${SRCPATH} to ${BUILDPATH}/${HEADERS_NAME}")
 
-  # copy header files into build area
-  file(GLOB_RECURSE headers_to_copy ${SRCPATH}/*.h ${SRCPATH}/*.hpp)
+  # Include .cc/.cpp so build/include matches install/include for in-tree
+  # tests that resolve kernel sources via cxx_header_path().
+  file(GLOB_RECURSE headers_to_copy
+       ${SRCPATH}/*.h ${SRCPATH}/*.hpp ${SRCPATH}/*.cc ${SRCPATH}/*.cpp)
   foreach(header ${headers_to_copy})
       file(RELATIVE_PATH rel_path ${SRCPATH} ${header})
 

docs/Building.md

@@ -1,6 +1,6 @@
 # Building the MLIR-AIE Codebase on Linux
 
-These instructions will guide you through everything required for building and executing a program on the Ryzen™ AI NPU, starting from a fresh bare-bones **Ubuntu 24.04** or **Ubuntu 24.10** install. It is possible to use **Ubuntu 22.04** however you must follow the documentation on the [xdna-driver](https://github.com/amd/xdna-driver) repository to configure the Linux kernel, driver and runtime for deployment. 
+These instructions will guide you through everything required for building and executing a program on the Ryzen™ AI NPU, starting from a fresh bare-bones **Ubuntu 24.04** or **Ubuntu 24.10** install. It is possible to use **Ubuntu 22.04** however you must follow the documentation on the [xdna-driver](https://github.com/amd/xdna-driver) repository to configure the Linux kernel, driver and runtime for deployment.
 
 ## Initial Setup
 
@@ -9,7 +9,7 @@ These instructions will guide you through everything required for building and e
 If starting from `Ubuntu 24.04` you may need to update the Linux kernel to 6.11+ by installing the Hardware Enablement (HWE) stack:
 
   ```bash
-  sudo apt update 
+  sudo apt update
   sudo apt install --install-recommends linux-generic-hwe-24.04
   sudo reboot
   ```
@@ -124,7 +124,7 @@ xrt-smi examine
 
 1. Install required Python packages:
    ```bash
-   # Install basic Python requirements 
+   # Install basic Python requirements
    python3 -m pip install -r python/requirements.txt
    ```
 
@@ -204,7 +204,7 @@ If the [upstream packages](#install-from-upstream-packages-ubuntu-2404) do not s
 ### Update BIOS:
 
 Be sure you have the latest BIOS for your laptop or mini PC, this will ensure the NPU (sometimes referred to as IPU) is enabled in the system. You may need to manually enable the NPU:
-   ```Advanced → CPU Configuration → IPU``` 
+   ```Advanced → CPU Configuration → IPU```
 
 > **NOTE:** Some manufacturers only provide Windows executables to update the BIOS, please do this before installing Ubuntu.
 

docs/README.md

@@ -108,8 +108,8 @@ Turn off SecureBoot (Allows for unsigned drivers to be installed):
 1. Install IRON library by installing the `mlir-aie` wheels:
 
    For installing the `mlir-aie` wheels, there are 3 options. Note that for whichever path you take,
-   it is important to sync the `mlir-aie` wheels version, the github repo commit, and the requirements versions. 
-   If you install from something other than the latest wheels, make sure 
+   it is important to sync the `mlir-aie` wheels version, the github repo commit, and the requirements versions.
+   If you install from something other than the latest wheels, make sure
    you use the repo commit -- and installation instructions -- from that point in time.
 
    1. **Latest:** For the latest wheels (not necessarily a release):

programming_examples/README.md

@@ -4,17 +4,15 @@
 // See https://llvm.org/LICENSE.txt for license information.
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
-// Copyright (C) 2024, Advanced Micro Devices, Inc.
+// Copyright (C) 2024-2026, Advanced Micro Devices, Inc.
 // 
 //===----------------------------------------------------------------------===//-->
 
 # <ins>Programming Examples</ins>
 
 These programming examples are provided so that application programmers can learn how to leverage the IRON design flow with mlir-aie python bindings, and the mlir-aie intermediate representation directly to build applications targeting AI Engines. 
 
-Each IRON example has one or more implementations:
-* `<example_name>.py` - These designs are generally written using a higher-level version of IRON
-* `<example_name>_placed.py` - These designs are generally written using a lower-level verion of IRON
+Most examples are a single `<example_name>.py` design driven by `@iron.jit` — one file describes the AIE-array dataflow, JIT-compiles to xclbin/insts, and runs end-to-end (or feeds the prebuilt artifacts to a C++ host).  A few examples additionally provide an `<example_name>_placed.py` variant written against a lower-level form of IRON for the cases where explicit tile/core placement is the pedagogical point.
 
 They are organized into the following directories:
 

programming_examples/algorithms/for_each.py

@@ -35,7 +35,13 @@ def main():
     initial_tensor = tensor.numpy().copy()
 
     # JIT compile the algorithm
-    iron.jit(for_each)(lambda a: a + 1, tensor, tile_size=16)
+    iron.jit(for_each)(
+        tensor,
+        func=lambda a: a + 1,
+        N=int(tensor.shape[0]),
+        dtype=tensor.dtype,
+        tile_size=16,
+    )
 
     # Check the correctness of the result
     e = np.equal(initial_tensor + 1, tensor.numpy())

programming_examples/algorithms/transform.py

@@ -36,7 +36,14 @@ def main():
     output = iron.zeros_like(input)
 
     # JIT compile the algorithm
-    iron.jit(transform)(lambda a: a + 1, input, output, tile_size=16)
+    iron.jit(transform)(
+        input,
+        output,
+        func=lambda a: a + 1,
+        N=int(input.shape[0]),
+        dtype=input.dtype,
+        tile_size=16,
+    )
 
     # Check the correctness of the result
     e = np.equal(input.numpy() + 1, output.numpy())

programming_examples/algorithms/transform_binary.py

@@ -36,7 +36,15 @@ def main():
     output = iron.zeros_like(input0)
 
     # JIT compile the algorithm
-    iron.jit(transform_binary)(lambda a, b: a + b, input0, input1, output, tile_size=16)
+    iron.jit(transform_binary)(
+        input0,
+        input1,
+        output,
+        func=lambda a, b: a + b,
+        N=int(input0.shape[0]),
+        dtype=input0.dtype,
+        tile_size=16,
+    )
 
     # Check the correctness of the result
     e = np.equal(input0.numpy() + input1.numpy(), output.numpy())

programming_examples/algorithms/transform_parallel.py

@@ -36,7 +36,14 @@ def main():
     output = iron.zeros_like(input)
 
     # JIT compile the algorithm
-    iron.jit(transform_parallel)(lambda a: a + 1, input, output, tile_size=16)
+    iron.jit(transform_parallel)(
+        input,
+        output,
+        func=lambda a: a + 1,
+        N=int(input.shape[0]),
+        dtype=input.dtype,
+        tile_size=16,
+    )
 
     # Check the correctness of the result
     e = np.equal(input.numpy() + 1, output.numpy())