[Iluvatar][TLE] Triton v3.6.x iluvatar backend and 5 TLE primitives support

.github/workflows/iluvatar3.6-build-and-test.yml

@@ -0,0 +1,86 @@
+name: Iluvatar3.6-Build-And-Test
+
+on:
+  push:
+    branches: [ "triton_v3.6.x" ]
+  pull_request:
+    branches: [ "triton_v3.6.x" ]
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  iluvatar-build-and-test:
+    runs-on: iluvatar3.6
+    if: ${{ github.repository == 'FlagTree/flagtree' || github.repository == 'flagos-ai/flagtree' }}
+    steps:
+      - name: Setup environment
+        shell: bash
+        run: |
+          source ~/env.sh
+          env | grep -E '^(http_proxy|https_proxy|all_proxy|no_proxy)=' >> $GITHUB_ENV || true
+
+      - name: Smart Checkout
+        uses: flagos-ai/FlagTree/.github/actions/smart-checkout@main
+        with:
+          checkout_version: 'v6'
+
+      - name: Check if backend-relevant files changed
+        id: check_backend
+        uses: flagos-ai/FlagTree/.github/actions/check-backend-changed@main
+        with:
+          backend: iluvatar
+
+      - name: FlagTree Build on Iluvatar
+        if: steps.check_backend.outputs.should_skip != 'true'
+        shell: bash
+        run: |
+          set -x
+          export FLAGTREE_BACKEND=iluvatar
+          export TRITON_OFFLINE_BUILD=1
+          export TRITON_BUILD_PROTON=OFF
+          export TRITON_ILU_BUILD_GLUON=OFF
+          MAX_JOBS=32 python3 -m pip install . --no-build-isolation
+
+      - name: FlagTree Test on Iluvatar
+        if: steps.check_backend.outputs.should_skip != 'true'
+        shell: bash
+        run: |
+          set -x
+          source ~/env.sh
+
+          pip3 install "pytest==9.1.1"
+
+          for pkg in hypothesis absl-py scipy pytest-forked; do
+            pip3 list "$pkg" | grep "$pkg" || pip3 install "$pkg"
+          done
+
+          if [[ "$(uname -m)" == "aarch64" ]]; then
+            libgomp_path=$(find /usr/lib /usr/lib64 /lib /lib64 -type f -name 'libgomp.so*' 2>/dev/null | head -n 1)
+            if [[ -n "$libgomp_path" ]]; then
+              export LD_PRELOAD="$libgomp_path${LD_PRELOAD:+:$LD_PRELOAD}"
+            fi
+          fi
+
+          # triton unit test
+          UMD_CUDAMODULELOADING=0 pytest -v third_party/iluvatar/python/test/unit/language/test_core.py
+          pytest -v third_party/iluvatar/python/test/unit/language/test_subprocess.py
+          pytest -v third_party/iluvatar/python/test/unit/operators/test_blocksparse.py
+          pytest -v third_party/iluvatar/python/test/unit/operators/test_cross_entropy.py
+          pytest -v third_party/iluvatar/python/test/unit/operators/test_dot_trans.py
+          pytest -v third_party/iluvatar/python/test/unit/operators/test_flash_attention.py
+          pytest -v third_party/iluvatar/python/test/unit/operators/test_inductor.py
+          pytest -v third_party/iluvatar/python/test/unit/operators/test_matmul.py
+          pytest -v third_party/iluvatar/python/test/unit/test_debug.py
+
+          # TLE unit test
+          pytest -v python/test/tle/integration/test_tle_local_store.py
+          pytest -v python/test/tle/unit/test_tle_gpu_local_ptr.py
+          pytest -v python/test/tle/unit/test_extract_tile_static_index.py
+          pytest -v python/test/tle/unit/test_extract_tile_dynamic_index.py
+          pytest -v python/test/tle/unit/test_insert_tile_static_index.py
+          pytest -v python/test/tle/unit/test_insert_tile_dynamic_index.py
+
+          # TLE tutorials unit tests
+          python3 python/tutorials/tle/01-fft.py --only_unit_test

.github/workflows/iluvatar3.6-delivery.yml

@@ -0,0 +1,44 @@
+name: CD-Iluvatar3.6
+
+on:
+  workflow_dispatch:
+    inputs:
+      WHL_VER:
+        description: '0.6.0.dev1 < 0.6.0a1 < 0.6.0b1 < 0.6.0rc1 < 0.6.0 < 0.6.0.post1'
+        required: true
+        default: '0.6.0'
+      PYTHON_VER:
+        required: true
+        default: '3.10'
+      PLATFORM:
+        required: true
+        default: 'x86_64'
+      FLAGTREE_PYPI_KEY:
+        required: false
+  workflow_call:
+    inputs:
+      WHL_VER:
+        type: string
+        required: true
+        default: '0.6.0'
+      PYTHON_VER:
+        type: string
+        required: true
+        default: '3.10'
+      PLATFORM:
+        type: string
+        required: true
+        default: 'x86_64'
+      FLAGTREE_PYPI_KEY:
+        type: string
+        required: false
+
+jobs:
+  call-delivery:
+    uses: flagos-ai/flagtree/.github/workflows/iluvatar3.6-delivery.yml@main
+    with:
+      WHL_VER: ${{ inputs.WHL_VER }}
+      PYTHON_VER: ${{ inputs.PYTHON_VER }}
+      PLATFORM: ${{ inputs.PLATFORM }}
+      FLAGTREE_PYPI_KEY: ${{ inputs.FLAGTREE_PYPI_KEY }}
+    secrets: inherit

.gitignore

@@ -32,6 +32,12 @@ third_party/xpu/python/*.egg-info
 third_party/nvidia/backend/flagcx*
 third_party/nvidia/backend/lib/libflagcx*
 
+# Backends iluvatar
+third_party/iluvatar/python/*.egg-info
+third_party/iluvatar/.triton/
+third_party/iluvatar/bin/FileCheck
+third_party/iluvatar/logs/
+
 # Backends copied from submodules
 python/triton/backends/*
 !python/triton/backends/__init__.py

CMakeLists.txt

@@ -37,8 +37,11 @@ if(NOT FLAGTREE_BACKEND)
   add_definitions(-D__AMD__)
 elseif(FLAGTREE_BACKEND STREQUAL "iluvatar")
   add_definitions(-D__ILUVATAR__)
-  remove_definitions(-D_GLIBCXX_USE_CXX11_ABI=1)
-  add_definitions(-D_GLIBCXX_USE_CXX11_ABI=0)
+  set(FLAGTREE_TLE OFF)
+  set(FLAGTREE_ILUVATAR_TLE ON)
+  add_definitions(-D__ILUVATAR_TLE__)
+  remove_definitions(-D__TLE__)
+  list(REMOVE_ITEM LLVM_TABLEGEN_FLAGS -D__TLE__)
 elseif(FLAGTREE_BACKEND STREQUAL "mthreads")
   set(ENV{PATH} "$ENV{LLVM_SYSPATH}/bin:$ENV{PATH}")
   set(CMAKE_C_COMPILER clang)
@@ -295,6 +298,8 @@ if(NOT MSVC)
       # Suppress visibility warnings in gluon_ir.cc (GCC 13+ -Wattributes on pybind11 hidden types)
       # Also suppress -Wcomment for generated TritonGPUAttrDefs.h.inc (ASCII diagrams in TableGen output)
       set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Werror -Wno-attributes -Wno-comment -Wno-error=odr")
+    elseif(FLAGTREE_BACKEND STREQUAL "iluvatar")
+      set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Werror -Wno-covered-switch-default -Wno-deprecated-declarations -Wno-attributes")
     else()
       set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Werror -Wno-covered-switch-default")
     endif()
@@ -349,6 +354,20 @@ if (FLAGTREE_BACKEND MATCHES "^(xpu|cambricon|aipu|tsingmicro|enflame|rpu|thrive
       endif()
     endforeach()
   endif()
+elseif(FLAGTREE_BACKEND STREQUAL "iluvatar")
+  set(TRITON_CORE_SOURCE_DIR ${CMAKE_CURRENT_SOURCE_DIR}/third_party/iluvatar)
+  set(TRITON_CORE_BINARY_DIR ${CMAKE_CURRENT_BINARY_DIR}/third_party/iluvatar)
+  option(TRITON_BUILD_GLUON "Build the Gluon IR Python bindings (gluon_ir.cc)" OFF)
+  include_directories(${TRITON_CORE_SOURCE_DIR}/include)
+  include_directories(${TRITON_CORE_BINARY_DIR}/include)
+  include_directories(${TRITON_CORE_SOURCE_DIR}/backend/include)
+  include_directories(${TRITON_CORE_BINARY_DIR}/backend/include)
+  if(FLAGTREE_ILUVATAR_TLE)
+    include_directories(${TRITON_CORE_SOURCE_DIR}/tle/include)
+    include_directories(${TRITON_CORE_BINARY_DIR}/tle/include)
+  endif()
+  add_subdirectory(${TRITON_CORE_SOURCE_DIR}/include ${TRITON_CORE_BINARY_DIR}/include)
+  add_subdirectory(${TRITON_CORE_SOURCE_DIR}/lib ${TRITON_CORE_BINARY_DIR}/lib)
 elseif(NOT FLAGTREE_BACKEND)
   add_subdirectory(include)
   add_subdirectory(lib)
@@ -410,6 +429,11 @@ if(TRITON_BUILD_PYTHON_MODULE)
   elseif(FLAGTREE_BACKEND AND FLAGTREE_BACKEND STREQUAL "mthreads")
     include_directories(${PROJECT_BINARY_DIR}/third_party/${FLAGTREE_BACKEND})
     add_subdirectory(third_party/mthreads/proton/Dialect)
+  elseif(FLAGTREE_BACKEND AND FLAGTREE_BACKEND STREQUAL "iluvatar")
+    if (TRITON_BUILD_PROTON)
+      list(APPEND TRITON_PLUGIN_NAMES "proton")
+      add_subdirectory(third_party/proton/Dialect)
+    endif()
   else()
     list(APPEND TRITON_PLUGIN_NAMES "proton")
     add_subdirectory(third_party/proton/Dialect)
@@ -621,24 +645,42 @@ if(TRITON_BUILD_PYTHON_MODULE)
 
   set(TRITON_BACKENDS_TUPLE "(${TRITON_BACKENDS_TUPLE})")
   add_compile_definitions(TRITON_BACKENDS_TUPLE=${TRITON_BACKENDS_TUPLE})
-  add_library(triton SHARED ${PYTHON_SRC_PATH}/main.cc
-                  ${PYTHON_SRC_PATH}/ir.cc
-                  ${PYTHON_SRC_PATH}/passes.cc
-                  ${PYTHON_SRC_PATH}/interpreter.cc
-                  ${PYTHON_SRC_PATH}/llvm.cc)
-  if(EXISTS "${PYTHON_SRC_PATH}/gluon_ir.cc")
-    target_sources(triton PRIVATE ${PYTHON_SRC_PATH}/gluon_ir.cc)
-  endif()
-  if(EXISTS "${PYTHON_SRC_PATH}/linear_layout.cc")
-    target_sources(triton PRIVATE ${PYTHON_SRC_PATH}/linear_layout.cc)
-  endif()
-  if(EXISTS "${PYTHON_SRC_PATH}/specialize.cc")
-    target_sources(triton PRIVATE ${PYTHON_SRC_PATH}/specialize.cc)
-  endif()
-  if(FLAGTREE_BACKEND STREQUAL "xpu")
-    target_sources(triton PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/third_party/xpu/python/src/mlir_pass_abi_shim.cc)
-    add_dependencies(triton TritonAMDGPUTableGen TritonAMDGPUAttrDefsIncGen)
-    target_link_libraries(triton PRIVATE TritonAMDGPUIR TritonAMDUtils)
+  if(FLAGTREE_BACKEND STREQUAL "iluvatar")
+    set(TRITON_ILU_PYTHON_SRCS
+                    ${PYTHON_SRC_PATH}/main.cc
+                    ${PYTHON_SRC_PATH}/ir.cc
+                    ${PYTHON_SRC_PATH}/linear_layout.cc
+                    ${PYTHON_SRC_PATH}/passes.cc
+                    ${PYTHON_SRC_PATH}/interpreter.cc
+                    ${PYTHON_SRC_PATH}/llvm.cc
+                    ${PYTHON_SRC_PATH}/specialize.cc)
+    if(TRITON_BUILD_GLUON)
+      list(APPEND TRITON_ILU_PYTHON_SRCS ${PYTHON_SRC_PATH}/gluon_ir.cc)
+    endif()
+    add_library(triton SHARED ${TRITON_ILU_PYTHON_SRCS})
+    if(TRITON_BUILD_GLUON)
+      target_compile_definitions(triton PRIVATE TRITON_BUILD_GLUON)
+    endif()
+  else()
+    add_library(triton SHARED ${PYTHON_SRC_PATH}/main.cc
+                    ${PYTHON_SRC_PATH}/ir.cc
+                    ${PYTHON_SRC_PATH}/passes.cc
+                    ${PYTHON_SRC_PATH}/interpreter.cc
+                    ${PYTHON_SRC_PATH}/llvm.cc)
+    if(EXISTS "${PYTHON_SRC_PATH}/gluon_ir.cc")
+      target_sources(triton PRIVATE ${PYTHON_SRC_PATH}/gluon_ir.cc)
+    endif()
+    if(EXISTS "${PYTHON_SRC_PATH}/linear_layout.cc")
+      target_sources(triton PRIVATE ${PYTHON_SRC_PATH}/linear_layout.cc)
+    endif()
+    if(EXISTS "${PYTHON_SRC_PATH}/specialize.cc")
+      target_sources(triton PRIVATE ${PYTHON_SRC_PATH}/specialize.cc)
+    endif()
+    if(FLAGTREE_BACKEND STREQUAL "xpu")
+      target_sources(triton PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/third_party/xpu/python/src/mlir_pass_abi_shim.cc)
+      add_dependencies(triton TritonAMDGPUTableGen TritonAMDGPUAttrDefsIncGen)
+      target_link_libraries(triton PRIVATE TritonAMDGPUIR TritonAMDUtils)
+    endif()
   endif()
 
   # Link triton with its dependencies
@@ -714,6 +756,12 @@ if(NOT FLAGTREE_BACKEND OR FLAGTREE_BACKEND MATCHES "^(aipu|tsingmicro|enflame|r
     flagtree_add_tle_generated_header_dependencies()
   endif()
   add_subdirectory(test)
+elseif(FLAGTREE_BACKEND STREQUAL "iluvatar")
+  option(FLAGTREE_ILUVATAR_BUILD_BIN "Build third_party/iluvatar/bin tools and lit tests" OFF)
+  if(FLAGTREE_ILUVATAR_BUILD_BIN)
+    add_subdirectory(${TRITON_CORE_SOURCE_DIR}/bin ${TRITON_CORE_BINARY_DIR}/bin)
+    add_subdirectory(test)
+  endif()
 endif()
 
 if(TRITON_BUILD_UT)

python/setup_tools/setup_helper.py

@@ -512,9 +512,9 @@ def uninstall_triton():
 
 # iluvatar
 cache.store(
-    file="iluvatar-llvm18-x86_64",
+    file="iluvatar-llvm22-x86_64",
     condition=("iluvatar" == flagtree_backend),
-    url="https://baai-cp-web.ks3-cn-beijing.ksyuncs.com/trans/iluvatar-llvm18-x86_64_v0.3.0.tar.gz",
+    url="https://baai-cp-web.ks3-cn-beijing.ksyuncs.com/trans/iluvatar-llvm22-x86_64_v0.6.0.tar.gz",
     pre_hook=lambda: check_env('LLVM_SYSPATH'),
     post_hook=set_llvm_env,
 )