Add Rocm execution provider for AMD GPUs

.github/workflows/linux-rocm.yaml

@@ -0,0 +1,205 @@
+name: linux-rocm
+
+on:
+  push:
+    branches:
+      - master
+      - rocm
+    tags:
+      - 'v[0-9]+.[0-9]+.[0-9]+*'
+    paths:
+      - '.github/workflows/linux-rocm.yaml'
+      - '.github/scripts/test-online-transducer.sh'
+      - '.github/scripts/test-online-paraformer.sh'
+      - '.github/scripts/test-offline-transducer.sh'
+      - '.github/scripts/test-offline-ctc.sh'
+      - '.github/scripts/test-online-ctc.sh'
+      - '.github/scripts/test-offline-tts.sh'
+      - 'CMakeLists.txt'
+      - 'cmake/**'
+      - 'sherpa-onnx/csrc/*'
+      - 'sherpa-onnx/c-api/*'
+      - 'c-api-examples/**'
+  pull_request:
+    branches:
+      - master
+    paths:
+      - '.github/workflows/linux-rocm.yaml'
+      - '.github/scripts/test-online-transducer.sh'
+      - '.github/scripts/test-online-paraformer.sh'
+      - '.github/scripts/test-offline-transducer.sh'
+      - '.github/scripts/test-offline-ctc.sh'
+      - '.github/scripts/test-online-ctc.sh'
+      - '.github/scripts/test-online-ctc.sh'
+      - '.github/scripts/test-offline-tts.sh'
+      - 'CMakeLists.txt'
+      - 'cmake/**'
+      - 'sherpa-onnx/csrc/*'
+      - 'sherpa-onnx/c-api/*'
+
+  workflow_dispatch:
+
+concurrency:
+  group: linux-rocm-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  linux_rocm:
+    runs-on: ${{ matrix.os }}
+    name: ${{ matrix.build_type }}
+    strategy:
+      fail-fast: false
+      matrix:
+        os: [ubuntu-latest]
+        # build_type: [Release, Debug]
+        build_type: [Release]
+
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+
+      - name: ccache
+        uses: hendrikmuhs/[email protected]
+        with:
+          key: ${{ matrix.os }}-${{ matrix.build_type }}-rocm
+
+      - name: Configure CMake
+        shell: bash
+        run: |
+          export CMAKE_CXX_COMPILER_LAUNCHER=ccache
+          export PATH="/usr/lib/ccache:/usr/local/opt/ccache/libexec:$PATH"
+          cmake --version
+
+          mkdir build
+          cd build
+          cmake \
+            -D CMAKE_BUILD_TYPE=${{ matrix.build_type }} \
+            -D CMAKE_INSTALL_PREFIX=./install \
+            -D BUILD_SHARED_LIBS=ON \
+            -D SHERPA_ONNX_ENABLE_ROCM=ON \
+            ..
+
+      - name: Build sherpa-onnx for ubuntu
+        shell: bash
+        run: |
+          export PATH="/usr/lib/ccache:/usr/local/opt/ccache/libexec:$PATH"
+
+          cd build
+          make -j2
+          make install
+
+          ls -lh lib
+          ls -lh bin
+
+          echo "----"
+          ls -lh install/lib
+
+          echo "----"
+          ls -lh install/bin
+
+      - name: Display dependencies of sherpa-onnx for linux
+        shell: bash
+        run: |
+          file build/bin/sherpa-onnx
+          readelf -d build/bin/sherpa-onnx
+          echo "----"
+          ldd build/bin/sherpa-onnx
+
+      - name: Test spoken language identification
+        shell: bash
+        run: |
+          export PATH=$PWD/build/bin:$PATH
+          export EXE=sherpa-onnx-offline-language-identification
+
+          .github/scripts/test-spoken-language-identification.sh
+
+      - name: Test online CTC
+        shell: bash
+        run: |
+          export PATH=$PWD/build/bin:$PATH
+          export EXE=sherpa-onnx
+
+          .github/scripts/test-online-ctc.sh
+
+      - name: Test offline TTS
+        shell: bash
+        run: |
+          export PATH=$PWD/build/bin:$PATH
+          export EXE=sherpa-onnx-offline-tts
+
+          .github/scripts/test-offline-tts.sh
+
+      - name: Test online paraformer
+        shell: bash
+        run: |
+          export PATH=$PWD/build/bin:$PATH
+          export EXE=sherpa-onnx
+
+          .github/scripts/test-online-paraformer.sh
+
+
+      - name: Test offline Whisper
+        shell: bash
+        run: |
+          export PATH=$PWD/build/bin:$PATH
+          export EXE=sherpa-onnx-offline
+
+          .github/scripts/test-offline-whisper.sh
+
+      - name: Test offline CTC
+        shell: bash
+        run: |
+          export PATH=$PWD/build/bin:$PATH
+          export EXE=sherpa-onnx-offline
+
+          .github/scripts/test-offline-ctc.sh
+
+      - name: Test offline transducer
+        shell: bash
+        run: |
+          export PATH=$PWD/build/bin:$PATH
+          export EXE=sherpa-onnx-offline
+
+          .github/scripts/test-offline-transducer.sh
+
+      - name: Test online transducer
+        shell: bash
+        run: |
+          export PATH=$PWD/build/bin:$PATH
+          export EXE=sherpa-onnx
+
+          .github/scripts/test-online-transducer.sh
+
+      - name: Test online transducer (C API)
+        shell: bash
+        run: |
+          export PATH=$PWD/build/bin:$PATH
+          export EXE=decode-file-c-api
+
+          .github/scripts/test-online-transducer.sh
+
+      - name: Copy files
+        if: (github.repository_owner == 'csukuangfj' || github.repository_owner == 'k2-fsa') && github.event_name == 'push' && contains(github.ref, 'refs/tags/')
+        shell: bash
+        run: |
+          SHERPA_ONNX_VERSION=v$(grep "SHERPA_ONNX_VERSION" ./CMakeLists.txt  | cut -d " " -f 2  | cut -d '"' -f 2)
+
+          dst=sherpa-onnx-${SHERPA_ONNX_VERSION}-linux-x64-rocm
+          mkdir $dst
+
+          cp -a build/install/bin $dst/
+          cp -a build/install/lib $dst/
+          cp -a build/install/include $dst/
+
+          tree $dst
+
+          tar cjvf ${dst}.tar.bz2 $dst
+
+      - name: Release pre-compiled binaries and libs for linux x64
+        if: (github.repository_owner == 'csukuangfj' || github.repository_owner == 'k2-fsa') && github.event_name == 'push' && contains(github.ref, 'refs/tags/')
+        uses: svenstaro/upload-release-action@v2
+        with:
+          file_glob: true
+          overwrite: true
+          file: sherpa-onnx-*linux-x64-rocm.tar.bz2

CMakeLists.txt

@@ -29,7 +29,8 @@ option(SHERPA_ONNX_ENABLE_PORTAUDIO "Whether to build with portaudio" ON)
 option(SHERPA_ONNX_ENABLE_JNI "Whether to build JNI internface" OFF)
 option(SHERPA_ONNX_ENABLE_C_API "Whether to build C API" ON)
 option(SHERPA_ONNX_ENABLE_WEBSOCKET "Whether to build webscoket server/client" ON)
-option(SHERPA_ONNX_ENABLE_GPU "Enable ONNX Runtime GPU support" OFF)
+option(SHERPA_ONNX_ENABLE_GPU "Enable ONNX Runtime NVIDIA GPU support" OFF)
+option(SHERPA_ONNX_ENABLE_ROCM "Enable ONNX Runtime AMD GPU support" OFF)
 option(SHERPA_ONNX_ENABLE_WASM "Whether to enable WASM" OFF)
 option(SHERPA_ONNX_ENABLE_WASM_TTS "Whether to enable WASM for TTS" OFF)
 option(SHERPA_ONNX_ENABLE_WASM_ASR "Whether to enable WASM for ASR" OFF)
@@ -82,6 +83,10 @@ if(SHERPA_ONNX_ENABLE_JNI AND NOT BUILD_SHARED_LIBS)
 endif()
 
 if(SHERPA_ONNX_ENABLE_GPU)
+  if(SHERPA_ONNX_ENABLE_ROCM)
+    message(FATAL_ERROR "Both SHERPA_ONNX_ENABLE_GPU and SHERPA_ONNX_ENABLE_ROCM are ON. Please set at most one of them to ON.")
+  endif()
+
   message(WARNING "\
 Compiling for NVIDIA GPU is enabled. Please make sure cudatoolkit
 is installed on your system. Otherwise, you will get errors at runtime.
@@ -94,6 +99,16 @@ to install CUDA toolkit if you have not installed it.")
   endif()
 endif()
 
+if(SHERPA_ONNX_ENABLE_ROCM)
+  message(WARNING "\
+Compiling for AMD GPU is enabled. Please make sure ROCm
+is installed on your system. Otherwise, you will get errors at runtime.")
+  if(NOT BUILD_SHARED_LIBS)
+    message(STATUS "Set BUILD_SHARED_LIBS to ON since SHERPA_ONNX_ENABLE_ROCM is ON")
+    set(BUILD_SHARED_LIBS ON CACHE BOOL "" FORCE)
+  endif()
+endif()
+
 # see https://cmake.org/cmake/help/latest/prop_tgt/MSVC_RUNTIME_LIBRARY.html
 # https://stackoverflow.com/questions/14172856/compile-with-mt-instead-of-md-using-cmake
 if(MSVC)
@@ -117,6 +132,7 @@ message(STATUS "SHERPA_ONNX_ENABLE_JNI ${SHERPA_ONNX_ENABLE_JNI}")
 message(STATUS "SHERPA_ONNX_ENABLE_C_API ${SHERPA_ONNX_ENABLE_C_API}")
 message(STATUS "SHERPA_ONNX_ENABLE_WEBSOCKET ${SHERPA_ONNX_ENABLE_WEBSOCKET}")
 message(STATUS "SHERPA_ONNX_ENABLE_GPU ${SHERPA_ONNX_ENABLE_GPU}")
+message(STATUS "SHERPA_ONNX_ENABLE_ROCM ${SHERPA_ONNX_ENABLE_ROCM}")
 message(STATUS "SHERPA_ONNX_ENABLE_WASM ${SHERPA_ONNX_ENABLE_WASM}")
 message(STATUS "SHERPA_ONNX_ENABLE_WASM_TTS ${SHERPA_ONNX_ENABLE_WASM_TTS}")
 message(STATUS "SHERPA_ONNX_ENABLE_WASM_ASR ${SHERPA_ONNX_ENABLE_WASM_ASR}")

cmake/onnxruntime-linux-x86_64-rocm.cmake

@@ -0,0 +1,101 @@
+# Copyright (c)  2022-2023  Xiaomi Corporation
+message(STATUS "CMAKE_SYSTEM_NAME: ${CMAKE_SYSTEM_NAME}")
+message(STATUS "CMAKE_SYSTEM_PROCESSOR: ${CMAKE_SYSTEM_PROCESSOR}")
+
+if(NOT CMAKE_SYSTEM_NAME STREQUAL Linux)
+  message(FATAL_ERROR "This file is for Linux only. Given: ${CMAKE_SYSTEM_NAME}")
+endif()
+
+if(NOT CMAKE_SYSTEM_PROCESSOR STREQUAL x86_64)
+  message(FATAL_ERROR "This file is for x86_64 only. Given: ${CMAKE_SYSTEM_PROCESSOR}")
+endif()
+
+if(NOT BUILD_SHARED_LIBS)
+  message(FATAL_ERROR "This file is for building shared libraries. BUILD_SHARED_LIBS: ${BUILD_SHARED_LIBS}")
+endif()
+
+if(NOT SHERPA_ONNX_ENABLE_ROCM)
+  message(FATAL_ERROR "This file is for AMD GPU only. Given SHERPA_ONNX_ENABLE_ROCM: ${SHERPA_ONNX_ENABLE_ROCM}")
+endif()
+
+set(onnxruntime_URL  "https://github.com/csukuangfj/onnxruntime-libs/releases/download/v1.18.1/onnxruntime-linux-x64-rocm-Release-1.18.1.zip")
+set(onnxruntime_URL2 "https://hub.nuaa.cf/csukuangfj/onnxruntime-libs/releases/download/v1.18.1/onnxruntime-linux-x64-rocm-Release-1.18.1.zip")
+set(onnxruntime_HASH "SHA256=fe6674d0d4d72d9361667de2ada44a81591c3769c63a87421636317590be659e")
+
+# If you don't have access to the Internet,
+# please download onnxruntime to one of the following locations.
+# You can add more if you want.
+set(possible_file_locations
+  $ENV{HOME}/Downloads/onnxruntime-linux-x64-rocm-Release-1.18.1.zip
+  ${CMAKE_SOURCE_DIR}/onnxruntime-linux-x64-rocm-Release-1.18.1.zip
+  ${CMAKE_BINARY_DIR}/onnxruntime-linux-x64-rocm-Release-1.18.1.zip
+  /tmp/onnxruntime-linux-x64-rocm-Release-1.18.1.zip
+  /star-fj/fangjun/download/github/onnxruntime-linux-x64-rocm-Release-1.18.1.zip
+)
+
+foreach(f IN LISTS possible_file_locations)
+  if(EXISTS ${f})
+    set(onnxruntime_URL  "${f}")
+    file(TO_CMAKE_PATH "${onnxruntime_URL}" onnxruntime_URL)
+    message(STATUS "Found local downloaded onnxruntime: ${onnxruntime_URL}")
+    set(onnxruntime_URL2)
+    break()
+  endif()
+endforeach()
+
+FetchContent_Declare(onnxruntime
+  URL
+    ${onnxruntime_URL}
+    ${onnxruntime_URL2}
+  URL_HASH          ${onnxruntime_HASH}
+)
+
+FetchContent_GetProperties(onnxruntime)
+if(NOT onnxruntime_POPULATED)
+  message(STATUS "Downloading onnxruntime from ${onnxruntime_URL}")
+  FetchContent_Populate(onnxruntime)
+endif()
+message(STATUS "onnxruntime is downloaded to ${onnxruntime_SOURCE_DIR}")
+
+find_library(location_onnxruntime onnxruntime
+  PATHS
+  "${onnxruntime_SOURCE_DIR}/lib"
+  NO_CMAKE_SYSTEM_PATH
+)
+
+message(STATUS "location_onnxruntime: ${location_onnxruntime}")
+
+add_library(onnxruntime SHARED IMPORTED)
+
+set_target_properties(onnxruntime PROPERTIES
+  IMPORTED_LOCATION ${location_onnxruntime}
+  INTERFACE_INCLUDE_DIRECTORIES "${onnxruntime_SOURCE_DIR}/include"
+)
+
+find_library(location_onnxruntime_rocm_lib onnxruntime_providers_rocm
+  PATHS
+  "${onnxruntime_SOURCE_DIR}/lib"
+  NO_CMAKE_SYSTEM_PATH
+)
+
+add_library(onnxruntime_providers_rocm SHARED IMPORTED)
+set_target_properties(onnxruntime_providers_rocm PROPERTIES
+  IMPORTED_LOCATION ${location_onnxruntime_rocm_lib}
+)
+message(STATUS "location_onnxruntime_rocm_lib: ${location_onnxruntime_rocm_lib}")
+
+# for libonnxruntime_providers_shared.so
+find_library(location_onnxruntime_providers_shared_lib onnxruntime_providers_shared
+  PATHS
+  "${onnxruntime_SOURCE_DIR}/lib"
+  NO_CMAKE_SYSTEM_PATH
+)
+add_library(onnxruntime_providers_shared SHARED IMPORTED)
+set_target_properties(onnxruntime_providers_shared PROPERTIES
+  IMPORTED_LOCATION ${location_onnxruntime_providers_shared_lib}
+)
+message(STATUS "location_onnxruntime_providers_shared_lib: ${location_onnxruntime_providers_shared_lib}")
+
+file(GLOB onnxruntime_lib_files "${onnxruntime_SOURCE_DIR}/lib/libonnxruntime*")
+message(STATUS "onnxruntime lib files: ${onnxruntime_lib_files}")
+install(FILES ${onnxruntime_lib_files} DESTINATION lib)

cmake/onnxruntime.cmake

@@ -27,6 +27,8 @@ function(download_onnxruntime)
   elseif(CMAKE_SYSTEM_NAME STREQUAL Linux AND CMAKE_SYSTEM_PROCESSOR STREQUAL x86_64)
     if(SHERPA_ONNX_ENABLE_GPU)
       include(onnxruntime-linux-x86_64-gpu)
+    elseif(SHERPA_ONNX_ENABLE_ROCM)
+      include(onnxruntime-linux-x86_64-rocm)
     elseif(BUILD_SHARED_LIBS)
       include(onnxruntime-linux-x86_64)
     else()

sherpa-onnx/csrc/CMakeLists.txt

@@ -190,6 +190,13 @@ if(SHERPA_ONNX_ENABLE_GPU)
   )
 endif()
 
+if(SHERPA_ONNX_ENABLE_ROCM)
+  target_link_libraries(sherpa-onnx-core
+    onnxruntime_providers_rocm
+    onnxruntime_providers_shared
+  )
+endif()
+
 if(BUILD_SHARED_LIBS)
   target_link_libraries(sherpa-onnx-core onnxruntime)
 else()