[lilac] updated to 1.9.1-4

4 files changed, 110 insertions, 106 deletions

diff --git a/.SRCINFO b/.SRCINFO
index 924dd0acc346..54884c7d0929 100644
--- a/.SRCINFO
+++ b/.SRCINFO
@@ -1,20 +1,10 @@
 pkgbase = python-onnxruntime
 	pkgdesc = Cross-platform, high performance scoring engine for ML models
 	pkgver = 1.9.1
-	pkgrel = 1
+	pkgrel = 4
 	url = https://github.com/microsoft/onnxruntime
 	arch = x86_64
 	license = MIT
-	checkdepends = python-pytest
-	checkdepends = python-pytorch
-	checkdepends = python-h5py
-	checkdepends = python-pandas
-	checkdepends = python-psutil
-	checkdepends = python-tqdm
-	checkdepends = python-sympy
-	checkdepends = python-torchvision
-	checkdepends = tensorboard
-	checkdepends = python-cerberus
 	makedepends = git
 	makedepends = cmake
 	makedepends = gtest
@@ -37,6 +27,7 @@ pkgbase = python-onnxruntime
 	depends = python-onnx
 	depends = python-protobuf
 	depends = openmpi
+	depends = onednn
 	options = !lto
 	source = git+https://github.com/microsoft/onnxruntime#tag=v1.9.1
 	source = git+https://github.com/onnx/onnx.git
@@ -48,6 +39,7 @@ pkgbase = python-onnxruntime
 	source = pytorch_cpuinfo::git+https://github.com/pytorch/cpuinfo.git
 	source = build-fixes.patch
 	source = clang.patch
+	source = system-dnnl.diff
 	sha512sums = SKIP
 	sha512sums = SKIP
 	sha512sums = SKIP
@@ -58,10 +50,12 @@ pkgbase = python-onnxruntime
 	sha512sums = SKIP
 	sha512sums = 685f0235abed6e1277dd0eb9bda56c464d1987fe7fc90a3550e17ec70cc49fd15f34996a0e159f9622c4ca3e6bf29917fe51b7849342531fa2a6808d782f1e06
 	sha512sums = ad94af8bb25744b244c4f82e9a06189741f82b295a88523ca0e8005568fac710c2299d783989457e9cf96ef8da0593fb4f70c8792d416f44ab29d6493e204f13
+	sha512sums = 6735c7aca2ba2f1f2a5286eb064125bf7f2c68a575d572dd157769d15778ff3e717b3a53d696c767748229f23ee6c3a7c82679df1d86283d7c4dd0ec9103ae08
 
 pkgname = python-onnxruntime
 
 pkgname = python-onnxruntime-cuda
+	pkgdesc = Cross-platform, high performance scoring engine for ML models (CUDA execution provider)
 	depends = nsync
 	depends = re2
 	depends = python-flatbuffers
@@ -69,8 +63,8 @@ pkgname = python-onnxruntime-cuda
 	depends = python-onnx
 	depends = python-protobuf
 	depends = openmpi
+	depends = onednn
 	depends = cuda
 	depends = cudnn
 	depends = nccl
-	provides = python-onnxruntime=1.9.1
-	conflicts = python-onnxruntime
+	depends = python-onnxruntime
diff --git a/PKGBUILD b/PKGBUILD
index b7be61267c75..631cd223a318 100644
--- a/PKGBUILD
+++ b/PKGBUILD
@@ -1,16 +1,17 @@
 # Maintainer: Chih-Hsuan Yen <yan12125@gmail.com>
 
 pkgbase=python-onnxruntime
+# Not split DNNL EP to another package as it's needed unconditionally at runtime if built at compile time
+# https://github.com/microsoft/onnxruntime/blob/v1.9.1/onnxruntime/python/onnxruntime_pybind_state.cc#L533
 pkgname=(python-onnxruntime python-onnxruntime-cuda)
 pkgver=1.9.1
 pkgdesc='Cross-platform, high performance scoring engine for ML models'
-pkgrel=1
+pkgrel=4
 arch=(x86_64)
 url='https://github.com/microsoft/onnxruntime'
 license=(MIT)
-depends=(nsync re2 python-flatbuffers python-numpy python-onnx python-protobuf openmpi)
+depends=(nsync re2 python-flatbuffers python-numpy python-onnx python-protobuf openmpi onednn)
 makedepends=(git cmake gtest gmock pybind11 python-setuptools nlohmann-json chrono-date boost eigen flatbuffers cuda cudnn nccl clang)
-checkdepends=(python-pytest python-pytorch python-h5py python-pandas python-psutil python-tqdm python-sympy python-torchvision tensorboard python-cerberus)
 # not de-vendored libraries
 # onnx: needs shared libonnx (https://github.com/onnx/onnx/issues/3030)
 source=("git+https://github.com/microsoft/onnxruntime#tag=v$pkgver"
@@ -22,7 +23,8 @@ source=("git+https://github.com/microsoft/onnxruntime#tag=v$pkgver"
         "git+https://github.com/jarro2783/cxxopts.git"
         "pytorch_cpuinfo::git+https://github.com/pytorch/cpuinfo.git"
         build-fixes.patch
-        clang.patch)
+        clang.patch
+        system-dnnl.diff)
 sha512sums=('SKIP'
             'SKIP'
             'SKIP'
@@ -32,7 +34,8 @@ sha512sums=('SKIP'
             'SKIP'
             'SKIP'
             '685f0235abed6e1277dd0eb9bda56c464d1987fe7fc90a3550e17ec70cc49fd15f34996a0e159f9622c4ca3e6bf29917fe51b7849342531fa2a6808d782f1e06'
-            'ad94af8bb25744b244c4f82e9a06189741f82b295a88523ca0e8005568fac710c2299d783989457e9cf96ef8da0593fb4f70c8792d416f44ab29d6493e204f13')
+            'ad94af8bb25744b244c4f82e9a06189741f82b295a88523ca0e8005568fac710c2299d783989457e9cf96ef8da0593fb4f70c8792d416f44ab29d6493e204f13'
+            '6735c7aca2ba2f1f2a5286eb064125bf7f2c68a575d572dd157769d15778ff3e717b3a53d696c767748229f23ee6c3a7c82679df1d86283d7c4dd0ec9103ae08')
 # CUDA seems not working with LTO
 options+=('!lto')
 
@@ -44,10 +47,7 @@ prepare() {
 
   patch -Np1 -i ../build-fixes.patch
   patch -Np1 -i ../clang.patch
-
-  # 1.9.0 is marked as 1.10.0 https://github.com/microsoft/onnxruntime/blob/v1.9.0/VERSION_NUMBER
-  # Official wheels are not affected, though
-  echo $pkgver > VERSION_NUMBER
+  patch -Np1 -i ../system-dnnl.diff
 
   git submodule init
   for mod in onnx SafeInt optional-lite tensorboard dlpack cxxopts pytorch_cpuinfo; do
@@ -56,82 +56,60 @@ prepare() {
   done
 }
 
-_build() {
-  build_dir=$1
-  shift
-
+build() {
   cd "$srcdir"/onnxruntime
+
+  local cmake_args=(
+    -DCMAKE_INSTALL_PREFIX=/usr
+    -Donnxruntime_ENABLE_PYTHON=ON
+    -Donnxruntime_PREFER_SYSTEM_LIB=ON
+    -Donnxruntime_BUILD_SHARED_LIB=ON
+    -Donnxruntime_ENABLE_TRAINING=ON
+    -Donnxruntime_USE_MPI=ON
+    -Donnxruntime_USE_PREINSTALLED_EIGEN=ON
+    -Donnxruntime_USE_DNNL=ON
+    -Deigen_SOURCE_PATH=/usr/include/eigen3
+  )
+
   # Use protobuf-lite instead of full protobuf to workaround symbol conflicts
   # with onnx; see https://github.com/onnx/onnx/issues/1277 for details.
-  CC=/usr/bin/clang CXX=/usr/bin/clang++ \
-  cmake -B $build_dir -S cmake \
-    -DCMAKE_INSTALL_PREFIX=/usr \
-    -Donnxruntime_ENABLE_PYTHON=ON \
-    -DONNX_CUSTOM_PROTOC_EXECUTABLE=/usr/bin/protoc \
-    -Donnxruntime_PREFER_SYSTEM_LIB=ON \
-    -Donnxruntime_USE_FULL_PROTOBUF=OFF \
-    -Donnxruntime_BUILD_SHARED_LIB=ON \
-    -Donnxruntime_ENABLE_TRAINING=ON \
-    -Donnxruntime_USE_MPI=ON \
-    -Donnxruntime_USE_PREINSTALLED_EIGEN=ON \
-    -Deigen_SOURCE_PATH=/usr/include/eigen3 \
-    "$@"
-
-  cd $build_dir
-  make
-  python ../setup.py build
-}
+  cmake_args+=(
+    -DONNX_CUSTOM_PROTOC_EXECUTABLE=/usr/bin/protoc
+    -Donnxruntime_USE_FULL_PROTOBUF=OFF
+  )
 
-build() {
-  _build build
-
-  # 1. Use clang as GCC does not work. GCC 11 crashes with internal
-  #    compiler errors. GCC 10 does not work as some dependent packages
-  #    (ex: re2) are built with libstdc++ from GCC 11, and thus linking
-  #    onnxruntime with libstdc++ 10 fails.
-  # 2. Redefine ___is_signed to ___is_signed to workaround a regression
+  # 1. Redefine ___is_signed to ___is_signed to workaround a regression
   #    from CUDA 11.3 -> 11.3.1 [1].
+  # 2. Enable parallel builds for NVCC via -t0, which spawns multiple
+  #    cicc and ptxas processes for each nvcc invocation. The number of
+  #    total processes may be much larger than the number of cores - let
+  #    the scheduler handle it.
   # [1] https://forums.developer.nvidia.com/t/182176
-  _build build-cuda \
-    -DCMAKE_CUDA_HOST_COMPILER=/usr/bin/clang \
-    -DCMAKE_CUDA_FLAGS="-D__is_signed=___is_signed" \
-    -DCMAKE_CUDA_ARCHITECTURES="$_CUDA_ARCHITECTURES" \
-    -Donnxruntime_USE_CUDA=ON \
-    -Donnxruntime_CUDA_HOME=/opt/cuda \
-    -Donnxruntime_CUDNN_HOME=/usr \
+  cmake_args+=(
+    -DCMAKE_CUDA_HOST_COMPILER=/usr/bin/clang
+    -DCMAKE_CUDA_FLAGS="-D__is_signed=___is_signed -t0"
+    -DCMAKE_CUDA_ARCHITECTURES="$_CUDA_ARCHITECTURES"
+    -Donnxruntime_USE_CUDA=ON
+    -Donnxruntime_CUDA_HOME=/opt/cuda
+    -Donnxruntime_CUDNN_HOME=/usr
     -Donnxruntime_USE_NCCL=ON
-}
+  )
 
-_check() {
-  # Test models are no longer publicly available [1]
-  # [1] https://github.com/microsoft/onnxruntime/issues/7447
-  GTEST_FILTER='-*ModelTest*' ARGS="--rerun-failed --output-on-failure" make test
-  # launch_test.py seems a script, and orttraining_* include BERT tests, which require the
-  # transformers package, and failed even if the latter is installed.
-
-  # XXX: Some python tests failed (ex: [1]). In those tests, tested ONNX models are
-  # generated on the fly using the onnx python library (ex: [2]). When the latter
-  # is newer than the included onnx submodule, loading a tested ONNX model may fail
-  # as the IR version for a tested model may be higher than the IR version used in
-  # onnxruntime.
-  # [1] https://build.archlinuxcn.org/~imlonghao/log/python-onnxruntime/2021-08-07T12%3A17%3A01.html
-  # [2] https://github.com/microsoft/onnxruntime/blob/v1.8.2/onnxruntime/test/python/quantization/test_op_gemm.py#L28-L76
-  LD_LIBRARY_PATH="$PWD" pytest \
-    --ignore launch_test.py \
-    --ignore orttraining_run_bert_pretrain.py \
-    --ignore orttraining_run_frontend_batch_size_test.py \
-    --ignore transformers
-}
-
-check() {
-  cd "$srcdir"/onnxruntime/build
-  _check
+  # Use clang as GCC does not work. GCC 11 crashes with internal
+  # compiler errors. GCC 10 does not work as some dependent packages
+  # (ex: re2) are built with libstdc++ from GCC 11, and thus linking
+  # onnxruntime with libstdc++ 10 fails.
+  CC=/usr/bin/clang CXX=/usr/bin/clang++ \
+    cmake -B build -S cmake "${cmake_args[@]}" "$@"
 
-  cd "$srcdir"/onnxruntime/build-cuda
-  # _check # requires machines with CUDA-compatible devices
+  cd build
+  make
+  python ../setup.py build
 }
 
-_package() {
+package_python-onnxruntime() {
+  cd onnxruntime/build
+
   make install DESTDIR="$pkgdir"
 
   python ../setup.py install --root="$pkgdir" --skip-build --optimize=1
@@ -143,18 +121,17 @@ _package() {
   done
   # already installed by `make install`, and not useful as this path is not looked up by the linker
   rm -vf "$pkgdir/$PY_ORT_DIR"/capi/libonnxruntime_providers_*
-}
 
-package_python-onnxruntime() {
-  cd onnxruntime/build
-  _package
+  # installed as split packages
+  rm -vf "$pkgdir"/usr/lib/libonnxruntime_providers_cuda.so
 }
 
 package_python-onnxruntime-cuda() {
-  depends+=(cuda cudnn nccl)
-  conflicts=(python-onnxruntime)
-  provides=("python-onnxruntime=$pkgver")
+  depends+=(cuda cudnn nccl python-onnxruntime)
+  pkgdesc+=' (CUDA execution provider)'
 
-  cd onnxruntime/build-cuda
-  _package
+  cd onnxruntime/build
+  install -Dm755 libonnxruntime_providers_cuda.so -t "$pkgdir"/usr/lib
+  install -Ddm755 "$pkgdir"/usr/share/licenses
+  ln -s python-onnxruntime "$pkgdir"/usr/share/licenses/$pkgname
 }
diff --git a/notes.txt b/notes.txt
index cc19868adbdb..62efe9cd463b 100644
--- a/notes.txt
+++ b/notes.txt
@@ -1,10 +1,3 @@
-Merging two builds?
-    Since 1.9.0, "GPU package can be used on both CPU-only and GPU machines" [1]
-    Package size considerations: make libonnxruntime_providers_cuda.so a separate package
-    There will be warnings from the CUDA build on non-CUDA machines
-    /usr/lib/python3.9/site-packages/onnxruntime/capi/onnxruntime_inference_collection.py:352: UserWarning: Deprecation warning. This ORT build has ['CUDAExecutionProvider', 'CPUExecutionProvider'] enabled. The next release (ORT 1.10) will require explicitly setting the providers parameter (as opposed to the current behavior of providers getting set/registered by default based on the build flags) when instantiating InferenceSession.For example, onnxruntime.InferenceSession(..., providers=["CUDAExecutionProvider"], ...)
-      warnings.warn("Deprecation warning. This ORT build has {} enabled. ".format(available_providers) +
-    2021-09-23 23:47:08.573828136 [E:onnxruntime:Default, provider_bridge_ort.cc:944 Get] Failed to load library libonnxruntime_providers_cuda.so with error: libcublas.so.11: cannot open shared object file: No such file or directory
-    2021-09-23 23:47:08.574000870 [E:onnxruntime:Default, provider_bridge_ort.cc:944 Get] Failed to load library libonnxruntime_providers_cuda.so with error: libcublas.so.11: cannot open shared object file: No such file or directory
-
-    [1] https://github.com/microsoft/onnxruntime/releases/tag/v1.9.0
+Build system changes in 1.10
+    Build: respect onnxruntime_PREFER_SYSTEM_LIB for more things https://github.com/microsoft/onnxruntime/pull/9181
+    Remove optional-lite https://github.com/microsoft/onnxruntime/pull/9424
diff --git a/system-dnnl.diff b/system-dnnl.diff
new file mode 100644
index 000000000000..1444d8acaa1e
--- /dev/null
+++ b/system-dnnl.diff
@@ -0,0 +1,40 @@
+diff --git a/cmake/external/dnnl.cmake b/cmake/external/dnnl.cmake
+index 6a51a3d5d..a89635210 100644
+--- a/cmake/external/dnnl.cmake
++++ b/cmake/external/dnnl.cmake
+@@ -26,6 +26,13 @@ elseif(onnxruntime_USE_DNNL AND onnxruntime_DNNL_GPU_RUNTIME STREQUAL "ocl" AND
+ endif()
+ 
+ if (onnxruntime_USE_DNNL)
++if (onnxruntime_PREFER_SYSTEM_LIB)
++  # https://oneapi-src.github.io/oneDNN/dev_guide_transition_to_dnnl.html
++  find_package(dnnl CONFIG REQUIRED)
++  add_library(project_dnnl INTERFACE)
++  add_library(dnnl INTERFACE)
++  target_link_libraries(dnnl INTERFACE DNNL::dnnl)
++else ()
+   set(DNNL_SOURCE ${CMAKE_CURRENT_BINARY_DIR}/dnnl/src/dnnl/src)
+   set(DNNL_INSTALL ${CMAKE_CURRENT_BINARY_DIR}/dnnl/install)
+   set(DNNL_LIB_DIR ${DNNL_INSTALL}/${CMAKE_INSTALL_LIBDIR})
+@@ -55,3 +62,4 @@ if (onnxruntime_USE_DNNL)
+   )
+   link_directories(${DNNL_LIB_DIR})
+ endif()
++endif()
+diff --git a/cmake/onnxruntime_unittests.cmake b/cmake/onnxruntime_unittests.cmake
+index 6bdb2d03c..514faa375 100644
+--- a/cmake/onnxruntime_unittests.cmake
++++ b/cmake/onnxruntime_unittests.cmake
+@@ -744,10 +744,12 @@ add_custom_command(
+ if (NOT onnxruntime_ENABLE_TRAINING_TORCH_INTEROP)
+   if (onnxruntime_USE_DNNL)
+     list(APPEND onnx_test_libs dnnl)
++    if (NOT onnxruntime_PREFER_SYSTEM_LIB)
+     add_custom_command(
+       TARGET ${test_data_target} POST_BUILD
+       COMMAND ${CMAKE_COMMAND} -E copy ${DNNL_DLL_PATH} $<TARGET_FILE_DIR:${test_data_target}>
+       )
++    endif()
+   endif()
+   if(WIN32)
+     if (onnxruntime_USE_TVM)