python-onnxruntime: split C++ and Python libraries & some other changes

* Rename python-onnxruntime-cuda to onnxruntime-cuda - this split
  package actually does not contain Python libraries.
* Move optdepends to the correct package
* Drop flatbuffers fix after upstream adds the compatibility back [1]
* Improve onednn patch - fallback to bundled onednn if the system one
  is missing & wrap other usage of `DNNL_DLL_PATH`
* Add CUDA architectures introduced in CUDA 11.8. See [2]
* Refresh patches for 1.13

[1] https://github.com/google/flatbuffers/issues/7499
[2] https://github.com/archlinux/svntogit-community/commit/54642de5ba70f8c67ad6c16815dd31138ba47456

4 files changed, 117 insertions, 77 deletions

diff --git a/.SRCINFO b/.SRCINFO
index afbc0bca6376..a7bb02693ce1 100644
--- a/.SRCINFO
+++ b/.SRCINFO
@@ -1,6 +1,6 @@
 pkgbase = python-onnxruntime
 	pkgdesc = Cross-platform, high performance scoring engine for ML models
-	pkgver = 1.12.1
+	pkgver = 1.13.1
 	pkgrel = 1
 	url = https://github.com/microsoft/onnxruntime
 	arch = x86_64
@@ -14,6 +14,13 @@ pkgbase = python-onnxruntime
 	makedepends = boost
 	makedepends = eigen
 	makedepends = flatbuffers
+	makedepends = onednn
+	makedepends = python-coloredlogs
+	makedepends = python-flatbuffers
+	makedepends = python-numpy
+	makedepends = python-packaging
+	makedepends = python-protobuf
+	makedepends = python-sympy
 	makedepends = cuda
 	makedepends = cudnn
 	makedepends = nccl
@@ -21,25 +28,9 @@ pkgbase = python-onnxruntime
 	depends = nsync
 	depends = re2
 	depends = openmpi
-	depends = onednn
 	depends = libprotobuf-lite.so
-	depends = python-coloredlogs
-	depends = python-flatbuffers
-	depends = python-numpy
-	depends = python-packaging
-	depends = python-protobuf
-	depends = python-sympy
-	optdepends = python-onnx: for the backend API, quantization, orttraining, transformers and various tools
-	optdepends = python-psutil: for transformers
-	optdepends = python-py-cpuinfo: for transformers
-	optdepends = python-py3nvml: for transformers
-	optdepends = python-transformers: for transformers
-	optdepends = python-scipy: for transformers and various tools
-	optdepends = python-pytorch: for transformers, orttraining and various tools
-	optdepends = python-cerberus: for orttraining
-	optdepends = python-h5py: for orttraining
 	options = !lto
-	source = git+https://github.com/microsoft/onnxruntime#tag=v1.12.1
+	source = git+https://github.com/microsoft/onnxruntime#tag=v1.13.1
 	source = git+https://github.com/onnx/onnx.git
 	source = git+https://github.com/dcleblanc/SafeInt.git
 	source = git+https://github.com/tensorflow/tensorboard.git
@@ -56,16 +47,45 @@ pkgbase = python-onnxruntime
 	sha512sums = SKIP
 	sha512sums = SKIP
 	sha512sums = SKIP
-	sha512sums = 6ad02636e2cba8f2e82f5ac733856eee3016750e809ee18a41c4edc78bca761f30ac174d7d683f9b14b9f72310dd654811b1ecc9dda514e12bac6b7440c449c2
+	sha512sums = 8f0bd7ae59f86f002c88368a8c2852b9613363771aae61f91a90bfc13dcd3173e43d7988a59ccef86657cf6abfcc53837bbf445c216a7994a765a7e0770d0f5f
 	sha512sums = 7d55b0d4232183a81c20a5049f259872150536eed799d81a15e7f10b5c8b5279b443ba96d7b97c0e4338e95fc18c9d6f088e348fc7002256ee7170d25b27d80d
-	sha512sums = 6735c7aca2ba2f1f2a5286eb064125bf7f2c68a575d572dd157769d15778ff3e717b3a53d696c767748229f23ee6c3a7c82679df1d86283d7c4dd0ec9103ae08
+	sha512sums = ab48d27be98a88d3c361e1d0aac3b1e078096c0902ba7a543261a1c24faed0f1f44947a1b7ea1f264434cd2199b9d563d2447c14b6afbdf9900e68a65f7d2619
+
+pkgname = onnxruntime
+	depends = nsync
+	depends = re2
+	depends = openmpi
+	depends = libprotobuf-lite.so
+	depends = onednn
 
 pkgname = python-onnxruntime
+	depends = nsync
+	depends = re2
+	depends = openmpi
+	depends = libprotobuf-lite.so
+	depends = onnxruntime
+	depends = python-coloredlogs
+	depends = python-flatbuffers
+	depends = python-numpy
+	depends = python-packaging
+	depends = python-protobuf
+	depends = python-sympy
+	optdepends = python-onnx: for the backend API, quantization, orttraining, transformers and various tools
+	optdepends = python-psutil: for transformers
+	optdepends = python-py-cpuinfo: for transformers
+	optdepends = python-py3nvml: for transformers
+	optdepends = python-transformers: for transformers
+	optdepends = python-scipy: for transformers and various tools
+	optdepends = python-pytorch: for transformers, orttraining and various tools
+	optdepends = python-cerberus: for orttraining
+	optdepends = python-h5py: for orttraining
 
-pkgname = python-onnxruntime-cuda
+pkgname = onnxruntime-cuda
 	pkgdesc = Cross-platform, high performance scoring engine for ML models (CUDA execution provider)
 	depends = cuda
 	depends = cudnn
 	depends = nccl
 	depends = openmpi
 	depends = nsync
+	conflicts = python-onnxruntime-cuda
+	replaces = python-onnxruntime-cuda
diff --git a/PKGBUILD b/PKGBUILD
index 849fe5500982..d9fb7fc19a30 100644
--- a/PKGBUILD
+++ b/PKGBUILD
@@ -5,28 +5,16 @@ _ENABLE_CUDA=1
 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)
-pkgver=1.12.1
+pkgname=(onnxruntime python-onnxruntime)
+pkgver=1.13.1
 pkgdesc='Cross-platform, high performance scoring engine for ML models'
 pkgrel=1
 arch=(x86_64)
 url='https://github.com/microsoft/onnxruntime'
 license=(MIT)
-depends=(nsync re2 openmpi onednn libprotobuf-lite.so
-         python-coloredlogs python-flatbuffers python-numpy python-packaging python-protobuf python-sympy)
-makedepends=(git cmake pybind11 python-setuptools nlohmann-json chrono-date boost eigen flatbuffers)
-optdepends=(
-  # https://github.com/microsoft/onnxruntime/pull/9969
-  'python-onnx: for the backend API, quantization, orttraining, transformers and various tools'
-  'python-psutil: for transformers'
-  'python-py-cpuinfo: for transformers'
-  'python-py3nvml: for transformers'
-  'python-transformers: for transformers'
-  'python-scipy: for transformers and various tools'
-  'python-pytorch: for transformers, orttraining and various tools'
-  'python-cerberus: for orttraining'
-  'python-h5py: for orttraining'
-)
+depends=(nsync re2 openmpi libprotobuf-lite.so)
+makedepends=(git cmake pybind11 python-setuptools nlohmann-json chrono-date boost eigen flatbuffers onednn
+             python-coloredlogs python-flatbuffers python-numpy python-packaging python-protobuf python-sympy)
 # 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"
@@ -46,19 +34,19 @@ sha512sums=('SKIP'
             'SKIP'
             'SKIP'
             'SKIP'
-            '6ad02636e2cba8f2e82f5ac733856eee3016750e809ee18a41c4edc78bca761f30ac174d7d683f9b14b9f72310dd654811b1ecc9dda514e12bac6b7440c449c2'
+            '8f0bd7ae59f86f002c88368a8c2852b9613363771aae61f91a90bfc13dcd3173e43d7988a59ccef86657cf6abfcc53837bbf445c216a7994a765a7e0770d0f5f'
             '7d55b0d4232183a81c20a5049f259872150536eed799d81a15e7f10b5c8b5279b443ba96d7b97c0e4338e95fc18c9d6f088e348fc7002256ee7170d25b27d80d'
-            '6735c7aca2ba2f1f2a5286eb064125bf7f2c68a575d572dd157769d15778ff3e717b3a53d696c767748229f23ee6c3a7c82679df1d86283d7c4dd0ec9103ae08')
+            'ab48d27be98a88d3c361e1d0aac3b1e078096c0902ba7a543261a1c24faed0f1f44947a1b7ea1f264434cd2199b9d563d2447c14b6afbdf9900e68a65f7d2619')
 # CUDA seems not working with LTO
 options+=('!lto')
 
 if [[ $_ENABLE_CUDA = 1 ]]; then
-  pkgname+=(python-onnxruntime-cuda)
+  pkgname+=(onnxruntime-cuda)
   makedepends+=(cuda cudnn nccl gcc11)
 fi
 
 # Check PKGBUILDs of python-pytorch and tensorflow for CUDA architectures built by official packages
-_CUDA_ARCHITECTURES="52-real;53-real;60-real;61-real;62-real;70-real;72-real;75-real;80-real;86-real;86-virtual"
+_CUDA_ARCHITECTURES="52-real;53-real;60-real;61-real;62-real;70-real;72-real;75-real;80-real;86-real;87-real;89-real;90-real;90-virtual"
 
 prepare() {
   cd onnxruntime
@@ -86,6 +74,8 @@ build() {
     export CUDAHOSTCXX=$CXX
   fi
 
+  # Use -Donnxruntime_ENABLE_LAZY_TENSOR=OFF as it requires patched python-pytorch
+  # See: https://github.com/microsoft/onnxruntime/pull/10460 https://github.com/pytorch/pytorch/pulls/wschin
   local cmake_args=(
     -DCMAKE_INSTALL_PREFIX=/usr
     -Donnxruntime_ENABLE_PYTHON=ON
@@ -93,6 +83,7 @@ build() {
     -Donnxruntime_BUILD_SHARED_LIB=ON
     -Donnxruntime_BUILD_UNIT_TESTS=OFF
     -Donnxruntime_ENABLE_TRAINING=ON
+    -Donnxruntime_ENABLE_LAZY_TENSOR=OFF
     -Donnxruntime_USE_MPI=ON
     -Donnxruntime_USE_PREINSTALLED_EIGEN=ON
     -Donnxruntime_USE_DNNL=ON
@@ -131,31 +122,60 @@ build() {
   python ../setup.py build
 }
 
-package_python-onnxruntime() {
+package_onnxruntime() {
+  depends+=(onednn)
+
   cd onnxruntime/build
 
   DESTDIR="$pkgdir" cmake --install .
 
+  install -Ddm755 "$pkgdir"/usr/share/licenses
+  for f in LICENSE ThirdPartyNotices.txt ; do
+    install -Dm644 ../$f -t "$pkgdir"/usr/share/licenses/$pkgname
+  done
+
+  # installed as split packages
+  rm -vf "$pkgdir"/usr/lib/libonnxruntime_providers_cuda.so
+}
+
+package_python-onnxruntime() {
+  depends+=(onnxruntime python-coloredlogs python-flatbuffers python-numpy python-packaging python-protobuf python-sympy)
+  optdepends=(
+    # https://github.com/microsoft/onnxruntime/pull/9969
+    'python-onnx: for the backend API, quantization, orttraining, transformers and various tools'
+    'python-psutil: for transformers'
+    'python-py-cpuinfo: for transformers'
+    'python-py3nvml: for transformers'
+    'python-transformers: for transformers'
+    'python-scipy: for transformers and various tools'
+    'python-pytorch: for transformers, orttraining and various tools'
+    'python-cerberus: for orttraining'
+    'python-h5py: for orttraining'
+  )
+
+  cd onnxruntime/build
+
   python ../setup.py install --root="$pkgdir" --skip-build --optimize=1
 
   PY_ORT_DIR="$(python -c 'import site; print(site.getsitepackages()[0])')/onnxruntime"
-  install -Ddm755 "$pkgdir"/usr/share/licenses/$pkgname
-  for f in LICENSE ThirdPartyNotices.txt ; do
-    ln -s "$PY_ORT_DIR/$f" "$pkgdir"/usr/share/licenses/$pkgname/$f
-  done
   # already installed by `cmake --install`, and not useful as this path is not looked up by the linker
   rm -vf "$pkgdir/$PY_ORT_DIR"/capi/libonnxruntime_providers_*
 
+  install -Ddm755 "$pkgdir"/usr/share/licenses
+  ln -s onnxruntime "$pkgdir"/usr/share/licenses/$pkgname
+
   # installed as split packages
   rm -vf "$pkgdir"/usr/lib/libonnxruntime_providers_cuda.so
 }
 
-package_python-onnxruntime-cuda() {
+package_onnxruntime-cuda() {
   depends=(cuda cudnn nccl openmpi nsync)
+  conflicts=('python-onnxruntime-cuda')
+  replaces=('python-onnxruntime-cuda')
   pkgdesc+=' (CUDA execution provider)'
 
   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
+  ln -s onnxruntime "$pkgdir"/usr/share/licenses/$pkgname
 }
diff --git a/build-fixes.patch b/build-fixes.patch
index 2e86341c4d4b..a0b9b3c3e060 100644
--- a/build-fixes.patch
+++ b/build-fixes.patch
@@ -25,29 +25,12 @@ index a027c69e0..eb7608518 100644
    message("Use re2 from preinstalled system lib")
  else()
    add_subdirectory(external/re2 EXCLUDE_FROM_ALL)
-@@ -1421,11 +1421,11 @@
- #Dependencies end. In the next we'll enable "treat warning as error"
- 
- if (onnxruntime_PREFER_SYSTEM_LIB)
--  find_package(Flatbuffers)
-+  find_package(FlatBuffers)
+@@ -1421,7 +1421,7 @@
  endif()
--if (Flatbuffers_FOUND)
-+if (FlatBuffers_FOUND)
+ if (Flatbuffers_FOUND)
    message("Use flatbuffers from preinstalled system lib")
 -  add_library(flatbuffers ALIAS flatbuffers::flatbuffers)
 +  add_library(flatbuffers ALIAS flatbuffers::flatbuffers_shared)
  else()
    message("Use flatbuffers from submodule")
    # We do not need to build flatc for iOS or Android Cross Compile
---- a/setup.py	2022-07-22 17:00:19.638893453 +0800
-+++ b/setup.py	2022-07-22 17:02:00.686317628 +0800
-@@ -16,7 +16,7 @@
- 
- from setuptools import Extension, setup
- from setuptools.command.install import install as InstallCommandBase
--from wheel.vendored.packaging.tags import sys_tags
-+from packaging.tags import sys_tags
- 
- nightly_build = False
- package_name = "onnxruntime"
diff --git a/system-dnnl.diff b/system-dnnl.diff
index 1444d8acaa1e..aad0d878293d 100644
--- a/system-dnnl.diff
+++ b/system-dnnl.diff
@@ -1,35 +1,52 @@
 diff --git a/cmake/external/dnnl.cmake b/cmake/external/dnnl.cmake
-index 6a51a3d5d..a89635210 100644
+index 5c42216..bc0a63f 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
+@@ -26,6 +26,17 @@ 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)
++  find_package(dnnl CONFIG)
++endif ()
++if (TARGET DNNL::dnnl)
++  message("Use DNNL from preinstalled system lib")
 +  add_library(project_dnnl INTERFACE)
 +  add_library(dnnl INTERFACE)
 +  target_link_libraries(dnnl INTERFACE DNNL::dnnl)
 +else ()
++  message("Use DNNL from submodule")
    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)
+@@ -55,3 +66,4 @@ if (onnxruntime_USE_DNNL)
    )
    link_directories(${DNNL_LIB_DIR})
  endif()
 +endif()
+diff --git a/cmake/onnxruntime_python.cmake b/cmake/onnxruntime_python.cmake
+index b23d731..13ef7c3 100644
+--- a/cmake/onnxruntime_python.cmake
++++ b/cmake/onnxruntime_python.cmake
+@@ -752,7 +752,7 @@ if (onnxruntime_ENABLE_TRAINING)
+   endif()
+ endif()
+ 
+-if (onnxruntime_USE_DNNL)
++if (onnxruntime_USE_DNNL AND NOT TARGET DNNL::dnnl)
+   add_custom_command(
+     TARGET onnxruntime_pybind11_state POST_BUILD
+     COMMAND ${CMAKE_COMMAND} -E copy
 diff --git a/cmake/onnxruntime_unittests.cmake b/cmake/onnxruntime_unittests.cmake
-index 6bdb2d03c..514faa375 100644
+index c5e1dfe..a9cb311 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)
+@@ -836,10 +836,12 @@ if (NOT onnxruntime_ENABLE_TRAINING_TORCH_INTEROP)
+   endif()
    if (onnxruntime_USE_DNNL)
      list(APPEND onnx_test_libs dnnl)
-+    if (NOT onnxruntime_PREFER_SYSTEM_LIB)
++    if (NOT TARGET DNNL::dnnl)
      add_custom_command(
        TARGET ${test_data_target} POST_BUILD
        COMMAND ${CMAKE_COMMAND} -E copy ${DNNL_DLL_PATH} $<TARGET_FILE_DIR:${test_data_target}>
@@ -37,4 +54,4 @@ index 6bdb2d03c..514faa375 100644
 +    endif()
    endif()
    if(WIN32)
-     if (onnxruntime_USE_TVM)
+     if (onnxruntime_USE_NUPHAR_TVM)