[aarch64] ACL, OpenBLAS and mkldnn updates for PyTorch 2.2 (#1627)

Note# ~~This PR has a dependency on updating the oneDNN version to v3.3 (via ideep submodule to v3.3)~~
ideep submodule update is done, so, this PR can be merged anytime now.

This PR is for:
ACL - build with fixed format kernels 
OpenBLAS - upgrade the version to 0.3.25
numpy - upgrade version to 1.26.2
and mkldnn - cleanup the patches that are already upstreamed.

aarch64_linux/aarch64_ci_setup.sh

@@ -28,9 +28,9 @@ conda activate "${CONDA_ENV_NAME}"
 if [[ "$DESIRED_PYTHON"  == "3.8" ]]; then
     NUMPY_VERSION="1.24.4"
 else
-    NUMPY_VERSION="1.26.0"
+    NUMPY_VERSION="1.26.2"
 fi
-conda install -y -c conda-forge numpy==${NUMPY_VERSION} pyyaml==6.0.1 patchelf==0.17.2 pygit2==1.13.2 openblas==0.3.24 ninja==1.11.1 scons==4.5.2
+conda install -y -c conda-forge numpy==${NUMPY_VERSION} pyyaml==6.0.1 patchelf==0.17.2 pygit2==1.13.2 openblas==0.3.25 ninja==1.11.1 scons==4.5.2
 
 python --version
 conda --version

aarch64_linux/aarch64_wheel_ci_build.py

@@ -20,9 +20,9 @@ def build_ArmComputeLibrary(git_clone_flags: str = "") -> None:
     '''
     print('Building Arm Compute Library')
     os.system("cd / && mkdir /acl")
-    os.system(f"git clone https://github.com/ARM-software/ComputeLibrary.git -b v23.05.1 {git_clone_flags}")
+    os.system(f"git clone https://github.com/ARM-software/ComputeLibrary.git -b v23.08 {git_clone_flags}")
     os.system("cd ComputeLibrary; export acl_install_dir=/acl; "
-              "scons Werror=1 -j8 debug=0 neon=1 opencl=0 os=linux openmp=1 cppthreads=0 arch=armv8a multi_isa=1 build=native build_dir=$acl_install_dir/build; "
+              "scons Werror=1 -j8 debug=0 neon=1 opencl=0 os=linux openmp=1 cppthreads=0 arch=armv8a multi_isa=1 fixed_format_kernels=1 build=native build_dir=$acl_install_dir/build; "
               "cp -r arm_compute $acl_install_dir; "
               "cp -r include $acl_install_dir; "
               "cp -r utils $acl_install_dir; "
@@ -106,8 +106,6 @@ def parse_arguments():
         print("build pytorch without mkldnn backend")
 
     # work around to fix Raspberry pie crash
-    print("Applying mkl-dnn patch to fix Raspberry pie crash")
-    os.system("cd /pytorch/third_party/ideep/mkl-dnn && patch -p1 < /builder/mkldnn_fix/aarch64-fix-default-build-flags-to-armv8-a.patch")
     print("Applying mkl-dnn patch to fix readdir crash")
     os.system("cd /pytorch/third_party/ideep/mkl-dnn && patch -p1 < /builder/mkldnn_fix/aarch64-fix-readdir-crash.patch")
     os.system(f"cd /pytorch; {build_vars} python3 setup.py bdist_wheel")

aarch64_linux/build_aarch64_wheel.py

@@ -219,15 +219,15 @@ def install_condaforge_python(host: RemoteHost, python_version="3.8") -> None:
 
 def build_OpenBLAS(host: RemoteHost, git_clone_flags: str = "") -> None:
     print('Building OpenBLAS')
-    host.run_cmd(f"git clone https://github.com/xianyi/OpenBLAS -b v0.3.20 {git_clone_flags}")
+    host.run_cmd(f"git clone https://github.com/xianyi/OpenBLAS -b v0.3.25 {git_clone_flags}")
     make_flags = "NUM_THREADS=64 USE_OPENMP=1 NO_SHARED=1 DYNAMIC_ARCH=1 TARGET=ARMV8"
     host.run_cmd(f"pushd OpenBLAS && make {make_flags} -j8 && sudo make {make_flags} install && popd && rm -rf OpenBLAS")
 
 
 def build_ArmComputeLibrary(host: RemoteHost, git_clone_flags: str = "") -> None:
     print('Building Arm Compute Library')
-    acl_build_flags="debug=0 neon=1 opencl=0 os=linux openmp=1 cppthreads=0 arch=armv8a multi_isa=1 build=native"
-    host.run_cmd(f"git clone https://github.com/ARM-software/ComputeLibrary.git -b v23.05.1 {git_clone_flags}")
+    acl_build_flags="debug=0 neon=1 opencl=0 os=linux openmp=1 cppthreads=0 arch=armv8a multi_isa=1 fixed_format_kernels=1 build=native"
+    host.run_cmd(f"git clone https://github.com/ARM-software/ComputeLibrary.git -b v23.08 {git_clone_flags}")
     host.run_cmd(f"cd ComputeLibrary && scons Werror=1 -j8 {acl_build_flags}")
 
 
@@ -555,7 +555,6 @@ def start_build(host: RemoteHost, *,
         print("build pytorch with mkldnn+acl backend")
         build_vars += " USE_MKLDNN=ON USE_MKLDNN_ACL=ON"
         host.run_cmd(f"cd $HOME && git clone https://github.com/pytorch/builder.git")
-        host.run_cmd(f"cd $HOME/pytorch/third_party/ideep/mkl-dnn && patch -p1 < $HOME/builder/mkldnn_fix/aarch64-fix-default-build-flags-to-armv8-a.patch")
         host.run_cmd(f"cd $HOME/pytorch && export ACL_ROOT_DIR=$HOME/ComputeLibrary && {build_vars} python3 setup.py bdist_wheel{build_opts}")
         print('Repair the wheel')
         pytorch_wheel_name = host.list_dir("pytorch/dist")[0]