Updates for 25.02 release

- Updates changelog
- Removes pytorch/pytorch#139387 - Add prepacking
  for linear weights. Performance gains better realised by ideep reorder
  caching.
- Updates OpenBLAS build to use recent commit from develop which
  includes: OpenMathLib/OpenBLAS#5108

ML-Frameworks/pytorch-aarch64/CHANGELOG.md

@@ -7,18 +7,31 @@ where `YY` is the year, and `MM` the month of the increment.
 
 ## [unreleased]
 
+### Added
+
+### Changed
+
+### Removed
+
+### Fixed
+
+## [r25.02] 2025-02-11
+https://github.com/ARM-software/Tool-Solutions/tree/pytorch-aarch64--r25.02
+
 ### Added
 - Adds work-in-progress PyTorch PRs:
-  - 145942 3d05899222da2b93ed3d4c88c382d318e68eeec6 - Enable qlinear_dynamic path for AArch64 through Arm Compute Library directly.
-  - 146476 8cacbf8a58ba43bb51340ba69204be342b368cf5 - Improve KleidiAI 4 bit kernel performance
-  - 143666 8e5134e9c22cdb6150e425bee43015998ae55c59 - Extend Vec backend with SVE BF16
+  - https://github.com/pytorch/pytorch/pull/145942 - Enable qlinear_dynamic path for AArch64 through Arm Compute Library directly. Gives ~15% speed up on approach in Tool Solutions 24.12.
+  - https://github.com/pytorch/pytorch/pull/146476 - Improve KleidiAI 4 bit kernel performance. Greater than 10% performance improvment when calling INT4 KleidiAI kernels
+  - https://github.com/pytorch/pytorch/pull/143666 - Extend Vec backend with SVE BF16.
 - Adds work-in-progress oneDNN PRs:
-  - 2502 49ac258a43520562a196ba081a3c259ac3732df2 - cpu: aarch64: ip: Allow bf16 for ACL inner product
+  - https://github.com/oneapi-src/oneDNN/pull/2502 - cpu: aarch64: ip: Allow bf16 for ACL inner product. Gives speedups of ~170x for BERT_pytorch and ~160x for alexnet using bf16 compile mode.
 - Minor improvements to build process and logging.
+- OpenBLAS build from source at 1b85b6a396c94e78c9ba14aafcdfd5c5da5a8bb2 from develop branch.
+  This includes https://github.com/OpenMathLib/OpenBLAS/pull/5108 to add SBGEMM support for 256bit SVE.
 
 ### Changed
 - Updates hashes for:
-  - PyTorch to 8d4926e30a944320adf434016129cb6788eff79b, from viable/strict
+  - PyTorch to 8d4926e30a944320adf434016129cb6788eff79b (2.7.0.dev20250115), from viable/strict
   - ideep to e026f3b0318087fe19e2b062e8edf55bfe7a522c, from ideep_pytorch
   - oneDNN to 0fd3b73a25d11106e141ddefd19fcacc74f8bbfe, from main
   - Arm Compute Library to 6acccf1730b48c9a22155998fc4b2e0752472148. from main
@@ -31,11 +44,13 @@ where `YY` is the year, and `MM` the month of the increment.
 
 ### Removed
 - Removes patches that are now merged upstream.
+- Removes https://github.com/pytorch/pytorch/pull/139387 - Add prepacking for linear weights. Performance gains better realised by ideep reorder caching.
 
 ### Fixed
 - Addition of https://github.com/pytorch/pytorch/pull/145486 fixes illigal instruction on non-SVE targets.
 
 ## [r24.12] 2024-12-20
+https://github.com/ARM-software/Tool-Solutions/tree/pytorch-aarch64--r24.12
 
 ### Added
 - Adds torchao.

ML-Frameworks/pytorch-aarch64/build-wheel.sh

@@ -87,7 +87,7 @@ if ! docker container inspect $TORCH_BUILD_CONTAINER >/dev/null 2>&1 ; then
     docker exec -t $TORCH_BUILD_CONTAINER bash -c "yum install -y tbb tbb-devel"
 
     # This must be in this if block because it cannot handle being called twice
-    docker exec -t $TORCH_BUILD_CONTAINER bash -c "bash $UTILS/install_openblas.sh"
+    docker exec -t $TORCH_BUILD_CONTAINER bash -c "bash $UTILS/build_openblas.sh"
 
     echo "Storing torch build container id in $TORCH_BUILD_CONTAINER_ID_FILE for reuse: $TORCH_BUILD_CONTAINER"
     echo $TORCH_BUILD_CONTAINER > "$TORCH_BUILD_CONTAINER_ID_FILE"

ML-Frameworks/pytorch-aarch64/get-source.sh

@@ -37,12 +37,6 @@ git-shallow-clone https://github.com/pytorch/pytorch.git $PYTORCH_HASH
     apply-github-patch https://github.com/pytorch/pytorch 139887 eff3c11b1a31f725b50020ce32f6eddba17b5a94 # Use s8s8s8 for qlinear on aarch64 instead of u8s8u8 with mkl-dnn
     apply-github-patch https://github.com/pytorch/pytorch 136850 6d5aaff8434203f870d76d840158d6989ddd61d0 # Enable XNNPACK for quantized add
     apply-github-patch https://github.com/pytorch/pytorch 142391 8373846f441381a56e7abd905af84102aa52fc7b # parallelize sort
-    apply-github-patch https://github.com/pytorch/pytorch 139387 4140ee51b77d11966fc6d584f11eb15a45ec63c7 # Add prepacking for linear weights
-    apply-github-patch https://github.com/pytorch/pytorch 139387 c71443890e6d5538168d4fb2d760200fd9fef8d4 # Add prepacking for linear weights
-    apply-github-patch https://github.com/pytorch/pytorch 139387 3759c98aced224d9aa987e80db46a22355001cf5 # Add prepacking for linear weights
-    apply-github-patch https://github.com/pytorch/pytorch 139387 a7ff8b73fc5a44f8842853fe25a53695ecca92f2 # Add prepacking for linear weights
-    apply-github-patch https://github.com/pytorch/pytorch 139387 d4512a8ea1748ca15c751c2e87d177ef1c7094da # Add prepacking for linear weights
-    apply-github-patch https://github.com/pytorch/pytorch 139387 d4145ece750e0737b0d1e73677d658817c4b64da # Add prepacking for linear weights
     apply-github-patch https://github.com/pytorch/pytorch 140159 8d3404ec5972528f606fe605887ad2254a174fbc # cpu: aarch64: enable gemm-bf16f32
     apply-github-patch https://github.com/pytorch/pytorch 140159 ab4c191ef0de1e4eced6b4dd7b6e387f57034ad9 # cpu: aarch64: enable gemm-bf16f32
     apply-github-patch https://github.com/pytorch/pytorch 140159 879ca72d54559a388db315eed40803d2f1c827b7 # cpu: aarch64: enable gemm-bf16f32

ML-Frameworks/pytorch-aarch64/utils/install_openblas.sh → ML-Frameworks/pytorch-aarch64/utils/build_openblas.sh

@@ -25,16 +25,11 @@
 source /utils/helper.sh
 
 set -ex
+OPENBLAS_HASH="1b85b6a396c94e78c9ba14aafcdfd5c5da5a8bb2"
 OPENBLAS_CHECKOUT_DIR="OpenBLAS"
 
 cd /
-git clone https://github.com/OpenMathLib/OpenBLAS.git -b develop --depth 1
-(
-    cd $OPENBLAS_CHECKOUT_DIR
-    apply-github-patch https://github.com/OpenMathLib/OpenBLAS/ 5108 4379a6fbe37038082c657bba5be5c67331a0bd0b 
-    apply-github-patch https://github.com/OpenMathLib/OpenBLAS/ 5108 c748e6a33871f0dfa3bf6569c88a676c9a387411 
-    cd /
-)
+git-shallow-clone https://github.com/OpenMathLib/OpenBLAS.git $OPENBLAS_HASH
 
 OPENBLAS_BUILD_FLAGS="
 NUM_THREADS=128