From f30310efab884230d968149c8dad73bb3f999260 Mon Sep 17 00:00:00 2001 From: Siddhartha Menon Date: Thu, 28 Nov 2024 15:42:21 +0000 Subject: [PATCH] fix, build, docs: aarch64: mutex lock if the ACL kernel is not stateless This feature depends on ACL >= 24.11.1 so the build system and docs have been updated to reflect that. Signed-off-by: Siddhartha Menon --- .github/automation/build_acl.sh | 2 +- README.md | 2 +- cmake/ACL.cmake | 2 +- src/cpu/aarch64/matmul/acl_matmul.cpp | 13 +++++++++++-- src/cpu/aarch64/matmul/acl_matmul.hpp | 2 ++ 5 files changed, 16 insertions(+), 5 deletions(-) diff --git a/.github/automation/build_acl.sh b/.github/automation/build_acl.sh index 0c45cc5a291..7ed588618ff 100755 --- a/.github/automation/build_acl.sh +++ b/.github/automation/build_acl.sh @@ -28,7 +28,7 @@ source ${SCRIPT_DIR}/common_aarch64.sh ACL_CONFIG=${ACL_CONFIG:-"Release"} ACL_ROOT_DIR=${ACL_ROOT_DIR:-"${PWD}/ComputeLibrary"} -ACL_VERSION=${ACL_VERSION:-v24.09} +ACL_VERSION=${ACL_VERSION:-v24.11.1} ACL_ARCH=${ACL_ARCH:-"armv8.2-a"} ACL_REPO="https://github.com/ARM-software/ComputeLibrary.git" diff --git a/README.md b/README.md index 15ed426b58d..7030985f013 100644 --- a/README.md +++ b/README.md @@ -173,7 +173,7 @@ On a CPU based on Arm AArch64 architecture, oneDNN CPU engine can be built with machine learning applications and provides AArch64 optimized implementations of core functions. This functionality currently requires that ACL is downloaded and built separately. See [Build from Source] section of the Developer Guide for -details. oneDNN only supports Compute Library versions 24.09 or later. +details. oneDNN only supports Compute Library versions 24.11.1 or later. [Arm Compute Library (ACL)]: https://github.com/arm-software/ComputeLibrary diff --git a/cmake/ACL.cmake b/cmake/ACL.cmake index b3fcd5f2912..3a08f779a39 100644 --- a/cmake/ACL.cmake +++ b/cmake/ACL.cmake @@ -31,7 +31,7 @@ endif() find_package(ACL REQUIRED) -set(ACL_MINIMUM_VERSION "24.09") +set(ACL_MINIMUM_VERSION "24.11.1") if(ACL_FOUND) file(GLOB_RECURSE ACL_VERSION_FILE ${ACL_INCLUDE_DIR}/*/arm_compute_version.embed) diff --git a/src/cpu/aarch64/matmul/acl_matmul.cpp b/src/cpu/aarch64/matmul/acl_matmul.cpp index e4c2b98e5b2..2c3beb542cf 100644 --- a/src/cpu/aarch64/matmul/acl_matmul.cpp +++ b/src/cpu/aarch64/matmul/acl_matmul.cpp @@ -15,6 +15,7 @@ *******************************************************************************/ #include "cpu/aarch64/matmul/acl_matmul.hpp" +#include namespace dnnl { namespace impl { @@ -170,12 +171,20 @@ status_t acl_matmul_t::pd_t::init(engine_t *engine) { template status_t acl_matmul_t::execute_forward(const exec_ctx_t &ctx) const { - status_t status = status::success; auto src_base = CTX_IN_MEM(const data_t *, DNNL_ARG_SRC); auto wei_base = CTX_IN_MEM(const data_t *, DNNL_ARG_WEIGHTS); - auto amp = pd()->amp_; + const auto & = pd()->amp_; + + std::unique_lock locker {mtx_, std::defer_lock}; + + // Some of the underlying kernels used by ACL still require some state and + // are not safe to be called in parallel with different execution contexts. + // Eventually when all kernels are truly stateless, this guard can be + // removed. + if (!acl_obj_->asm_gemm.has_stateless_impl()) { locker.lock(); } + bool is_transA = amp.is_transA; bool is_transB = amp.is_transB; bool do_transC = amp.do_transC; diff --git a/src/cpu/aarch64/matmul/acl_matmul.hpp b/src/cpu/aarch64/matmul/acl_matmul.hpp index 246a96ddd9e..80ee3d85b56 100644 --- a/src/cpu/aarch64/matmul/acl_matmul.hpp +++ b/src/cpu/aarch64/matmul/acl_matmul.hpp @@ -17,6 +17,7 @@ #ifndef ACL_MATMUL_HPP #define ACL_MATMUL_HPP +#include #include "common/utils.hpp" #include "cpu/aarch64/acl_post_ops.hpp" #include "cpu/aarch64/matmul/acl_matmul_utils.hpp" @@ -60,6 +61,7 @@ struct acl_matmul_t : public primitive_t { const pd_t *pd() const { return (const pd_t *)primitive_t::pd().get(); } std::unique_ptr acl_obj_; + mutable std::mutex mtx_; }; // acl_matmul_t } // namespace matmul