From d7154dbf0f4a347f2f35f2475a893f1631c5ee1a Mon Sep 17 00:00:00 2001 From: Dana Zlotnik Date: Wed, 10 Nov 2021 11:50:58 +0200 Subject: Implement 1D Adaptive Workload Splitting in CPPScheduler Resolves COMPMID-4649 Change-Id: I941d2f8a40737ff05c49f6695a42884731ef2dc9 Signed-off-by: Dana Zlotnik Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/6656 Tested-by: Arm Jenkins Reviewed-by: SiCong Li Comments-Addressed: Arm Jenkins --- .../internal/CpuDepthwiseConv2dAssemblyWrapperKernel.cpp | 13 ++++++++++--- src/cpu/kernels/internal/CpuPool2dAssemblyWrapperKernel.cpp | 13 ++++++++++--- 2 files changed, 20 insertions(+), 6 deletions(-) (limited to 'src/cpu/kernels/internal') diff --git a/src/cpu/kernels/internal/CpuDepthwiseConv2dAssemblyWrapperKernel.cpp b/src/cpu/kernels/internal/CpuDepthwiseConv2dAssemblyWrapperKernel.cpp index a71864c10c..934e38b054 100644 --- a/src/cpu/kernels/internal/CpuDepthwiseConv2dAssemblyWrapperKernel.cpp +++ b/src/cpu/kernels/internal/CpuDepthwiseConv2dAssemblyWrapperKernel.cpp @@ -357,9 +357,16 @@ const char *CpuDepthwiseConv2dAssemblyWrapperKernel::name() const size_t CpuDepthwiseConv2dAssemblyWrapperKernel::get_mws(const CPUInfo &platform, size_t thread_count) const { - ARM_COMPUTE_UNUSED(platform, thread_count); - - return ICPPKernel::small_network_mws; + ARM_COMPUTE_UNUSED(thread_count); + // Tuning results that gave optimized results in performance investigation + if (platform.get_cpu_model() == CPUModel::A73 ) + { + return 10240; + } + else + { + return 9216; + } } } // namespace kernels } // namespace cpu diff --git a/src/cpu/kernels/internal/CpuPool2dAssemblyWrapperKernel.cpp b/src/cpu/kernels/internal/CpuPool2dAssemblyWrapperKernel.cpp index f9c11fd4bd..78ac134604 100644 --- a/src/cpu/kernels/internal/CpuPool2dAssemblyWrapperKernel.cpp +++ b/src/cpu/kernels/internal/CpuPool2dAssemblyWrapperKernel.cpp @@ -277,9 +277,16 @@ void CpuPool2dAssemblyWrapperKernel::create_arm_pooling_requant(const ITensorInf size_t CpuPool2dAssemblyWrapperKernel::get_mws(const CPUInfo &platform, size_t thread_count) const { - ARM_COMPUTE_UNUSED(platform, thread_count); - - return ICPPKernel::small_network_mws; + ARM_COMPUTE_UNUSED(thread_count); + // Tuning results that gave optimized results in performance investigation + if (platform.get_cpu_model() == CPUModel::A73 ) + { + return 10240; + } + else + { + return 9216; + } } } // namespace kernels } // namespace cpu -- cgit v1.2.1