ONCPUML-1072: Tuned MWS values (for N1, V1) for binary operators used by oneDNN
Added approximate values for MWS for the following binary operators:
Add, Sub, Mul, Min, Max, Div
Change-Id: I5c4c75511129982a3f44c038ee272f09598469de
Signed-off-by: Fadi Arafeh <fadi.arafeh@arm.com>
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/c/VisualCompute/ComputeLibrary/+/459609
Tested-by: bsgcomp <bsgcomp@arm.com>
Reviewed-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Comments-Addressed: bsgcomp <bsgcomp@arm.com>
Signed-off-by: fadara01 <fadi.arafeh@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8392
Reviewed-by: Gunes Bayir <gunes.bayir@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
diff --git a/src/cpu/kernels/CpuAddKernel.cpp b/src/cpu/kernels/CpuAddKernel.cpp
index 1648a46..ec210a4 100644
--- a/src/cpu/kernels/CpuAddKernel.cpp
+++ b/src/cpu/kernels/CpuAddKernel.cpp
@@ -33,6 +33,11 @@
#include "src/cpu/kernels/add/list.h"
#include <array>
+namespace
+{
+ static constexpr size_t default_mws_N1_fp32_neon = 24536;
+ static constexpr size_t default_mws_V1_fp32_neon = 40510;
+}
namespace arm_compute
{
namespace cpu
@@ -267,8 +272,41 @@
size_t CpuAddKernel::get_mws(const CPUInfo &platform, size_t thread_count) const
{
ARM_COMPUTE_UNUSED(thread_count);
- ARM_COMPUTE_UNUSED(platform);
+#if defined(ENABLE_FP32_KERNELS)
+ if(this->_run_method == &add_fp32_neon)
+ {
+ size_t mws = ICPPKernel::default_mws;
+ if(platform.get_cpu_model() == CPUModel::N1)
+ {
+ mws = default_mws_N1_fp32_neon;
+ }
+ else if(platform.get_cpu_model() == CPUModel::V1)
+ {
+ mws = default_mws_V1_fp32_neon;
+ }
+ else
+ {
+ return ICPPKernel::default_mws;
+ }
+
+ // tensor is 1D or was re-interpreted as 1D
+ if(this->window().shape().num_dimensions() == 1)
+ {
+ return mws;
+ }
+ else
+ {
+ // scale mws down by the number of elements along all the dimensions (x, z, w, etc) except the one
+ // that we parallelize along (the y dimension). This allows for parallelization when the Y_SIZE is small
+ // but the other sizes are large, which boosts performance.
+ mws = static_cast<size_t>(mws / (this->window().num_iterations_total() / this->window().num_iterations(1)));
+ return std::max(static_cast<size_t>(1), mws);
+ }
+ }
+#else /* ENABLE_FP32_KERNELS */
+ ARM_COMPUTE_UNUSED(platform);
+#endif /* ENABLE_FP32_KERNELS */
return ICPPKernel::default_mws;
}