From 7b9998d0fe1f98768b690ead10ebfa166d1b873d Mon Sep 17 00:00:00 2001 From: Manuel Bottini Date: Mon, 21 Oct 2019 17:59:07 +0100 Subject: COMPMID-1816: Use parallel reduction on 0 axis in CL ARG_MIN/ARG_MAX Introducing new CLArgMinMax kernel Change-Id: I0b8254207cc3859d19ceef9b6429cf5c1c586db0 Signed-off-by: Manuel Bottini Reviewed-on: https://review.mlplatform.org/c/2202 Comments-Addressed: Arm Jenkins Tested-by: Arm Jenkins Reviewed-by: Michalis Spyrou --- src/runtime/Utils.cpp | 17 +++++++++++++++++ 1 file changed, 17 insertions(+) (limited to 'src/runtime/Utils.cpp') diff --git a/src/runtime/Utils.cpp b/src/runtime/Utils.cpp index 70494be05c..2204ec11d7 100644 --- a/src/runtime/Utils.cpp +++ b/src/runtime/Utils.cpp @@ -25,6 +25,7 @@ #include "arm_compute/runtime/NEON/NEScheduler.h" +#include #include #include @@ -61,4 +62,20 @@ void schedule_kernel_on_ctx(IRuntimeContext *ctx, ICPPKernel *kernel, const ISch NEScheduler::get().schedule(kernel, hints); } } + +unsigned int calculate_number_of_stages_only_x_axis(size_t input_x_dimension, unsigned int axis) +{ + // We need only 1 stage for all axis except x-axis + if(axis != 0) + { + return 1; + } + // Calculate number of WGs. 16 elements per thread, 8 threads per WG + const auto num_of_wg = static_cast(ceil(input_x_dimension / 128.f)); + + // Calculate number of stages. First stage performs op and the rest reduction sum + // depending on the size of the input. Last stage should have only 1 WG. + const unsigned int num_of_stages = num_of_wg / 128 + 2; + return num_of_stages; +} } // namespace arm_compute -- cgit v1.2.1