From 658039bc4e06be34272eccf559a516a6b52f75f5 Mon Sep 17 00:00:00 2001 From: Georgios Pinitas Date: Fri, 15 Sep 2017 16:30:50 +0100 Subject: COMPMID-534: Add MemoryManager support in NEON functions Adds support for: -NECannyEdge -NEConvolution -NEDirectConvolution -NEGEMM -NEGEMMLowp -NEGaussian5x5 -NEHOGDescriptor -NEHOGGradient -NEL2Normalize -NELocallyConnectedLayer -NENormalizationLayer -NEScale -NESobel5x5 -NESobel7x7 Change-Id: I68e05aa6054372fa873a882633a15fb97882c00d Reviewed-on: http://mpd-gerrit.cambridge.arm.com/87926 Reviewed-by: Pablo Tello Tested-by: Kaizen --- src/runtime/NEON/functions/NEGEMM.cpp | 12 ++++++++++-- 1 file changed, 10 insertions(+), 2 deletions(-) (limited to 'src/runtime/NEON/functions/NEGEMM.cpp') diff --git a/src/runtime/NEON/functions/NEGEMM.cpp b/src/runtime/NEON/functions/NEGEMM.cpp index dfcb3954ea..85b283cd41 100644 --- a/src/runtime/NEON/functions/NEGEMM.cpp +++ b/src/runtime/NEON/functions/NEGEMM.cpp @@ -36,8 +36,8 @@ using namespace arm_compute; -NEGEMM::NEGEMM() - : _interleave_kernel(), _transpose_kernel(), _mm_kernel(), _ma_kernel(), _tmp_a(), _tmp_b(), _run_vector_matrix_multiplication(false), _run_addition(false) +NEGEMM::NEGEMM(std::shared_ptr memory_manager) + : _memory_group(std::move(memory_manager)), _interleave_kernel(), _transpose_kernel(), _mm_kernel(), _ma_kernel(), _tmp_a(), _tmp_b(), _run_vector_matrix_multiplication(false), _run_addition(false) { } @@ -85,6 +85,10 @@ void NEGEMM::configure(const ITensor *a, const ITensor *b, const ITensor *c, ITe _tmp_a.allocator()->init(info_a); _tmp_b.allocator()->init(info_b); + // Manage intermediate buffers + _memory_group.manage(&_tmp_a); + _memory_group.manage(&_tmp_b); + // Configure interleave kernel _interleave_kernel.configure(a, &_tmp_a); @@ -109,6 +113,8 @@ void NEGEMM::configure(const ITensor *a, const ITensor *b, const ITensor *c, ITe void NEGEMM::run() { + _memory_group.acquire(); + if(!_run_vector_matrix_multiplication) { // Run interleave kernel @@ -121,6 +127,8 @@ void NEGEMM::run() // Run matrix multiply kernel NEScheduler::get().schedule(&_mm_kernel, _run_vector_matrix_multiplication ? Window::DimX : Window::DimY); + _memory_group.release(); + // Run matrix addition kernel if(_run_addition) { -- cgit v1.2.1