From 1246b63ca04cb067f26ae860688647224d6ba24e Mon Sep 17 00:00:00 2001 From: Gian Marco Iodice Date: Wed, 16 Aug 2017 18:38:32 +0100 Subject: COMPMID-477 - Optimized Direct Convolution 3x3 and 5x5 (f32) for Bifrost. Each work-item computes 4x3 output elements in case of 3x3 convolution and 4x2 in case of 5x5 convolution Change-Id: I6ebbaff8b7e971c1f90d5845c0b58d2a40f39df5 Reviewed-on: http://mpd-gerrit.cambridge.arm.com/84345 Reviewed-by: Anthony Barbier Tested-by: Kaizen --- src/runtime/CL/functions/CLGEMM.cpp | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) (limited to 'src/runtime/CL/functions/CLGEMM.cpp') diff --git a/src/runtime/CL/functions/CLGEMM.cpp b/src/runtime/CL/functions/CLGEMM.cpp index 935e856333..e81d8a6b97 100644 --- a/src/runtime/CL/functions/CLGEMM.cpp +++ b/src/runtime/CL/functions/CLGEMM.cpp @@ -59,6 +59,8 @@ void CLGEMM::configure(const ICLTensor *a, const ICLTensor *b, const ICLTensor * ARM_COMPUTE_ERROR_ON_MSG(a->info()->dimension(0) != b->info()->dimension(1), "The product AB is defined only if the number of columns in A is equal to the number of rows in B"); + _mm_kernel.set_target(CLScheduler::get().target()); + // Check if the first input tensor is a vector. If so, all the kernels for reshaping the tensors can be skipped if(a->info()->dimension(1) != 1) { @@ -87,7 +89,6 @@ void CLGEMM::configure(const ICLTensor *a, const ICLTensor *b, const ICLTensor * _transpose_kernel.configure(b, &_tmp_b); // Configure matrix multiply kernel - _mm_kernel.set_target(CLScheduler::get().target()); _mm_kernel.configure(&_tmp_a, &_tmp_b, output, alpha); // Allocate intermediate tensors -- cgit v1.2.1