commit e52a3000d2c13bc1b66ca66b3d12b6b836982394
author Gian Marco Iodice <gianmarco.iodice@arm.com> Wed Apr 11 15:59:10 2018 +0100
committer Anthony Barbier <anthony.barbier@arm.com> Fri Nov 02 16:49:37 2018 +0000
tree 70e8ef5ba216762604f84228805aac9bd65747b6
parent dd03870b63784abe499761da2b26b209b33f2db2

COMPMID-1026 - Add support for 4x4 output tile in CLWinogradConvolutionLayer

The performance achieved can be found at the following confluence page:
https://confluence.arm.com/display/MLENG/GEMM-based+convolution+vs+Winograd-based+convolution+on+OpenCL

Change-Id: I4b690cfdd4eb4ff0cd17b14fdd49ccaa1d1dc85c
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/127729
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>

src/runtime/NEON/functions/NEConvolutionLayer.cpp

diff --git a/src/runtime/NEON/functions/NEConvolutionLayer.cpp b/src/runtime/NEON/functions/NEConvolutionLayer.cpp
index afc3545..b0603e9 100644
--- a/src/runtime/NEON/functions/NEConvolutionLayer.cpp
+++ b/src/runtime/NEON/functions/NEConvolutionLayer.cpp
@@ -109,10 +109,12 @@
     ARM_COMPUTE_ERROR_ON_NULLPTR(weights);
     ARM_COMPUTE_UNUSED(output);
     ARM_COMPUTE_UNUSED(weights_info);
-    ARM_COMPUTE_UNUSED(act_info);
 
-    if((input->data_type() == DataType::F32) && (weights->dimension(0) == 3) && (weights->dimension(1) == 3) && (weights->num_dimensions() <= 4) && (conv_info.stride().first == 1)
-       && (conv_info.stride().second == 1) && (dilation == Size2D(1U, 1U)))
+    const size_t idx_w = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::WIDTH);
+    const size_t idx_h = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::HEIGHT);
+
+    if((input->data_type() == DataType::F32) && (input->data_layout() == DataLayout::NCHW) && (weights->dimension(idx_w) == 3) && (weights->dimension(idx_h) == 3) && (weights->num_dimensions() <= 4)
+       && (conv_info.stride().first == 1) && (conv_info.stride().second == 1) && (dilation == Size2D(1U, 1U)) && (!act_info.enabled()))
     {
         //FIXME Until COMPMID-1041 is implemented Winograd is slower than GEMM on A53.
         if(Scheduler::get().cpu_info().get_cpu_model() != CPUModel::A53)