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 <anthony.barbier@arm.com>
Tested-by: Kaizen <jeremy.johnson+kaizengerrit@arm.com>

1 files changed, 8 insertions, 0 deletions

diff --git a/src/runtime/CL/functions/CLDirectConvolutionLayer.cpp b/src/runtime/CL/functions/CLDirectConvolutionLayer.cpp
index 65be417afb..6fafd9c7f1 100644
--- a/src/runtime/CL/functions/CLDirectConvolutionLayer.cpp
+++ b/src/runtime/CL/functions/CLDirectConvolutionLayer.cpp
@@ -38,13 +38,21 @@ CLDirectConvolutionLayer::CLDirectConvolutionLayer()
 
 void CLDirectConvolutionLayer::configure(ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info)
 {
+    // Set GPU target
+    _direct_conv_kernel.set_target(CLScheduler::get().target());
+
+    // Configure direct convolution
     _direct_conv_kernel.configure(input, weights, biases, output, conv_info);
 
+    // Configure border handler
     _input_border_handler.configure(input, _direct_conv_kernel.border_size(), BorderMode::CONSTANT, PixelValue(0));
 }
 
 void CLDirectConvolutionLayer::run()
 {
+    // Run border handler
     CLScheduler::get().enqueue(_input_border_handler, false);
+
+    // Run direct convolution
     CLScheduler::get().enqueue(_direct_conv_kernel);
 }