commit e34a1cd545f2c5374a80e1cebd1fe895c28fc7a4
author Raman Sarokin <sorokin@google.com> Fri May 01 11:26:13 2020 -0700
committer TensorFlower Gardener <gardener@tensorflow.org> Fri May 01 11:29:35 2020 -0700
tree c11488df7ba9d09bc8ebbcdc7a8d5e55ffbfc892
parent ae76544efc2731acb3f1ab401ffa1a35c02f57d6

Convolution performance improvement for Adreno.

PiperOrigin-RevId: 309441540
Change-Id: Ie70bc203b625078e8ec201d97289d8f999a67acb

tensorflow/lite/delegates/gpu/cl/kernels/conv_powervr.cc

diff --git a/tensorflow/lite/delegates/gpu/cl/kernels/conv_powervr.cc b/tensorflow/lite/delegates/gpu/cl/kernels/conv_powervr.cc
index eb4b222..9bb52b3 100644
--- a/tensorflow/lite/delegates/gpu/cl/kernels/conv_powervr.cc
+++ b/tensorflow/lite/delegates/gpu/cl/kernels/conv_powervr.cc
@@ -795,6 +795,13 @@
     conv_params.work_group_launch_order = int3(0, 1, 2);
     conv_params.fixed_work_group_size = false;
     conv_params.weights_upload_type = WeightsUploadType::GLOBAL_MEM;
+  } else if (device.IsAdreno()) {
+    conv_params.block_size = int3(2, 2, 1);
+    conv_params.work_group_size = int3(8, 2, 1);
+    conv_params.work_group_launch_order = int3(0, 1, 2);
+    conv_params.fixed_work_group_size = false;
+    conv_params.src_depth_loop_size = 1;
+    conv_params.weights_upload_type = WeightsUploadType::GLOBAL_MEM;
   } else {
     conv_params.block_size = int3(1, 1, 4);
     conv_params.work_group_size = int3(8, 2, 1);