experiments/STT-tensorflow
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Better FullyConnected/ConvTransposed selection for Intel.

Browse Source 
PiperOrigin-RevId: 316997745
Change-Id: I28befdd528917c3846ff6ae79b0f8427389dfc39
This commit is contained in:
Raman Sarokin 2020-06-17 17:18:45 -07:00 committed by TensorFlower Gardener
parent bae911298b
commit 1f05cc5973
4 changed files with 6 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
tensorflow/lite/delegates/gpu/cl/kernels/convolution_transposed_3x3.cc
View File

@ -269,7 +269,7 @@ ConvolutionTransposed3x3::ConvolutionTransposed3x3(
work_group_launch_order_(2, 0, 1) { work_group_launch_order_(2, 0, 1) {
if (device.IsPowerVR()) { if (device.IsPowerVR()) {
weights_upload_type_ = WeightsUploadType::LOCAL_MEM_ASYNC; weights_upload_type_ = WeightsUploadType::LOCAL_MEM_ASYNC;
} else if (device.IsNvidia()) { } else if (device.IsNvidia() || device.IsIntel()) {
weights_upload_type_ = WeightsUploadType::LOCAL_MEM_BY_THREADS; weights_upload_type_ = WeightsUploadType::LOCAL_MEM_BY_THREADS;
} else if (device.IsAMD()) { } else if (device.IsAMD()) {
weights_upload_type_ = WeightsUploadType::CONSTANT_MEM; weights_upload_type_ = WeightsUploadType::CONSTANT_MEM;

2
tensorflow/lite/delegates/gpu/cl/kernels/convolution_transposed_4x4.cc
View File

@ -270,7 +270,7 @@ ConvolutionTransposed4x4::ConvolutionTransposed4x4(
: GPUOperation(definition) { : GPUOperation(definition) {
if (device.IsPowerVR()) { if (device.IsPowerVR()) {
weights_upload_type_ = WeightsUploadType::LOCAL_MEM_ASYNC; weights_upload_type_ = WeightsUploadType::LOCAL_MEM_ASYNC;
} else if (device.IsNvidia()) { } else if (device.IsNvidia() || device.IsIntel()) {
weights_upload_type_ = WeightsUploadType::LOCAL_MEM_BY_THREADS; weights_upload_type_ = WeightsUploadType::LOCAL_MEM_BY_THREADS;
} else if (device.IsAMD()) { } else if (device.IsAMD()) {
weights_upload_type_ = WeightsUploadType::CONSTANT_MEM; weights_upload_type_ = WeightsUploadType::CONSTANT_MEM;

1
tensorflow/lite/delegates/gpu/cl/selectors/convolution_transposed_selector.cc
View File

@ -112,6 +112,7 @@ absl::Status SelectConvolutionTransposed(
case Vendor::POWERVR: case Vendor::POWERVR:
case Vendor::NVIDIA: case Vendor::NVIDIA:
case Vendor::AMD: case Vendor::AMD:
case Vendor::INTEL:
return SelectConvolutionTransposedPowerVR(attr, creation_context, op_def, return SelectConvolutionTransposedPowerVR(attr, creation_context, op_def,
ptr); ptr);
case Vendor::MALI: case Vendor::MALI:

3
tensorflow/lite/delegates/gpu/cl/selectors/fully_connected_selector.cc
View File

@ -109,6 +109,9 @@ absl::Status SelectFullyConnected(const FullyConnectedAttributes& attr,
return SelectFullyConnectedAdreno(attr, creation_context, op_def, return SelectFullyConnectedAdreno(attr, creation_context, op_def,
batch_size, ptr); batch_size, ptr);
case Vendor::POWERVR: case Vendor::POWERVR:
case Vendor::AMD:
case Vendor::NVIDIA:
case Vendor::INTEL:
return SelectFullyConnectedPowerVR(attr, creation_context, op_def, return SelectFullyConnectedPowerVR(attr, creation_context, op_def,
batch_size, ptr); batch_size, ptr);
case Vendor::MALI: case Vendor::MALI: