arm_compute v18.08
diff --git a/src/runtime/CL/functions/CLConvolutionLayer.cpp b/src/runtime/CL/functions/CLConvolutionLayer.cpp
index 47a8d5f..0014e71 100644
--- a/src/runtime/CL/functions/CLConvolutionLayer.cpp
+++ b/src/runtime/CL/functions/CLConvolutionLayer.cpp
@@ -43,17 +43,18 @@
}
void CLConvolutionLayer::configure(ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info, const WeightsInfo &weights_info,
- const Size2D &dilation, const ActivationLayerInfo &act_info, bool enable_fast_math)
+ const Size2D &dilation, const ActivationLayerInfo &act_info, bool enable_fast_math, unsigned int num_groups)
{
ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output);
ARM_COMPUTE_ERROR_THROW_ON(CLConvolutionLayer::validate(input->info(), weights->info(), ((biases != nullptr) ? biases->info() : nullptr), output->info(), conv_info, weights_info, dilation, act_info,
- enable_fast_math));
+ enable_fast_math, num_groups));
switch(CLConvolutionLayer::get_convolution_method(input->info(), weights->info(), output->info(), conv_info,
weights_info, act_info, CLScheduler::get().target(), dilation, enable_fast_math))
{
case ConvolutionMethod::WINOGRAD:
{
+ ARM_COMPUTE_ERROR_ON(num_groups != 1);
auto f = arm_compute::support::cpp14::make_unique<CLWinogradConvolutionLayer>(_memory_manager);
f->configure(input, weights, biases, output, conv_info, act_info, enable_fast_math);
_function = std::move(f);
@@ -61,6 +62,7 @@
}
case ConvolutionMethod::DIRECT:
{
+ ARM_COMPUTE_ERROR_ON(num_groups != 1);
auto f = arm_compute::support::cpp14::make_unique<CLDirectConvolutionLayer>();
f->configure(input, weights, biases, output, conv_info, act_info);
_function = std::move(f);
@@ -69,7 +71,7 @@
case ConvolutionMethod::GEMM:
{
auto f = arm_compute::support::cpp14::make_unique<CLGEMMConvolutionLayer>(_memory_manager);
- f->configure(input, weights, biases, output, conv_info, weights_info, dilation, act_info);
+ f->configure(input, weights, biases, output, conv_info, weights_info, dilation, act_info, num_groups);
_function = std::move(f);
break;
}
@@ -80,9 +82,10 @@
}
Status CLConvolutionLayer::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info,
- const WeightsInfo &weights_info, const Size2D &dilation, const ActivationLayerInfo &act_info, bool enable_fast_math)
+ const WeightsInfo &weights_info, const Size2D &dilation, const ActivationLayerInfo &act_info, bool enable_fast_math, unsigned int num_groups)
{
ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights, output);
+ ARM_COMPUTE_RETURN_ERROR_ON_MSG((num_groups != 1) && (input->data_layout() != DataLayout::NCHW), "Grouping (num_groups != 1) with NHWC data layout is not supported");
const GPUTarget gpu_target = CLScheduler::get().target();
@@ -91,19 +94,21 @@
case ConvolutionMethod::WINOGRAD:
{
//Validate Winograd
+ ARM_COMPUTE_RETURN_ERROR_ON_MSG(num_groups != 1, "Grouping (num_groups != 1) with CLWinogradConvolutionLayer is not supported");
ARM_COMPUTE_RETURN_ON_ERROR(CLWinogradConvolutionLayer::validate(input, weights, biases, output, conv_info, act_info, enable_fast_math));
break;
}
case ConvolutionMethod::DIRECT:
{
// Validate direct convolution layer
+ ARM_COMPUTE_RETURN_ERROR_ON_MSG(num_groups != 1, "Grouping (num_groups != 1) with CLDirectConvolutionLayer is not supported");
ARM_COMPUTE_RETURN_ON_ERROR(CLDirectConvolutionLayer::validate(input, weights, biases, output, conv_info, act_info));
break;
}
case ConvolutionMethod::GEMM:
{
// Validate gemm-based convolution layer
- ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMConvolutionLayer::validate(input, weights, biases, output, conv_info, weights_info, dilation, act_info));
+ ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMConvolutionLayer::validate(input, weights, biases, output, conv_info, weights_info, dilation, act_info, num_groups));
break;
}
default:
@@ -123,8 +128,47 @@
ARM_COMPUTE_UNUSED(weights_info);
ARM_COMPUTE_UNUSED(gpu_target);
+ 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);
const size_t idx_c = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::CHANNEL);
+ /* Input spatial dims, kernel size, IFM/OFM, conv info*/
+ using ConvolutionConfiguration = std::tuple<Size2D, Size2D, Size2D, PadStrideInfo, DataLayout>;
+ using ConfigurationMethod = std::pair<ConvolutionConfiguration, ConvolutionMethod>;
+
+ const std::vector<ConfigurationMethod> known_configs =
+ {
+ // Alexnet
+ ConfigurationMethod(ConvolutionConfiguration(Size2D(27U, 27U), Size2D(5U, 5U), Size2D(48U, 128U), PadStrideInfo(1U, 1U, 2U, 2U), DataLayout::NCHW), ConvolutionMethod::DIRECT),
+ // VGG16 / VGG19
+ ConfigurationMethod(ConvolutionConfiguration(Size2D(224U, 224U), Size2D(3U, 3U), Size2D(3U, 64U), PadStrideInfo(1U, 1U, 1U, 1U), DataLayout::NCHW), ConvolutionMethod::DIRECT),
+ // Mobilenet 224
+ ConfigurationMethod(ConvolutionConfiguration(Size2D(224U, 224U), Size2D(3U, 3U), Size2D(3U, 32U), PadStrideInfo(2U, 2U, 0U, 1U, 0U, 1U, DimensionRoundingType::FLOOR), DataLayout::NCHW), ConvolutionMethod::GEMM),
+ // Mobilenet 160
+ ConfigurationMethod(ConvolutionConfiguration(Size2D(160U, 160U), Size2D(3U, 3U), Size2D(3U, 24U), PadStrideInfo(2U, 2U, 0U, 1U, 0U, 1U, DimensionRoundingType::FLOOR), DataLayout::NCHW), ConvolutionMethod::GEMM),
+ // Mobilenet 224
+ ConfigurationMethod(ConvolutionConfiguration(Size2D(224U, 224U), Size2D(3U, 3U), Size2D(3U, 32U), PadStrideInfo(2U, 2U, 0U, 1U, 0U, 1U, DimensionRoundingType::FLOOR), DataLayout::NHWC), ConvolutionMethod::GEMM),
+ // Mobilenet 160
+ ConfigurationMethod(ConvolutionConfiguration(Size2D(160U, 160U), Size2D(3U, 3U), Size2D(3U, 24U), PadStrideInfo(2U, 2U, 0U, 1U, 0U, 1U, DimensionRoundingType::FLOOR), DataLayout::NHWC), ConvolutionMethod::GEMM),
+ };
+
+ const auto find_config = [&](ConfigurationMethod c)
+ {
+ const ConvolutionConfiguration config = c.first;
+ const PadStrideInfo info = std::get<3>(config);
+ const DataLayout data_layout = std::get<4>(config);
+
+ return std::get<0>(config) == Size2D(input->dimension(idx_w), input->dimension(idx_h)) && std::get<1>(config) == Size2D(weights->dimension(idx_w), weights->dimension(idx_h))
+ && std::get<2>(config) == Size2D(weights->dimension(idx_c), weights->dimension(3)) && info.pad_top() == conv_info.pad_top() && info.pad_right() == conv_info.pad_right()
+ && info.pad_bottom() == conv_info.pad_bottom() && info.pad_left() == conv_info.pad_left() && info.stride() == conv_info.stride() && (data_layout == input->data_layout());
+ };
+
+ std::vector<ConfigurationMethod>::const_iterator found;
+ if((found = std::find_if(known_configs.begin(), known_configs.end(), find_config)) != known_configs.end())
+ {
+ return (*found).second;
+ }
+
if(dilation != Size2D(1U, 1U) || (input->dimension(idx_c) < 16))
{
return ConvolutionMethod::GEMM;