aboutsummaryrefslogtreecommitdiff
diff options
context:
space:
mode:
Diffstat
-rw-r--r--arm_compute/runtime/CL/functions/CLGEMMConvolutionLayer.h4
-rw-r--r--src/graph/GraphBuilder.cpp2
-rw-r--r--src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp2
-rw-r--r--src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp45
-rw-r--r--tests/validation/CL/ConvolutionLayer.cpp4
5 files changed, 37 insertions, 20 deletions
diff --git a/arm_compute/runtime/CL/functions/CLGEMMConvolutionLayer.h b/arm_compute/runtime/CL/functions/CLGEMMConvolutionLayer.h
index 7c272a348b..8538d83c2b 100644
--- a/arm_compute/runtime/CL/functions/CLGEMMConvolutionLayer.h
+++ b/arm_compute/runtime/CL/functions/CLGEMMConvolutionLayer.h
@@ -40,6 +40,7 @@
#include "arm_compute/runtime/CL/functions/CLGEMM.h"
#include "arm_compute/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.h"
#include "arm_compute/runtime/CL/functions/CLGEMMLowpOutputStage.h"
+#include "arm_compute/runtime/CL/functions/CLReshapeLayer.h"
#include "arm_compute/runtime/IMemoryManager.h"
#include <memory>
@@ -88,7 +89,7 @@ private:
* -# @ref CLGEMMLowpMatrixMultiplyCore (if the data type is QASYMM8)
* -# @ref CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint (if the data type is QASYMM8)
* -# @ref CLArithmeticAdditionKernel (if biases != nullptr and we have a 1x1 convolution with the NHWC data layout)
- * -# @ref CLCol2ImKernel (if NCHW data layout)
+ * -# @ref CLCol2ImKernel (if NCHW data layout) or @ref CLReshapeLayer (if NHWC with QASYMM8)
*/
class CLGEMMConvolutionLayer : public IFunction
{
@@ -182,6 +183,7 @@ private:
CLCol2ImKernel _col2im_kernel;
CLActivationLayer _activationlayer_function;
CLArithmeticAdditionKernel _add_bias_kernel;
+ CLReshapeLayer _reshape_layer;
const ICLTensor *_original_weights;
diff --git a/src/graph/GraphBuilder.cpp b/src/graph/GraphBuilder.cpp
index 7b09ec98f5..fa78024e71 100644
--- a/src/graph/GraphBuilder.cpp
+++ b/src/graph/GraphBuilder.cpp
@@ -380,7 +380,7 @@ NodeID GraphBuilder::add_depthwise_convolution_node(Graph &g, NodeParams params,
if(has_bias)
{
TensorDescriptor b_desc = input_tensor_desc;
- b_desc.shape = TensorShape(b_desc.shape.z());
+ b_desc.shape = TensorShape(get_dimension_size(input_tensor_desc, DataLayoutDimension::CHANNEL));
b_nid = add_const_node_with_name(g, params, "Bias", b_desc, std::move(bias_accessor));
}
diff --git a/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp b/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp
index 1d89b84097..76451af9b1 100644
--- a/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp
+++ b/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp
@@ -73,7 +73,7 @@ Status CLDepthwiseConvolutionLayer3x3::validate(const ITensorInfo *input, const
ActivationLayerInfo act_info, GPUTarget gpu_target)
{
ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights, output);
- ARM_COMPUTE_RETURN_ERROR_ON(input->data_layout() != DataLayout::NCHW && input->data_layout() != DataLayout::NHWC);
+ ARM_COMPUTE_RETURN_ERROR_ON(input->data_layout() == DataLayout::UNKNOWN);
if(input->data_layout() == DataLayout::NCHW)
{
diff --git a/src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp b/src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp
index 49549a0ad0..ca6157ef13 100644
--- a/src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp
+++ b/src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp
@@ -91,8 +91,8 @@ void CLConvolutionLayerReshapeWeights::run()
CLGEMMConvolutionLayer::CLGEMMConvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager)
: _memory_group(memory_manager), _reshape_weights(), _im2col_kernel(), _mm_gemm(memory_manager), _mm_gemmlowp(memory_manager), _gemmlowp_output_stage(), _col2im_kernel(), _activationlayer_function(),
- _add_bias_kernel(), _original_weights(nullptr), _im2col_output(), _weights_reshaped(), _gemm_output(), _tmp_output(), _data_layout(DataLayout::NCHW), _append_bias(false), _skip_im2col(false),
- _is_quantized(false), _is_activationlayer_enabled(false), _is_prepared(false)
+ _add_bias_kernel(), _reshape_layer(), _original_weights(nullptr), _im2col_output(), _weights_reshaped(), _gemm_output(), _tmp_output(), _data_layout(DataLayout::NCHW), _append_bias(false),
+ _skip_im2col(false), _is_quantized(false), _is_activationlayer_enabled(false), _is_prepared(false)
{
}
@@ -143,14 +143,13 @@ Status CLGEMMConvolutionLayer::validate_mm(const ITensorInfo *input, const ITens
weights_qa->set_quantization_info(QuantizationInfo(weights_quantization_info.scale, -weights_quantization_info.offset));
// Perform validation step on GEMMLowp
- CLGEMMLowpMatrixMultiplyCore::validate(input_qa.get(), weights_qa.get(), output, gemm_info);
+ return CLGEMMLowpMatrixMultiplyCore::validate(input_qa.get(), weights_qa.get(), output, gemm_info);
}
else
{
// Perform validation step on Matrix multiply function
- CLGEMM::validate(input, weights, nullptr, output, 1.0f, 0.0f, gemm_info);
+ return CLGEMM::validate(input, weights, nullptr, output, 1.0f, 0.0f, gemm_info);
}
- return Status{};
}
void CLGEMMConvolutionLayer::configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info, const WeightsInfo &weights_info,
@@ -283,9 +282,17 @@ void CLGEMMConvolutionLayer::configure(const ICLTensor *input, const ICLTensor *
if(!is_nhwc || _is_quantized)
{
- // Configure and tune Col2Im
- _col2im_kernel.configure(_is_quantized ? gemm_output_staged_to_use : gemm_output_to_use, output, std::make_pair(conv_w, conv_h));
- CLScheduler::get().tune_kernel_static(_col2im_kernel);
+ if(input->info()->data_layout() == DataLayout::NCHW)
+ {
+ // Configure and tune Col2Im
+ _col2im_kernel.configure(_is_quantized ? gemm_output_staged_to_use : gemm_output_to_use, output, std::make_pair(conv_w, conv_h));
+ CLScheduler::get().tune_kernel_static(_col2im_kernel);
+ }
+ else
+ {
+ // Configure reshape layer
+ _reshape_layer.configure(_is_quantized ? gemm_output_staged_to_use : gemm_output_to_use, output);
+ }
}
if(!is_nhwc || _is_quantized)
@@ -316,8 +323,6 @@ Status CLGEMMConvolutionLayer::validate(const ITensorInfo *input, const ITensorI
ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32);
ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights);
ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(input, weights);
- ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->data_type() == DataType::QASYMM8 && input->data_layout() == DataLayout::NHWC,
- "NHWC is unsupported for QASYMM8!");
const DataLayout data_layout = input->data_layout();
const DataType data_type = input->data_type();
@@ -378,7 +383,7 @@ Status CLGEMMConvolutionLayer::validate(const ITensorInfo *input, const ITensorI
// Output tensor auto inizialitation if not yet initialized
ARM_COMPUTE_RETURN_ON_ERROR(CLConvolutionLayerReshapeWeights::validate(weights, is_quantized ? nullptr : biases, nullptr));
- weights_reshaped_info = TensorInfo(compute_weights_reshaped_shape(*weights, append_bias), 1, data_type);
+ weights_reshaped_info = TensorInfo(compute_weights_reshaped_shape(*weights, (append_bias && !skip_im2col)), 1, data_type);
weights_to_use = &weights_reshaped_info;
if(!skip_im2col)
@@ -431,9 +436,12 @@ Status CLGEMMConvolutionLayer::validate(const ITensorInfo *input, const ITensorI
// Validate Col2Im
if(!is_nhwc || is_quantized)
{
- ARM_COMPUTE_RETURN_ON_ERROR(CLCol2ImKernel::validate(is_quantized ? gemm_output_staged_to_use : gemm_output_to_use,
- output,
- std::make_pair(conv_w, conv_h)));
+ if(input->data_layout() == DataLayout::NCHW)
+ {
+ ARM_COMPUTE_RETURN_ON_ERROR(CLCol2ImKernel::validate(is_quantized ? gemm_output_staged_to_use : gemm_output_to_use,
+ output,
+ std::make_pair(conv_w, conv_h)));
+ }
}
//Validate Activation Layer
@@ -480,7 +488,14 @@ void CLGEMMConvolutionLayer::run()
// Reshape output matrix
if(_data_layout == DataLayout::NCHW || _is_quantized)
{
- CLScheduler::get().enqueue(_col2im_kernel, false);
+ if(_data_layout == DataLayout::NCHW)
+ {
+ CLScheduler::get().enqueue(_col2im_kernel, false);
+ }
+ else
+ {
+ _reshape_layer.run();
+ }
}
//Run Activation Layer if enabled
diff --git a/tests/validation/CL/ConvolutionLayer.cpp b/tests/validation/CL/ConvolutionLayer.cpp
index 4ea2eb81a5..54fdc0c386 100644
--- a/tests/validation/CL/ConvolutionLayer.cpp
+++ b/tests/validation/CL/ConvolutionLayer.cpp
@@ -263,7 +263,7 @@ TEST_SUITE(QASYMM8)
FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(combine(datasets::SmallConvolutionLayerDataset(),
framework::dataset::make("ReshapeWeights", { true })),
framework::dataset::make("DataType", DataType::QASYMM8)),
- framework::dataset::make("DataLayout", { DataLayout::NCHW })),
+ framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })),
framework::dataset::make("QuantizationInfo", { QuantizationInfo(2.f / 255.f, 10) })),
QuantizedActivationFunctionsDataset))
{
@@ -273,7 +273,7 @@ FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMConvolutionLayerQuantizedFixture<uint8_t>
FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(combine(datasets::LargeConvolutionLayerDataset(),
framework::dataset::make("ReshapeWeights", { true })),
framework::dataset::make("DataType", DataType::QASYMM8)),
- framework::dataset::make("DataLayout", { DataLayout::NCHW })),
+ framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })),
framework::dataset::make("QuantizationInfo", { QuantizationInfo(2.f / 255.f, 0) })),
QuantizedActivationFunctionsDataset))
{
generated by cgit v1.2.1 (git 2.23.0) at 2024-05-20 20:52:02 +0000