COMPMID-2600: Implement a new and generic depthwise convolution for CL QASYMM8 NHWC
The NCHW case is supported at function level by permuting the
inputs/outputs to NHWC.
This patch also removes CLDirectConvolutionLayerOutputStageKernel which
is deprecated and some kernels which were only used in the generic case
of depthwise convolution.
Change-Id: I91e0f02d0a2f4a4a352e08c248e648944137fe68
Signed-off-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-on: https://review.mlplatform.org/c/2056
Reviewed-by: Giorgio Arena <giorgio.arena@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
diff --git a/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp b/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp
index f01b58a..d9c2115 100644
--- a/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp
+++ b/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp
@@ -246,28 +246,44 @@
}
}
-CLDepthwiseConvolutionLayer::CLDepthwiseConvolutionLayer()
- : _im2col_kernel(), _weights_reshape_kernel(), _v2mm_kernel(), _vector_to_tensor_kernel(), _output_stage_kernel(), _activationlayer_function(), _v2mm_input_fill_border(), _v2mm_weights_fill_border(),
- _input_reshaped(), _weights_reshaped(), _v2mm_output(), _output_reshaped(), _is_prepared(false), _is_quantized(false), _is_activationlayer_enabled(false), _original_weights(nullptr),
- _optimised_function(nullptr)
+CLDepthwiseConvolutionLayer::CLDepthwiseConvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager)
+ : _memory_group(std::move(memory_manager)),
+ _optimised_function(nullptr),
+ _dwc_native_kernel(),
+ _permute_input_to_nhwc(),
+ _permute_weights_to_nhwc(),
+ _permute_output_to_nchw(),
+ _permuted_input(),
+ _permuted_weights(),
+ _permuted_output(),
+ _original_weights(),
+ _needs_permute(false),
+ _is_prepared(false)
{
}
void CLDepthwiseConvolutionLayer::configure(ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info,
unsigned int depth_multiplier, const ActivationLayerInfo &act_info, const Size2D &dilation)
{
- ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32);
- ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights);
- ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_LAYOUT(input, output);
+ ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output);
+ ARM_COMPUTE_ERROR_THROW_ON(CLDepthwiseConvolutionLayer::validate(input->info(),
+ weights->info(),
+ biases != nullptr ? biases->info() : nullptr,
+ output->info(),
+ conv_info,
+ depth_multiplier,
+ act_info,
+ dilation));
const size_t idx_w = get_data_layout_dimension_index(input->info()->data_layout(), DataLayoutDimension::WIDTH);
const size_t idx_h = get_data_layout_dimension_index(input->info()->data_layout(), DataLayoutDimension::HEIGHT);
- ARM_COMPUTE_ERROR_ON(weights->info()->dimension(idx_w) + (weights->info()->dimension(idx_w) - 1) * (dilation.x() - 1) > input->info()->dimension(idx_w) + conv_info.pad_left() + conv_info.pad_right());
- ARM_COMPUTE_ERROR_ON(weights->info()->dimension(idx_h) + (weights->info()->dimension(idx_h) - 1) * (dilation.y() - 1) > input->info()->dimension(idx_h) + conv_info.pad_top() + conv_info.pad_bottom());
-
const bool can_run_optimised_3x3_kernel = (weights->info()->dimension(idx_w) == 3) && (weights->info()->dimension(idx_h) == 3);
+ _needs_permute = false;
+ _is_prepared = false;
+ _original_weights = weights;
+
if(bool(can_run_optimised_3x3_kernel))
{
auto f = arm_compute::support::cpp14::make_unique<CLDepthwiseConvolutionLayer3x3>();
@@ -276,103 +292,46 @@
}
else
{
- const size_t idx_c = get_data_layout_dimension_index(input->info()->data_layout(), DataLayoutDimension::CHANNEL);
+ _needs_permute = input->info()->data_layout() == DataLayout::NCHW;
- const size_t weights_w = weights->info()->dimension(idx_w);
- const size_t weights_h = weights->info()->dimension(idx_h);
- const size_t weights_z = weights->info()->dimension(idx_c);
-
- _is_prepared = false;
- _original_weights = weights;
- _is_quantized = is_data_type_quantized_asymmetric(input->info()->data_type());
-
- bool append_bias = (biases != nullptr) && !_is_quantized;
- const GPUTarget gpu_target = CLScheduler::get().target();
-
- // Calculate output shape
- TensorShape output_shape = shape_calculator::compute_depthwise_convolution_shape(*input->info(), *weights->info(), conv_info, depth_multiplier, dilation);
-
- // Output auto inizialitation if not yet initialized
- auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(output_shape));
- ARM_COMPUTE_ERROR_ON_MISMATCHING_DIMENSIONS(output->info()->tensor_shape(), output_shape);
-
- // Output width and height
- const unsigned int conv_w = output_shape[idx_w];
- const unsigned int conv_h = output_shape[idx_h];
-
- // Set up intermediate tensors
- const size_t patch_size = weights_w * weights_h + ((append_bias) ? 1 : 0);
- const size_t conv_size = conv_w * conv_h;
-
- const UniformQuantizationInfo iq_info = input->info()->quantization_info().uniform();
- const UniformQuantizationInfo wq_info = weights->info()->quantization_info().uniform();
- const UniformQuantizationInfo oq_info = output->info()->quantization_info().uniform();
-
- // Im2Col configuration
- TensorShape shape_im2col = input->info()->tensor_shape();
- shape_im2col.set(0, patch_size);
- shape_im2col.set(1, conv_size);
- shape_im2col.set(2, weights_z);
- _input_reshaped.allocator()->init(input->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(shape_im2col));
- _im2col_kernel.set_target(gpu_target);
- _im2col_kernel.configure(input, &_input_reshaped, Size2D(weights_w, weights_h), conv_info, append_bias, depth_multiplier, dilation);
- CLScheduler::get().tune_kernel_static(_im2col_kernel);
-
- // Weights reshape configuration
- const TensorShape shape_weights_reshape(patch_size, weights_z);
- _weights_reshaped.allocator()->init(weights->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(shape_weights_reshape));
- _weights_reshape_kernel.configure(weights, &_weights_reshaped, append_bias ? biases : nullptr);
-
- // GEMV configuration
- DataType v2mm_dt = (input->info()->data_type() == DataType::QASYMM8) ? DataType::S32 : input->info()->data_type();
- TensorShape shape_v2mm_out = input->info()->tensor_shape();
- shape_v2mm_out.set(0, conv_size * weights_z);
- shape_v2mm_out.set(1, 1);
- shape_v2mm_out.set(2, 1);
- _v2mm_output.allocator()->init(input->info()->clone()->set_is_resizable(true).reset_padding().set_data_type(v2mm_dt).set_tensor_shape(shape_v2mm_out));
- _v2mm_kernel.set_target(gpu_target);
- _v2mm_kernel.configure(&_input_reshaped, &_weights_reshaped, &_v2mm_output);
- CLScheduler::get().tune_kernel_static(_v2mm_kernel);
- _output_reshaped.allocator()->init(_v2mm_output.info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(output_shape));
- _vector_to_tensor_kernel.configure(&_v2mm_output, (_is_quantized) ? &_output_reshaped : output, conv_w, conv_h);
-
- // Output staged configuration
- if(_is_quantized)
+ ICLTensor *input_to_use = input;
+ const ICLTensor *weights_to_use = weights;
+ ICLTensor *output_to_use = output;
+ if(_needs_permute)
{
- const UniformQuantizationInfo output_quant_info = (output->info()->total_size() == 0) ? iq_info : oq_info;
+ _memory_group.manage(&_permuted_input);
+ _memory_group.manage(&_permuted_output);
- int output_multiplier = 0;
- int output_shift = 0;
- const float multiplier = iq_info.scale * wq_info.scale / output_quant_info.scale;
- quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift);
- _output_stage_kernel.configure(&_output_reshaped, biases, output, output_multiplier, output_shift, output_quant_info.offset);
- _output_reshaped.allocator()->allocate();
+ // Configure the function to transform the input tensor from NCHW -> NHWC
+ _permute_input_to_nhwc.configure(input, &_permuted_input, PermutationVector(2U, 0U, 1U));
+ _permuted_input.info()->set_data_layout(DataLayout::NHWC);
+
+ // Configure the function to transform the weights tensor from IHW -> HWI
+ _permute_weights_to_nhwc.configure(weights, &_permuted_weights, PermutationVector(2U, 0U, 1U));
+ _permuted_weights.info()->set_data_layout(DataLayout::NHWC);
+
+ // Set output quantization info before dwc kernel configure
+ _permuted_output.info()->set_quantization_info(output->info()->quantization_info());
+
+ input_to_use = &_permuted_input;
+ weights_to_use = &_permuted_weights;
+ output_to_use = &_permuted_output;
}
- // Fill borders on inputs
- PixelValue zero_in(static_cast<int32_t>(0));
- PixelValue zero_w(static_cast<int32_t>(0));
- if(_is_quantized)
+ DWCWeightsKernelInfo dwc_weights_info;
+ dwc_weights_info.n0 = (depth_multiplier == 1) ? 8 : 1;
+ DWCKernelInfo dwc_info;
+ dwc_info.activation_info = act_info;
+ _dwc_native_kernel.configure(input_to_use, weights_to_use, biases, output_to_use, dwc_weights_info, dwc_info, conv_info, depth_multiplier, dilation);
+
+ if(_needs_permute)
{
- zero_in = PixelValue(static_cast<int32_t>(iq_info.offset));
- zero_w = PixelValue(static_cast<int32_t>(wq_info.offset));
- }
- BorderSize border_size = _v2mm_kernel.border_size();
- _v2mm_input_fill_border.configure(&_input_reshaped, border_size, BorderMode::CONSTANT, zero_in);
+ _permuted_input.allocator()->allocate();
- border_size.bottom = 0;
- _v2mm_weights_fill_border.configure(&_weights_reshaped, border_size, BorderMode::CONSTANT, zero_w);
-
- // Allocate intermediate tensors
- _input_reshaped.allocator()->allocate();
- _v2mm_output.allocator()->allocate();
-
- //Configure Activation Layer
- _is_activationlayer_enabled = act_info.enabled();
-
- if(_is_activationlayer_enabled)
- {
- _activationlayer_function.configure(output, nullptr, act_info);
+ // Configure the function to transform the convoluted output to NCHW format
+ _permuted_output.info()->set_data_layout(DataLayout::NCHW);
+ _permute_output_to_nchw.configure(&_permuted_output, output, PermutationVector(1U, 2U, 0U));
+ _permuted_output.allocator()->allocate();
}
}
}
@@ -380,6 +339,8 @@
Status CLDepthwiseConvolutionLayer::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info,
unsigned int depth_multiplier, const ActivationLayerInfo &act_info, const Size2D &dilation)
{
+ ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(input, output);
+
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);
@@ -390,60 +351,36 @@
if(!can_run_optimised_3x3_kernel)
{
- const size_t idx_c = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::CHANNEL);
+ DWCWeightsKernelInfo dwc_weights_info;
+ dwc_weights_info.n0 = (depth_multiplier == 1) ? 8 : 1;
+ DWCKernelInfo dwc_info;
+ dwc_info.activation_info = act_info;
- ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights, output);
- ARM_COMPUTE_RETURN_ERROR_ON((input->dimension(idx_c) * depth_multiplier) != weights->dimension(idx_c));
+ const bool needs_permute = input->data_layout() == DataLayout::NCHW;
- const bool is_quantized = is_data_type_quantized_asymmetric(input->data_type());
- const bool append_bias = (biases != nullptr) && !is_quantized;
- const TensorShape output_shape = shape_calculator::compute_depthwise_convolution_shape(*input, *weights, conv_info, depth_multiplier, dilation);
- const size_t weights_w = weights->dimension(idx_w);
- const size_t weights_h = weights->dimension(idx_h);
- const size_t weights_z = weights->dimension(idx_c);
- const unsigned int conv_w = output_shape[idx_w];
- const unsigned int conv_h = output_shape[idx_h];
- const size_t patch_size = weights_w * weights_h + ((append_bias) ? 1 : 0);
- const size_t conv_size = conv_w * conv_h;
-
- TensorShape shape_im2col = input->tensor_shape();
- shape_im2col.set(0, patch_size);
- shape_im2col.set(1, conv_size);
- shape_im2col.set(2, weights_z);
- TensorInfo input_reshaped(input->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(shape_im2col));
- ARM_COMPUTE_RETURN_ON_ERROR(CLDepthwiseIm2ColKernel::validate(input, &input_reshaped, Size2D(weights_w, weights_h), conv_info, append_bias, depth_multiplier, dilation));
-
- const TensorShape shape_weights_reshape(patch_size, weights_z);
- TensorInfo weights_reshaped(weights->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(shape_weights_reshape));
- ARM_COMPUTE_RETURN_ON_ERROR(CLDepthwiseConvolutionLayerReshapeWeightsGenericKernel::validate(weights, &weights_reshaped, append_bias ? biases : nullptr));
-
- DataType v2mm_dt = (input->data_type() == DataType::QASYMM8) ? DataType::S32 : input->data_type();
- TensorShape shape_v2mm_out = input->tensor_shape();
- shape_v2mm_out.set(0, conv_size * weights_z);
- shape_v2mm_out.set(1, 1);
- shape_v2mm_out.set(2, 1);
- TensorInfo v2mm_output(input->clone()->set_is_resizable(true).reset_padding().set_data_type(v2mm_dt).set_tensor_shape(shape_v2mm_out));
- ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMMatrixVectorMultiplyKernel::validate(&input_reshaped, &weights_reshaped, &v2mm_output));
-
- TensorInfo output_reshaped(v2mm_output.clone()->set_is_resizable(true).reset_padding().set_tensor_shape(output_shape));
- ARM_COMPUTE_RETURN_ON_ERROR(CLDepthwiseVectorToTensorKernel::validate(&v2mm_output, (is_quantized) ? &output_reshaped : output, conv_w, conv_h));
-
- if(is_quantized)
+ if(needs_permute)
{
- const UniformQuantizationInfo iq_info = input->quantization_info().uniform();
- const UniformQuantizationInfo wq_info = weights->quantization_info().uniform();
- const UniformQuantizationInfo oq_info = (output->total_size() == 0) ? iq_info : output->quantization_info().uniform();
+ TensorShape permuted_input_shape = input->tensor_shape();
+ TensorShape permuted_weights_shape = weights->tensor_shape();
+ TensorShape permuted_output_shape = shape_calculator::compute_depthwise_convolution_shape(*input, *weights, conv_info, depth_multiplier, dilation);
- const float multiplier = iq_info.scale * wq_info.scale / oq_info.scale;
- ARM_COMPUTE_UNUSED(multiplier);
- ARM_COMPUTE_RETURN_ERROR_ON(multiplier > 1.0f);
- ARM_COMPUTE_RETURN_ON_ERROR(CLDirectConvolutionLayerOutputStageKernel::validate(&output_reshaped, biases, output));
+ permute(permuted_input_shape, PermutationVector(2U, 0U, 1U));
+ permute(permuted_weights_shape, PermutationVector(2U, 0U, 1U));
+ permute(permuted_output_shape, PermutationVector(2U, 0U, 1U));
+
+ const TensorInfo permuted_input = input->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(permuted_input_shape).set_data_layout(DataLayout::NHWC);
+ const TensorInfo permuted_weights = weights->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(permuted_weights_shape).set_data_layout(DataLayout::NHWC);
+ const TensorInfo permuted_output = output->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(permuted_output_shape).set_data_layout(DataLayout::NHWC);
+
+ ARM_COMPUTE_RETURN_ON_ERROR(CLPermute::validate(input, &permuted_input, PermutationVector(2U, 0U, 1U)));
+ ARM_COMPUTE_RETURN_ON_ERROR(CLPermute::validate(weights, &permuted_weights, PermutationVector(2U, 0U, 1U)));
+ ARM_COMPUTE_RETURN_ON_ERROR(CLDepthwiseConvolutionLayerNativeKernel::validate(&permuted_input, &permuted_weights, biases, &permuted_output, dwc_weights_info,
+ dwc_info, conv_info, depth_multiplier, dilation));
+ ARM_COMPUTE_RETURN_ON_ERROR(CLPermute::validate(&permuted_output, output, PermutationVector(1U, 2U, 0U)));
}
-
- // Validate Activation Layer
- if(act_info.enabled())
+ else
{
- ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayer::validate(output, nullptr, act_info));
+ ARM_COMPUTE_RETURN_ON_ERROR(CLDepthwiseConvolutionLayerNativeKernel::validate(input, weights, biases, output, dwc_weights_info, dwc_info, conv_info, depth_multiplier, dilation));
}
}
else
@@ -457,23 +394,22 @@
{
prepare();
+ MemoryGroupResourceScope scope_mg(_memory_group);
+
if(_optimised_function != nullptr)
{
_optimised_function->run();
}
else
{
- CLScheduler::get().enqueue(_im2col_kernel);
- CLScheduler::get().enqueue(_v2mm_input_fill_border);
- CLScheduler::get().enqueue(_v2mm_kernel);
- CLScheduler::get().enqueue(_vector_to_tensor_kernel);
- if(_is_quantized)
+ if(_needs_permute)
{
- CLScheduler::get().enqueue(_output_stage_kernel);
+ _permute_input_to_nhwc.run();
}
- if(_is_activationlayer_enabled)
+ CLScheduler::get().enqueue(_dwc_native_kernel);
+ if(_needs_permute)
{
- _activationlayer_function.run();
+ _permute_output_to_nchw.run();
}
}
}
@@ -484,21 +420,17 @@
{
_optimised_function->prepare();
}
- else
+ else if(!_is_prepared)
{
- if(!_is_prepared)
+ if(_needs_permute)
{
ARM_COMPUTE_ERROR_ON(!_original_weights->is_used());
- // Run weights reshaping and mark original weights tensor as unused
- _weights_reshaped.allocator()->allocate();
- CLScheduler::get().enqueue(_weights_reshape_kernel);
- CLScheduler::get().enqueue(_v2mm_weights_fill_border);
+ _permuted_weights.allocator()->allocate();
+ _permute_weights_to_nhwc.run();
_original_weights->mark_as_unused();
-
- CLScheduler::get().queue().finish();
- _is_prepared = true;
}
+ _is_prepared = true;
}
}
} // namespace arm_compute