arm_compute v18.11
diff --git a/src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp b/src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp
index 92d04d6..4694aa7 100644
--- a/src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp
+++ b/src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp
@@ -91,16 +91,21 @@
}
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(), _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)
+ : _memory_group(memory_manager), _reshape_weights(), _im2col_kernel(), _mm_gemm(memory_manager), _mm_gemmlowp(memory_manager), _col2im_kernel(), _activationlayer_function(), _add_bias_kernel(),
+ _original_weights(nullptr), _im2col_output(), _weights_reshaped(), _gemm_output(), _data_layout(DataLayout::NCHW), _append_bias(false), _skip_im2col(false), _skip_col2im(false), _is_quantized(false),
+ _is_activationlayer_enabled(false), _is_prepared(false)
{
}
-void CLGEMMConvolutionLayer::configure_mm(const ICLTensor *input, const ICLTensor *weights, ICLTensor *output, int gemm_3d_depth)
+void CLGEMMConvolutionLayer::configure_mm(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const GEMMLowpOutputStageInfo &gemmlowp_output_stage,
+ int gemm_3d_depth)
{
ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights);
- ARM_COMPUTE_ERROR_THROW_ON(validate_mm(input->info(), weights->info(), output->info(), gemm_3d_depth, _skip_im2col));
+ ARM_COMPUTE_ERROR_THROW_ON(validate_mm(input->info(), weights->info(), biases != nullptr ? biases->info() : nullptr, output->info(), gemmlowp_output_stage, gemm_3d_depth, _skip_im2col));
+
+ const GEMMInfo &gemm_info = GEMMInfo(false, false, true /* Reshape weights only for the first run */,
+ gemm_3d_depth, _skip_im2col /* Reinterpret the input as 3D if im2col is skipped */,
+ false, gemmlowp_output_stage);
if(_is_quantized)
{
@@ -112,7 +117,7 @@
input->info()->set_quantization_info(QuantizationInfo(input_quantization_info.scale, -input_quantization_info.offset));
weights->info()->set_quantization_info(QuantizationInfo(weights_quantization_info.scale, -weights_quantization_info.offset));
- _mm_gemmlowp.configure(input, weights, output, GEMMInfo(false, false, true /* Reshape weights only for the first run*/));
+ _mm_gemmlowp.configure(input, weights, biases, output, gemm_info);
// Revert back QuantizatioInfo as input and weights could be used in other convolution layers
input->info()->set_quantization_info(input_quantization_info);
@@ -121,16 +126,19 @@
else
{
// Configure matrix multiply function
- _mm_gemm.configure(input, weights, nullptr, output, 1.0f, 0.0f, GEMMInfo(false, false, true /* Reshape weights only for the first run*/, gemm_3d_depth,
- _skip_im2col /* Reinterpret the input as 3D if im2col is skipped */));
+ _mm_gemm.configure(input, weights, nullptr, output, 1.0f, 0.0f, gemm_info);
}
}
-Status CLGEMMConvolutionLayer::validate_mm(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *output, int gemm_3d_depth, bool skip_im2col)
+Status CLGEMMConvolutionLayer::validate_mm(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output,
+ const GEMMLowpOutputStageInfo &gemmlowp_output_stage, int gemm_3d_depth, bool skip_im2col)
{
const bool is_quantized = is_data_type_quantized_asymmetric(input->data_type());
- const GEMMInfo &gemm_info = GEMMInfo(false, false, true /* Reshape weights only for the first run */, gemm_3d_depth, skip_im2col /* Reinterpret the input as 3D if im2col is skipped */);
+ const GEMMInfo &gemm_info = GEMMInfo(false, false, true /* Reshape weights only for the first run */,
+ gemm_3d_depth, skip_im2col /* Reinterpret the input as 3D if im2col is skipped */,
+ false, gemmlowp_output_stage);
+
if(is_quantized)
{
// Since we need negative offsets for computing convolution, we need to change QuantizationInfo()
@@ -144,7 +152,7 @@
weights_qa->set_quantization_info(QuantizationInfo(weights_quantization_info.scale, -weights_quantization_info.offset));
// Perform validation step on GEMMLowp
- return CLGEMMLowpMatrixMultiplyCore::validate(input_qa.get(), weights_qa.get(), output, gemm_info);
+ return CLGEMMLowpMatrixMultiplyCore::validate(input_qa.get(), weights_qa.get(), biases, output, gemm_info);
}
else
{
@@ -177,21 +185,21 @@
const unsigned int kernel_width = weights->info()->dimension(idx_width);
const unsigned int kernel_height = weights->info()->dimension(idx_height);
- _is_prepared = weights_info.retain_internal_weights();
- _original_weights = weights;
- _is_quantized = is_data_type_quantized_asymmetric(input->info()->data_type());
- _data_layout = data_layout;
- _skip_im2col = (data_layout == DataLayout::NHWC && kernel_width == 1 && kernel_height == 1 && conv_info.stride().first == 1 && conv_info.stride().second == 1) && !_is_quantized;
- _append_bias = (biases != nullptr) && (!_is_quantized);
+ _is_prepared = weights_info.retain_internal_weights();
+ _original_weights = weights;
+ _is_quantized = is_data_type_quantized_asymmetric(input->info()->data_type());
+ _data_layout = data_layout;
+ _skip_im2col = (data_layout == DataLayout::NHWC && kernel_width == 1 && kernel_height == 1 && conv_info.stride().first == 1 && conv_info.stride().second == 1);
+ _skip_col2im = data_layout == DataLayout::NHWC;
+ _append_bias = (biases != nullptr) && (!_is_quantized);
+ _is_activationlayer_enabled = act_info.enabled();
// Set the GPU target for im2col and col2im
_im2col_kernel.set_target(CLScheduler::get().target());
_col2im_kernel.set_target(CLScheduler::get().target());
- bool is_nhwc = _data_layout == DataLayout::NHWC;
- const ICLTensor *gemm_input_to_use = input;
- ICLTensor *gemm_output_to_use = output;
- ICLTensor *gemm_output_staged_to_use = output;
+ const ICLTensor *gemm_input_to_use = input;
+ ICLTensor *gemm_output_to_use = output;
const ICLTensor *biases_to_use = (_append_bias && !_skip_im2col) ? biases : nullptr;
@@ -238,17 +246,18 @@
}
// Create GEMM output tensor
- if(!is_nhwc || _is_quantized)
+ if(!_skip_col2im)
{
- // Calculate GEMM output shape
- TensorShape shape_gemm = _im2col_output.info()->tensor_shape();
+ TensorShape shape_gemm;
+
+ // If we cannot skip col2im it means we run im2col as well
+ shape_gemm = _im2col_output.info()->tensor_shape();
shape_gemm.set(0, mat_weights_cols);
shape_gemm.set(1, conv_w * conv_h);
- // GEMM output should be S32 for acquiring raw integer accumulator without quantized postprocessing for quantized asymmetric input.
- const DataType gemm_data_type = _is_quantized ? DataType::S32 : data_type;
- TensorInfo info_gemm(shape_gemm, 1, gemm_data_type);
- info_gemm.set_quantization_info(output->info()->quantization_info());
+ // FIXME: input->clone() doesn't work with subtensors for grouped convolutions.
+ TensorInfo info_gemm(shape_gemm, 1, data_type);
+ info_gemm.set_quantization_info(output->info()->quantization_info()).set_data_layout(input->info()->data_layout());
_gemm_output.allocator()->init(info_gemm);
_memory_group.manage(&_gemm_output);
@@ -256,56 +265,76 @@
gemm_output_to_use = &_gemm_output;
}
- // Configure and tune GEMM
- configure_mm(gemm_input_to_use, &_weights_reshaped, gemm_output_to_use, (data_layout == DataLayout::NHWC) ? conv_h : 1);
-
- if(!_skip_im2col)
- {
- _im2col_output.allocator()->allocate();
- }
+ GEMMLowpOutputStageInfo gemmlowp_output_stage;
+ gemmlowp_output_stage.type = GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT;
+ gemmlowp_output_stage.gemmlowp_offset = 0;
+ gemmlowp_output_stage.gemmlowp_multiplier = 0;
+ gemmlowp_output_stage.gemmlowp_shift = 0;
// Configure output stage for quantized case
if(_is_quantized)
{
const QuantizationInfo output_quant_info = (output->info()->total_size() == 0) ? input->info()->quantization_info() : output->info()->quantization_info();
- float multiplier = input->info()->quantization_info().scale * weights->info()->quantization_info().scale / output_quant_info.scale;
- int output_multiplier, output_shift;
+ const float multiplier = (input->info()->quantization_info().scale * weights->info()->quantization_info().scale) / output_quant_info.scale;
+ int output_multiplier = 0;
+ int output_shift = 0;
quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift);
- _memory_group.manage(&_tmp_output);
- gemm_output_staged_to_use = &_tmp_output;
+ int min_activation = 0;
+ int max_activation = 0;
- _gemmlowp_output_stage.configure(gemm_output_to_use, biases, gemm_output_staged_to_use, output_multiplier, output_shift, output_quant_info.offset);
+ const std::set<ActivationLayerInfo::ActivationFunction> supported_acts = { ActivationLayerInfo::ActivationFunction::RELU,
+ ActivationLayerInfo::ActivationFunction::BOUNDED_RELU,
+ ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU
+ };
+
+ if(_is_activationlayer_enabled && supported_acts.count(act_info.activation()) != 0)
+ {
+ const int a_const_int = output_quant_info.quantize(act_info.a(), RoundingPolicy::TO_NEAREST_UP);
+ const int b_const_int = output_quant_info.quantize(act_info.b(), RoundingPolicy::TO_NEAREST_UP);
+
+ min_activation = act_info.activation() != ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU ? output_quant_info.offset : b_const_int;
+ max_activation = act_info.activation() == ActivationLayerInfo::ActivationFunction::RELU ? 255 : a_const_int;
+
+ // If the activation layer is RELU, BOUNDED_RELU or LU_BOUNDED_RELU, we can use the GEMMLowp output stage to perform this operation
+ _is_activationlayer_enabled = false;
+ }
+
+ // Set the GEMMLowp output stage info
+ gemmlowp_output_stage.gemmlowp_offset = output_quant_info.offset;
+ gemmlowp_output_stage.gemmlowp_multiplier = output_multiplier;
+ gemmlowp_output_stage.gemmlowp_shift = output_shift;
+ gemmlowp_output_stage.gemmlowp_min_bound = min_activation;
+ gemmlowp_output_stage.gemmlowp_max_bound = max_activation;
}
- if(!is_nhwc || _is_quantized)
+ // Configure and tune GEMM
+ // In case of NHWC, we need to run GEMM3D (gemm_3d_depth != 0) in order to avoid reshaping the output matrix
+ const unsigned int gemm_3d_depth = (data_layout == DataLayout::NHWC) ? conv_h : 0;
+
+ configure_mm(gemm_input_to_use, &_weights_reshaped, biases, gemm_output_to_use, gemmlowp_output_stage, gemm_3d_depth);
+
+ if(!_skip_im2col)
{
- 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), num_groups);
- 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);
- }
+ _im2col_output.allocator()->allocate();
}
- if(!is_nhwc || _is_quantized)
+ if(!_skip_col2im)
{
- _tmp_output.allocator()->allocate();
+ // Configure and tune Col2Im
+ _col2im_kernel.configure(gemm_output_to_use, output, Size2D(conv_w, conv_h), num_groups);
+ CLScheduler::get().tune_kernel_static(_col2im_kernel);
+ }
+
+ if(!_skip_col2im)
+ {
_gemm_output.allocator()->allocate();
}
ARM_COMPUTE_ERROR_ON_MSG((output->info()->dimension(idx_width) != conv_w) || (output->info()->dimension(idx_height) != conv_h),
"Output shape does not match the expected one");
- //Configure Activation Layer
- _is_activationlayer_enabled = act_info.enabled();
-
if(_is_activationlayer_enabled)
{
_activationlayer_function.configure(output, nullptr, act_info);
@@ -336,16 +365,16 @@
const unsigned int kernel_width = weights->dimension(idx_width);
const unsigned int kernel_height = weights->dimension(idx_height);
- TensorInfo im2col_reshaped_info, info_gemm, tmp_info, weights_reshaped_info;
- const ITensorInfo *gemm_input_to_use = input;
- const ITensorInfo *gemm_output_to_use = output;
- const ITensorInfo *gemm_output_staged_to_use = output;
- const ITensorInfo *weights_to_use = weights;
+ TensorInfo im2col_reshaped_info, info_gemm, weights_reshaped_info;
+ const ITensorInfo *gemm_input_to_use = input;
+ const ITensorInfo *gemm_output_to_use = output;
+ const ITensorInfo *weights_to_use = weights;
- const bool is_nhwc = data_layout == DataLayout::NHWC;
- const bool is_quantized = is_data_type_quantized_asymmetric(data_type);
- const bool skip_im2col = (data_layout == DataLayout::NHWC && kernel_width == 1 && kernel_height == 1 && conv_info.stride().first == 1 && conv_info.stride().second == 1) && !is_quantized;
- const bool append_bias = (biases != nullptr) && (!is_quantized);
+ const bool is_quantized = is_data_type_quantized_asymmetric(data_type);
+ const bool append_bias = (biases != nullptr) && (!is_quantized);
+ const bool skip_im2col = (data_layout == DataLayout::NHWC && kernel_width == 1 && kernel_height == 1 && conv_info.stride().first == 1 && conv_info.stride().second == 1);
+ const bool skip_col2im = data_layout == DataLayout::NHWC;
+ bool is_activationlayer_enabled = act_info.enabled();
ARM_COMPUTE_RETURN_ERROR_ON((weights->dimension(idx_channel) * num_groups) != input->dimension(idx_channel));
ARM_COMPUTE_RETURN_ERROR_ON(weights->num_dimensions() > 4);
@@ -407,47 +436,76 @@
}
// Create GEMM output tensor
- if(!is_nhwc || is_quantized)
+ if(!skip_col2im)
{
- TensorShape shape_gemm = gemm_input_to_use->tensor_shape();
+ TensorShape shape_gemm;
+
+ shape_gemm = gemm_input_to_use->tensor_shape();
shape_gemm.set(0, mat_weights_cols);
shape_gemm.set(1, conv_w * conv_h);
- const DataType gemm_data_type = is_quantized ? DataType::S32 : data_type;
- // GEMM output should be S32 for acquiring raw integer accumulator without quantized postprocessing for quantized asymmetric input.
- info_gemm = TensorInfo(shape_gemm, 1, gemm_data_type);
- info_gemm.set_quantization_info(output->quantization_info());
+
+ info_gemm = TensorInfo(shape_gemm, 1, data_type);
+ info_gemm.set_quantization_info(output->quantization_info()).set_data_layout(input->data_layout());
gemm_output_to_use = &info_gemm;
}
- ARM_COMPUTE_RETURN_ON_ERROR(validate_mm(gemm_input_to_use, weights_to_use, gemm_output_to_use, (data_layout == DataLayout::NHWC) ? conv_h : 1, skip_im2col));
+ GEMMLowpOutputStageInfo gemmlowp_output_stage;
+ gemmlowp_output_stage.type = GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT;
+ gemmlowp_output_stage.gemmlowp_offset = 0;
+ gemmlowp_output_stage.gemmlowp_multiplier = 0;
+ gemmlowp_output_stage.gemmlowp_shift = 0;
if(is_quantized)
{
- float multiplier = input->quantization_info().scale * weights_to_use->quantization_info().scale / output->quantization_info().scale;
- int output_multiplier, output_shift;
- quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift);
+ const QuantizationInfo output_quant_info = (output->total_size() == 0) ? input->quantization_info() : output->quantization_info();
- tmp_info = TensorInfo(gemm_output_to_use->tensor_shape(), 1, DataType::QASYMM8);
- tmp_info.set_quantization_info(output->quantization_info());
- gemm_output_staged_to_use = &tmp_info;
+ const float multiplier = (input->quantization_info().scale * weights->quantization_info().scale) / output_quant_info.scale;
+ int output_multiplier = 0;
+ int output_shift = 0;
- // Validate output stage for quantized case
- CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint::validate(gemm_output_to_use, biases, gemm_output_staged_to_use, output->quantization_info().offset);
+ ARM_COMPUTE_RETURN_ON_ERROR(quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift));
+
+ int min_activation = 0;
+ int max_activation = 0;
+
+ const std::set<ActivationLayerInfo::ActivationFunction> supported_acts = { ActivationLayerInfo::ActivationFunction::RELU,
+ ActivationLayerInfo::ActivationFunction::BOUNDED_RELU,
+ ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU
+ };
+
+ if(is_activationlayer_enabled && supported_acts.count(act_info.activation()) != 0)
+ {
+ const int a_const_int = output_quant_info.quantize(act_info.a(), RoundingPolicy::TO_NEAREST_UP);
+ const int b_const_int = output_quant_info.quantize(act_info.b(), RoundingPolicy::TO_NEAREST_UP);
+
+ min_activation = act_info.activation() != ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU ? output_quant_info.offset : b_const_int;
+ max_activation = act_info.activation() == ActivationLayerInfo::ActivationFunction::RELU ? 255 : a_const_int;
+
+ // If the activation layer is RELU, BOUNDED_RELU or LU_BOUNDED_RELU, we can use the GEMMLowp output stage to perform this operation
+ is_activationlayer_enabled = false;
+ }
+
+ // Set the GEMMLowp output stage info
+ gemmlowp_output_stage.gemmlowp_offset = output_quant_info.offset;
+ gemmlowp_output_stage.gemmlowp_multiplier = output_multiplier;
+ gemmlowp_output_stage.gemmlowp_shift = output_shift;
+ gemmlowp_output_stage.gemmlowp_min_bound = min_activation;
+ gemmlowp_output_stage.gemmlowp_max_bound = max_activation;
}
+ // In case of NHWC, we need to run GEMM3D (gemm_3d_depth != 0) in order to avoid reshaping the output matrix
+ const unsigned int gemm_3d_depth = (data_layout == DataLayout::NHWC) ? conv_h : 0;
+
+ ARM_COMPUTE_RETURN_ON_ERROR(validate_mm(gemm_input_to_use, weights_to_use, biases, gemm_output_to_use, gemmlowp_output_stage, gemm_3d_depth, skip_im2col));
+
// Validate Col2Im
- if(!is_nhwc || is_quantized)
+ if(!skip_col2im)
{
- 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), num_groups));
- }
+ ARM_COMPUTE_RETURN_ON_ERROR(CLCol2ImKernel::validate(gemm_output_to_use, output, Size2D(conv_w, conv_h), num_groups));
}
//Validate Activation Layer
- if(act_info.enabled())
+ if(is_activationlayer_enabled)
{
ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayer::validate(output, nullptr, act_info));
}
@@ -472,9 +530,6 @@
{
// Run gemmlowp
_mm_gemmlowp.run();
-
- // Run output stage
- _gemmlowp_output_stage.run();
}
else
{
@@ -488,16 +543,9 @@
}
// Reshape output matrix
- if(_data_layout == DataLayout::NCHW || _is_quantized)
+ if(!_skip_col2im)
{
- if(_data_layout == DataLayout::NCHW)
- {
- CLScheduler::get().enqueue(_col2im_kernel, false);
- }
- else
- {
- _reshape_layer.run();
- }
+ CLScheduler::get().enqueue(_col2im_kernel, false);
}
//Run Activation Layer if enabled