COMPMID-790 - NEON: Add QASYMM8 support to Convolution
Change-Id: Iec82a91ad351cfe8d07d0976a24bd42f4703177a
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/116833
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
diff --git a/src/runtime/NEON/functions/NEConvolutionLayer.cpp b/src/runtime/NEON/functions/NEConvolutionLayer.cpp
index 8f7d940..bb685c6 100644
--- a/src/runtime/NEON/functions/NEConvolutionLayer.cpp
+++ b/src/runtime/NEON/functions/NEConvolutionLayer.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -29,6 +29,7 @@
#include "arm_compute/core/Size2D.h"
#include "arm_compute/core/Utils.h"
#include "arm_compute/core/Validate.h"
+#include "arm_compute/core/utils/quantization/AsymmHelpers.h"
#include "arm_compute/runtime/NEON/NEScheduler.h"
#include "support/ToolchainSupport.h"
@@ -46,10 +47,10 @@
{
namespace
{
-TensorShape get_reshaped_weights_shape(const ITensorInfo *weights, bool has_bias)
+TensorShape get_reshaped_weights_shape(const ITensorInfo *weights, bool append_bias)
{
const unsigned int mat_weights_cols = weights->dimension(3);
- const unsigned int mat_weights_rows = weights->dimension(0) * weights->dimension(1) * weights->dimension(2) + (has_bias ? 1 : 0);
+ const unsigned int mat_weights_rows = weights->dimension(0) * weights->dimension(1) * weights->dimension(2) + (append_bias ? 1 : 0);
return TensorShape(mat_weights_cols, mat_weights_rows);
}
} // namespace
@@ -69,14 +70,16 @@
transpose1xW));
// Check if bias are present, if yes they will be embedded to the weights matrix
- const bool _has_bias = (biases != nullptr);
+ const bool append_biases = (biases != nullptr) && !is_data_type_quantized_asymmetric(weights->info()->data_type());
+ //const unsigned bias_element = (append_biases) ? 1 : 0;
+ const ITensor *biases_to_use = (append_biases) ? biases : nullptr;
_transpose1xW = transpose1xW;
if(transpose1xW)
{
// Create tensor to store the reshaped weights
- TensorInfo info_wr = weights->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(get_reshaped_weights_shape(weights->info(), _has_bias));
+ TensorInfo info_wr = weights->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(get_reshaped_weights_shape(weights->info(), append_biases));
_weights_reshaped.allocator()->init(info_wr);
_memory_group.manage(&_weights_reshaped);
@@ -88,30 +91,35 @@
}
else
{
- _weights_reshape_kernel.configure(weights, biases, output);
+ _weights_reshape_kernel.configure(weights, biases_to_use, output);
}
+
+ output->info()->set_quantization_info(weights->info()->quantization_info());
}
Status NEConvolutionLayerReshapeWeights::validate(const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, bool transpose1xW)
{
- ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(weights, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32);
- ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(weights, output);
- ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(weights, output);
+ ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(weights, 1, DataType::QS8, DataType::QASYMM8, DataType::QS16, DataType::F16, DataType::F32);
ARM_COMPUTE_RETURN_ERROR_ON(weights->num_dimensions() > 4);
-
- if(biases != nullptr)
+ if(!is_data_type_quantized_asymmetric(weights->data_type()))
{
+ ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(weights, output);
+ ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(weights, output);
+ }
+ // Check if bias are present, if yes they will be embedded to the weights matrix
+ const bool append_bias = (biases != nullptr);
+
+ if(append_bias)
+ {
+ ARM_COMPUTE_ERROR_ON(is_data_type_quantized_asymmetric(weights->data_type()));
ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(weights, biases);
ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(weights, biases);
ARM_COMPUTE_RETURN_ERROR_ON(biases->dimension(0) != weights->dimension(3));
ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1);
}
- // Check if bias are present, if yes they will be embedded to the weights matrix
- const bool has_bias = (biases != nullptr);
-
// Checks performed when biases are present
- if(has_bias)
+ if(append_bias)
{
ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(weights, biases);
ARM_COMPUTE_RETURN_ERROR_ON(biases->dimension(0) != weights->dimension(3));
@@ -120,7 +128,7 @@
if(transpose1xW)
{
- TensorInfo weights_reshaped = weights->clone()->set_tensor_shape(get_reshaped_weights_shape(weights, has_bias));
+ TensorInfo weights_reshaped = weights->clone()->set_tensor_shape(get_reshaped_weights_shape(weights, append_bias));
ARM_COMPUTE_RETURN_ON_ERROR(NEWeightsReshapeKernel::validate(weights, biases, &weights_reshaped));
ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMTranspose1xWKernel::validate(&weights_reshaped, output));
}
@@ -148,10 +156,10 @@
namespace
{
-TensorShape get_reshaped_weights_shape_conv(const ITensorInfo *weights, bool has_bias, bool is_fully_connected_convolution)
+TensorShape get_reshaped_weights_shape_conv(const ITensorInfo *weights, bool append_bias, bool is_fully_connected_convolution)
{
unsigned int mat_weights_cols = weights->dimension(3);
- unsigned int mat_weights_rows = weights->dimension(0) * weights->dimension(1) * weights->dimension(2) + (has_bias ? 1 : 0);
+ unsigned int mat_weights_rows = weights->dimension(0) * weights->dimension(1) * weights->dimension(2) + (append_bias ? 1 : 0);
if(is_fully_connected_convolution)
{
@@ -167,47 +175,86 @@
}
Status validate_and_initialize_values(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const PadStrideInfo &conv_info, const WeightsInfo &weights_info, DataType &dt,
- bool &has_bias,
- bool &are_weights_reshaped, unsigned int &kernel_width, unsigned int &kernel_height, bool &is_fully_connected_convolution, unsigned int &mat_weights_cols, unsigned int &mat_weights_rows,
+ bool &append_bias,
+ bool &are_weights_reshaped, unsigned int &kernel_width, unsigned int &kernel_height,
+ bool &is_fully_connected_convolution, bool &is_interleaved_transposed, bool &is_quantized,
+ unsigned int &mat_weights_cols, unsigned int &mat_weights_rows,
unsigned int &conv_w, unsigned int &conv_h)
{
- ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32);
+ ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QASYMM8, DataType::QS16, DataType::F16, DataType::F32);
ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights);
ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, weights);
ARM_COMPUTE_RETURN_ERROR_ON(!weights_info.are_reshaped() && weights->dimension(2) != input->dimension(2));
ARM_COMPUTE_RETURN_ERROR_ON(weights->num_dimensions() > 4);
+ ARM_COMPUTE_ERROR_ON(weights_info.are_reshaped() && is_data_type_quantized_asymmetric(input->data_type()));
+
+ dt = input->data_type();
+ is_quantized = is_data_type_quantized_asymmetric(dt);
if(biases != nullptr)
{
- ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases);
+ if(is_quantized)
+ {
+ ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(biases, 1, DataType::S32);
+ }
+ else
+ {
+ ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases);
+ }
ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, biases);
ARM_COMPUTE_RETURN_ERROR_ON(!weights_info.are_reshaped() && biases->dimension(0) != weights->dimension(3));
ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1);
}
- dt = input->data_type();
- has_bias = (biases != nullptr);
+ append_bias = (biases != nullptr) && (!is_quantized);
are_weights_reshaped = weights_info.are_reshaped();
kernel_width = (are_weights_reshaped) ? weights_info.kernel_size().first : weights->dimension(0);
kernel_height = (are_weights_reshaped) ? weights_info.kernel_size().second : weights->dimension(1);
mat_weights_cols = weights->dimension(3);
- mat_weights_rows = weights->dimension(0) * weights->dimension(1) * weights->dimension(2) + (has_bias ? 1 : 0);
+ mat_weights_rows = weights->dimension(0) * weights->dimension(1) * weights->dimension(2) + (append_bias ? 1 : 0);
std::tie(conv_w, conv_h) = scaled_dimensions(input->dimension(0), input->dimension(1), kernel_width, kernel_height,
conv_info);
+ // Check if its a "fully connected" convolution
is_fully_connected_convolution = ((conv_w == 1) && (conv_h == 1));
+ is_interleaved_transposed = (!is_fully_connected_convolution && !is_quantized);
return Status{};
}
} // namespace
NEConvolutionLayer::NEConvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager)
- : _memory_group(std::move(memory_manager)), _input_im2col_kernel(), _input_interleave_kernel(), _reshape_weights(), _mm_kernel(), _mm_optimised_kernel(nullptr), _output_col2im_kernel(),
- _input_im2col_reshaped(), _input_interleaved_reshaped(), _weights_reshaped(), _gemm_output(), _workspace(), _has_bias(false), _is_fully_connected_convolution(false), _are_weights_reshaped(false)
+ : _memory_group(std::move(memory_manager)), _input_im2col_kernel(), _input_interleave_kernel(), _reshape_weights(), _mm_kernel(), _mm_optimised_kernel(nullptr), _mm_gemmlowp(memory_manager),
+ _gemmlowp_output_stage(), _output_col2im_kernel(), _input_im2col_reshaped(), _input_interleaved_reshaped(), _weights_reshaped(), _gemm_output(), _tmp_output(), _workspace(), _append_bias(false),
+ _is_fully_connected_convolution(false), _are_weights_reshaped(false), _is_quantized(false), _is_interleaved_transposed(false)
{
}
+void NEConvolutionLayer::configure_mm(const ITensor *input, const ITensor *weights, ITensor *output)
+{
+ if(_is_quantized)
+ {
+ // Since we need negative offsets for computing convolution, we need to change QuantizationInfo()
+ // Extract and negate input and weights offset
+ const QuantizationInfo input_quantization_info = input->info()->quantization_info();
+ const QuantizationInfo weights_quantization_info = weights->info()->quantization_info();
+
+ 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*/));
+
+ // Revert back QuantizatioInfo as input and weights could be used in other convolution layers
+ input->info()->set_quantization_info(input_quantization_info);
+ weights->info()->set_quantization_info(weights_quantization_info);
+ }
+ else
+ {
+ _mm_kernel.configure(input, weights, output, 1.f);
+ }
+}
+
void NEConvolutionLayer::configure(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info, const WeightsInfo &weights_info)
{
// Perform validate step
@@ -221,14 +268,15 @@
unsigned int conv_w = 0;
unsigned int conv_h = 0;
- Status status = validate_and_initialize_values(input->info(), weights->info(), (biases == nullptr) ? nullptr : biases->info(), conv_info, weights_info, dt, _has_bias, _are_weights_reshaped,
+ Status status = validate_and_initialize_values(input->info(), weights->info(), (biases == nullptr) ? nullptr : biases->info(), conv_info, weights_info, dt, _append_bias, _are_weights_reshaped,
kernel_width, kernel_height,
- _is_fully_connected_convolution,
+ _is_fully_connected_convolution, _is_interleaved_transposed, _is_quantized,
mat_weights_cols, mat_weights_rows, conv_w, conv_h);
ARM_COMPUTE_ERROR_THROW_ON(status);
const unsigned int fixed_point_position = input->info()->fixed_point_position();
+ const ITensor *biases_to_use = (_append_bias) ? biases : nullptr;
#if defined(__arm__)
if(NEScheduler::get().cpu_info().CPU == CPUTarget::ARMV7 && dt == DataType::F32)
@@ -264,7 +312,7 @@
{
if(_are_weights_reshaped)
{
- if(_is_fully_connected_convolution)
+ if(_is_fully_connected_convolution || _is_quantized)
{
mat_weights_cols = weights_info.num_kernels();
mat_weights_rows = weights->info()->dimension(1);
@@ -273,14 +321,14 @@
{
const unsigned int transpose_width = 16 / input->info()->element_size();
mat_weights_cols = weights_info.num_kernels();
- mat_weights_rows = weights->info()->dimension(0) / transpose_width + (_has_bias ? 1 : 0);
+ mat_weights_rows = weights->info()->dimension(0) / transpose_width + (_append_bias ? 1 : 0);
}
}
else
{
TensorShape reshaped_weights_shape;
- if(_is_fully_connected_convolution)
+ if(_is_fully_connected_convolution || _is_quantized)
{
reshaped_weights_shape = TensorShape{ mat_weights_cols, mat_weights_rows };
}
@@ -294,7 +342,7 @@
// Create tensor to store the reshaped weights
_weights_reshaped.allocator()->init(TensorInfo(reshaped_weights_shape, 1, dt, fixed_point_position));
- _reshape_weights.configure(weights, biases, &_weights_reshaped, !_is_fully_connected_convolution /* 1xW transpose */);
+ _reshape_weights.configure(weights, biases_to_use, &_weights_reshaped, _is_interleaved_transposed /* 1xW transpose */);
weights = &_weights_reshaped;
}
}
@@ -324,12 +372,18 @@
TensorShape shape_gemm(_input_im2col_reshaped.info()->tensor_shape());
shape_gemm.set(0, mat_weights_cols);
shape_gemm.set(1, mat_input_rows);
- _gemm_output.allocator()->init(_input_im2col_reshaped.info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(shape_gemm));
+ const DataType gemm_data_type = _is_quantized ? DataType::S32 : dt;
+ // GEMM output should be S32 for acquiring raw integer accumulator without quantized postprocessing for quantized asymmetric input.
+ TensorInfo info_gemm(shape_gemm, 1, gemm_data_type, input->info()->fixed_point_position());
+ info_gemm.set_quantization_info(output->info()->quantization_info());
+ _gemm_output.allocator()->init(info_gemm);
_memory_group.manage(&_gemm_output);
// Configure kernels
- _input_im2col_kernel.configure(input, &_input_im2col_reshaped, Size2D(kernel_width, kernel_height), conv_info, _has_bias);
+ // Configure im2col
+ _input_im2col_kernel.configure(input, &_input_im2col_reshaped, Size2D(kernel_width, kernel_height), conv_info, _append_bias);
+ // Configure matrix multiply
#if defined(__arm__) || defined(__aarch64__)
if(_mm_optimised_kernel != nullptr)
{
@@ -357,22 +411,44 @@
else
#endif /* defined(__arm__) || defined(__aarch64__) */
{
- if(_is_fully_connected_convolution)
+ if(_is_interleaved_transposed)
{
- _mm_kernel.configure(&_input_im2col_reshaped, weights, &_gemm_output, 1.0f);
+ // Configure GEMMInterleave4x4. _input_interleaved_reshaped will be auto configured in the kernel
+ _memory_group.manage(&_input_interleaved_reshaped);
+ _input_interleave_kernel.configure(&_input_im2col_reshaped, &_input_interleaved_reshaped);
+
+ // Configure GEMM
+ configure_mm(&_input_interleaved_reshaped, weights, &_gemm_output);
+ _input_interleaved_reshaped.allocator()->allocate();
}
else
{
- _input_interleave_kernel.configure(&_input_im2col_reshaped, &_input_interleaved_reshaped);
- _mm_kernel.configure(&_input_interleaved_reshaped, weights, &_gemm_output, 1.0f);
- _input_interleaved_reshaped.allocator()->allocate();
+ configure_mm(&_input_im2col_reshaped, weights, &_gemm_output);
}
}
_input_im2col_reshaped.allocator()->allocate();
- _output_col2im_kernel.configure(&_gemm_output, output, Size2D(conv_w, conv_h));
+
+ // Configure output stage for quantized case
+ if(_is_quantized)
+ {
+ float multiplier = input->info()->quantization_info().scale * weights->info()->quantization_info().scale / output->info()->quantization_info().scale;
+ int output_multiplier, output_shift;
+ quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift);
+ _memory_group.manage(&_tmp_output);
+ _gemmlowp_output_stage.configure(&_gemm_output, biases, &_tmp_output, output_multiplier, output_shift, output->info()->quantization_info().offset);
+ }
+
+ // Configure Col2Im
+ _output_col2im_kernel.configure(_is_quantized ? &_tmp_output : &_gemm_output, output, Size2D(conv_w, conv_h));
+ if(_is_quantized)
+ {
+ _tmp_output.allocator()->allocate();
+ }
_gemm_output.allocator()->allocate();
+ ARM_COMPUTE_ERROR_ON_MSG((output->info()->dimension(0) != conv_w) || (output->info()->dimension(1) != conv_h), "Output shape does not match the expected one");
+
// Allocate intermediate tensor
if(!_are_weights_reshaped)
{
@@ -384,9 +460,11 @@
const WeightsInfo &weights_info)
{
DataType dt{};
- bool has_bias{};
+ bool append_bias{};
bool are_weights_reshaped{};
bool is_fully_connected_convolution{};
+ bool is_interleaved_transposed{};
+ bool is_quantized{};
unsigned int kernel_width = 0;
unsigned int kernel_height = 0;
unsigned int mat_weights_cols = 0;
@@ -394,8 +472,8 @@
unsigned int conv_w = 0;
unsigned int conv_h = 0;
- Status status = validate_and_initialize_values(input, weights, biases, conv_info, weights_info, dt, has_bias, are_weights_reshaped, kernel_width, kernel_height,
- is_fully_connected_convolution, mat_weights_cols, mat_weights_rows,
+ Status status = validate_and_initialize_values(input, weights, biases, conv_info, weights_info, dt, append_bias, are_weights_reshaped, kernel_width, kernel_height,
+ is_fully_connected_convolution, is_interleaved_transposed, is_quantized, mat_weights_cols, mat_weights_rows,
conv_w, conv_h);
ARM_COMPUTE_RETURN_ON_ERROR(status);
@@ -428,7 +506,7 @@
TensorShape reshaped_weights_shape{ mat_weights_cols, mat_weights_rows };
// Create tensor to store the reshaped weights
- reshaped_weights->set_tensor_shape(get_reshaped_weights_shape_conv(weights, has_bias, is_fully_connected_convolution));
+ reshaped_weights->set_tensor_shape(get_reshaped_weights_shape_conv(weights, append_bias, is_fully_connected_convolution));
ARM_COMPUTE_RETURN_ON_ERROR(NEConvolutionLayerReshapeWeights::validate(weights, biases, reshaped_weights.get(), !is_fully_connected_convolution /* 1xW transpose */));
weights = reshaped_weights.get();
}
@@ -439,13 +517,13 @@
{
const unsigned int transpose_width = 16 / input->element_size();
mat_weights_cols = weights_info.num_kernels();
- mat_weights_rows = weights->dimension(0) / transpose_width + (has_bias ? 1 : 0);
+ mat_weights_rows = weights->dimension(0) / transpose_width + (append_bias ? 1 : 0);
}
else
{
TensorShape reshaped_weights_shape;
- if(is_fully_connected_convolution)
+ if(is_fully_connected_convolution || is_quantized)
{
reshaped_weights_shape = TensorShape{ mat_weights_cols, mat_weights_rows };
}
@@ -458,7 +536,7 @@
}
// Create tensor to store the reshaped weights
- reshaped_weights->set_tensor_shape(get_reshaped_weights_shape_conv(weights, has_bias, is_fully_connected_convolution));
+ reshaped_weights->set_tensor_shape(get_reshaped_weights_shape_conv(weights, append_bias, is_fully_connected_convolution));
ARM_COMPUTE_RETURN_ON_ERROR(NEConvolutionLayerReshapeWeights::validate(weights, biases, reshaped_weights.get(), !is_fully_connected_convolution /* 1xW transpose */));
weights = reshaped_weights.get();
}
@@ -472,7 +550,7 @@
shape_im2col.set(1, mat_input_rows);
shape_im2col.set(2, 1);
TensorInfo im2_col_info = input->clone()->set_tensor_shape(shape_im2col);
- ARM_COMPUTE_RETURN_ON_ERROR(NEIm2ColKernel::validate(input, &im2_col_info, Size2D(weights->dimension(0), weights->dimension(1)), conv_info, has_bias));
+ ARM_COMPUTE_RETURN_ON_ERROR(NEIm2ColKernel::validate(input, &im2_col_info, Size2D(weights->dimension(0), weights->dimension(1)), conv_info, append_bias));
// Create GEMM output tensor
TensorShape shape_gemm(im2_col_info.tensor_shape());
@@ -481,7 +559,7 @@
TensorInfo gemm_output_info = input->clone()->set_tensor_shape(shape_gemm);
// Validate GEMM interleave and multiply
- if(!is_fully_connected_convolution)
+ if(is_interleaved_transposed)
{
TensorShape shape_interleaved = shape_im2col;
shape_interleaved.set(0, shape_interleaved.x() * 4);
@@ -523,13 +601,27 @@
}
else
{
- if(!_is_fully_connected_convolution)
+ if(_is_interleaved_transposed)
{
// Run interleave
NEScheduler::get().schedule(&_input_interleave_kernel, Window::DimY);
}
- NEScheduler::get().schedule(&_mm_kernel, Window::DimY);
+ // Runs matrix multiply on reshaped matrices
+ if(_is_quantized)
+ {
+ _mm_gemmlowp.run();
+ }
+ else
+ {
+ NEScheduler::get().schedule(&_mm_kernel, Window::DimY);
+ }
+ }
+
+ // Run output stage for quantized case
+ if(_is_quantized)
+ {
+ _gemmlowp_output_stage.run();
}
// Reshape output matrix