COMPMID-856: CL Depthwise Convolution QASYMM8 support

Change-Id: Ic6097e7cf160e8b829fb521b7b99d9a57d9799d3
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/118774
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>

1 files changed, 43 insertions, 18 deletions

diff --git a/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp b/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp
index 02273fe08b..fcf3969515 100644
--- a/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp
+++ b/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -25,10 +25,12 @@
 
 #include "arm_compute/core/CL/ICLTensor.h"
 #include "arm_compute/core/PixelValue.h"
+#include "arm_compute/core/utils/quantization/AsymmHelpers.h"
 #include "arm_compute/runtime/CL/CLScheduler.h"
 #include "support/ToolchainSupport.h"
 
 using namespace arm_compute;
+using namespace arm_compute::misc;
 
 CLDepthwiseConvolutionLayer3x3::CLDepthwiseConvolutionLayer3x3()
     : _kernel(), _border_handler()
@@ -59,14 +61,14 @@ void CLDepthwiseConvolutionLayer3x3::run()
 }
 
 CLDepthwiseConvolutionLayer::CLDepthwiseConvolutionLayer()
-    : _im2col_kernel(), _weights_reshape_kernel(), _v2mm_kernel(), _vector_to_tensor_kernel(), _v2mm_input_fill_border(), _v2mm_weights_fill_border(), _input_reshaped(), _weights_reshaped(),
-      _v2mm_output()
+    : _im2col_kernel(), _weights_reshape_kernel(), _v2mm_kernel(), _vector_to_tensor_kernel(), _output_stage_kernel(), _v2mm_input_fill_border(), _v2mm_weights_fill_border(), _input_reshaped(),
+      _weights_reshaped(), _v2mm_output(), _output_reshaped(), _is_quantized(false)
 {
 }
 
 void CLDepthwiseConvolutionLayer::configure(ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info)
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F32);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights);
     ARM_COMPUTE_ERROR_ON(input->info()->dimension(2) != weights->info()->dimension(2));
 
@@ -74,15 +76,17 @@ void CLDepthwiseConvolutionLayer::configure(ICLTensor *input, const ICLTensor *w
     const size_t weights_h = weights->info()->dimension(1);
     const size_t weights_z = weights->info()->dimension(2);
 
-    const bool      has_bias   = (biases != nullptr);
-    const GPUTarget gpu_target = CLScheduler::get().target();
+    _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();
 
     unsigned int conv_w = 0;
     unsigned int conv_h = 0;
     std::tie(conv_w, conv_h) = scaled_dimensions(input->info()->dimension(0), input->info()->dimension(1), weights_w, weights_h, conv_info);
 
     // Set up intermediate tensors
-    const size_t patch_size = weights_w * weights_h + ((has_bias) ? 1 : 0);
+    const size_t patch_size = weights_w * weights_h + ((append_bias) ? 1 : 0);
     const size_t conv_size  = conv_w * conv_h;
 
     // Im2Col configuration
@@ -90,33 +94,49 @@ void CLDepthwiseConvolutionLayer::configure(ICLTensor *input, const ICLTensor *w
     shape_im2col.set(0, patch_size);
     shape_im2col.set(1, conv_size);
     shape_im2col.set(2, weights_z);
-    const TensorInfo info_im2col(shape_im2col, 1, input->info()->data_type(), input->info()->fixed_point_position());
-    _input_reshaped.allocator()->init(info_im2col);
+    _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, has_bias);
+    _im2col_kernel.configure(input, &_input_reshaped, Size2D(weights_w, weights_h), conv_info, append_bias);
 
     // Weights reshape configuration
     const TensorShape shape_weights_reshape(patch_size, weights_z);
-    const TensorInfo  info_weights_reshape(shape_weights_reshape, 1, weights->info()->data_type(), weights->info()->fixed_point_position());
-    _weights_reshaped.allocator()->init(info_weights_reshape);
-    _weights_reshape_kernel.configure(weights, &_weights_reshaped, biases);
+    _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);
-    const TensorInfo info_v2mm_out(shape_v2mm_out, 1, input->info()->data_type(), input->info()->fixed_point_position());
-    _v2mm_output.allocator()->init(info_v2mm_out);
+    _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);
-    _vector_to_tensor_kernel.configure(&_v2mm_output, output, conv_w, conv_h);
+    _vector_to_tensor_kernel.configure(&_v2mm_output, (_is_quantized) ? &_output_reshaped : output, conv_w, conv_h);
+
+    // Output staged configuration
+    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);
+        _output_stage_kernel.configure(&_output_reshaped, biases, output, output_multiplier, output_shift, output->info()->quantization_info().offset);
+        _output_reshaped.allocator()->allocate();
+    }
 
+    // Fill borders on inputs
+    PixelValue zero_in(static_cast<int32_t>(0));
+    PixelValue zero_w(static_cast<int32_t>(0));
+    if(_is_quantized)
+    {
+        zero_in = PixelValue(static_cast<int32_t>(input->info()->quantization_info().offset));
+        zero_w  = PixelValue(static_cast<int32_t>(weights->info()->quantization_info().offset));
+    }
     BorderSize border_size = _v2mm_kernel.border_size();
-    _v2mm_input_fill_border.configure(&_input_reshaped, border_size, BorderMode::CONSTANT, PixelValue(0));
+    _v2mm_input_fill_border.configure(&_input_reshaped, border_size, BorderMode::CONSTANT, zero_in);
 
     border_size.bottom = 0;
-    _v2mm_weights_fill_border.configure(&_weights_reshaped, border_size, BorderMode::CONSTANT, PixelValue(0));
+    _v2mm_weights_fill_border.configure(&_weights_reshaped, border_size, BorderMode::CONSTANT, zero_w);
 
     // Allocate intermediate tensors
     _input_reshaped.allocator()->allocate();
@@ -135,4 +155,9 @@ void CLDepthwiseConvolutionLayer::run()
     CLScheduler::get().enqueue(_v2mm_kernel);
 
     CLScheduler::get().enqueue(_vector_to_tensor_kernel);
+
+    if(_is_quantized)
+    {
+        CLScheduler::get().enqueue(_output_stage_kernel);
+    }
 }