commit 4c5469b192665c94118a8a558787cb9cec2d0765
author Georgios Pinitas <georgios.pinitas@arm.com> Tue May 21 13:32:43 2019 +0100
committer Georgios Pinitas <georgios.pinitas@arm.com> Mon Jun 03 14:51:29 2019 +0000
tree 168aa969de8243bdbb1f25247dd9f54d037ae32c
parent 43a129e94df41f9ac8bc78b702da5a387ada0494

COMPMID-2225: Add interface support for new quantized data types.

Add support for:
-QSYMM8, 8-bit quantized symmetric
-QSYMM8_PER_CHANNEL, 8-bit quantized symmetric with per channel quantization

Change-Id: I00c4ff98e44af37419470af61419ee95d0de2463
Signed-off-by: Georgios Pinitas <georgios.pinitas@arm.com>
Reviewed-on: https://review.mlplatform.org/c/1236
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp

diff --git a/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp b/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp
index 3bb69b1..4bc8439 100644
--- a/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp
+++ b/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp
@@ -72,7 +72,7 @@
         _memory_group.manage(&_accumulator);
         _accumulator.allocator()->init(TensorInfo(accum_shape, 1, DataType::S32, output->info()->quantization_info()));
         _accumulator.info()->set_data_layout(accum_layout);
-        zero_value = PixelValue(static_cast<uint32_t>(input->info()->quantization_info().offset));
+        zero_value = PixelValue(static_cast<uint32_t>(input->info()->quantization_info().uniform().offset));
     }
 
     if(!_is_nchw)
@@ -109,13 +109,15 @@
     // Configure biases accumulation
     if(_is_quantized)
     {
-        const QuantizationInfo output_quant_info = (output->info()->total_size() == 0) ? input->info()->quantization_info() : output->info()->quantization_info();
+        const UniformQuantizationInfo iq_info = input->info()->quantization_info().uniform();
+        const UniformQuantizationInfo wq_info = weights->info()->quantization_info().uniform();
+        const UniformQuantizationInfo oq_info = (output->info()->total_size() == 0) ? iq_info : output->info()->quantization_info().uniform();
 
-        float multiplier = input->info()->quantization_info().scale * weights->info()->quantization_info().scale / output_quant_info.scale;
+        float multiplier = (iq_info.scale * wq_info.scale) / oq_info.scale;
         int   output_multiplier;
         int   output_shift;
         quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift);
-        _output_stage_kernel.configure(&_accumulator, biases, _is_nchw ? output : &_permuted_output, output_multiplier, output_shift, output_quant_info.offset);
+        _output_stage_kernel.configure(&_accumulator, biases, _is_nchw ? output : &_permuted_output, output_multiplier, output_shift, oq_info.offset);
         _accumulator.allocator()->allocate();
     }
     else if(_has_bias)
@@ -459,13 +461,15 @@
     // Output staged configuration
     if(_is_quantized)
     {
-        const QuantizationInfo output_quant_info = output->info()->quantization_info();
+        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();
 
-        float multiplier = input->info()->quantization_info().scale * weights->info()->quantization_info().scale / output_quant_info.scale;
+        float multiplier = (iq_info.scale * wq_info.scale) / oq_info.scale;
         int   output_multiplier;
         int   output_shift;
         quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift);
-        _output_stage_kernel.configure(&_output_reshaped, biases, output_to_use, output_multiplier, output_shift, output_quant_info.offset);
+        _output_stage_kernel.configure(&_output_reshaped, biases, output_to_use, output_multiplier, output_shift, oq_info.offset);
         _output_reshaped.allocator()->allocate();
     }
 
@@ -483,8 +487,8 @@
     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));
+        zero_in = PixelValue(static_cast<int32_t>(input->info()->quantization_info().uniform().offset));
+        zero_w  = PixelValue(static_cast<int32_t>(weights->info()->quantization_info().uniform().offset));
     }
     BorderSize border_size = _v2mm_kernel.border_size();
     _v2mm_input_fill_border.configure(&_input_reshaped, border_size, BorderMode::CONSTANT, zero_in);