COMPMID-1376: Add support for QASYMM8 in CLDeconvolutionLayer
Change-Id: I13ec79b6668e2b9559d3fa789ae0b51ab6975289
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/139126
Reviewed-by: Michalis Spyrou <michalis.spyrou@arm.com>
Tested-by: Jenkins <bsgcomp@arm.com>
diff --git a/src/core/CL/cl_kernels/deconvolution_layer.cl b/src/core/CL/cl_kernels/deconvolution_layer.cl
index e15482c..e5169f9 100644
--- a/src/core/CL/cl_kernels/deconvolution_layer.cl
+++ b/src/core/CL/cl_kernels/deconvolution_layer.cl
@@ -25,7 +25,7 @@
/** This function applies upsample on an input image.
*
- * @param[in] src_ptr Pointer to the source image. Supported data types: F16/F32
+ * @param[in] src_ptr Pointer to the source image. Supported data types: QASYMM8/F16/F32
* @param[in] src_stride_x Stride of the source image in X dimension (in bytes)
* @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] src_stride_y Stride of the source image in Y dimension (in bytes)
@@ -33,7 +33,7 @@
* @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes)
* @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] src_offset_first_element_in_bytes The offset of the first element in the source image
- * @param[out] dst_ptr Pointer to the destination image. Supported data types: F16/F32
+ * @param[out] dst_ptr Pointer to the destination image. Supported data types: same as @p src_ptr
* @param[in] dst_stride_x Stride of the destination image in X dimension (in bytes)
* @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] dst_stride_y Stride of the destination image in Y dimension (in bytes)
diff --git a/src/core/CL/kernels/CLDeconvolutionLayerUpsampleKernel.cpp b/src/core/CL/kernels/CLDeconvolutionLayerUpsampleKernel.cpp
index c6a0031..be3a926 100644
--- a/src/core/CL/kernels/CLDeconvolutionLayerUpsampleKernel.cpp
+++ b/src/core/CL/kernels/CLDeconvolutionLayerUpsampleKernel.cpp
@@ -43,7 +43,7 @@
{
ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
- ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32);
+ ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32);
ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
ARM_COMPUTE_RETURN_ERROR_ON(output->dimension(0) == 0);
ARM_COMPUTE_RETURN_ERROR_ON(output->dimension(1) == 0);
diff --git a/src/runtime/CL/functions/CLDeconvolutionLayer.cpp b/src/runtime/CL/functions/CLDeconvolutionLayer.cpp
index 40562b5..3f5b8c9 100644
--- a/src/runtime/CL/functions/CLDeconvolutionLayer.cpp
+++ b/src/runtime/CL/functions/CLDeconvolutionLayer.cpp
@@ -47,7 +47,7 @@
unsigned int inner_border_right, unsigned int inner_border_top, const WeightsInfo &weights_info)
{
ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights, output);
- ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32);
+ ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32);
ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(0) != weights->dimension(1));
ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(0) < 1);
ARM_COMPUTE_RETURN_ERROR_ON(!info.padding_is_symmetric());
@@ -67,7 +67,14 @@
if(bias != nullptr)
{
- ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, bias);
+ if(is_data_type_quantized_asymmetric(input->data_type()))
+ {
+ ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(bias, 1, DataType::S32);
+ }
+ else
+ {
+ ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, bias);
+ }
}
ARM_COMPUTE_RETURN_ERROR_ON_MSG(output->dimension(Window::DimX) != output_shape.x(), "Output's width is invalid.");
@@ -98,7 +105,7 @@
const TensorShape output_shape = deconvolution_output_shape(out_dims, input->info()->tensor_shape(), weights->info()->tensor_shape());
// Output auto initialization if not yet initialized
- auto_init_if_empty(*output->info(), output_shape, 1, input->info()->data_type());
+ auto_init_if_empty(*output->info(), output_shape, 1, input->info()->data_type(), input->info()->quantization_info());
// Perform validation step
ARM_COMPUTE_ERROR_THROW_ON(CLDeconvolutionLayer::validate(input->info(), weights->info(), bias == nullptr ? nullptr : bias->info(), output->info(), info, inner_border_right, inner_border_top));
@@ -108,13 +115,13 @@
_memory_group.manage(&_scaled_output);
// configure scale function
- // Init and allocate intermmidiate tensor for output, same size as input but the first two axis are the same as the output tensor
+ // Init and allocate intermediate tensor for output, same size as input but the first two axis are the same as the output tensor
TensorShape scale_out_shape(input->info()->tensor_shape());
const unsigned int out_x = input->info()->dimension(0) + (input->info()->dimension(0) - 1) * (stride_x - 1) + inner_border_right + 2 * info.pad().first;
const unsigned int out_y = input->info()->dimension(1) + (input->info()->dimension(1) - 1) * (stride_y - 1) + inner_border_top + 2 * info.pad().second;
scale_out_shape.set(0, out_x);
scale_out_shape.set(1, out_y);
- TensorInfo scale_out_info(scale_out_shape, 1, input->info()->data_type());
+ TensorInfo scale_out_info(scale_out_shape, 1, input->info()->data_type(), input->info()->quantization_info());
_scaled_output.allocator()->init(scale_out_info);
_scale_f.configure(input, &_scaled_output, BorderSize(inner_border_top, inner_border_right), info);
diff --git a/src/runtime/CL/functions/CLDeconvolutionLayerUpsample.cpp b/src/runtime/CL/functions/CLDeconvolutionLayerUpsample.cpp
index 13a24f8..ce8667d 100644
--- a/src/runtime/CL/functions/CLDeconvolutionLayerUpsample.cpp
+++ b/src/runtime/CL/functions/CLDeconvolutionLayerUpsample.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2017, 2018 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -57,7 +57,15 @@
void CLDeconvolutionLayerUpsample::run()
{
_output->map(CLScheduler::get().queue(), true);
- memset(_output->buffer(), 0, _output->info()->total_size());
+ if(is_data_type_quantized_asymmetric(_output->info()->data_type()))
+ {
+ const uint8_t quantized_zero = _output->info()->quantization_info().offset;
+ std::fill_n(_output->buffer(), _output->info()->total_size(), quantized_zero);
+ }
+ else
+ {
+ memset(_output->buffer(), 0, _output->info()->total_size());
+ }
_output->unmap(CLScheduler::get().queue());
CLScheduler::get().enqueue(_upsample, false);