COMPMID-2383: DepthwiseConv use output qinfo
There was a problem in depthwise_conv where when needing to run
permute it was calling auto_init to initialise the temporary tensor
with input's qinfo instead of output's qinfo. This patch fixes
this and adds some tests to exercise the cases where we have different
qinfos
Change-Id: I396fbb4d73f6aa5efe03dc2dd727b3e13154d6f5
Signed-off-by: Pablo Tello <pablo.tello@arm.com>
Reviewed-on: https://review.mlplatform.org/c/1270
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Giuseppe Rossini <giuseppe.rossini@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
diff --git a/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp b/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp
index e912740..1608277 100644
--- a/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp
+++ b/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp
@@ -90,6 +90,7 @@
// Configure the function to transform the weights tensor from HWI -> IHW
_permute_weights_to_nchw.configure(weights, &_permuted_weights, PermutationVector(1U, 2U, 0U));
_permuted_weights.info()->set_data_layout(DataLayout::NCHW);
+ _permuted_output.info()->set_quantization_info(output->info()->quantization_info());
input_to_use = &_permuted_input;
weights_to_use = &_permuted_weights;
diff --git a/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp b/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp
index 4bc8439..43288ec 100644
--- a/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp
+++ b/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp
@@ -87,6 +87,7 @@
// Configure the function to transform the weights tensor from HWI -> IHW
_permute_weights.configure(weights, &_permuted_weights, PermutationVector(1U, 2U, 0U));
_permuted_weights.info()->set_data_layout(DataLayout::NCHW);
+ _permuted_output.info()->set_quantization_info(output->info()->quantization_info());
// Configure depthwise
_dwc_kernel.configure(&_permuted_input, &_permuted_weights, (_is_quantized) ? &_accumulator : &_permuted_output, conv_info, depth_multiplier, dilation);
@@ -164,6 +165,9 @@
_permute_weights.configure(weights, &_permuted_weights, PermutationVector(2U, 0U, 1U));
_permuted_weights.info()->set_data_layout(DataLayout::NHWC);
+ _permuted_output.info()->set_data_layout(DataLayout::NHWC);
+ _permuted_output.info()->set_quantization_info(output->info()->quantization_info());
+
// Configure optimized depthwise
_dwc_optimized_func.configure(&_permuted_input, &_permuted_weights, biases, &_permuted_output, conv_info, depth_multiplier, act_info_to_use);
@@ -423,6 +427,7 @@
permute(output_shape, PermutationVector(1U, 2U, 0U));
_permuted_output.allocator()->init(output->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(output_shape));
_permuted_output.info()->set_data_layout(DataLayout::NCHW);
+ _permuted_output.info()->set_quantization_info(output->info()->quantization_info());
output_to_use = &_permuted_output;
}
diff --git a/src/runtime/NEON/functions/assembly/NEDepthwiseConvolutionAssemblyDispatch.cpp b/src/runtime/NEON/functions/assembly/NEDepthwiseConvolutionAssemblyDispatch.cpp
index 0499d99..5f57bbf 100644
--- a/src/runtime/NEON/functions/assembly/NEDepthwiseConvolutionAssemblyDispatch.cpp
+++ b/src/runtime/NEON/functions/assembly/NEDepthwiseConvolutionAssemblyDispatch.cpp
@@ -176,7 +176,7 @@
// Output auto inizialitation if not yet initialized
const TensorShape output_shape = misc::shape_calculator::compute_depthwise_convolution_shape(*input->info(), *weights->info(), conv_info, depth_multiplier);
- auto_init_if_empty(*output->info(), input->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(output_shape));
+ auto_init_if_empty(*output->info(), input->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(output_shape).set_quantization_info(output->info()->quantization_info()));
_input = input;
_weights = weights;