COMPMID-926 Add depth multiplier support to NEON/CL/GLES depthwise convolution
Change-Id: I03f32c62350e5ea43e77bb15fc5a832d83719e3b
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/126657
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele DiGiorgio <michele.digiorgio@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
diff --git a/src/core/GLES_COMPUTE/cs_shaders/depthwise_convolution3x3.cs b/src/core/GLES_COMPUTE/cs_shaders/depthwise_convolution3x3.cs
index adfc126..134cc10 100644
--- a/src/core/GLES_COMPUTE/cs_shaders/depthwise_convolution3x3.cs
+++ b/src/core/GLES_COMPUTE/cs_shaders/depthwise_convolution3x3.cs
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -108,6 +108,8 @@
uint z_index = gl_GlobalInvocationID.z;
TENSOR_ITERATOR_ADVANCE_IN_BYTES(weights_iter, z_index * weights_attrs.stride_z);
+ src_iter.current_offset_in_bytes -= int((z_index - z_index / uint(DEPTH_MULTIPLIER)) * src_attrs.step_z);
+
vec4 w[3];
w[0] = LOAD_UNPACK4_CURRENT_ITEM_HALF(weights_ptr, weights_iter);
w[1] = LOAD_UNPACK4_HALF(weights_ptr, TENSOR3D_OFFSET(weights_iter, 0, 1, 0));
@@ -263,6 +265,8 @@
uint z_index = gl_GlobalInvocationID.z;
TENSOR_ITERATOR_ADVANCE_IN_BYTES(weights_iter, z_index * weights_attrs.stride_z);
+ src_iter.current_offset_in_bytes -= int((z_index - z_index / uint(DEPTH_MULTIPLIER)) * src_attrs.step_z);
+
vec4 w[3];
w[0] = LOAD_UNPACK4_CURRENT_ITEM_HALF(weights_ptr, weights_iter);
w[1] = LOAD_UNPACK4_HALF(weights_ptr, TENSOR3D_OFFSET(weights_iter, 0, 1, 0));
diff --git a/src/core/GLES_COMPUTE/kernels/GCDepthwiseConvolutionLayer3x3Kernel.cpp b/src/core/GLES_COMPUTE/kernels/GCDepthwiseConvolutionLayer3x3Kernel.cpp
index 9343268..c237409 100644
--- a/src/core/GLES_COMPUTE/kernels/GCDepthwiseConvolutionLayer3x3Kernel.cpp
+++ b/src/core/GLES_COMPUTE/kernels/GCDepthwiseConvolutionLayer3x3Kernel.cpp
@@ -33,31 +33,10 @@
#include "arm_compute/core/TensorInfo.h"
#include "arm_compute/core/Types.h"
#include "arm_compute/core/Utils.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
using namespace arm_compute;
-
-namespace
-{
-/** Calculates expected output shape dimension
- *
- * @param[in] Input shape
- *
- * @return Expected output shape
- */
-TensorShape get_output_shape(TensorShape input_shape, TensorShape weights_shape, PadStrideInfo conv_info)
-{
- unsigned int output_width = 0;
- unsigned int output_height = 0;
-
- std::tie(output_width, output_height) = scaled_dimensions(input_shape.x(), input_shape.y(), weights_shape.x(), weights_shape.y(), conv_info);
-
- TensorShape output_shape = input_shape;
- output_shape.set(0, output_width);
- output_shape.set(1, output_height);
-
- return output_shape;
-}
-} // namespace
+using namespace arm_compute::misc::shape_calculator;
GCDepthwiseConvolutionLayer3x3Kernel::GCDepthwiseConvolutionLayer3x3Kernel()
: _border_size(0), _input(), _output(), _weights(), _biases(), _conv_stride_x(0), _conv_stride_y(0), _conv_pad_left(0), _conv_pad_top(0), _lws(gles::NDRange(1U, 1U, 1U))
@@ -69,7 +48,8 @@
return _border_size;
}
-void GCDepthwiseConvolutionLayer3x3Kernel::configure(const IGCTensor *input, const IGCTensor *weights, const IGCTensor *biases, IGCTensor *output, const PadStrideInfo &conv_info)
+void GCDepthwiseConvolutionLayer3x3Kernel::configure(const IGCTensor *input, const IGCTensor *weights, const IGCTensor *biases, IGCTensor *output, const PadStrideInfo &conv_info,
+ unsigned int depth_multiplier)
{
ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16);
ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights);
@@ -83,7 +63,7 @@
}
// Get convolved dimensions
- TensorShape output_shape = get_output_shape(input->info()->tensor_shape(), weights->info()->tensor_shape(), conv_info);
+ const TensorShape output_shape = compute_depthwise_convolution_shape(*input->info(), *weights->info(), conv_info, depth_multiplier);
// Output auto inizialitation if not yet initialized
auto_init_if_empty(*output->info(),
@@ -93,6 +73,7 @@
input->info()->fixed_point_position());
ARM_COMPUTE_ERROR_ON_MISMATCHING_DIMENSIONS(output->info()->tensor_shape(), output_shape);
+ ARM_COMPUTE_ERROR_ON(output->info()->dimension(2) != weights->info()->dimension(2));
_input = input;
_output = output;
@@ -108,6 +89,7 @@
ARM_COMPUTE_ERROR_ON(_conv_stride_x < 1 || _conv_stride_x > 3);
std::set<std::string> options;
+ options.emplace("#define DEPTH_MULTIPLIER " + support::cpp11::to_string(depth_multiplier));
options.emplace("#define LOCAL_SIZE_X " + support::cpp11::to_string(_lws[0]));
options.emplace("#define LOCAL_SIZE_Y " + support::cpp11::to_string(_lws[1]));
options.emplace("#define LOCAL_SIZE_Z " + support::cpp11::to_string(_lws[2]));