COMPMID-643: Add bias to CLDepthwiseConvolution.
Change-Id: Ibfe7b8c1172d10cbcae7971fe86b82090519d31d
Reviewed-on: http://mpd-gerrit.cambridge.arm.com/92798
Tested-by: Kaizen <jeremy.johnson+kaizengerrit@arm.com>
Reviewed-by: Jaroslaw Rzepecki <jaroslaw.rzepecki@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
diff --git a/src/core/CL/kernels/CLDepthwiseConvolution3x3Kernel.cpp b/src/core/CL/kernels/CLDepthwiseConvolution3x3Kernel.cpp
index 6e56835..2d0c416 100644
--- a/src/core/CL/kernels/CLDepthwiseConvolution3x3Kernel.cpp
+++ b/src/core/CL/kernels/CLDepthwiseConvolution3x3Kernel.cpp
@@ -37,7 +37,7 @@
using namespace arm_compute;
CLDepthwiseConvolution3x3Kernel::CLDepthwiseConvolution3x3Kernel()
- : _border_size(0), _input(), _output(), _weights(), _conv_stride_x(0), _conv_stride_y(0), _conv_pad_x(0), _conv_pad_y(0)
+ : _border_size(0), _input(), _output(), _weights(), _biases(), _conv_stride_x(0), _conv_stride_y(0), _conv_pad_x(0), _conv_pad_y(0)
{
}
@@ -46,13 +46,20 @@
return _border_size;
}
-void CLDepthwiseConvolution3x3Kernel::configure(const ICLTensor *input, ICLTensor *output, const ICLTensor *weights, const PadStrideInfo &conv_info)
+void CLDepthwiseConvolution3x3Kernel::configure(const ICLTensor *input, ICLTensor *output, const ICLTensor *weights, const ICLTensor *biases, 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(output, 1, DataType::F32);
ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(weights, 1, DataType::F32);
ARM_COMPUTE_ERROR_ON(weights->info()->dimension(0) != 3 || weights->info()->dimension(1) != 3);
+ if(biases != nullptr)
+ {
+ ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(weights, biases);
+ ARM_COMPUTE_ERROR_ON(biases->info()->dimension(0) != weights->info()->dimension(2));
+ ARM_COMPUTE_ERROR_ON(biases->info()->num_dimensions() > 1);
+ }
+
std::pair<unsigned int, unsigned int> expected_output = scaled_dimensions(input->info()->tensor_shape().x(), input->info()->tensor_shape().y(),
weights->info()->tensor_shape().x(), weights->info()->tensor_shape().y(),
conv_info);
@@ -64,6 +71,7 @@
_input = input;
_output = output;
_weights = weights;
+ _biases = biases;
_conv_stride_x = conv_info.stride().first;
_conv_stride_y = conv_info.stride().second;
_conv_pad_x = conv_info.pad().first;
@@ -73,6 +81,10 @@
// Set build options
ARM_COMPUTE_ERROR_ON(_conv_stride_x < 1 || _conv_stride_x > 3);
std::set<std::string> options{ "-DCONV_STRIDE_X=" + support::cpp11::to_string(_conv_stride_x) };
+ if(_biases != nullptr)
+ {
+ options.emplace("-DHAS_BIAS");
+ }
_kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("depthwise_convolution_3x3", options));
@@ -111,6 +123,15 @@
slice_weights.set_dimension_step(Window::DimX, 0);
slice_weights.set_dimension_step(Window::DimY, 0);
+ // Set biases
+ if(_biases != nullptr)
+ {
+ unsigned int idx = 3 * num_arguments_per_3D_tensor();
+ Window slice_biases;
+ slice_biases.use_tensor_dimensions(_biases->info()->tensor_shape());
+ add_1D_tensor_argument(idx, _biases, slice_biases);
+ }
+
do
{
unsigned int idx = 0;
diff --git a/src/core/CL/kernels/CLDepthwiseIm2ColKernel.cpp b/src/core/CL/kernels/CLDepthwiseIm2ColKernel.cpp
index 5c7fe7e..743cd4a 100644
--- a/src/core/CL/kernels/CLDepthwiseIm2ColKernel.cpp
+++ b/src/core/CL/kernels/CLDepthwiseIm2ColKernel.cpp
@@ -41,13 +41,13 @@
{
}
-void CLDepthwiseIm2ColKernel::configure(const ICLTensor *input, ICLTensor *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info)
+void CLDepthwiseIm2ColKernel::configure(const ICLTensor *input, ICLTensor *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias)
{
ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32);
ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, output);
ARM_COMPUTE_ERROR_ON(input->info()->dimension(2) != output->info()->dimension(2));
- ARM_COMPUTE_ERROR_ON(output->info()->dimension(0) != (kernel_dims.width * kernel_dims.height));
+ ARM_COMPUTE_ERROR_ON(output->info()->dimension(0) != (kernel_dims.width * kernel_dims.height + ((has_bias) ? 1 : 0)));
_input = input;
_output = output;
@@ -66,7 +66,10 @@
build_opts.emplace("-DSRC_HEIGHT=" + support::cpp11::to_string(input->info()->dimension(1)));
build_opts.emplace("-DKERNEL_WIDTH=" + support::cpp11::to_string(kernel_dims.width));
build_opts.emplace("-DKERNEL_HEIGHT=" + support::cpp11::to_string(kernel_dims.height));
-
+ if(has_bias)
+ {
+ build_opts.emplace("-DHAS_BIAS");
+ }
_kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("depthwise_im2col", build_opts));
// Configure kernel window
diff --git a/src/core/CL/kernels/CLDepthwiseWeightsReshapeKernel.cpp b/src/core/CL/kernels/CLDepthwiseWeightsReshapeKernel.cpp
index 68de68b..81dd6b4 100644
--- a/src/core/CL/kernels/CLDepthwiseWeightsReshapeKernel.cpp
+++ b/src/core/CL/kernels/CLDepthwiseWeightsReshapeKernel.cpp
@@ -35,19 +35,28 @@
using namespace arm_compute;
CLDepthwiseWeightsReshapeKernel::CLDepthwiseWeightsReshapeKernel()
- : _input(nullptr), _output(nullptr)
+ : _input(nullptr), _biases(nullptr), _output(nullptr)
{
}
-void CLDepthwiseWeightsReshapeKernel::configure(const ICLTensor *input, ICLTensor *output)
+void CLDepthwiseWeightsReshapeKernel::configure(const ICLTensor *input, ICLTensor *output, const ICLTensor *biases)
{
ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32);
ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, output);
ARM_COMPUTE_ERROR_ON(input->info()->dimension(2) != output->info()->dimension(1));
- ARM_COMPUTE_ERROR_ON(output->info()->dimension(0) != input->info()->dimension(0) * input->info()->dimension(1));
+ ARM_COMPUTE_ERROR_ON(output->info()->dimension(0) != (input->info()->dimension(0) * input->info()->dimension(1) + ((biases != nullptr) ? 1 : 0)));
+
+ if(biases != nullptr)
+ {
+ ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases);
+ ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, biases);
+ ARM_COMPUTE_ERROR_ON(biases->info()->dimension(0) != input->info()->dimension(2));
+ ARM_COMPUTE_ERROR_ON(biases->info()->num_dimensions() > 1);
+ }
_input = input;
+ _biases = biases;
_output = output;
// Create kernel
@@ -55,6 +64,10 @@
build_opts.emplace("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()));
build_opts.emplace("-DSRC_WIDTH=" + support::cpp11::to_string(input->info()->dimension(0)));
+ if(_biases != nullptr)
+ {
+ build_opts.emplace("-DHAS_BIAS");
+ }
_kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("depthwise_weights_reshape", build_opts));
@@ -84,6 +97,15 @@
slice_out.set(Window::DimX, Window::Dimension(0, 0, 0));
slice_out.set(Window::DimY, Window::Dimension(0, 0, 0));
+ // Set biases
+ if(_biases != nullptr)
+ {
+ unsigned int idx = num_arguments_per_3D_tensor() + num_arguments_per_2D_tensor();
+ Window slice_biases;
+ slice_biases.use_tensor_dimensions(_biases->info()->tensor_shape());
+ add_1D_tensor_argument(idx, _biases, slice_biases);
+ }
+
do
{
unsigned int idx = 0;