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>

3 files changed, 17 insertions, 10 deletions

diff --git a/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.cpp b/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.cpp
index de68ceda11..1997a901fe 100644
--- a/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.cpp
+++ b/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.cpp
@@ -50,6 +50,7 @@ BorderSize CLDepthwiseConvolutionLayer3x3NCHWKernel::border_size() const
 }
 
 void CLDepthwiseConvolutionLayer3x3NCHWKernel::configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info,
+                                                         unsigned int        depth_multiplier,
                                                          ActivationLayerInfo act_info)
 {
     ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32);
@@ -73,7 +74,7 @@ void CLDepthwiseConvolutionLayer3x3NCHWKernel::configure(const ICLTensor *input,
     }
 
     // Get convolved dimensions
-    const TensorShape output_shape = compute_depthwise_convolution_shape(*input->info(), *weights->info(), 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(),
@@ -84,6 +85,7 @@ void CLDepthwiseConvolutionLayer3x3NCHWKernel::configure(const ICLTensor *input,
                        input->info()->quantization_info());
 
     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;
@@ -98,6 +100,7 @@ void CLDepthwiseConvolutionLayer3x3NCHWKernel::configure(const ICLTensor *input,
     // Set build options
     ARM_COMPUTE_ERROR_ON(_conv_stride_x < 1 || _conv_stride_x > 3);
     CLBuildOptions build_opts;
+    build_opts.add_option("-DDEPTH_MULTIPLIER=" + support::cpp11::to_string(depth_multiplier));
     build_opts.add_option("-DCONV_STRIDE_X=" + support::cpp11::to_string(_conv_stride_x));
     build_opts.add_option_if(_biases != nullptr, "-DHAS_BIAS");
 
diff --git a/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.cpp b/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.cpp
index d783b9e159..a02b84fba1 100644
--- a/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.cpp
+++ b/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.cpp
@@ -41,7 +41,7 @@ using namespace arm_compute::misc::shape_calculator;
 
 namespace
 {
-Status validate_arguments(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info,
+Status validate_arguments(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info, unsigned int depth_multiplier,
                           const ActivationLayerInfo &act_info)
 {
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8);
@@ -50,6 +50,7 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *weights,
                                     && (act_info.activation() != ActivationLayerInfo::ActivationFunction::RELU),
                                     "For QASYMM8 only relu, lower bounded relu and lower-upper bounded relu are supported");
     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights);
+    ARM_COMPUTE_RETURN_ERROR_ON(depth_multiplier > 1); // COMPMID-1071 Add depth multiplier support for NHWC
     ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(1) != 3 || weights->dimension(2) != 3);
 
     if(biases != nullptr)
@@ -61,7 +62,7 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *weights,
 
     if(output->total_size() != 0)
     {
-        const TensorShape output_shape = compute_depthwise_convolution_shape(*input, *weights, conv_info);
+        const TensorShape output_shape = compute_depthwise_convolution_shape(*input, *weights, conv_info, depth_multiplier);
         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), output_shape);
     }
 
@@ -105,12 +106,13 @@ BorderSize CLDepthwiseConvolutionLayer3x3NHWCKernel::border_size() const
 }
 
 void CLDepthwiseConvolutionLayer3x3NHWCKernel::configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info,
+                                                         unsigned int        depth_multiplier,
                                                          ActivationLayerInfo act_info)
 {
     ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output);
 
     // Get convolved dimensions
-    const TensorShape output_shape = compute_depthwise_convolution_shape(*input->info(), *weights->info(), 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(),
@@ -120,7 +122,7 @@ void CLDepthwiseConvolutionLayer3x3NHWCKernel::configure(const ICLTensor *input,
                        input->info()->fixed_point_position(),
                        input->info()->quantization_info());
 
-    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), weights->info(), (biases != nullptr) ? biases->info() : nullptr, output->info(), conv_info, act_info));
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), weights->info(), (biases != nullptr) ? biases->info() : nullptr, output->info(), conv_info, depth_multiplier, act_info));
 
     const unsigned int conv_stride_x = conv_info.stride().first;
     ARM_COMPUTE_ERROR_ON(conv_stride_x < 1 || conv_stride_x > 2);
@@ -208,9 +210,10 @@ void CLDepthwiseConvolutionLayer3x3NHWCKernel::configure(const ICLTensor *input,
 }
 
 Status CLDepthwiseConvolutionLayer3x3NHWCKernel::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info,
+                                                          unsigned int        depth_multiplier,
                                                           ActivationLayerInfo act_info)
 {
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, weights, biases, output, conv_info, act_info));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, weights, biases, output, conv_info, depth_multiplier, act_info));
     ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), weights->clone().get(), output->clone().get(), conv_info).first);
 
     return Status{};
diff --git a/src/core/CL/kernels/CLDepthwiseIm2ColKernel.cpp b/src/core/CL/kernels/CLDepthwiseIm2ColKernel.cpp
index a0784dcad6..0aef52f791 100644
--- a/src/core/CL/kernels/CLDepthwiseIm2ColKernel.cpp
+++ b/src/core/CL/kernels/CLDepthwiseIm2ColKernel.cpp
@@ -42,13 +42,13 @@ CLDepthwiseIm2ColKernel::CLDepthwiseIm2ColKernel()
 {
 }
 
-void CLDepthwiseIm2ColKernel::configure(const ICLTensor *input, ICLTensor *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias)
+void CLDepthwiseIm2ColKernel::configure(const ICLTensor *input, ICLTensor *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, unsigned int depth_multiplier)
 {
     ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, 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(is_data_type_quantized_asymmetric(input->info()->data_type()) && has_bias);
-    ARM_COMPUTE_ERROR_ON(input->info()->dimension(2) != output->info()->dimension(2));
+    ARM_COMPUTE_ERROR_ON((input->info()->dimension(2) * depth_multiplier) != output->info()->dimension(2));
     ARM_COMPUTE_ERROR_ON(output->info()->dimension(0) != (kernel_dims.width * kernel_dims.height + ((has_bias) ? 1 : 0)));
 
     _input  = input;
@@ -68,6 +68,7 @@ void CLDepthwiseIm2ColKernel::configure(const ICLTensor *input, ICLTensor *outpu
     build_opts.add_option("-DSRC_HEIGHT=" + support::cpp11::to_string(input->info()->dimension(1)));
     build_opts.add_option("-DKERNEL_WIDTH=" + support::cpp11::to_string(kernel_dims.width));
     build_opts.add_option("-DKERNEL_HEIGHT=" + support::cpp11::to_string(kernel_dims.height));
+    build_opts.add_option("-DDEPTH_MULTIPLIER=" + support::cpp11::to_string(depth_multiplier));
     build_opts.add_option_if(has_bias, "-DHAS_BIAS");
     build_opts.add_option_if_else(is_data_type_quantized_asymmetric(input->info()->data_type()),
                                   "-DPAD_VALUE=" + support::cpp11::to_string(input->info()->quantization_info().offset),
@@ -85,8 +86,8 @@ void CLDepthwiseIm2ColKernel::configure(const ICLTensor *input, ICLTensor *outpu
     }
 
     // Configure  kernel window
-    Window win = calculate_max_window(*input->info(), Steps());
-    // The CLDepthwiseIm2ColKernel doesn't need padding so update_window_and_padding() can be skipped
+    Window win = calculate_max_window(*output->info(), Steps());
+    // CLDepthwiseIm2ColKernel doesn't need padding so update_window_and_padding() can be skipped
     output->info()->set_valid_region(ValidRegion(Coordinates(), output->info()->tensor_shape()));
 
     ICLKernel::configure(win);