COMPMID-2048: Add support for dilation in NEDepthwiseConvolution.

Change-Id: If9941e770779fbf918ba5ff0573da9378078b969
Signed-off-by: Usama Arif <usama.arif@arm.com>
Reviewed-on: https://review.mlplatform.org/c/999
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Pablo Marquez <pablo.tello@arm.com>

2 files changed, 82 insertions, 51 deletions

diff --git a/src/core/NEON/kernels/NEDepthwiseConvolutionLayer3x3Kernel.cpp b/src/core/NEON/kernels/NEDepthwiseConvolutionLayer3x3Kernel.cpp
index ec672e0fc4..fdafc2da90 100644
--- a/src/core/NEON/kernels/NEDepthwiseConvolutionLayer3x3Kernel.cpp
+++ b/src/core/NEON/kernels/NEDepthwiseConvolutionLayer3x3Kernel.cpp
@@ -49,7 +49,7 @@ class convolver_3x3
 {
 public:
     static void convolve(const Window &window, unsigned int num_elems_written_per_iteration,
-                         const ITensor *input, const ITensor *weights, ITensor *output, const PadStrideInfo &conv_info, unsigned int depth_multiplier)
+                         const ITensor *input, const ITensor *weights, ITensor *output, const PadStrideInfo &conv_info, unsigned int depth_multiplier, const Size2D &dilation)
     {
         const int input_offset   = -input->info()->quantization_info().offset;
         const int weights_offset = -weights->info()->quantization_info().offset;
@@ -57,6 +57,7 @@ public:
         const int          input_stride_x  = input->info()->strides_in_bytes().x();
         const int          input_stride_y  = input->info()->strides_in_bytes().y();
         const int          input_stride_z  = input->info()->strides_in_bytes().z();
+        const int          input_stride_w  = input->info()->strides_in_bytes()[3];
         const int          output_stride_y = output->info()->strides_in_bytes().y();
         const int          kernel_stride_y = weights->info()->strides_in_bytes().y();
         const int          kernel_stride_z = weights->info()->strides_in_bytes().z();
@@ -74,9 +75,10 @@ public:
 
         // setup input window for the iterator
         Window window_in = window;
-        // we just want execute_window_loop to iterate over the dimensions > 2, so we set the first 2 dimensions to 0
+        // Iteration of input is taken care of in execute_window_loop
         window_in.set(Window::DimX, Window::Dimension(0, 0, 0));
         window_in.set(Window::DimY, Window::Dimension(0, 0, 0));
+        window_in.set(Window::DimZ, Window::Dimension(0, 0, 0));
 
         Window window_k = calculate_max_window(*weights->info(), Steps(1u));
 
@@ -91,7 +93,7 @@ public:
             int ih = 0;
             int oh = 0;
 
-            const uint8_t *input_ptr        = in.ptr() - conv_pad_x * input_stride_x - conv_pad_y * input_stride_y - (id.z() - id.z() / depth_multiplier) * input_stride_z;
+            const uint8_t *input_ptr        = in.ptr() - conv_pad_x * input_stride_x - conv_pad_y * input_stride_y + (id.z() / depth_multiplier) * input_stride_z + input_stride_w * id[3];
             const uint8_t *ptr_weights_base = weights_ptr + id.z() * kernel_stride_z;
 
             const auto ptr_weights_r0 = reinterpret_cast<const T1 *>(ptr_weights_base);
@@ -104,45 +106,54 @@ public:
             for(ih = 0, oh = 0; oh < output_h; ++oh, ih += conv_stride_y)
             {
                 auto in_top = reinterpret_cast<const T1 *>(input_ptr + (ih + 0) * input_stride_y);
-                auto in_mid = reinterpret_cast<const T1 *>(input_ptr + (ih + 1) * input_stride_y);
-                auto in_low = reinterpret_cast<const T1 *>(input_ptr + (ih + 2) * input_stride_y);
-                auto p_out  = reinterpret_cast<T2 *>(out.ptr() + oh * output_stride_y);
+                auto in_mid = reinterpret_cast<const T1 *>(input_ptr + (ih + dilation.y()) * input_stride_y);
+                auto in_low = reinterpret_cast<const T1 *>(input_ptr + (ih + 2 * dilation.y()) * input_stride_y); //uint8
+                auto p_out  = reinterpret_cast<T2 *>(out.ptr() + oh * output_stride_y);                           //int32
 
                 for(int ow = 0; ow < output_w; ow += num_elems_written_per_iteration,
                     in_top += delta_input, in_mid += delta_input, in_low += delta_input,
                     p_out += num_elems_written_per_iteration)
                 {
-                    auto vres = detail::convolve_3x3<stridex>(in_top, in_mid, in_low, vw_r0, vw_r1, vw_r2, input_offset);
-                    detail::store_results<stridex>(p_out, vres);
+                    if(dilation == Size2D(1U, 1U))
+                    {
+                        auto vres = detail::convolve_3x3<stridex>(in_top, in_mid, in_low, vw_r0, vw_r1, vw_r2, input_offset);
+                        detail::store_results<stridex>(p_out, vres);
+                    }
+                    else
+                    {
+                        auto vres = detail::convolve_3x3_dilation<stridex>(in_top, in_mid, in_low, vw_r0, vw_r1, vw_r2, dilation.x(), input_offset);
+                        detail::store_results<stridex>(p_out, vres);
+                    }
                 }
             }
         },
-        in, out);
+        out);
     }
 };
 
 template <typename T1, typename T2>
 inline void convolve_3x3(const Window &window, unsigned int num_elems_written_per_iteration,
-                         const ITensor *input, const ITensor *weights, ITensor *output, const PadStrideInfo &conv_info, unsigned int depth_multiplier)
+                         const ITensor *input, const ITensor *weights, ITensor *output,
+                         const PadStrideInfo &conv_info, unsigned int depth_multiplier, const Size2D &dilation)
 {
     const unsigned int conv_stride_x = std::get<0>(conv_info.stride());
     switch(conv_stride_x)
     {
         case 1:
-            convolver_3x3<T1, T2, 1>::convolve(window, num_elems_written_per_iteration, input, weights, output, conv_info, depth_multiplier);
+            convolver_3x3<T1, T2, 1>::convolve(window, num_elems_written_per_iteration, input, weights, output, conv_info, depth_multiplier, dilation);
             break;
         case 2:
-            convolver_3x3<T1, T2, 2>::convolve(window, num_elems_written_per_iteration, input, weights, output, conv_info, depth_multiplier);
+            convolver_3x3<T1, T2, 2>::convolve(window, num_elems_written_per_iteration, input, weights, output, conv_info, depth_multiplier, dilation);
             break;
         case 3:
-            convolver_3x3<T1, T2, 3>::convolve(window, num_elems_written_per_iteration, input, weights, output, conv_info, depth_multiplier);
+            convolver_3x3<T1, T2, 3>::convolve(window, num_elems_written_per_iteration, input, weights, output, conv_info, depth_multiplier, dilation);
             break;
         default:
             ARM_COMPUTE_ERROR("Not implemented");
     }
 }
 
-Status validate_arguments(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *output, const PadStrideInfo &conv_info, unsigned int depth_multiplier)
+Status validate_arguments(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *output, const PadStrideInfo &conv_info, unsigned int depth_multiplier, const Size2D &dilation)
 {
     ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(input);
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32);
@@ -157,7 +168,7 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *weights,
 
     if(output->total_size() != 0)
     {
-        const TensorShape output_shape = misc::shape_calculator::compute_depthwise_convolution_shape(*input, *weights, conv_info, depth_multiplier);
+        const TensorShape output_shape = misc::shape_calculator::compute_depthwise_convolution_shape(*input, *weights, conv_info, depth_multiplier, dilation);
         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), output_shape);
 
         if(is_data_type_quantized_asymmetric(input->data_type()))
@@ -173,13 +184,14 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *weights,
     return Status{};
 }
 
-std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *weights, ITensorInfo *output, const PadStrideInfo &conv_info, unsigned int depth_multiplier)
+std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *weights, ITensorInfo *output, const PadStrideInfo &conv_info, unsigned int depth_multiplier,
+                                                        const Size2D &dilation)
 {
     Window win;
     bool   window_changed = false;
 
     // Get convolved dimensions
-    const TensorShape output_shape = misc::shape_calculator::compute_depthwise_convolution_shape(*input, *weights, conv_info, depth_multiplier);
+    const TensorShape output_shape = misc::shape_calculator::compute_depthwise_convolution_shape(*input, *weights, conv_info, depth_multiplier, dilation);
     const DataType    output_dt    = (input->data_type() == DataType::QASYMM8) ? DataType::S32 : input->data_type();
 
     // Output auto inizialitation if not yet initialized
@@ -197,15 +209,16 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITen
     switch(input->data_type())
     {
         case DataType::QASYMM8:
-            num_elems_read_per_iteration = 16;
+            num_elems_read_per_iteration = 16 + 15 * (dilation.x() - 1);
             break;
 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
         case DataType::F16:
-            num_elems_read_per_iteration = 24;
+            num_elems_written_per_iteration = 32 >> conv_stride_x;
+            num_elems_read_per_iteration    = 24 + 23 * (dilation.x() - 1);
             break;
 #endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
         case DataType::F32:
-            num_elems_read_per_iteration = 12;
+            num_elems_read_per_iteration = 12 + 11 * (dilation.x() - 1);
             break;
         default:
             ARM_COMPUTE_ERROR("Data type not supported.");
@@ -214,7 +227,7 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITen
     // Configure kernel window
     win = calculate_max_window(*output, Steps(num_elems_written_per_iteration));
 
-    AccessWindowRectangle  input_access(input, -conv_pad_left, -conv_pad_top, num_elems_read_per_iteration, 3, conv_stride_x, conv_stride_y);
+    AccessWindowRectangle  input_access(input, -conv_pad_left, -conv_pad_top, num_elems_read_per_iteration, 3 + 2 * (dilation.y() - 1), conv_stride_x, conv_stride_y);
     AccessWindowStatic     weights_access(weights, 0, 0, 3, 3);
     AccessWindowHorizontal output_access(output, 0, num_elems_written_per_iteration);
 
@@ -227,7 +240,7 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITen
 } // namespace
 
 NEDepthwiseConvolutionLayer3x3Kernel::NEDepthwiseConvolutionLayer3x3Kernel()
-    : _border_size(0), _input(), _output(), _weights(), _conv_info(), _num_elems_written_per_iteration(0), _depth_multiplier(1)
+    : _border_size(0), _input(), _output(), _weights(), _conv_info(), _num_elems_written_per_iteration(0), _depth_multiplier(1), _dilation()
 {
 }
 
@@ -236,28 +249,41 @@ BorderSize NEDepthwiseConvolutionLayer3x3Kernel::border_size() const
     return _border_size;
 }
 
-void NEDepthwiseConvolutionLayer3x3Kernel::configure(const ITensor *input, const ITensor *weights, ITensor *output, const PadStrideInfo &conv_info, unsigned int depth_multiplier)
+void NEDepthwiseConvolutionLayer3x3Kernel::configure(const ITensor *input, const ITensor *weights, ITensor *output, const PadStrideInfo &conv_info, unsigned int depth_multiplier,
+                                                     const Size2D &dilation)
 {
     ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output);
-    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), weights->info(), output->info(), conv_info, depth_multiplier));
-
-    _input                           = input;
-    _output                          = output;
-    _weights                         = weights;
-    _conv_info                       = conv_info;
-    _depth_multiplier                = depth_multiplier;
-    _num_elems_written_per_iteration = 16 >> _conv_info.stride().first;
-    _border_size                     = BorderSize(_conv_info.pad_top(), _conv_info.pad_right(), _conv_info.pad_bottom(), _conv_info.pad_left());
-
-    auto win_config = validate_and_configure_window(_input->info(), _weights->info(), _output->info(), _conv_info, _depth_multiplier);
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), weights->info(), output->info(), conv_info, depth_multiplier, dilation));
+
+    _input            = input;
+    _output           = output;
+    _weights          = weights;
+    _conv_info        = conv_info;
+    _depth_multiplier = depth_multiplier;
+    switch(input->info()->data_type())
+    {
+        case DataType::QASYMM8:
+        case DataType::F32:
+            _num_elems_written_per_iteration = 16 >> _conv_info.stride().first;
+            break;
+        case DataType::F16:
+            _num_elems_written_per_iteration = 32 >> _conv_info.stride().first;
+            break;
+        default:
+            ARM_COMPUTE_ERROR("Data type not supported.");
+    }
+    _border_size    = BorderSize(_conv_info.pad_top(), _conv_info.pad_right(), _conv_info.pad_bottom(), _conv_info.pad_left());
+    _dilation       = dilation;
+    auto win_config = validate_and_configure_window(_input->info(), _weights->info(), _output->info(), _conv_info, _depth_multiplier, dilation);
     ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
     INEKernel::configure(win_config.second);
 }
 
-Status NEDepthwiseConvolutionLayer3x3Kernel::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *output, const PadStrideInfo &conv_info, unsigned int depth_multiplier)
+Status NEDepthwiseConvolutionLayer3x3Kernel::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *output, const PadStrideInfo &conv_info, unsigned int depth_multiplier,
+                                                      const Size2D &dilation)
 {
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, weights, output, conv_info, depth_multiplier));
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), weights->clone().get(), output->clone().get(), conv_info, depth_multiplier).first);
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, weights, output, conv_info, depth_multiplier, dilation));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), weights->clone().get(), output->clone().get(), conv_info, depth_multiplier, dilation).first);
     return Status{};
 }
 
@@ -272,14 +298,14 @@ void NEDepthwiseConvolutionLayer3x3Kernel::run(const Window &window, const Threa
     {
 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
         case DataType::F16:
-            convolve_3x3<float16_t, float16_t>(window, _num_elems_written_per_iteration, _input, _weights, _output, _conv_info, _depth_multiplier);
+            convolve_3x3<float16_t, float16_t>(window, _num_elems_written_per_iteration, _input, _weights, _output, _conv_info, _depth_multiplier, _dilation);
             break;
 #endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
         case DataType::F32:
-            convolve_3x3<float, float>(window, _num_elems_written_per_iteration, _input, _weights, _output, _conv_info, _depth_multiplier);
+            convolve_3x3<float, float>(window, _num_elems_written_per_iteration, _input, _weights, _output, _conv_info, _depth_multiplier, _dilation);
             break;
         case DataType::QASYMM8:
-            convolve_3x3<uint8_t, int32_t>(window, _num_elems_written_per_iteration, _input, _weights, _output, _conv_info, _depth_multiplier);
+            convolve_3x3<uint8_t, int32_t>(window, _num_elems_written_per_iteration, _input, _weights, _output, _conv_info, _depth_multiplier, _dilation);
             break;
         default:
             ARM_COMPUTE_ERROR("Not implemented");
diff --git a/src/core/NEON/kernels/NEDepthwiseIm2ColKernel.cpp b/src/core/NEON/kernels/NEDepthwiseIm2ColKernel.cpp
index 62373e348b..88f8b31a35 100644
--- a/src/core/NEON/kernels/NEDepthwiseIm2ColKernel.cpp
+++ b/src/core/NEON/kernels/NEDepthwiseIm2ColKernel.cpp
@@ -38,7 +38,8 @@ using namespace arm_compute;
 
 namespace
 {
-Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, unsigned int depth_multiplier)
+Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output,
+                          const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, unsigned int depth_multiplier, const Size2D &dilation)
 {
     ARM_COMPUTE_UNUSED(conv_info);
     //Note: ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(input) is not needed here as this kernel doesn't use NEON FP16 instructions.
@@ -48,6 +49,7 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, c
     ARM_COMPUTE_RETURN_ERROR_ON((input->dimension(2) * depth_multiplier) != output->dimension(2));
     ARM_COMPUTE_RETURN_ERROR_ON(output->dimension(0) != (kernel_dims.width * kernel_dims.height + ((has_bias) ? 1 : 0)));
     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(input, output);
+    ARM_COMPUTE_RETURN_ERROR_ON((dilation.x() < 1) || dilation.y() < 1);
 
     return Status{};
 }
@@ -84,7 +86,7 @@ void NEDepthwiseIm2ColKernel::run_generic(const Window &window)
     Iterator out(_output, window_out);
 
     const int full_length   = input_w + pad_left + pad_right;
-    const int max_initial_x = stride_x * (((full_length - _kernel_dims.width) / stride_x) + 1);
+    const int max_initial_x = stride_x * (((full_length - (_kernel_dims.width + (_kernel_dims.width - 1) * (_dilation.x() - 1))) / stride_x) + 1);
 
     // Define pad value
     auto zero = static_cast<T>(0);
@@ -103,12 +105,12 @@ void NEDepthwiseIm2ColKernel::run_generic(const Window &window)
         // Get pointers
         const uint8_t *const input_ptr  = in.ptr() + id.z() / _depth_multiplier * input_stride_z;
         auto                 output_ptr = reinterpret_cast<T *>(out.ptr());
-        const int            height     = src_y + _kernel_dims.height;
-        const int            width      = src_x + _kernel_dims.width;
+        const int            height     = src_y + (_kernel_dims.height + (_kernel_dims.height - 1) * (_dilation.y() - 1));
+        const int            width      = src_x + (_kernel_dims.width + (_kernel_dims.width - 1) * (_dilation.x() - 1));
 
-        for(int y = src_y; y < height; ++y)
+        for(int y = src_y; y < height; y += _dilation.y())
         {
-            for(int x = src_x; x < width; ++x, ++output_ptr)
+            for(int x = src_x; x < width; x += _dilation.x(), ++output_ptr)
             {
                 if(x < 0 || x >= input_w || y < 0 || y >= input_h)
                 {
@@ -130,15 +132,16 @@ void NEDepthwiseIm2ColKernel::run_generic(const Window &window)
 }
 
 NEDepthwiseIm2ColKernel::NEDepthwiseIm2ColKernel()
-    : _func(nullptr), _input(nullptr), _output(nullptr), _kernel_dims(), _conv_info(), _has_bias(), _depth_multiplier(1)
+    : _func(nullptr), _input(nullptr), _output(nullptr), _kernel_dims(), _conv_info(), _has_bias(), _depth_multiplier(1), _dilation()
 {
 }
 
-void NEDepthwiseIm2ColKernel::configure(const ITensor *input, ITensor *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, unsigned int depth_multiplier)
+void NEDepthwiseIm2ColKernel::configure(const ITensor *input, ITensor *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, unsigned int depth_multiplier,
+                                        const Size2D &dilation)
 {
     ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
 
-    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), kernel_dims, conv_info, has_bias, depth_multiplier));
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), kernel_dims, conv_info, has_bias, depth_multiplier, dilation));
 
     _input            = input;
     _output           = output;
@@ -146,6 +149,7 @@ void NEDepthwiseIm2ColKernel::configure(const ITensor *input, ITensor *output, c
     _conv_info        = conv_info;
     _has_bias         = has_bias;
     _depth_multiplier = depth_multiplier;
+    _dilation         = dilation;
 
     // Configure kernel window
     Window win = calculate_max_window(*input->info(), Steps());
@@ -172,10 +176,11 @@ void NEDepthwiseIm2ColKernel::configure(const ITensor *input, ITensor *output, c
     INEKernel::configure(win);
 }
 
-Status NEDepthwiseIm2ColKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, unsigned int depth_multiplier)
+Status NEDepthwiseIm2ColKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, unsigned int depth_multiplier,
+                                         const Size2D &dilation)
 {
     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, kernel_dims, conv_info, has_bias, depth_multiplier));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, kernel_dims, conv_info, has_bias, depth_multiplier, dilation));
     return Status{};
 }