Port DepthwiseConvolution to new API

Resolves: COMPMID-4185

Change-Id: Ib5f22356356a022d567bb18d44ea272b62d10ebf
Signed-off-by: Michalis Spyrou <michalis.spyrou@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5424
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

4 files changed, 29 insertions, 186 deletions

diff --git a/arm_compute/core/Types.h b/arm_compute/core/Types.h
index 53333ff608..b1f340d18e 100644
--- a/arm_compute/core/Types.h
+++ b/arm_compute/core/Types.h
@@ -1861,6 +1861,19 @@ private:
     bool _broadcast_bias;
 };
 
+struct ConvolutionInfo
+{
+    ConvolutionInfo() = default;
+    ConvolutionInfo(const PadStrideInfo &pad_stride_info, unsigned int depth_multiplier, const ActivationLayerInfo &act_info, const Size2D &dilation)
+        : pad_stride_info(pad_stride_info), depth_multiplier(depth_multiplier), act_info(act_info), dilation(dilation)
+    {
+    }
+    PadStrideInfo       pad_stride_info{};        /**< Convolution info (Pads, strides,...) */
+    unsigned int        depth_multiplier{ 1 };    /**< Multiplier to apply to input's depth to retrieve the output depth. Defaults to 1 */
+    ActivationLayerInfo act_info{};               /**< Fused activation to apply after convolution. */
+    Size2D              dilation{ Size2D(1, 1) }; /**< Dilation, in elements, across x and y. Defaults to (1, 1). */
+};
+
 struct DepthwiseConvolutionReshapeInfo
 {
     unsigned int c0{ 1 };            /**< Number of channels processed by the depth-wise convolution */
diff --git a/arm_compute/core/utils/misc/ShapeCalculator.h b/arm_compute/core/utils/misc/ShapeCalculator.h
index 56038dd853..ba37f9a61e 100644
--- a/arm_compute/core/utils/misc/ShapeCalculator.h
+++ b/arm_compute/core/utils/misc/ShapeCalculator.h
@@ -435,16 +435,13 @@ inline TensorShape compute_transposed_shape(const ITensorInfo &input)
 
 /** Calculate the depthwise convolution output shape of a tensor
  *
- * @param[in] input            Input tensor info
- * @param[in] weights          Weights tensor info
- * @param[in] conv_info        Padding and stride information to use for the convolution.
- * @param[in] depth_multiplier Multiplier to apply to the input's depth in order to retrieve the output's depth.
- * @param[in] dilation         Dilation, in elements, across x and y. Defaults to (1, 1).
+ * @param[in] input   Input tensor info
+ * @param[in] weights Weights tensor info
+ * @param[in] info    Convolution info
  *
  * @return the calculated shape
  */
-inline TensorShape compute_depthwise_convolution_shape(const ITensorInfo &input, const ITensorInfo &weights, PadStrideInfo conv_info, unsigned int depth_multiplier, const Size2D &dilation = Size2D(1U,
-                                                       1U))
+inline TensorShape compute_depthwise_convolution_shape(const ITensorInfo &input, const ITensorInfo &weights, const ConvolutionInfo &info)
 {
     const TensorShape input_shape{ input.tensor_shape() };
     const TensorShape weights_shape{ weights.tensor_shape() };
@@ -462,12 +459,12 @@ inline TensorShape compute_depthwise_convolution_shape(const ITensorInfo &input,
     unsigned int output_height = 0;
     std::tie(output_width, output_height) = scaled_dimensions(input_shape[width_idx], input_shape[height_idx],
                                                               weights_shape[weights_width_idx], weights_shape[weights_height_idx],
-                                                              conv_info, dilation);
+                                                              info.pad_stride_info, info.dilation);
 
     TensorShape output_shape{ input_shape };
     output_shape.set(width_idx, output_width);
     output_shape.set(height_idx, output_height);
-    output_shape.set(channel_idx, input_shape[channel_idx] * depth_multiplier);
+    output_shape.set(channel_idx, input_shape[channel_idx] * info.depth_multiplier);
 
     return output_shape;
 }
diff --git a/arm_compute/runtime/NEON/functions/NEDepthwiseConvolutionLayer.h b/arm_compute/runtime/NEON/functions/NEDepthwiseConvolutionLayer.h
index 9aa8f04eb8..c74b2a93ee 100644
--- a/arm_compute/runtime/NEON/functions/NEDepthwiseConvolutionLayer.h
+++ b/arm_compute/runtime/NEON/functions/NEDepthwiseConvolutionLayer.h
@@ -24,9 +24,10 @@
 #ifndef ARM_COMPUTE_NEDEPTHWISECONVOLUTION_H
 #define ARM_COMPUTE_NEDEPTHWISECONVOLUTION_H
 
+#include "arm_compute/runtime/IMemoryManager.h"
+#include "arm_compute/runtime/MemoryGroup.h"
 #include "arm_compute/runtime/NEON/functions/NEActivationLayer.h"
 #include "arm_compute/runtime/NEON/functions/NEPermute.h"
-#include "arm_compute/runtime/NEON/functions/assembly/NEDepthwiseConvolutionAssemblyDispatch.h"
 #include <memory>
 
 namespace arm_compute
@@ -91,25 +92,6 @@ public:
     void prepare() override;
 
 private:
-    /** Static function to choose the best depthwise convolution function for @ref NEDepthwiseConvolutionLayer
-     *
-     * @param[in] input            Source tensor info. Data type supported: QASYMM8/QASYMM8_SIGNED/F16/F32
-     * @param[in] weights          Weights tensor info. These are 3D tensors with shape [kernel_x, kernel_y, IFM].
-     *                             Data type supported: Same as @p input or QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL when @p input is QASYMM8/QASYMM8_SIGNED.
-     * @param[in] biases           Biases tensor info. A 1D tensor with shape [IFM]. Must be nullptr if not needed.
-     *                             Data type supported: Same as @p input, S32 when input is QASYMM8/QASYMM8_SIGNED.
-     * @param[in] output           Destination tensor. Data type supported: same as @p input.
-     * @param[in] conv_info        Padding and stride information to use for the convolution.
-     * @param[in] depth_multiplier (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1.
-     * @param[in] act_info         (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU for 3x3 quantized are supported.
-     * @param[in] dilation         (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
-     *
-     * @return a Depthwise Convolution Function
-     */
-    static DepthwiseConvolutionFunction get_depthwiseconvolution_function(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output,
-                                                                          const PadStrideInfo &conv_info, unsigned int depth_multiplier = 1,
-                                                                          ActivationLayerInfo act_info = ActivationLayerInfo(), const Size2D &dilation = Size2D(1U, 1U));
-
     /** Basic function to execute optimized depthwise convolution routines. This function calls the following kernels:
     *
     * @note At the moment 3x3 and 5x5 convolution of stride 1, 2 are supported
@@ -173,23 +155,9 @@ private:
         void prepare() override;
 
     private:
-        MemoryGroup                            _memory_group;
-        NEDepthwiseConvolutionAssemblyDispatch _dwc_optimized_func;
-        NEPermute                              _permute_input;
-        NEPermute                              _permute_weights;
-        NEPermute                              _permute_output;
-        NEActivationLayer                      _activationlayer_function;
-        Tensor                                 _accumulator;
-        Tensor                                 _permuted_input;
-        Tensor                                 _permuted_weights;
-        Tensor                                 _permuted_output;
-        const ITensor                         *_original_weights;
-        bool                                   _has_bias;
-        bool                                   _is_quantized;
-        bool                                   _is_nchw;
-        bool                                   _permute;
-        bool                                   _is_activationlayer_enabled;
-        bool                                   _is_prepared;
+        MemoryGroup _memory_group;
+        struct Impl;
+        std::unique_ptr<Impl> _impl;
     };
 
     /** Basic function to execute a generic depthwise convolution. This function calls the following kernel:
@@ -248,26 +216,14 @@ private:
 
         // Inherited methods overriden:
         void run() override;
-        void prepare() override;
 
     private:
-        std::unique_ptr<NEDepthwiseConvolutionLayerNativeKernel> _depthwise_conv_kernel;
-        NEPermute                                                _permute_input;
-        NEPermute                                                _permute_weights;
-        NEPermute                                                _permute_output;
-        NEActivationLayer                                        _activationlayer_function;
-        Tensor                                                   _permuted_input;
-        Tensor                                                   _permuted_weights;
-        Tensor                                                   _permuted_output;
-        bool                                                     _is_prepared;
-        bool                                                     _is_nchw;
-        bool                                                     _is_activationlayer_enabled;
-        const ITensor                                           *_original_weights;
+        struct Impl;
+        std::unique_ptr<Impl> _impl;
     };
-
-    DepthwiseConvolutionFunction                 _depth_conv_func;
-    NEDepthwiseConvolutionLayerOptimizedInternal _func_optimized;
-    NEDepthwiseConvolutionLayerGeneric           _func_generic;
+    MemoryGroup _memory_group;
+    struct Impl;
+    std::unique_ptr<Impl> _impl;
 };
 } // namespace arm_compute
 #endif /* ARM_COMPUTE_NEDEPTHWISECONVOLUTION_H */
diff --git a/arm_compute/runtime/NEON/functions/assembly/NEDepthwiseConvolutionAssemblyDispatch.h b/arm_compute/runtime/NEON/functions/assembly/NEDepthwiseConvolutionAssemblyDispatch.h
deleted file mode 100644
index 7f63717b02..0000000000
--- a/arm_compute/runtime/NEON/functions/assembly/NEDepthwiseConvolutionAssemblyDispatch.h
+++ /dev/null
@@ -1,123 +0,0 @@
-/*
- * Copyright (c) 2019 Arm Limited.
- *
- * SPDX-License-Identifier: MIT
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to
- * deal in the Software without restriction, including without limitation the
- * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
- * sell copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-#ifndef ARM_COMPUTE_NEDEPTHWISECONVOLUTIONASSEMBLYDISPATCH_H
-#define ARM_COMPUTE_NEDEPTHWISECONVOLUTIONASSEMBLYDISPATCH_H
-
-#include "arm_compute/runtime/IFunction.h"
-
-#include "arm_compute/runtime/IMemoryManager.h"
-#include "arm_compute/runtime/MemoryGroup.h"
-#include "arm_compute/runtime/Tensor.h"
-
-namespace arm_compute
-{
-/** Depthwise convolution assembly kernel glue */
-class NEDepthwiseConvolutionAssemblyDispatch : public IFunction
-{
-public:
-    /** Default constructor
-     *
-     * @param[in,out] memory_manager Memory manager to use
-     */
-    NEDepthwiseConvolutionAssemblyDispatch(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
-    /** Prevent instances of this class from being copied (As this class contains pointers) */
-    NEDepthwiseConvolutionAssemblyDispatch(const NEDepthwiseConvolutionAssemblyDispatch &) = delete;
-    /** Default move constructor */
-    NEDepthwiseConvolutionAssemblyDispatch(NEDepthwiseConvolutionAssemblyDispatch &&) = default;
-    /** Prevent instances of this class from being copied (As this class contains pointers) */
-    NEDepthwiseConvolutionAssemblyDispatch &operator=(const NEDepthwiseConvolutionAssemblyDispatch &) = delete;
-    /** Default move assignment operator */
-    NEDepthwiseConvolutionAssemblyDispatch &operator=(NEDepthwiseConvolutionAssemblyDispatch &&) = default;
-    /** Default destructor */
-    ~NEDepthwiseConvolutionAssemblyDispatch();
-    /** Initialize the function's source, destination, kernels and border_size.
-     *
-     * @note Supports only NHWC format
-     *
-     * @param[in]  input            Source tensor. Data type supported: QASYMM8/F16/F32. (Written to only for border filling).
-     * @param[in]  weights          Weights tensor. These are 3D tensors with shape [W, H, IFM]. Data type supported: Same as @p input.
-     * @param[in]  bias             (Optional) Biases tensor. A 1D tensor with shape [IFM]. Must be nullptr if not needed.
-     *                              Data type supported: Same as @p input.
-     * @param[out] output           Destination tensor. Data type supported: same as @p input.
-     * @param[in]  conv_info        Padding and stride information to use for the convolution.
-     * @param[in]  depth_multiplier (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1.
-     * @param[in]  act_info         (Optional) Activation layer information in case of a fused activation.
-     * @param[in]  dilation         (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
-     */
-    void configure(const ITensor *input, const ITensor *weights, const ITensor *bias, ITensor *output,
-                   const PadStrideInfo &conv_info, unsigned int depth_multiplier = 1, const ActivationLayerInfo &act_info = ActivationLayerInfo(),
-                   const Size2D &dilation = Size2D(1, 1));
-    /** Static function to check if given info will lead to a valid configuration of @ref NEDepthwiseConvolutionAssemblyDispatch
-     *
-     * @note Supports only NHWC format
-     *
-     * @param[in]  input            Source tensor. Data type supported: QASYMM8/F16/F32. (Written to only for border filling).
-     * @param[in]  weights          Weights tensor. These are 3D tensors with shape [W, H, IFM]. Data type supported: Same as @p input.
-     * @param[in]  bias             (Optional) Biases tensor. A 1D tensor with shape [IFM]. Must be nullptr if not needed.
-     *                              Data type supported: Same as @p input.
-     * @param[out] output           Destination tensor. Data type supported: same as @p input.
-     * @param[in]  conv_info        Padding and stride information to use for the convolution.
-     * @param[in]  depth_multiplier (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1.
-     * @param[in]  act_info         (Optional) Activation layer information in case of a fused activation.
-     * @param[in]  dilation         (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
-     *
-     * @return An error status
-     */
-    static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *bias, const ITensorInfo *output,
-                           const PadStrideInfo &conv_info, unsigned int depth_multiplier = 1, const ActivationLayerInfo &act_info = ActivationLayerInfo(),
-                           const Size2D &dilation = Size2D(1, 1));
-    /** Check if the optimized kernel can be used for the given kernel sizes and strides
-     *
-     * @warning Even if this return true the inputs and outputs might need to get permuted as the only layout supported is NHWC
-     *
-     * @param[in] input            Input tensor info.
-     * @param[in] weights          Weights tensor info.
-     * @param[in] conv_info        Convolution layer metadata.
-     * @param[in] depth_multiplier (Optional) Depth multiplier to be used.
-     * @param[in] dilation         (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
-     *
-     * @return True if the assembly kernel could be used else false. Note that transformations of input/output could be needed.
-     */
-    static bool is_optimized_supported(const ITensorInfo *input, const ITensorInfo *weights, PadStrideInfo conv_info, unsigned int depth_multiplier = 1, const Size2D &dilation = Size2D(1, 1));
-
-    // Inherited methods overridden:
-    void run() override;
-    void prepare() override;
-
-private:
-    struct LocalImpl;
-
-private:
-    MemoryGroup                _memory_group;
-    const ITensor             *_input;
-    const ITensor             *_weights;
-    const ITensor             *_bias;
-    ITensor                   *_output;
-    Tensor                     _packed_weights;
-    Tensor                     _workspace;
-    bool                       _is_prepared;
-    std::unique_ptr<LocalImpl> _pImpl;
-};
-} // namespace arm_compute
-#endif /* ARM_COMPUTE_NEDEPTHWISECONVOLUTIONASSEMBLYDISPATCH_H */