Rework OpenCL Depthwise Convolution

- Remove dedicated kernels for NCHW. Now we only use NHWC with permute
- Remove specialized kernels for 3x3 NHWC
- Simplify CLDepthwiseConvolutionLayer.cpp to call just the native
  implementation for both floating-point and quantized data types
- Develop two parametric opencl kernels for depthwise convolution layer NHWC
  (floating-point and quantized)
- Add support to export the weights to cl_image
- Extend test for depthwise convolution on opencl

Resolves COMPMID-4417

Change-Id: Ibe533f79c2860f9cac8e921895d5a8f947753a5c
Signed-off-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5893
Reviewed-by: Giorgio Arena <giorgio.arena@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

1 files changed, 28 insertions, 228 deletions

diff --git a/arm_compute/runtime/CL/functions/CLDepthwiseConvolutionLayer.h b/arm_compute/runtime/CL/functions/CLDepthwiseConvolutionLayer.h
index bbb00a1ebc..01ddae12bb 100644
--- a/arm_compute/runtime/CL/functions/CLDepthwiseConvolutionLayer.h
+++ b/arm_compute/runtime/CL/functions/CLDepthwiseConvolutionLayer.h
@@ -33,13 +33,14 @@
 namespace arm_compute
 {
 class CLCompileContext;
-class CLFillBorderKernel;
 class CLDepthwiseConvolutionLayerNativeKernel;
-class CLDepthwiseConvolutionLayer3x3NCHWKernel;
-class CLDepthwiseConvolutionLayer3x3NHWCKernel;
 class ICLTensor;
 
 /** Function to execute a depthwise convolution
+ *
+ * -# @ref CLDepthwiseConvolutionLayerNativeKernel
+ * -# @ref CLPermute (if the data layout is NCHW)
+ *
  */
 class CLDepthwiseConvolutionLayer : public IFunction
 {
@@ -102,239 +103,38 @@ public:
      *
      * @return a status
      */
-    static Status validate(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));
+    static Status validate(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));
 
     // Inherited methods overriden:
     void run() override;
     void prepare() override;
 
-private:
-    /** Static function to choose the best depthwise convolution function for @ref CLDepthwiseConvolutionLayer
-     *
-     * @param[in] input            Source tensor info. Data type supported: QASYMM8/FP16/FP32. Data layout supported: NHWC, NCHW
-     * @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/QSYMM8_PER_CHANNEL when @p input is QASYMM8.
-     * @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.
-     * @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 QASYMM8 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 a depthwise convolution for kernel size 3x3xC (when data layout NCHW) or Cx3x3 (when data layout NHWC). This function calls the following OpenCL kernels:
-    *
-    * -# @ref CLDepthwiseConvolutionLayer3x3NCHWKernel (if data_layout == NCHW)
-    * -# @ref CLDepthwiseConvolutionLayer3x3NHWCKernel (if data_layout == NHWC)
-    * -# @ref CLFillBorderKernel (if pad_x or pad_y > 0)
-    *
-    */
-    class CLDepthwiseConvolutionLayerInternal3x3 : public IFunction
-    {
-    public:
-        /** Default constructor */
-        CLDepthwiseConvolutionLayerInternal3x3(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
-        /** Prevent instances of this class from being copied (As this class contains pointers) */
-        CLDepthwiseConvolutionLayerInternal3x3(const CLDepthwiseConvolutionLayerInternal3x3 &) = delete;
-        /** Default move constructor */
-        CLDepthwiseConvolutionLayerInternal3x3(CLDepthwiseConvolutionLayerInternal3x3 &&) = default;
-        /** Prevent instances of this class from being copied (As this class contains pointers) */
-        CLDepthwiseConvolutionLayerInternal3x3 &operator=(const CLDepthwiseConvolutionLayerInternal3x3 &) = delete;
-        /** Default move assignment operator */
-        CLDepthwiseConvolutionLayerInternal3x3 &operator=(CLDepthwiseConvolutionLayerInternal3x3 &&) = default;
-        /** Initialize the function's source, destination, conv and border_size.
-         *
-         * @param[in, out] input            Source tensor. Data type supported: QASYMM8/F16/F32. (Written to only for border filling).
-         * @param[in]      weights          Weights tensor. A 3D tensor with shape [3, 3, IFM].
-         *                                  Data type supported: Same as @p input or QASYMM8/QSYMM8_PER_CHANNEL when @p input is QASYMM8.
-         * @param[in]      biases           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. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU for 3x3 QASYMM8 supported.
-         * @param[in]      dilation         (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
-         */
-        void configure(ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info, unsigned int depth_multiplier = 1,
-                       ActivationLayerInfo act_info = ActivationLayerInfo(), const Size2D &dilation = Size2D(1U, 1U));
-        /** Initialize the function's source, destination, conv and border_size.
-         *
-         * @param[in]      compile_context  The compile context to be used.
-         * @param[in, out] input            Source tensor. Data type supported: QASYMM8/F16/F32. (Written to only for border filling).
-         * @param[in]      weights          Weights tensor. A 3D tensor with shape [3, 3, IFM].
-         *                                  Data type supported: Same as @p input or QASYMM8/QSYMM8_PER_CHANNEL when @p input is QASYMM8.
-         * @param[in]      biases           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. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU for 3x3 QASYMM8 supported.
-         * @param[in]      dilation         (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
-         */
-        void configure(const CLCompileContext &compile_context, ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info,
-                       unsigned int depth_multiplier = 1, ActivationLayerInfo act_info = ActivationLayerInfo(), const Size2D &dilation = Size2D(1U, 1U));
-
-        /** Static function to check if given info will lead to a valid configuration of @ref CLDepthwiseConvolutionLayer3x3
-         *
-         * @param[in] input            Source tensor info. Data type supported: QASYMM8 for all layouts, F16/F32 for NCHW.
-         * @param[in] weights          Weights tensor info. A 3D tensor with shape [3, 3, IFM].
-         *                             Data type supported: Same as @p input or QASYMM8/QSYMM8_PER_CHANNEL when @p input is QASYMM8.
-         * @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.
-         * @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 QASYMM8 supported.
-         * @param[in] dilation         (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
-         *
-         * @return a status
-         */
-        static Status validate(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));
-
-        // Inherited methods overriden:
-        void run() override;
-        void prepare() override;
-
-        void set_memory_group(std::shared_ptr<IMemoryManager> memory_manager)
-        {
-            _memory_group = MemoryGroup(std::move(memory_manager));
-        };
-
-    private:
-        MemoryGroup                                               _memory_group;
-        std::unique_ptr<CLDepthwiseConvolutionLayer3x3NCHWKernel> _kernel_nchw;
-        std::unique_ptr<CLDepthwiseConvolutionLayer3x3NHWCKernel> _kernel_nhwc;
-        std::unique_ptr<CLFillBorderKernel>                       _border_handler;
-        CLPermute                                                 _permute_input_to_nchw;
-        CLPermute                                                 _permute_weights_to_nchw;
-        CLPermute                                                 _permute_output_to_nhwc;
-        CLTensor                                                  _permuted_input;
-        CLTensor                                                  _permuted_weights;
-        CLTensor                                                  _permuted_output;
-        CLTensor                                                  _output_multipliers;
-        CLTensor                                                  _output_shifts;
-        const ITensor                                            *_original_weights;
-        const ITensor                                            *_input;
-        const ITensor                                            *_output;
-        bool                                                      _needs_permute;
-        bool                                                      _is_prepared;
-        bool                                                      _is_quantized;
-        bool                                                      _is_nhwc;
-    };
-
-    /** Basic function to execute a generic depthwise convolution. This function calls the following OpenCL kernels:
-     *
-     * -# @ref CLDepthwiseConvolutionLayerNativeKernel
-     * -# @ref CLPermute (x 3) if the data layout is NCHW
-     *
-     */
-    class CLDepthwiseConvolutionLayerGeneric : public IFunction
+    void set_memory_group(std::shared_ptr<IMemoryManager> memory_manager)
     {
-    public:
-        /** Default constructor */
-        CLDepthwiseConvolutionLayerGeneric(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
-        /** Prevent instances of this class from being copied (As this class contains pointers) */
-        CLDepthwiseConvolutionLayerGeneric(const CLDepthwiseConvolutionLayerGeneric &) = delete;
-        /** Default move constructor */
-        CLDepthwiseConvolutionLayerGeneric(CLDepthwiseConvolutionLayerGeneric &&) = default;
-        /** Prevent instances of this class from being copied (As this class contains pointers) */
-        CLDepthwiseConvolutionLayerGeneric &operator=(const CLDepthwiseConvolutionLayerGeneric &) = delete;
-        /** Default move assignment operator */
-        CLDepthwiseConvolutionLayerGeneric &operator=(CLDepthwiseConvolutionLayerGeneric &&) = default;
-        /** Initialize the function's source, destination, weights and convolution information.
-         *
-         * @param[in, out] input            Source tensor. Data type supported: QASYMM8/QASYMM8_SIGNED/F32. (Written to only for border filling).
-         * @param[in]      weights          Weights tensor. 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.
-         * @param[in]      biases           Biases tensor. 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[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(ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info,
-                       unsigned int depth_multiplier = 1, const ActivationLayerInfo &act_info = ActivationLayerInfo(), const Size2D &dilation = Size2D(1U, 1U));
-        /** Initialize the function's source, destination, weights and convolution information.
-         *
-         * @param[in]      compile_context  The compile context to be used.
-         * @param[in, out] input            Source tensor. Data type supported: QASYMM8/QASYMM8_SIGNED/F32. (Written to only for border filling).
-         * @param[in]      weights          Weights tensor. 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.
-         * @param[in]      biases           Biases tensor. 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[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 CLCompileContext &compile_context, ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info,
-                       unsigned int depth_multiplier = 1, const ActivationLayerInfo &act_info = ActivationLayerInfo(), const Size2D &dilation = Size2D(1U, 1U));
-
-        /** Static function to check if given info will lead to a valid configuration of @ref CLDepthwiseConvolutionLayerGeneric
-         *
-         * @param[in] input            Source tensor info. Data type supported: QASYMM8/QASYMM8_SIGNED/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.
-         * @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.
-         * @param[in] dilation         (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
-         *
-         * @return a status
-         */
-        static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info,
-                               unsigned int depth_multiplier = 1, const ActivationLayerInfo &act_info = ActivationLayerInfo(), const Size2D &dilation = Size2D(1U, 1U));
-
-        // Inherited methods overriden:
-        void run() override;
-        void prepare() override;
-
-        void set_memory_group(std::shared_ptr<IMemoryManager> memory_manager)
-        {
-            _memory_group = MemoryGroup(std::move(memory_manager));
-        };
-
-    private:
-        MemoryGroup _memory_group;
-
-        std::unique_ptr<CLDepthwiseConvolutionLayerNativeKernel> _dwc_native_kernel;
-        CLPermute                                                _permute_input_to_nhwc;
-        CLPermute                                                _permute_weights_to_nhwc;
-        CLPermute                                                _permute_output_to_nchw;
-
-        CLTensor       _permuted_input;
-        CLTensor       _permuted_weights;
-        CLTensor       _permuted_output;
-        CLTensor       _output_multipliers;
-        CLTensor       _output_shifts;
-        const ITensor *_original_weights;
-        const ITensor *_input;
-        const ITensor *_output;
-
-        bool _needs_permute;
-        bool _is_prepared;
-        bool _is_quantized;
+        _memory_group = MemoryGroup(std::move(memory_manager));
     };
 
-    std::shared_ptr<IMemoryManager> _memory_manager;
-
-    DepthwiseConvolutionFunction           _depth_conv_func;
-    CLDepthwiseConvolutionLayerInternal3x3 _func_3x3;
-    CLDepthwiseConvolutionLayerGeneric     _func_generic;
+private:
+    MemoryGroup _memory_group;
+
+    std::unique_ptr<CLDepthwiseConvolutionLayerNativeKernel> _dwc_native_kernel;
+    CLPermute                                                _permute_input_to_nhwc;
+    CLPermute                                                _permute_weights_to_nhwc;
+    CLPermute                                                _permute_output_to_nchw;
+
+    CLTensor       _permuted_input;
+    CLTensor       _permuted_weights;
+    CLTensor       _permuted_output;
+    CLTensor       _output_multipliers;
+    CLTensor       _output_shifts;
+    const ITensor *_original_weights;
+    const ITensor *_input;
+    const ITensor *_output;
+
+    bool _needs_permute;
+    bool _is_prepared;
+    bool _is_quantized;
 };
 } // namespace arm_compute
 #endif /*ARM_COMPUTE_CLDEPTHWISECONVOLUTION_H */