COMPMID-1635: Optimize CLDeconvolutionLayer - Part III

Change-Id: Id2661e093a669ef3eaf2a5116cd278a80c1d5a89
Signed-off-by: giuros01 <giuseppe.rossini@arm.com>
Reviewed-on: https://review.mlplatform.org/c/935
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Comments-Addressed: Gian Marco Iodice <gianmarco.iodice@arm.com>
Reviewed-by: Isabella Gottardi <isabella.gottardi@arm.com>
Comments-Addressed: Isabella Gottardi <isabella.gottardi@arm.com>
Tested-by: Isabella Gottardi <isabella.gottardi@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

7 files changed, 255 insertions, 5 deletions

diff --git a/arm_compute/core/CL/CLKernelLibrary.h b/arm_compute/core/CL/CLKernelLibrary.h
index c1999b45e1..741e47c65f 100644
--- a/arm_compute/core/CL/CLKernelLibrary.h
+++ b/arm_compute/core/CL/CLKernelLibrary.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2018 ARM Limited.
+ * Copyright (c) 2016-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -333,7 +333,7 @@ public:
      * @param[in] built_program_name Name of the program
      * @param[in] program            Built program to add to the cache
      */
-    void add_built_program(const std::string &built_program_name, cl::Program program);
+    void add_built_program(const std::string &built_program_name, const cl::Program &program);
 
     /** Returns true if FP16 is supported by the CL device
      *
diff --git a/arm_compute/core/CL/CLKernels.h b/arm_compute/core/CL/CLKernels.h
index 57498715c8..3f5a7dc241 100644
--- a/arm_compute/core/CL/CLKernels.h
+++ b/arm_compute/core/CL/CLKernels.h
@@ -48,6 +48,7 @@
 #include "arm_compute/core/CL/kernels/CLCopyKernel.h"
 #include "arm_compute/core/CL/kernels/CLCropKernel.h"
 #include "arm_compute/core/CL/kernels/CLDeconvolutionLayerUpsampleKernel.h"
+#include "arm_compute/core/CL/kernels/CLDeconvolutionReshapeOutputKernel.h"
 #include "arm_compute/core/CL/kernels/CLDepthConcatenateLayerKernel.h"
 #include "arm_compute/core/CL/kernels/CLDepthConvertLayerKernel.h"
 #include "arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.h"
diff --git a/arm_compute/core/CL/kernels/CLDeconvolutionReshapeOutputKernel.h b/arm_compute/core/CL/kernels/CLDeconvolutionReshapeOutputKernel.h
new file mode 100644
index 0000000000..4b6284cfff
--- /dev/null
+++ b/arm_compute/core/CL/kernels/CLDeconvolutionReshapeOutputKernel.h
@@ -0,0 +1,93 @@
+/*
+ * 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_CLDECONVOLUTIONLAYERRESHAPEOUTPUTKERNEL_H__
+#define __ARM_COMPUTE_CLDECONVOLUTIONLAYERRESHAPEOUTPUTKERNEL_H__
+
+#include "arm_compute/core/CL/ICLSimpleKernel.h"
+
+namespace arm_compute
+{
+class ICLTensor;
+
+/** Interface for the OpenCL kernel to be used for reshaping the tensor before returning the result of deconvolution.
+ *
+ * The input tensor to this OpenCL kernel is expected to be the result of a @ref CLGEMM operation between the Deconvolution input and the Deconvolution filter.
+ *
+ * The input tensor should have the following shape: [filter_width * filter_height * ofms, width, height, batch_size]
+ *
+ * The output tensor should have the following shape: [stride_x * (input_width - 1) + filter_width - 2 * padx, stride_y * (input_height - 1) + filter_height - 2 * pady, ofms, batch_size]
+ *
+ * For example, given a tensor with dimensions [4, 2, 2] this function returns a tensor with dimensions [1, 4, 4].
+ *
+ */
+class CLDeconvolutionReshapeOutputKernel : public ICLSimpleKernel
+{
+public:
+    /** Default constructor */
+    CLDeconvolutionReshapeOutputKernel();
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    CLDeconvolutionReshapeOutputKernel(const CLDeconvolutionReshapeOutputKernel &) = delete;
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    CLDeconvolutionReshapeOutputKernel &operator=(const CLDeconvolutionReshapeOutputKernel &) = delete;
+    /** Allow instances of this class to be moved */
+    CLDeconvolutionReshapeOutputKernel(CLDeconvolutionReshapeOutputKernel &&) = default;
+    /** Allow instances of this class to be moved */
+    CLDeconvolutionReshapeOutputKernel &operator=(CLDeconvolutionReshapeOutputKernel &&) = default;
+    /** Default destructor */
+    ~CLDeconvolutionReshapeOutputKernel() = default;
+
+    /** Initialise the kernel's source and destination.
+     *
+     * @param[in]  input        Input tensor. Supported data types: F16/F32.
+     * @param[in]  bias         Bias tensor to be added directly during the reshape operation. Supported data types: same as @p input.  Supported data layouts: same as @p input.
+     * @param[out] output       Output tensor with the following shape: [stride_x * (input_width - 1) + filter_width - 2 * padx, stride_y * (input_height - 1) + filter_height - 2 * pady, ofms, batch_size]
+     *                          Supported data types: same as @p input.  Supported data layouts: same as @p input.
+     * @param[in]  input_info   Deconvolution input tensor info. Supported data types: same as @p input.  Supported data layouts: same as @p input.
+     * @param[in]  weights_info Deconvolution weights tensor info. Supported data types: same as @p input.  Supported data layouts: same as @p input.
+     * @param[in]  deconv_info  Contains padding and policies to be used in the deconvolution, this is described in @ref PadStrideInfo. This kernel supports only stride_x = weights.width && stride_y = weights.height. Moreover, padding is not supported.
+     */
+    void configure(const ICLTensor *input, const ICLTensor *bias, ICLTensor *output, const ITensorInfo *input_info, const ITensorInfo *weights_info, const PadStrideInfo &deconv_info);
+
+    /** Static function to check if given info will lead to a valid configuration of @ref  CLDeconvolutionReshapeOutputKernel.
+     *
+     * @param[in] input        GEMM output tensor info to be reshaped. Supported data types: F16/F32.
+     * @param[in] bias         (Optional) Optional bias tensor info to be added directly during the reshape operation. Supported data types: same as @p input.  Supported data layouts: same as @p input.
+     * @param[in] output       Reshaped output tensor info. Supported data types: same as @p input.  Supported data layouts: same as @p input.
+     * @param[in] input_info   Original input tensor info. Supported data types: same as @p input.  Supported data layouts: same as @p input.
+     * @param[in] weights_info Original weights tensor info output. Supported data types: same as @p input.  Supported data layouts: same as @p input.
+     * @param[in] deconv_info  Contains padding and policies to be used in the deconvolution, this is described in @ref PadStrideInfo. This kernel supports only stride_x = weights.width && stride_y = weights.height. Moreover, padding is not supported.
+     *
+     * @return a Status
+     */
+    static Status validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, const ITensorInfo *input_info, const ITensorInfo *weights_info, const PadStrideInfo &deconv_info);
+
+    // Inherited methods overridden:
+    void run(const Window &window, cl::CommandQueue &queue) override;
+
+private:
+    bool             _add_bias;
+    const ICLTensor *_bias;
+};
+} // namespace arm_compute
+#endif /*__ARM_COMPUTE_CLDECONVOLUTIONLAYERRESHAPEOUTPUTKERNEL_H__ */
diff --git a/arm_compute/core/Types.h b/arm_compute/core/Types.h
index 24e91bd3c5..a2dfbb7d82 100644
--- a/arm_compute/core/Types.h
+++ b/arm_compute/core/Types.h
@@ -134,6 +134,13 @@ enum class ConvolutionMethod
     FFT       /**< Convolution using FFT */
 };
 
+/** Available DeconvolutionMethod*/
+enum class DeconvolutionMethod
+{
+    GEMM,   /**< Deconvolution using GEMM */
+    DIRECT, /**< Direct deconvolution */
+};
+
 /** Padding mode to use for PadLayer */
 enum class PaddingMode
 {
diff --git a/arm_compute/runtime/CL/CLFunctions.h b/arm_compute/runtime/CL/CLFunctions.h
index a4fcdc27ac..129be4b307 100644
--- a/arm_compute/runtime/CL/CLFunctions.h
+++ b/arm_compute/runtime/CL/CLFunctions.h
@@ -77,6 +77,7 @@
 #include "arm_compute/runtime/CL/functions/CLFuseBatchNormalization.h"
 #include "arm_compute/runtime/CL/functions/CLGEMM.h"
 #include "arm_compute/runtime/CL/functions/CLGEMMConvolutionLayer.h"
+#include "arm_compute/runtime/CL/functions/CLGEMMDeconvolutionLayer.h"
 #include "arm_compute/runtime/CL/functions/CLGEMMInterleave4x4.h"
 #include "arm_compute/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.h"
 #include "arm_compute/runtime/CL/functions/CLGEMMLowpOutputStage.h"
diff --git a/arm_compute/runtime/CL/functions/CLDeconvolutionLayer.h b/arm_compute/runtime/CL/functions/CLDeconvolutionLayer.h
index b613708c50..e5b406ee5e 100644
--- a/arm_compute/runtime/CL/functions/CLDeconvolutionLayer.h
+++ b/arm_compute/runtime/CL/functions/CLDeconvolutionLayer.h
@@ -25,6 +25,7 @@
 #define __ARM_COMPUTE_CLDECONVOLUTIONLAYER_H__
 
 #include "arm_compute/runtime/CL/functions/CLDirectDeconvolutionLayer.h"
+#include "arm_compute/runtime/CL/functions/CLGEMMDeconvolutionLayer.h"
 #include "arm_compute/runtime/IFunction.h"
 #include "arm_compute/runtime/IMemoryManager.h"
 
@@ -34,6 +35,7 @@ namespace arm_compute
 {
 /** Basic function to compute the deconvolution layer. This function calls the following OpenCL kernels/functions:
  *
+ * -# @ref CLGEMMDeconvolutionLayer
  * -# @ref CLDirectDeconvolutionLayer
  */
 class CLDeconvolutionLayer : public IFunction
@@ -44,7 +46,7 @@ public:
 
     /** Set the input, weights, biases and output tensors.
      *
-     * @deprecated This method is deprecated and will be removed in release 19.05
+     * @note This method will be deprecated in the next release.
      *
      * @param[in,out] input              Input tensor. 3 lower dimensions represent a single input, and an optional 4th dimension for batch of inputs. Data types supported: QASYMM8/F16/F32.
      * @param[in]     weights            The 4d weights with dimensions [width, height, IFM, OFM]. Data type supported: Same as @p input.
@@ -60,7 +62,7 @@ public:
                    unsigned int inner_border_right, unsigned int inner_border_top, const WeightsInfo &weights_info = WeightsInfo());
     /** Static function to check if given info will lead to a valid configuration of @ref CLDeconvolutionLayer
      *
-     * @deprecated This method is deprecated and will be removed in release 19.05
+     * @note This method will be deprecated in the next release.
      *
      * @param[in] input              Input tensor info. 3 lower dimensions represent a single input, and an optional 4th dimension for batch of inputs. Data types supported: QASYMM8/F16/F32.
      * @param[in] weights            The 4d weights info with dimensions [width, height, IFM, OFM]. Data type supported: Same as @p input.
@@ -101,6 +103,8 @@ public:
     static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *bias, ITensorInfo *output, const PadStrideInfo &deconv_info,
                            const WeightsInfo &weights_info = WeightsInfo());
 
+    static DeconvolutionMethod get_deconvolution_method(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *bias, ITensorInfo *output, const PadStrideInfo &deconv_info,
+                                                        const WeightsInfo &weights_info);
     // Inherited methods overridden:
     void run() override;
     void prepare() override;
@@ -109,5 +113,5 @@ private:
     std::shared_ptr<IMemoryManager> _memory_manager;
     std::unique_ptr<IFunction>      _function;
 };
-}
+} // namespace arm_compute
 #endif /* __ARM_COMPUTE_CLDECONVOLUTIONLAYER_H__ */
diff --git a/arm_compute/runtime/CL/functions/CLGEMMDeconvolutionLayer.h b/arm_compute/runtime/CL/functions/CLGEMMDeconvolutionLayer.h
new file mode 100644
index 0000000000..b28fa0f3f5
--- /dev/null
+++ b/arm_compute/runtime/CL/functions/CLGEMMDeconvolutionLayer.h
@@ -0,0 +1,144 @@
+/*
+ * 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_CLGEMMDECONVOLUTIONLAYER_H__
+#define __ARM_COMPUTE_CLGEMMDECONVOLUTIONLAYER_H__
+
+#include "arm_compute/core/CL/kernels/CLDeconvolutionReshapeOutputKernel.h"
+#include "arm_compute/runtime/CL/CLMemoryGroup.h"
+#include "arm_compute/runtime/CL/CLTensor.h"
+#include "arm_compute/runtime/CL/functions/CLConvolutionLayer.h"
+#include "arm_compute/runtime/CL/functions/CLPermute.h"
+#include "arm_compute/runtime/CL/functions/CLReshapeLayer.h"
+#include "arm_compute/runtime/CL/functions/CLSlice.h"
+#include "arm_compute/runtime/CL/functions/CLTranspose.h"
+#include "arm_compute/runtime/IFunction.h"
+#include "arm_compute/runtime/IMemoryManager.h"
+
+#include <memory>
+
+namespace arm_compute
+{
+class ICLTensor;
+/** Function to run the deconvolution layer through a call to GEMM.
+ *
+ * Deconvolution Layer is the backward pass of Convolution Layer. First we transform the input depending on the stride and pad info and then perform a 1x1
+ * convolution pass. Input stride defines how many zeroes we should put between each element of the input, pad is the amount of padding and finally a is a user
+ * specified value where a < stride - 1, that increases the padding top and right of the input image.
+ *
+ *  The relation between input to output is as follows:
+ *  \f[
+ *       width\_output = (width\_input - 1) \cdot stride\_x - 2 \cdot padding\_x + kernel\_x
+ *  \f]
+ *  \f[
+ *       height\_output = (height\_input - 1) \cdot stride\_y - 2 \cdot padding\_y + kernel\_y
+ *  \f]
+ *
+ *  where:
+ *      width_input is the size of the first input dimension.
+ *      height_input is the size of the second input dimension.
+ *      width_output is the size of the first output dimension.
+ *      height_output is the size of the second output dimension.
+ *      kernel_x and kernel_y are the convolution sizes in x and y.
+ *      stride_x and stride_y is the input stride of the first and second dimension.
+ *
+ * The weights used by Deconvolution are supposed to be the same as the ones used for Convolution.
+ *
+ * This function calls the following OpenCL kernels/functions:
+ *
+ * -# @ref CLGEMMLowpMatrixMultiplyCore
+ * -# @ref CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint
+ * -# @ref CLPermute
+ * -# @ref CLPermute
+ * -# @ref CLReshapeLayer
+ * -# @ref CLTranspose
+ * -# @ref CLDeconvolutionReshapeOutputKernel
+ * -# @ref CLSlice
+ */
+class CLGEMMDeconvolutionLayer : public IFunction
+{
+public:
+    /** Constructor */
+    CLGEMMDeconvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    CLGEMMDeconvolutionLayer(const CLGEMMDeconvolutionLayer &) = delete;
+    /** Default move constructor */
+    CLGEMMDeconvolutionLayer(CLGEMMDeconvolutionLayer &&) = default;
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    CLGEMMDeconvolutionLayer &operator=(const CLGEMMDeconvolutionLayer &) = delete;
+    /** Default move assignment operator */
+    CLGEMMDeconvolutionLayer &operator=(CLGEMMDeconvolutionLayer &&) = default;
+    /** Set the input, weights, biases and output tensors.
+     *
+     * @param[in,out] input       Input tensor. 3 lower dimensions represent a single input, and an optional 4th dimension for batch of inputs. Data types supported: F16/F32. Data layout supported: NHWC
+     * @param[in]     weights     The 4d weights with dimensions [width, height, IFM, OFM]. Data type supported: Same as @p input. Data layout supported: same as @p input.
+     * @param[in]     bias        (Optional) The biases have one dimension. Data type supported: Same as @p input. Data layout supported: same as @p input.
+     * @param[out]    output      Output tensor. The output has the same number of dimensions as the @p input. Data layout supported: same as @p input.
+     * @param[in]     deconv_info Contains padding and policies to be used in the deconvolution, this is described in @ref PadStrideInfo. This function supports only stride_x = weights.width && stride_y = weights.height. Moreover, padding is not supported.
+     */
+    void configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *bias, ICLTensor *output, const PadStrideInfo &deconv_info);
+    /** Static function to check if given info will lead to a valid configuration of @ref CLDeconvolutionLayer
+     *
+     * @param[in] input       Input tensor info. 3 lower dimensions represent a single input, and an optional 4th dimension for batch of inputs. Data types supported: F16/F32. Data layout supported: NHWC
+     * @param[in] weights     The 4d weights info with dimensions [width, height, IFM, OFM]. Data type supported: Same as @p input. Data layout supported: same as @p input.
+     * @param[in] bias        (Optional) The biases have one dimension. Data type supported: Same as @p input. Data layout supported: same as @p input.
+     * @param[in] output      Output tensor info. The output has the same number of dimensions as the @p input. Data layout supported: same as @p input.
+     * @param[in] deconv_info Contains padding and policies to be used in the deconvolution, this is described in @ref PadStrideInfo.
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *bias, const ITensorInfo *output, const PadStrideInfo &deconv_info);
+
+    // Inherited methods overridden:
+    void run() override;
+    void prepare() override;
+
+private:
+    CLMemoryGroup _memory_group;
+
+    CLGEMM                                              _mm_gemm;
+    CLGEMMLowpMatrixMultiplyCore                        _mm_gemmlowp;
+    CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint _gemmlowp_output_stage;
+    CLPermute                                           _permute_input_to_nhwc;
+    CLPermute                                           _permute_weights_to_nhwc;
+    CLReshapeLayer                                      _reshape_weights;
+    CLTranspose                                         _transpose_weights;
+    CLDeconvolutionReshapeOutputKernel                  _deconv_reshape;
+    CLSlice                                             _slice_gemm;
+
+    CLTensor _gemmlowp_final;
+    CLTensor _reshaped_weights;
+    CLTensor _reshaped_weights_t;
+    CLTensor _permuted_input;
+    CLTensor _permuted_weights;
+    CLTensor _gemm_output;
+    CLTensor _slice_gemm_input;
+
+    const ICLTensor *_original_weights;
+    bool             _is_prepared;
+    bool             _padded_input;
+    bool             _is_nchw;
+    bool             _is_quantized;
+};
+} // namespace arm_compute
+#endif /* __ARM_COMPUTE_CLGEMMDECONVOLUTIONLAYER_H__ */