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>
diff --git a/arm_compute/core/CL/CLKernelLibrary.h b/arm_compute/core/CL/CLKernelLibrary.h
index c1999b4..741e47c 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 @@
* @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 5749871..3f5a7dc 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 0000000..4b6284c
--- /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 24e91bd..a2dfbb7 100644
--- a/arm_compute/core/Types.h
+++ b/arm_compute/core/Types.h
@@ -134,6 +134,13 @@
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
{