COMPMID-935 Implementing Convolution with Winograd on OpenCL (part 3)

Change-Id: I51f92f30602fb0a02314f344fa67061f448694bf
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/122793
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Giorgio Arena <giorgio.arena@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>

5 files changed, 159 insertions, 0 deletions

diff --git a/arm_compute/core/CL/CLKernels.h b/arm_compute/core/CL/CLKernels.h
index e7fa74d00e..ca2cb0411a 100644
--- a/arm_compute/core/CL/CLKernels.h
+++ b/arm_compute/core/CL/CLKernels.h
@@ -109,5 +109,6 @@
 #include "arm_compute/core/CL/kernels/CLWarpAffineKernel.h"
 #include "arm_compute/core/CL/kernels/CLWarpPerspectiveKernel.h"
 #include "arm_compute/core/CL/kernels/CLWeightsReshapeKernel.h"
+#include "arm_compute/core/CL/kernels/CLWinogradInputTransformKernel.h"
 
 #endif /* __ARM_COMPUTE_CLKERNELS_H__ */
diff --git a/arm_compute/core/CL/kernels/CLWinogradInputTransformKernel.h b/arm_compute/core/CL/kernels/CLWinogradInputTransformKernel.h
new file mode 100644
index 0000000000..15cd6e2649
--- /dev/null
+++ b/arm_compute/core/CL/kernels/CLWinogradInputTransformKernel.h
@@ -0,0 +1,79 @@
+/*
+ * Copyright (c) 2018 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_CLWINOGRADINPUTTRANSFORMKERNEL_H__
+#define __ARM_COMPUTE_CLWINOGRADINPUTTRANSFORMKERNEL_H__
+
+#include "arm_compute/core/CL/ICLKernel.h"
+
+namespace arm_compute
+{
+class ICLTensor;
+
+/** OpenCL kernel to perform Winograd input transform.*/
+class CLWinogradInputTransformKernel : public ICLKernel
+{
+public:
+    /** Default constructor */
+    CLWinogradInputTransformKernel();
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    CLWinogradInputTransformKernel(const CLWinogradInputTransformKernel &) = delete;
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    CLWinogradInputTransformKernel &operator=(const CLWinogradInputTransformKernel &) = delete;
+    /** Allow instances of this class to be moved */
+    CLWinogradInputTransformKernel(CLWinogradInputTransformKernel &&) = default;
+    /** Allow instances of this class to be moved */
+    CLWinogradInputTransformKernel &operator=(CLWinogradInputTransformKernel &&) = default;
+    /** Set the input and output of the kernel.
+     *
+     * @param[in] input       The input tensor to permute. Data types supported: F32
+     * @param[in] output      The output tensor. Data types supported: Same as @p input
+     * @param[in] conv_info   Contains padding and stride information described in @ref PadStrideInfo. Currently only unit strides are supported.
+     * @param[in] kernel_dims Kernel dimensions. Currently only 3x3 kernels are supported
+     */
+    void configure(const ICLTensor *input, ICLTensor *output, const PadStrideInfo &conv_info, const Size2D &kernel_dims);
+    /** Static function to check if given info will lead to a valid configuration of @ref CLWinogradInputTransformKernel
+     *
+     * @param[in] input       First tensor input info. Data types supported: F32.
+     * @param[in] output      Output tensor info. Data types supported: same as @p input.
+     * @param[in] conv_info   Contains padding and stride information described in @ref PadStrideInfo. Currently only unit strides are supported.
+     * @param[in] kernel_dims Kernel dimensions. Currently only 3x3 kernels are supported
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *input, const ITensorInfo *output, const PadStrideInfo &conv_info, const Size2D &kernel_dims);
+
+    // Inherited methods overridden:
+    void run(const Window &window, cl::CommandQueue &queue) override;
+    BorderSize border_size() const override;
+
+private:
+    BorderSize       _border_size;
+    const ICLTensor *_input;
+    ICLTensor       *_output;
+    int              _num_tiles_x;
+    int              _num_tiles_y;
+    unsigned int     _step_z;
+};
+} // arm_compute
+#endif /*__ARM_COMPUTE_CLWINOGRADINPUTTRANSFORMKERNEL_H__ */
diff --git a/arm_compute/core/utils/misc/ShapeCalculator.h b/arm_compute/core/utils/misc/ShapeCalculator.h
index e21e5cd0d6..2919625511 100644
--- a/arm_compute/core/utils/misc/ShapeCalculator.h
+++ b/arm_compute/core/utils/misc/ShapeCalculator.h
@@ -194,6 +194,24 @@ inline TensorShape compute_fully_connected_reshaped_weights_shape(const ITensorI
 
     return output_shape;
 }
+
+inline TensorShape compute_winograd_input_transform_shape(const ITensorInfo &input, const PadStrideInfo &conv_info, const Size2D &kernel_size)
+{
+    // Compute height
+    const unsigned int num_tiles_x = std::ceil((input.tensor_shape().x() - (kernel_size.width - 1) + conv_info.pad_left() + conv_info.pad_right()) / 2.f);
+    const unsigned int num_tiles_y = std::ceil((input.tensor_shape().y() - (kernel_size.height - 1) + conv_info.pad_top() + conv_info.pad_bottom()) / 2.f);
+
+    const unsigned int width  = input.tensor_shape()[get_data_layout_dimension_index(input.data_layout(), DataLayoutDimension::CHANNEL)];
+    const unsigned int height = num_tiles_x * num_tiles_y;
+    const unsigned int depth  = 16; // COMPMID-990
+
+    TensorShape output_shape{ input.tensor_shape() };
+    output_shape.set(0, width);
+    output_shape.set(1, height);
+    output_shape.set(2, depth);
+
+    return output_shape;
+}
 } // namespace shape_calculator
 } // namespace misc
 } // namespace arm_compute
diff --git a/arm_compute/runtime/CL/CLFunctions.h b/arm_compute/runtime/CL/CLFunctions.h
index a5bbc41a17..7c2377aa09 100644
--- a/arm_compute/runtime/CL/CLFunctions.h
+++ b/arm_compute/runtime/CL/CLFunctions.h
@@ -107,5 +107,6 @@
 #include "arm_compute/runtime/CL/functions/CLTranspose.h"
 #include "arm_compute/runtime/CL/functions/CLWarpAffine.h"
 #include "arm_compute/runtime/CL/functions/CLWarpPerspective.h"
+#include "arm_compute/runtime/CL/functions/CLWinogradInputTransform.h"
 
 #endif /* __ARM_COMPUTE_CLFUNCTIONS_H__ */
diff --git a/arm_compute/runtime/CL/functions/CLWinogradInputTransform.h b/arm_compute/runtime/CL/functions/CLWinogradInputTransform.h
new file mode 100644
index 0000000000..54b8bdecba
--- /dev/null
+++ b/arm_compute/runtime/CL/functions/CLWinogradInputTransform.h
@@ -0,0 +1,60 @@
+/*
+ * Copyright (c) 2018 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_CLWINOGRADINPUTTRANSFORM_H__
+#define __ARM_COMPUTE_CLWINOGRADINPUTTRANSFORM_H__
+
+#include "arm_compute/core/Types.h"
+#include "arm_compute/runtime/CL/ICLSimpleFunction.h"
+
+#include <cstdint>
+
+namespace arm_compute
+{
+class ICLTensor;
+
+/** Basic function to execute a @ref CLWinogradInputTransformKernel. */
+class CLWinogradInputTransform : public ICLSimpleFunction
+{
+public:
+    /** Set the input and output tensors.
+     *
+     * @param[in] input       The input tensor to transform. Data types supported: F32
+     * @param[in] output      The output tensor. Data types supported: Same as @p input
+     * @param[in] conv_info   Contains padding and stride information described in @ref PadStrideInfo. Currently only unit strides are supported.
+     * @param[in] kernel_dims Kernel dimensions. Currently only 3x3 kernels are supported
+     */
+    void configure(ICLTensor *input, ICLTensor *output, const PadStrideInfo &conv_info, const Size2D &kernel_dims);
+    /**  Static function to check if given info will lead to a valid configuration of @ref CLWinogradInputTransform.
+     *
+     * @param[in] input       First tensor input info. Data types supported: F32.
+     * @param[in] output      Output tensor info. Data types supported: same as @p input.
+     * @param[in] conv_info   Contains padding and stride information described in @ref PadStrideInfo. Currently only unit strides are supported.
+     * @param[in] kernel_dims Kernel dimensions. Currently only 3x3 kernels are supported
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *input, const ITensorInfo *output, const PadStrideInfo &conv_info, const Size2D &kernel_dims);
+};
+}
+#endif /*__ARM_COMPUTE_CLWINOGRADINPUTTRANSFORM_H__ */