COMPMID-935 - Implementing Convolution with Winograd on OpenCL (part 2)

Implemented Winograd Filter Transform 3x3 on OpenCL

Change-Id: I8f2b2dd938c5c000ef7ce392a37fb7b8b4202a4e
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/122708
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Tested-by: Jenkins <bsgcomp@arm.com>

3 files changed, 99 insertions, 1 deletions

diff --git a/arm_compute/core/CL/CLKernels.h b/arm_compute/core/CL/CLKernels.h
index ca2cb0411a..ef629c2e81 100644
--- a/arm_compute/core/CL/CLKernels.h
+++ b/arm_compute/core/CL/CLKernels.h
@@ -109,6 +109,7 @@
 #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/CLWinogradFilterTransformKernel.h"
 #include "arm_compute/core/CL/kernels/CLWinogradInputTransformKernel.h"
 
 #endif /* __ARM_COMPUTE_CLKERNELS_H__ */
diff --git a/arm_compute/core/CL/kernels/CLWinogradFilterTransformKernel.h b/arm_compute/core/CL/kernels/CLWinogradFilterTransformKernel.h
new file mode 100644
index 0000000000..ec5e51482a
--- /dev/null
+++ b/arm_compute/core/CL/kernels/CLWinogradFilterTransformKernel.h
@@ -0,0 +1,74 @@
+/*
+ * 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_CLWINOGRADFILTERTRANSFORMKERNEL_H__
+#define __ARM_COMPUTE_CLWINOGRADFILTERTRANSFORMKERNEL_H__
+
+#include "arm_compute/core/CL/ICLKernel.h"
+
+namespace arm_compute
+{
+class ICLTensor;
+
+/** Interface for the Winograd filter transform kernel. */
+class CLWinogradFilterTransformKernel : public ICLKernel
+{
+public:
+    /** Default constructor */
+    CLWinogradFilterTransformKernel();
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    CLWinogradFilterTransformKernel(const CLWinogradFilterTransformKernel &) = delete;
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    CLWinogradFilterTransformKernel &operator=(const CLWinogradFilterTransformKernel &) = delete;
+    /** Allow instances of this class to be moved */
+    CLWinogradFilterTransformKernel(CLWinogradFilterTransformKernel &&) = default;
+    /** Allow instances of this class to be moved */
+    CLWinogradFilterTransformKernel &operator=(CLWinogradFilterTransformKernel &&) = default;
+    /** Default destructor */
+    ~CLWinogradFilterTransformKernel() = default;
+    /** Set the input and output tensor.
+     *
+     * @param[in]  input  Source tensor. The input is a 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM] (NCHW data layout).
+     *                    kernel_x must be 3 and equal to kernel_y. Data types supported: F32.
+     * @param[out] output Destination tensor. The output is a 3D tensor with dimensions [OFM, IFM, 16]. Data type supported: same as @p input
+     */
+    void configure(const ICLTensor *input, ICLTensor *output);
+    /** Static function to check if given info will lead to a valid configuration of @ref CLWinogradFilterTransformKernel
+     *
+     * @param[in] input  Source tensor info. The input is a 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM] (NCHW data layout).
+     *                   kernel_x must be 3 and equal to kernel_y. Data types supported: F32.
+     * @param[in] output Destination tensor info. The output is a 3D tensor with dimensions [OFM, IFM, 16]. Data type supported: same as @p input
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *input, const ITensorInfo *output);
+
+    // Inherited methods overridden:
+    void run(const Window &window, cl::CommandQueue &queue) override;
+
+private:
+    const ICLTensor *_input;
+    ICLTensor       *_output;
+};
+} // namespace arm_compute
+#endif /*__ARM_COMPUTE_CLWINOGRADFILTERTRANSFORMKERNEL_H__ */
diff --git a/arm_compute/core/utils/misc/ShapeCalculator.h b/arm_compute/core/utils/misc/ShapeCalculator.h
index 354f60d016..9cb8023463 100644
--- a/arm_compute/core/utils/misc/ShapeCalculator.h
+++ b/arm_compute/core/utils/misc/ShapeCalculator.h
@@ -184,7 +184,7 @@ inline TensorShape compute_fully_connected_reshaped_weights_shape(const ITensorI
         output_shape = compute_transposed_shape(*input);
     }
 
-    // If the we run multiple batches we need 1xW transpose, too.
+    // If we run multiple batches we need 1xW transpose, too.
     if(is_batched_fc_layer)
     {
         output_shape = compute_transposed_shape(input->clone()->set_tensor_shape(output_shape));
@@ -193,6 +193,29 @@ inline TensorShape compute_fully_connected_reshaped_weights_shape(const ITensorI
 
     return output_shape;
 }
+
+inline TensorShape compute_winograd_filter_transform_shape(const ITensorInfo &input)
+{
+    // COMPMID-984 (giaiod01)
+    TensorShape tensor_shape{ input.tensor_shape() };
+
+    if(input.data_layout() == DataLayout::NCHW)
+    {
+        tensor_shape.remove_dimension(0);
+        tensor_shape.set(Window::DimX, input.dimension(3));
+        tensor_shape.set(Window::DimY, input.dimension(2));
+        tensor_shape.set(Window::DimZ, 16);
+    }
+    else
+    {
+        tensor_shape.remove_dimension(1);
+        tensor_shape.set(Window::DimY, input.dimension(2));
+        tensor_shape.set(Window::DimZ, 16);
+    }
+
+    return tensor_shape;
+}
+
 inline TensorShape compute_winograd_input_transform_shape(const ITensorInfo &input, const PadStrideInfo &conv_info, const Size2D &kernel_size)
 {
     // Compute height