APPBROWSER-400: Implement the tensorshift kernel for fixing DC's alignment issue on OpenGL ES

Change-Id: I7a8489bb0fddc72899ea165e414ee87bdbfb45b3
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/118106
Reviewed-by: Joel Liang <joel.liang@arm.com>
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>

5 files changed, 157 insertions, 4 deletions

diff --git a/arm_compute/core/GLES_COMPUTE/GCKernels.h b/arm_compute/core/GLES_COMPUTE/GCKernels.h
index a1f3c278c4..40312d121a 100644
--- a/arm_compute/core/GLES_COMPUTE/GCKernels.h
+++ b/arm_compute/core/GLES_COMPUTE/GCKernels.h
@@ -47,6 +47,7 @@
 #include "arm_compute/core/GLES_COMPUTE/kernels/GCPoolingLayerKernel.h"
 #include "arm_compute/core/GLES_COMPUTE/kernels/GCScaleKernel.h"
 #include "arm_compute/core/GLES_COMPUTE/kernels/GCSoftmaxLayerKernel.h"
+#include "arm_compute/core/GLES_COMPUTE/kernels/GCTensorShiftKernel.h"
 #include "arm_compute/core/GLES_COMPUTE/kernels/GCTransposeKernel.h"
 #include "arm_compute/core/GLES_COMPUTE/kernels/GCWeightsReshapeKernel.h"
 
diff --git a/arm_compute/core/GLES_COMPUTE/kernels/GCTensorShiftKernel.h b/arm_compute/core/GLES_COMPUTE/kernels/GCTensorShiftKernel.h
new file mode 100644
index 0000000000..5f108764b4
--- /dev/null
+++ b/arm_compute/core/GLES_COMPUTE/kernels/GCTensorShiftKernel.h
@@ -0,0 +1,83 @@
+/*
+ * Copyright (c) 2017-2018 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_GCTENSORSHIFTKERNEL_H__
+#define __ARM_COMPUTE_GCTENSORSHIFTKERNEL_H__
+
+#include "arm_compute/core/GLES_COMPUTE/IGCKernel.h"
+#include "arm_compute/core/GLES_COMPUTE/OpenGLES.h"
+
+namespace arm_compute
+{
+class IGCTensor;
+/** Interface for the kernel to shift valid data on a tensor.
+ *
+ * For example shifting 3x3 valid data with padding of 1 to right:
+ * @f[
+ * \left( \begin{array}{ccccc}
+ * 0   & 0   & 0   & 0 & 0 \\
+ * a00 & a01 & a02 & 0 & 0 \\
+ * a10 & a11 & a12 & 0 & 0 \\
+ * a20 & a21 & a22 & 0 & 0 \\
+ * 0   & 0   & 0   & 0 & 0 \\
+ * \end{array} \right)
+ * =
+ * \left( \begin{array}{ccccc}
+ * 0  & 0   & 0   & 0   & 0 \\
+ * 0  & a00 & a01 & a02 & 0 \\
+ * 0  & a10 & a11 & a12 & 0 \\
+ * 0  & a20 & a21 & a22 & 0 \\
+ * 0  & 0   & 0   & 0   & 0 \\
+ * \end{array} \right)
+ * @f]
+ */
+class GCTensorShiftKernel : public IGCKernel
+{
+public:
+    /** Default constructor */
+    GCTensorShiftKernel();
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    GCTensorShiftKernel(const GCTensorShiftKernel &) = delete;
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    GCTensorShiftKernel &operator=(const GCTensorShiftKernel &) = delete;
+    /** Allow instances of this class to be moved */
+    GCTensorShiftKernel(GCTensorShiftKernel &&) = default;
+    /** Allow instances of this class to be moved */
+    GCTensorShiftKernel &operator=(GCTensorShiftKernel &&) = default;
+    /** Default destructor */
+    ~GCTensorShiftKernel() = default;
+    /** Set the input of the kernel.
+     *
+     * @param[in,out] input Source tensor. Data types supported: F16/F32
+     */
+    void configure(IGCTensor *input);
+
+    // Inherited methods overridden:
+    void run(const Window &window) override;
+
+private:
+    IGCTensor    *_input;
+    gles::NDRange _lws;
+};
+}
+#endif /*__ARM_COMPUTE_GCTENSORSHIFTKERNEL_H__ */
diff --git a/arm_compute/runtime/GLES_COMPUTE/GCFunctions.h b/arm_compute/runtime/GLES_COMPUTE/GCFunctions.h
index bbd8218722..6f338568c2 100644
--- a/arm_compute/runtime/GLES_COMPUTE/GCFunctions.h
+++ b/arm_compute/runtime/GLES_COMPUTE/GCFunctions.h
@@ -45,6 +45,7 @@
 #include "arm_compute/runtime/GLES_COMPUTE/functions/GCPoolingLayer.h"
 #include "arm_compute/runtime/GLES_COMPUTE/functions/GCScale.h"
 #include "arm_compute/runtime/GLES_COMPUTE/functions/GCSoftmaxLayer.h"
+#include "arm_compute/runtime/GLES_COMPUTE/functions/GCTensorShift.h"
 #include "arm_compute/runtime/GLES_COMPUTE/functions/GCTranspose.h"
 
 #endif /* __ARM_COMPUTE_GCFUNCTIONS_H__ */
diff --git a/arm_compute/runtime/GLES_COMPUTE/functions/GCDirectConvolutionLayer.h b/arm_compute/runtime/GLES_COMPUTE/functions/GCDirectConvolutionLayer.h
index 5472bdb9ea..c6b948be1f 100644
--- a/arm_compute/runtime/GLES_COMPUTE/functions/GCDirectConvolutionLayer.h
+++ b/arm_compute/runtime/GLES_COMPUTE/functions/GCDirectConvolutionLayer.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -24,9 +24,12 @@
 #ifndef __ARM_COMPUTE_GCDIRECTCONVOLUTIONLAYER_H__
 #define __ARM_COMPUTE_GCDIRECTCONVOLUTIONLAYER_H__
 
+#include "arm_compute/core/GLES_COMPUTE/IGCKernel.h"
 #include "arm_compute/core/GLES_COMPUTE/kernels/GCDirectConvolutionLayerKernel.h"
+#include "arm_compute/core/GLES_COMPUTE/kernels/GCFillBorderKernel.h"
+#include "arm_compute/core/GLES_COMPUTE/kernels/GCTensorShiftKernel.h"
 #include "arm_compute/core/Types.h"
-#include "arm_compute/runtime/GLES_COMPUTE/IGCSimpleFunction.h"
+#include "arm_compute/runtime/IFunction.h"
 
 #include <memory>
 
@@ -34,14 +37,20 @@ namespace arm_compute
 {
 class IGCTensor;
 
-/** Basic function to execute direct convolution function:
+/** Basic function to execute direct convolution function. This function calls the following kernels:
+ *
+ * -# @ref GCDirectConvolutionLayerKernel
+ * -# @ref GCFillBorderKernel
+ * -# @ref GCTensorShiftKernel
  *
  * @note Supported kernel size: 1x1, 3x3, and 5x5
  * @note This OpenGL ES implementation works with stride_x = 1 and 2
  */
-class GCDirectConvolutionLayer : public IGCSimpleFunction
+class GCDirectConvolutionLayer : public IFunction
 {
 public:
+    /** Default constructor */
+    GCDirectConvolutionLayer();
     /** Set the input and output tensors.
      *
      * @param[in]  input     Source tensor. 3 lower dimensions represent a single input [width, height, IFM],
@@ -54,6 +63,14 @@ public:
      * @param[in]  conv_info Contains padding and stride information described in @ref PadStrideInfo.
      */
     void configure(const IGCTensor *input, const IGCTensor *weights, const IGCTensor *biases, IGCTensor *output, const PadStrideInfo &conv_info);
+
+    // Inherited methods overridden:
+    void run() override final;
+
+private:
+    std::unique_ptr<IGCKernel> _kernel;
+    GCFillBorderKernel         _border_handler;
+    GCTensorShiftKernel        _shift_handler;
 };
 }
 #endif /* __ARM_COMPUTE_GCDIRECTCONVOLUTIONLAYER_H__ */
diff --git a/arm_compute/runtime/GLES_COMPUTE/functions/GCTensorShift.h b/arm_compute/runtime/GLES_COMPUTE/functions/GCTensorShift.h
new file mode 100644
index 0000000000..dfcec57044
--- /dev/null
+++ b/arm_compute/runtime/GLES_COMPUTE/functions/GCTensorShift.h
@@ -0,0 +1,51 @@
+/*
+ * Copyright (c) 2017-2018 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_GCTENSORSHIFT_H__
+#define __ARM_COMPUTE_GCTENSORSHIFT_H__
+
+#include "arm_compute/core/GLES_COMPUTE/kernels/GCTensorShiftKernel.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/runtime/GLES_COMPUTE/IGCSimpleFunction.h"
+
+#include <memory>
+
+namespace arm_compute
+{
+class IGCTensor;
+
+/** Basic function to execute shift function for tensor. This function applies to fix alignment issue on OpenGL ES:
+ *
+ * @note This alignment issue is introduced by limits of compute shader which requires 32/64/128bit alignment for data access on OpenGL ES
+ */
+class GCTensorShift : public IGCSimpleFunction
+{
+public:
+    /** Initialise the kernel's input, output.
+     *
+     * @param[in,out] input Source tensor. Data types supported: F16/F32.
+     */
+    void configure(IGCTensor *input);
+};
+}
+#endif /* __ARM_COMPUTE_GCTENSORSHIFT_H__ */