COMPMID-811 Add NHWC data format support for CL depthwise convolution QASYMM8

Change-Id: I89de432f3fbcba7abf9e1d4f8396a4334b4fa2c2
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/118324
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>

6 files changed, 170 insertions, 27 deletions

diff --git a/arm_compute/core/CL/CLKernels.h b/arm_compute/core/CL/CLKernels.h
index 9481e144df..67b7a8bb8f 100644
--- a/arm_compute/core/CL/CLKernels.h
+++ b/arm_compute/core/CL/CLKernels.h
@@ -46,7 +46,8 @@
 #include "arm_compute/core/CL/kernels/CLDeconvolutionLayerUpsampleKernel.h"
 #include "arm_compute/core/CL/kernels/CLDepthConcatenateLayerKernel.h"
 #include "arm_compute/core/CL/kernels/CLDepthConvertLayerKernel.h"
-#include "arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayer3x3Kernel.h"
+#include "arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.h"
+#include "arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.h"
 #include "arm_compute/core/CL/kernels/CLDepthwiseIm2ColKernel.h"
 #include "arm_compute/core/CL/kernels/CLDepthwiseVectorToTensorKernel.h"
 #include "arm_compute/core/CL/kernels/CLDepthwiseWeightsReshapeKernel.h"
@@ -113,5 +114,6 @@
 #include "arm_compute/core/CL/kernels/CLWinogradFilterTransformKernel.h"
 #include "arm_compute/core/CL/kernels/CLWinogradInputTransformKernel.h"
 #include "arm_compute/core/CL/kernels/CLWinogradOutputTransformKernel.h"
+#include "arm_compute/core/CL/kernels/ICLDepthwiseConvolutionLayer3x3Kernel.h"
 
 #endif /* __ARM_COMPUTE_CLKERNELS_H__ */
diff --git a/arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.h b/arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.h
new file mode 100644
index 0000000000..0f3f4bfc76
--- /dev/null
+++ b/arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.h
@@ -0,0 +1,61 @@
+/*
+ * 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_CLDEPTHWISECONVOLUTIONNCHWKERNEL3x3_H__
+#define __ARM_COMPUTE_CLDEPTHWISECONVOLUTIONNCHWKERNEL3x3_H__
+
+#include "arm_compute/core/CL/kernels/ICLDepthwiseConvolutionLayer3x3Kernel.h"
+
+namespace arm_compute
+{
+class ICLTensor;
+
+/** Interface for the kernel to run a 3x3 depthwise convolution on a tensor when the data layout is NCHW.
+ */
+class CLDepthwiseConvolutionLayer3x3NCHWKernel : public ICLDepthwiseConvolutionLayer3x3Kernel
+{
+public:
+    /** Default constructor */
+    CLDepthwiseConvolutionLayer3x3NCHWKernel();
+    /** Initialize the function's source, destination, conv and border_size.
+     *
+     * @param[in]  input     Source tensor. DataType supported: QASYMM8/F16/F32.
+     * @param[in]  weights   Weights tensor. A 3D tensor with dimensions [3, 3, IFM]. Data type supported: Same as @p input.
+     * @param[in]  biases    (Optional) Biases tensor. A 1D tensor with dimensions [IFM]. Must be nullptr if not needed.
+     *                       Data type supported: Same as @p input.
+     * @param[out] output    Destination tensor. Data type supported: Same as @p input.
+     * @param[in]  conv_info Padding and stride information to use for the convolution.
+     * @param[in]  act_info  (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU for QASYMM8 supported.
+     */
+    void configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info,
+                   ActivationLayerInfo act_info) override;
+
+    void run(const Window &window, cl::CommandQueue &queue) override;
+    BorderSize border_size() const override;
+
+private:
+    unsigned int _conv_stride_x;
+    unsigned int _conv_pad_top;
+};
+} // namespace arm_compute
+#endif /*__ARM_COMPUTE_CLDEPTHWISECONVOLUTIONNCHWKERNEL3x3_H__ */
diff --git a/arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.h b/arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.h
new file mode 100644
index 0000000000..4ecc07af6a
--- /dev/null
+++ b/arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.h
@@ -0,0 +1,75 @@
+/*
+ * 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_CLDEPTHWISECONVOLUTIONNHWCKERNEL3x3_H__
+#define __ARM_COMPUTE_CLDEPTHWISECONVOLUTIONNHWCKERNEL3x3_H__
+
+#include "arm_compute/core/CL/kernels/ICLDepthwiseConvolutionLayer3x3Kernel.h"
+
+namespace arm_compute
+{
+class ICLTensor;
+
+/** Interface for the kernel to run a 3x3 depthwise convolution on a tensor when the data layout is NHWC.
+ */
+class CLDepthwiseConvolutionLayer3x3NHWCKernel : public ICLDepthwiseConvolutionLayer3x3Kernel
+{
+public:
+    /** Default constructor */
+    CLDepthwiseConvolutionLayer3x3NHWCKernel();
+    /** Default move assignment operator. */
+    /** Initialize the function's source, destination, conv and border_size.
+     *
+     * @param[in]  input     Source tensor. DataType supported: QASYMM8.
+     * @param[in]  weights   Weights tensor. A 3D tensor with dimensions [IFM, 3, 3]. Data type supported: Same as @p input.
+     * @param[in]  biases    (Optional) Biases tensor. A 1D tensor with dimensions [IFM]. Must be nullptr if not needed.
+     *                       Data type supported: Same as @p input.
+     * @param[out] output    Destination tensor. Data type supported: Same as @p input.
+     * @param[in]  conv_info Padding and stride information to use for the convolution.
+     * @param[in]  act_info  (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU are supported.
+     */
+    void configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info,
+                   ActivationLayerInfo act_info) override;
+    /** Static function to check if given info will lead to a valid configuration of @ref CLDepthwiseConvolutionLayer3x3NHWCKernel
+     *
+     * @param[in] input     Source tensor. DataType supported: QASYMM8.
+     * @param[in] weights   Weights tensor. A 3D tensor with dimensions [IFM, 3, 3]. Data type supported: Same as @p input.
+     * @param[in] biases    (Optional) Biases tensor. A 1D tensor with dimensions [IFM]. Must be nullptr if not needed.
+     *                      Data type supported: Same as @p input.
+     * @param[in] output    Destination tensor. Data type supported: Same as @p input.
+     * @param[in] conv_info Padding and stride information to use for the convolution.
+     * @param[in] act_info  (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU are supported.
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info,
+                           ActivationLayerInfo act_info = ActivationLayerInfo());
+
+    void run(const Window &window, cl::CommandQueue &queue) override;
+    BorderSize border_size() const override;
+
+private:
+    unsigned int _num_rows_processed_per_iteration;
+};
+} // namespace arm_compute
+#endif /*__ARM_COMPUTE_CLDEPTHWISECONVOLUTIONNHWCKERNEL3x3_H__ */
diff --git a/arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayer3x3Kernel.h b/arm_compute/core/CL/kernels/ICLDepthwiseConvolutionLayer3x3Kernel.h
index 5f72cf70ed..f02ba331b5 100644
--- a/arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayer3x3Kernel.h
+++ b/arm_compute/core/CL/kernels/ICLDepthwiseConvolutionLayer3x3Kernel.h
@@ -21,8 +21,8 @@
  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  */
-#ifndef __ARM_COMPUTE_CLDEPTHWISECONVOLUTIONKERNEL3x3_H__
-#define __ARM_COMPUTE_CLDEPTHWISECONVOLUTIONKERNEL3x3_H__
+#ifndef __ARM_COMPUTE_ICLDEPTHWISECONVOLUTIONKERNEL3x3_H__
+#define __ARM_COMPUTE_ICLDEPTHWISECONVOLUTIONKERNEL3x3_H__
 
 #include "arm_compute/core/CL/ICLKernel.h"
 
@@ -32,19 +32,22 @@ class ICLTensor;
 
 /** Interface for the kernel to run a 3x3 depthwise convolution on a tensor.
  */
-class CLDepthwiseConvolutionLayer3x3Kernel : public ICLKernel
+class ICLDepthwiseConvolutionLayer3x3Kernel : public ICLKernel
 {
 public:
     /** Default constructor */
-    CLDepthwiseConvolutionLayer3x3Kernel();
+    ICLDepthwiseConvolutionLayer3x3Kernel()
+        : _border_size(0), _input(), _output(), _weights(), _biases(), _conv_stride_y(1), _conv_pad_left(0)
+    {
+    }
     /** Prevent instances of this class from being copied (As this class contains pointers) */
-    CLDepthwiseConvolutionLayer3x3Kernel(const CLDepthwiseConvolutionLayer3x3Kernel &) = delete;
+    ICLDepthwiseConvolutionLayer3x3Kernel(const ICLDepthwiseConvolutionLayer3x3Kernel &) = delete;
     /** Prevent instances of this class from being copied (As this class contains pointers) */
-    CLDepthwiseConvolutionLayer3x3Kernel &operator=(const CLDepthwiseConvolutionLayer3x3Kernel &) = delete;
+    ICLDepthwiseConvolutionLayer3x3Kernel &operator=(const ICLDepthwiseConvolutionLayer3x3Kernel &) = delete;
     /** Default Move Constructor. */
-    CLDepthwiseConvolutionLayer3x3Kernel(CLDepthwiseConvolutionLayer3x3Kernel &&) = default;
+    ICLDepthwiseConvolutionLayer3x3Kernel(ICLDepthwiseConvolutionLayer3x3Kernel &&) = default;
     /** Default move assignment operator */
-    CLDepthwiseConvolutionLayer3x3Kernel &operator=(CLDepthwiseConvolutionLayer3x3Kernel &&) = default;
+    ICLDepthwiseConvolutionLayer3x3Kernel &operator=(ICLDepthwiseConvolutionLayer3x3Kernel &&) = default;
     /** Initialize the function's source, destination, conv and border_size.
      *
      * @param[in]  input     Source tensor. DataType supported: QASYMM8/F16/F32.
@@ -55,22 +58,17 @@ public:
      * @param[in]  conv_info Padding and stride information to use for the convolution.
      * @param[in]  act_info  (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU for QASYMM8 supported.
      */
-    void configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info, ActivationLayerInfo act_info = ActivationLayerInfo());
+    virtual void configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info,
+                           ActivationLayerInfo act_info = ActivationLayerInfo()) = 0;
 
-    // Inherited methods overridden:
-    void run(const Window &window, cl::CommandQueue &queue) override;
-    BorderSize border_size() const override;
-
-private:
+protected:
     BorderSize       _border_size;
     const ICLTensor *_input;
     ICLTensor       *_output;
     const ICLTensor *_weights;
     const ICLTensor *_biases;
-    unsigned int     _conv_stride_x;
     unsigned int     _conv_stride_y;
     unsigned int     _conv_pad_left;
-    unsigned int     _conv_pad_top;
 };
 } // namespace arm_compute
-#endif /*__ARM_COMPUTE_CLDEPTHWISECONVOLUTIONKERNEL3x3_H__ */
+#endif /*__ARM_COMPUTE_ICLDEPTHWISECONVOLUTIONKERNEL3x3_H__ */
diff --git a/arm_compute/core/utils/misc/ShapeCalculator.h b/arm_compute/core/utils/misc/ShapeCalculator.h
index e174227302..b91e52a657 100644
--- a/arm_compute/core/utils/misc/ShapeCalculator.h
+++ b/arm_compute/core/utils/misc/ShapeCalculator.h
@@ -130,15 +130,19 @@ inline TensorShape compute_depthwise_convolution_shape(const ITensorInfo &input,
     const TensorShape input_shape{ input.tensor_shape() };
     const TensorShape weights_shape{ weights.tensor_shape() };
 
+    const DataLayout data_layout = input.data_layout();
+    const int        width_idx   = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
+    const int        height_idx  = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
+
     unsigned int output_width  = 0;
     unsigned int output_height = 0;
-    std::tie(output_width, output_height) = scaled_dimensions(input_shape.x(), input_shape.y(),
-                                                              weights_shape.x(), weights_shape.y(),
+    std::tie(output_width, output_height) = scaled_dimensions(input_shape[width_idx], input_shape[height_idx],
+                                                              weights_shape[width_idx], weights_shape[height_idx],
                                                               conv_info);
 
     TensorShape output_shape{ input_shape };
-    output_shape.set(0, output_width);
-    output_shape.set(1, output_height);
+    output_shape.set(width_idx, output_width);
+    output_shape.set(height_idx, output_height);
 
     return output_shape;
 }
diff --git a/arm_compute/runtime/CL/functions/CLDepthwiseConvolutionLayer.h b/arm_compute/runtime/CL/functions/CLDepthwiseConvolutionLayer.h
index d6fc8f0fcc..82947bc7e6 100644
--- a/arm_compute/runtime/CL/functions/CLDepthwiseConvolutionLayer.h
+++ b/arm_compute/runtime/CL/functions/CLDepthwiseConvolutionLayer.h
@@ -24,13 +24,15 @@
 #ifndef __ARM_COMPUTE_CLDEPTHWISECONVOLUTION_H__
 #define __ARM_COMPUTE_CLDEPTHWISECONVOLUTION_H__
 
-#include "arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayer3x3Kernel.h"
+#include "arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.h"
+#include "arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.h"
 #include "arm_compute/core/CL/kernels/CLDepthwiseIm2ColKernel.h"
 #include "arm_compute/core/CL/kernels/CLDepthwiseVectorToTensorKernel.h"
 #include "arm_compute/core/CL/kernels/CLDepthwiseWeightsReshapeKernel.h"
 #include "arm_compute/core/CL/kernels/CLDirectConvolutionLayerOutputStageKernel.h"
 #include "arm_compute/core/CL/kernels/CLFillBorderKernel.h"
 #include "arm_compute/core/CL/kernels/CLGEMMMatrixVectorMultiplyKernel.h"
+#include "arm_compute/core/CL/kernels/ICLDepthwiseConvolutionLayer3x3Kernel.h"
 #include "arm_compute/core/Types.h"
 #include "arm_compute/runtime/CL/CLTensor.h"
 #include "arm_compute/runtime/IFunction.h"
@@ -39,9 +41,10 @@ namespace arm_compute
 {
 class ICLTensor;
 
-/** Basic function to execute a depthwise convolution for kernel size 3x3xC. This function calls the following OpenCL kernels:
+/** Basic function to execute a depthwise convolution for kernel size 3x3xC (when data layout NCHW) or Cx3x3 (when data layout NHWC). This function calls the following OpenCL kernels:
  *
- * -# @ref CLDepthwiseConvolutionLayer3x3Kernel
+ * -# @ref CLDepthwiseConvolutionLayer3x3NCHWKernel (if data_layout == NCHW)
+ * -# @ref CLDepthwiseConvolutionLayer3x3NHWCKernel (if data_layout == NHWC)
  * -# @ref CLFillBorderKernel (if pad_x or pad_y > 0)
  *
  */
@@ -66,8 +69,8 @@ public:
     void run() override;
 
 private:
-    CLDepthwiseConvolutionLayer3x3Kernel _kernel;
-    CLFillBorderKernel                   _border_handler;
+    std::unique_ptr<ICLDepthwiseConvolutionLayer3x3Kernel> _kernel;
+    CLFillBorderKernel                                     _border_handler;
 };
 
 /** Basic function to execute a generic depthwise convolution. This function calls the following OpenCL kernels: