COMPMID-452 CL Depthwise Separable Convolution Layer kernel implementation, validation and benchmarking for 3x3xC depthwise filter and DataType::F32.

Change-Id: I95c0c87709763cdbf58d0de66025eac86e30791b
Reviewed-on: http://mpd-gerrit.cambridge.arm.com/82768
Tested-by: Kaizen <jeremy.johnson+kaizengerrit@arm.com>
Reviewed-by: Steven Niu <steven.niu@arm.com>

2 files changed, 144 insertions, 0 deletions

diff --git a/arm_compute/runtime/CL/functions/CLDepthwiseConvolution.h b/arm_compute/runtime/CL/functions/CLDepthwiseConvolution.h
new file mode 100644
index 0000000000..cc11f9cc5a
--- /dev/null
+++ b/arm_compute/runtime/CL/functions/CLDepthwiseConvolution.h
@@ -0,0 +1,68 @@
+/*
+ * Copyright (c) 2017 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_CL_DEPTHWISE_CONVOLUTION_H__
+#define __ARM_COMPUTE_CL_DEPTHWISE_CONVOLUTION_H__
+
+#include "arm_compute/core/CL/kernels/CLDepthwiseConvolutionKernel.h"
+#include "arm_compute/core/CL/kernels/CLFillBorderKernel.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/runtime/CL/CLTensor.h"
+#include "arm_compute/runtime/CL/ICLSimpleFunction.h"
+#include "arm_compute/runtime/IFunction.h"
+
+#include <cstdint>
+
+namespace arm_compute
+{
+class ICLTensor;
+
+/** Basic function to execute depthwise convolution. This function calls the following OpenCL kernels:
+ *
+ * -# @ref CLFillBorderKernel (executed if border_mode == CONSTANT or border_mode == REPLICATE)
+ * -# @ref CLDepthwiseConvolutionKernel
+ *
+ */
+class CLDepthwiseConvolution : public IFunction
+{
+public:
+    /** Default constructor */
+    CLDepthwiseConvolution();
+    /** Initialize the function's source, destination, conv and border_size.
+     *
+     * @param[in]  input     Source tensor. DataType supported: F32. (Written to only for border filling).
+     * @param[out] output    Destination tensor. DataType supported: F32.
+     * @param[in]  weights   Weights tensor. These are 3D tensors with dimensions [3, 3, IFM]. Data type supported: Same as @p input.
+     * @param[in]  conv_info Padding and stride information to use for the convolution. DataType supported: F32.
+     */
+    void configure(ICLTensor *input, ICLTensor *output, const ICLTensor *weights, const PadStrideInfo &conv_info);
+
+    // Inherited methods overriden:
+    void run() override;
+
+private:
+    CLDepthwiseConvolutionKernel _kernel;
+    CLFillBorderKernel           _border_handler;
+};
+}
+#endif /*__ARM_COMPUTE_CL_DEPTHWISE_CONVOLUTION_H__ */
diff --git a/arm_compute/runtime/CL/functions/CLDepthwiseSeparableConvolutionLayer.h b/arm_compute/runtime/CL/functions/CLDepthwiseSeparableConvolutionLayer.h
new file mode 100644
index 0000000000..7dabed1814
--- /dev/null
+++ b/arm_compute/runtime/CL/functions/CLDepthwiseSeparableConvolutionLayer.h
@@ -0,0 +1,76 @@
+/*
+ * Copyright (c) 2017 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_CL_DEPTHWISE_SEPARABLE_CONVOLUTION_H__
+#define __ARM_COMPUTE_CL_DEPTHWISE_SEPARABLE_CONVOLUTION_H__
+
+#include "arm_compute/core/Types.h"
+#include "arm_compute/runtime/CL/CLTensor.h"
+#include "arm_compute/runtime/CL/ICLSimpleFunction.h"
+#include "arm_compute/runtime/CL/functions/CLDepthwiseConvolution.h"
+#include "arm_compute/runtime/CL/functions/CLDirectConvolutionLayer.h"
+#include "arm_compute/runtime/IFunction.h"
+
+#include <cstdint>
+
+namespace arm_compute
+{
+class ICLTensor;
+
+/** Basic function to execute depthwise convolution. This function calls the following OpenCL kernels and function:
+ *
+ * -# @ref CLDepthwiseConvolution
+ * -# @ref CLDirectConvolutionLayer
+ *
+ */
+class CLDepthwiseSeparableConvolutionLayer : public IFunction
+{
+public:
+    /** Default constructor */
+    CLDepthwiseSeparableConvolutionLayer();
+    /** Set the input and output tensors.
+     *
+     * @param[in]  input               Source tensor. 3 lower dimensions represent a single input [width, height, IFM],
+     *                                 while every optional dimension from 4 and above represent a batch of inputs. Data types supported: F32.
+     * @param[in]  depthwise_weights   Depthwise convolution weights tensor. These are 3D tensors with dimensions [kernel_x, kernel_y, IFM]. Data type supported: Same as @p input.
+     * @param[out] depthwise_out       Depthwise destination tensor.
+     * @param[in]  pointwise_weights   Pointwise convolution weights tensor. These are 4D tensors with dimensions [1, 1, IFM, OFM]. Data type supported: Same as @p input.
+     * @param[in]  biases              Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p weights.
+     * @param[out] output              Destination tensor. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs.
+     *                                 Data types supported: Same as @p input.
+     * @param[in]  depthwise_conv_info Contains padding and stride information described in @ref PadStrideInfo for depthwise convolution.
+     * @param[in]  pointwise_conv_info Contains padding and stride information described in @ref PadStrideInfo for pointwise convolution.
+     */
+    void configure(ICLTensor *input, const ICLTensor *depthwise_weights, ICLTensor *depthwise_out, const ICLTensor *pointwise_weights, const ICLTensor *biases, ICLTensor *output,
+                   const PadStrideInfo &depthwise_conv_info,
+                   const PadStrideInfo &pointwise_conv_info);
+
+    // Inherited methods overriden:
+    void run() override;
+
+private:
+    CLDepthwiseConvolution   _depthwise_conv;
+    CLDirectConvolutionLayer _pointwise_conv;
+};
+}
+#endif /*__ARM_COMPUTE_CL_DEPTHWISE_SEPARABLE_CONVOLUTION_H__ */