COMPMID-1050 CL/NEON: Create a function to convert the 2D weights of FC layer from NHWC to NCHW and viceversa

Change-Id: If77ffeb92b6eb883e5d2d2c97c2c4d1d23d17c8d
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/129257
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>

8 files changed, 301 insertions, 1 deletions

diff --git a/arm_compute/core/CL/CLKernels.h b/arm_compute/core/CL/CLKernels.h
index 2a9900d9fc..9316c6e489 100644
--- a/arm_compute/core/CL/CLKernels.h
+++ b/arm_compute/core/CL/CLKernels.h
@@ -42,6 +42,7 @@
 #include "arm_compute/core/CL/kernels/CLChannelShuffleLayerKernel.h"
 #include "arm_compute/core/CL/kernels/CLCol2ImKernel.h"
 #include "arm_compute/core/CL/kernels/CLColorConvertKernel.h"
+#include "arm_compute/core/CL/kernels/CLConvertFullyConnectedWeightsKernel.h"
 #include "arm_compute/core/CL/kernels/CLConvolutionKernel.h"
 #include "arm_compute/core/CL/kernels/CLCopyKernel.h"
 #include "arm_compute/core/CL/kernels/CLDeconvolutionLayerUpsampleKernel.h"
diff --git a/arm_compute/core/CL/kernels/CLConvertFullyConnectedWeightsKernel.h b/arm_compute/core/CL/kernels/CLConvertFullyConnectedWeightsKernel.h
new file mode 100644
index 0000000000..fe24aa9d8c
--- /dev/null
+++ b/arm_compute/core/CL/kernels/CLConvertFullyConnectedWeightsKernel.h
@@ -0,0 +1,81 @@
+/*
+ * 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_CLCONVERTFULLYCONNECTEDWEIGHTSKERNEL_H__
+#define __ARM_COMPUTE_CLCONVERTFULLYCONNECTEDWEIGHTSKERNEL_H__
+
+#include "arm_compute/core/CL/ICLKernel.h"
+
+namespace arm_compute
+{
+class ICLTensor;
+
+/** Interface to convert the 2D Fully Connected weights from NCHW to NHWC or vice versa.
+ *
+ * @note This function can be applied to the 2D weights used by a Fully Connected layer if:
+ *       - It follows a Convolution layer
+ *       - The data layout used by the network does not match the one the model has been trained in.
+ *
+ * @note This function assumes the weights are already reshaped (transposed)
+ */
+class CLConvertFullyConnectedWeightsKernel : public ICLKernel
+{
+public:
+    /** Default constructor */
+    CLConvertFullyConnectedWeightsKernel();
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    CLConvertFullyConnectedWeightsKernel(const CLConvertFullyConnectedWeightsKernel &) = delete;
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    CLConvertFullyConnectedWeightsKernel &operator=(const CLConvertFullyConnectedWeightsKernel &) = delete;
+    /** Allow instances of this class to be moved */
+    CLConvertFullyConnectedWeightsKernel(CLConvertFullyConnectedWeightsKernel &&) = default;
+    /** Allow instances of this class to be moved */
+    CLConvertFullyConnectedWeightsKernel &operator=(CLConvertFullyConnectedWeightsKernel &&) = default;
+    /** Default destructor */
+    ~CLConvertFullyConnectedWeightsKernel() = default;
+    /** Set the input and output tensor.
+     *
+     * @param[in]  input                Source weights tensor to convert. Must be 2 dimensional. Data types supported: U8/S8/QS8/QASYMM8/U16/S16/QS16/U32/S32/QS32/F16/F32.
+     * @param[out] output               The converted weights tensor. Shape and Data Type: Same as @p input.
+     * @param[in]  original_input_shape Shape of the original input tensor (the one entering fully connected layer). Must be in NCHW format.
+     * @param[in]  data_layout          The data layout the weights have been trained in.
+     */
+    void configure(const ICLTensor *input, ICLTensor *output, const TensorShape &original_input_shape, DataLayout data_layout);
+    /** Static function to check if given info will lead to a valid configuration of @ref CLConvertFullyConnectedWeightsKernel
+     *
+     * @param[in] input                Source weights tensor info to convert. Must be 2 dimensional. Data types supported: U8/S8/QS8/QASYMM8/U16/S16/QS16/U32/S32/QS32/F16/F32.
+     * @param[in] output               The converted weights tensor info. Shape and Data Type: Same as @p input.
+     * @param[in] original_input_shape Shape of the original input tensor (the one entering fully connected layer). Must be in NCHW format.
+     * @param[in] data_layout          The data layout the weights have been trained in.
+     */
+    static Status validate(const ITensorInfo *input, const ITensorInfo *output, const TensorShape &original_input_shape, DataLayout data_layout);
+
+    // Inherited methods overridden:
+    void run(const Window &window, cl::CommandQueue &queue) override;
+
+private:
+    const ICLTensor *_input;
+    ICLTensor       *_output;
+};
+} // namespace arm_compute
+#endif /*__ARM_COMPUTE_CLCONVERTFULLYCONNECTEDWEIGHTSKERNEL_H__ */
diff --git a/arm_compute/core/NEON/NEKernels.h b/arm_compute/core/NEON/NEKernels.h
index 0e271efa49..b9c54b2c25 100644
--- a/arm_compute/core/NEON/NEKernels.h
+++ b/arm_compute/core/NEON/NEKernels.h
@@ -41,6 +41,7 @@
 #include "arm_compute/core/NEON/kernels/NEChannelExtractKernel.h"
 #include "arm_compute/core/NEON/kernels/NECol2ImKernel.h"
 #include "arm_compute/core/NEON/kernels/NEColorConvertKernel.h"
+#include "arm_compute/core/NEON/kernels/NEConvertFullyConnectedWeightsKernel.h"
 #include "arm_compute/core/NEON/kernels/NEConvolutionKernel.h"
 #include "arm_compute/core/NEON/kernels/NECumulativeDistributionKernel.h"
 #include "arm_compute/core/NEON/kernels/NEDepthConcatenateLayerKernel.h"
diff --git a/arm_compute/core/NEON/kernels/NEConvertFullyConnectedWeightsKernel.h b/arm_compute/core/NEON/kernels/NEConvertFullyConnectedWeightsKernel.h
new file mode 100644
index 0000000000..65ce764246
--- /dev/null
+++ b/arm_compute/core/NEON/kernels/NEConvertFullyConnectedWeightsKernel.h
@@ -0,0 +1,94 @@
+/*
+ * 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_NECONVERTFULLYCONNECTEDWEIGHTSKERNEL_H__
+#define __ARM_COMPUTE_NECONVERTFULLYCONNECTEDWEIGHTSKERNEL_H__
+
+#include "arm_compute/core/NEON/INEKernel.h"
+
+namespace arm_compute
+{
+class ITensor;
+
+/** Interface to convert the 2D Fully Connected weights from NCHW to NHWC or vice versa.
+ *
+ * @note This function can be applied to the 2D weights used by a Fully Connected layer if:
+ *       - It follows a Convolution layer
+ *       - The data layout used by the network does not match the one the model has been trained in.
+ *
+ * @note This function assumes the weights are already reshaped (transposed)
+ */
+class NEConvertFullyConnectedWeightsKernel : public INEKernel
+{
+public:
+    const char *name() const override
+    {
+        return "NEConvertFullyConnectedWeightsKernel";
+    }
+    /** Default constructor */
+    NEConvertFullyConnectedWeightsKernel();
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    NEConvertFullyConnectedWeightsKernel(const NEConvertFullyConnectedWeightsKernel &) = delete;
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    NEConvertFullyConnectedWeightsKernel &operator=(const NEConvertFullyConnectedWeightsKernel &) = delete;
+    /** Allow instances of this class to be moved */
+    NEConvertFullyConnectedWeightsKernel(NEConvertFullyConnectedWeightsKernel &&) = default;
+    /** Allow instances of this class to be moved */
+    NEConvertFullyConnectedWeightsKernel &operator=(NEConvertFullyConnectedWeightsKernel &&) = default;
+    /** Default destructor */
+    ~NEConvertFullyConnectedWeightsKernel() = default;
+    /** Set the input and output tensor.
+     *
+     * @param[in]  input                Source weights tensor to convert. Must be 2 dimensional. Data types supported: U8/S8/QS8/QASYMM8/U16/S16/QS16/U32/S32/QS32/F16/F32.
+     * @param[out] output               The converted weights tensor. Shape and Data Type: Same as @p input.
+     * @param[in]  original_input_shape Shape of the original input tensor (the one entering fully connected layer). Must be in NCHW format.
+     * @param[in]  data_layout          The data layout the weights have been trained in.
+     */
+    void configure(const ITensor *input, ITensor *output, const TensorShape &original_input_shape, DataLayout data_layout);
+    /** Static function to check if given info will lead to a valid configuration of @ref NEConvertFullyConnectedWeightsKernel
+     *
+     * @param[in] input                Source weights tensor info to convert. Must be 2 dimensional. Data types supported: U8/S8/QS8/QASYMM8/U16/S16/QS16/U32/S32/QS32/F16/F32.
+     * @param[in] output               The converted weights tensor info. Shape and Data Type: Same as @p input.
+     * @param[in] original_input_shape Shape of the original input tensor (the one entering fully connected layer). Must be in NCHW format.
+     * @param[in] data_layout          The data layout the weights have been trained in.
+     */
+    static Status validate(const ITensorInfo *input, const ITensorInfo *output, const TensorShape &original_input_shape, DataLayout data_layout);
+
+    // Inherited methods overridden:
+    void run(const Window &window, const ThreadInfo &info) override;
+
+private:
+    /** Template function to run the permute
+     *
+     * @param[in] window Region on which to execute the kernel. (Must be a valid region of the window returned by window()).
+     */
+    template <typename T>
+    void run_convert_fc_weights(const Window &window);
+
+    const ITensor *_input;
+    ITensor       *_output;
+    unsigned int   _factor1; /*  equals to the number of elements per original input plane if @p data_layout == NCHW; its number of channels otherwise */
+    unsigned int   _factor2; /*  equals to the number of elements per original input plane if @p data_layout == NHWC; its number of channels otherwise */
+};
+} // namespace arm_compute
+#endif /*__ARM_COMPUTE_NECONVERTFULLYCONNECTEDWEIGHTSKERNEL_H__ */
diff --git a/arm_compute/runtime/CL/CLFunctions.h b/arm_compute/runtime/CL/CLFunctions.h
index 53e49936df..0d9407bfe4 100644
--- a/arm_compute/runtime/CL/CLFunctions.h
+++ b/arm_compute/runtime/CL/CLFunctions.h
@@ -41,6 +41,7 @@
 #include "arm_compute/runtime/CL/functions/CLChannelExtract.h"
 #include "arm_compute/runtime/CL/functions/CLChannelShuffleLayer.h"
 #include "arm_compute/runtime/CL/functions/CLColorConvert.h"
+#include "arm_compute/runtime/CL/functions/CLConvertFullyConnectedWeights.h"
 #include "arm_compute/runtime/CL/functions/CLConvolution.h"
 #include "arm_compute/runtime/CL/functions/CLConvolutionLayer.h"
 #include "arm_compute/runtime/CL/functions/CLCopy.h"
diff --git a/arm_compute/runtime/CL/functions/CLConvertFullyConnectedWeights.h b/arm_compute/runtime/CL/functions/CLConvertFullyConnectedWeights.h
new file mode 100644
index 0000000000..d8eef34c62
--- /dev/null
+++ b/arm_compute/runtime/CL/functions/CLConvertFullyConnectedWeights.h
@@ -0,0 +1,56 @@
+/*
+ * 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_CLCONVERTFULLYCONNECTEDWEIGHTS_H__
+#define __ARM_COMPUTE_CLCONVERTFULLYCONNECTEDWEIGHTS_H__
+
+#include "arm_compute/core/CL/kernels/CLConvertFullyConnectedWeightsKernel.h"
+#include "arm_compute/runtime/CL/ICLSimpleFunction.h"
+
+namespace arm_compute
+{
+class ICLTensor;
+
+/** Basic function to run @ref CLConvertFullyConnectedWeightsKernel. */
+class CLConvertFullyConnectedWeights : public ICLSimpleFunction
+{
+public:
+    /** Initialize the function.
+     *
+     * @param[in]  input                Source weights tensor to convert. Must be 2 dimensional. Data types supported: U8/S8/QS8/QASYMM8/U16/S16/QS16/U32/S32/QS32/F16/F32.
+     * @param[out] output               The converted weights tensor. Shape and Data Type: Same as @p input.
+     * @param[in]  original_input_shape Shape of the original input tensor (the one entering fully connected layer). Must be in NCHW format.
+     * @param[in]  data_layout          The data layout the weights have been trained in.
+     */
+    void configure(const ICLTensor *input, ICLTensor *output, const TensorShape &original_input_shape, DataLayout data_layout);
+    /** Static function to check if given info will lead to a valid configuration of @ref CLConvertFullyConnectedWeights
+     *
+     * @param[in] input                Source weights tensor info to convert. Must be 2 dimensional. Data types supported: U8/S8/QS8/QASYMM8/U16/S16/QS16/U32/S32/QS32/F16/F32.
+     * @param[in] output               The converted weights tensor info. Shape and Data Type: Same as @p input.
+     * @param[in] original_input_shape Shape of the original input tensor (the one entering fully connected layer). Must be in NCHW format.
+     * @param[in] data_layout          The data layout the weights have been trained in.
+     */
+    static Status validate(const ITensorInfo *input, const ITensorInfo *output, const TensorShape &original_input_shape, DataLayout data_layout);
+};
+} // namespace arm_compute
+#endif /* __ARM_COMPUTE_CLCONVERTFULLYCONNECTEDWEIGHTS_H__ */
diff --git a/arm_compute/runtime/NEON/NEFunctions.h b/arm_compute/runtime/NEON/NEFunctions.h
index 8091acd258..bd4097c6d3 100644
--- a/arm_compute/runtime/NEON/NEFunctions.h
+++ b/arm_compute/runtime/NEON/NEFunctions.h
@@ -41,6 +41,7 @@
 #include "arm_compute/runtime/NEON/functions/NEChannelExtract.h"
 #include "arm_compute/runtime/NEON/functions/NECol2Im.h"
 #include "arm_compute/runtime/NEON/functions/NEColorConvert.h"
+#include "arm_compute/runtime/NEON/functions/NEConvertFullyConnectedWeights.h"
 #include "arm_compute/runtime/NEON/functions/NEConvolution.h"
 #include "arm_compute/runtime/NEON/functions/NEConvolutionLayer.h"
 #include "arm_compute/runtime/NEON/functions/NEDeconvolutionLayer.h"
@@ -111,4 +112,4 @@
 #include "arm_compute/runtime/NEON/functions/NEWarpPerspective.h"
 #include "arm_compute/runtime/NEON/functions/NEWinogradConvolutionLayer.h"
 
-#endif /* __ARM_COMPUTE_NEFUNCTIONS_H__ */
+#endif /* __ARM_COMPUTE_NEFUNCTIONS_H__ */
\ No newline at end of file
diff --git a/arm_compute/runtime/NEON/functions/NEConvertFullyConnectedWeights.h b/arm_compute/runtime/NEON/functions/NEConvertFullyConnectedWeights.h
new file mode 100644
index 0000000000..bdb157f30b
--- /dev/null
+++ b/arm_compute/runtime/NEON/functions/NEConvertFullyConnectedWeights.h
@@ -0,0 +1,65 @@
+/*
+ * 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_NECONVERTFULLYCONNECTEDWEIGHTS_H__
+#define __ARM_COMPUTE_NECONVERTFULLYCONNECTEDWEIGHTS_H__
+
+#include "arm_compute/core/NEON/kernels/NEConvertFullyConnectedWeightsKernel.h"
+#include "arm_compute/runtime/IFunction.h"
+#include "arm_compute/runtime/NEON/NEScheduler.h"
+
+namespace arm_compute
+{
+class ITensor;
+
+/** Basic function to run @ref NEConvertFullyConnectedWeightsKernel. */
+class NEConvertFullyConnectedWeights : public IFunction
+{
+public:
+    /** Default constructor */
+    NEConvertFullyConnectedWeights();
+    /** Initialize the function.
+     *
+     * @param[in]  input                Source weights tensor to convert. Must be 2 dimensional. Data types supported: U8/S8/QS8/QASYMM8/U16/S16/QS16/U32/S32/QS32/F16/F32.
+     * @param[out] output               The converted weights tensor. Shape and Data Type: Same as @p input.
+     * @param[in]  original_input_shape Shape of the original input tensor (the one entering fully connected layer). Must be in NCHW format.
+     * @param[in]  data_layout          The data layout the weights have been trained in.
+     */
+    void configure(const ITensor *input, ITensor *output, const TensorShape &original_input_shape, DataLayout data_layout);
+    /** Static function to check if given info will lead to a valid configuration of @ref NEConvertFullyConnectedWeights
+     *
+     * @param[in] input                Source weights tensor info to convert. Must be 2 dimensional. Data types supported: U8/S8/QS8/QASYMM8/U16/S16/QS16/U32/S32/QS32/F16/F32.
+     * @param[in] output               The converted weights tensor info. Shape and Data Type: Same as @p input.
+     * @param[in] original_input_shape Shape of the original input tensor (the one entering fully connected layer). Must be in NCHW format.
+     * @param[in] data_layout          The data layout the weights have been trained in.
+     */
+    static Status validate(const ITensorInfo *input, const ITensorInfo *output, const TensorShape &original_input_shape, DataLayout data_layout);
+
+    // Inherited methods overriden:
+    void run() override;
+
+private:
+    NEConvertFullyConnectedWeightsKernel _kernel;
+};
+}
+#endif /* __ARM_COMPUTE_NECONVERTFULLYCONNECTEDWEIGHTS_H__ */