COMPMID-845: Create a ConvolutionLayer for CL

Change-Id: Ifcc406d2d0a99c911d6b6c875657b0e0028255d5
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/119148
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>

1 files changed, 41 insertions, 91 deletions

diff --git a/arm_compute/runtime/CL/functions/CLConvolutionLayer.h b/arm_compute/runtime/CL/functions/CLConvolutionLayer.h
index f6672cef1d..53d59c3176 100644
--- a/arm_compute/runtime/CL/functions/CLConvolutionLayer.h
+++ b/arm_compute/runtime/CL/functions/CLConvolutionLayer.h
@@ -26,71 +26,18 @@
 
 #include "arm_compute/runtime/IFunction.h"
 
-#include "arm_compute/core/CL/kernels/CLCol2ImKernel.h"
-#include "arm_compute/core/CL/kernels/CLFillBorderKernel.h"
-#include "arm_compute/core/CL/kernels/CLGEMMInterleave4x4Kernel.h"
-#include "arm_compute/core/CL/kernels/CLGEMMMatrixMultiplyKernel.h"
-#include "arm_compute/core/CL/kernels/CLGEMMTranspose1xWKernel.h"
-#include "arm_compute/core/CL/kernels/CLIm2ColKernel.h"
-#include "arm_compute/core/CL/kernels/CLWeightsReshapeKernel.h"
-#include "arm_compute/core/Types.h"
-#include "arm_compute/runtime/CL/CLMemoryGroup.h"
-#include "arm_compute/runtime/CL/CLTensor.h"
-#include "arm_compute/runtime/CL/functions/CLGEMM.h"
-#include "arm_compute/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.h"
-#include "arm_compute/runtime/CL/functions/CLGEMMLowpOutputStage.h"
+#include "arm_compute/runtime/CL/functions/CLDirectConvolutionLayer.h"
+#include "arm_compute/runtime/CL/functions/CLGEMMConvolutionLayer.h"
 #include "arm_compute/runtime/IMemoryManager.h"
 
 #include <memory>
 
 namespace arm_compute
 {
-class ICLTensor;
-
-/** Function to reshape and transpose the weights. This function calls the following kernels:
- * -# @ref CLWeightsReshapeKernel
- * -# @ref CLGEMMTranspose1xWKernel
- */
-class CLConvolutionLayerReshapeWeights : public IFunction
-{
-public:
-    /** Constructor */
-    CLConvolutionLayerReshapeWeights(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
-    /** Set the input and output tensors.
-     *
-     * @param[in]  weights      Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM].
-     *                          Data type supported: QS8/QASYMM8/QS16/F16/F32.
-     * @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. Data types supported: Same as @p weights.
-     * @param[in]  transpose1xW True if the weights are to undergo a 1xW transposition after reshaping (in case of GEMM operation), false otherwise.
-     *                          Data types supported: Same as @p weights.
-     */
-    void configure(const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, bool transpose1xW);
-    // Inherited methods overridden:
-    void run() override;
-
-private:
-    CLMemoryGroup            _memory_group;
-    CLWeightsReshapeKernel   _weights_reshape_kernel;
-    CLGEMMTranspose1xWKernel _weights_transposed_kernel;
-    CLTensor                 _weights_reshaped;
-    bool                     _transpose1xW;
-};
-
 /** Basic function to compute the convolution layer. This function calls the following OpenCL kernels/functions:
  *
- * Note: weights already reshaped for quantized asymmetric is not supported
- *
- * -# @ref CLIm2ColKernel
- * -# @ref CLGEMMLowpMatrixMultiplyCore (if quantized asymmetric)
- * -# @ref CLGEMMLowpQuantizeDownInt32ToUint8Scale (if quantized asymmetric)
- * -# @ref CLCol2ImKernel
- *
- * if the weights are already reshaped:
- * -# @ref CLGEMMInterleave4x4Kernel
- * -# @ref CLGEMMMatrixMultiplyKernel
- * else
- * -# @ref CLGEMM
+ * -# @ref CLGEMMConvolutionLayer
+ * -# @ref CLDirectConvolutionLayer
  */
 class CLConvolutionLayer : public IFunction
 {
@@ -108,46 +55,49 @@ public:
      * @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]  conv_info    Contains padding and stride information described in @ref PadStrideInfo.
-     * @param[in]  weights_info Specifies if the weights tensor has been reshaped with CLWeightsReshapeKernel. If this is not part of the fully connected layer the weights
-     *                          tensor has also been transposed with CLGEMMTranspose1xWKernel. Data type supported: Same as @p input.
+     * @param[in]  weights_info Specifies if the weights tensor has been reshaped with CLWeightsReshapeKernel. Data type supported: Same as @p input.
+     */
+    void configure(ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info, const WeightsInfo &weights_info = WeightsInfo());
+    /** Static function to check if given info will lead to a valid configuration of @ref CLConvolutionLayer
+     *
+     * @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: QS8/QASYMM8/QS16/F16/F32.
+     * @param[in] weights      Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, 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 input.
+     * @param[in] 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] conv_info    Contains padding and stride information described in @ref PadStrideInfo.
+     * @param[in] weights_info Specifies if the weights tensor has been reshaped with CLWeightsReshapeKernel. Data type supported: Same as @p input.
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info,
+                           const WeightsInfo &weights_info = WeightsInfo());
+    /** Static function to check if given info will return the convolution called by @ref CLConvolutionLayer
+     *
+     * @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: QS8/QASYMM8/QS16/F16/F32.
+     * @param[in] weights      Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, 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 input.
+     * @param[in] 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] conv_info    Contains padding and stride information described in @ref PadStrideInfo.
+     * @param[in] weights_info Specifies if the weights tensor has been reshaped with CLWeightsReshapeKernel. Data type supported: Same as @p input.
+     * @param[in] gpu_target   Specifies the @p GPUTarget.
+     *
+     * @return a status
      */
-    void configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info, const WeightsInfo &weights_info = WeightsInfo());
+    static ConvolutionMethod get_convolution_method(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info,
+                                                    const WeightsInfo &weights_info, const GPUTarget gpu_target);
 
     // Inherited methods overridden:
     void run() override;
 
 private:
-    /** Configures the appropriate matrix multiply routine
-     *
-     * @param input                     Input tensor. Data types supported: QS8/QASYMM8/QS16/F16/F32.
-     * @param weights                   Weights tensor. Data type supported: Same as @p input.
-     * @param output                    Output tensor. Data types supported: Same as @p input,
-     *                                                 except for input of QASYMM8 type where output should be of S32 type.
-     * @param is_interleaved_transposed Flag that signals if matrix is interleaved transposed
-     */
-    void configure_mm(const ICLTensor *input, const ICLTensor *weights, ICLTensor *output, bool is_interleaved_transposed, bool are_weights_reshaped);
-
-private:
-    CLMemoryGroup                                       _memory_group;
-    CLConvolutionLayerReshapeWeights                    _reshape_weights;
-    CLIm2ColKernel                                      _im2col_kernel;
-    CLGEMMInterleave4x4Kernel                           _interleave_kernel;
-    CLGEMMMatrixMultiplyKernel                          _mm_kernel;
-    CLGEMM                                              _mm_gemm;
-    CLGEMMLowpMatrixMultiplyCore                        _mm_gemmlowp;
-    CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint _gemmlowp_output_stage;
-    CLCol2ImKernel                                      _col2im_kernel;
-
-    CLTensor _im2col_output;
-    CLTensor _interleave_output;
-    CLTensor _weights_reshaped;
-    CLTensor _weights_transposed;
-    CLTensor _gemm_output;
-    CLTensor _tmp_output;
-
-    bool _are_weights_reshaped;
-    bool _is_quantized;
-    bool _is_interleaved_transposed;
+    std::shared_ptr<IMemoryManager> _memory_manager;
+    std::unique_ptr<IFunction>      _function; /**< Function to run */
 };
 }
 #endif /* __ARM_COMPUTE_CLCONVOLUTIONLAYER_H__ */