Port NEGEMMConvolutionLayer

Details:
port NEWeightsReshapeKernel to CpuWeightsReshapeKernel
port NEGEMMConvolutionLayer to CpuGEMMConvolutionLayer

Resolves: COMPMID-4509

Change-Id: I3c7051e2c3f6d808a7ccb898aad70e5b221b9dc3
Signed-off-by: Manuel Bottini <manuel.bottini@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5938
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>

1 files changed, 7 insertions, 194 deletions

diff --git a/arm_compute/runtime/NEON/functions/NEGEMMConvolutionLayer.h b/arm_compute/runtime/NEON/functions/NEGEMMConvolutionLayer.h
index e3b7d91187..2ebb80bef2 100644
--- a/arm_compute/runtime/NEON/functions/NEGEMMConvolutionLayer.h
+++ b/arm_compute/runtime/NEON/functions/NEGEMMConvolutionLayer.h
@@ -27,143 +27,21 @@
 #include "arm_compute/runtime/IFunction.h"
 
 #include "arm_compute/core/Types.h"
+#include "arm_compute/runtime/IFunction.h"
+#include "arm_compute/runtime/IMemoryManager.h"
 #include "arm_compute/runtime/IWeightsManager.h"
 #include "arm_compute/runtime/MemoryGroup.h"
-#include "arm_compute/runtime/NEON/functions/NEGEMM.h"
-#include "arm_compute/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.h"
-#include "arm_compute/runtime/NEON/functions/NEGEMMLowpOutputStage.h"
-#include "arm_compute/runtime/NEON/functions/NEReshapeLayer.h"
-#include "arm_compute/runtime/Tensor.h"
 
 #include <memory>
 
 namespace arm_compute
 {
 class ITensor;
-class NEWeightsReshapeKernel;
-namespace cpu
-{
-namespace kernels
-{
-class CpuIm2ColKernel;
-class CpuCol2ImKernel;
-} // namespace kernels
-} // namespace cpu
-
-/** Function to reshape the weights. This function calls the following kernel:
- * -# @ref NEWeightsReshapeKernel
- */
-class NEConvolutionLayerReshapeWeights : public IFunction
-{
-public:
-    /** Constructor */
-    NEConvolutionLayerReshapeWeights() noexcept;
-    /** Prevent instances of this class from being copied (As this class contains pointers) */
-    NEConvolutionLayerReshapeWeights(const NEConvolutionLayerReshapeWeights &) = delete;
-    /** Prevent instances of this class from being moved (As this class contains non movable objects) */
-    NEConvolutionLayerReshapeWeights(NEConvolutionLayerReshapeWeights &&) = delete;
-    /** Prevent instances of this class from being copied (As this class contains pointers) */
-    NEConvolutionLayerReshapeWeights &operator=(const NEConvolutionLayerReshapeWeights &) = delete;
-    /** Prevent instances of this class from being moved (As this class contains non movable objects) */
-    NEConvolutionLayerReshapeWeights &operator=(NEConvolutionLayerReshapeWeights &&) = delete;
-    /** Default destructor */
-    ~NEConvolutionLayerReshapeWeights();
-    /** 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: All.
-     * @param[in]  biases  Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM].
-     *                     Data type supported: same as @p weights.
-     *                     @warning Appending biases to weights reshaped matrix is not supported for quantized asymmetric types.
-     * @param[out] output  Destination tensor. Data types supported: same as @p weights.
-     */
-    void configure(const ITensor *weights, const ITensor *biases, ITensor *output);
-    /** Static function to check if given info will lead to a valid configuration of @ref NEConvolutionLayerReshapeWeights
-     *
-     * @param[in] weights Weights tensor info. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM].
-     *                    Data type supported: All.
-     * @param[in] biases  Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM].
-     *                    Data type supported: same as @p weights.
-     *                    @warning Appending biases to weights reshaped matrix is not supported for quantized asymmetric types.
-     * @param[in] output  Destination tensor. Data types supported: same as @p weights.
-     *
-     * @return an error status
-     */
-    static Status validate(const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output);
-
-    // Inherited methods overridden:
-    void run() override;
-
-private:
-    std::unique_ptr<NEWeightsReshapeKernel> _weights_reshape_kernel;
-};
-
-namespace weights_transformations
-{
-/** Basic function to manage the reshape weights generated from @ref NEConvolutionLayerReshapeWeights */
-class NEConvolutionLayerReshapeWeightsTransform : public ITransformWeights
-{
-public:
-    /** Constructor */
-    NEConvolutionLayerReshapeWeightsTransform() = default;
-    /** Prevent instances of this class from being copied (As this class contains pointers) */
-    NEConvolutionLayerReshapeWeightsTransform(const NEConvolutionLayerReshapeWeightsTransform &) = delete;
-    /** Prevent instances of this class from being copied (As this class contains pointers) */
-    NEConvolutionLayerReshapeWeightsTransform &operator=(const NEConvolutionLayerReshapeWeightsTransform &) = delete;
-    /** Prevent instances of this class from being moved (As this class contains non movable objects) */
-    NEConvolutionLayerReshapeWeightsTransform(NEConvolutionLayerReshapeWeightsTransform &&) = delete;
-    /** Prevent instances of this class from being moved (As this class contains non movable objects) */
-    NEConvolutionLayerReshapeWeightsTransform &operator=(NEConvolutionLayerReshapeWeightsTransform &&) = delete;
-    /** Default destructor */
-    ~NEConvolutionLayerReshapeWeightsTransform() = default;
-    void configure(const ITensor *input, const ITensor *biases)
-    {
-        _bias_bit = (biases != nullptr) ? 1 : 0;
-        _func.configure(input, biases, &_output);
-    }
-
-    void run() override
-    {
-        _output.allocator()->allocate();
-        _func.run();
-        _reshape_run = true;
-    }
-
-    ITensor *get_weights() override
-    {
-        return &_output;
-    }
-
-    void release() override
-    {
-        _output.allocator()->free();
-    }
-
-    uint32_t uid() override
-    {
-        return ((0x8) | (_bias_bit << 7));
-    }
-
-    bool is_reshape_run()
-    {
-        return _reshape_run;
-    }
-
-private:
-    Tensor                           _output{};
-    NEConvolutionLayerReshapeWeights _func{};
-    int32_t                          _bias_bit{ 0 };
-};
-} // namespace weights_transformations
+class ITensorInfo;
 
 /** Basic function to compute the convolution layer. This function calls the following kernels/functions:
  *
- * -# @ref cpu::kernels::CpuIm2ColKernel
- * -# @ref NEGEMM (if the data type is BFLOAT16/FP16/FP32)
- * -# @ref NEGEMMLowpMatrixMultiplyCore (if the data type is QASYMM8/QASYMM8_SIGNED)
- * -# @ref NEGEMMLowpOutputStage (if the data type is QASYMM8/QASYMM8_SIGNED)
- * -# @ref NEArithmeticAddition (if biases != nullptr and we have a 1x1 convolution with the NHWC data layout)
- * -# @ref cpu::kernels::CpuCol2ImKernel (if NCHW data layout)
+ * -# @ref cpu::CpuGemmConvolution
  *
  */
 class NEGEMMConvolutionLayer : public IFunction
@@ -244,73 +122,8 @@ public:
     void prepare() override;
 
 private:
-    /** Configures the appropriate matrix multiply routine
-     *
-     * @param[in]  input         Input tensor. Data types supported: QASYMM8/QASYMM8_SIGNED/BFLOAT16/F16/F32.
-     * @param[in]  weights       Weights tensor. Data type supported: QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL/BFLOAT16/F16/F32.
-     * @param[in]  biases        Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM].
-     *                           Data type supported: Should match @p input data type, except for input of QASYMM8/QASYMM8_SIGNED type where biases should be of S32 type.
-     * @param[out] output        Output tensor. Data types supported: Same as @p input,
-     *                           except for input of QASYMM8/QASYMM8_SIGNED type where output should be of S32 type.
-     * @param[in]  act_info      (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU supported.
-     * @param[in]  gemm_3d_depth (Optional) Depth of GEMM 3D (Defaults to 1)
-     */
-    void configure_mm(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const ActivationLayerInfo &act_info = ActivationLayerInfo(), int gemm_3d_depth = 1);
-    /** Static function to check if given info will lead to a valid configuration of @ref NEGEMMConvolutionLayer matrix multiply routines
-     *
-     * @param[in] input         Input tensor info. Data types supported: QASYMM8/QASYMM8_SIGNED/BFLOAT16/F16/F32.
-     * @param[in] weights       Weights tensor info. Data type supported: QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL/BFLOAT16/F16/F32.
-     * @param[in] biases        Biases tensor info. Shared biases supported. Biases are 1D tensor with dimensions [OFM].
-     *                          Data type supported: Should match @p input data type, except for input of QASYMM8/QASYMM8_SIGNED type where biases should be of S32 type.
-     * @param[in] output        Output tensor info. Data types supported: Same as @p input,
-     *                          except for input of QASYMM8/QASYMM8_SIGNED type where output should be of S32 type.
-     * @param[in] act_info      (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU supported.
-     * @param[in] gemm_3d_depth (Optional) Depth of GEMM 3D (Defaults to 1)
-     * @param[in] skip_im2col   (Optional) Flag which specifies if im2col has to be skipped. i.e. 1x1 convolution with NHWC data layout. (Default to false)
-     *
-     * @return a status
-     */
-    static Status validate_mm(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const ActivationLayerInfo &act_info = ActivationLayerInfo(),
-                              int gemm_3d_depth = 1, bool skip_im2col = false);
-    /** Static function to check if GEMM3D is supported in @ref NEGEMM or in @ref NEGEMMLowpMatrixMultiplyCore
-     *
-     * @param[in] input_info    Input tensor info. Data types supported: QASYMM8/QASYMM8_SIGNED/BFLOAT16/F16/F32.
-     * @param[in] weights_info  Weights tensor info. Data types supported: QASYMM8/QASYMM8_SIGNED/BFLOAT16/F16/F32.
-     * @param[in] act_info      Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU supported.
-     * @param[in] gemm_3d_depth Depth of GEMM 3D
-     * @param[in] skip_im2col   Flag which specifies if im2col has to be skipped. i.e. 1x1 convolution with NHWC data layout
-     *
-     * @return a status
-     */
-    static Status validate_gemm3d(const ITensorInfo *input_info, const ITensorInfo *weights_info, const ActivationLayerInfo &act_info, int gemm_3d_depth, bool skip_im2col);
-
-private:
-    MemoryGroup                                                        _memory_group;
-    IWeightsManager                                                   *_weights_manager;
-    NEConvolutionLayerReshapeWeights                                   _reshape_weights;
-    weights_transformations::NEConvolutionLayerReshapeWeightsTransform _reshape_weights_managed;
-    std::unique_ptr<cpu::kernels::CpuIm2ColKernel>                     _im2col_kernel;
-    NEGEMM                                                             _mm_gemm;
-    NEGEMMLowpMatrixMultiplyCore                                       _mm_gemmlowp;
-    std::unique_ptr<cpu::kernels::CpuCol2ImKernel>                     _col2im_kernel;
-    NEReshapeLayer                                                     _reshape_layer;
-
-    const ITensor *_input;
-    const ITensor *_original_weights;
-    ITensor       *_original_output;
-
-    Tensor _im2col_output;
-    Tensor _weights_reshaped;
-    Tensor _gemm_output;
-    Tensor _gemm_output_3d;
-    Tensor _tmp_output;
-
-    DataLayout _data_layout;
-
-    bool _skip_im2col;
-    bool _skip_col2im;
-    bool _is_quantized;
-    bool _is_prepared;
+    struct Impl;
+    std::unique_ptr<Impl> _impl;
 };
 } // namespace arm_compute
-#endif /* ARM_COMPUTE_NECONVOLUTIONGEMMLAYER_H */
+#endif /* ARM_COMPUTE_NEGEMMCONVOLUTIONLAYER_H */