COMPMID-1694: Fuse offset contribution with the output stage when we use NEGEMMLowpMatrixMultiplyCore

Change-Id: Ic1a681e4cc03e1eba3bf8485d9cdb17b3e926047
Signed-off-by: giuros01 <giuseppe.rossini@arm.com>
Reviewed-on: https://review.mlplatform.org/c/561
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

3 files changed, 57 insertions, 42 deletions

diff --git a/arm_compute/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.h b/arm_compute/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.h
index 4345ff267b..67b22821da 100644
--- a/arm_compute/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.h
+++ b/arm_compute/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.h
@@ -71,7 +71,7 @@ public:
      *  This kernel performs the following computations:
      *
      *  -# Convert a values from QASYMM8 to int32 and add a_offset to each of them.
-     *  -# Convert b values from QASYMM8 to int32 add b_offset to each of them.
+     *  -# Convert b values from QASYMM8 to int32 and add b_offset to each of them.
      *  -# Compute the matrix product of the resulting a * b in int32.
      *  -# Quantize to uint8 if gemm_info.gemmlowp_output_stage != NONE
      *
diff --git a/arm_compute/runtime/NEON/functions/NEGEMMConvolutionLayer.h b/arm_compute/runtime/NEON/functions/NEGEMMConvolutionLayer.h
index 6df7af0d86..ace924f146 100644
--- a/arm_compute/runtime/NEON/functions/NEGEMMConvolutionLayer.h
+++ b/arm_compute/runtime/NEON/functions/NEGEMMConvolutionLayer.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2018 ARM Limited.
+ * Copyright (c) 2017-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -151,44 +151,51 @@ private:
      *
      * @param[in]  input         Input tensor. Data types supported: QASYMM8/F16/F32.
      * @param[in]  weights       Weights tensor. 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: Should match @p input data type, except for input of QASYMM8 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 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, ITensor *output, int gemm_3d_depth = 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. Data types supported: QASYMM8/F16/F32.
      * @param[in] weights       Weights tensor. 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: Should match @p input data type, except for input of QASYMM8 type where biases should be of S32 type.
      * @param[in] output        Output tensor. Data types supported: Same as @p input,
      *                          except for input of QASYMM8 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 *output, int gemm_3d_depth = 1, bool skip_im2col = false);
+    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] data_type     Input data type
+     * @param[in] input_info    Input tensor info. Data types supported: QASYMM8/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(DataType data_type, int gemm_3d_depth, bool skip_im2col);
+    static Status validate_gemm3d(const ITensorInfo *input_info, const ActivationLayerInfo &act_info, int gemm_3d_depth, bool skip_im2col);
 
 private:
-    MemoryGroup                                         _memory_group;
-    NEConvolutionLayerReshapeWeights                    _reshape_weights;
-    NEIm2ColKernel                                      _im2col_kernel;
-    NEGEMM                                              _mm_gemm;
-    NEGEMMLowpMatrixMultiplyCore                        _mm_gemmlowp;
-    NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint _gemmlowp_output_stage;
-    NECol2ImKernel                                      _col2im_kernel;
-    NEActivationLayer                                   _activationlayer_function;
-    NEArithmeticAdditionKernel                          _add_bias_kernel;
-    NEReshapeLayer                                      _reshape_layer;
+    MemoryGroup                      _memory_group;
+    NEConvolutionLayerReshapeWeights _reshape_weights;
+    NEIm2ColKernel                   _im2col_kernel;
+    NEGEMM                           _mm_gemm;
+    NEGEMMLowpMatrixMultiplyCore     _mm_gemmlowp;
+    NECol2ImKernel                   _col2im_kernel;
+    NEActivationLayer                _activationlayer_function;
+    NEArithmeticAdditionKernel       _add_bias_kernel;
+    NEReshapeLayer                   _reshape_layer;
 
     const ITensor *_original_weights;
 
diff --git a/arm_compute/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.h b/arm_compute/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.h
index 682475c824..d3b27e4faf 100644
--- a/arm_compute/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.h
+++ b/arm_compute/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2018 ARM Limited.
+ * Copyright (c) 2017-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -26,6 +26,7 @@
 
 #include "arm_compute/core/NEON/INEKernel.h"
 #include "arm_compute/core/NEON/kernels/NEGEMMLowpOffsetContributionKernel.h"
+#include "arm_compute/core/NEON/kernels/NEGEMMLowpOffsetContributionOutputStageKernel.h"
 #include "arm_compute/core/NEON/kernels/NEGEMMLowpReductionKernel.h"
 #include "arm_compute/runtime/IFunction.h"
 #include "arm_compute/runtime/IMemoryManager.h"
@@ -73,20 +74,24 @@ public:
      *  -# Convert b values from QASYMM8 to int32 add b_offset to each of them.
      *  -# Compute the matrix product of the resulting a * b in int32.
      *
+     * @note The @p output type is S32 if @p gemm_info.type == GEMMLowpOutputStageType::NONE. It is QASYMM8 otherwise
+     *
      * @param[in]  a         First input tensor  (Matrix A). Data type supported: QASYMM8.
      * @param[in]  b         Second input tensor (Matrix B). Data type supported: same as @p a
      * @param[in]  c         Third input tensor  (Matrix C). It can be a nullptr. Data type supported: S32
-     * @param[out] output    Output tensor. Data type supported: Data type supported: S32
+     * @param[out] output    Output tensor. Data type supported: Data type supported: S32/QASYMM8
      * @param[in]  gemm_info (Optional) Specifies if the matrix A and/or matrix B have been reshaped and
      *                       if the reshape of matrix B should be executed only for the first run
      */
     void configure(const ITensor *a, const ITensor *b, const ITensor *c, ITensor *output, const GEMMInfo &gemm_info = GEMMInfo());
     /** Static function to check if given info will lead to a valid configuration of @ref NEGEMMLowpMatrixMultiplyCore
      *
-     * @param[in] a         First input tensor  (Matrix A). Data type supported: QASYMM8.
-     * @param[in] b         Second input tensor (Matrix B). Data type supported: same as @p a
-     * @param[in] c         Third input tensor  (Matrix C). It can be a nullptr. Data type supported: S32
-     * @param[in] output    Output tensor. Data type supported: Data type supported: S32
+     * @note The @p output type is S32 if @p gemm_info.type == GEMMLowpOutputStageType::NONE. It is QASYMM8 otherwise
+     *
+     * @param[in] a         First input tensor info  (Matrix A). Data type supported: QASYMM8.
+     * @param[in] b         Second input tensor info (Matrix B). Data type supported: same as @p a
+     * @param[in] c         Third input tensor  info (Matrix C). It can be a nullptr. Data type supported: S32
+     * @param[in] output    Output tensor info. Data type supported: Data type supported: S32/QASYMM8
      * @param[in] gemm_info (Optional) Specifies if the matrix A and/or matrix B have been reshaped and
      *                      if the reshape of matrix B should be executed only for the first run
      *
@@ -99,25 +104,28 @@ public:
     void prepare() override;
 
 private:
-    MemoryGroup                        _memory_group;
-    NEGEMMAssemblyDispatch             _asm_glue;
-    std::unique_ptr<INEKernel>         _mm_kernel;
-    std::unique_ptr<INEKernel>         _mtx_a_reshape_kernel;
-    std::unique_ptr<INEKernel>         _mtx_b_reshape_kernel;
-    NEGEMMLowpMatrixAReductionKernel   _mtx_a_reduction_kernel;
-    NEGEMMLowpMatrixBReductionKernel   _mtx_b_reduction_kernel;
-    NEGEMMLowpOffsetContributionKernel _offset_contribution_kernel;
-    Tensor                             _vector_sum_col;
-    Tensor                             _vector_sum_row;
-    Tensor                             _tmp_a;
-    Tensor                             _tmp_b;
-    const ITensor                     *_original_b;
-    int32_t                            _a_offset;
-    int32_t                            _b_offset;
-    bool                               _run_vector_matrix_multiplication;
-    bool                               _dot_product_path;
-    bool                               _reshape_b_only_on_first_run;
-    bool                               _is_prepared;
+    MemoryGroup                                   _memory_group;
+    NEGEMMAssemblyDispatch                        _asm_glue;
+    std::unique_ptr<INEKernel>                    _mm_kernel;
+    std::unique_ptr<INEKernel>                    _mtx_a_reshape_kernel;
+    std::unique_ptr<INEKernel>                    _mtx_b_reshape_kernel;
+    NEGEMMLowpMatrixAReductionKernel              _mtx_a_reduction_kernel;
+    NEGEMMLowpMatrixBReductionKernel              _mtx_b_reduction_kernel;
+    NEGEMMLowpOffsetContributionKernel            _offset_contribution_kernel;
+    NEGEMMLowpOffsetContributionOutputStageKernel _offset_contribution_output_stage_kernel;
+    Tensor                                        _vector_sum_col;
+    Tensor                                        _vector_sum_row;
+    Tensor                                        _tmp_a;
+    Tensor                                        _tmp_b;
+    Tensor                                        _mm_result_s32;
+    const ITensor                                *_original_b;
+    int32_t                                       _a_offset;
+    int32_t                                       _b_offset;
+    bool                                          _run_vector_matrix_multiplication;
+    bool                                          _dot_product_path;
+    bool                                          _reshape_b_only_on_first_run;
+    bool                                          _is_prepared;
+    bool                                          _fuse_output_stage;
 };
-}
+} // namespace arm_compute
 #endif /*__ARM_COMPUTE_NEGEMMLOWPMATRIXMULTIPLYCORE_H__ */