COMPMID-1979: Fuse Activation Function in CLGEMM - part 4

Fused activation function in CLGEMM

Change-Id: I644fdf09349325c0b3a2cd5fef2a3ea2c974149d
Signed-off-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Reviewed-on: https://review.mlplatform.org/c/1640
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

3 files changed, 29 insertions, 21 deletions

diff --git a/arm_compute/core/Types.h b/arm_compute/core/Types.h
index b4d94eced4..2c17f273a5 100644
--- a/arm_compute/core/Types.h
+++ b/arm_compute/core/Types.h
@@ -1775,7 +1775,8 @@ public:
           _gemmlowp_output_stage(),
           _fp_mixed_precision(false),
           _broadcast_bias(false),
-          _pretranpose_B(true)
+          _pretranpose_B(true),
+          _activation_info()
     {
     }
     /** Constructor
@@ -1791,9 +1792,11 @@ public:
      * @param[in] gemmlowp_output_stage       (Optional) GEMMLowp Output stage info
      * @param[in] fp_mixed_precision          (Optional) Use wider accumulators (32 bit instead of 16 for FP16) to improve accuracy.
      * @param[in] broadcast_bias              (Optional) Broadcast the shape of the bias tensor from a vector to a matrix.
+     * @param[in] activation_info             (Optional) Activation to apply after the matrix multiplication
      */
     GEMMInfo(bool is_a_reshaped, bool is_b_reshaped, bool reshape_b_only_on_first_run, int depth_output_gemm3d = 0, bool reinterpret_input_as_3d = false, bool retain_internal_weights = false,
-             GEMMLowpOutputStageInfo gemmlowp_output_stage = GEMMLowpOutputStageInfo(), bool fp_mixed_precision = false, bool broadcast_bias = false) noexcept
+             GEMMLowpOutputStageInfo gemmlowp_output_stage = GEMMLowpOutputStageInfo(), bool fp_mixed_precision = false, bool broadcast_bias = false,
+             const ActivationLayerInfo &activation_info = ActivationLayerInfo()) noexcept
         : _is_a_reshaped(is_a_reshaped),
           _is_b_reshaped(is_b_reshaped),
           _reshape_b_only_on_first_run(reshape_b_only_on_first_run),
@@ -1803,7 +1806,8 @@ public:
           _gemmlowp_output_stage(gemmlowp_output_stage),
           _fp_mixed_precision(fp_mixed_precision),
           _broadcast_bias(broadcast_bias),
-          _pretranpose_B(reshape_b_only_on_first_run)
+          _pretranpose_B(reshape_b_only_on_first_run),
+          _activation_info(activation_info)
     {
     }
     /** Flag which specifies if the matrix A has been reshaped
@@ -1896,6 +1900,14 @@ public:
     {
         _pretranpose_B = flag;
     }
+    /** Activation layer to apply after the matrix multiplication
+     *
+     * @return ActivationLayerInfo object
+     */
+    ActivationLayerInfo activation_info() const
+    {
+        return _activation_info;
+    }
 
 private:
     bool                    _is_a_reshaped;
@@ -1908,6 +1920,7 @@ private:
     bool                    _fp_mixed_precision;
     bool                    _broadcast_bias;
     bool                    _pretranpose_B;
+    ActivationLayerInfo     _activation_info;
 };
 
 /** Winograd information */
diff --git a/arm_compute/runtime/CL/functions/CLGEMM.h b/arm_compute/runtime/CL/functions/CLGEMM.h
index 8c462fa4cb..e2a92a8a37 100644
--- a/arm_compute/runtime/CL/functions/CLGEMM.h
+++ b/arm_compute/runtime/CL/functions/CLGEMM.h
@@ -127,7 +127,6 @@ private:
 
     CLMemoryGroup                             _memory_group;
     CLGEMMMatrixMultiplyKernel                _mm_kernel;
-    CLGEMMMatrixAdditionKernel                _ma_kernel;
     CLGEMMReshapeLHSMatrixKernel              _reshape_lhs_kernel;
     CLGEMMReshapeRHSMatrixKernel              _reshape_rhs_kernel;
     CLGEMMMatrixMultiplyReshapedKernel        _mm_reshaped_kernel;
@@ -135,7 +134,6 @@ private:
     CLTensor                                  _tmp_a;
     CLTensor                                  _tmp_b;
     const ICLTensor                          *_original_b;
-    bool                                      _run_addition;
     bool                                      _reshape_b_only_on_first_run;
     bool                                      _is_prepared;
     GEMMType                                  _gemm_type;
diff --git a/arm_compute/runtime/CL/functions/CLGEMMConvolutionLayer.h b/arm_compute/runtime/CL/functions/CLGEMMConvolutionLayer.h
index e9a3f9bf2b..027727c7f7 100644
--- a/arm_compute/runtime/CL/functions/CLGEMMConvolutionLayer.h
+++ b/arm_compute/runtime/CL/functions/CLGEMMConvolutionLayer.h
@@ -163,8 +163,10 @@ private:
      *                                       except for input of QASYMM8 type where output should be of S32 type.
      * @param[in]      gemmlowp_output_stage GEMMLowp output stage info
      * @param[in]      gemm_3d_depth         Depth of GEMM 3D
+     * @param[in]      act_info              Activation to apply after the matrix multiplication
      */
-    void configure_mm(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const GEMMLowpOutputStageInfo &gemmlowp_output_stage, int gemm_3d_depth = 1);
+    void configure_mm(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const GEMMLowpOutputStageInfo &gemmlowp_output_stage, int gemm_3d_depth,
+                      const ActivationLayerInfo &act_info);
     /** Static function to check if given info will lead to a valid configuration of @ref CLGEMMConvolutionLayer matrix multiply routines
      *
      * @param[in] input                 Input tensor. Data types supported: QASYMM8/F16/F32.
@@ -176,22 +178,21 @@ private:
      * @param[in] gemmlowp_output_stage GEMMLowp output stage info
      * @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.
-     * @param[in] run_addition          Flag which specifies if @ref CLGEMMMatrixMatrixMultiplyAddition to be run.
+     * @param[in] act_info              Activation to apply after the matrix multiplication
      *
      * @return a status
      */
     static Status validate_mm(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const GEMMLowpOutputStageInfo &gemmlowp_output_stage,
-                              int gemm_3d_depth, bool skip_im2col, bool run_addition);
+                              int gemm_3d_depth, bool skip_im2col, const ActivationLayerInfo &act_info);
 
 private:
-    CLMemoryGroup                        _memory_group;
-    CLConvolutionLayerReshapeWeights     _reshape_weights;
-    CLIm2ColKernel                       _im2col_kernel;
-    CLGEMM                               _mm_gemm;
-    CLGEMMLowpMatrixMultiplyCore         _mm_gemmlowp;
-    CLCol2ImKernel                       _col2im_kernel;
-    CLActivationLayer                    _activationlayer_function;
-    CLSaturatedArithmeticOperationKernel _add_bias_kernel;
+    CLMemoryGroup                    _memory_group;
+    CLConvolutionLayerReshapeWeights _reshape_weights;
+    CLIm2ColKernel                   _im2col_kernel;
+    CLGEMM                           _mm_gemm;
+    CLGEMMLowpMatrixMultiplyCore     _mm_gemmlowp;
+    CLCol2ImKernel                   _col2im_kernel;
+    CLActivationLayer                _activationlayer_function;
 
     const ICLTensor *_original_weights;
 
@@ -199,15 +200,11 @@ private:
     CLTensor _weights_reshaped;
     CLTensor _gemm_output;
 
-    DataLayout _data_layout;
-
-    bool _append_bias;
     bool _skip_im2col;
     bool _skip_col2im;
     bool _is_quantized;
-    bool _is_activationlayer_enabled;
+    bool _fuse_activation;
     bool _is_prepared;
-    bool _run_addition;
 };
 } // namespace arm_compute
 #endif /* __ARM_COMPUTE_CLGEMMCONVOLUTIONLAYER_H__ */