COMPMID-1706: Fuse the bias addition within CLGEMM

Change-Id: I378f2023f4fa010f195f76716ac07aa86279bfae
Signed-off-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-on: https://review.mlplatform.org/280
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>

3 files changed, 25 insertions, 15 deletions

diff --git a/arm_compute/core/CL/kernels/CLGEMMMatrixMultiplyKernel.h b/arm_compute/core/CL/kernels/CLGEMMMatrixMultiplyKernel.h
index 797bda86cf..724a7d67e6 100644
--- a/arm_compute/core/CL/kernels/CLGEMMMatrixMultiplyKernel.h
+++ b/arm_compute/core/CL/kernels/CLGEMMMatrixMultiplyKernel.h
@@ -30,12 +30,14 @@ namespace arm_compute
 {
 class ICLTensor;
 
-/** OpenCL kernel to multiply two input matrices "A" and "B" . All elements of the output matrix will be multiplied by alpha
+/** OpenCL kernel to multiply two input matrices "A" and "B" and add a vector "C" if provided. All elements of the output matrix will be multiplied by alpha. In case vector C is passed, it will be added to the previous result (a broadcast addition will be performed).
  *
  * @note If the input tensors @p input0 and @p input1 have been reshaped respectively with @ref CLGEMMReshapeLHSMatrixKernel" and @ref CLGEMMReshapeRHSMatrixKernel,
  *       the flag @p is_interleaved_transposed must be set to true
  *
- * @attention The second input tensor must have at least 2 dimensions (matrix)
+ * @attention Vector C (@p input2) must be 1D. A broadcast addition is performed.
+ *
+ * @attention @p input1 tensor must have at least 2 dimensions (matrix)
  *
  */
 class CLGEMMMatrixMultiplyKernel : public ICLKernel
@@ -55,21 +57,25 @@ public:
      *
      * @param[in]  input0                    Input tensor containing the Matrix A. Data types supported: F16/F32
      * @param[in]  input1                    Input tensor containing the Matrix B. Data type supported: same as @p input0
+     * @param[in]  input2                    Input tensor containing the Vector C. Can be nullptr. Data type supported: same as @p input0
      * @param[out] output                    Output tensor to store the result of matrix multiplication. Data type supported: same as @p input0
      * @param[in]  alpha                     Weight of the matrix product
+     * @param[in]  beta                      (Optional) Weight of vector C. Default value is 0. Only beta = 1 is currently supported.
      * @param[in]  is_interleaved_transposed (Optional) True if input0 and input1 have been reshaped respectively using @ref CLGEMMReshapeLHSMatrixKernel and @ref CLGEMMReshapeRHSMatrixKernel
      * @param[in]  reshape_info              (Optional) GEMM reshape info. If is_interleaved_transposed = true, this object must contain the information to understand how the matrix A and matrix B have been reshaped
      * @param[in]  fp_mixed_precision        (Optional) Use wider accumulators (32 bit instead of 16 for FP16) to improve accuracy
      *
      */
-    void configure(const ICLTensor *input0, const ICLTensor *input1, ICLTensor *output, float alpha, bool is_interleaved_transposed = true, const GEMMReshapeInfo &reshape_info = GEMMReshapeInfo(),
-                   bool fp_mixed_precision = false);
+    void configure(const ICLTensor *input0, const ICLTensor *input1, const ICLTensor *input2, ICLTensor *output, float alpha, float beta = 0.f,
+                   bool is_interleaved_transposed = true, const GEMMReshapeInfo &reshape_info = GEMMReshapeInfo(), bool fp_mixed_precision = false);
     /** Static function to check if given info will lead to a valid configuration of @ref CLGEMMMatrixMultiplyKernel
      *
-     * @param[in] input0                    Input tensor containing the Matrix A. Data types supported: F16/F32
-     * @param[in] input1                    Input tensor containing the Matrix B. Data type supported: same as @p input0
+     * @param[in] input0                    Input tensor containing the Matrix A info. Data types supported: F16/F32
+     * @param[in] input1                    Input tensor containing the Matrix B info. Data type supported: same as @p input0
+     * @param[in] input2                    Input tensor containing the Vector C info. Can be nullptr. Data type supported: same as @p input0
      * @param[in] output                    Output tensor to store the result of matrix multiplication. Data type supported: same as @p input0
      * @param[in] alpha                     Weight of the matrix product
+     * @param[in] beta                      Weight of vector C. Default value is 0. Only beta = 1 is currently supported.
      * @param[in] is_interleaved_transposed True if input0 and input1 have been reshaped respectively using @ref CLGEMMReshapeLHSMatrixKernel and @ref CLGEMMReshapeRHSMatrixKernel
      * @param[in] reshape_info              GEMM reshape info. If is_interleaved_transposed = true, this object must contain the information to understand how the matrix A and matrix B have been reshaped
      * @param[in] gpu_target                GPU Target
@@ -77,8 +83,8 @@ public:
      *
      * @return a status
      */
-    static Status validate(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *output, float alpha, bool is_interleaved_transposed, const GEMMReshapeInfo &reshape_info,
-                           GPUTarget gpu_target, bool fp_mixed_precision = false);
+    static Status validate(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, float alpha, float beta,
+                           bool is_interleaved_transposed, const GEMMReshapeInfo &reshape_info, GPUTarget gpu_target, bool fp_mixed_precision = false);
 
     // Inherited methods overridden:
     void run(const Window &window, cl::CommandQueue &queue) override;
@@ -86,10 +92,12 @@ public:
 public:
     const ICLTensor *_input0;
     const ICLTensor *_input1;
+    const ICLTensor *_input2;
     ICLTensor       *_output;
     bool             _slide_matrix_b;
     bool             _reinterpret_input_as_3d;
     bool             _reinterpret_output_as_3d;
+    bool             _has_vec_c;
 };
 } // namespace arm_compute
 #endif /* __ARM_COMPUTE_CLGEMMMATRIXMULTIPLYKERNEL_H__ */
diff --git a/arm_compute/runtime/CL/functions/CLFullyConnectedLayer.h b/arm_compute/runtime/CL/functions/CLFullyConnectedLayer.h
index d6d88cec55..e800dd7cbb 100644
--- a/arm_compute/runtime/CL/functions/CLFullyConnectedLayer.h
+++ b/arm_compute/runtime/CL/functions/CLFullyConnectedLayer.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2018 ARM Limited.
+ * Copyright (c) 2017-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -136,7 +136,7 @@ private:
     CLGEMM                                              _mm_gemm;
     CLGEMMLowpMatrixMultiplyCore                        _mm_gemmlowp;
     CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint _gemmlowp_output_stage;
-    CLGEMMMatrixAccumulateBiasesKernel                  _accumulate_biases_kernel;
+    CLGEMMMatrixAccumulateBiasesKernel                  _accumulate_biases_kernel; // TODO(COMPMID-1889): Use CLGEMM to add bias in CLFullyConnectedLayer
     CLTensor                                            _flatten_output;
     CLTensor                                            _gemmlowp_output;
     CLTensor                                            _converted_weights_output;
diff --git a/arm_compute/runtime/CL/functions/CLGEMMConvolutionLayer.h b/arm_compute/runtime/CL/functions/CLGEMMConvolutionLayer.h
index d7694a8328..b304576f33 100644
--- a/arm_compute/runtime/CL/functions/CLGEMMConvolutionLayer.h
+++ b/arm_compute/runtime/CL/functions/CLGEMMConvolutionLayer.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2018 ARM Limited.
+ * Copyright (c) 2017-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -163,7 +163,7 @@ private:
      * @param[in, 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]      gemmlowp_output_stage GEMMLowp output stage info
-     * @param[in]      gemm_3d_depth         (Optional) Depth of GEMM 3D (Defaults to 1)
+     * @param[in]      gemm_3d_depth         Depth of GEMM 3D
      */
     void configure_mm(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const GEMMLowpOutputStageInfo &gemmlowp_output_stage, int gemm_3d_depth = 1);
     /** Static function to check if given info will lead to a valid configuration of @ref CLGEMMConvolutionLayer matrix multiply routines
@@ -175,13 +175,14 @@ private:
      * @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] gemmlowp_output_stage GEMMLowp output stage info
-     * @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)
+     * @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.
      *
      * @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 = 1, bool skip_im2col = false);
+                              int gemm_3d_depth, bool skip_im2col, bool run_addition);
 
 private:
     CLMemoryGroup                        _memory_group;
@@ -207,6 +208,7 @@ private:
     bool _is_quantized;
     bool _is_activationlayer_enabled;
     bool _is_prepared;
+    bool _run_addition;
 };
 } // namespace arm_compute
 #endif /* __ARM_COMPUTE_CLGEMMCONVOLUTIONLAYER_H__ */