COMPMID-1979: Fuse Activation Function in CLGEMM - part 3

Fused beta*bias in in the old cl gemm kernels
Fused activation function in the old cl gemm kernels

Change-Id: I695fb9189e6d4792010abd256784624982d17d79
Signed-off-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Reviewed-on: https://review.mlplatform.org/c/1587
Reviewed-by: Giuseppe Rossini <giuseppe.rossini@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

1 files changed, 10 insertions, 8 deletions

diff --git a/arm_compute/core/CL/kernels/CLGEMMMatrixMultiplyKernel.h b/arm_compute/core/CL/kernels/CLGEMMMatrixMultiplyKernel.h
index 724a7d67e6..8e6e07973c 100644
--- a/arm_compute/core/CL/kernels/CLGEMMMatrixMultiplyKernel.h
+++ b/arm_compute/core/CL/kernels/CLGEMMMatrixMultiplyKernel.h
@@ -30,13 +30,12 @@ namespace arm_compute
 {
 class ICLTensor;
 
-/** 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).
+/** OpenCL kernel to multiply two input matrices "A" and "B" and add a martix "C" if provided. All elements of the output matrix will be multiplied by alpha. In case matrix C is passed, it will be added to the previous result.
+ *  For the matrix C, the broadcast addition is supported if the flag "broadcast_bias" is set in the GEMMReshapeInfo object
  *
  * @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 Vector C (@p input2) must be 1D. A broadcast addition is performed.
- *
  * @attention @p input1 tensor must have at least 2 dimensions (matrix)
  *
  */
@@ -57,22 +56,23 @@ 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[in]  input2                    Input tensor containing the Matrix C (bias). 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
+     * @param[in]  activation_info           (Optional) Activation to apply after the matrix multiplication
      *
      */
     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);
+                   bool is_interleaved_transposed = true, const GEMMReshapeInfo &reshape_info = GEMMReshapeInfo(), bool fp_mixed_precision = false, const ActivationLayerInfo &activation_info = ActivationLayerInfo());
     /** 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 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] input2                    Input tensor containing the Matrix C (bias) 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.
@@ -80,11 +80,12 @@ public:
      * @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
      * @param[in] fp_mixed_precision        (Optional) Use wider accumulators (32 bit instead of 16 for FP16) to improve accuracy
+     * @param[in] activation_info           (Optional) Activation to apply after the matrix multiplication
      *
      * @return a status
      */
     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);
+                           bool is_interleaved_transposed, const GEMMReshapeInfo &reshape_info, GPUTarget gpu_target, bool fp_mixed_precision = false, const ActivationLayerInfo &activation_info = ActivationLayerInfo());
 
     // Inherited methods overridden:
     void run(const Window &window, cl::CommandQueue &queue) override;
@@ -97,7 +98,8 @@ public:
     bool             _slide_matrix_b;
     bool             _reinterpret_input_as_3d;
     bool             _reinterpret_output_as_3d;
-    bool             _has_vec_c;
+    bool             _add_bias;
+    bool             _broadcast_bias;
 };
 } // namespace arm_compute
 #endif /* __ARM_COMPUTE_CLGEMMMATRIXMULTIPLYKERNEL_H__ */