COMPMID-3237: Extend GEMMLowpReduction kernels to multiply reductions by a scalar value

Change-Id: If2a242f52aea753591525d30a4cb64c1a766bf8d
Signed-off-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/2881
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Sang-Hoon Park <sang-hoon.park@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

4 files changed, 74 insertions, 33 deletions

diff --git a/arm_compute/core/KernelDescriptors.h b/arm_compute/core/KernelDescriptors.h
index 58400b190b..d9d3e1a4d8 100644
--- a/arm_compute/core/KernelDescriptors.h
+++ b/arm_compute/core/KernelDescriptors.h
@@ -124,5 +124,27 @@ struct InstanceNormalizationLayerKernelInfo
     float epsilon;             /**< Lower bound value for the normalization. Defaults to 1e-12 */
     bool  use_mixed_precision; /**< Use mixed precision in case of FP16 execution. Defaults to true */
 };
+
+struct GEMMLowpReductionKernelInfo
+{
+    /** Default constructor */
+    GEMMLowpReductionKernelInfo() = default;
+    /** Constructor
+     *
+     * @param[in] k             Number of matrix columns/rows.
+     * @param[in] is_reshaped   True if the input tensor has been reshaped.
+     * @param[in] scalar        Scalar value to multiply each reduced column/row by.
+     * @param[in] mul_by_scalar True if each column/row reduction has to be multiplied by a scalar value.
+     */
+    GEMMLowpReductionKernelInfo(int32_t k, bool is_reshaped, int32_t scalar, bool mul_by_scalar)
+        : k(k), is_reshaped(is_reshaped), scalar(scalar), mul_by_scalar(mul_by_scalar)
+    {
+    }
+
+    int32_t k{ 0 };                 /**< Number of matrix columns/rows */
+    bool    is_reshaped{ false };   /**< True if the input tensor has been reshaped */
+    int32_t scalar{ 0 };            /**< Scalar value to multiply each reduced column/row by */
+    bool    mul_by_scalar{ false }; /**< True if each column/row reduction has to be multiplied by a scalar value */
+};
 } // namespace arm_compute
 #endif /* ARM_COMPUTE_CORE_KERNEL_DESCRIPTORS_H */
diff --git a/arm_compute/core/NEON/kernels/NEGEMMLowpReductionKernel.h b/arm_compute/core/NEON/kernels/NEGEMMLowpReductionKernel.h
index fb781aea28..1e472f5252 100644
--- a/arm_compute/core/NEON/kernels/NEGEMMLowpReductionKernel.h
+++ b/arm_compute/core/NEON/kernels/NEGEMMLowpReductionKernel.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2019 ARM Limited.
+ * Copyright (c) 2017-2020 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -28,7 +28,9 @@
 
 namespace arm_compute
 {
+// Forward declarations
 class ITensor;
+struct GEMMLowpReductionKernelInfo;
 
 /** Common interface for all NEON reduction kernels */
 class INEGEMMLowpReductionKernel : public INEKernel
@@ -47,18 +49,23 @@ public:
 
     /** Initialise the kernel's input and output.
      *
-     * @param[in]  input       Input tensor. Data type supported: QASYMM8/QASYMM8_SIGNED
-     * @param[out] output      Output row-vector of sums of all the entries in each row/col of input tensor. Data type supported: S32
-     * @param[in]  k           Number of matrix A columns (or matrix B rows)
-     * @param[in]  is_reshaped True if the input tensor has been reshaped
+     * @param[in]  input  Input tensor. Data type supported: QASYMM8/QASYMM8_SIGNED
+     * @param[out] output Output row-vector of sums of all the entries in each row/col of input tensor. Data type supported: S32
+     * @param[in]  info   Kernel metadata:
+     *                    - k            Number of matrix columns/rows depending on the type of reduction.
+     *                    - is_reshaped  True if the matrix has been reshaped.
+     *                    - scalar       Scalar value to multiply each reduced column/row by.
+     *                    - mul_byscalar True if each reduced column/row must be multiplied by a scalar value.
      */
-    virtual void configure(const ITensor *input, ITensor *output, int32_t k, bool is_reshaped) = 0;
+    virtual void configure(const ITensor *input, ITensor *output, const GEMMLowpReductionKernelInfo &info) = 0;
 
 protected:
     const ITensor *_input;
     ITensor       *_output;
     int32_t        _k;
     bool           _is_reshaped;
+    int32_t        _scalar;
+    bool           _mul_by_scalar;
 };
 
 /** NEON kernel used to compute the row-vectors of sums of all the entries in each row of Matrix A.
@@ -75,22 +82,28 @@ public:
     }
     /** Initialise the kernel's input and output.
      *
-     * @param[in]  mtx_a             Input tensor. Data type supported: QASYMM8/QASYMM8_SIGNED
-     * @param[out] vector_sum_row    Output row-vector of sums of all the entries in each row of mtx_a. Data type supported: S32
-     * @param[in]  num_mtx_a_cols    Number of matrix A columns
-     * @param[in]  is_interleaved4x4 True if the matrix A has been interleaved4x4
+     * @param[in]  mtx_a          Input tensor. Data type supported: QASYMM8/QASYMM8_SIGNED
+     * @param[out] vector_sum_row Output row-vector of sums of all the entries in each row of mtx_a. Data type supported: S32
+     * @param[in]  info           Kernel metadata:
+     *                            - k            (num_mtx_a_cols) Number of matrix A columns
+     *                            - is_reshaped  (is_interleaved4x4) True if the matrix A has been interleaved4x4
+     *                            - scalar       Scalar value to multiply each reduced row by.
+     *                            - mul_byscalar True if each reduced column must be multiplied by a scalar value.
      */
-    void configure(const ITensor *mtx_a, ITensor *vector_sum_row, int32_t num_mtx_a_cols, bool is_interleaved4x4) override;
+    void configure(const ITensor *mtx_a, ITensor *vector_sum_row, const GEMMLowpReductionKernelInfo &info) override;
     /** Static function to check if given info will lead to a valid configuration of @ref NEGEMMLowpMatrixAReductionKernel
      *
-     * @param[in] mtx_a             Input tensor. Data type supported: QASYMM8/QASYMM8_SIGNED
-     * @param[in] vector_sum_row    Output row-vector of sums of all the entries in each row of mtx_a. Data type supported: S32
-     * @param[in] num_mtx_a_cols    Number of matrix A columns
-     * @param[in] is_interleaved4x4 True if the matrix A has been interleaved4x4
+     * @param[in] mtx_a          Input tensor. Data type supported: QASYMM8/QASYMM8_SIGNED
+     * @param[in] vector_sum_row Output row-vector of sums of all the entries in each row of mtx_a. Data type supported: S32
+     * @param[in] info           Kernel metadata:
+     *                           - k            (num_mtx_a_cols) Number of matrix A columns
+     *                           - is_reshaped  (is_interleaved4x4) True if the matrix A has been interleaved4x4
+     *                           - scalar       Scalar value to multiply each reduced row by.
+     *                           - mul_byscalar True if each reduced column must be multiplied by a scalar value.
      *
      * @return a status
      */
-    static Status validate(const ITensorInfo *mtx_a, const ITensorInfo *vector_sum_row, int32_t num_mtx_a_cols, bool is_interleaved4x4);
+    static Status validate(const ITensorInfo *mtx_a, const ITensorInfo *vector_sum_row, const GEMMLowpReductionKernelInfo &info);
 
     // Inherited methods overridden:
     void run(const Window &window, const ThreadInfo &info) override;
@@ -118,22 +131,28 @@ public:
     }
     /** Initialise the kernel's input and output.
      *
-     * @param[in]  mtx_b            Input tensor. Data type supported: Data type supported: QASYMM8/QASYMM8_SIGNED
-     * @param[out] vector_sum_col   Output row-vector of sums of all the entries in each column of mtx_b. Data type supported: S32
-     * @param[in]  num_mtx_b_rows   Number of matrix B rows
-     * @param[in]  is_transposed1xW True if the input tensor is transposed 1xW
+     * @param[in]  mtx_b          Input tensor. Data type supported: Data type supported: QASYMM8/QASYMM8_SIGNED
+     * @param[out] vector_sum_col Output row-vector of sums of all the entries in each column of mtx_b. Data type supported: S32
+     * @param[in]  info           Kernel metadata:
+     *                            - k            (num_mtx_b_rows) Number of matrix B rows.
+     *                            - is_reshaped  (is_transposed1xW) True if the input tensor is transposed 1xW.
+     *                            - scalar       Scalar value to multiply each reduced row by.
+     *                            - mul_byscalar True if each reduced row must be multiplied by a scalar value.
      */
-    void configure(const ITensor *mtx_b, ITensor *vector_sum_col, int32_t num_mtx_b_rows, bool is_transposed1xW) override;
+    void configure(const ITensor *mtx_b, ITensor *vector_sum_col, const GEMMLowpReductionKernelInfo &info) override;
     /** Static function to check if given info will lead to a valid configuration of @ref NEGEMMLowpMatrixBReductionKernel
      *
-     * @param[in] mtx_b            Input tensor. Data type supported: Data type supported: QASYMM8/QASYMM8_SIGNED
-     * @param[in] vector_sum_col   Output row-vector of sums of all the entries in each column of mtx_b. Data type supported: S32
-     * @param[in] num_mtx_b_rows   Number of matrix B rows
-     * @param[in] is_transposed1xW True if the input tensor is transposed 1xW
+     * @param[in] mtx_b          Input tensor. Data type supported: Data type supported: QASYMM8/QASYMM8_SIGNED
+     * @param[in] vector_sum_col Output row-vector of sums of all the entries in each column of mtx_b. Data type supported: S32
+     * @param[in] info           Kernel metadata:
+     *                           - k            (num_mtx_b_rows) Number of matrix B rows.
+     *                           - is_reshaped  (is_transposed1xW) True if the input tensor is transposed 1xW.
+     *                           - scalar       Scalar value to multiply each reduced row by.
+     *                           - mul_byscalar True if each reduced row must be multiplied by a scalar value.
      *
      * @return a status
      */
-    static Status validate(const ITensorInfo *mtx_b, const ITensorInfo *vector_sum_col, int32_t num_mtx_b_rows, bool is_transposed1xW);
+    static Status validate(const ITensorInfo *mtx_b, const ITensorInfo *vector_sum_col, const GEMMLowpReductionKernelInfo &info);
 
     // Inherited methods overridden:
     void run(const Window &window, const ThreadInfo &info) override;
diff --git a/arm_compute/runtime/NEON/functions/NEGEMM.h b/arm_compute/runtime/NEON/functions/NEGEMM.h
index c87e806d0c..8dc6b88bb0 100644
--- a/arm_compute/runtime/NEON/functions/NEGEMM.h
+++ b/arm_compute/runtime/NEON/functions/NEGEMM.h
@@ -74,7 +74,7 @@ public:
      * @note GEMM: General Matrix Multiply - [alpha * A * B + beta * C].
      * @note GEMM: The tensors a, b, c, d must have the same data type. You should not mix data types when calling this function.
      *
-     * @param[in]  a         First input tensor  (Matrix A or Vector A). Data type supported: BLOAT16/F16/F32
+     * @param[in]  a         First input tensor  (Matrix A or Vector A). Data type supported: BFLOAT16/F16/F32
      * @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 if just the multiplication between @p a and @p b is needed. Data type supported: same as @p a
      * @param[out] d         Output tensor. Data type supported: same as @p a
@@ -86,7 +86,7 @@ public:
     void configure(const ITensor *a, const ITensor *b, const ITensor *c, ITensor *d, float alpha, float beta, const GEMMInfo &gemm_info = GEMMInfo());
     /** Static function to check if given info will lead to a valid configuration of @ref NEGEMM.
      *
-     * @param[in]  a         First input tensor info  (Matrix or Vector A). Data types supported: BLOAT16/F16/F32
+     * @param[in]  a         First input tensor info  (Matrix or Vector A). Data types supported: BFLOAT16/F16/F32
      * @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 if just the multiplication between @p a and @p b is needed. Data type supported: same as @p a.
      * @param[out] output    Output tensor info. Data type supported: same as @p a
diff --git a/arm_compute/runtime/NEON/functions/NEGEMMConvolutionLayer.h b/arm_compute/runtime/NEON/functions/NEGEMMConvolutionLayer.h
index 5368384b19..e7da1006e0 100644
--- a/arm_compute/runtime/NEON/functions/NEGEMMConvolutionLayer.h
+++ b/arm_compute/runtime/NEON/functions/NEGEMMConvolutionLayer.h
@@ -66,9 +66,9 @@ public:
      * @param[in]  weights Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM].
      *                     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: Same as @p weights, S32 if @p weights is QASYMM8/QASYMM8_SIGNED, FP32 if @p weights is BLOAT16
+     *                     Data type supported: Same as @p weights, S32 if @p weights is QASYMM8/QASYMM8_SIGNED, FP32 if @p weights is BFLOAT16
      * @param[out] output  Destination tensor.
-     *                     Data types supported: Same as @p weights, FP32 if @p weights is BLOAT16
+     *                     Data types supported: Same as @p weights, FP32 if @p weights is BFLOAT16
      */
     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
@@ -76,9 +76,9 @@ public:
      * @param[in] weights Weights tensor info. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM].
      *                    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: Same as @p weights, S32 if @p weights is QASYMM8/QASYMM8_SIGNED, FP32 if @p weights is BLOAT16
+     *                    Data type supported: Same as @p weights, S32 if @p weights is QASYMM8/QASYMM8_SIGNED, FP32 if @p weights is BFLOAT16
      * @param[in] output  Destination tensor.
-     *                    Data types supported: Same as @p weights FP32 if @p weights is BLOAT16
+     *                    Data types supported: Same as @p weights FP32 if @p weights is BFLOAT16
      *
      * @return an error status
      */
@@ -140,7 +140,7 @@ private:
 /** Basic function to compute the convolution layer. This function calls the following NEON kernels/functions:
  *
  * -# @ref NEIm2ColKernel
- * -# @ref NEGEMM (if the data type is BLOAT16/FP16/FP32)
+ * -# @ref NEGEMM (if the data type is BFLOAT16/FP16/FP32)
  * -# @ref NEGEMMLowpMatrixMultiplyCore (if the data type is QASYMM8/QASYMM8_SIGNED)
  * -# @ref NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint (if the data type is QASYMM8/QASYMM8_SIGNED)
  * -# @ref NEArithmeticAdditionKernel (if biases != nullptr and we have a 1x1 convolution with the NHWC data layout)