Compute Library
 21.02
NEGEMMLowpOffsetContributionOutputStageKernel.h
Go to the documentation of this file.
1 /*
2  * Copyright (c) 2019-2021 Arm Limited.
3  *
4  * SPDX-License-Identifier: MIT
5  *
6  * Permission is hereby granted, free of charge, to any person obtaining a copy
7  * of this software and associated documentation files (the "Software"), to
8  * deal in the Software without restriction, including without limitation the
9  * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
10  * sell copies of the Software, and to permit persons to whom the Software is
11  * furnished to do so, subject to the following conditions:
12  *
13  * The above copyright notice and this permission notice shall be included in all
14  * copies or substantial portions of the Software.
15  *
16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
19  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22  * SOFTWARE.
23  */
24 #ifndef ARM_COMPUTE_NEGEMMLOWPOFFSETCONTRIBUTIONOUTPUTSTAGEKERNEL_H
25 #define ARM_COMPUTE_NEGEMMLOWPOFFSETCONTRIBUTIONOUTPUTSTAGEKERNEL_H
26 
28 
29 namespace arm_compute
30 {
31 class ITensor;
32 
33 /** Neon kernel used to add the offset contribution and perform the output stage after @ref NEGEMMLowpMatrixMultiplyKernel.
34  *
35  * The computation is performed in-place
36  *
37  * This kernel takes a final int32 accumulator value (the output of @ref NEGEMMLowpMatrixMultiplyKernel),
38  * and adds to it the offset contribution of matrix A and matrix B in-place.
39  *
40  * The output stage can perform either QuantizeDownInt32ToUint8Scale or QuantizeDownInt32ToUint8ScaleByFixedPoint for Uint8.
41  * The output stage can perform either QuantizeDownInt32ToInt8Scale or QuantizeDownInt32ToInt8ScaleByFixedPoint for Int8.
42  *
43  * For QuantizeDownInt32ToUint8Scale/QuantizeDownInt32ToInt8Scale the final result is:
44  *
45  * ((mm_result'[i][k] + result_offset) * result_mult_int) >> result_shift
46  *
47  * For QuantizeDownInt32ToUint8ScaleByFixedPoint/QuantizeDownInt32ToInt8ScaleByFixedPoint the final result is:
48  *
49  * (FixedPointMul(mm_result'[i][k], result_fixedpoint_multiplier) >> result_shift) + result_offset_after_shift
50  *
51  * where FixedPointMul(x, y) is the nearest integer to the following
52  * mathematical expression, evaluated without overflow or intermediate rounding:
53  *
54  * (x * y) / 2^31
55  *
56  * and mm_result'[i][k] = mm_result[i][k] +
57  * (vector_sum_col[k] * a_offset) +
58  * (vector_sum_row[i] * b_offset) +
59  * (a_offset * b_offset * k)
60  */
61 
63 {
64 public:
65  const char *name() const override
66  {
67  return "NEGEMMLowpOffsetContributionOutputStageKernel";
68  }
69  /** Constructor */
71  /** Prevent instances of this class from being copied (As this class contains pointers)*/
73  /** Prevent instances of this class from being copied (As this class contains pointers)*/
75  /** Allow instances of this class to be moved */
77  /** Allow instances of this class to be moved */
79  /** Default destructor */
81  /** Initialise the kernel's input and output.
82  *
83  * @param[in] mm_result Input tensor containing the result of @ref NEGEMMLowpMatrixMultiplyKernel. Data type supported: S32
84  * @param[in] vector_sum_col Input row-vector of sums of all the entries in each column of matrix B.
85  * Note: vector_sum_col can be a nullptr in case a_offset = 0. Data type supported: same as @p mm_result
86  * @param[in] vector_sum_row Input row-vector of sums of all the entries in each row of matrix A.
87  * @param[in] bias Biases tensor. Only shared biases supported and it can be a nullptr if the addition of biases is not required.
88  * Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p mm_result.
89  * @param[out] output Output tensor containing the final quantized result. Data type supported: QASYMM8/QASYMM8_SIGNED
90  * @param[in] k Number of matrix A columns or Matrix B rows
91  * @param[in] a_offset Offset to be added to each element of the matrix A.
92  * @param[in] b_offset Offset to be added to each element of the matrix B.
93  * @param[in] output_stage GEMMLowp output stage info, providing the type of quantization and the necessary parameters.
94  */
95  void configure(const ITensor *mm_result, const ITensor *vector_sum_col, const ITensor *vector_sum_row, const ITensor *bias, ITensor *output, int32_t k, int32_t a_offset, int32_t b_offset,
96  GEMMLowpOutputStageInfo output_stage);
97  /** Static function to check if given info will lead to a valid configuration of @ref NEGEMMLowpOffsetContributionOutputStageKernel
98  *
99  * @param[in] mm_result Input tensor info containing the result of @ref NEGEMMLowpMatrixMultiplyKernel. Data type supported: S32
100  * @param[in] vector_sum_col Tensor info for the input row-vector of sums of all the entries in each column of matrix B.
101  * Note: vector_sum_col can be a nullptr in case a_offset = 0. Data type supported: same as @p mm_result
102  * @param[in] vector_sum_row Tensor info for the input row-vector of sums of all the entries in each row of matrix A.
103  * Note: vector_sum_row can be a nullptr in case b_offset = 0. Data type supported: same as @p mm_result
104  * @param[in] bias Biases tensor info. Only shared biases supported and it can be a nullptr if the addition of biases is not required.
105  * Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p mm_result.
106  * @param[in] output Output tensor info containing the final quantized result. Data type supported: QASYMM8/QASYMM8_SIGNED
107  * @param[in] a_offset Offset to be added to each element of the matrix A.
108  * @param[in] b_offset Offset to be added to each element of the matrix B.
109  * @param[in] output_stage GEMMLowp output stage info, providing the type of quantization and the necessary parameters.
110  *
111  * @return a status
112  */
113  static Status validate(const ITensorInfo *mm_result, const ITensorInfo *vector_sum_col, const ITensorInfo *vector_sum_row, const ITensorInfo *bias, const ITensorInfo *output, int32_t a_offset,
114  int32_t b_offset,
115  GEMMLowpOutputStageInfo output_stage);
116 
117  // Inherited methods overridden:
118  void run(const Window &window, const ThreadInfo &info) override;
119 
120 private:
121  /** Function to use for the particular tensors passed to configure() */
122  const ITensor *_vector_sum_col;
123  const ITensor *_vector_sum_row;
124  const ITensor *_bias;
125  const ITensor *_mm_result;
126  ITensor *_output;
127  int32_t _a_offset;
128  int32_t _b_offset;
129  int32_t _k_offset;
130  bool _slide_vector_sum_col;
131  GEMMLowpOutputStageInfo _output_stage;
132 };
133 } // namespace arm_compute
134 
135 #endif /* ARM_COMPUTE_NEGEMMLOWPOFFSETCONTRIBUTIONOUTPUTSTAGEKERNEL_H */
const Window & window() const
The maximum window the kernel can be executed on.
Definition: IKernel.cpp:28
void run(const Window &window, const ThreadInfo &info) override
Execute the kernel on the passed window.
Common interface for all kernels implemented in C++.
Definition: ICPPKernel.h:38
Store the tensor's metadata.
Definition: ITensorInfo.h:40
Status class.
Definition: Error.h:52
Interface for Neon tensor.
Definition: ITensor.h:36
Copyright (c) 2017-2021 Arm Limited.
GEMMLowp output stage info.
Definition: Types.h:1952
void configure(const ITensor *mm_result, const ITensor *vector_sum_col, const ITensor *vector_sum_row, const ITensor *bias, ITensor *output, int32_t k, int32_t a_offset, int32_t b_offset, GEMMLowpOutputStageInfo output_stage)
Initialise the kernel's input and output.
NEGEMMLowpOffsetContributionOutputStageKernel & operator=(const NEGEMMLowpOffsetContributionOutputStageKernel &)=delete
Prevent instances of this class from being copied (As this class contains pointers) ...
ScaleKernelInfo info(interpolation_policy, default_border_mode, PixelValue(), sampling_policy, false)
Information about executing thread and CPU.
Definition: CPPTypes.h:235
Neon kernel used to add the offset contribution and perform the output stage after NEGEMMLowpMatrixMu...
~NEGEMMLowpOffsetContributionOutputStageKernel()=default
Default destructor.
static Status validate(const ITensorInfo *mm_result, const ITensorInfo *vector_sum_col, const ITensorInfo *vector_sum_row, const ITensorInfo *bias, const ITensorInfo *output, int32_t a_offset, int32_t b_offset, GEMMLowpOutputStageInfo output_stage)
Static function to check if given info will lead to a valid configuration of NEGEMMLowpOffsetContribu...
Describe a multidimensional execution window.
Definition: Window.h:39