Compute Library
 21.08
GEMMReshapeRHSMatrix.cpp
Go to the documentation of this file.
1 /*
2  * Copyright (c) 2018-2020 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 #include "GEMMReshapeRHSMatrix.h"
25 
26 #include "arm_compute/core/Types.h"
27 
29 
30 #include <algorithm>
31 #include <cmath>
32 #include <cstring>
33 
34 namespace arm_compute
35 {
36 namespace test
37 {
38 namespace validation
39 {
40 namespace reference
41 {
42 template <typename T>
44 {
45  ARM_COMPUTE_ERROR_ON(in.shape().num_dimensions() > 3);
46 
48 
49  // Initialize the output tensor with zero
50  std::memset(&out[0], 0, out.num_elements() * sizeof(T));
51 
52  const unsigned int N = in.shape()[0];
53  const unsigned int K = in.shape()[1];
54  const unsigned int B = in.shape()[2];
55 
56  const unsigned int num_tiles_x = std::ceil(N / static_cast<float>(rhs_info.n0));
57  const unsigned int num_tiles_y = std::ceil(K / static_cast<float>(rhs_info.k0));
58 
59  const TensorShape tile_dims(rhs_info.n0, rhs_info.k0);
60  const TensorShape tile_dims_transposed(rhs_info.k0, rhs_info.n0);
61 
62  // Simple tensor for the input tile
63  SimpleTensor<T> src_tile{ tile_dims, in.data_type() };
64 
65  // Simple tensor for the input tile
66  SimpleTensor<T> src_tile_transposed{ tile_dims_transposed, in.data_type() };
67 
68  // Simple tensor to use when storing the values
69  SimpleTensor<T> *tile_to_use = rhs_info.transpose ? &src_tile_transposed : &src_tile;
70 
71  const unsigned int offset_output_x = rhs_info.interleave ? tile_to_use->shape()[0] : tile_to_use->shape()[0] * tile_to_use->shape()[1];
72  const unsigned int step_output_x = rhs_info.interleave ? tile_to_use->shape()[0] * rhs_info.h0 : tile_to_use->shape()[0];
73 #ifdef ARM_COMPUTE_OPENMP
74  #pragma omp parallel for schedule(dynamic, 1) collapse(3)
75 #endif /* _OPENMP */
76  for(unsigned int z = 0; z < B; ++z)
77  {
78  for(unsigned int y = 0; y < num_tiles_y; ++y)
79  {
80  for(unsigned int x = 0; x < num_tiles_x; ++x)
81  {
82  // Get the tile from the input tensor
83  get_tile<T>(in, src_tile, Coordinates(x * rhs_info.n0, y * rhs_info.k0, z, 0));
84 
85  if(rhs_info.transpose)
86  {
87  // Transpose matrix
88  transpose_matrix<T>(src_tile, src_tile_transposed);
89  }
90 
91  // Store
92  const unsigned int offset_output = (y * rhs_info.k0 * rhs_info.n0 * rhs_info.h0) + ((x % rhs_info.h0) * offset_output_x) + ((x / rhs_info.h0) * out.shape()[0]) + (z * out.shape()[0] * out.shape()[1]);
93 
94  for(unsigned int i = 0; i < tile_to_use->shape()[1]; ++i)
95  {
96  const unsigned int offset_tile = i * tile_to_use->shape()[0];
97 
98  // Copy per row
99  std::copy(&(*tile_to_use)[offset_tile], &(*tile_to_use)[offset_tile + tile_to_use->shape()[0]], &out[offset_output + i * step_output_x]);
100  }
101  }
102  }
103  }
104 
105  return out;
106 }
108 template SimpleTensor<short> gemm_reshape_rhs_matrix(const SimpleTensor<short> &in, const TensorShape &output_shape, const GEMMRHSMatrixInfo &rhs_info);
109 template SimpleTensor<char> gemm_reshape_rhs_matrix(const SimpleTensor<char> &in, const TensorShape &output_shape, const GEMMRHSMatrixInfo &rhs_info);
110 } // namespace reference
111 } // namespace validation
112 } // namespace test
113 } // namespace arm_compute
Shape of a tensor.
Definition: TensorShape.h:39
DataType data_type() const override
Data type of the tensor.
Definition: SimpleTensor.h:357
unsigned int h0
Number of horizontal blocks of size (k0xn0) stored on the same output row.
Definition: Types.h:1928
#define ARM_COMPUTE_ERROR_ON(cond)
If the condition is true then an error message is printed and an exception thrown.
Definition: Error.h:466
TensorShape shape() const override
Shape of the tensor.
Definition: SimpleTensor.h:320
SimpleTensor< T > copy(const SimpleTensor< T > &src, const TensorShape &output_shape)
Definition: Copy.cpp:37
bool transpose
True if the (k0xn0) block has to be transposed before been stored.
Definition: Types.h:1929
Copyright (c) 2017-2021 Arm Limited.
unsigned int k0
Number of partial accumulations performed by the matrix multiplication.
Definition: Types.h:1927
GEMM RHS (Right Hand Side) matrix information.
Definition: Types.h:1919
unsigned int n0
Number of columns processed by the matrix multiplication.
Definition: Types.h:1926
Coordinates of an item.
Definition: Coordinates.h:37
unsigned int N
Simple tensor object that stores elements in a consecutive chunk of memory.
Definition: SimpleTensor.h:58
bool interleave
True if the h0 (k0xn0) blocks have to be interleaved in the output row.
Definition: Types.h:1930
SimpleTensor< T > gemm_reshape_rhs_matrix(const SimpleTensor< T > &in, const TensorShape &output_shape, const GEMMRHSMatrixInfo &rhs_info)
unsigned int K