1 /*
2 * Copyright (c) 2016-2020 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 "src/core/NEON/kernels/NEGEMMTranspose1xWKernel.h"
25
26 #include "arm_compute/core/ITensor.h"
27 #include "arm_compute/core/TensorInfo.h"
28 #include "arm_compute/core/Validate.h"
29 #include "arm_compute/core/Window.h"
30 #include "src/core/AccessWindowStatic.h"
31 #include "src/core/NEON/INEKernel.h"
32 #include "src/core/helpers/AutoConfiguration.h"
33 #include "src/core/helpers/WindowHelpers.h"
34
35 #include <arm_neon.h>
36
37 namespace arm_compute
38 {
39 namespace
40 {
get_output_shape(const ITensorInfo * input)41 TensorShape get_output_shape(const ITensorInfo *input)
42 {
43 TensorShape output_shape{ input->tensor_shape() };
44 const size_t transpose_w = 16 / input->element_size();
45 output_shape.set(0, input->dimension(1) * transpose_w);
46 output_shape.set(1, static_cast<size_t>(std::ceil((input->dimension(0) / static_cast<float>(transpose_w)))));
47 return output_shape;
48 }
49
validate_arguments(const ITensorInfo * input,const ITensorInfo * output)50 Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output)
51 {
52 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input);
53 ARM_COMPUTE_RETURN_ERROR_ON(input->data_type() == DataType::UNKNOWN);
54 //Note: ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(input) is not needed here as this kernel doesn't use NEON FP16 instructions.
55
56 if(output->total_size() != 0)
57 {
58 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), get_output_shape(input));
59 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
60 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(input, output);
61 }
62
63 return Status{};
64 }
65 } // namespace
66
configure(const ITensor * input,ITensor * output)67 void NEGEMMTranspose1xWKernel::configure(const ITensor *input, ITensor *output)
68 {
69 ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
70
71 // Output tensor auto inizialitation if not yet initialized
72 auto_init_if_empty(*output->info(), get_output_shape(input->info()), 1, input->info()->data_type());
73
74 // Perform validate step
75 ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info()));
76
77 _input = input;
78 _output = output;
79
80 const size_t vector_size = 16 / input->info()->element_size();
81
82 // Configure kernel window
83 Window win = calculate_max_window(*input->info(), Steps(vector_size));
84
85 Coordinates coord;
86 coord.set_num_dimensions(output->info()->num_dimensions());
87 output->info()->set_valid_region(ValidRegion(coord, output->info()->tensor_shape()));
88
89 INEKernel::configure(win);
90 }
91
validate(const ITensorInfo * input,const ITensorInfo * output)92 Status NEGEMMTranspose1xWKernel::validate(const ITensorInfo *input, const ITensorInfo *output)
93 {
94 ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output));
95
96 return Status{};
97 }
98
run(const Window & window,const ThreadInfo & info)99 void NEGEMMTranspose1xWKernel::run(const Window &window, const ThreadInfo &info)
100 {
101 ARM_COMPUTE_UNUSED(info);
102 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
103 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INESimpleKernel::window(), window);
104
105 /*
106 * Following an example of how the transposition1xW works when the input data type is F32
107 *
108 * |a00 a01 a02 a03|
109 * |a10 a11 a12 a13|
110 * |a20 a21 a22 a23| = | a00 a01 a02 a03 || a10 a11 a12 a13 || a20 a21 a22 a23 || a30 a31 a32 a33 |
111 * |a30 a31 a32 a33|
112 *
113 * The output matrix will have the following shape: [ height * W, ceil(width / W) ], where W = (16 / element size of the tensor)
114 */
115
116 // Set window for output tensor. Set to 0 the X and Y dimensions in order to allow multi-threading implementation and future batched matrix multiplications
117 Window win_out(window);
118 win_out.set(Window::DimX, Window::Dimension(0, 0, 0));
119 win_out.set(Window::DimY, Window::Dimension(0, 0, 0));
120
121 Iterator in(_input, window);
122 Iterator out(_output, win_out);
123
124 const size_t in_width = _input->info()->dimension(0);
125 const size_t element_size = _input->info()->element_size();
126 const size_t out_stride = _output->info()->strides_in_bytes()[1];
127 const size_t vector_size = 16 / element_size;
128
129 execute_window_loop(window, [&](const Coordinates & id)
130 {
131 const uint8_t *in_ptr = in.ptr();
132 uint8_t *const out_ptr = out.ptr() + (id.y() * vector_size) * element_size + (id.x() / vector_size) * out_stride;
133
134 for(size_t k = 0; k < vector_size; ++k)
135 {
136 // If the input width is not multiple of W, we fill the reference with 0s
137 if((id.x() + k) >= in_width)
138 {
139 std::memset(out_ptr + k * element_size, 0, element_size);
140 }
141 else
142 {
143 std::memcpy(out_ptr + k * element_size, in_ptr + k * element_size, element_size);
144 }
145 }
146 },
147 in, out);
148 }
149 } // namespace arm_compute
150