1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
|
/*
* Copyright (c) 2017-2020 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to
* deal in the Software without restriction, including without limitation the
* rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
* sell copies of the Software, and to permit persons to whom the Software is
* furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in all
* copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*/
#include "arm_compute/core/GLES_COMPUTE/kernels/GCGEMMTranspose1xWKernel.h"
#include "arm_compute/core/AccessWindowTranspose.h"
#include "arm_compute/core/Error.h"
#include "arm_compute/core/GLES_COMPUTE/GCHelpers.h"
#include "arm_compute/core/GLES_COMPUTE/GCKernelLibrary.h"
#include "arm_compute/core/GLES_COMPUTE/IGCTensor.h"
#include "arm_compute/core/GLES_COMPUTE/OpenGLES.h"
#include "arm_compute/core/Helpers.h"
#include "arm_compute/core/Types.h"
#include "arm_compute/core/Validate.h"
#include "arm_compute/core/Window.h"
#include "support/StringSupport.h"
#include <cmath>
using namespace arm_compute;
void GCGEMMTranspose1xWKernel::configure(const IGCTensor *input, IGCTensor *output)
{
ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32);
ARM_COMPUTE_ERROR_ON_NULLPTR(output);
TensorShape output_shape{ input->info()->tensor_shape() };
const size_t transpose_w = 16 / input->info()->element_size();
output_shape.set(0, input->info()->dimension(1) * transpose_w);
output_shape.set(1, static_cast<size_t>(std::ceil((input->info()->dimension(0) / static_cast<float>(transpose_w)))));
// Output tensor auto inizialitation if not yet initialized
auto_init_if_empty(*output->info(), output_shape, 1, input->info()->data_type());
ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
ARM_COMPUTE_ERROR_ON_MISMATCHING_DIMENSIONS(output->info()->tensor_shape(), output_shape);
const unsigned int num_elems_processed_per_iteration = 16 / input->info()->element_size();
const int scale_x = num_elems_processed_per_iteration;
_input = input;
_output = output;
std::set<std::string> build_opts;
std::string dt_name = (input->info()->data_type() == DataType::F32) ? "DATA_TYPE_FP32" : "DATA_TYPE_FP16";
build_opts.emplace(("#define " + dt_name));
build_opts.emplace("#define LOCAL_SIZE_X " + support::cpp11::to_string(1));
build_opts.emplace("#define LOCAL_SIZE_Y " + support::cpp11::to_string(1));
build_opts.emplace("#define LOCAL_SIZE_Z " + support::cpp11::to_string(1));
/*
* Following an example of how the transposition1xW works when the input data type is F32
*
* |a00 a01 a02 a03|
* |a10 a11 a12 a13|
* |a20 a21 a22 a23| = | a00 a01 a02 a03 || a10 a11 a12 a13 || a20 a21 a22 a23 || a30 a31 a32 a33 |
* |a30 a31 a32 a33|
*
* The output matrix will have the following shape: [ height * W, ceil(width / W) ], where W = (16 / element size of the tensor)
*/
// Create kernel
build_opts.emplace("#define GEMM_TRANSPOSE1xW");
_kernel = GCKernelLibrary::get().create_kernel("gemm_transpose1x4", build_opts);
// Configure window
Window win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration));
ARM_COMPUTE_ERROR_ON_MSG((win.x().end() / scale_x) == 0, "Transposed shape would be 0 in the second dimension");
AccessWindowHorizontal input_access(input->info(), 0, num_elems_processed_per_iteration);
AccessWindowTranspose output_access(output->info(), 0, 0, num_elems_processed_per_iteration, 1, scale_x, 1.f / scale_x);
update_window_and_padding(win, input_access, output_access);
output->info()->set_valid_region(ValidRegion(Coordinates(), output->info()->tensor_shape()));
IGCKernel::configure(win);
}
void GCGEMMTranspose1xWKernel::run(const Window &window)
{
ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IGCKernel::window(), window);
_kernel.use();
// Output is transposed
Window out_window(window);
out_window.set(Window::DimX, window.y());
out_window.set(Window::DimY, window.x());
Window in_slice = window.first_slice_window_2D();
Window out_slice = out_window.first_slice_window_2D();
do
{
unsigned int idx = 0;
add_2D_tensor_argument(idx, _input, 1, in_slice);
add_2D_tensor_argument(idx, _output, 2, out_slice);
_kernel.update_shader_params();
enqueue(*this, in_slice);
}
while(window.slide_window_slice_2D(in_slice) && out_window.slide_window_slice_2D(out_slice));
}
|
