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
126
127
|
/*
* Copyright (c) 2017 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/NEON/kernels/arm32/NEGEMMAArch32Kernel.h"
#include "arm_compute/core/AccessWindowStatic.h"
#include "arm_compute/core/AccessWindowTranspose.h"
#include "arm_compute/core/Error.h"
#include "arm_compute/core/Helpers.h"
#include "arm_compute/core/IAccessWindow.h"
#include "arm_compute/core/ITensor.h"
#include "arm_compute/core/NEON/NEFixedPoint.h"
#include "arm_compute/core/TensorInfo.h"
#include "arm_compute/core/Types.h"
#include "arm_compute/core/Utils.h"
#include "arm_compute/core/Validate.h"
#include "arm_compute/core/Window.h"
#include "support/ToolchainSupport.h"
namespace arm_compute
{
#include "arm_compute/core/NEON/kernels/assembly/gemm_interleaved.hpp"
#include "arm_compute/core/NEON/kernels/assembly/kernels/a32_sgemm_8x6.hpp"
} // namespace arm_compute
#include <arm_neon.h>
#include <cstddef>
#include <cstdint>
#include <tuple>
namespace arm_compute
{
void NEGEMMAArch32Kernel::internal_configure(const ITensor *input0, const ITensor *input1, ITensor *output, ITensor *workspace, float alpha, float beta, bool is_transposed_0, bool is_transposed_1)
{
ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::F32);
ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input1, output);
ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input0, input1, output);
_input0 = input0;
_input1 = input1;
_output = output;
_workspace = workspace;
_alpha = alpha;
_beta = beta;
_is_transposed_0 = is_transposed_0;
_is_transposed_1 = is_transposed_1;
// Configure kernel window
Window win = calculate_max_window(*output->info());
AccessWindowRectangle output_access(output->info(), 0, 0, 8, 6);
const int input0_access_end = ceil_to_multiple(input0->info()->tensor_shape().x(), 6);
const int input1_access_end = ceil_to_multiple(input1->info()->tensor_shape().x(), 8);
update_window_and_padding(win,
AccessWindowStatic(input0->info(), 0, 0, input0_access_end, input0->info()->tensor_shape().y()),
AccessWindowStatic(input1->info(), 0, 0, input1_access_end, input1->info()->tensor_shape().y()),
output_access);
INEKernel::configure(win);
}
void NEGEMMAArch32Kernel::run(const Window &window, const ThreadInfo &info)
{
ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
const int lda = _input0->info()->strides_in_bytes().y() / sizeof(float);
const int ldb = _input1->info()->strides_in_bytes().y() / sizeof(float);
const int ldc = _output->info()->strides_in_bytes().y() / sizeof(float);
const auto in1_ptr = reinterpret_cast<const float *>(_input1->buffer());
const int M = std::min(_output->info()->tensor_shape().y(), static_cast<size_t>(window.y().end())) - window.y().start();
const int N = _output->info()->tensor_shape().x();
const int K = _input0->info()->tensor_shape().x();
// Only iterate over batches
Window win(window);
win.set(0, Window::Dimension(0, 1, 1));
win.set(1, Window::Dimension(0, 1, 1));
Iterator in0(_input0, window);
Iterator out(_output, window);
GemmInterleaved<sgemm_8x6, float, float> gemm(&info.cpu_info, M, N, K, _is_transposed_0, _is_transposed_1);
constexpr size_t alignment = 4096;
const size_t offset = (gemm.get_working_size() + alignment - 1) * info.thread_id;
void *workspace = _workspace->buffer() + offset;
size_t workspace_size = _workspace->info()->total_size();
if(support::cpp11::align(alignment, gemm.get_working_size(), workspace, workspace_size) == nullptr)
{
ARM_COMPUTE_ERROR("Not enough space to align buffer!");
}
execute_window_loop(win, [&](const Coordinates & id)
{
gemm.execute(reinterpret_cast<const float *>(in0.ptr()), lda,
reinterpret_cast<const float *>(in1_ptr), ldb,
reinterpret_cast<float *>(out.ptr()), ldc,
_alpha, _beta, workspace);
},
in0, out);
}
} // namespace arm_compute
|
