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
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
|
/*
* Copyright (c) 2017-2018 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/runtime/NEON/functions/NEGEMM.h"
#include "arm_compute/core/Error.h"
#include "arm_compute/core/Helpers.h"
#include "arm_compute/core/ITensor.h"
#include "arm_compute/core/TensorInfo.h"
#include "arm_compute/core/Types.h"
#include "arm_compute/core/Validate.h"
#include "arm_compute/runtime/NEON/AssemblyHelper.h"
#include "arm_compute/runtime/NEON/NEScheduler.h"
#include "arm_compute/runtime/TensorAllocator.h"
#include "support/ToolchainSupport.h"
#include <cmath>
namespace arm_compute
{
NEGEMM::NEGEMM(std::shared_ptr<IMemoryManager> memory_manager)
: _memory_group(std::move(memory_manager)), _interleave_kernel(), _transpose_kernel(), _mm_kernel(), _asm_glue(), _ma_kernel(), _tmp_a(), _tmp_b(), _workspace(), _B_pretransposed(),
_original_b(nullptr), _run_vector_matrix_multiplication(false), _run_addition(false), _reshape_b_only_on_first_run(false), _is_prepared(false)
{
}
void NEGEMM::configure(const ITensor *a, const ITensor *b, const ITensor *c, ITensor *d, float alpha, float beta, const GEMMInfo &gemm_info)
{
ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(a, 1, DataType::F32, DataType::F16, DataType::QS8, DataType::QS16);
ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(a, b, d);
ARM_COMPUTE_ERROR_ON_MSG(a->info()->dimension(0) != b->info()->dimension(1), "The product AB is defined only if the number of columns in A is equal to the number of rows in B");
ARM_COMPUTE_ERROR_ON_MSG(gemm_info.is_a_reshaped(), "Matrix A already reshaped is not supported");
ARM_COMPUTE_ERROR_ON_MSG(gemm_info.is_b_reshaped(), "Matrix B already reshaped is not supported");
if(c != nullptr)
{
ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(c, 1, DataType::F32, DataType::F16, DataType::QS8, DataType::QS16);
ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(a, c);
ARM_COMPUTE_ERROR_ON_MSG(a->info()->dimension(1) != c->info()->dimension(1), "The C matrix must have the same number of rows as the matrix A");
ARM_COMPUTE_ERROR_ON_MSG(b->info()->dimension(0) != c->info()->dimension(0), "The C matrix must have the same number of columns as the matrix B");
ARM_COMPUTE_ERROR_ON_MSG(c->info()->dimension(0) != d->info()->dimension(0), "The C matrix must have the same number of rows as the output matrix");
ARM_COMPUTE_ERROR_ON_MSG(c->info()->dimension(1) != d->info()->dimension(1), "The C matrix must have the same number of columns as the output matrix");
}
// Check if we need to reshape the matrix B only on the first run
_is_prepared = false;
_reshape_b_only_on_first_run = gemm_info.reshape_b_only_on_first_run();
_run_vector_matrix_multiplication = a->info()->dimension(1) < 2;
_original_b = b;
_asm_glue._optimised_kernel = nullptr;
const bool run_optimised = a->info()->data_type() == DataType::F32 && (c == nullptr || beta == 0.f)
&& setup_assembly_kernel(a, b, d, alpha, beta, _reshape_b_only_on_first_run, _workspace, _B_pretransposed, _memory_group, _asm_glue);
// Check if the first input tensor is a vector.
// If so, all the kernels for reshaping the tensors can be skipped
if(_run_vector_matrix_multiplication)
{
if(!run_optimised)
{
// Configure the matrix multiply kernel
_mm_kernel.configure(a, b, d, alpha, false);
}
// Configure matrix addition kernel
if(beta != 0 && c != nullptr)
{
_ma_kernel.configure(c, d, beta);
_run_addition = true;
}
}
else
{
if(!run_optimised)
{
TensorShape shape_tmp_a = a->info()->tensor_shape();
TensorShape shape_tmp_b = b->info()->tensor_shape();
shape_tmp_a.set(0, a->info()->dimension(0) * 4);
shape_tmp_a.set(1, std::ceil(a->info()->dimension(1) / 4.0f));
const unsigned int transpose_w = 16 / data_size_from_type(b->info()->data_type());
shape_tmp_b.set(0, b->info()->dimension(1) * transpose_w);
shape_tmp_b.set(1, std::ceil(b->info()->dimension(0) / static_cast<float>(transpose_w)));
TensorInfo info_a(shape_tmp_a, 1, a->info()->data_type(), a->info()->fixed_point_position());
TensorInfo info_b(shape_tmp_b, 1, b->info()->data_type(), a->info()->fixed_point_position());
_tmp_a.allocator()->init(info_a);
_tmp_b.allocator()->init(info_b);
// Manage intermediate buffers
_memory_group.manage(&_tmp_a);
if(!_reshape_b_only_on_first_run)
{
_memory_group.manage(&_tmp_b);
}
int m = a->info()->dimension(1);
int n = b->info()->dimension(0);
int k = a->info()->dimension(0);
// Configure interleave kernel
_interleave_kernel.configure(a, &_tmp_a);
// Configure transpose kernel
_transpose_kernel.configure(b, &_tmp_b);
// Configure matrix multiplication kernel
_mm_kernel.configure(&_tmp_a, &_tmp_b, d, alpha, true, GEMMReshapeInfo(m, n, k));
// Allocate once the all configure methods have been called
_tmp_a.allocator()->allocate();
if(!_reshape_b_only_on_first_run)
{
_tmp_b.allocator()->allocate();
}
// Configure matrix addition kernel
if(beta != 0 && c != nullptr)
{
_ma_kernel.configure(c, d, beta);
_run_addition = true;
}
}
}
}
void NEGEMM::run()
{
prepare();
if(_asm_glue._optimised_kernel != nullptr)
{
_memory_group.acquire();
_asm_glue.run();
_memory_group.release();
}
else
{
_memory_group.acquire();
if(!_run_vector_matrix_multiplication)
{
// Run interleave kernel
NEScheduler::get().schedule(&_interleave_kernel, Window::DimY);
if(!_reshape_b_only_on_first_run)
{
// Run transpose kernel
NEScheduler::get().schedule(&_transpose_kernel, Window::DimY);
}
}
NEScheduler::get().schedule(&_mm_kernel, _run_vector_matrix_multiplication ? Window::DimX : Window::DimY);
_memory_group.release();
// Run matrix addition kernel
if(_run_addition)
{
NEScheduler::get().schedule(&_ma_kernel, Window::DimY);
}
}
}
void NEGEMM::prepare()
{
if(!_is_prepared)
{
if(_asm_glue._optimised_kernel)
{
ARM_COMPUTE_ERROR_ON(!_original_b->is_used());
_asm_glue.prepare();
_original_b->mark_as_unused();
}
else if(_reshape_b_only_on_first_run && !_run_vector_matrix_multiplication && !_asm_glue._optimised_kernel)
{
ARM_COMPUTE_ERROR_ON(!_original_b->is_used());
_tmp_b.allocator()->allocate();
NEScheduler::get().schedule(&_transpose_kernel, Window::DimY);
_original_b->mark_as_unused();
}
_is_prepared = true;
}
}
} // namespace arm_compute
|
