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
211
212
213
214
215
216
|
/*
* 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/core/CL/kernels/CLGEMMLowpMatrixMultiplyKernel.h"
#include "arm_compute/core/AccessWindowStatic.h"
#include "arm_compute/core/CL/CLHelpers.h"
#include "arm_compute/core/CL/CLKernelLibrary.h"
#include "arm_compute/core/CL/ICLTensor.h"
#include "arm_compute/core/CL/OpenCL.h"
#include "arm_compute/core/Error.h"
#include "arm_compute/core/Helpers.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"
#include <cstddef>
#include <cstdint>
#include <tuple>
using namespace arm_compute;
namespace arm_compute
{
class Coordinates;
} // namespace arm_compute
namespace
{
using ElementsProcessed = Steps;
Status validate_arguments(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *output, bool is_interleaved_transposed)
{
ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::QASYMM8);
ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input1, 1, DataType::QASYMM8);
ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::S32);
if(!is_interleaved_transposed)
{
ARM_COMPUTE_RETURN_ERROR_ON(input0->dimension(0) != input1->dimension(1));
}
return Status{};
}
std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input0, ITensorInfo *input1, ITensorInfo *output, bool is_interleaved_transposed,
ElementsProcessed &num_elements_processed)
{
unsigned int &num_elems_processed_per_iteration_x = num_elements_processed[0];
unsigned int &num_elems_processed_per_iteration_y = num_elements_processed[1];
Window win{};
bool window_changed = false;
// Check if the output tensor is a vector. If so,the kernel runs the vector-matrix multiplication
if(is_interleaved_transposed)
{
// Configure window
num_elems_processed_per_iteration_x = 16;
num_elems_processed_per_iteration_y = 4;
constexpr unsigned int num_elems_read_per_iteration_input0 = 4;
constexpr unsigned int num_elems_read_per_iteration_input1 = 16;
win = calculate_max_window(*output, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
AccessWindowRectangle input0_access(input0, 0, 0, num_elems_read_per_iteration_input0, 1);
AccessWindowRectangle input1_access(input1, 0, 0, num_elems_read_per_iteration_input1, 1);
AccessWindowRectangle output_access(output, 0, 0, num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y);
window_changed = update_window_and_padding(win, input0_access, input1_access, output_access);
output_access.set_valid_region(win, ValidRegion(Coordinates(0, 0), output->tensor_shape()));
}
else
{
// Special case for 1xN, 2xN, 3xN and 4xN input0 tensor. num_elems_processed_per_iteration_x
num_elems_processed_per_iteration_x = 4;
num_elems_processed_per_iteration_y = std::min(static_cast<int>(output->dimension(1)), 5);
// Configure window
win = calculate_max_window(*output, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
AccessWindowStatic input0_access(input0, 0, 0, input0->dimension(0), ceil_to_multiple(input0->dimension(1), num_elems_processed_per_iteration_y));
AccessWindowStatic input1_access(input1, 0, 0, ceil_to_multiple(input1->dimension(0), num_elems_processed_per_iteration_x), input1->dimension(1));
AccessWindowRectangle output_access(output, 0, 0, num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y);
window_changed = update_window_and_padding(win, input0_access, input1_access, output_access);
Coordinates coord;
coord.set_num_dimensions(output->num_dimensions());
output_access.set_valid_region(win, ValidRegion(coord, output->tensor_shape()));
}
Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
return std::make_pair(err, win);
}
} // namespace
CLGEMMLowpMatrixMultiplyKernel::CLGEMMLowpMatrixMultiplyKernel()
: _input0(nullptr), _input1(nullptr), _output(nullptr)
{
}
void CLGEMMLowpMatrixMultiplyKernel::configure(const ICLTensor *input0, const ICLTensor *input1, ICLTensor *output, bool is_interleaved_transposed)
{
ARM_COMPUTE_ERROR_ON_NULLPTR(input0, input1, output);
ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input0->info(), input1->info(), output->info(), is_interleaved_transposed));
_input0 = input0;
_input1 = input1;
_output = output;
ElementsProcessed num_elements_processed{};
// Get target architecture
GPUTarget arch_target = get_arch_from_target(get_target());
// Configure kernel window
auto win_config = validate_and_configure_window(input0->info(), input1->info(), output->info(), is_interleaved_transposed, num_elements_processed);
ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
ICLKernel::configure(win_config.second);
// Create build options
CLBuildOptions build_opts;
std::string kernel_name(" ");
if(is_interleaved_transposed)
{
build_opts.add_option("-DCOLS_B=" + support::cpp11::to_string(input1->info()->dimension(0)));
kernel_name = "gemmlowp_mm_interleaved_transposed";
}
else
{
build_opts.add_option("-DCOLS_A=" + support::cpp11::to_string(input0->info()->dimension(0)));
build_opts.add_option("-DNUM_ELEMS_PROCESSED_PER_THREAD_X=" + support::cpp11::to_string(num_elements_processed.x()));
build_opts.add_option("-DNUM_ELEMS_PROCESSED_PER_THREAD_Y=" + support::cpp11::to_string(num_elements_processed.y()));
kernel_name = "gemmlowp_mm_" + string_from_target(arch_target);
}
// Create kernel
_kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts.options()));
// Set config_id for enabling LWS tuning
_config_id = "gemmlowp_";
_config_id += (is_interleaved_transposed ? "reshaped_" : "");
_config_id += lower_string(string_from_data_type(input0->info()->data_type()));
_config_id += "_";
_config_id += support::cpp11::to_string(output->info()->dimension(1));
_config_id += "_";
_config_id += support::cpp11::to_string(output->info()->dimension(0));
_config_id += "_";
_config_id += (is_interleaved_transposed ? support::cpp11::to_string(input1->info()->dimension(0)) : support::cpp11::to_string(input1->info()->dimension(1)));
}
Status CLGEMMLowpMatrixMultiplyKernel::validate(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *output, bool is_interleaved_transposed)
{
ElementsProcessed num_elements_processed{};
ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input0, input1, output, is_interleaved_transposed));
ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input0->clone().get(),
input1->clone().get(),
output->clone().get(),
is_interleaved_transposed,
num_elements_processed)
.first);
return Status{};
}
void CLGEMMLowpMatrixMultiplyKernel::run(const Window &window, cl::CommandQueue &queue)
{
ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
Window slice = window.first_slice_window_2D();
Window slice_matrix_b = slice;
slice_matrix_b.set(Window::DimX, Window::Dimension(0, _input1->info()->dimension(0), 1));
slice_matrix_b.set(Window::DimY, Window::Dimension(0, _input1->info()->dimension(1), 1));
slice_matrix_b.set(Window::DimZ, Window::Dimension(0, 1, 1));
do
{
Window slice_b = slice;
// Don't slice matrix B along the z dimension if matrix B has just 2 dimensions and matrix A more than 2
// This scenario can happen when the the matrix multiplication is used to perform a convolution operation
if(_input1->info()->num_dimensions() < 3)
{
slice_b = slice_matrix_b;
}
unsigned int idx = 0;
add_2D_tensor_argument(idx, _input0, slice);
add_2D_tensor_argument(idx, _input1, slice_b);
add_2D_tensor_argument(idx, _output, slice);
enqueue(queue, *this, slice, _lws_hint);
}
while(window.slide_window_slice_2D(slice));
}
|
