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 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/GCGEMMMatrixMultiplyKernel.h"
#include "arm_compute/core/AccessWindowStatic.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/Utils.h"
#include "arm_compute/core/Validate.h"
#include "arm_compute/core/Window.h"
#include <set>
#include <string>
using namespace arm_compute;
GCGEMMMatrixMultiplyKernel::GCGEMMMatrixMultiplyKernel()
: _input0(nullptr), _input1(nullptr), _output(nullptr)
{
}
void GCGEMMMatrixMultiplyKernel::configure(const IGCTensor *input0, const IGCTensor *input1, IGCTensor *output, float alpha, bool is_interleaved_transposed)
{
ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::F32, DataType::F16);
ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input1, output);
if(!is_interleaved_transposed)
{
ARM_COMPUTE_ERROR_ON(input0->info()->dimension(0) != input1->info()->dimension(1));
}
_input0 = input0;
_input1 = input1;
_output = output;
std::set<std::string> build_opts;
Window win;
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));
build_opts.emplace("#define COLS_A " + support::cpp11::to_string(input0->info()->dimension(0)));
build_opts.emplace("#define COLS_B " + support::cpp11::to_string(input1->info()->dimension(0)));
build_opts.emplace("#define ALPHA " + float_to_string_with_full_precision(alpha));
// Check if the output tensor is a vector. If so,the kernel runs the vector-matrix multiplication
if(is_interleaved_transposed)
{
switch(input0->info()->data_type())
{
case DataType::F16:
build_opts.emplace("#define DATA_TYPE_FP16");
break;
case DataType::F32:
build_opts.emplace("#define DATA_TYPE_FP32");
break;
default:
ARM_COMPUTE_ERROR("Current data type is not supported");
break;
}
build_opts.emplace("#define GEMM_MM_INTERLEAVED_TRANSPOSED");
// Create kernel
_kernel = GCKernelLibrary::get().create_kernel(("gemm_mm_interleaved_transposed"), build_opts);
// Configure window kernel
const unsigned int num_elems_processed_per_iteration_x = max_gc_vector_width / data_size_from_type(input0->info()->data_type());
constexpr unsigned int num_elems_processed_per_iteration_y = 4;
win = calculate_max_window(*output->info(), Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
AccessWindowRectangle input0_access(input0->info(), 0, 0, num_elems_processed_per_iteration_y, 1, 1.f, 0.25f);
AccessWindowTranspose input1_access(input1->info(), 0, 0, num_elems_processed_per_iteration_x, 1, 0.f, 0.25f);
AccessWindowRectangle output_access(output->info(), 0, 0, num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y);
update_window_and_padding(win, input0_access, input1_access, output_access);
output_access.set_valid_region(win, ValidRegion(Coordinates(0, 0), output->info()->tensor_shape()));
}
else
{
ARM_COMPUTE_ERROR_ON(input0->info()->dimension(0) != input1->info()->dimension(1));
// Special case for 1xN, 2xN, 3xN and 4xN input0 tensor
unsigned int num_elems_processed_per_iteration_x;
unsigned int num_elems_processed_per_iteration_y;
switch(input0->info()->data_type())
{
case DataType::F16:
num_elems_processed_per_iteration_x = 4;
num_elems_processed_per_iteration_y = 1;
build_opts.emplace("#define DATA_TYPE_FP16");
break;
case DataType::F32:
num_elems_processed_per_iteration_x = max_gc_vector_width / data_size_from_type(input0->info()->data_type());
num_elems_processed_per_iteration_y = std::min(static_cast<int>(output->info()->dimension(1)), 4);
build_opts.emplace("#define DATA_TYPE_FP32");
break;
default:
ARM_COMPUTE_ERROR("Current data type is not supported");
break;
}
build_opts.emplace("#define GEMM_MM_FLOATING_POINT");
build_opts.emplace("#define NUM_ELEMS_PROCESSED_PER_THREAD_X " + support::cpp11::to_string(num_elems_processed_per_iteration_x));
build_opts.emplace("#define NUM_ELEMS_PROCESSED_PER_THREAD_Y " + support::cpp11::to_string(num_elems_processed_per_iteration_y));
// Create kernel
_kernel = GCKernelLibrary::get().create_kernel("gemm_mm_floating_point", build_opts);
win = calculate_max_window(*output->info(), Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
AccessWindowStatic input0_access(input0->info(), 0, 0, ceil_to_multiple(input0->info()->dimension(0), num_elems_processed_per_iteration_x), ceil_to_multiple(input0->info()->dimension(1),
num_elems_processed_per_iteration_y));
AccessWindowStatic input1_access(input1->info(), 0, 0, ceil_to_multiple(input1->info()->dimension(0), num_elems_processed_per_iteration_x), input1->info()->dimension(1));
AccessWindowRectangle output_access(output->info(), 0, 0, num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y);
update_window_and_padding(win, input0_access, input1_access, output_access);
Coordinates coord;
coord.set_num_dimensions(output->info()->num_dimensions());
output_access.set_valid_region(win, ValidRegion(coord, output->info()->tensor_shape()));
}
_kernel.clear_params();
_kernel.set_shader_params_binding_point(0);
IGCKernel::configure(win);
}
void GCGEMMMatrixMultiplyKernel::run(const Window &window)
{
ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IGCKernel::window(), window);
_kernel.use();
Window slice = window.first_slice_window_2D();
Window slice_matrix_b = slice;
slice_matrix_b.set(Window::DimX, Window::Dimension(0, 1, 1));
slice_matrix_b.set(Window::DimY, 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;
switch(_input0->info()->data_type())
{
case DataType::F16:
add_2D_tensor_argument(idx, _input0, BufferParam(1, 2), slice);
add_2D_tensor_argument(idx, _input1, BufferParam(2, 3), slice_b);
add_2D_tensor_argument(idx, _output, BufferParam(3, 3), slice);
break;
case DataType::F32:
add_2D_tensor_argument(idx, _input0, BufferParam(1, 2), slice);
add_2D_tensor_argument(idx, _input1, BufferParam(2, 2), slice_b);
add_2D_tensor_argument(idx, _output, BufferParam(3, 2), slice);
break;
default:
ARM_COMPUTE_ERROR("Current data type is not supported");
break;
}
_kernel.update_shader_params();
enqueue(*this, slice);
}
while(window.slide_window_slice_2D(slice));
}
|
