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
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
|
/*
* 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/NEGEMMLowpMatrixMultiplyKernel.h"
#include "arm_compute/core/AccessWindowStatic.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/Utils.h"
#include "arm_compute/core/Validate.h"
#include "arm_compute/core/Window.h"
#include <arm_neon.h>
#include <cstddef>
#include <cstdint>
#include <tuple>
using namespace arm_compute;
namespace arm_compute
{
class Coordinates;
} // namespace arm_compute
NEGEMMLowpMatrixMultiplyKernel::NEGEMMLowpMatrixMultiplyKernel()
: _input0(nullptr), _input1(nullptr), _output(nullptr), _slide_matrix_b(true)
{
}
void NEGEMMLowpMatrixMultiplyKernel::configure(const ITensor *input0, const ITensor *input1, ITensor *output)
{
ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::QASYMM8);
ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input1);
ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::S32);
// Check if matrix B should be slidden or not
// Don't slide 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
TensorShape in0_shape = input0->info()->tensor_shape();
TensorShape in1_shape = input1->info()->tensor_shape();
TensorShape out_shape = output->info()->tensor_shape();
in0_shape.collapse(2);
in1_shape.collapse(2);
out_shape.collapse(2);
ARM_COMPUTE_ERROR_ON_MSG(in0_shape[2] != out_shape[2], "Output tensor must have the same number of batches of input0 tensor");
ARM_COMPUTE_ERROR_ON_MSG(in1_shape[2] != 1 && in0_shape[2] != in1_shape[2], "Input1 tensor must have the same number of batches of input0 or the number of batches must be set to 1");
_input0 = input0;
_input1 = input1;
_output = output;
_slide_matrix_b = in1_shape[2] != 1;
constexpr unsigned int num_elems_processed_per_iteration_x = 16;
constexpr unsigned int num_elems_processed_per_iteration_y = 4;
Window win = calculate_max_window(*output->info(), Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
AccessWindowStatic in0_access(input0->info(), 0, 0, ceil_to_multiple(input0->info()->dimension(0), 8), input0->info()->dimension(1));
AccessWindowHorizontal in1_access(input1->info(), 0, num_elems_processed_per_iteration_x);
AccessWindowRectangle output_access(output->info(), 0, 0, num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y);
update_window_and_padding(win, in0_access, in1_access, output_access);
output_access.set_valid_region(win, ValidRegion(Coordinates(0, 0), output->info()->tensor_shape()));
INEKernel::configure(win);
}
void NEGEMMLowpMatrixMultiplyKernel::run(const Window &window, const ThreadInfo &info)
{
ARM_COMPUTE_UNUSED(info);
ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
const size_t in_b_stride = _input1->info()->strides_in_bytes()[1];
const size_t out_stride = _output->info()->strides_in_bytes()[1] / _output->info()->element_size();
// Set step_x and step_y for matrix A. Scale by a factor of 4 the Y range as the input interleaved matrix A has 4 times less the rows of the output matrix
Window win_a(window);
win_a.set(Window::DimX, Window::Dimension(0, 0, 0));
win_a.set(Window::DimY, Window::Dimension(window.y().start() / 4, window.y().end() / 4, 1));
// Set step_x and step_y for matrix B. Scale by a factor of 16 the X range as the input transposed matrix A has 16 times less the columns of the output matrix
Window win_b;
// 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(_slide_matrix_b)
{
win_b = window;
}
win_b.set(Window::DimX, Window::Dimension(window.x().start() / 16, window.x().end() / 16, in_b_stride));
win_b.set(Window::DimY, Window::Dimension(0, 0, 0));
// The step x and step y for the output matrix has been already set using in configure()
Iterator ina(_input0, win_a);
Iterator inb(_input1, win_b);
Iterator out(_output, window);
const int width_b = _input1->info()->dimension(0);
// The implementation assumes that the matrix A and Matrix B have been reshaped respectively with NEGEMMInterleave4x4 and NEGEMMTranspose1xW
// The reshaping of the matrices helps to have a cache friendly implementation and helps to avoid the data re-arrangements needed for computing 16x4 elements per iteration
// All the values needed for computing a single 4x4 block will be read from consecutive memory positions
execute_window_loop(window, [&](const Coordinates & id)
{
const uint8_t *mtx_a0 = ina.ptr();
const uint8_t *mtx_b0 = inb.ptr();
// Note: Since the input are all positives, we can use uint32_t
// Accumulators for the block 0
uint32x4x4_t c0 =
{
{
vdupq_n_u32(0),
vdupq_n_u32(0),
vdupq_n_u32(0),
vdupq_n_u32(0)
}
};
// Accumulators for the block 1
uint32x4x4_t c1 =
{
{
vdupq_n_u32(0),
vdupq_n_u32(0),
vdupq_n_u32(0),
vdupq_n_u32(0)
}
};
// Accumulators for the block 2
uint32x4x4_t c2 =
{
{
vdupq_n_u32(0),
vdupq_n_u32(0),
vdupq_n_u32(0),
vdupq_n_u32(0)
}
};
// Accumulators for the block 3
uint32x4x4_t c3 =
{
{
vdupq_n_u32(0),
vdupq_n_u32(0),
vdupq_n_u32(0),
vdupq_n_u32(0)
}
};
for(int k = 0; k < width_b; k += 16, mtx_a0 += 4, mtx_b0 += 16)
{
const uint8x8_t a00_u8 = vld1_u8(mtx_a0);
const uint8x16_t b00_u8 = vld1q_u8(mtx_b0);
// Convert a00_s8 to uint16_t and get the lower part
const uint16x4_t a00_u16 = vget_low_u16(vmovl_u8(a00_u8));
// Convert b00_s8 to uint16_t
const uint16x4x4_t b00_u16 =
{
{
vget_low_u16(vmovl_u8(vget_low_u8(b00_u8))),
vget_high_u16(vmovl_u8(vget_low_u8(b00_u8))),
vget_low_u16(vmovl_u8(vget_high_u8(b00_u8))),
vget_high_u16(vmovl_u8(vget_high_u8(b00_u8)))
}
};
// 4x4 block 0
c0.val[0] = vmlal_lane_u16(c0.val[0], b00_u16.val[0], a00_u16, 0);
c0.val[1] = vmlal_lane_u16(c0.val[1], b00_u16.val[1], a00_u16, 0);
c0.val[2] = vmlal_lane_u16(c0.val[2], b00_u16.val[2], a00_u16, 0);
c0.val[3] = vmlal_lane_u16(c0.val[3], b00_u16.val[3], a00_u16, 0);
// 4x4 block 1
c1.val[0] = vmlal_lane_u16(c1.val[0], b00_u16.val[0], a00_u16, 1);
c1.val[1] = vmlal_lane_u16(c1.val[1], b00_u16.val[1], a00_u16, 1);
c1.val[2] = vmlal_lane_u16(c1.val[2], b00_u16.val[2], a00_u16, 1);
c1.val[3] = vmlal_lane_u16(c1.val[3], b00_u16.val[3], a00_u16, 1);
// 4x4 block 2
c2.val[0] = vmlal_lane_u16(c2.val[0], b00_u16.val[0], a00_u16, 2);
c2.val[1] = vmlal_lane_u16(c2.val[1], b00_u16.val[1], a00_u16, 2);
c2.val[2] = vmlal_lane_u16(c2.val[2], b00_u16.val[2], a00_u16, 2);
c2.val[3] = vmlal_lane_u16(c2.val[3], b00_u16.val[3], a00_u16, 2);
// 4x4 block 3
c3.val[0] = vmlal_lane_u16(c3.val[0], b00_u16.val[0], a00_u16, 3);
c3.val[1] = vmlal_lane_u16(c3.val[1], b00_u16.val[1], a00_u16, 3);
c3.val[2] = vmlal_lane_u16(c3.val[2], b00_u16.val[2], a00_u16, 3);
c3.val[3] = vmlal_lane_u16(c3.val[3], b00_u16.val[3], a00_u16, 3);
}
auto mtx_out = reinterpret_cast<int32_t *>(out.ptr());
vst1q_s32(mtx_out + 0 * out_stride + 0, vreinterpretq_s32_u32(c0.val[0]));
vst1q_s32(mtx_out + 0 * out_stride + 4, vreinterpretq_s32_u32(c0.val[1]));
vst1q_s32(mtx_out + 0 * out_stride + 8, vreinterpretq_s32_u32(c0.val[2]));
vst1q_s32(mtx_out + 0 * out_stride + 12, vreinterpretq_s32_u32(c0.val[3]));
vst1q_s32(mtx_out + 1 * out_stride + 0, vreinterpretq_s32_u32(c1.val[0]));
vst1q_s32(mtx_out + 1 * out_stride + 4, vreinterpretq_s32_u32(c1.val[1]));
vst1q_s32(mtx_out + 1 * out_stride + 8, vreinterpretq_s32_u32(c1.val[2]));
vst1q_s32(mtx_out + 1 * out_stride + 12, vreinterpretq_s32_u32(c1.val[3]));
vst1q_s32(mtx_out + 2 * out_stride + 0, vreinterpretq_s32_u32(c2.val[0]));
vst1q_s32(mtx_out + 2 * out_stride + 4, vreinterpretq_s32_u32(c2.val[1]));
vst1q_s32(mtx_out + 2 * out_stride + 8, vreinterpretq_s32_u32(c2.val[2]));
vst1q_s32(mtx_out + 2 * out_stride + 12, vreinterpretq_s32_u32(c2.val[3]));
vst1q_s32(mtx_out + 3 * out_stride + 0, vreinterpretq_s32_u32(c3.val[0]));
vst1q_s32(mtx_out + 3 * out_stride + 4, vreinterpretq_s32_u32(c3.val[1]));
vst1q_s32(mtx_out + 3 * out_stride + 8, vreinterpretq_s32_u32(c3.val[2]));
vst1q_s32(mtx_out + 3 * out_stride + 12, vreinterpretq_s32_u32(c3.val[3]));
},
ina, inb, out);
}
|
