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
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
|
/*
* Copyright (c) 2021 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.
*/
#ifndef SRC_CORE_NEON_KERNELS_CONV3D_QUANTIZED_H
#define SRC_CORE_NEON_KERNELS_CONV3D_QUANTIZED_H
#include "arm_compute/core/Types.h"
#include "arm_compute/core/utils/misc/Traits.h"
#include "arm_compute/core/utils/quantization/AsymmHelpers.h"
#include "arm_compute/runtime/FunctionDescriptors.h"
#include "src/core/helpers/WindowHelpers.h"
#include "src/core/NEON/NEAsymm.h"
#include "src/core/NEON/wrapper/wrapper.h"
namespace arm_compute
{
namespace cpu
{
template <typename T>
void directconv3d_quantized_neon_ndhwc(const ITensor *src0,
const ITensor *src1,
const ITensor *src2,
ITensor *dst,
const Conv3dInfo &conv_info,
const Window &window)
{
const ITensor *src = src0;
const ITensor *weights = src1;
const ITensor *biases = src2;
using vtype = wrapper::traits::neon_bitvector<T, wrapper::traits::BitWidth::W128>;
using vector_type = typename vtype::type;
using tag_type = typename vtype::tag_type;
constexpr int num_elems_read_per_iteration = 16 / sizeof(T);
using q16_t = typename wrapper::traits::promote_t<T>;
using q32_t = typename wrapper::traits::promote_t<q16_t>;
using q32x4_t = typename wrapper::traits::neon_vector<q32_t, 4>::type;
const int32_t input_offset = -src->info()->quantization_info().uniform().offset;
const float input_scale = src->info()->quantization_info().uniform().scale;
const int32_t weights_offset = -weights->info()->quantization_info().uniform().offset;
const float weights_scale = weights->info()->quantization_info().uniform().scale;
const int32_t output_offset = dst->info()->quantization_info().uniform().offset;
const float output_scale = dst->info()->quantization_info().uniform().scale;
int32_t output_multiplier = 0;
int32_t output_shift = 0;
const float multiplier = input_scale * weights_scale / output_scale;
arm_compute::quantization::calculate_quantized_multiplier(multiplier, &output_multiplier, &output_shift);
// Scalar quantities (N D H W Cin)
const int element_size = src->info()->element_size();
const int input_stride_w = src->info()->strides_in_bytes().y() / element_size;
const int input_stride_h = src->info()->strides_in_bytes().z() / element_size;
const int input_stride_d = src->info()->strides_in_bytes()[3] / element_size;
const int input_stride_n = src->info()->strides_in_bytes()[4] / element_size;
const int input_dim_w = src->info()->dimension(1);
const int input_dim_h = src->info()->dimension(2);
const int input_dim_d = src->info()->dimension(3);
// Kernel info (D H W Cin Cout)
const unsigned int kernel_stride_w = weights->info()->strides_in_bytes()[2] / element_size;
const unsigned int kernel_stride_h = weights->info()->strides_in_bytes()[3] / element_size;
const unsigned int kernel_stride_d = weights->info()->strides_in_bytes()[4] / element_size;
const int kernel_dim_w = weights->info()->dimension(2);
const int kernel_dim_h = weights->info()->dimension(3);
const int kernel_dim_d = weights->info()->dimension(4);
// Convolution padding and stride
const int conv_pad_top = conv_info.padding.top;
const int conv_pad_left = conv_info.padding.left;
const int conv_pad_front = conv_info.padding.front;
const int conv_stride_w = conv_info.stride.width;
const int conv_stride_h = conv_info.stride.height;
const int conv_stride_d = conv_info.stride.depth;
// Setup input window for the output iterator
Window window_out = window;
window_out.set(Window::DimX, Window::Dimension(0, 1, 1));
// Setup input window for the weights iterator
Window window_w = calculate_max_window(*weights->info(), Steps());
window_w.set(Window::DimY, Window::Dimension(0, 1, 1));
window_w.set(Window::DimZ, Window::Dimension(0, 1, 1));
window_w.set(Window::DimW, Window::Dimension(0, 1, 1));
window_w.set(4, Window::Dimension(0, 1, 1));
Iterator out(dst, window_out);
Iterator wei(weights, window_w);
const int32_t *biases_ptr = nullptr;
if (biases != nullptr)
{
biases_ptr = reinterpret_cast<int32_t *>(biases->buffer() + biases->info()->offset_first_element_in_bytes());
}
execute_window_loop(
window_out,
[&](const Coordinates &id)
{
// We are computing the theoretical input starting points
const int in_w_start_t = static_cast<int>(id.y()) * conv_stride_w - conv_pad_left;
const int in_h_start_t = static_cast<int>(id.z()) * conv_stride_h - conv_pad_top;
const int in_d_start_t = static_cast<int>(id[3]) * conv_stride_d - conv_pad_front;
const int in_w_end_t = in_w_start_t + kernel_dim_w;
const int in_h_end_t = in_h_start_t + kernel_dim_h;
const int in_d_end_t = in_d_start_t + kernel_dim_d;
// We are computing the valid initial and ending input points by checking the borders
const int in_w_start = std::max(in_w_start_t, 0);
const int in_h_start = std::max(in_h_start_t, 0);
const int in_d_start = std::max(in_d_start_t, 0);
const int in_w_end = std::min(in_w_end_t, input_dim_w);
const int in_h_end = std::min(in_h_end_t, input_dim_h);
const int in_d_end = std::min(in_d_end_t, input_dim_d);
// We use the input points to select the valid weight points to use
const int wei_w_start = in_w_start - in_w_start_t;
const int wei_h_start = in_h_start - in_h_start_t;
const int wei_d_start = in_d_start - in_d_start_t;
const int wei_w_end = kernel_dim_w - (in_w_end_t - in_w_end);
const int wei_h_end = kernel_dim_h - (in_h_end_t - in_h_end);
const int wei_d_end = kernel_dim_d - (in_d_end_t - in_d_end);
const int index_c_out_end = weights->info()->dimension(0);
const int index_c_in_end = weights->info()->dimension(1);
const T *const in_ptr_start =
reinterpret_cast<const T *>(src->buffer() + src->info()->offset_first_element_in_bytes()) +
id[4] * input_stride_n;
execute_window_loop(
window_w,
[&](const Coordinates &id_w)
{
/*
* This is the loop in the weights, and it goes along OFM (output feature map)
*/
const auto weights_ptr_start = reinterpret_cast<const T *>(wei.ptr());
int32_t acc = static_cast<int32_t>(0);
T *out_ptr = reinterpret_cast<T *>(out.ptr());
for (int index_wei_d = wei_d_start, index_in_d = in_d_start; index_wei_d < wei_d_end;
++index_wei_d, ++index_in_d)
{
const auto in_ptr_d = in_ptr_start + index_in_d * input_stride_d;
const auto weights_ptr_d = weights_ptr_start + index_wei_d * kernel_stride_d;
for (int index_wei_h = wei_h_start, index_in_h = in_h_start; index_wei_h < wei_h_end;
++index_wei_h, ++index_in_h)
{
const T *const in_ptr_row = in_ptr_d + index_in_h * input_stride_h;
const T *const weights_ptr_row = weights_ptr_d + index_wei_h * kernel_stride_h;
for (int index_wei_w = wei_w_start, index_in_w = in_w_start; index_wei_w < wei_w_end;
++index_wei_w, ++index_in_w)
{
const T *in_ptr_mover = in_ptr_row + index_in_w * input_stride_w;
const T *weights_ptr_mover = weights_ptr_row + index_wei_w * kernel_stride_w;
int index_c_in = 0;
vector_type w_vec = wrapper::vdup_n(static_cast<T>(0), tag_type());
q32x4_t acc_q32_0 = wrapper::vdup_n(static_cast<q32_t>(0), tag_type());
q32x4_t acc_q32_1 = wrapper::vdup_n(static_cast<q32_t>(0), tag_type());
q32x4_t acc_q32_2 = wrapper::vdup_n(static_cast<q32_t>(0), tag_type());
q32x4_t acc_q32_3 = wrapper::vdup_n(static_cast<q32_t>(0), tag_type());
for (; index_c_in <= index_c_in_end - num_elems_read_per_iteration;
index_c_in += num_elems_read_per_iteration,
in_ptr_mover += num_elems_read_per_iteration)
{
const auto src_vec = wrapper::vloadq(in_ptr_mover);
//Load Cin weights
for (int k = 0; k < num_elems_read_per_iteration;
++k, weights_ptr_mover += index_c_out_end)
{
w_vec = wrapper::vsetlane(*weights_ptr_mover, w_vec, k);
}
q32x4_t src_q32_0 = wrapper::vdup_n(static_cast<q32_t>(input_offset), tag_type());
q32x4_t src_q32_1 = wrapper::vdup_n(static_cast<q32_t>(input_offset), tag_type());
q32x4_t src_q32_2 = wrapper::vdup_n(static_cast<q32_t>(input_offset), tag_type());
q32x4_t src_q32_3 = wrapper::vdup_n(static_cast<q32_t>(input_offset), tag_type());
q32x4_t wei_q32_0 = wrapper::vdup_n(static_cast<q32_t>(weights_offset), tag_type());
q32x4_t wei_q32_1 = wrapper::vdup_n(static_cast<q32_t>(weights_offset), tag_type());
q32x4_t wei_q32_2 = wrapper::vdup_n(static_cast<q32_t>(weights_offset), tag_type());
q32x4_t wei_q32_3 = wrapper::vdup_n(static_cast<q32_t>(weights_offset), tag_type());
const auto src_q16_0 = wrapper::vmovl(wrapper::vgetlow(src_vec));
const auto src_q16_1 = wrapper::vmovl(wrapper::vgethigh(src_vec));
const auto wei_q16_0 = wrapper::vmovl(wrapper::vgetlow(w_vec));
const auto wei_q16_1 = wrapper::vmovl(wrapper::vgethigh(w_vec));
src_q32_0 = wrapper::vadd(src_q32_0, wrapper::vmovl(wrapper::vgetlow(src_q16_0)));
src_q32_1 = wrapper::vadd(src_q32_1, wrapper::vmovl(wrapper::vgethigh(src_q16_0)));
src_q32_2 = wrapper::vadd(src_q32_2, wrapper::vmovl(wrapper::vgetlow(src_q16_1)));
src_q32_3 = wrapper::vadd(src_q32_3, wrapper::vmovl(wrapper::vgethigh(src_q16_1)));
wei_q32_0 = wrapper::vadd(wei_q32_0, wrapper::vmovl(wrapper::vgetlow(wei_q16_0)));
wei_q32_1 = wrapper::vadd(wei_q32_1, wrapper::vmovl(wrapper::vgethigh(wei_q16_0)));
wei_q32_2 = wrapper::vadd(wei_q32_2, wrapper::vmovl(wrapper::vgetlow(wei_q16_1)));
wei_q32_3 = wrapper::vadd(wei_q32_3, wrapper::vmovl(wrapper::vgethigh(wei_q16_1)));
acc_q32_0 = wrapper::vmla(acc_q32_0, wei_q32_0, src_q32_0);
acc_q32_1 = wrapper::vmla(acc_q32_1, wei_q32_1, src_q32_1);
acc_q32_2 = wrapper::vmla(acc_q32_2, wei_q32_2, src_q32_2);
acc_q32_3 = wrapper::vmla(acc_q32_3, wei_q32_3, src_q32_3);
}
#if defined(__aarch64__)
acc += wrapper::vaddv(acc_q32_0);
acc += wrapper::vaddv(acc_q32_1);
acc += wrapper::vaddv(acc_q32_2);
acc += wrapper::vaddv(acc_q32_3);
#else // __aarch64__
auto temp = wrapper::vpadd(wrapper::vgethigh(acc_q32_0), wrapper::vgetlow(acc_q32_0));
temp = wrapper::vpadd(temp, temp);
acc += wrapper::vgetlane(temp, 0);
temp = wrapper::vpadd(wrapper::vgethigh(acc_q32_1), wrapper::vgetlow(acc_q32_1));
temp = wrapper::vpadd(temp, temp);
acc += wrapper::vgetlane(temp, 0);
temp = wrapper::vpadd(wrapper::vgethigh(acc_q32_2), wrapper::vgetlow(acc_q32_2));
temp = wrapper::vpadd(temp, temp);
acc += wrapper::vgetlane(temp, 0);
temp = wrapper::vpadd(wrapper::vgethigh(acc_q32_3), wrapper::vgetlow(acc_q32_3));
temp = wrapper::vpadd(temp, temp);
acc += wrapper::vgetlane(temp, 0);
#endif // __aarch64__
for (; index_c_in < index_c_in_end;
++index_c_in, ++in_ptr_mover, weights_ptr_mover += index_c_out_end)
{
const auto src_val = *(in_ptr_mover) + input_offset;
const auto w_val = *(weights_ptr_mover) + weights_offset;
acc += src_val * w_val;
}
}
}
}
if (biases)
{
acc += *reinterpret_cast<const int32_t *>(biases_ptr + id_w[0]);
}
T out_val =
finalize_quantization(acc, output_multiplier, output_shift, output_offset, T(0), T(0), false);
*(reinterpret_cast<T *>(out_ptr + id_w[0])) = out_val;
},
wei);
},
out);
}
} // namespace cpu
} // namespace arm_compute
#endif // SRC_CORE_NEON_KERNELS_CONV3D_QUANTIZED_H
|
