1 /*
2 * Copyright (c) 2021 Arm Limited.
3 *
4 * SPDX-License-Identifier: MIT
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a copy
7 * of this software and associated documentation files (the "Software"), to
8 * deal in the Software without restriction, including without limitation the
9 * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
10 * sell copies of the Software, and to permit persons to whom the Software is
11 * furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice shall be included in all
14 * copies or substantial portions of the Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
19 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22 * SOFTWARE.
23 */
24 #ifndef SRC_CORE_NEON_KERNELS_CONV3D_QUANTIZED_H
25 #define SRC_CORE_NEON_KERNELS_CONV3D_QUANTIZED_H
26
27 #include "arm_compute/core/Types.h"
28 #include "arm_compute/core/utils/misc/Traits.h"
29 #include "arm_compute/core/utils/quantization/AsymmHelpers.h"
30 #include "arm_compute/runtime/FunctionDescriptors.h"
31 #include "src/core/NEON/NEAsymm.h"
32 #include "src/core/NEON/wrapper/wrapper.h"
33 #include "src/core/helpers/WindowHelpers.h"
34
35 namespace arm_compute
36 {
37 namespace cpu
38 {
39 template <typename T>
directconv3d_quantized_neon_ndhwc(const ITensor * src0,const ITensor * src1,const ITensor * src2,ITensor * dst,const Conv3dInfo & conv_info,const Window & window)40 void directconv3d_quantized_neon_ndhwc(const ITensor *src0, const ITensor *src1, const ITensor *src2, ITensor *dst, const Conv3dInfo &conv_info, const Window &window)
41 {
42 const ITensor *src = src0;
43 const ITensor *weights = src1;
44 const ITensor *biases = src2;
45
46 using vtype = wrapper::traits::neon_bitvector<T, wrapper::traits::BitWidth::W128>;
47 using vector_type = typename vtype::type;
48 using tag_type = typename vtype::tag_type;
49 constexpr int num_elems_read_per_iteration = 16 / sizeof(T);
50 using q16_t = typename wrapper::traits::promote_t<T>;
51 using q32_t = typename wrapper::traits::promote_t<q16_t>;
52 using q32x4_t = typename wrapper::traits::neon_vector<q32_t, 4>::type;
53
54 const int32_t input_offset = -src->info()->quantization_info().uniform().offset;
55 const float input_scale = src->info()->quantization_info().uniform().scale;
56 const int32_t weights_offset = -weights->info()->quantization_info().uniform().offset;
57 const float weights_scale = weights->info()->quantization_info().uniform().scale;
58 const int32_t output_offset = dst->info()->quantization_info().uniform().offset;
59 const float output_scale = dst->info()->quantization_info().uniform().scale;
60
61 int32_t output_multiplier = 0;
62 int32_t output_shift = 0;
63 const float multiplier = input_scale * weights_scale / output_scale;
64 arm_compute::quantization::calculate_quantized_multiplier(multiplier, &output_multiplier, &output_shift);
65
66 // Scalar quantities (N D H W Cin)
67 const int element_size = src->info()->element_size();
68 const int input_stride_w = src->info()->strides_in_bytes().y() / element_size;
69 const int input_stride_h = src->info()->strides_in_bytes().z() / element_size;
70 const int input_stride_d = src->info()->strides_in_bytes()[3] / element_size;
71 const int input_stride_n = src->info()->strides_in_bytes()[4] / element_size;
72 const int input_dim_w = src->info()->dimension(1);
73 const int input_dim_h = src->info()->dimension(2);
74 const int input_dim_d = src->info()->dimension(3);
75
76 // Kernel info (D H W Cin Cout)
77 const unsigned int kernel_stride_w = weights->info()->strides_in_bytes()[2] / element_size;
78 const unsigned int kernel_stride_h = weights->info()->strides_in_bytes()[3] / element_size;
79 const unsigned int kernel_stride_d = weights->info()->strides_in_bytes()[4] / element_size;
80 const int kernel_dim_w = weights->info()->dimension(2);
81 const int kernel_dim_h = weights->info()->dimension(3);
82 const int kernel_dim_d = weights->info()->dimension(4);
83
84 // Convolution padding and stride
85 const int conv_pad_top = conv_info.padding.top;
86 const int conv_pad_left = conv_info.padding.left;
87 const int conv_pad_front = conv_info.padding.front;
88 const int conv_stride_w = conv_info.stride.width;
89 const int conv_stride_h = conv_info.stride.height;
90 const int conv_stride_d = conv_info.stride.depth;
91
92 // Setup input window for the output iterator
93 Window window_out = window;
94 window_out.set(Window::DimX, Window::Dimension(0, 1, 1));
95
96 // Setup input window for the weights iterator
97 Window window_w = calculate_max_window(*weights->info(), Steps());
98 window_w.set(Window::DimY, Window::Dimension(0, 1, 1));
99 window_w.set(Window::DimZ, Window::Dimension(0, 1, 1));
100 window_w.set(Window::DimW, Window::Dimension(0, 1, 1));
101 window_w.set(4, Window::Dimension(0, 1, 1));
102
103 Iterator out(dst, window_out);
104 Iterator wei(weights, window_w);
105
106 const int32_t *biases_ptr = nullptr;
107 if(biases != nullptr)
108 {
109 biases_ptr = reinterpret_cast<int32_t *>(biases->buffer() + biases->info()->offset_first_element_in_bytes());
110 }
111 execute_window_loop(window_out, [&](const Coordinates & id)
112 {
113 // We are computing the theoretical input starting points
114 const int in_w_start_t = static_cast<int>(id.y()) * conv_stride_w - conv_pad_left;
115 const int in_h_start_t = static_cast<int>(id.z()) * conv_stride_h - conv_pad_top;
116 const int in_d_start_t = static_cast<int>(id[3]) * conv_stride_d - conv_pad_front;
117 const int in_w_end_t = in_w_start_t + kernel_dim_w;
118 const int in_h_end_t = in_h_start_t + kernel_dim_h;
119 const int in_d_end_t = in_d_start_t + kernel_dim_d;
120
121 // We are computing the valid initial and ending input points by checking the borders
122 const int in_w_start = std::max(in_w_start_t, 0);
123 const int in_h_start = std::max(in_h_start_t, 0);
124 const int in_d_start = std::max(in_d_start_t, 0);
125 const int in_w_end = std::min(in_w_end_t, input_dim_w);
126 const int in_h_end = std::min(in_h_end_t, input_dim_h);
127 const int in_d_end = std::min(in_d_end_t, input_dim_d);
128
129 // We use the input points to select the valid weight points to use
130 const int wei_w_start = in_w_start - in_w_start_t;
131 const int wei_h_start = in_h_start - in_h_start_t;
132 const int wei_d_start = in_d_start - in_d_start_t;
133 const int wei_w_end = kernel_dim_w - (in_w_end_t - in_w_end);
134 const int wei_h_end = kernel_dim_h - (in_h_end_t - in_h_end);
135 const int wei_d_end = kernel_dim_d - (in_d_end_t - in_d_end);
136
137 const int index_c_out_end = weights->info()->dimension(0);
138 const int index_c_in_end = weights->info()->dimension(1);
139 const T *const in_ptr_start = reinterpret_cast<const T *>(src->buffer() + src->info()->offset_first_element_in_bytes()) + id[4] * input_stride_n;
140
141 execute_window_loop(window_w, [&](const Coordinates & id_w)
142 {
143 /*
144 * This is the loop in the weights, and it goes along OFM (output feature map)
145 */
146 const auto weights_ptr_start = reinterpret_cast<const T *>(wei.ptr());
147 int32_t acc = static_cast<int32_t>(0);
148 T *out_ptr = reinterpret_cast<T *>(out.ptr());
149 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)
150 {
151 const auto in_ptr_d = in_ptr_start + index_in_d * input_stride_d;
152 const auto weights_ptr_d = weights_ptr_start + index_wei_d * kernel_stride_d;
153 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)
154 {
155 const T *const in_ptr_row = in_ptr_d + index_in_h * input_stride_h;
156 const T *const weights_ptr_row = weights_ptr_d + index_wei_h * kernel_stride_h;
157 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)
158 {
159 const T *in_ptr_mover = in_ptr_row + index_in_w * input_stride_w;
160 const T *weights_ptr_mover = weights_ptr_row + index_wei_w * kernel_stride_w;
161 int index_c_in = 0;
162 vector_type w_vec = wrapper::vdup_n(static_cast<T>(0), tag_type());
163
164 q32x4_t acc_q32_0 = wrapper::vdup_n(static_cast<q32_t>(0), tag_type());
165 q32x4_t acc_q32_1 = wrapper::vdup_n(static_cast<q32_t>(0), tag_type());
166 q32x4_t acc_q32_2 = wrapper::vdup_n(static_cast<q32_t>(0), tag_type());
167 q32x4_t acc_q32_3 = wrapper::vdup_n(static_cast<q32_t>(0), tag_type());
168
169 for(; index_c_in <= index_c_in_end - num_elems_read_per_iteration;
170 index_c_in += num_elems_read_per_iteration, in_ptr_mover += num_elems_read_per_iteration)
171 {
172 const auto src_vec = wrapper::vloadq(in_ptr_mover);
173 //Load Cin weights
174 for(int k = 0; k < num_elems_read_per_iteration; ++k, weights_ptr_mover += index_c_out_end)
175 {
176 w_vec = wrapper::vsetlane(*weights_ptr_mover, w_vec, k);
177 }
178 q32x4_t src_q32_0 = wrapper::vdup_n(static_cast<q32_t>(input_offset), tag_type());
179 q32x4_t src_q32_1 = wrapper::vdup_n(static_cast<q32_t>(input_offset), tag_type());
180 q32x4_t src_q32_2 = wrapper::vdup_n(static_cast<q32_t>(input_offset), tag_type());
181 q32x4_t src_q32_3 = wrapper::vdup_n(static_cast<q32_t>(input_offset), tag_type());
182
183 q32x4_t wei_q32_0 = wrapper::vdup_n(static_cast<q32_t>(weights_offset), tag_type());
184 q32x4_t wei_q32_1 = wrapper::vdup_n(static_cast<q32_t>(weights_offset), tag_type());
185 q32x4_t wei_q32_2 = wrapper::vdup_n(static_cast<q32_t>(weights_offset), tag_type());
186 q32x4_t wei_q32_3 = wrapper::vdup_n(static_cast<q32_t>(weights_offset), tag_type());
187
188 const auto src_q16_0 = wrapper::vmovl(wrapper::vgetlow(src_vec));
189 const auto src_q16_1 = wrapper::vmovl(wrapper::vgethigh(src_vec));
190 const auto wei_q16_0 = wrapper::vmovl(wrapper::vgetlow(w_vec));
191 const auto wei_q16_1 = wrapper::vmovl(wrapper::vgethigh(w_vec));
192
193 src_q32_0 = wrapper::vadd(src_q32_0, wrapper::vmovl(wrapper::vgetlow(src_q16_0)));
194 src_q32_1 = wrapper::vadd(src_q32_1, wrapper::vmovl(wrapper::vgethigh(src_q16_0)));
195 src_q32_2 = wrapper::vadd(src_q32_2, wrapper::vmovl(wrapper::vgetlow(src_q16_1)));
196 src_q32_3 = wrapper::vadd(src_q32_3, wrapper::vmovl(wrapper::vgethigh(src_q16_1)));
197
198 wei_q32_0 = wrapper::vadd(wei_q32_0, wrapper::vmovl(wrapper::vgetlow(wei_q16_0)));
199 wei_q32_1 = wrapper::vadd(wei_q32_1, wrapper::vmovl(wrapper::vgethigh(wei_q16_0)));
200 wei_q32_2 = wrapper::vadd(wei_q32_2, wrapper::vmovl(wrapper::vgetlow(wei_q16_1)));
201 wei_q32_3 = wrapper::vadd(wei_q32_3, wrapper::vmovl(wrapper::vgethigh(wei_q16_1)));
202
203 acc_q32_0 = wrapper::vmla(acc_q32_0, wei_q32_0, src_q32_0);
204 acc_q32_1 = wrapper::vmla(acc_q32_1, wei_q32_1, src_q32_1);
205 acc_q32_2 = wrapper::vmla(acc_q32_2, wei_q32_2, src_q32_2);
206 acc_q32_3 = wrapper::vmla(acc_q32_3, wei_q32_3, src_q32_3);
207 }
208 #if defined(__aarch64__)
209 acc += wrapper::vaddv(acc_q32_0);
210 acc += wrapper::vaddv(acc_q32_1);
211 acc += wrapper::vaddv(acc_q32_2);
212 acc += wrapper::vaddv(acc_q32_3);
213 #else // __aarch64__
214 auto temp = wrapper::vpadd(wrapper::vgethigh(acc_q32_0), wrapper::vgetlow(acc_q32_0));
215 temp = wrapper::vpadd(temp, temp);
216 acc += wrapper::vgetlane(temp, 0);
217
218 temp = wrapper::vpadd(wrapper::vgethigh(acc_q32_1), wrapper::vgetlow(acc_q32_1));
219 temp = wrapper::vpadd(temp, temp);
220 acc += wrapper::vgetlane(temp, 0);
221
222 temp = wrapper::vpadd(wrapper::vgethigh(acc_q32_2), wrapper::vgetlow(acc_q32_2));
223 temp = wrapper::vpadd(temp, temp);
224 acc += wrapper::vgetlane(temp, 0);
225
226 temp = wrapper::vpadd(wrapper::vgethigh(acc_q32_3), wrapper::vgetlow(acc_q32_3));
227 temp = wrapper::vpadd(temp, temp);
228 acc += wrapper::vgetlane(temp, 0);
229
230 #endif // __aarch64__
231
232 for(; index_c_in < index_c_in_end; ++index_c_in, ++in_ptr_mover, weights_ptr_mover += index_c_out_end)
233 {
234 const auto src_val = *(in_ptr_mover) + input_offset;
235 const auto w_val = *(weights_ptr_mover) + weights_offset;
236 acc += src_val * w_val;
237 }
238 }
239 }
240 }
241
242 if(biases)
243 {
244 acc += *reinterpret_cast<const int32_t *>(biases_ptr + id_w[0]);
245 }
246
247 T out_val = finalize_quantization(acc, output_multiplier, output_shift, output_offset, T(0), T(0), false);
248 *(reinterpret_cast<T *>(out_ptr + id_w[0])) = out_val;
249 },
250 wei);
251 },
252 out);
253 }
254 } // namespace cpu
255 } // namespace arm_compute
256 #endif // SRC_CORE_NEON_KERNELS_CONV3D_QUANTIZED_H