1 /*
2 * Copyright (c) 2020-2022 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
25 #include "arm_compute/core/Helpers.h"
26 #include "arm_compute/core/Window.h"
27 #include "src/core/NEON/NEAsymm.h"
28 #include "src/core/NEON/NEMath.h"
29 #include "src/core/NEON/wrapper/wrapper.h"
30
31 #include <arm_neon.h>
32 #include <cmath>
33 #include <cstddef>
34 #include <cstdint>
35
36 namespace arm_compute
37 {
38 namespace cpu
39 {
neon_qasymm8_activation(const ITensor * src,ITensor * dst,const ActivationLayerInfo & act_info,const Window & window)40 void neon_qasymm8_activation(const ITensor *src, ITensor *dst, const ActivationLayerInfo &act_info, const Window &window)
41 {
42 constexpr int window_step_x = 16;
43 const auto window_start_x = static_cast<int>(window.x().start());
44 const auto window_end_x = static_cast<int>(window.x().end());
45 const ActivationLayerInfo::ActivationFunction act = act_info.activation();
46
47 Window win_collapsed = window.collapse_if_possible(window, Window::DimZ);
48 win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1));
49
50 Iterator input(src, win_collapsed);
51 Iterator output(dst, win_collapsed);
52
53 const UniformQuantizationInfo qi_in = src->info()->quantization_info().uniform();
54 const UniformQuantizationInfo qi_out = dst->info()->quantization_info().uniform();
55 const qasymm8x16_t va = vdupq_n_u8(quantize_qasymm8(act_info.a(), qi_in));
56 const qasymm8x16_t vb = vdupq_n_u8(quantize_qasymm8(act_info.b(), qi_in));
57 const qasymm8_t a = quantize_qasymm8(act_info.a(), qi_in);
58 const qasymm8_t b = quantize_qasymm8(act_info.b(), qi_in);
59 const qasymm8_t const_0 = quantize_qasymm8(0.f, qi_in);
60 const qasymm8x16_t vconst_0 = vdupq_n_u8(const_0);
61 const auto vconst_1 = vdupq_n_f32(1.f);
62
63 #ifndef __aarch64__
64 const auto vconst_0_f32 = vdupq_n_f32(0);
65 #else // #ifndef __aarch64__
66 const auto const_inv_2 = vdupq_n_f32(0.5f);
67 const auto const_inv_sqrt_2 = vdupq_n_f32(0.70710678118f);
68 #endif // __aarch64__
69 const float32x4_t va_f32 = vdupq_n_f32(act_info.a());
70 const float32x4_t vb_f32 = vdupq_n_f32(act_info.b());
71 const float a_f32 = act_info.a();
72 const float b_f32 = act_info.b();
73
74 #ifndef __aarch64__
75 const auto const_6_f32 = vdupq_n_f32(6.f);
76 const auto const_0_f32 = vdupq_n_f32(0.f);
77 const auto const_3_f32 = vdupq_n_f32(3.f);
78 const auto const_inv_6_f32 = vdupq_n_f32(0.166666667f);
79 #endif // __aarch64__
80
81 // Initialise scale/offset for re-quantization
82 float s = qi_in.scale / qi_out.scale;
83 float o = -qi_in.offset * s + qi_out.offset;
84 float32x4_t vs = vdupq_n_f32(s);
85 float32x4_t vo = vdupq_n_f32(o);
86
87 execute_window_loop(win_collapsed, [&](const Coordinates &)
88 {
89 const auto input_ptr = reinterpret_cast<const qasymm8_t *>(input.ptr());
90 const auto output_ptr = reinterpret_cast<qasymm8_t *>(output.ptr());
91
92 wrapper::traits::neon_bitvector_t<qasymm8_t, wrapper::traits::BitWidth::W128> tmp;
93
94 // Compute S elements per iteration
95 int x = window_start_x;
96 for(; x <= (window_end_x - window_step_x); x += window_step_x)
97 {
98 const auto vin = wrapper::vloadq(input_ptr + x);
99 if(act == ActivationLayerInfo::ActivationFunction::RELU)
100 {
101 // Perform activation
102 tmp = vmaxq_u8(vconst_0, vin);
103 // Re-quantize to new output space
104 tmp = vmlaq_qasymm8(tmp, vs, vo);
105 }
106 else if(act == ActivationLayerInfo::ActivationFunction::BOUNDED_RELU)
107 {
108 // Perform activation
109 tmp = vminq_u8(va, vmaxq_u8(vconst_0, vin));
110 // Re-quantize to new output space
111 tmp = vmlaq_qasymm8(tmp, vs, vo);
112 }
113 else if(act == ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU)
114 {
115 // Perform activation
116 tmp = vminq_u8(va, vmaxq_u8(vb, vin));
117 // Re-quantize to new output space
118 tmp = vmlaq_qasymm8(tmp, vs, vo);
119 }
120 #ifndef __aarch64__ // LUT-based implementation is used for aarch64 instead.
121 else if(act == ActivationLayerInfo::ActivationFunction::LOGISTIC)
122 {
123 // De-quantize
124 const auto vin_deq = vdequantize(vin, qi_in);
125 // Perform activation
126 const float32x4x4_t tmp_dep =
127 {
128 {
129 wrapper::vdiv(vconst_1, wrapper::vadd(vconst_1, wrapper::vexpq(wrapper::vneg(vin_deq.val[0])))),
130 wrapper::vdiv(vconst_1, wrapper::vadd(vconst_1, wrapper::vexpq(wrapper::vneg(vin_deq.val[1])))),
131 wrapper::vdiv(vconst_1, wrapper::vadd(vconst_1, wrapper::vexpq(wrapper::vneg(vin_deq.val[2])))),
132 wrapper::vdiv(vconst_1, wrapper::vadd(vconst_1, wrapper::vexpq(wrapper::vneg(vin_deq.val[3])))),
133 }
134 };
135 // Re-quantize to new output space
136 tmp = vquantize(tmp_dep, qi_out);
137 }
138 #endif // __aarch64__
139 else if(act == ActivationLayerInfo::ActivationFunction::TANH)
140 {
141 // De-quantize
142 const auto vin_deq = vdequantize(vin, qi_in);
143 // Perform activation
144 const float32x4x4_t tmp_dep =
145 {
146 {
147 wrapper::vmul(va_f32, wrapper::vtanh(wrapper::vmul(vin_deq.val[0], vb_f32))),
148 wrapper::vmul(va_f32, wrapper::vtanh(wrapper::vmul(vin_deq.val[1], vb_f32))),
149 wrapper::vmul(va_f32, wrapper::vtanh(wrapper::vmul(vin_deq.val[2], vb_f32))),
150 wrapper::vmul(va_f32, wrapper::vtanh(wrapper::vmul(vin_deq.val[3], vb_f32))),
151 }
152 };
153 // Re-quantize to new output space
154 tmp = vquantize(tmp_dep, qi_out);
155 }
156 #ifndef __aarch64__ // LUT-based implementation is used for aarch64 instead.
157 else if(act == ActivationLayerInfo::ActivationFunction::HARD_SWISH)
158 {
159 // De-quantize
160 const auto vin_deq = vdequantize(vin, qi_in);
161 // Perform activation
162 const float32x4x4_t tmp_dep =
163 {
164 {
165 wrapper::vmul(vin_deq.val[0], wrapper::vmul(const_inv_6_f32, wrapper::vmin(const_6_f32, wrapper::vmax(const_0_f32, wrapper::vadd(vin_deq.val[0], const_3_f32))))),
166 wrapper::vmul(vin_deq.val[1], wrapper::vmul(const_inv_6_f32, wrapper::vmin(const_6_f32, wrapper::vmax(const_0_f32, wrapper::vadd(vin_deq.val[1], const_3_f32))))),
167 wrapper::vmul(vin_deq.val[2], wrapper::vmul(const_inv_6_f32, wrapper::vmin(const_6_f32, wrapper::vmax(const_0_f32, wrapper::vadd(vin_deq.val[2], const_3_f32))))),
168 wrapper::vmul(vin_deq.val[3], wrapper::vmul(const_inv_6_f32, wrapper::vmin(const_6_f32, wrapper::vmax(const_0_f32, wrapper::vadd(vin_deq.val[3], const_3_f32))))),
169 }
170 };
171 // Re-quantize to new output space
172 tmp = vquantize(tmp_dep, qi_out);
173 }
174 else if(act == ActivationLayerInfo::ActivationFunction::LEAKY_RELU)
175 {
176 const auto vin_deq = vdequantize(vin, qi_in);
177
178 const uint32x4x4_t pos_mask =
179 {
180 {
181 wrapper::vcgt(vin_deq.val[0], vconst_0_f32),
182 wrapper::vcgt(vin_deq.val[1], vconst_0_f32),
183 wrapper::vcgt(vin_deq.val[2], vconst_0_f32),
184 wrapper::vcgt(vin_deq.val[3], vconst_0_f32),
185 }
186 };
187
188 const float32x4x4_t tmp_dep =
189 {
190 {
191 wrapper::vbsl(pos_mask.val[0], vin_deq.val[0], wrapper::vmul(va_f32, vin_deq.val[0])),
192 wrapper::vbsl(pos_mask.val[1], vin_deq.val[1], wrapper::vmul(va_f32, vin_deq.val[1])),
193 wrapper::vbsl(pos_mask.val[2], vin_deq.val[2], wrapper::vmul(va_f32, vin_deq.val[2])),
194 wrapper::vbsl(pos_mask.val[3], vin_deq.val[3], wrapper::vmul(va_f32, vin_deq.val[3])),
195 }
196 };
197
198 tmp = vquantize(tmp_dep, qi_out);
199 }
200 #else // #ifndef __aarch64__
201 else if (act == ActivationLayerInfo::ActivationFunction::GELU)
202 {
203 const auto vin_deq = vdequantize(vin, qi_in);
204 // Perform activation
205 const float32x4x4_t tmp_dep =
206 {
207 {
208 wrapper::vmul(vin_deq.val[0], wrapper::vmul(const_inv_2, wrapper::vadd(vconst_1, wrapper::verf(wrapper::vmul(vin_deq.val[0], const_inv_sqrt_2))))),
209 wrapper::vmul(vin_deq.val[1], wrapper::vmul(const_inv_2, wrapper::vadd(vconst_1, wrapper::verf(wrapper::vmul(vin_deq.val[1], const_inv_sqrt_2))))),
210 wrapper::vmul(vin_deq.val[2], wrapper::vmul(const_inv_2, wrapper::vadd(vconst_1, wrapper::verf(wrapper::vmul(vin_deq.val[2], const_inv_sqrt_2))))),
211 wrapper::vmul(vin_deq.val[3], wrapper::vmul(const_inv_2, wrapper::vadd(vconst_1, wrapper::verf(wrapper::vmul(vin_deq.val[3], const_inv_sqrt_2))))),
212 }
213 };
214 // Re-quantize to new output space
215 tmp = vquantize(tmp_dep, qi_out);
216 }
217 #endif // __aarch64__
218 else
219 {
220 ARM_COMPUTE_ERROR("Unsupported activation function");
221 }
222 wrapper::vstore(output_ptr + x, tmp);
223 }
224
225 // Compute left-over elements
226 for(; x < window_end_x; ++x)
227 {
228 qasymm8_t in = *(reinterpret_cast<const qasymm8_t *>(input_ptr + x));
229 qasymm8_t tmp = 0;
230 if(act == ActivationLayerInfo::ActivationFunction::RELU)
231 {
232 tmp = std::max(const_0, in);
233 tmp = utility::clamp<int32_t, qasymm8_t>(tmp * s + o);
234 }
235 else if(act == ActivationLayerInfo::ActivationFunction::BOUNDED_RELU)
236 {
237 tmp = std::min(a, std::max(const_0, in));
238 tmp = utility::clamp<int32_t, qasymm8_t>(tmp * s + o);
239 }
240 else if(act == ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU)
241 {
242 tmp = std::min(a, std::max(b, in));
243 tmp = utility::clamp<int32_t, qasymm8_t>(tmp * s + o);
244 }
245 #ifndef __aarch64__ // LUT-based implementation is used for aarch64 instead.
246 else if(act == ActivationLayerInfo::ActivationFunction::LOGISTIC)
247 {
248 float tmp_f = dequantize_qasymm8(in, qi_in);
249 tmp_f = 1.f / (1.f + std::exp(-tmp_f));
250 tmp = quantize_qasymm8(tmp_f, qi_out);
251 }
252 #endif // __aarch64__
253 else if(act == ActivationLayerInfo::ActivationFunction::TANH)
254 {
255 float tmp_f = dequantize_qasymm8(in, qi_in);
256 tmp_f = a_f32 * std::tanh(b_f32 * tmp_f);
257 tmp = quantize_qasymm8(tmp_f, qi_out);
258 }
259 #ifndef __aarch64__ // LUT-based implementation is used for aarch64 instead.
260 else if(act == ActivationLayerInfo::ActivationFunction::HARD_SWISH)
261 {
262 float tmp_f = dequantize_qasymm8(in, qi_in);
263 tmp_f = tmp_f * ((std::min(std::max((tmp_f + 3), 0.0f), 6.0f)) * 0.166666667f);
264 tmp = quantize_qasymm8(tmp_f, qi_out);
265 }
266 else if(act == ActivationLayerInfo::ActivationFunction::LEAKY_RELU)
267 {
268 float tmp_f = dequantize_qasymm8(in, qi_in);
269 tmp_f = tmp_f > 0 ? tmp_f : tmp_f * a_f32;
270 tmp = quantize_qasymm8(tmp_f, qi_out);
271 }
272 else if(act == ActivationLayerInfo::ActivationFunction::GELU)
273 {
274 float tmp_f = dequantize_qasymm8(in, qi_in);
275 tmp = tmp_f * 0.5f * (1.0f + std::erff(in / 1.41421356237f));
276 tmp = quantize_qasymm8(tmp_f, qi_out);
277 }
278 #endif // __aarch64__
279 else
280 {
281 ARM_COMPUTE_ERROR("Unsupported activation function");
282 }
283 *(output_ptr + x) = tmp;
284 }
285 },
286 input, output);
287 }
288 } // namespace cpu
289 } // namespace arm_compute
290