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 #include "arm_compute/core/Helpers.h"
25 #include "arm_compute/core/Window.h"
26 #include "src/core/NEON/wrapper/wrapper.h"
27
28 #include <cmath>
29 #include <cstddef>
30
31 #include "src/core/NEON/SVEAsymm.h"
32 #include "src/core/NEON/SVEMath.h"
33 #include <arm_sve.h>
34
35 namespace arm_compute
36 {
37 namespace cpu
38 {
sve2_qasymm8_signed_activation(const ITensor * src,ITensor * dst,const ActivationLayerInfo & act_info,const Window & window)39 void sve2_qasymm8_signed_activation(const ITensor *src, ITensor *dst, const ActivationLayerInfo &act_info, const Window &window)
40 {
41 const auto window_start_x = static_cast<int>(window.x().start());
42 const auto window_end_x = static_cast<int>(window.x().end());
43 const ActivationLayerInfo::ActivationFunction act = act_info.activation();
44
45 Window win_collapsed = window.collapse_if_possible(window, Window::DimZ);
46 win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1));
47
48 Iterator input(src, win_collapsed);
49 Iterator output(dst, win_collapsed);
50
51 const UniformQuantizationInfo qi_in = src->info()->quantization_info().uniform();
52 const UniformQuantizationInfo qi_out = dst->info()->quantization_info().uniform();
53 const auto va = svdup_n_s8(quantize_qasymm8_signed(act_info.a(), qi_in));
54 const auto vb = svdup_n_s8(quantize_qasymm8_signed(act_info.b(), qi_in));
55 const auto const_0 = quantize_qasymm8_signed(0.f, qi_in);
56 const auto vconst_0 = svdup_n_s8(const_0);
57 const auto vconst_1 = svdup_n_f32(1.f);
58 const auto va_f32 = svdup_n_f32(act_info.a());
59 const auto vb_f32 = svdup_n_f32(act_info.b());
60 const auto const_6_f32 = svdup_n_f32(6.f);
61 const auto const_0_f32 = svdup_n_f32(0.f);
62 const auto const_3_f32 = svdup_n_f32(3.f);
63 const auto const_inv_6_f32 = svdup_n_f32(0.166666667f);
64
65 // Initialise scale/offset for re-quantization
66 bool requant = true;
67 if(qi_in.scale == qi_out.scale && qi_in.offset == qi_out.offset)
68 {
69 requant = false;
70 }
71 float s = qi_in.scale / qi_out.scale;
72 float o = -qi_in.offset * s + qi_out.offset;
73 auto vs = svdup_n_f32(s);
74 auto vo = svdup_n_f32(o);
75
76 // Initialise scale/offset for re-quantization with int32_t
77 const auto voffset_in = svdup_n_s32(qi_in.offset);
78 int32_t s_s32 = round(s * (1 << 8), arm_compute::RoundingPolicy::TO_NEAREST_EVEN);
79 int32_t o_s32 = round(o * (1 << 8), arm_compute::RoundingPolicy::TO_NEAREST_EVEN);
80 const auto vs_s32 = svdup_n_s32(s_s32);
81 const auto vo_s32 = svdup_n_s32(o_s32);
82
83 // Initialise scale/offset for re-quantization for leaky relu
84 int32_t s_leaky_s32 = round(s * act_info.a() * (1 << 8), arm_compute::RoundingPolicy::TO_NEAREST_EVEN);
85 int32_t o_leaky_s32 = round((-qi_in.offset * s * act_info.a() + qi_out.offset) * (1 << 8),
86 arm_compute::RoundingPolicy::TO_NEAREST_EVEN);
87 const auto vs_leaky_s32 = svdup_n_s32(s_leaky_s32);
88 const auto vo_leaky_s32 = svdup_n_s32(o_leaky_s32);
89
90 execute_window_loop(win_collapsed, [&](const Coordinates &)
91 {
92 const auto input_ptr = reinterpret_cast<const int8_t *>(input.ptr());
93 const auto output_ptr = reinterpret_cast<int8_t *>(output.ptr());
94
95 svint8_t tmp;
96
97 int x = window_start_x;
98 svbool_t pg = svwhilelt_b8(x, window_end_x);
99 do
100 {
101 const auto vin = svld1_s8(pg, input_ptr + x);
102 if(act == ActivationLayerInfo::ActivationFunction::RELU)
103 {
104 // Perform activation
105 tmp = svmax_s8_z(pg, vconst_0, vin);
106 // Re-quantize to new output space
107 tmp = requant ? svmla_qasymm8_signed_z(pg, tmp, vs, vo) : tmp;
108 }
109 else if(act == ActivationLayerInfo::ActivationFunction::BOUNDED_RELU)
110 {
111 // Perform activation
112 tmp = svmin_s8_z(pg, va, svmax_s8_z(pg, vconst_0, vin));
113 // Re-quantize to new output space
114 tmp = requant ? svmla_qasymm8_signed_z(pg, tmp, vs, vo) : tmp;
115 }
116 else if(act == ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU)
117 {
118 // Perform activation
119 tmp = svmin_s8_z(pg, va, svmax_s8_z(pg, vb, vin));
120 // Re-quantize to new output space
121 tmp = requant ? svmla_qasymm8_signed_z(pg, tmp, vs, vo) : tmp;
122 }
123 else if(act == ActivationLayerInfo::ActivationFunction::LOGISTIC)
124 {
125 // De-quantize
126 const auto vin_deq = svdequantize_z(pg, vin, qi_in);
127 // Perform activation
128 const svfloat32x4_t tmp_dep = svcreate4_f32(
129 svdiv_f32_z(pg, vconst_1, svadd_f32_z(pg, vconst_1, svexp_f32_z(pg, svneg_f32_z(pg, svget4_f32(vin_deq, 0))))),
130 svdiv_f32_z(pg, vconst_1, svadd_f32_z(pg, vconst_1, svexp_f32_z(pg, svneg_f32_z(pg, svget4_f32(vin_deq, 1))))),
131 svdiv_f32_z(pg, vconst_1, svadd_f32_z(pg, vconst_1, svexp_f32_z(pg, svneg_f32_z(pg, svget4_f32(vin_deq, 2))))),
132 svdiv_f32_z(pg, vconst_1, svadd_f32_z(pg, vconst_1, svexp_f32_z(pg, svneg_f32_z(pg, svget4_f32(vin_deq, 3))))));
133 // Re-quantize to new output space
134 tmp = svquantize_signed_z(pg, tmp_dep, qi_out);
135 }
136 else if(act == ActivationLayerInfo::ActivationFunction::TANH)
137 {
138 // De-quantize
139 const auto vin_deq = svdequantize_z(pg, vin, qi_in);
140 // Perform activation
141 const svfloat32x4_t tmp_dep = svcreate4_f32(
142 svmul_f32_z(pg, va_f32, svtanh_f32_z(pg, svmul_f32_z(pg, svget4_f32(vin_deq, 0), vb_f32))),
143 svmul_f32_z(pg, va_f32, svtanh_f32_z(pg, svmul_f32_z(pg, svget4_f32(vin_deq, 1), vb_f32))),
144 svmul_f32_z(pg, va_f32, svtanh_f32_z(pg, svmul_f32_z(pg, svget4_f32(vin_deq, 2), vb_f32))),
145 svmul_f32_z(pg, va_f32, svtanh_f32_z(pg, svmul_f32_z(pg, svget4_f32(vin_deq, 3), vb_f32))));
146 // Re-quantize to new output space
147 tmp = svquantize_signed_z(pg, tmp_dep, qi_out);
148 }
149 else if(act == ActivationLayerInfo::ActivationFunction::HARD_SWISH)
150 {
151 // De-quantize
152 const auto vin_deq = svdequantize_z(pg, vin, qi_in);
153 // Perform activation
154 const svfloat32x4_t tmp_dep = svcreate4_f32(
155 svmul_f32_z(pg, svget4_f32(vin_deq, 0), svmul_f32_z(pg, const_inv_6_f32, svmin_f32_z(pg, const_6_f32, svmax_f32_z(pg, const_0_f32, svadd_f32_z(pg, svget4_f32(vin_deq, 0), const_3_f32))))),
156 svmul_f32_z(pg, svget4_f32(vin_deq, 1), svmul_f32_z(pg, const_inv_6_f32, svmin_f32_z(pg, const_6_f32, svmax_f32_z(pg, const_0_f32, svadd_f32_z(pg, svget4_f32(vin_deq, 1), const_3_f32))))),
157 svmul_f32_z(pg, svget4_f32(vin_deq, 2), svmul_f32_z(pg, const_inv_6_f32, svmin_f32_z(pg, const_6_f32, svmax_f32_z(pg, const_0_f32, svadd_f32_z(pg, svget4_f32(vin_deq, 2), const_3_f32))))),
158 svmul_f32_z(pg, svget4_f32(vin_deq, 3), svmul_f32_z(pg, const_inv_6_f32, svmin_f32_z(pg, const_6_f32, svmax_f32_z(pg, const_0_f32, svadd_f32_z(pg, svget4_f32(vin_deq, 3), const_3_f32))))));
159 // Re-quantize to new output space
160 tmp = svquantize_signed_z(pg, tmp_dep, qi_out);
161 }
162 else if(act == ActivationLayerInfo::ActivationFunction::LEAKY_RELU)
163 {
164 svbool_t p0, p1, p2, p3;
165 svint32x4_t tmp_dep;
166
167 // Expand to int32
168 const svint32x4_t vin_s32 = svcreate4_s32(
169 svmovlb_s32(svmovlb_s16(vin)),
170 svmovlt_s32(svmovlb_s16(vin)),
171 svmovlb_s32(svmovlt_s16(vin)),
172 svmovlt_s32(svmovlt_s16(vin)));
173
174 // Compare elements to input offset
175 if(qi_in.scale >= 0)
176 {
177 p0 = svcmplt_s32(pg, svget4_s32(vin_s32, 0), voffset_in);
178 p1 = svcmplt_s32(pg, svget4_s32(vin_s32, 1), voffset_in);
179 p2 = svcmplt_s32(pg, svget4_s32(vin_s32, 2), voffset_in);
180 p3 = svcmplt_s32(pg, svget4_s32(vin_s32, 3), voffset_in);
181 }
182 else
183 {
184 p0 = svcmpgt_s32(pg, svget4_s32(vin_s32, 0), voffset_in);
185 p1 = svcmpgt_s32(pg, svget4_s32(vin_s32, 1), voffset_in);
186 p2 = svcmpgt_s32(pg, svget4_s32(vin_s32, 2), voffset_in);
187 p3 = svcmpgt_s32(pg, svget4_s32(vin_s32, 3), voffset_in);
188 }
189
190 // Multiply negative elements and requantize if necessary
191 if(requant)
192 {
193 tmp_dep = svcreate4_s32(
194 svasr_n_s32_m(pg, svmla_s32_m(pg, svsel(p0, vo_leaky_s32, vo_s32), svget4_s32(vin_s32, 0), svsel(p0, vs_leaky_s32, vs_s32)), 8),
195 svasr_n_s32_m(pg, svmla_s32_m(pg, svsel(p1, vo_leaky_s32, vo_s32), svget4_s32(vin_s32, 1), svsel(p1, vs_leaky_s32, vs_s32)), 8),
196 svasr_n_s32_m(pg, svmla_s32_m(pg, svsel(p2, vo_leaky_s32, vo_s32), svget4_s32(vin_s32, 2), svsel(p2, vs_leaky_s32, vs_s32)), 8),
197 svasr_n_s32_m(pg, svmla_s32_m(pg, svsel(p3, vo_leaky_s32, vo_s32), svget4_s32(vin_s32, 3), svsel(p3, vs_leaky_s32, vs_s32)), 8));
198 }
199 else
200 {
201 tmp_dep = svcreate4_s32(
202 svasr_n_s32_m(p0, svmad_s32_m(p0, svget4_s32(vin_s32, 0), vs_leaky_s32, vo_leaky_s32), 8),
203 svasr_n_s32_m(p1, svmad_s32_m(p1, svget4_s32(vin_s32, 1), vs_leaky_s32, vo_leaky_s32), 8),
204 svasr_n_s32_m(p2, svmad_s32_m(p2, svget4_s32(vin_s32, 2), vs_leaky_s32, vo_leaky_s32), 8),
205 svasr_n_s32_m(p3, svmad_s32_m(p3, svget4_s32(vin_s32, 3), vs_leaky_s32, vo_leaky_s32), 8));
206 }
207
208 // Convert uint32 vectors to uint16 vectors (with saturation)
209 const auto v_low_s16 = svqxtnt_s32(svqxtnb_s32(svget4_s32(tmp_dep, 0)), svget4_s32(tmp_dep, 1));
210 const auto v_high_s16 = svqxtnt_s32(svqxtnb_s32(svget4_s32(tmp_dep, 2)), svget4_s32(tmp_dep, 3));
211
212 // convert uint16 vectors to uint8 vectors (with saturation)
213 tmp = svqxtnt_s16(svqxtnb_s16(v_low_s16), v_high_s16);
214 }
215 else
216 {
217 ARM_COMPUTE_ERROR("Unsupported activation function");
218 }
219
220 svst1_s8(pg, output_ptr + x, tmp);
221
222 x += svcntb();
223 pg = svwhilelt_b8(x, window_end_x);
224
225 }
226 while(svptest_any(svptrue_b8(), pg));
227 },
228 input, output);
229 }
230 } // namespace cpu
231 } // namespace arm_compute
232