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
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_activation(const ITensor * src,ITensor * dst,const ActivationLayerInfo & act_info,const Window & window)39 void sve2_qasymm8_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_u8(quantize_qasymm8(act_info.a(), qi_in));
54 const auto vb = svdup_n_u8(quantize_qasymm8(act_info.b(), qi_in));
55 const auto const_0 = quantize_qasymm8(0.f, qi_in);
56 const auto vconst_0 = svdup_n_u8(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
61
62 // Initialise scale/offset for re-quantization
63 bool requant = true;
64 if(qi_in.scale == qi_out.scale && qi_in.offset == qi_out.offset)
65 {
66 requant = false;
67 }
68 float s = qi_in.scale / qi_out.scale;
69 float o = -qi_in.offset * s + qi_out.offset;
70 auto vs = svdup_n_f32(s);
71 auto vo = svdup_n_f32(o);
72
73 // Initialise scale/offset for re-quantization with int32_t
74 const auto voffset_in = svdup_n_s32(qi_in.offset);
75 int32_t s_s32 = round(s * (1 << 8), arm_compute::RoundingPolicy::TO_NEAREST_EVEN);
76 int32_t o_s32 = round(o * (1 << 8), arm_compute::RoundingPolicy::TO_NEAREST_EVEN);
77 const auto vs_s32 = svdup_n_s32(s_s32);
78 const auto vo_s32 = svdup_n_s32(o_s32);
79
80 // Initialise scale/offset for re-quantization for leaky relu
81 int32_t s_leaky_s32 = round(s * act_info.a() * (1 << 8), arm_compute::RoundingPolicy::TO_NEAREST_EVEN);
82 int32_t o_leaky_s32 = round((-qi_in.offset * s * act_info.a() + qi_out.offset) * (1 << 8),
83 arm_compute::RoundingPolicy::TO_NEAREST_EVEN);
84 const auto vs_leaky_s32 = svdup_n_s32(s_leaky_s32);
85 const auto vo_leaky_s32 = svdup_n_s32(o_leaky_s32);
86
87 execute_window_loop(win_collapsed, [&](const Coordinates &)
88 {
89 const auto input_ptr = reinterpret_cast<const uint8_t *>(input.ptr());
90 const auto output_ptr = reinterpret_cast<uint8_t *>(output.ptr());
91
92 svuint8_t tmp;
93
94 int x = window_start_x;
95 svbool_t pg = svwhilelt_b8(x, window_end_x);
96 do
97 {
98 const auto vin = svld1_u8(pg, input_ptr + x);
99 if(act == ActivationLayerInfo::ActivationFunction::RELU)
100 {
101 // Perform activation
102 tmp = svmax_u8_z(pg, vconst_0, vin);
103 // Re-quantize to new output space
104 tmp = requant ? svmla_qasymm8_z(pg, tmp, vs, vo) : tmp;
105 }
106 else if(act == ActivationLayerInfo::ActivationFunction::BOUNDED_RELU)
107 {
108 // Perform activation
109 tmp = svmin_u8_z(pg, va, svmax_u8_z(pg, vconst_0, vin));
110 // Re-quantize to new output space
111 tmp = requant ? svmla_qasymm8_z(pg, tmp, vs, vo) : tmp;
112 }
113 else if(act == ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU)
114 {
115 // Perform activation
116 tmp = svmin_u8_z(pg, va, svmax_u8_z(pg, vb, vin));
117 // Re-quantize to new output space
118 tmp = svmla_qasymm8_z(pg, tmp, vs, vo);
119 }
120 else if(act == ActivationLayerInfo::ActivationFunction::LOGISTIC)
121 {
122 // De-quantize
123 const auto vin_deq = svdequantize_z(pg, vin, qi_in);
124 // Perform activation
125 const svfloat32x4_t tmp_dep = svcreate4_f32(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))))),
126 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))))),
127 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))))),
128 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))))));
129
130 // Re-quantize to new output space
131 tmp = svquantize_z(pg, tmp_dep, qi_out);
132 }
133 else if(act == ActivationLayerInfo::ActivationFunction::TANH)
134 {
135 // De-quantize
136 const auto vin_deq = svdequantize_z(pg, vin, qi_in);
137 // Perform activation
138 const svfloat32x4_t tmp_dep = svcreate4_f32(svmul_f32_z(pg, va_f32, svtanh_f32_z(pg, svmul_f32_z(pg, svget4_f32(vin_deq, 0), vb_f32))),
139 svmul_f32_z(pg, va_f32, svtanh_f32_z(pg, svmul_f32_z(pg, svget4_f32(vin_deq, 1), vb_f32))),
140 svmul_f32_z(pg, va_f32, svtanh_f32_z(pg, svmul_f32_z(pg, svget4_f32(vin_deq, 2), vb_f32))),
141 svmul_f32_z(pg, va_f32, svtanh_f32_z(pg, svmul_f32_z(pg, svget4_f32(vin_deq, 3), vb_f32))));
142
143 // Re-quantize to new output space
144 tmp = svquantize_z(pg, tmp_dep, qi_out);
145 }
146 else if(act == ActivationLayerInfo::ActivationFunction::LEAKY_RELU)
147 {
148 svbool_t p0, p1, p2, p3;
149 svint32x4_t tmp_dep;
150
151 // Expand to int32
152 const svint32x4_t vin_s32 = svcreate4_s32(
153 svreinterpret_s32_u32(svmovlb_u32(svmovlb_u16(vin))),
154 svreinterpret_s32_u32(svmovlt_u32(svmovlb_u16(vin))),
155 svreinterpret_s32_u32(svmovlb_u32(svmovlt_u16(vin))),
156 svreinterpret_s32_u32(svmovlt_u32(svmovlt_u16(vin))));
157
158 // Compare elements to input offset
159 if(qi_in.scale >= 0)
160 {
161 p0 = svcmplt_s32(pg, svget4_s32(vin_s32, 0), voffset_in);
162 p1 = svcmplt_s32(pg, svget4_s32(vin_s32, 1), voffset_in);
163 p2 = svcmplt_s32(pg, svget4_s32(vin_s32, 2), voffset_in);
164 p3 = svcmplt_s32(pg, svget4_s32(vin_s32, 3), voffset_in);
165 }
166 else
167 {
168 p0 = svcmpgt_s32(pg, svget4_s32(vin_s32, 0), voffset_in);
169 p1 = svcmpgt_s32(pg, svget4_s32(vin_s32, 1), voffset_in);
170 p2 = svcmpgt_s32(pg, svget4_s32(vin_s32, 2), voffset_in);
171 p3 = svcmpgt_s32(pg, svget4_s32(vin_s32, 3), voffset_in);
172 }
173
174 // Multiply negative elements and requantize if necessary
175 if(requant)
176 {
177 tmp_dep = svcreate4_s32(
178 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),
179 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),
180 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),
181 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));
182 }
183 else
184 {
185 tmp_dep = svcreate4_s32(
186 svasr_n_s32_m(p0, svmad_s32_m(p0, svget4_s32(vin_s32, 0), vs_leaky_s32, vo_leaky_s32), 8),
187 svasr_n_s32_m(p1, svmad_s32_m(p1, svget4_s32(vin_s32, 1), vs_leaky_s32, vo_leaky_s32), 8),
188 svasr_n_s32_m(p2, svmad_s32_m(p2, svget4_s32(vin_s32, 2), vs_leaky_s32, vo_leaky_s32), 8),
189 svasr_n_s32_m(p3, svmad_s32_m(p3, svget4_s32(vin_s32, 3), vs_leaky_s32, vo_leaky_s32), 8));
190 }
191
192 // Convert uint32 vectors to uint16 vectors (with saturation)
193 const auto v_low_u16 = svqxtunt_s32(svqxtunb_s32(svget4_s32(tmp_dep, 0)), svget4_s32(tmp_dep, 1));
194 const auto v_high_u16 = svqxtunt_s32(svqxtunb_s32(svget4_s32(tmp_dep, 2)), svget4_s32(tmp_dep, 3));
195
196 // convert uint16 vectors to uint8 vectors (with saturation)
197 tmp = svqxtnt_u16(svqxtnb_u16(v_low_u16), v_high_u16);
198 }
199 else
200 {
201 ARM_COMPUTE_ERROR("Unsupported activation function");
202 }
203
204 svst1_u8(pg, output_ptr + x, tmp);
205
206 x += svcntb();
207 pg = svwhilelt_b8(x, window_end_x);
208
209 }
210 while(svptest_any(svptrue_b8(), pg));
211
212 },
213 input, output);
214 }
215 } // namespace cpu
216 } // namespace arm_compute
217