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/ITensorPack.h"
26 #include "arm_compute/core/Window.h"
27 #include "arm_compute/core/experimental/Types.h"
28 #include "src/core/NEON/NEMath.h"
29 #include "src/core/NEON/NESymm.h"
30 #include "src/core/NEON/wrapper/wrapper.h"
31
32 #include <arm_neon.h>
33 #include <cmath>
34 #include <cstddef>
35
36 namespace arm_compute
37 {
38 namespace cpu
39 {
neon_qsymm16_activation(const ITensor * src,ITensor * dst,const ActivationLayerInfo & act_info,const Window & window)40 void neon_qsymm16_activation(const ITensor *src, ITensor *dst, const ActivationLayerInfo &act_info, const Window &window)
41 {
42 constexpr int window_step_x = 8;
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 auto vconst_1 = vdupq_n_f32(1.f);
56 const float32x4_t va_f32 = vdupq_n_f32(act_info.a());
57 const float32x4_t vb_f32 = vdupq_n_f32(act_info.b());
58 const float a_f32 = act_info.a();
59 const float b_f32 = act_info.b();
60
61 execute_window_loop(win_collapsed, [&](const Coordinates &)
62 {
63 const auto input_ptr = reinterpret_cast<const qsymm16_t *>(input.ptr());
64 const auto output_ptr = reinterpret_cast<qsymm16_t *>(output.ptr());
65
66 wrapper::traits::neon_bitvector_t<qsymm16_t, wrapper::traits::BitWidth::W128> tmp;
67 ARM_COMPUTE_UNUSED(tmp);
68
69 // Compute S elements per iteration
70 int x = window_start_x;
71 for(; x <= (window_end_x - window_step_x); x += window_step_x)
72 {
73 const auto vin = wrapper::vloadq(input_ptr + x);
74 if(act == ActivationLayerInfo::ActivationFunction::LOGISTIC)
75 {
76 // De-quantize
77 const auto vin_deq = vdequantize_int16(vin, qi_in.scale);
78 // Perform activation
79 const float32x4x2_t tmp_dep =
80 {
81 {
82 wrapper::vdiv(vconst_1, wrapper::vadd(vconst_1, wrapper::vexpq(wrapper::vneg(vin_deq.val[0])))),
83 wrapper::vdiv(vconst_1, wrapper::vadd(vconst_1, wrapper::vexpq(wrapper::vneg(vin_deq.val[1])))),
84 }
85 };
86 // Re-quantize to new output space
87 tmp = vquantize_int16(tmp_dep, qi_out.scale);
88 }
89 else if(act == ActivationLayerInfo::ActivationFunction::TANH)
90 {
91 // De-quantize
92 const auto vin_deq = vdequantize_int16(vin, qi_in.scale);
93 // Perform activation
94 const float32x4x2_t tmp_dep =
95 {
96 {
97 wrapper::vmul(va_f32, wrapper::vtanh(wrapper::vmul(vin_deq.val[0], vb_f32))),
98 wrapper::vmul(va_f32, wrapper::vtanh(wrapper::vmul(vin_deq.val[1], vb_f32))),
99 }
100 };
101 // Re-quantize to new output space
102 tmp = vquantize_int16(tmp_dep, qi_out.scale);
103 }
104
105 else if(act == ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU)
106 {
107 // De-quantize
108 const auto vin_deq = vdequantize_int16(vin, qi_in.scale);
109 // Perform activation
110 const float32x4x2_t tmp_dep =
111 {
112 {
113 wrapper::vmin(va_f32, wrapper::vmax(vb_f32, vin_deq.val[0])),
114 wrapper::vmin(va_f32, wrapper::vmax(vb_f32, vin_deq.val[1]))
115 }
116 };
117 // Re-quantize to new output space
118 tmp = vquantize_int16(tmp_dep, qi_out.scale);
119 }
120 else
121 {
122 ARM_COMPUTE_ERROR("Unsupported activation function");
123 }
124 wrapper::vstore(output_ptr + x, tmp);
125 }
126
127 // Compute left-over elements
128 for(; x < window_end_x; ++x)
129 {
130 qsymm16_t in = *(reinterpret_cast<const qsymm16_t *>(input_ptr + x));
131 qsymm16_t tmp = 0;
132 if(act == ActivationLayerInfo::ActivationFunction::LOGISTIC)
133 {
134 float tmp_f = dequantize_qsymm16(in, qi_in.scale);
135 tmp_f = 1.f / (1.f + std::exp(-tmp_f));
136 tmp = quantize_qsymm16(tmp_f, qi_out);
137 }
138 else if(act == ActivationLayerInfo::ActivationFunction::TANH)
139 {
140 float tmp_f = dequantize_qsymm16(in, qi_in.scale);
141 tmp_f = a_f32 * std::tanh(b_f32 * tmp_f);
142 tmp = quantize_qsymm16(tmp_f, qi_out);
143 }
144 else if(act == ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU)
145 {
146 float tmp_f = dequantize_qsymm16(in, qi_in.scale);
147 tmp_f = std::min<float>(a_f32, std::max<float>(b_f32, tmp_f));
148 tmp = quantize_qsymm16(tmp_f, qi_out);
149 }
150 else
151 {
152 ARM_COMPUTE_ERROR("Unsupported activation function");
153 }
154 *(output_ptr + x) = tmp;
155 }
156 },
157 input, output);
158 }
159 } // namespace cpu
160 } // namespace arm_compute
161