1 /*
2 * Copyright (c) 2021-2022 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 "src/cpu/kernels/softmax/generic/sve2/impl.h"
26 #include "arm_compute/core/Types.h"
27 #include "src/core/NEON/wrapper/wrapper.h"
28
29 namespace arm_compute
30 {
31 namespace cpu
32 {
33 template <typename ScalarType>
sve2_softmax_logits_1d_quantized(const ITensor * in,const ITensor * max,void * const tmp,ITensor * out,float beta,bool is_log,const Window & window)34 void sve2_softmax_logits_1d_quantized(const ITensor *in, const ITensor *max, void *const tmp,
35 ITensor *out, float beta, bool is_log, const Window &window)
36 {
37 const int start_x = in->info()->valid_region().anchor.x();
38 const int input_width = in->info()->valid_region().shape.x();
39
40 const float scale_beta = -beta * in->info()->quantization_info().uniform().scale;
41 const auto scale_beta_vec = svdup_n_f32(scale_beta);
42
43 Iterator in_it(in, window);
44 Iterator max_it(max, window);
45 Iterator out_it(out, window);
46 const auto all_true_pg = wrapper::svptrue<ScalarType>();
47 using SVEType = typename wrapper::traits::sve_vector<ScalarType>::type;
48
49 const int inc_1 = static_cast<int>(svcntw());
50 const int inc_2 = static_cast<int>(2 * svcntw());
51 const int inc_3 = static_cast<int>(3 * svcntw());
52
53 execute_window_loop(window, [&](const Coordinates &)
54 {
55 /* Get pointers */
56 const auto in_ptr = reinterpret_cast<const ScalarType *>(in_it.ptr()) + start_x;
57 const auto out_ptr = reinterpret_cast<ScalarType *>(out_it.ptr()) + start_x;
58 const auto tmp_ptr = reinterpret_cast<float *>(tmp);
59
60 float sum{};
61
62 /* Compute exponentials and sum */
63 {
64 /* Get max value */
65 const auto max_val = *reinterpret_cast<const ScalarType *>(max_it.ptr());
66 const auto vec_max = wrapper::svdup_n(max_val);
67
68 /* Init sum to zero */
69 auto vec_sum_0 = svdup_n_f32(0.f);
70 auto vec_sum_1 = svdup_n_f32(0.f);
71 auto vec_sum_2 = svdup_n_f32(0.f);
72 auto vec_sum_3 = svdup_n_f32(0.f);
73
74 /* Loop over row and compute exponentials and sum */
75 int x = 0;
76 svbool_t pg = wrapper::svwhilelt<ScalarType>(x, input_width);
77 svbool_t pg_0 = svunpklo(svunpklo(pg));
78 svbool_t pg_1 = svunpkhi(svunpklo(pg));
79 svbool_t pg_2 = svunpklo(svunpkhi(pg));
80 svbool_t pg_3 = svunpkhi(svunpkhi(pg));
81 do
82 {
83 const auto vec_elements = svld1(pg, in_ptr + x);
84 const auto vec_elements_sub = svreinterpret_u8(svsub_z(pg, vec_max, vec_elements));
85
86 auto vec_elements_flt_0 = svcvt_f32_z(pg_0, svunpklo(svunpklo(vec_elements_sub)));
87 auto vec_elements_flt_1 = svcvt_f32_z(pg_1, svunpkhi(svunpklo(vec_elements_sub)));
88 auto vec_elements_flt_2 = svcvt_f32_z(pg_2, svunpklo(svunpkhi(vec_elements_sub)));
89 auto vec_elements_flt_3 = svcvt_f32_z(pg_3, svunpkhi(svunpkhi(vec_elements_sub)));
90
91 if(is_log)
92 {
93 vec_elements_flt_0 = svmul_f32_z(pg_0, vec_elements_flt_0, scale_beta_vec);
94 vec_elements_flt_1 = svmul_f32_z(pg_1, vec_elements_flt_1, scale_beta_vec);
95 vec_elements_flt_2 = svmul_f32_z(pg_2, vec_elements_flt_2, scale_beta_vec);
96 vec_elements_flt_3 = svmul_f32_z(pg_3, vec_elements_flt_3, scale_beta_vec);
97 vec_sum_0 = svadd_f32_m(pg_0, vec_sum_0, svexp_f32_z(pg_0, vec_elements_flt_0));
98 vec_sum_1 = svadd_f32_m(pg_1, vec_sum_1, svexp_f32_z(pg_1, vec_elements_flt_1));
99 vec_sum_2 = svadd_f32_m(pg_2, vec_sum_2, svexp_f32_z(pg_2, vec_elements_flt_2));
100 vec_sum_3 = svadd_f32_m(pg_3, vec_sum_3, svexp_f32_z(pg_3, vec_elements_flt_3));
101 }
102 else
103 {
104 vec_elements_flt_0 = svexp_f32_z(pg_0, svmul_f32_z(pg_0, vec_elements_flt_0, scale_beta_vec));
105 vec_elements_flt_1 = svexp_f32_z(pg_1, svmul_f32_z(pg_1, vec_elements_flt_1, scale_beta_vec));
106 vec_elements_flt_2 = svexp_f32_z(pg_2, svmul_f32_z(pg_2, vec_elements_flt_2, scale_beta_vec));
107 vec_elements_flt_3 = svexp_f32_z(pg_3, svmul_f32_z(pg_3, vec_elements_flt_3, scale_beta_vec));
108 vec_sum_0 = svadd_f32_m(pg_0, vec_sum_0, vec_elements_flt_0);
109 vec_sum_1 = svadd_f32_m(pg_1, vec_sum_1, vec_elements_flt_1);
110 vec_sum_2 = svadd_f32_m(pg_2, vec_sum_2, vec_elements_flt_2);
111 vec_sum_3 = svadd_f32_m(pg_3, vec_sum_3, vec_elements_flt_3);
112 }
113
114 svst1_f32(pg_0, tmp_ptr + x, vec_elements_flt_0);
115 svst1_f32(pg_1, tmp_ptr + x + inc_1, vec_elements_flt_1);
116 svst1_f32(pg_2, tmp_ptr + x + inc_2, vec_elements_flt_2);
117 svst1_f32(pg_3, tmp_ptr + x + inc_3, vec_elements_flt_3);
118
119 x += wrapper::svcnt<ScalarType>();
120 pg = wrapper::svwhilelt<ScalarType>(x, input_width);
121 pg_0 = svunpklo(svunpklo(pg));
122 pg_1 = svunpkhi(svunpklo(pg));
123 pg_2 = svunpklo(svunpkhi(pg));
124 pg_3 = svunpkhi(svunpkhi(pg));
125 }
126 while(svptest_any(all_true_pg, pg));
127
128 /* Reduce sum */
129 const auto vec_sum = svadd_f32_z(all_true_pg, svadd_f32_z(all_true_pg, vec_sum_0, vec_sum_1), svadd_f32_z(all_true_pg, vec_sum_2, vec_sum_3));
130 sum = svaddv_f32(all_true_pg, vec_sum);
131
132 /* Run remaining elements */
133 x = 0;
134 if(is_log)
135 {
136 sum = std::log(sum);
137 }
138 else
139 {
140 sum = 256.f / sum;
141 }
142 }
143
144 /* Normalize exponentials */
145 {
146 constexpr bool is_qasymm8_signed = std::is_same<ScalarType, qasymm8_signed_t>::value;
147 /* Loop over row and compute softmax */
148 int x = 0;
149 svbool_t pg = wrapper::svwhilelt<ScalarType>(x, input_width);
150 svbool_t pg_0 = svunpklo(svunpklo(pg));
151 svbool_t pg_1 = svunpkhi(svunpklo(pg));
152 svbool_t pg_2 = svunpklo(svunpkhi(pg));
153 svbool_t pg_3 = svunpkhi(svunpkhi(pg));
154 do
155 {
156 auto vec_in_0 = svld1_f32(pg_0, tmp_ptr + x);
157 auto vec_in_1 = svld1_f32(pg_1, tmp_ptr + x + inc_1);
158 auto vec_in_2 = svld1_f32(pg_2, tmp_ptr + x + inc_2);
159 auto vec_in_3 = svld1_f32(pg_3, tmp_ptr + x + inc_3);
160
161 svfloat32_t res_0{};
162 svfloat32_t res_1{};
163 svfloat32_t res_2{};
164 svfloat32_t res_3{};
165
166 if(is_log)
167 {
168 res_0 = svsub_f32_z(pg_0, vec_in_0, svdup_n_f32(sum));
169 res_1 = svsub_f32_z(pg_1, vec_in_1, svdup_n_f32(sum));
170 res_2 = svsub_f32_z(pg_2, vec_in_2, svdup_n_f32(sum));
171 res_3 = svsub_f32_z(pg_3, vec_in_3, svdup_n_f32(sum));
172 }
173 else
174 {
175 res_0 = svmul_f32_z(pg_0, vec_in_0, svdup_n_f32(sum));
176 res_1 = svmul_f32_z(pg_1, vec_in_1, svdup_n_f32(sum));
177 res_2 = svmul_f32_z(pg_2, vec_in_2, svdup_n_f32(sum));
178 res_3 = svmul_f32_z(pg_3, vec_in_3, svdup_n_f32(sum));
179
180 if(is_qasymm8_signed)
181 {
182 const auto offset_vec = svdup_n_f32(128.f);
183 res_0 = svsub_z(pg_0, res_0, offset_vec);
184 res_1 = svsub_z(pg_1, res_1, offset_vec);
185 res_2 = svsub_z(pg_2, res_2, offset_vec);
186 res_3 = svsub_z(pg_3, res_3, offset_vec);
187 }
188 }
189
190 // Store value
191 const auto out = convert_float_to_int<SVEType>(res_0, res_1, res_2, res_3);
192 svst1(pg, out_ptr + x, out);
193 x += wrapper::svcnt<ScalarType>();
194 pg = wrapper::svwhilelt<ScalarType>(x, input_width);
195 pg_0 = svunpklo(svunpklo(pg));
196 pg_1 = svunpkhi(svunpklo(pg));
197 pg_2 = svunpklo(svunpkhi(pg));
198 pg_3 = svunpkhi(svunpkhi(pg));
199 }
200 while(svptest_any(all_true_pg, pg));
201 }
202 },
203 in_it, max_it, out_it);
204 }
205
206 template void sve2_softmax_logits_1d_quantized<qasymm8_signed_t>(const ITensor *in, const ITensor *max, void *const tmp,
207 ITensor *out, float beta, bool is_log, const Window &window);
208 template void sve2_softmax_logits_1d_quantized<qasymm8_t>(const ITensor *in, const ITensor *max, void *const tmp,
209 ITensor *out, float beta, bool is_log, const Window &window);
210 } // namespace cpu
211 } // namespace arm_compute
212