1 // Auto-generated file. Do not edit!
2 // Template: src/f32-vscaleextexp/avx2-p5.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2019 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9
10 #include <assert.h>
11
12 #include <immintrin.h>
13
14 #include <xnnpack/common.h>
15 #include <xnnpack/vscaleextexp.h>
16
17
18 static const int32_t mask_table[14] = {-1, -1, -1, -1, -1, -1, -1, 0, 0, 0, 0, 0, 0, 0};
19
xnn_f32_vscaleextexp_ukernel__avx2_p5_x56(size_t elements,const float * x,float * y,float scale_value,float scale_exp)20 void xnn_f32_vscaleextexp_ukernel__avx2_p5_x56(
21 size_t elements,
22 const float* x,
23 float* y,
24 float scale_value,
25 float scale_exp)
26 {
27 assert(elements % sizeof(float) == 0);
28
29 const __m256 vlog2e = _mm256_set1_ps(0x1.715476p+0f);
30 const __m256 vminus_ln2_hi = _mm256_set1_ps(-0x1.62E43p-1f);
31 const __m256 vminus_ln2_lo = _mm256_set1_ps(0x1.05C61p-29f);
32
33 // The smallest elements such that 2**elements is considered non-negligible.
34 // For smaller elements, 2**elements is replaced with zero.
35 const __m256 vmin_exponent = _mm256_set1_ps(-127.0f);
36 const __m256 vmagic_bias = _mm256_set1_ps(0x1.8000FEp23f);
37
38 const __m256 vc0 = _mm256_set1_ps(1.0f);
39 const __m256 vc1 = _mm256_set1_ps(0x1.FFFFF6p-1f);
40 const __m256 vc2 = _mm256_set1_ps(0x1.FFFDC6p-2f);
41 const __m256 vc3 = _mm256_set1_ps(0x1.555A80p-3f);
42 const __m256 vc4 = _mm256_set1_ps(0x1.573A1Ap-5f);
43 const __m256 vc5 = _mm256_set1_ps(0x1.0F9F9Cp-7f);
44
45 const __m256 vscalev = _mm256_set1_ps(scale_value);
46 const __m256 vscalee = _mm256_set1_ps(scale_exp);
47
48 for (; elements >= 56 * sizeof(float); elements -= 56 * sizeof(float)) {
49 // Load 56 (7x8) inputs at a time.
50 const __m256 vx0 = _mm256_loadu_ps(x);
51 const __m256 vx1 = _mm256_loadu_ps(x + 8);
52 const __m256 vx2 = _mm256_loadu_ps(x + 16);
53 const __m256 vx3 = _mm256_loadu_ps(x + 24);
54 const __m256 vx4 = _mm256_loadu_ps(x + 32);
55 const __m256 vx5 = _mm256_loadu_ps(x + 40);
56 const __m256 vx6 = _mm256_loadu_ps(x + 48);
57 x += 56;
58
59 // Compute reduced argument elements := round(x / log(2)).
60 const __m256 vn0 = _mm256_round_ps(_mm256_mul_ps(vx0, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
61 const __m256 vn1 = _mm256_round_ps(_mm256_mul_ps(vx1, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
62 const __m256 vn2 = _mm256_round_ps(_mm256_mul_ps(vx2, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
63 const __m256 vn3 = _mm256_round_ps(_mm256_mul_ps(vx3, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
64 const __m256 vn4 = _mm256_round_ps(_mm256_mul_ps(vx4, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
65 const __m256 vn5 = _mm256_round_ps(_mm256_mul_ps(vx5, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
66 const __m256 vn6 = _mm256_round_ps(_mm256_mul_ps(vx6, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
67
68 // Compute reduced argument t := x - elements * log(2).
69 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
70 __m256 vt0 = _mm256_fmadd_ps(vn0, vminus_ln2_hi, vx0);
71 __m256 vt1 = _mm256_fmadd_ps(vn1, vminus_ln2_hi, vx1);
72 __m256 vt2 = _mm256_fmadd_ps(vn2, vminus_ln2_hi, vx2);
73 __m256 vt3 = _mm256_fmadd_ps(vn3, vminus_ln2_hi, vx3);
74 __m256 vt4 = _mm256_fmadd_ps(vn4, vminus_ln2_hi, vx4);
75 __m256 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_hi, vx5);
76 __m256 vt6 = _mm256_fmadd_ps(vn6, vminus_ln2_hi, vx6);
77
78 vt0 = _mm256_fmadd_ps(vn0, vminus_ln2_lo, vt0);
79 vt1 = _mm256_fmadd_ps(vn1, vminus_ln2_lo, vt1);
80 vt2 = _mm256_fmadd_ps(vn2, vminus_ln2_lo, vt2);
81 vt3 = _mm256_fmadd_ps(vn3, vminus_ln2_lo, vt3);
82 vt4 = _mm256_fmadd_ps(vn4, vminus_ln2_lo, vt4);
83 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_lo, vt5);
84 vt6 = _mm256_fmadd_ps(vn6, vminus_ln2_lo, vt6);
85
86 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
87 __m256 vp0 = _mm256_fmadd_ps(vc5, vt0, vc4);
88 __m256 vp1 = _mm256_fmadd_ps(vc5, vt1, vc4);
89 __m256 vp2 = _mm256_fmadd_ps(vc5, vt2, vc4);
90 __m256 vp3 = _mm256_fmadd_ps(vc5, vt3, vc4);
91 __m256 vp4 = _mm256_fmadd_ps(vc5, vt4, vc4);
92 __m256 vp5 = _mm256_fmadd_ps(vc5, vt5, vc4);
93 __m256 vp6 = _mm256_fmadd_ps(vc5, vt6, vc4);
94
95 vp0 = _mm256_fmadd_ps(vp0, vt0, vc3);
96 vp1 = _mm256_fmadd_ps(vp1, vt1, vc3);
97 vp2 = _mm256_fmadd_ps(vp2, vt2, vc3);
98 vp3 = _mm256_fmadd_ps(vp3, vt3, vc3);
99 vp4 = _mm256_fmadd_ps(vp4, vt4, vc3);
100 vp5 = _mm256_fmadd_ps(vp5, vt5, vc3);
101 vp6 = _mm256_fmadd_ps(vp6, vt6, vc3);
102
103 vp0 = _mm256_fmadd_ps(vp0, vt0, vc2);
104 vp1 = _mm256_fmadd_ps(vp1, vt1, vc2);
105 vp2 = _mm256_fmadd_ps(vp2, vt2, vc2);
106 vp3 = _mm256_fmadd_ps(vp3, vt3, vc2);
107 vp4 = _mm256_fmadd_ps(vp4, vt4, vc2);
108 vp5 = _mm256_fmadd_ps(vp5, vt5, vc2);
109 vp6 = _mm256_fmadd_ps(vp6, vt6, vc2);
110
111 vp0 = _mm256_fmadd_ps(vp0, vt0, vc1);
112 vp1 = _mm256_fmadd_ps(vp1, vt1, vc1);
113 vp2 = _mm256_fmadd_ps(vp2, vt2, vc1);
114 vp3 = _mm256_fmadd_ps(vp3, vt3, vc1);
115 vp4 = _mm256_fmadd_ps(vp4, vt4, vc1);
116 vp5 = _mm256_fmadd_ps(vp5, vt5, vc1);
117 vp6 = _mm256_fmadd_ps(vp6, vt6, vc1);
118
119 vp0 = _mm256_fmadd_ps(vp0, vt0, vc0);
120 vp1 = _mm256_fmadd_ps(vp1, vt1, vc0);
121 vp2 = _mm256_fmadd_ps(vp2, vt2, vc0);
122 vp3 = _mm256_fmadd_ps(vp3, vt3, vc0);
123 vp4 = _mm256_fmadd_ps(vp4, vt4, vc0);
124 vp5 = _mm256_fmadd_ps(vp5, vt5, vc0);
125 vp6 = _mm256_fmadd_ps(vp6, vt6, vc0);
126
127 // Multiply "extended" floating-point numbers in ("mantissa", "exponent") representation where
128 // - vnX is "exponent"
129 // - vpX is "mantissa"
130 //
131 // exp2(ae) * av * exp2(be) * bv =
132 // = exp2(ae + be) * (av * bv)
133 __m256 vf0 = _mm256_mul_ps(vp0, vscalev);
134 __m256 vf1 = _mm256_mul_ps(vp1, vscalev);
135 __m256 vf2 = _mm256_mul_ps(vp2, vscalev);
136 __m256 vf3 = _mm256_mul_ps(vp3, vscalev);
137 __m256 vf4 = _mm256_mul_ps(vp4, vscalev);
138 __m256 vf5 = _mm256_mul_ps(vp5, vscalev);
139 __m256 vf6 = _mm256_mul_ps(vp6, vscalev);
140
141 __m256 ve0 = _mm256_add_ps(vn0, vscalee);
142 __m256 ve1 = _mm256_add_ps(vn1, vscalee);
143 __m256 ve2 = _mm256_add_ps(vn2, vscalee);
144 __m256 ve3 = _mm256_add_ps(vn3, vscalee);
145 __m256 ve4 = _mm256_add_ps(vn4, vscalee);
146 __m256 ve5 = _mm256_add_ps(vn5, vscalee);
147 __m256 ve6 = _mm256_add_ps(vn6, vscalee);
148
149 // For computational efficiency, replace exp2(e) with 0.0f when e <= -127.0.
150 // This replacement is done in two steps:
151 // 1. Clamp minimum e at -127.0.
152 // 2. Map e to scale factor 0.0 when e == -127.0
153 ve0 = _mm256_max_ps(ve0, vmin_exponent);
154 ve1 = _mm256_max_ps(ve1, vmin_exponent);
155 ve2 = _mm256_max_ps(ve2, vmin_exponent);
156 ve3 = _mm256_max_ps(ve3, vmin_exponent);
157 ve4 = _mm256_max_ps(ve4, vmin_exponent);
158 ve5 = _mm256_max_ps(ve5, vmin_exponent);
159 ve6 = _mm256_max_ps(ve6, vmin_exponent);
160
161 // Convert exponents into scale factors:
162 // - s = exp2(e) when e > -127.0
163 // - s = 0.0 when e <= -127.0
164 const __m256 vs0 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve0, vmagic_bias)), 23));
165 const __m256 vs1 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve1, vmagic_bias)), 23));
166 const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve2, vmagic_bias)), 23));
167 const __m256 vs3 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve3, vmagic_bias)), 23));
168 const __m256 vs4 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve4, vmagic_bias)), 23));
169 const __m256 vs5 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve5, vmagic_bias)), 23));
170 const __m256 vs6 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve6, vmagic_bias)), 23));
171
172 // Multiply "mantissa" by the scale factor.
173 vf0 = _mm256_mul_ps(vf0, vs0);
174 vf1 = _mm256_mul_ps(vf1, vs1);
175 vf2 = _mm256_mul_ps(vf2, vs2);
176 vf3 = _mm256_mul_ps(vf3, vs3);
177 vf4 = _mm256_mul_ps(vf4, vs4);
178 vf5 = _mm256_mul_ps(vf5, vs5);
179 vf6 = _mm256_mul_ps(vf6, vs6);
180
181 // Store 56 (7x8) outputs at a time.
182 _mm256_storeu_ps(y, vf0);
183 _mm256_storeu_ps(y + 8, vf1);
184 _mm256_storeu_ps(y + 16, vf2);
185 _mm256_storeu_ps(y + 24, vf3);
186 _mm256_storeu_ps(y + 32, vf4);
187 _mm256_storeu_ps(y + 40, vf5);
188 _mm256_storeu_ps(y + 48, vf6);
189 y += 56;
190 }
191
192 for (; elements >= 8 * sizeof(float); elements -= 8 * sizeof(float)) {
193 // Load 8 inputs at a time.
194 const __m256 vx = _mm256_loadu_ps(x);
195 x += 8;
196
197 // Compute reduced argument elements := round(x / log(2)).
198 const __m256 vn = _mm256_round_ps(_mm256_mul_ps(vx, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
199
200 // Compute reduced argument t := x - elements * log(2).
201 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
202 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2_hi, vx);
203 vt = _mm256_fmadd_ps(vn, vminus_ln2_lo, vt);
204
205 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
206 __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
207 vp = _mm256_fmadd_ps(vp, vt, vc3);
208 vp = _mm256_fmadd_ps(vp, vt, vc2);
209 vp = _mm256_fmadd_ps(vp, vt, vc1);
210 vp = _mm256_fmadd_ps(vp, vt, vc0);
211
212 // Multiply "extended" floating-point numbers in ("mantissa", "exponent") representation.
213 __m256 vf = _mm256_mul_ps(vp, vscalev);
214 __m256 ve = _mm256_add_ps(vn, vscalee);
215
216 // For computational efficiency, replace exp2(e) with 0.0f when e <= -127.0.
217 ve = _mm256_max_ps(ve, vmin_exponent);
218
219 // Convert exponents into scale factors.
220 const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve, vmagic_bias)), 23));
221
222 // Multiply "mantissa" by the scale factor.
223 vf = _mm256_mul_ps(vf, vs);
224
225 // Store 8 results at a time.
226 _mm256_storeu_ps(y, vf);
227 y += 8;
228 }
229 if XNN_UNLIKELY(elements != 0) {
230 assert(elements >= 1 * sizeof(float));
231 assert(elements <= 7 * sizeof(float));
232 const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - elements));
233
234 // Load up to 7 inputs at a time.
235 const __m256 vx = _mm256_maskload_ps(x, vmask);
236
237 // Compute reduced argument elements := round(x / log(2)).
238 const __m256 vn = _mm256_round_ps(_mm256_mul_ps(vx, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
239
240 // Compute reduced argument t := x - elements * log(2).
241 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
242 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2_hi, vx);
243 vt = _mm256_fmadd_ps(vn, vminus_ln2_lo, vt);
244
245 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
246 __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
247 vp = _mm256_fmadd_ps(vp, vt, vc3);
248 vp = _mm256_fmadd_ps(vp, vt, vc2);
249 vp = _mm256_fmadd_ps(vp, vt, vc1);
250 vp = _mm256_fmadd_ps(vp, vt, vc0);
251
252 // Multiply "extended" floating-point numbers in ("mantissa", "exponent") representation.
253 __m256 vf = _mm256_mul_ps(vp, vscalev);
254 __m256 ve = _mm256_add_ps(vn, vscalee);
255
256 // For computational efficiency, replace exp2(e) with 0.0f when e <= -127.0.
257 ve = _mm256_max_ps(ve, vmin_exponent);
258
259 // Convert exponents into scale factors.
260 const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve, vmagic_bias)), 23));
261
262 // Multiply "mantissa" by the scale factor.
263 vf = _mm256_mul_ps(vf, vs);
264
265 // Store up to 7 inputs at a time.
266 _mm256_maskstore_ps(y, vmask, vf);
267 }
268 _mm256_zeroupper();
269 }
270