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_x80(size_t elements,const float * x,float * y,float scale_value,float scale_exp)20 void xnn_f32_vscaleextexp_ukernel__avx2_p5_x80(
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 >= 80 * sizeof(float); elements -= 80 * sizeof(float)) {
49 // Load 80 (10x8) 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 const __m256 vx7 = _mm256_loadu_ps(x + 56);
58 const __m256 vx8 = _mm256_loadu_ps(x + 64);
59 const __m256 vx9 = _mm256_loadu_ps(x + 72);
60 x += 80;
61
62 // Compute reduced argument elements := round(x / log(2)).
63 const __m256 vn0 = _mm256_round_ps(_mm256_mul_ps(vx0, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
64 const __m256 vn1 = _mm256_round_ps(_mm256_mul_ps(vx1, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
65 const __m256 vn2 = _mm256_round_ps(_mm256_mul_ps(vx2, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
66 const __m256 vn3 = _mm256_round_ps(_mm256_mul_ps(vx3, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
67 const __m256 vn4 = _mm256_round_ps(_mm256_mul_ps(vx4, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
68 const __m256 vn5 = _mm256_round_ps(_mm256_mul_ps(vx5, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
69 const __m256 vn6 = _mm256_round_ps(_mm256_mul_ps(vx6, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
70 const __m256 vn7 = _mm256_round_ps(_mm256_mul_ps(vx7, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
71 const __m256 vn8 = _mm256_round_ps(_mm256_mul_ps(vx8, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
72 const __m256 vn9 = _mm256_round_ps(_mm256_mul_ps(vx9, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
73
74 // Compute reduced argument t := x - elements * log(2).
75 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
76 __m256 vt0 = _mm256_fmadd_ps(vn0, vminus_ln2_hi, vx0);
77 __m256 vt1 = _mm256_fmadd_ps(vn1, vminus_ln2_hi, vx1);
78 __m256 vt2 = _mm256_fmadd_ps(vn2, vminus_ln2_hi, vx2);
79 __m256 vt3 = _mm256_fmadd_ps(vn3, vminus_ln2_hi, vx3);
80 __m256 vt4 = _mm256_fmadd_ps(vn4, vminus_ln2_hi, vx4);
81 __m256 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_hi, vx5);
82 __m256 vt6 = _mm256_fmadd_ps(vn6, vminus_ln2_hi, vx6);
83 __m256 vt7 = _mm256_fmadd_ps(vn7, vminus_ln2_hi, vx7);
84 __m256 vt8 = _mm256_fmadd_ps(vn8, vminus_ln2_hi, vx8);
85 __m256 vt9 = _mm256_fmadd_ps(vn9, vminus_ln2_hi, vx9);
86
87 vt0 = _mm256_fmadd_ps(vn0, vminus_ln2_lo, vt0);
88 vt1 = _mm256_fmadd_ps(vn1, vminus_ln2_lo, vt1);
89 vt2 = _mm256_fmadd_ps(vn2, vminus_ln2_lo, vt2);
90 vt3 = _mm256_fmadd_ps(vn3, vminus_ln2_lo, vt3);
91 vt4 = _mm256_fmadd_ps(vn4, vminus_ln2_lo, vt4);
92 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_lo, vt5);
93 vt6 = _mm256_fmadd_ps(vn6, vminus_ln2_lo, vt6);
94 vt7 = _mm256_fmadd_ps(vn7, vminus_ln2_lo, vt7);
95 vt8 = _mm256_fmadd_ps(vn8, vminus_ln2_lo, vt8);
96 vt9 = _mm256_fmadd_ps(vn9, vminus_ln2_lo, vt9);
97
98 // Compute degree-5 polynomial approxiatmion for exp(t) on [-log(2)/2, log(2)/2].
99 __m256 vp0 = _mm256_fmadd_ps(vc5, vt0, vc4);
100 __m256 vp1 = _mm256_fmadd_ps(vc5, vt1, vc4);
101 __m256 vp2 = _mm256_fmadd_ps(vc5, vt2, vc4);
102 __m256 vp3 = _mm256_fmadd_ps(vc5, vt3, vc4);
103 __m256 vp4 = _mm256_fmadd_ps(vc5, vt4, vc4);
104 __m256 vp5 = _mm256_fmadd_ps(vc5, vt5, vc4);
105 __m256 vp6 = _mm256_fmadd_ps(vc5, vt6, vc4);
106 __m256 vp7 = _mm256_fmadd_ps(vc5, vt7, vc4);
107 __m256 vp8 = _mm256_fmadd_ps(vc5, vt8, vc4);
108 __m256 vp9 = _mm256_fmadd_ps(vc5, vt9, vc4);
109
110 vp0 = _mm256_fmadd_ps(vp0, vt0, vc3);
111 vp1 = _mm256_fmadd_ps(vp1, vt1, vc3);
112 vp2 = _mm256_fmadd_ps(vp2, vt2, vc3);
113 vp3 = _mm256_fmadd_ps(vp3, vt3, vc3);
114 vp4 = _mm256_fmadd_ps(vp4, vt4, vc3);
115 vp5 = _mm256_fmadd_ps(vp5, vt5, vc3);
116 vp6 = _mm256_fmadd_ps(vp6, vt6, vc3);
117 vp7 = _mm256_fmadd_ps(vp7, vt7, vc3);
118 vp8 = _mm256_fmadd_ps(vp8, vt8, vc3);
119 vp9 = _mm256_fmadd_ps(vp9, vt9, vc3);
120
121 vp0 = _mm256_fmadd_ps(vp0, vt0, vc2);
122 vp1 = _mm256_fmadd_ps(vp1, vt1, vc2);
123 vp2 = _mm256_fmadd_ps(vp2, vt2, vc2);
124 vp3 = _mm256_fmadd_ps(vp3, vt3, vc2);
125 vp4 = _mm256_fmadd_ps(vp4, vt4, vc2);
126 vp5 = _mm256_fmadd_ps(vp5, vt5, vc2);
127 vp6 = _mm256_fmadd_ps(vp6, vt6, vc2);
128 vp7 = _mm256_fmadd_ps(vp7, vt7, vc2);
129 vp8 = _mm256_fmadd_ps(vp8, vt8, vc2);
130 vp9 = _mm256_fmadd_ps(vp9, vt9, vc2);
131
132 vp0 = _mm256_fmadd_ps(vp0, vt0, vc1);
133 vp1 = _mm256_fmadd_ps(vp1, vt1, vc1);
134 vp2 = _mm256_fmadd_ps(vp2, vt2, vc1);
135 vp3 = _mm256_fmadd_ps(vp3, vt3, vc1);
136 vp4 = _mm256_fmadd_ps(vp4, vt4, vc1);
137 vp5 = _mm256_fmadd_ps(vp5, vt5, vc1);
138 vp6 = _mm256_fmadd_ps(vp6, vt6, vc1);
139 vp7 = _mm256_fmadd_ps(vp7, vt7, vc1);
140 vp8 = _mm256_fmadd_ps(vp8, vt8, vc1);
141 vp9 = _mm256_fmadd_ps(vp9, vt9, vc1);
142
143 vp0 = _mm256_fmadd_ps(vp0, vt0, vc0);
144 vp1 = _mm256_fmadd_ps(vp1, vt1, vc0);
145 vp2 = _mm256_fmadd_ps(vp2, vt2, vc0);
146 vp3 = _mm256_fmadd_ps(vp3, vt3, vc0);
147 vp4 = _mm256_fmadd_ps(vp4, vt4, vc0);
148 vp5 = _mm256_fmadd_ps(vp5, vt5, vc0);
149 vp6 = _mm256_fmadd_ps(vp6, vt6, vc0);
150 vp7 = _mm256_fmadd_ps(vp7, vt7, vc0);
151 vp8 = _mm256_fmadd_ps(vp8, vt8, vc0);
152 vp9 = _mm256_fmadd_ps(vp9, vt9, vc0);
153
154 // Multiply "extended" floating-point numbers in ("mantissa", "exponent") representation where
155 // - vnX is "exponent"
156 // - vpX is "mantissa"
157 //
158 // exp2(ae) * av * exp2(be) * bv =
159 // = exp2(ae + be) * (av * bv)
160 __m256 vf0 = _mm256_mul_ps(vp0, vscalev);
161 __m256 vf1 = _mm256_mul_ps(vp1, vscalev);
162 __m256 vf2 = _mm256_mul_ps(vp2, vscalev);
163 __m256 vf3 = _mm256_mul_ps(vp3, vscalev);
164 __m256 vf4 = _mm256_mul_ps(vp4, vscalev);
165 __m256 vf5 = _mm256_mul_ps(vp5, vscalev);
166 __m256 vf6 = _mm256_mul_ps(vp6, vscalev);
167 __m256 vf7 = _mm256_mul_ps(vp7, vscalev);
168 __m256 vf8 = _mm256_mul_ps(vp8, vscalev);
169 __m256 vf9 = _mm256_mul_ps(vp9, vscalev);
170
171 __m256 ve0 = _mm256_add_ps(vn0, vscalee);
172 __m256 ve1 = _mm256_add_ps(vn1, vscalee);
173 __m256 ve2 = _mm256_add_ps(vn2, vscalee);
174 __m256 ve3 = _mm256_add_ps(vn3, vscalee);
175 __m256 ve4 = _mm256_add_ps(vn4, vscalee);
176 __m256 ve5 = _mm256_add_ps(vn5, vscalee);
177 __m256 ve6 = _mm256_add_ps(vn6, vscalee);
178 __m256 ve7 = _mm256_add_ps(vn7, vscalee);
179 __m256 ve8 = _mm256_add_ps(vn8, vscalee);
180 __m256 ve9 = _mm256_add_ps(vn9, vscalee);
181
182 // For computational efficiency, replace exp2(e) with 0.0f when e <= -127.0.
183 // This replacement is done in two steps:
184 // 1. Clamp minimum e at -127.0.
185 // 2. Map e to scale factor 0.0 when e == -127.0
186 ve0 = _mm256_max_ps(ve0, vmin_exponent);
187 ve1 = _mm256_max_ps(ve1, vmin_exponent);
188 ve2 = _mm256_max_ps(ve2, vmin_exponent);
189 ve3 = _mm256_max_ps(ve3, vmin_exponent);
190 ve4 = _mm256_max_ps(ve4, vmin_exponent);
191 ve5 = _mm256_max_ps(ve5, vmin_exponent);
192 ve6 = _mm256_max_ps(ve6, vmin_exponent);
193 ve7 = _mm256_max_ps(ve7, vmin_exponent);
194 ve8 = _mm256_max_ps(ve8, vmin_exponent);
195 ve9 = _mm256_max_ps(ve9, vmin_exponent);
196
197 // Convert exponents into scale factors:
198 // - s = exp2(e) when e > -127.0
199 // - s = 0.0 when e <= -127.0
200 const __m256 vs0 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve0, vmagic_bias)), 23));
201 const __m256 vs1 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve1, vmagic_bias)), 23));
202 const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve2, vmagic_bias)), 23));
203 const __m256 vs3 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve3, vmagic_bias)), 23));
204 const __m256 vs4 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve4, vmagic_bias)), 23));
205 const __m256 vs5 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve5, vmagic_bias)), 23));
206 const __m256 vs6 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve6, vmagic_bias)), 23));
207 const __m256 vs7 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve7, vmagic_bias)), 23));
208 const __m256 vs8 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve8, vmagic_bias)), 23));
209 const __m256 vs9 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve9, vmagic_bias)), 23));
210
211 // Multiply "mantissa" by the scale factor.
212 vf0 = _mm256_mul_ps(vf0, vs0);
213 vf1 = _mm256_mul_ps(vf1, vs1);
214 vf2 = _mm256_mul_ps(vf2, vs2);
215 vf3 = _mm256_mul_ps(vf3, vs3);
216 vf4 = _mm256_mul_ps(vf4, vs4);
217 vf5 = _mm256_mul_ps(vf5, vs5);
218 vf6 = _mm256_mul_ps(vf6, vs6);
219 vf7 = _mm256_mul_ps(vf7, vs7);
220 vf8 = _mm256_mul_ps(vf8, vs8);
221 vf9 = _mm256_mul_ps(vf9, vs9);
222
223 // Store 80 (10x8) outputs at a time.
224 _mm256_storeu_ps(y, vf0);
225 _mm256_storeu_ps(y + 8, vf1);
226 _mm256_storeu_ps(y + 16, vf2);
227 _mm256_storeu_ps(y + 24, vf3);
228 _mm256_storeu_ps(y + 32, vf4);
229 _mm256_storeu_ps(y + 40, vf5);
230 _mm256_storeu_ps(y + 48, vf6);
231 _mm256_storeu_ps(y + 56, vf7);
232 _mm256_storeu_ps(y + 64, vf8);
233 _mm256_storeu_ps(y + 72, vf9);
234 y += 80;
235 }
236
237 for (; elements >= 8 * sizeof(float); elements -= 8 * sizeof(float)) {
238 // Load 8 inputs at a time.
239 const __m256 vx = _mm256_loadu_ps(x);
240 x += 8;
241
242 // Compute reduced argument elements := round(x / log(2)).
243 const __m256 vn = _mm256_round_ps(_mm256_mul_ps(vx, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
244
245 // Compute reduced argument t := x - elements * log(2).
246 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
247 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2_hi, vx);
248 vt = _mm256_fmadd_ps(vn, vminus_ln2_lo, vt);
249
250 // Compute degree-5 polynomial approxiatmion for exp(t) on [-log(2)/2, log(2)/2].
251 __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
252 vp = _mm256_fmadd_ps(vp, vt, vc3);
253 vp = _mm256_fmadd_ps(vp, vt, vc2);
254 vp = _mm256_fmadd_ps(vp, vt, vc1);
255 vp = _mm256_fmadd_ps(vp, vt, vc0);
256
257 // Multiply "extended" floating-point numbers in ("mantissa", "exponent") representation.
258 __m256 vf = _mm256_mul_ps(vp, vscalev);
259 __m256 ve = _mm256_add_ps(vn, vscalee);
260
261 // For computational efficiency, replace exp2(e) with 0.0f when e <= -127.0.
262 ve = _mm256_max_ps(ve, vmin_exponent);
263
264 // Convert exponents into scale factors.
265 const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve, vmagic_bias)), 23));
266
267 // Multiply "mantissa" by the scale factor.
268 vf = _mm256_mul_ps(vf, vs);
269
270 // Store 8 results at a time.
271 _mm256_storeu_ps(y, vf);
272 y += 8;
273 }
274 if XNN_UNLIKELY(elements != 0) {
275 assert(elements >= 1 * sizeof(float));
276 assert(elements <= 7 * sizeof(float));
277 const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - elements));
278
279 // Load up to 7 inputs at a time.
280 const __m256 vx = _mm256_maskload_ps(x, vmask);
281
282 // Compute reduced argument elements := round(x / log(2)).
283 const __m256 vn = _mm256_round_ps(_mm256_mul_ps(vx, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
284
285 // Compute reduced argument t := x - elements * log(2).
286 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
287 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2_hi, vx);
288 vt = _mm256_fmadd_ps(vn, vminus_ln2_lo, vt);
289
290 // Compute degree-5 polynomial approxiatmion for exp(t) on [-log(2)/2, log(2)/2].
291 __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
292 vp = _mm256_fmadd_ps(vp, vt, vc3);
293 vp = _mm256_fmadd_ps(vp, vt, vc2);
294 vp = _mm256_fmadd_ps(vp, vt, vc1);
295 vp = _mm256_fmadd_ps(vp, vt, vc0);
296
297 // Multiply "extended" floating-point numbers in ("mantissa", "exponent") representation.
298 __m256 vf = _mm256_mul_ps(vp, vscalev);
299 __m256 ve = _mm256_add_ps(vn, vscalee);
300
301 // For computational efficiency, replace exp2(e) with 0.0f when e <= -127.0.
302 ve = _mm256_max_ps(ve, vmin_exponent);
303
304 // Convert exponents into scale factors.
305 const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve, vmagic_bias)), 23));
306
307 // Multiply "mantissa" by the scale factor.
308 vf = _mm256_mul_ps(vf, vs);
309
310 // Store up to 7 inputs at a time.
311 _mm256_maskstore_ps(y, vmask, vf);
312 }
313 _mm256_zeroupper();
314 }
315