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_x96(size_t elements,const float * x,float * y,float scale_value,float scale_exp)20 void xnn_f32_vscaleextexp_ukernel__avx2_p5_x96(
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 >= 96 * sizeof(float); elements -= 96 * sizeof(float)) {
49 // Load 96 (12x8) 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 const __m256 vx10 = _mm256_loadu_ps(x + 80);
61 const __m256 vx11 = _mm256_loadu_ps(x + 88);
62 x += 96;
63
64 // Compute reduced argument elements := round(x / log(2)).
65 const __m256 vn0 = _mm256_round_ps(_mm256_mul_ps(vx0, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
66 const __m256 vn1 = _mm256_round_ps(_mm256_mul_ps(vx1, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
67 const __m256 vn2 = _mm256_round_ps(_mm256_mul_ps(vx2, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
68 const __m256 vn3 = _mm256_round_ps(_mm256_mul_ps(vx3, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
69 const __m256 vn4 = _mm256_round_ps(_mm256_mul_ps(vx4, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
70 const __m256 vn5 = _mm256_round_ps(_mm256_mul_ps(vx5, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
71 const __m256 vn6 = _mm256_round_ps(_mm256_mul_ps(vx6, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
72 const __m256 vn7 = _mm256_round_ps(_mm256_mul_ps(vx7, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
73 const __m256 vn8 = _mm256_round_ps(_mm256_mul_ps(vx8, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
74 const __m256 vn9 = _mm256_round_ps(_mm256_mul_ps(vx9, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
75 const __m256 vn10 = _mm256_round_ps(_mm256_mul_ps(vx10, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
76 const __m256 vn11 = _mm256_round_ps(_mm256_mul_ps(vx11, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
77
78 // Compute reduced argument t := x - elements * log(2).
79 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
80 __m256 vt0 = _mm256_fmadd_ps(vn0, vminus_ln2_hi, vx0);
81 __m256 vt1 = _mm256_fmadd_ps(vn1, vminus_ln2_hi, vx1);
82 __m256 vt2 = _mm256_fmadd_ps(vn2, vminus_ln2_hi, vx2);
83 __m256 vt3 = _mm256_fmadd_ps(vn3, vminus_ln2_hi, vx3);
84 __m256 vt4 = _mm256_fmadd_ps(vn4, vminus_ln2_hi, vx4);
85 __m256 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_hi, vx5);
86 __m256 vt6 = _mm256_fmadd_ps(vn6, vminus_ln2_hi, vx6);
87 __m256 vt7 = _mm256_fmadd_ps(vn7, vminus_ln2_hi, vx7);
88 __m256 vt8 = _mm256_fmadd_ps(vn8, vminus_ln2_hi, vx8);
89 __m256 vt9 = _mm256_fmadd_ps(vn9, vminus_ln2_hi, vx9);
90 __m256 vt10 = _mm256_fmadd_ps(vn10, vminus_ln2_hi, vx10);
91 __m256 vt11 = _mm256_fmadd_ps(vn11, vminus_ln2_hi, vx11);
92
93 vt0 = _mm256_fmadd_ps(vn0, vminus_ln2_lo, vt0);
94 vt1 = _mm256_fmadd_ps(vn1, vminus_ln2_lo, vt1);
95 vt2 = _mm256_fmadd_ps(vn2, vminus_ln2_lo, vt2);
96 vt3 = _mm256_fmadd_ps(vn3, vminus_ln2_lo, vt3);
97 vt4 = _mm256_fmadd_ps(vn4, vminus_ln2_lo, vt4);
98 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_lo, vt5);
99 vt6 = _mm256_fmadd_ps(vn6, vminus_ln2_lo, vt6);
100 vt7 = _mm256_fmadd_ps(vn7, vminus_ln2_lo, vt7);
101 vt8 = _mm256_fmadd_ps(vn8, vminus_ln2_lo, vt8);
102 vt9 = _mm256_fmadd_ps(vn9, vminus_ln2_lo, vt9);
103 vt10 = _mm256_fmadd_ps(vn10, vminus_ln2_lo, vt10);
104 vt11 = _mm256_fmadd_ps(vn11, vminus_ln2_lo, vt11);
105
106 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
107 __m256 vp0 = _mm256_fmadd_ps(vc5, vt0, vc4);
108 __m256 vp1 = _mm256_fmadd_ps(vc5, vt1, vc4);
109 __m256 vp2 = _mm256_fmadd_ps(vc5, vt2, vc4);
110 __m256 vp3 = _mm256_fmadd_ps(vc5, vt3, vc4);
111 __m256 vp4 = _mm256_fmadd_ps(vc5, vt4, vc4);
112 __m256 vp5 = _mm256_fmadd_ps(vc5, vt5, vc4);
113 __m256 vp6 = _mm256_fmadd_ps(vc5, vt6, vc4);
114 __m256 vp7 = _mm256_fmadd_ps(vc5, vt7, vc4);
115 __m256 vp8 = _mm256_fmadd_ps(vc5, vt8, vc4);
116 __m256 vp9 = _mm256_fmadd_ps(vc5, vt9, vc4);
117 __m256 vp10 = _mm256_fmadd_ps(vc5, vt10, vc4);
118 __m256 vp11 = _mm256_fmadd_ps(vc5, vt11, vc4);
119
120 vp0 = _mm256_fmadd_ps(vp0, vt0, vc3);
121 vp1 = _mm256_fmadd_ps(vp1, vt1, vc3);
122 vp2 = _mm256_fmadd_ps(vp2, vt2, vc3);
123 vp3 = _mm256_fmadd_ps(vp3, vt3, vc3);
124 vp4 = _mm256_fmadd_ps(vp4, vt4, vc3);
125 vp5 = _mm256_fmadd_ps(vp5, vt5, vc3);
126 vp6 = _mm256_fmadd_ps(vp6, vt6, vc3);
127 vp7 = _mm256_fmadd_ps(vp7, vt7, vc3);
128 vp8 = _mm256_fmadd_ps(vp8, vt8, vc3);
129 vp9 = _mm256_fmadd_ps(vp9, vt9, vc3);
130 vp10 = _mm256_fmadd_ps(vp10, vt10, vc3);
131 vp11 = _mm256_fmadd_ps(vp11, vt11, vc3);
132
133 vp0 = _mm256_fmadd_ps(vp0, vt0, vc2);
134 vp1 = _mm256_fmadd_ps(vp1, vt1, vc2);
135 vp2 = _mm256_fmadd_ps(vp2, vt2, vc2);
136 vp3 = _mm256_fmadd_ps(vp3, vt3, vc2);
137 vp4 = _mm256_fmadd_ps(vp4, vt4, vc2);
138 vp5 = _mm256_fmadd_ps(vp5, vt5, vc2);
139 vp6 = _mm256_fmadd_ps(vp6, vt6, vc2);
140 vp7 = _mm256_fmadd_ps(vp7, vt7, vc2);
141 vp8 = _mm256_fmadd_ps(vp8, vt8, vc2);
142 vp9 = _mm256_fmadd_ps(vp9, vt9, vc2);
143 vp10 = _mm256_fmadd_ps(vp10, vt10, vc2);
144 vp11 = _mm256_fmadd_ps(vp11, vt11, vc2);
145
146 vp0 = _mm256_fmadd_ps(vp0, vt0, vc1);
147 vp1 = _mm256_fmadd_ps(vp1, vt1, vc1);
148 vp2 = _mm256_fmadd_ps(vp2, vt2, vc1);
149 vp3 = _mm256_fmadd_ps(vp3, vt3, vc1);
150 vp4 = _mm256_fmadd_ps(vp4, vt4, vc1);
151 vp5 = _mm256_fmadd_ps(vp5, vt5, vc1);
152 vp6 = _mm256_fmadd_ps(vp6, vt6, vc1);
153 vp7 = _mm256_fmadd_ps(vp7, vt7, vc1);
154 vp8 = _mm256_fmadd_ps(vp8, vt8, vc1);
155 vp9 = _mm256_fmadd_ps(vp9, vt9, vc1);
156 vp10 = _mm256_fmadd_ps(vp10, vt10, vc1);
157 vp11 = _mm256_fmadd_ps(vp11, vt11, vc1);
158
159 vp0 = _mm256_fmadd_ps(vp0, vt0, vc0);
160 vp1 = _mm256_fmadd_ps(vp1, vt1, vc0);
161 vp2 = _mm256_fmadd_ps(vp2, vt2, vc0);
162 vp3 = _mm256_fmadd_ps(vp3, vt3, vc0);
163 vp4 = _mm256_fmadd_ps(vp4, vt4, vc0);
164 vp5 = _mm256_fmadd_ps(vp5, vt5, vc0);
165 vp6 = _mm256_fmadd_ps(vp6, vt6, vc0);
166 vp7 = _mm256_fmadd_ps(vp7, vt7, vc0);
167 vp8 = _mm256_fmadd_ps(vp8, vt8, vc0);
168 vp9 = _mm256_fmadd_ps(vp9, vt9, vc0);
169 vp10 = _mm256_fmadd_ps(vp10, vt10, vc0);
170 vp11 = _mm256_fmadd_ps(vp11, vt11, vc0);
171
172 // Multiply "extended" floating-point numbers in ("mantissa", "exponent") representation where
173 // - vnX is "exponent"
174 // - vpX is "mantissa"
175 //
176 // exp2(ae) * av * exp2(be) * bv =
177 // = exp2(ae + be) * (av * bv)
178 __m256 vf0 = _mm256_mul_ps(vp0, vscalev);
179 __m256 vf1 = _mm256_mul_ps(vp1, vscalev);
180 __m256 vf2 = _mm256_mul_ps(vp2, vscalev);
181 __m256 vf3 = _mm256_mul_ps(vp3, vscalev);
182 __m256 vf4 = _mm256_mul_ps(vp4, vscalev);
183 __m256 vf5 = _mm256_mul_ps(vp5, vscalev);
184 __m256 vf6 = _mm256_mul_ps(vp6, vscalev);
185 __m256 vf7 = _mm256_mul_ps(vp7, vscalev);
186 __m256 vf8 = _mm256_mul_ps(vp8, vscalev);
187 __m256 vf9 = _mm256_mul_ps(vp9, vscalev);
188 __m256 vf10 = _mm256_mul_ps(vp10, vscalev);
189 __m256 vf11 = _mm256_mul_ps(vp11, vscalev);
190
191 __m256 ve0 = _mm256_add_ps(vn0, vscalee);
192 __m256 ve1 = _mm256_add_ps(vn1, vscalee);
193 __m256 ve2 = _mm256_add_ps(vn2, vscalee);
194 __m256 ve3 = _mm256_add_ps(vn3, vscalee);
195 __m256 ve4 = _mm256_add_ps(vn4, vscalee);
196 __m256 ve5 = _mm256_add_ps(vn5, vscalee);
197 __m256 ve6 = _mm256_add_ps(vn6, vscalee);
198 __m256 ve7 = _mm256_add_ps(vn7, vscalee);
199 __m256 ve8 = _mm256_add_ps(vn8, vscalee);
200 __m256 ve9 = _mm256_add_ps(vn9, vscalee);
201 __m256 ve10 = _mm256_add_ps(vn10, vscalee);
202 __m256 ve11 = _mm256_add_ps(vn11, vscalee);
203
204 // For computational efficiency, replace exp2(e) with 0.0f when e <= -127.0.
205 // This replacement is done in two steps:
206 // 1. Clamp minimum e at -127.0.
207 // 2. Map e to scale factor 0.0 when e == -127.0
208 ve0 = _mm256_max_ps(ve0, vmin_exponent);
209 ve1 = _mm256_max_ps(ve1, vmin_exponent);
210 ve2 = _mm256_max_ps(ve2, vmin_exponent);
211 ve3 = _mm256_max_ps(ve3, vmin_exponent);
212 ve4 = _mm256_max_ps(ve4, vmin_exponent);
213 ve5 = _mm256_max_ps(ve5, vmin_exponent);
214 ve6 = _mm256_max_ps(ve6, vmin_exponent);
215 ve7 = _mm256_max_ps(ve7, vmin_exponent);
216 ve8 = _mm256_max_ps(ve8, vmin_exponent);
217 ve9 = _mm256_max_ps(ve9, vmin_exponent);
218 ve10 = _mm256_max_ps(ve10, vmin_exponent);
219 ve11 = _mm256_max_ps(ve11, vmin_exponent);
220
221 // Convert exponents into scale factors:
222 // - s = exp2(e) when e > -127.0
223 // - s = 0.0 when e <= -127.0
224 const __m256 vs0 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve0, vmagic_bias)), 23));
225 const __m256 vs1 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve1, vmagic_bias)), 23));
226 const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve2, vmagic_bias)), 23));
227 const __m256 vs3 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve3, vmagic_bias)), 23));
228 const __m256 vs4 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve4, vmagic_bias)), 23));
229 const __m256 vs5 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve5, vmagic_bias)), 23));
230 const __m256 vs6 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve6, vmagic_bias)), 23));
231 const __m256 vs7 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve7, vmagic_bias)), 23));
232 const __m256 vs8 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve8, vmagic_bias)), 23));
233 const __m256 vs9 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve9, vmagic_bias)), 23));
234 const __m256 vs10 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve10, vmagic_bias)), 23));
235 const __m256 vs11 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve11, vmagic_bias)), 23));
236
237 // Multiply "mantissa" by the scale factor.
238 vf0 = _mm256_mul_ps(vf0, vs0);
239 vf1 = _mm256_mul_ps(vf1, vs1);
240 vf2 = _mm256_mul_ps(vf2, vs2);
241 vf3 = _mm256_mul_ps(vf3, vs3);
242 vf4 = _mm256_mul_ps(vf4, vs4);
243 vf5 = _mm256_mul_ps(vf5, vs5);
244 vf6 = _mm256_mul_ps(vf6, vs6);
245 vf7 = _mm256_mul_ps(vf7, vs7);
246 vf8 = _mm256_mul_ps(vf8, vs8);
247 vf9 = _mm256_mul_ps(vf9, vs9);
248 vf10 = _mm256_mul_ps(vf10, vs10);
249 vf11 = _mm256_mul_ps(vf11, vs11);
250
251 // Store 96 (12x8) outputs at a time.
252 _mm256_storeu_ps(y, vf0);
253 _mm256_storeu_ps(y + 8, vf1);
254 _mm256_storeu_ps(y + 16, vf2);
255 _mm256_storeu_ps(y + 24, vf3);
256 _mm256_storeu_ps(y + 32, vf4);
257 _mm256_storeu_ps(y + 40, vf5);
258 _mm256_storeu_ps(y + 48, vf6);
259 _mm256_storeu_ps(y + 56, vf7);
260 _mm256_storeu_ps(y + 64, vf8);
261 _mm256_storeu_ps(y + 72, vf9);
262 _mm256_storeu_ps(y + 80, vf10);
263 _mm256_storeu_ps(y + 88, vf11);
264 y += 96;
265 }
266
267 for (; elements >= 8 * sizeof(float); elements -= 8 * sizeof(float)) {
268 // Load 8 inputs at a time.
269 const __m256 vx = _mm256_loadu_ps(x);
270 x += 8;
271
272 // Compute reduced argument elements := round(x / log(2)).
273 const __m256 vn = _mm256_round_ps(_mm256_mul_ps(vx, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
274
275 // Compute reduced argument t := x - elements * log(2).
276 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
277 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2_hi, vx);
278 vt = _mm256_fmadd_ps(vn, vminus_ln2_lo, vt);
279
280 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
281 __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
282 vp = _mm256_fmadd_ps(vp, vt, vc3);
283 vp = _mm256_fmadd_ps(vp, vt, vc2);
284 vp = _mm256_fmadd_ps(vp, vt, vc1);
285 vp = _mm256_fmadd_ps(vp, vt, vc0);
286
287 // Multiply "extended" floating-point numbers in ("mantissa", "exponent") representation.
288 __m256 vf = _mm256_mul_ps(vp, vscalev);
289 __m256 ve = _mm256_add_ps(vn, vscalee);
290
291 // For computational efficiency, replace exp2(e) with 0.0f when e <= -127.0.
292 ve = _mm256_max_ps(ve, vmin_exponent);
293
294 // Convert exponents into scale factors.
295 const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve, vmagic_bias)), 23));
296
297 // Multiply "mantissa" by the scale factor.
298 vf = _mm256_mul_ps(vf, vs);
299
300 // Store 8 results at a time.
301 _mm256_storeu_ps(y, vf);
302 y += 8;
303 }
304 if XNN_UNLIKELY(elements != 0) {
305 assert(elements >= 1 * sizeof(float));
306 assert(elements <= 7 * sizeof(float));
307 const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - elements));
308
309 // Load up to 7 inputs at a time.
310 const __m256 vx = _mm256_maskload_ps(x, vmask);
311
312 // Compute reduced argument elements := round(x / log(2)).
313 const __m256 vn = _mm256_round_ps(_mm256_mul_ps(vx, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
314
315 // Compute reduced argument t := x - elements * log(2).
316 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
317 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2_hi, vx);
318 vt = _mm256_fmadd_ps(vn, vminus_ln2_lo, vt);
319
320 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
321 __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
322 vp = _mm256_fmadd_ps(vp, vt, vc3);
323 vp = _mm256_fmadd_ps(vp, vt, vc2);
324 vp = _mm256_fmadd_ps(vp, vt, vc1);
325 vp = _mm256_fmadd_ps(vp, vt, vc0);
326
327 // Multiply "extended" floating-point numbers in ("mantissa", "exponent") representation.
328 __m256 vf = _mm256_mul_ps(vp, vscalev);
329 __m256 ve = _mm256_add_ps(vn, vscalee);
330
331 // For computational efficiency, replace exp2(e) with 0.0f when e <= -127.0.
332 ve = _mm256_max_ps(ve, vmin_exponent);
333
334 // Convert exponents into scale factors.
335 const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve, vmagic_bias)), 23));
336
337 // Multiply "mantissa" by the scale factor.
338 vf = _mm256_mul_ps(vf, vs);
339
340 // Store up to 7 inputs at a time.
341 _mm256_maskstore_ps(y, vmask, vf);
342 }
343 _mm256_zeroupper();
344 }
345