Debian Codegraph @

src/f32-vscaleexpminusmax/gen/f32-vscaleexpminusmax-avx2-p5-u88.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/f32-vscaleexpminusmax/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/vscaleexpminusmax.h"
16
17
18 void xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_u88(
19 size_t batch,
20 const float* input,
21 float* output,
22 float scale,
23 float max)
24 {
25 assert(batch != 0);
26 assert(batch % sizeof(float) == 0);
27 assert(input != NULL);
28 assert(output != NULL);
29
30 static const int32_t mask_table[14] = {-1, -1, -1, -1, -1, -1, -1, 0, 0, 0, 0, 0, 0, 0};
31
32 const __m256 vmagic_bias = _mm256_set1_ps(0x1.8000FEp23f);
33 // The smallest x for which expf(x) is normalized.
34 const __m256 vdenorm_cutoff = _mm256_set1_ps(-0x1.5D589Ep6f);
35 const __m256 vlog2e = _mm256_set1_ps(0x1.715476p+0f);
36 const __m256 vminus_ln2_hi = _mm256_set1_ps(-0x1.62E43p-1f);
37 const __m256 vminus_ln2_lo = _mm256_set1_ps(0x1.05C61p-29f);
38
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 vscale = _mm256_set1_ps(scale);
46 const __m256 vi_max = _mm256_set1_ps(max);
47
48 for (; batch >= 88 * sizeof(float); batch -= 88 * sizeof(float)) {
49 // Load 88 (11x8) inputs at a time.
50 const __m256 vi0 = _mm256_loadu_ps(input);
51 const __m256 vi1 = _mm256_loadu_ps(input + 8);
52 const __m256 vi2 = _mm256_loadu_ps(input + 16);
53 const __m256 vi3 = _mm256_loadu_ps(input + 24);
54 const __m256 vi4 = _mm256_loadu_ps(input + 32);
55 const __m256 vi5 = _mm256_loadu_ps(input + 40);
56 const __m256 vi6 = _mm256_loadu_ps(input + 48);
57 const __m256 vi7 = _mm256_loadu_ps(input + 56);
58 const __m256 vi8 = _mm256_loadu_ps(input + 64);
59 const __m256 vi9 = _mm256_loadu_ps(input + 72);
60 const __m256 vi10 = _mm256_loadu_ps(input + 80);
61 input += 88;
62
63 // Subtract maximum input x := i - i_max. This implies x <= 0.
64 const __m256 vx0 = _mm256_sub_ps(vi0, vi_max);
65 const __m256 vx1 = _mm256_sub_ps(vi1, vi_max);
66 const __m256 vx2 = _mm256_sub_ps(vi2, vi_max);
67 const __m256 vx3 = _mm256_sub_ps(vi3, vi_max);
68 const __m256 vx4 = _mm256_sub_ps(vi4, vi_max);
69 const __m256 vx5 = _mm256_sub_ps(vi5, vi_max);
70 const __m256 vx6 = _mm256_sub_ps(vi6, vi_max);
71 const __m256 vx7 = _mm256_sub_ps(vi7, vi_max);
72 const __m256 vx8 = _mm256_sub_ps(vi8, vi_max);
73 const __m256 vx9 = _mm256_sub_ps(vi9, vi_max);
74 const __m256 vx10 = _mm256_sub_ps(vi10, vi_max);
75
76 // Compute reduced argument batch := round(x / log(2)).
77 __m256 vn0 = _mm256_fmadd_ps(vx0, vlog2e, vmagic_bias);
78 __m256 vn1 = _mm256_fmadd_ps(vx1, vlog2e, vmagic_bias);
79 __m256 vn2 = _mm256_fmadd_ps(vx2, vlog2e, vmagic_bias);
80 __m256 vn3 = _mm256_fmadd_ps(vx3, vlog2e, vmagic_bias);
81 __m256 vn4 = _mm256_fmadd_ps(vx4, vlog2e, vmagic_bias);
82 __m256 vn5 = _mm256_fmadd_ps(vx5, vlog2e, vmagic_bias);
83 __m256 vn6 = _mm256_fmadd_ps(vx6, vlog2e, vmagic_bias);
84 __m256 vn7 = _mm256_fmadd_ps(vx7, vlog2e, vmagic_bias);
85 __m256 vn8 = _mm256_fmadd_ps(vx8, vlog2e, vmagic_bias);
86 __m256 vn9 = _mm256_fmadd_ps(vx9, vlog2e, vmagic_bias);
87 __m256 vn10 = _mm256_fmadd_ps(vx10, vlog2e, vmagic_bias);
88
89 // Create a floating-point number s (scale) such that s == 2**batch for inputs which don't cause underflow, i.e.
90 // -87.33642 <= x <= 0.0, and -126 <= batch <= 0 accordingly.
91 const __m256 vs0 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn0), 23));
92 const __m256 vs1 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn1), 23));
93 const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn2), 23));
94 const __m256 vs3 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn3), 23));
95 const __m256 vs4 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn4), 23));
96 const __m256 vs5 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn5), 23));
97 const __m256 vs6 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn6), 23));
98 const __m256 vs7 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn7), 23));
99 const __m256 vs8 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn8), 23));
100 const __m256 vs9 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn9), 23));
101 const __m256 vs10 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn10), 23));
102
103 // Subtract the large number back to get final batch := round(x / log(2)).
104 vn0 = _mm256_sub_ps(vn0, vmagic_bias);
105 vn1 = _mm256_sub_ps(vn1, vmagic_bias);
106 vn2 = _mm256_sub_ps(vn2, vmagic_bias);
107 vn3 = _mm256_sub_ps(vn3, vmagic_bias);
108 vn4 = _mm256_sub_ps(vn4, vmagic_bias);
109 vn5 = _mm256_sub_ps(vn5, vmagic_bias);
110 vn6 = _mm256_sub_ps(vn6, vmagic_bias);
111 vn7 = _mm256_sub_ps(vn7, vmagic_bias);
112 vn8 = _mm256_sub_ps(vn8, vmagic_bias);
113 vn9 = _mm256_sub_ps(vn9, vmagic_bias);
114 vn10 = _mm256_sub_ps(vn10, vmagic_bias);
115
116 // Compute reduced argument t := x - batch * log(2).
117 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
118 __m256 vt0 = _mm256_fmadd_ps(vn0, vminus_ln2_hi, vx0);
119 __m256 vt1 = _mm256_fmadd_ps(vn1, vminus_ln2_hi, vx1);
120 __m256 vt2 = _mm256_fmadd_ps(vn2, vminus_ln2_hi, vx2);
121 __m256 vt3 = _mm256_fmadd_ps(vn3, vminus_ln2_hi, vx3);
122 __m256 vt4 = _mm256_fmadd_ps(vn4, vminus_ln2_hi, vx4);
123 __m256 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_hi, vx5);
124 __m256 vt6 = _mm256_fmadd_ps(vn6, vminus_ln2_hi, vx6);
125 __m256 vt7 = _mm256_fmadd_ps(vn7, vminus_ln2_hi, vx7);
126 __m256 vt8 = _mm256_fmadd_ps(vn8, vminus_ln2_hi, vx8);
127 __m256 vt9 = _mm256_fmadd_ps(vn9, vminus_ln2_hi, vx9);
128 __m256 vt10 = _mm256_fmadd_ps(vn10, vminus_ln2_hi, vx10);
129
130 vt0 = _mm256_fmadd_ps(vn0, vminus_ln2_lo, vt0);
131 vt1 = _mm256_fmadd_ps(vn1, vminus_ln2_lo, vt1);
132 vt2 = _mm256_fmadd_ps(vn2, vminus_ln2_lo, vt2);
133 vt3 = _mm256_fmadd_ps(vn3, vminus_ln2_lo, vt3);
134 vt4 = _mm256_fmadd_ps(vn4, vminus_ln2_lo, vt4);
135 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_lo, vt5);
136 vt6 = _mm256_fmadd_ps(vn6, vminus_ln2_lo, vt6);
137 vt7 = _mm256_fmadd_ps(vn7, vminus_ln2_lo, vt7);
138 vt8 = _mm256_fmadd_ps(vn8, vminus_ln2_lo, vt8);
139 vt9 = _mm256_fmadd_ps(vn9, vminus_ln2_lo, vt9);
140 vt10 = _mm256_fmadd_ps(vn10, vminus_ln2_lo, vt10);
141
142 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
143 __m256 vp0 = _mm256_fmadd_ps(vc5, vt0, vc4);
144 __m256 vp1 = _mm256_fmadd_ps(vc5, vt1, vc4);
145 __m256 vp2 = _mm256_fmadd_ps(vc5, vt2, vc4);
146 __m256 vp3 = _mm256_fmadd_ps(vc5, vt3, vc4);
147 __m256 vp4 = _mm256_fmadd_ps(vc5, vt4, vc4);
148 __m256 vp5 = _mm256_fmadd_ps(vc5, vt5, vc4);
149 __m256 vp6 = _mm256_fmadd_ps(vc5, vt6, vc4);
150 __m256 vp7 = _mm256_fmadd_ps(vc5, vt7, vc4);
151 __m256 vp8 = _mm256_fmadd_ps(vc5, vt8, vc4);
152 __m256 vp9 = _mm256_fmadd_ps(vc5, vt9, vc4);
153 __m256 vp10 = _mm256_fmadd_ps(vc5, vt10, vc4);
154
155 vp0 = _mm256_fmadd_ps(vp0, vt0, vc3);
156 vp1 = _mm256_fmadd_ps(vp1, vt1, vc3);
157 vp2 = _mm256_fmadd_ps(vp2, vt2, vc3);
158 vp3 = _mm256_fmadd_ps(vp3, vt3, vc3);
159 vp4 = _mm256_fmadd_ps(vp4, vt4, vc3);
160 vp5 = _mm256_fmadd_ps(vp5, vt5, vc3);
161 vp6 = _mm256_fmadd_ps(vp6, vt6, vc3);
162 vp7 = _mm256_fmadd_ps(vp7, vt7, vc3);
163 vp8 = _mm256_fmadd_ps(vp8, vt8, vc3);
164 vp9 = _mm256_fmadd_ps(vp9, vt9, vc3);
165 vp10 = _mm256_fmadd_ps(vp10, vt10, vc3);
166
167 vp0 = _mm256_fmadd_ps(vp0, vt0, vc2);
168 vp1 = _mm256_fmadd_ps(vp1, vt1, vc2);
169 vp2 = _mm256_fmadd_ps(vp2, vt2, vc2);
170 vp3 = _mm256_fmadd_ps(vp3, vt3, vc2);
171 vp4 = _mm256_fmadd_ps(vp4, vt4, vc2);
172 vp5 = _mm256_fmadd_ps(vp5, vt5, vc2);
173 vp6 = _mm256_fmadd_ps(vp6, vt6, vc2);
174 vp7 = _mm256_fmadd_ps(vp7, vt7, vc2);
175 vp8 = _mm256_fmadd_ps(vp8, vt8, vc2);
176 vp9 = _mm256_fmadd_ps(vp9, vt9, vc2);
177 vp10 = _mm256_fmadd_ps(vp10, vt10, vc2);
178
179 vp0 = _mm256_fmadd_ps(vp0, vt0, vc1);
180 vp1 = _mm256_fmadd_ps(vp1, vt1, vc1);
181 vp2 = _mm256_fmadd_ps(vp2, vt2, vc1);
182 vp3 = _mm256_fmadd_ps(vp3, vt3, vc1);
183 vp4 = _mm256_fmadd_ps(vp4, vt4, vc1);
184 vp5 = _mm256_fmadd_ps(vp5, vt5, vc1);
185 vp6 = _mm256_fmadd_ps(vp6, vt6, vc1);
186 vp7 = _mm256_fmadd_ps(vp7, vt7, vc1);
187 vp8 = _mm256_fmadd_ps(vp8, vt8, vc1);
188 vp9 = _mm256_fmadd_ps(vp9, vt9, vc1);
189 vp10 = _mm256_fmadd_ps(vp10, vt10, vc1);
190
191 // Reconstruct the final f value:
192 // f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
193 // = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
194 // = s + (t * s) * p
195 vt0 = _mm256_mul_ps(vt0, vs0);
196 vt1 = _mm256_mul_ps(vt1, vs1);
197 vt2 = _mm256_mul_ps(vt2, vs2);
198 vt3 = _mm256_mul_ps(vt3, vs3);
199 vt4 = _mm256_mul_ps(vt4, vs4);
200 vt5 = _mm256_mul_ps(vt5, vs5);
201 vt6 = _mm256_mul_ps(vt6, vs6);
202 vt7 = _mm256_mul_ps(vt7, vs7);
203 vt8 = _mm256_mul_ps(vt8, vs8);
204 vt9 = _mm256_mul_ps(vt9, vs9);
205 vt10 = _mm256_mul_ps(vt10, vs10);
206
207 __m256 vf0 = _mm256_fmadd_ps(vt0, vp0, vs0);
208 __m256 vf1 = _mm256_fmadd_ps(vt1, vp1, vs1);
209 __m256 vf2 = _mm256_fmadd_ps(vt2, vp2, vs2);
210 __m256 vf3 = _mm256_fmadd_ps(vt3, vp3, vs3);
211 __m256 vf4 = _mm256_fmadd_ps(vt4, vp4, vs4);
212 __m256 vf5 = _mm256_fmadd_ps(vt5, vp5, vs5);
213 __m256 vf6 = _mm256_fmadd_ps(vt6, vp6, vs6);
214 __m256 vf7 = _mm256_fmadd_ps(vt7, vp7, vs7);
215 __m256 vf8 = _mm256_fmadd_ps(vt8, vp8, vs8);
216 __m256 vf9 = _mm256_fmadd_ps(vt9, vp9, vs9);
217 __m256 vf10 = _mm256_fmadd_ps(vt10, vp10, vs10);
218
219 // For inputs below zero cutoff, replace output with +0.0f.
220 // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
221 vf0 = _mm256_andnot_ps(_mm256_cmp_ps(vx0, vdenorm_cutoff, _CMP_LT_OS), vf0);
222 vf1 = _mm256_andnot_ps(_mm256_cmp_ps(vx1, vdenorm_cutoff, _CMP_LT_OS), vf1);
223 vf2 = _mm256_andnot_ps(_mm256_cmp_ps(vx2, vdenorm_cutoff, _CMP_LT_OS), vf2);
224 vf3 = _mm256_andnot_ps(_mm256_cmp_ps(vx3, vdenorm_cutoff, _CMP_LT_OS), vf3);
225 vf4 = _mm256_andnot_ps(_mm256_cmp_ps(vx4, vdenorm_cutoff, _CMP_LT_OS), vf4);
226 vf5 = _mm256_andnot_ps(_mm256_cmp_ps(vx5, vdenorm_cutoff, _CMP_LT_OS), vf5);
227 vf6 = _mm256_andnot_ps(_mm256_cmp_ps(vx6, vdenorm_cutoff, _CMP_LT_OS), vf6);
228 vf7 = _mm256_andnot_ps(_mm256_cmp_ps(vx7, vdenorm_cutoff, _CMP_LT_OS), vf7);
229 vf8 = _mm256_andnot_ps(_mm256_cmp_ps(vx8, vdenorm_cutoff, _CMP_LT_OS), vf8);
230 vf9 = _mm256_andnot_ps(_mm256_cmp_ps(vx9, vdenorm_cutoff, _CMP_LT_OS), vf9);
231 vf10 = _mm256_andnot_ps(_mm256_cmp_ps(vx10, vdenorm_cutoff, _CMP_LT_OS), vf10);
232
233 // Multiply by scale.
234 vf0 = _mm256_mul_ps(vf0, vscale);
235 vf1 = _mm256_mul_ps(vf1, vscale);
236 vf2 = _mm256_mul_ps(vf2, vscale);
237 vf3 = _mm256_mul_ps(vf3, vscale);
238 vf4 = _mm256_mul_ps(vf4, vscale);
239 vf5 = _mm256_mul_ps(vf5, vscale);
240 vf6 = _mm256_mul_ps(vf6, vscale);
241 vf7 = _mm256_mul_ps(vf7, vscale);
242 vf8 = _mm256_mul_ps(vf8, vscale);
243 vf9 = _mm256_mul_ps(vf9, vscale);
244 vf10 = _mm256_mul_ps(vf10, vscale);
245
246 // Store 88 (11x8) outputs at a time.
247 _mm256_storeu_ps(output, vf0);
248 _mm256_storeu_ps(output + 8, vf1);
249 _mm256_storeu_ps(output + 16, vf2);
250 _mm256_storeu_ps(output + 24, vf3);
251 _mm256_storeu_ps(output + 32, vf4);
252 _mm256_storeu_ps(output + 40, vf5);
253 _mm256_storeu_ps(output + 48, vf6);
254 _mm256_storeu_ps(output + 56, vf7);
255 _mm256_storeu_ps(output + 64, vf8);
256 _mm256_storeu_ps(output + 72, vf9);
257 _mm256_storeu_ps(output + 80, vf10);
258 output += 88;
259 }
260 for (; batch >= 8 * sizeof(float); batch -= 8 * sizeof(float)) {
261 // Load 8 inputs at a time.
262 const __m256 vi = _mm256_loadu_ps(input);
263 input += 8;
264
265 // Subtract maximum input x := i - i_max. This implies x <= 0.
266 const __m256 vx = _mm256_sub_ps(vi, vi_max);
267
268 // Compute reduced argument batch := round(x / log(2)).
269 __m256 vn = _mm256_fmadd_ps(vx, vlog2e, vmagic_bias);
270
271 // Create a floating-point number s (scale) such that s == 2**batch for inputs which don't cause underflow, i.e.
272 // -87.33642 <= x <= 0.0, and -126 <= batch <= 0 accordingly.
273 const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
274
275 // Subtract the large number back to get final batch := round(x / log(2)).
276 vn = _mm256_sub_ps(vn, vmagic_bias);
277
278 // Compute reduced argument t := x - batch * log(2).
279 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
280 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2_hi, vx);
281 vt = _mm256_fmadd_ps(vn, vminus_ln2_lo, vt);
282
283 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
284 __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
285 vp = _mm256_fmadd_ps(vp, vt, vc3);
286 vp = _mm256_fmadd_ps(vp, vt, vc2);
287 vp = _mm256_fmadd_ps(vp, vt, vc1);
288
289 // Reconstruct the final f value:
290 // f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
291 // = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
292 // = s + (t * s) * p
293 vt = _mm256_mul_ps(vt, vs);
294 __m256 vf = _mm256_fmadd_ps(vt, vp, vs);
295
296 // For inputs below zero cutoff, replace output with +0.0f.
297 // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
298 vf = _mm256_andnot_ps(_mm256_cmp_ps(vx, vdenorm_cutoff, _CMP_LT_OS), vf);
299
300 // Multiply by scale.
301 vf = _mm256_mul_ps(vf, vscale);
302
303 // Store 64 (8x8) outputs at a time.
304 _mm256_storeu_ps(output, vf);
305 output += 8;
306 }
307 if (batch != 0) {
308 assert(batch >= 1 * sizeof(float));
309 assert(batch <= 7 * sizeof(float));
310 const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - batch));
311
312 // Load up to 7 inputs at a time.
313 const __m256 vi = _mm256_maskload_ps(input, vmask);
314
315 // Subtract maximum input x := i - i_max. This implies x <= 0.
316 const __m256 vx = _mm256_sub_ps(vi, vi_max);
317
318 // Compute reduced argument batch := round(x / log(2)).
319 __m256 vn = _mm256_fmadd_ps(vx, vlog2e, vmagic_bias);
320
321 // Create a floating-point number s (scale) such that s == 2**batch for inputs which don't cause underflow, i.e.
322 // -87.33642 <= x <= 0.0, and -126 <= batch <= 0 accordingly.
323 const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
324
325 // Subtract the large number back to get final batch := round(x / log(2)).
326 vn = _mm256_sub_ps(vn, vmagic_bias);
327
328 // Compute reduced argument t := x - batch * log(2).
329 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
330 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2_hi, vx);
331 vt = _mm256_fmadd_ps(vn, vminus_ln2_lo, vt);
332
333 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
334 __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
335 vp = _mm256_fmadd_ps(vp, vt, vc3);
336 vp = _mm256_fmadd_ps(vp, vt, vc2);
337 vp = _mm256_fmadd_ps(vp, vt, vc1);
338
339 // Reconstruct the final f value:
340 // f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
341 // = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
342 // = s + (t * s) * p
343 vt = _mm256_mul_ps(vt, vs);
344 __m256 vf = _mm256_fmadd_ps(vt, vp, vs);
345
346 // For inputs below zero cutoff, replace output with +0.0f.
347 // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
348 vf = _mm256_andnot_ps(_mm256_cmp_ps(vx, vdenorm_cutoff, _CMP_LT_OS), vf);
349
350 // Multiply by scale.
351 vf = _mm256_mul_ps(vf, vscale);
352
353 // Store up to 7 outputs at a time.
354 _mm256_maskstore_ps(output, vmask, vf);
355 }
356 }