• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 static const int32_t mask_table[14] = {-1, -1, -1, -1, -1, -1, -1, 0, 0, 0, 0, 0, 0, 0};
19 
xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x88(size_t elements,const float * input,float * output,float scale,float max)20 void xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x88(
21     size_t elements,
22     const float* input,
23     float* output,
24     float scale,
25     float max)
26 {
27   assert(elements % sizeof(float) == 0);
28 
29   const __m256 vmagic_bias = _mm256_set1_ps(0x1.8000FEp23f);
30   // The smallest x for which expf(x) is normalized.
31   const __m256 vdenorm_cutoff = _mm256_set1_ps(-0x1.5D589Ep6f);
32   const __m256 vlog2e = _mm256_set1_ps(0x1.715476p+0f);
33   const __m256 vminus_ln2_hi = _mm256_set1_ps(-0x1.62E43p-1f);
34   const __m256 vminus_ln2_lo = _mm256_set1_ps(0x1.05C61p-29f);
35 
36   const __m256 vc1 = _mm256_set1_ps(0x1.FFFFF6p-1f);
37   const __m256 vc2 = _mm256_set1_ps(0x1.FFFDC6p-2f);
38   const __m256 vc3 = _mm256_set1_ps(0x1.555A80p-3f);
39   const __m256 vc4 = _mm256_set1_ps(0x1.573A1Ap-5f);
40   const __m256 vc5 = _mm256_set1_ps(0x1.0F9F9Cp-7f);
41 
42   const __m256 vscale = _mm256_set1_ps(scale);
43   const __m256 vi_max = _mm256_set1_ps(max);
44 
45   for (; elements >= 88 * sizeof(float); elements -= 88 * sizeof(float)) {
46     // Load 88 (11x8) inputs at a time.
47     const __m256 vi0 = _mm256_loadu_ps(input);
48     const __m256 vi1 = _mm256_loadu_ps(input + 8);
49     const __m256 vi2 = _mm256_loadu_ps(input + 16);
50     const __m256 vi3 = _mm256_loadu_ps(input + 24);
51     const __m256 vi4 = _mm256_loadu_ps(input + 32);
52     const __m256 vi5 = _mm256_loadu_ps(input + 40);
53     const __m256 vi6 = _mm256_loadu_ps(input + 48);
54     const __m256 vi7 = _mm256_loadu_ps(input + 56);
55     const __m256 vi8 = _mm256_loadu_ps(input + 64);
56     const __m256 vi9 = _mm256_loadu_ps(input + 72);
57     const __m256 vi10 = _mm256_loadu_ps(input + 80);
58     input += 88;
59 
60     // Subtract maximum input x := i - i_max. This implies x <= 0.
61     const __m256 vx0 = _mm256_sub_ps(vi0, vi_max);
62     const __m256 vx1 = _mm256_sub_ps(vi1, vi_max);
63     const __m256 vx2 = _mm256_sub_ps(vi2, vi_max);
64     const __m256 vx3 = _mm256_sub_ps(vi3, vi_max);
65     const __m256 vx4 = _mm256_sub_ps(vi4, vi_max);
66     const __m256 vx5 = _mm256_sub_ps(vi5, vi_max);
67     const __m256 vx6 = _mm256_sub_ps(vi6, vi_max);
68     const __m256 vx7 = _mm256_sub_ps(vi7, vi_max);
69     const __m256 vx8 = _mm256_sub_ps(vi8, vi_max);
70     const __m256 vx9 = _mm256_sub_ps(vi9, vi_max);
71     const __m256 vx10 = _mm256_sub_ps(vi10, vi_max);
72 
73     // Compute reduced argument elements := round(x / log(2)).
74     __m256 vn0 = _mm256_fmadd_ps(vx0, vlog2e, vmagic_bias);
75     __m256 vn1 = _mm256_fmadd_ps(vx1, vlog2e, vmagic_bias);
76     __m256 vn2 = _mm256_fmadd_ps(vx2, vlog2e, vmagic_bias);
77     __m256 vn3 = _mm256_fmadd_ps(vx3, vlog2e, vmagic_bias);
78     __m256 vn4 = _mm256_fmadd_ps(vx4, vlog2e, vmagic_bias);
79     __m256 vn5 = _mm256_fmadd_ps(vx5, vlog2e, vmagic_bias);
80     __m256 vn6 = _mm256_fmadd_ps(vx6, vlog2e, vmagic_bias);
81     __m256 vn7 = _mm256_fmadd_ps(vx7, vlog2e, vmagic_bias);
82     __m256 vn8 = _mm256_fmadd_ps(vx8, vlog2e, vmagic_bias);
83     __m256 vn9 = _mm256_fmadd_ps(vx9, vlog2e, vmagic_bias);
84     __m256 vn10 = _mm256_fmadd_ps(vx10, vlog2e, vmagic_bias);
85 
86     // Create a floating-point number s (scale) such that s == 2**elements for inputs which don't cause underflow, i.e.
87     // -87.33642 <= x <= 0.0, and -126 <= elements <= 0 accordingly.
88     const __m256 vs0 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn0), 23));
89     const __m256 vs1 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn1), 23));
90     const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn2), 23));
91     const __m256 vs3 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn3), 23));
92     const __m256 vs4 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn4), 23));
93     const __m256 vs5 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn5), 23));
94     const __m256 vs6 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn6), 23));
95     const __m256 vs7 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn7), 23));
96     const __m256 vs8 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn8), 23));
97     const __m256 vs9 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn9), 23));
98     const __m256 vs10 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn10), 23));
99 
100     // Subtract the large number back to get final elements := round(x / log(2)).
101     vn0 = _mm256_sub_ps(vn0, vmagic_bias);
102     vn1 = _mm256_sub_ps(vn1, vmagic_bias);
103     vn2 = _mm256_sub_ps(vn2, vmagic_bias);
104     vn3 = _mm256_sub_ps(vn3, vmagic_bias);
105     vn4 = _mm256_sub_ps(vn4, vmagic_bias);
106     vn5 = _mm256_sub_ps(vn5, vmagic_bias);
107     vn6 = _mm256_sub_ps(vn6, vmagic_bias);
108     vn7 = _mm256_sub_ps(vn7, vmagic_bias);
109     vn8 = _mm256_sub_ps(vn8, vmagic_bias);
110     vn9 = _mm256_sub_ps(vn9, vmagic_bias);
111     vn10 = _mm256_sub_ps(vn10, vmagic_bias);
112 
113     // Compute reduced argument t := x - elements * log(2).
114     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
115     __m256 vt0 = _mm256_fmadd_ps(vn0, vminus_ln2_hi, vx0);
116     __m256 vt1 = _mm256_fmadd_ps(vn1, vminus_ln2_hi, vx1);
117     __m256 vt2 = _mm256_fmadd_ps(vn2, vminus_ln2_hi, vx2);
118     __m256 vt3 = _mm256_fmadd_ps(vn3, vminus_ln2_hi, vx3);
119     __m256 vt4 = _mm256_fmadd_ps(vn4, vminus_ln2_hi, vx4);
120     __m256 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_hi, vx5);
121     __m256 vt6 = _mm256_fmadd_ps(vn6, vminus_ln2_hi, vx6);
122     __m256 vt7 = _mm256_fmadd_ps(vn7, vminus_ln2_hi, vx7);
123     __m256 vt8 = _mm256_fmadd_ps(vn8, vminus_ln2_hi, vx8);
124     __m256 vt9 = _mm256_fmadd_ps(vn9, vminus_ln2_hi, vx9);
125     __m256 vt10 = _mm256_fmadd_ps(vn10, vminus_ln2_hi, vx10);
126 
127     vt0 = _mm256_fmadd_ps(vn0, vminus_ln2_lo, vt0);
128     vt1 = _mm256_fmadd_ps(vn1, vminus_ln2_lo, vt1);
129     vt2 = _mm256_fmadd_ps(vn2, vminus_ln2_lo, vt2);
130     vt3 = _mm256_fmadd_ps(vn3, vminus_ln2_lo, vt3);
131     vt4 = _mm256_fmadd_ps(vn4, vminus_ln2_lo, vt4);
132     vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_lo, vt5);
133     vt6 = _mm256_fmadd_ps(vn6, vminus_ln2_lo, vt6);
134     vt7 = _mm256_fmadd_ps(vn7, vminus_ln2_lo, vt7);
135     vt8 = _mm256_fmadd_ps(vn8, vminus_ln2_lo, vt8);
136     vt9 = _mm256_fmadd_ps(vn9, vminus_ln2_lo, vt9);
137     vt10 = _mm256_fmadd_ps(vn10, vminus_ln2_lo, vt10);
138 
139     // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
140     __m256 vp0 = _mm256_fmadd_ps(vc5, vt0, vc4);
141     __m256 vp1 = _mm256_fmadd_ps(vc5, vt1, vc4);
142     __m256 vp2 = _mm256_fmadd_ps(vc5, vt2, vc4);
143     __m256 vp3 = _mm256_fmadd_ps(vc5, vt3, vc4);
144     __m256 vp4 = _mm256_fmadd_ps(vc5, vt4, vc4);
145     __m256 vp5 = _mm256_fmadd_ps(vc5, vt5, vc4);
146     __m256 vp6 = _mm256_fmadd_ps(vc5, vt6, vc4);
147     __m256 vp7 = _mm256_fmadd_ps(vc5, vt7, vc4);
148     __m256 vp8 = _mm256_fmadd_ps(vc5, vt8, vc4);
149     __m256 vp9 = _mm256_fmadd_ps(vc5, vt9, vc4);
150     __m256 vp10 = _mm256_fmadd_ps(vc5, vt10, vc4);
151 
152     vp0 = _mm256_fmadd_ps(vp0, vt0, vc3);
153     vp1 = _mm256_fmadd_ps(vp1, vt1, vc3);
154     vp2 = _mm256_fmadd_ps(vp2, vt2, vc3);
155     vp3 = _mm256_fmadd_ps(vp3, vt3, vc3);
156     vp4 = _mm256_fmadd_ps(vp4, vt4, vc3);
157     vp5 = _mm256_fmadd_ps(vp5, vt5, vc3);
158     vp6 = _mm256_fmadd_ps(vp6, vt6, vc3);
159     vp7 = _mm256_fmadd_ps(vp7, vt7, vc3);
160     vp8 = _mm256_fmadd_ps(vp8, vt8, vc3);
161     vp9 = _mm256_fmadd_ps(vp9, vt9, vc3);
162     vp10 = _mm256_fmadd_ps(vp10, vt10, vc3);
163 
164     vp0 = _mm256_fmadd_ps(vp0, vt0, vc2);
165     vp1 = _mm256_fmadd_ps(vp1, vt1, vc2);
166     vp2 = _mm256_fmadd_ps(vp2, vt2, vc2);
167     vp3 = _mm256_fmadd_ps(vp3, vt3, vc2);
168     vp4 = _mm256_fmadd_ps(vp4, vt4, vc2);
169     vp5 = _mm256_fmadd_ps(vp5, vt5, vc2);
170     vp6 = _mm256_fmadd_ps(vp6, vt6, vc2);
171     vp7 = _mm256_fmadd_ps(vp7, vt7, vc2);
172     vp8 = _mm256_fmadd_ps(vp8, vt8, vc2);
173     vp9 = _mm256_fmadd_ps(vp9, vt9, vc2);
174     vp10 = _mm256_fmadd_ps(vp10, vt10, vc2);
175 
176     vp0 = _mm256_fmadd_ps(vp0, vt0, vc1);
177     vp1 = _mm256_fmadd_ps(vp1, vt1, vc1);
178     vp2 = _mm256_fmadd_ps(vp2, vt2, vc1);
179     vp3 = _mm256_fmadd_ps(vp3, vt3, vc1);
180     vp4 = _mm256_fmadd_ps(vp4, vt4, vc1);
181     vp5 = _mm256_fmadd_ps(vp5, vt5, vc1);
182     vp6 = _mm256_fmadd_ps(vp6, vt6, vc1);
183     vp7 = _mm256_fmadd_ps(vp7, vt7, vc1);
184     vp8 = _mm256_fmadd_ps(vp8, vt8, vc1);
185     vp9 = _mm256_fmadd_ps(vp9, vt9, vc1);
186     vp10 = _mm256_fmadd_ps(vp10, vt10, vc1);
187 
188     // Reconstruct the final f value:
189     //   f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
190     //     = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
191     //     = s + (t * s) * p
192     vt0 = _mm256_mul_ps(vt0, vs0);
193     vt1 = _mm256_mul_ps(vt1, vs1);
194     vt2 = _mm256_mul_ps(vt2, vs2);
195     vt3 = _mm256_mul_ps(vt3, vs3);
196     vt4 = _mm256_mul_ps(vt4, vs4);
197     vt5 = _mm256_mul_ps(vt5, vs5);
198     vt6 = _mm256_mul_ps(vt6, vs6);
199     vt7 = _mm256_mul_ps(vt7, vs7);
200     vt8 = _mm256_mul_ps(vt8, vs8);
201     vt9 = _mm256_mul_ps(vt9, vs9);
202     vt10 = _mm256_mul_ps(vt10, vs10);
203 
204     __m256 vf0 = _mm256_fmadd_ps(vt0, vp0, vs0);
205     __m256 vf1 = _mm256_fmadd_ps(vt1, vp1, vs1);
206     __m256 vf2 = _mm256_fmadd_ps(vt2, vp2, vs2);
207     __m256 vf3 = _mm256_fmadd_ps(vt3, vp3, vs3);
208     __m256 vf4 = _mm256_fmadd_ps(vt4, vp4, vs4);
209     __m256 vf5 = _mm256_fmadd_ps(vt5, vp5, vs5);
210     __m256 vf6 = _mm256_fmadd_ps(vt6, vp6, vs6);
211     __m256 vf7 = _mm256_fmadd_ps(vt7, vp7, vs7);
212     __m256 vf8 = _mm256_fmadd_ps(vt8, vp8, vs8);
213     __m256 vf9 = _mm256_fmadd_ps(vt9, vp9, vs9);
214     __m256 vf10 = _mm256_fmadd_ps(vt10, vp10, vs10);
215 
216     // For inputs below zero cutoff, replace output with +0.0f.
217     // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
218     vf0 = _mm256_andnot_ps(_mm256_cmp_ps(vx0, vdenorm_cutoff, _CMP_LT_OS), vf0);
219     vf1 = _mm256_andnot_ps(_mm256_cmp_ps(vx1, vdenorm_cutoff, _CMP_LT_OS), vf1);
220     vf2 = _mm256_andnot_ps(_mm256_cmp_ps(vx2, vdenorm_cutoff, _CMP_LT_OS), vf2);
221     vf3 = _mm256_andnot_ps(_mm256_cmp_ps(vx3, vdenorm_cutoff, _CMP_LT_OS), vf3);
222     vf4 = _mm256_andnot_ps(_mm256_cmp_ps(vx4, vdenorm_cutoff, _CMP_LT_OS), vf4);
223     vf5 = _mm256_andnot_ps(_mm256_cmp_ps(vx5, vdenorm_cutoff, _CMP_LT_OS), vf5);
224     vf6 = _mm256_andnot_ps(_mm256_cmp_ps(vx6, vdenorm_cutoff, _CMP_LT_OS), vf6);
225     vf7 = _mm256_andnot_ps(_mm256_cmp_ps(vx7, vdenorm_cutoff, _CMP_LT_OS), vf7);
226     vf8 = _mm256_andnot_ps(_mm256_cmp_ps(vx8, vdenorm_cutoff, _CMP_LT_OS), vf8);
227     vf9 = _mm256_andnot_ps(_mm256_cmp_ps(vx9, vdenorm_cutoff, _CMP_LT_OS), vf9);
228     vf10 = _mm256_andnot_ps(_mm256_cmp_ps(vx10, vdenorm_cutoff, _CMP_LT_OS), vf10);
229 
230     // Multiply by scale.
231     vf0 = _mm256_mul_ps(vf0, vscale);
232     vf1 = _mm256_mul_ps(vf1, vscale);
233     vf2 = _mm256_mul_ps(vf2, vscale);
234     vf3 = _mm256_mul_ps(vf3, vscale);
235     vf4 = _mm256_mul_ps(vf4, vscale);
236     vf5 = _mm256_mul_ps(vf5, vscale);
237     vf6 = _mm256_mul_ps(vf6, vscale);
238     vf7 = _mm256_mul_ps(vf7, vscale);
239     vf8 = _mm256_mul_ps(vf8, vscale);
240     vf9 = _mm256_mul_ps(vf9, vscale);
241     vf10 = _mm256_mul_ps(vf10, vscale);
242 
243     // Store 88 (11x8) outputs at a time.
244     _mm256_storeu_ps(output, vf0);
245     _mm256_storeu_ps(output + 8, vf1);
246     _mm256_storeu_ps(output + 16, vf2);
247     _mm256_storeu_ps(output + 24, vf3);
248     _mm256_storeu_ps(output + 32, vf4);
249     _mm256_storeu_ps(output + 40, vf5);
250     _mm256_storeu_ps(output + 48, vf6);
251     _mm256_storeu_ps(output + 56, vf7);
252     _mm256_storeu_ps(output + 64, vf8);
253     _mm256_storeu_ps(output + 72, vf9);
254     _mm256_storeu_ps(output + 80, vf10);
255     output += 88;
256   }
257   for (; elements >= 8 * sizeof(float); elements -= 8 * sizeof(float)) {
258     // Load 8 inputs at a time.
259     const __m256 vi = _mm256_loadu_ps(input);
260     input += 8;
261 
262     // Subtract maximum input x := i - i_max. This implies x <= 0.
263     const __m256 vx = _mm256_sub_ps(vi, vi_max);
264 
265     // Compute reduced argument elements := round(x / log(2)).
266     __m256 vn = _mm256_fmadd_ps(vx, vlog2e, vmagic_bias);
267 
268     // Create a floating-point number s (scale) such that s == 2**elements for inputs which don't cause underflow, i.e.
269     // -87.33642 <= x <= 0.0, and -126 <= elements <= 0 accordingly.
270     const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
271 
272     // Subtract the large number back to get final elements := round(x / log(2)).
273     vn = _mm256_sub_ps(vn, vmagic_bias);
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 
286     // Reconstruct the final f value:
287     //   f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
288     //     = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
289     //     = s + (t * s) * p
290     vt = _mm256_mul_ps(vt, vs);
291     __m256 vf = _mm256_fmadd_ps(vt, vp, vs);
292 
293     // For inputs below zero cutoff, replace output with +0.0f.
294     // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
295     vf = _mm256_andnot_ps(_mm256_cmp_ps(vx, vdenorm_cutoff, _CMP_LT_OS), vf);
296 
297     // Multiply by scale.
298     vf = _mm256_mul_ps(vf, vscale);
299 
300     // Store 64 (8x8) outputs at a time.
301     _mm256_storeu_ps(output, vf);
302     output += 8;
303   }
304   if (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 vi = _mm256_maskload_ps(input, vmask);
311 
312     // Subtract maximum input x := i - i_max. This implies x <= 0.
313     const __m256 vx = _mm256_sub_ps(vi, vi_max);
314 
315     // Compute reduced argument elements := round(x / log(2)).
316     __m256 vn = _mm256_fmadd_ps(vx, vlog2e, vmagic_bias);
317 
318     // Create a floating-point number s (scale) such that s == 2**elements for inputs which don't cause underflow, i.e.
319     // -87.33642 <= x <= 0.0, and -126 <= elements <= 0 accordingly.
320     const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
321 
322     // Subtract the large number back to get final elements := round(x / log(2)).
323     vn = _mm256_sub_ps(vn, vmagic_bias);
324 
325     // Compute reduced argument t := x - elements * log(2).
326     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
327     __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2_hi, vx);
328     vt = _mm256_fmadd_ps(vn, vminus_ln2_lo, vt);
329 
330     // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
331     __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
332     vp = _mm256_fmadd_ps(vp, vt, vc3);
333     vp = _mm256_fmadd_ps(vp, vt, vc2);
334     vp = _mm256_fmadd_ps(vp, vt, vc1);
335 
336     // Reconstruct the final f value:
337     //   f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
338     //     = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
339     //     = s + (t * s) * p
340     vt = _mm256_mul_ps(vt, vs);
341     __m256 vf = _mm256_fmadd_ps(vt, vp, vs);
342 
343     // For inputs below zero cutoff, replace output with +0.0f.
344     // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
345     vf = _mm256_andnot_ps(_mm256_cmp_ps(vx, vdenorm_cutoff, _CMP_LT_OS), vf);
346 
347     // Multiply by scale.
348     vf = _mm256_mul_ps(vf, vscale);
349 
350     // Store up to 7 outputs at a time.
351     _mm256_maskstore_ps(output, vmask, vf);
352   }
353 }
354