xref: /aosp_15_r20/external/XNNPACK/src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-vscaleexpminusmax/avx512f-p5-scalef.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/intrinsics-polyfill.h>
15 #include <xnnpack/vscaleexpminusmax.h>
16 
17 
xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x112(size_t elements,const float * input,float * output,float scale,float max)18 void xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x112(
19     size_t elements,
20     const float* input,
21     float* output,
22     float scale,
23     float max)
24 {
25   assert(elements % sizeof(float) == 0);
26 
27   const __m512 vlog2e = _mm512_set1_ps(0x1.715476p+0f);
28   const __m512 vminus_ln2_hi = _mm512_set1_ps(-0x1.62E43p-1f);
29   const __m512 vminus_ln2_lo = _mm512_set1_ps(0x1.05C61p-29f);
30 
31   const __m512 vc0 = _mm512_set1_ps(1.0f);
32   const __m512 vc1 = _mm512_set1_ps(0x1.FFFFF6p-1f);
33   const __m512 vc2 = _mm512_set1_ps(0x1.FFFDC6p-2f);
34   const __m512 vc3 = _mm512_set1_ps(0x1.555A80p-3f);
35   const __m512 vc4 = _mm512_set1_ps(0x1.573A1Ap-5f);
36   const __m512 vc5 = _mm512_set1_ps(0x1.0F9F9Cp-7f);
37 
38   const __m512 vscale = _mm512_set1_ps(scale);
39   const __m512 vi_max = _mm512_set1_ps(max);
40 
41   for (; elements >= 112 * sizeof(float); elements -= 112 * sizeof(float)) {
42     // Load 112 (7x16) inputs at a time.
43     const __m512 vi0 = _mm512_loadu_ps(input);
44     const __m512 vi1 = _mm512_loadu_ps(input + 16);
45     const __m512 vi2 = _mm512_loadu_ps(input + 32);
46     const __m512 vi3 = _mm512_loadu_ps(input + 48);
47     const __m512 vi4 = _mm512_loadu_ps(input + 64);
48     const __m512 vi5 = _mm512_loadu_ps(input + 80);
49     const __m512 vi6 = _mm512_loadu_ps(input + 96);
50     input += 112;
51 
52     // Subtract maximum input x := i - i_max.
53     const __m512 vx0 = _mm512_sub_ps(vi0, vi_max);
54     const __m512 vx1 = _mm512_sub_ps(vi1, vi_max);
55     const __m512 vx2 = _mm512_sub_ps(vi2, vi_max);
56     const __m512 vx3 = _mm512_sub_ps(vi3, vi_max);
57     const __m512 vx4 = _mm512_sub_ps(vi4, vi_max);
58     const __m512 vx5 = _mm512_sub_ps(vi5, vi_max);
59     const __m512 vx6 = _mm512_sub_ps(vi6, vi_max);
60 
61     // Compute reduced argument elements := round(x / log(2)).
62     __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0);
63     __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0);
64     __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0);
65     __m512 vn3 = _mm512_roundscale_ps(_mm512_mul_ps(vx3, vlog2e), 0);
66     __m512 vn4 = _mm512_roundscale_ps(_mm512_mul_ps(vx4, vlog2e), 0);
67     __m512 vn5 = _mm512_roundscale_ps(_mm512_mul_ps(vx5, vlog2e), 0);
68     __m512 vn6 = _mm512_roundscale_ps(_mm512_mul_ps(vx6, vlog2e), 0);
69 
70     // Compute reduced argument t := x - elements * log(2).
71     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
72     __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_hi, vx0);
73     __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_hi, vx1);
74     __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_hi, vx2);
75     __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_hi, vx3);
76     __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_hi, vx4);
77     __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_hi, vx5);
78     __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_hi, vx6);
79 
80     vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0);
81     vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_lo, vt1);
82     vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_lo, vt2);
83     vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_lo, vt3);
84     vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_lo, vt4);
85     vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_lo, vt5);
86     vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_lo, vt6);
87 
88     // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
89     __m512 vp0 = _mm512_fmadd_ps(vc5, vt0, vc4);
90     __m512 vp1 = _mm512_fmadd_ps(vc5, vt1, vc4);
91     __m512 vp2 = _mm512_fmadd_ps(vc5, vt2, vc4);
92     __m512 vp3 = _mm512_fmadd_ps(vc5, vt3, vc4);
93     __m512 vp4 = _mm512_fmadd_ps(vc5, vt4, vc4);
94     __m512 vp5 = _mm512_fmadd_ps(vc5, vt5, vc4);
95     __m512 vp6 = _mm512_fmadd_ps(vc5, vt6, vc4);
96 
97     vp0 = _mm512_fmadd_ps(vp0, vt0, vc3);
98     vp1 = _mm512_fmadd_ps(vp1, vt1, vc3);
99     vp2 = _mm512_fmadd_ps(vp2, vt2, vc3);
100     vp3 = _mm512_fmadd_ps(vp3, vt3, vc3);
101     vp4 = _mm512_fmadd_ps(vp4, vt4, vc3);
102     vp5 = _mm512_fmadd_ps(vp5, vt5, vc3);
103     vp6 = _mm512_fmadd_ps(vp6, vt6, vc3);
104 
105     vp0 = _mm512_fmadd_ps(vp0, vt0, vc2);
106     vp1 = _mm512_fmadd_ps(vp1, vt1, vc2);
107     vp2 = _mm512_fmadd_ps(vp2, vt2, vc2);
108     vp3 = _mm512_fmadd_ps(vp3, vt3, vc2);
109     vp4 = _mm512_fmadd_ps(vp4, vt4, vc2);
110     vp5 = _mm512_fmadd_ps(vp5, vt5, vc2);
111     vp6 = _mm512_fmadd_ps(vp6, vt6, vc2);
112 
113     vp0 = _mm512_fmadd_ps(vp0, vt0, vc1);
114     vp1 = _mm512_fmadd_ps(vp1, vt1, vc1);
115     vp2 = _mm512_fmadd_ps(vp2, vt2, vc1);
116     vp3 = _mm512_fmadd_ps(vp3, vt3, vc1);
117     vp4 = _mm512_fmadd_ps(vp4, vt4, vc1);
118     vp5 = _mm512_fmadd_ps(vp5, vt5, vc1);
119     vp6 = _mm512_fmadd_ps(vp6, vt6, vc1);
120 
121     vp0 = _mm512_fmadd_ps(vp0, vt0, vc0);
122     vp1 = _mm512_fmadd_ps(vp1, vt1, vc0);
123     vp2 = _mm512_fmadd_ps(vp2, vt2, vc0);
124     vp3 = _mm512_fmadd_ps(vp3, vt3, vc0);
125     vp4 = _mm512_fmadd_ps(vp4, vt4, vc0);
126     vp5 = _mm512_fmadd_ps(vp5, vt5, vc0);
127     vp6 = _mm512_fmadd_ps(vp6, vt6, vc0);
128 
129     // Reconstruct the final f value:
130     //   f = 2**elements * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
131     //     = 2**elements * p
132     __m512 vf0 = _mm512_scalef_ps(vp0, vn0);
133     __m512 vf1 = _mm512_scalef_ps(vp1, vn1);
134     __m512 vf2 = _mm512_scalef_ps(vp2, vn2);
135     __m512 vf3 = _mm512_scalef_ps(vp3, vn3);
136     __m512 vf4 = _mm512_scalef_ps(vp4, vn4);
137     __m512 vf5 = _mm512_scalef_ps(vp5, vn5);
138     __m512 vf6 = _mm512_scalef_ps(vp6, vn6);
139 
140     // Multiply by scale.
141     vf0 = _mm512_mul_ps(vf0, vscale);
142     vf1 = _mm512_mul_ps(vf1, vscale);
143     vf2 = _mm512_mul_ps(vf2, vscale);
144     vf3 = _mm512_mul_ps(vf3, vscale);
145     vf4 = _mm512_mul_ps(vf4, vscale);
146     vf5 = _mm512_mul_ps(vf5, vscale);
147     vf6 = _mm512_mul_ps(vf6, vscale);
148 
149     // Store 112 (7x16) outputs at a time.
150     _mm512_storeu_ps(output, vf0);
151     _mm512_storeu_ps(output + 0, vf0);
152     _mm512_storeu_ps(output + 16, vf1);
153     _mm512_storeu_ps(output + 32, vf2);
154     _mm512_storeu_ps(output + 48, vf3);
155     _mm512_storeu_ps(output + 64, vf4);
156     _mm512_storeu_ps(output + 80, vf5);
157     _mm512_storeu_ps(output + 96, vf6);
158     output += 112;
159   }
160   for (; elements >= 16 * sizeof(float); elements -= 16 * sizeof(float)) {
161     // Load 16 inputs at a time.
162     const __m512 vi = _mm512_loadu_ps(input);
163     input += 16;
164 
165     // Subtract maximum input x := i - i_max.
166     const __m512 vx = _mm512_sub_ps(vi, vi_max);
167 
168     // Compute reduced argument elements := round(x / log(2)).
169     __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
170 
171     // Compute reduced argument t := x - elements * log(2).
172     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
173     __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);
174     vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
175 
176     // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
177     __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
178     vp = _mm512_fmadd_ps(vp, vt, vc3);
179     vp = _mm512_fmadd_ps(vp, vt, vc2);
180     vp = _mm512_fmadd_ps(vp, vt, vc1);
181     vp = _mm512_fmadd_ps(vp, vt, vc0);
182 
183     // Reconstruct the final f value:
184     //   f = 2**elements * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
185     //     = 2**elements * p
186     __m512 vf = _mm512_scalef_ps(vp, vn);
187 
188     // Multiply by scale.
189     vf = _mm512_mul_ps(vf, vscale);
190 
191     // Store 16 outputs at a time.
192     _mm512_storeu_ps(output, vf);
193     output += 16;
194   }
195   if (elements != 0) {
196     // Prepare mask for valid 32-bit elements (depends on elements).
197     elements >>= 2 /* log2(sizeof(float)) */;
198     const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << elements) - UINT32_C(1)));
199 
200     // Load up to 15 inputs at a time.
201     const __m512 vi = _mm512_mask_loadu_ps(_mm512_undefined_ps(), vmask, input);
202 
203     // Subtract maximum input x := i - i_max.
204     const __m512 vx = _mm512_sub_ps(vi, vi_max);
205 
206     // Compute reduced argument elements := round(x / log(2)).
207     __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
208 
209     // Compute reduced argument t := x - elements * log(2).
210     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
211     __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);
212     vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
213 
214     // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
215     __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
216     vp = _mm512_fmadd_ps(vp, vt, vc3);
217     vp = _mm512_fmadd_ps(vp, vt, vc2);
218     vp = _mm512_fmadd_ps(vp, vt, vc1);
219     vp = _mm512_fmadd_ps(vp, vt, vc0);
220 
221     // Reconstruct the final f value:
222     //   f = 2**elements * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
223     //     = 2**elements * p
224     __m512 vf = _mm512_scalef_ps(vp, vn);
225 
226     // Multiply by scale.
227     vf = _mm512_mul_ps(vf, vscale);
228 
229     // Store up to 15 outputs at a time.
230     _mm512_mask_storeu_ps(output, vmask, vf);
231   }
232 }
233