xref: /aosp_15_r20/external/XNNPACK/src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-vscaleextexp/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/common.h>
15 #include <xnnpack/intrinsics-polyfill.h>
16 #include <xnnpack/vscaleextexp.h>
17 
18 
xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x96(size_t elements,const float * x,float * y,float scale_value,float scale_exp)19 void xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x96(
20     size_t elements,
21     const float* x,
22     float* y,
23     float scale_value,
24     float scale_exp)
25 {
26   assert(elements % sizeof(float) == 0);
27 
28   const __m512 vlog2e = _mm512_set1_ps(0x1.715476p+0f);
29   const __m512 vminus_ln2_hi = _mm512_set1_ps(-0x1.62E43p-1f);
30   const __m512 vminus_ln2_lo = _mm512_set1_ps(0x1.05C61p-29f);
31 
32   const __m512 vc0 = _mm512_set1_ps(1.0f);
33   const __m512 vc1 = _mm512_set1_ps(0x1.FFFFF6p-1f);
34   const __m512 vc2 = _mm512_set1_ps(0x1.FFFDC6p-2f);
35   const __m512 vc3 = _mm512_set1_ps(0x1.555A80p-3f);
36   const __m512 vc4 = _mm512_set1_ps(0x1.573A1Ap-5f);
37   const __m512 vc5 = _mm512_set1_ps(0x1.0F9F9Cp-7f);
38 
39   const __m512 vscalev = _mm512_set1_ps(scale_value);
40   const __m512 vscalee = _mm512_set1_ps(scale_exp);
41 
42   for (; elements >= 96 * sizeof(float); elements -= 96 * sizeof(float)) {
43     // Load 96 (6x16) inputs at a time.
44     const __m512 vx0 = _mm512_loadu_ps(x);
45     const __m512 vx1 = _mm512_loadu_ps(x + 16);
46     const __m512 vx2 = _mm512_loadu_ps(x + 32);
47     const __m512 vx3 = _mm512_loadu_ps(x + 48);
48     const __m512 vx4 = _mm512_loadu_ps(x + 64);
49     const __m512 vx5 = _mm512_loadu_ps(x + 80);
50     x += 96;
51 
52     // Compute reduced argument elements := round(x / log(2)).
53     const __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0);
54     const __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0);
55     const __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0);
56     const __m512 vn3 = _mm512_roundscale_ps(_mm512_mul_ps(vx3, vlog2e), 0);
57     const __m512 vn4 = _mm512_roundscale_ps(_mm512_mul_ps(vx4, vlog2e), 0);
58     const __m512 vn5 = _mm512_roundscale_ps(_mm512_mul_ps(vx5, vlog2e), 0);
59 
60     // Compute reduced argument t := x - elements * log(2).
61     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
62     __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_hi, vx0);
63     __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_hi, vx1);
64     __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_hi, vx2);
65     __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_hi, vx3);
66     __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_hi, vx4);
67     __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_hi, vx5);
68 
69     vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0);
70     vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_lo, vt1);
71     vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_lo, vt2);
72     vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_lo, vt3);
73     vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_lo, vt4);
74     vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_lo, vt5);
75 
76     // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
77     __m512 vp0 = _mm512_fmadd_ps(vc5, vt0, vc4);
78     __m512 vp1 = _mm512_fmadd_ps(vc5, vt1, vc4);
79     __m512 vp2 = _mm512_fmadd_ps(vc5, vt2, vc4);
80     __m512 vp3 = _mm512_fmadd_ps(vc5, vt3, vc4);
81     __m512 vp4 = _mm512_fmadd_ps(vc5, vt4, vc4);
82     __m512 vp5 = _mm512_fmadd_ps(vc5, vt5, vc4);
83 
84     vp0 = _mm512_fmadd_ps(vp0, vt0, vc3);
85     vp1 = _mm512_fmadd_ps(vp1, vt1, vc3);
86     vp2 = _mm512_fmadd_ps(vp2, vt2, vc3);
87     vp3 = _mm512_fmadd_ps(vp3, vt3, vc3);
88     vp4 = _mm512_fmadd_ps(vp4, vt4, vc3);
89     vp5 = _mm512_fmadd_ps(vp5, vt5, vc3);
90 
91     vp0 = _mm512_fmadd_ps(vp0, vt0, vc2);
92     vp1 = _mm512_fmadd_ps(vp1, vt1, vc2);
93     vp2 = _mm512_fmadd_ps(vp2, vt2, vc2);
94     vp3 = _mm512_fmadd_ps(vp3, vt3, vc2);
95     vp4 = _mm512_fmadd_ps(vp4, vt4, vc2);
96     vp5 = _mm512_fmadd_ps(vp5, vt5, vc2);
97 
98     vp0 = _mm512_fmadd_ps(vp0, vt0, vc1);
99     vp1 = _mm512_fmadd_ps(vp1, vt1, vc1);
100     vp2 = _mm512_fmadd_ps(vp2, vt2, vc1);
101     vp3 = _mm512_fmadd_ps(vp3, vt3, vc1);
102     vp4 = _mm512_fmadd_ps(vp4, vt4, vc1);
103     vp5 = _mm512_fmadd_ps(vp5, vt5, vc1);
104 
105     vp0 = _mm512_fmadd_ps(vp0, vt0, vc0);
106     vp1 = _mm512_fmadd_ps(vp1, vt1, vc0);
107     vp2 = _mm512_fmadd_ps(vp2, vt2, vc0);
108     vp3 = _mm512_fmadd_ps(vp3, vt3, vc0);
109     vp4 = _mm512_fmadd_ps(vp4, vt4, vc0);
110     vp5 = _mm512_fmadd_ps(vp5, vt5, vc0);
111 
112     // Multiply "extended" floating-point numbers in ("mantissa", "exponent") representation where
113     //  - vnX is "exponent"
114     //  - vpX is "mantissa"
115     //
116     // exp2(ae) * av * exp2(be) * bv =
117     //   = exp2(ae + be) * (av * bv)
118     __m512 vf0 = _mm512_mul_ps(vp0, vscalev);
119     __m512 vf1 = _mm512_mul_ps(vp1, vscalev);
120     __m512 vf2 = _mm512_mul_ps(vp2, vscalev);
121     __m512 vf3 = _mm512_mul_ps(vp3, vscalev);
122     __m512 vf4 = _mm512_mul_ps(vp4, vscalev);
123     __m512 vf5 = _mm512_mul_ps(vp5, vscalev);
124 
125     const __m512 ve0 = _mm512_add_ps(vn0, vscalee);
126     const __m512 ve1 = _mm512_add_ps(vn1, vscalee);
127     const __m512 ve2 = _mm512_add_ps(vn2, vscalee);
128     const __m512 ve3 = _mm512_add_ps(vn3, vscalee);
129     const __m512 ve4 = _mm512_add_ps(vn4, vscalee);
130     const __m512 ve5 = _mm512_add_ps(vn5, vscalee);
131 
132     // Multiply "mantissa" by the exp2("exponent").
133     vf0 = _mm512_scalef_ps(vf0, ve0);
134     vf1 = _mm512_scalef_ps(vf1, ve1);
135     vf2 = _mm512_scalef_ps(vf2, ve2);
136     vf3 = _mm512_scalef_ps(vf3, ve3);
137     vf4 = _mm512_scalef_ps(vf4, ve4);
138     vf5 = _mm512_scalef_ps(vf5, ve5);
139 
140     // Store 128 (8x16) results at a time.
141     _mm512_storeu_ps(y, vf0);
142     _mm512_storeu_ps(y + 0, vf0);
143     _mm512_storeu_ps(y + 16, vf1);
144     _mm512_storeu_ps(y + 32, vf2);
145     _mm512_storeu_ps(y + 48, vf3);
146     _mm512_storeu_ps(y + 64, vf4);
147     _mm512_storeu_ps(y + 80, vf5);
148     y += 96;
149   }
150 
151   for (; elements >= 16 * sizeof(float); elements -= 16 * sizeof(float)) {
152     // Load 16 inputs at a time.
153     const __m512 vx = _mm512_loadu_ps(x);
154     x += 16;
155 
156     // Compute reduced argument elements := round(x / log(2)).
157     const __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
158 
159     // Compute reduced argument t := x - elements * log(2).
160     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
161     __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);
162     vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
163 
164     // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
165     __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
166     vp = _mm512_fmadd_ps(vp, vt, vc3);
167     vp = _mm512_fmadd_ps(vp, vt, vc2);
168     vp = _mm512_fmadd_ps(vp, vt, vc1);
169     vp = _mm512_fmadd_ps(vp, vt, vc0);
170 
171     // Multiply "extended" floating-point numbers in ("mantissa", "exponent") representation.
172     __m512 vf = _mm512_mul_ps(vp, vscalev);
173     const __m512 ve = _mm512_add_ps(vn, vscalee);
174 
175     // Multiply "mantissa" by the exp2("exponent").
176     vf = _mm512_scalef_ps(vf, ve);
177 
178     // Store 16 results at a time.
179     _mm512_storeu_ps(y, vf);
180     y += 16;
181   }
182   if XNN_UNLIKELY(elements != 0) {
183     // Prepare mask for valid 32-bit elements (depends on elements).
184     elements >>= 2 /* log2(sizeof(float)) */;
185     const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << elements) - UINT32_C(1)));
186 
187     // Load up to 15 inputs at a time.
188     const __m512 vx = _mm512_maskz_loadu_ps(vmask, x);
189 
190     // Compute reduced argument elements := round(x / log(2)).
191     const __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
192 
193     // Compute reduced argument t := x - elements * log(2).
194     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
195     __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);
196     vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
197 
198     // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
199     __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
200     vp = _mm512_fmadd_ps(vp, vt, vc3);
201     vp = _mm512_fmadd_ps(vp, vt, vc2);
202     vp = _mm512_fmadd_ps(vp, vt, vc1);
203     vp = _mm512_fmadd_ps(vp, vt, vc0);
204 
205     // Multiply "extended" floating-point numbers in ("mantissa", "exponent") representation.
206     __m512 vf = _mm512_mul_ps(vp, vscalev);
207     const __m512 ve = _mm512_add_ps(vn, vscalee);
208 
209     // Multiply "mantissa" by the exp2("exponent").
210     vf = _mm512_scalef_ps(vf, ve);
211 
212     // Store up to 15 results at a time.
213     _mm512_mask_storeu_ps(y, vmask, vf);
214   }
215   _mm256_zeroupper();
216 }
217