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_x128(size_t elements,const float * x,float * y,float scale_value,float scale_exp)19 void xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x128(
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 >= 128 * sizeof(float); elements -= 128 * sizeof(float)) {
43     // Load 128 (8x16) 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     const __m512 vx6 = _mm512_loadu_ps(x + 96);
51     const __m512 vx7 = _mm512_loadu_ps(x + 112);
52     x += 128;
53 
54     // Compute reduced argument elements := round(x / log(2)).
55     const __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0);
56     const __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0);
57     const __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0);
58     const __m512 vn3 = _mm512_roundscale_ps(_mm512_mul_ps(vx3, vlog2e), 0);
59     const __m512 vn4 = _mm512_roundscale_ps(_mm512_mul_ps(vx4, vlog2e), 0);
60     const __m512 vn5 = _mm512_roundscale_ps(_mm512_mul_ps(vx5, vlog2e), 0);
61     const __m512 vn6 = _mm512_roundscale_ps(_mm512_mul_ps(vx6, vlog2e), 0);
62     const __m512 vn7 = _mm512_roundscale_ps(_mm512_mul_ps(vx7, vlog2e), 0);
63 
64     // Compute reduced argument t := x - elements * log(2).
65     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
66     __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_hi, vx0);
67     __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_hi, vx1);
68     __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_hi, vx2);
69     __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_hi, vx3);
70     __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_hi, vx4);
71     __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_hi, vx5);
72     __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_hi, vx6);
73     __m512 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_hi, vx7);
74 
75     vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0);
76     vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_lo, vt1);
77     vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_lo, vt2);
78     vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_lo, vt3);
79     vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_lo, vt4);
80     vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_lo, vt5);
81     vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_lo, vt6);
82     vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_lo, vt7);
83 
84     // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
85     __m512 vp0 = _mm512_fmadd_ps(vc5, vt0, vc4);
86     __m512 vp1 = _mm512_fmadd_ps(vc5, vt1, vc4);
87     __m512 vp2 = _mm512_fmadd_ps(vc5, vt2, vc4);
88     __m512 vp3 = _mm512_fmadd_ps(vc5, vt3, vc4);
89     __m512 vp4 = _mm512_fmadd_ps(vc5, vt4, vc4);
90     __m512 vp5 = _mm512_fmadd_ps(vc5, vt5, vc4);
91     __m512 vp6 = _mm512_fmadd_ps(vc5, vt6, vc4);
92     __m512 vp7 = _mm512_fmadd_ps(vc5, vt7, vc4);
93 
94     vp0 = _mm512_fmadd_ps(vp0, vt0, vc3);
95     vp1 = _mm512_fmadd_ps(vp1, vt1, vc3);
96     vp2 = _mm512_fmadd_ps(vp2, vt2, vc3);
97     vp3 = _mm512_fmadd_ps(vp3, vt3, vc3);
98     vp4 = _mm512_fmadd_ps(vp4, vt4, vc3);
99     vp5 = _mm512_fmadd_ps(vp5, vt5, vc3);
100     vp6 = _mm512_fmadd_ps(vp6, vt6, vc3);
101     vp7 = _mm512_fmadd_ps(vp7, vt7, vc3);
102 
103     vp0 = _mm512_fmadd_ps(vp0, vt0, vc2);
104     vp1 = _mm512_fmadd_ps(vp1, vt1, vc2);
105     vp2 = _mm512_fmadd_ps(vp2, vt2, vc2);
106     vp3 = _mm512_fmadd_ps(vp3, vt3, vc2);
107     vp4 = _mm512_fmadd_ps(vp4, vt4, vc2);
108     vp5 = _mm512_fmadd_ps(vp5, vt5, vc2);
109     vp6 = _mm512_fmadd_ps(vp6, vt6, vc2);
110     vp7 = _mm512_fmadd_ps(vp7, vt7, vc2);
111 
112     vp0 = _mm512_fmadd_ps(vp0, vt0, vc1);
113     vp1 = _mm512_fmadd_ps(vp1, vt1, vc1);
114     vp2 = _mm512_fmadd_ps(vp2, vt2, vc1);
115     vp3 = _mm512_fmadd_ps(vp3, vt3, vc1);
116     vp4 = _mm512_fmadd_ps(vp4, vt4, vc1);
117     vp5 = _mm512_fmadd_ps(vp5, vt5, vc1);
118     vp6 = _mm512_fmadd_ps(vp6, vt6, vc1);
119     vp7 = _mm512_fmadd_ps(vp7, vt7, 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     vp7 = _mm512_fmadd_ps(vp7, vt7, vc0);
129 
130     // Multiply "extended" floating-point numbers in ("mantissa", "exponent") representation where
131     //  - vnX is "exponent"
132     //  - vpX is "mantissa"
133     //
134     // exp2(ae) * av * exp2(be) * bv =
135     //   = exp2(ae + be) * (av * bv)
136     __m512 vf0 = _mm512_mul_ps(vp0, vscalev);
137     __m512 vf1 = _mm512_mul_ps(vp1, vscalev);
138     __m512 vf2 = _mm512_mul_ps(vp2, vscalev);
139     __m512 vf3 = _mm512_mul_ps(vp3, vscalev);
140     __m512 vf4 = _mm512_mul_ps(vp4, vscalev);
141     __m512 vf5 = _mm512_mul_ps(vp5, vscalev);
142     __m512 vf6 = _mm512_mul_ps(vp6, vscalev);
143     __m512 vf7 = _mm512_mul_ps(vp7, vscalev);
144 
145     const __m512 ve0 = _mm512_add_ps(vn0, vscalee);
146     const __m512 ve1 = _mm512_add_ps(vn1, vscalee);
147     const __m512 ve2 = _mm512_add_ps(vn2, vscalee);
148     const __m512 ve3 = _mm512_add_ps(vn3, vscalee);
149     const __m512 ve4 = _mm512_add_ps(vn4, vscalee);
150     const __m512 ve5 = _mm512_add_ps(vn5, vscalee);
151     const __m512 ve6 = _mm512_add_ps(vn6, vscalee);
152     const __m512 ve7 = _mm512_add_ps(vn7, vscalee);
153 
154     // Multiply "mantissa" by the exp2("exponent").
155     vf0 = _mm512_scalef_ps(vf0, ve0);
156     vf1 = _mm512_scalef_ps(vf1, ve1);
157     vf2 = _mm512_scalef_ps(vf2, ve2);
158     vf3 = _mm512_scalef_ps(vf3, ve3);
159     vf4 = _mm512_scalef_ps(vf4, ve4);
160     vf5 = _mm512_scalef_ps(vf5, ve5);
161     vf6 = _mm512_scalef_ps(vf6, ve6);
162     vf7 = _mm512_scalef_ps(vf7, ve7);
163 
164     // Store 128 (8x16) results at a time.
165     _mm512_storeu_ps(y, vf0);
166     _mm512_storeu_ps(y + 0, vf0);
167     _mm512_storeu_ps(y + 16, vf1);
168     _mm512_storeu_ps(y + 32, vf2);
169     _mm512_storeu_ps(y + 48, vf3);
170     _mm512_storeu_ps(y + 64, vf4);
171     _mm512_storeu_ps(y + 80, vf5);
172     _mm512_storeu_ps(y + 96, vf6);
173     _mm512_storeu_ps(y + 112, vf7);
174     y += 128;
175   }
176 
177   for (; elements >= 16 * sizeof(float); elements -= 16 * sizeof(float)) {
178     // Load 16 inputs at a time.
179     const __m512 vx = _mm512_loadu_ps(x);
180     x += 16;
181 
182     // Compute reduced argument elements := round(x / log(2)).
183     const __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
184 
185     // Compute reduced argument t := x - elements * log(2).
186     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
187     __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);
188     vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
189 
190     // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
191     __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
192     vp = _mm512_fmadd_ps(vp, vt, vc3);
193     vp = _mm512_fmadd_ps(vp, vt, vc2);
194     vp = _mm512_fmadd_ps(vp, vt, vc1);
195     vp = _mm512_fmadd_ps(vp, vt, vc0);
196 
197     // Multiply "extended" floating-point numbers in ("mantissa", "exponent") representation.
198     __m512 vf = _mm512_mul_ps(vp, vscalev);
199     const __m512 ve = _mm512_add_ps(vn, vscalee);
200 
201     // Multiply "mantissa" by the exp2("exponent").
202     vf = _mm512_scalef_ps(vf, ve);
203 
204     // Store 16 results at a time.
205     _mm512_storeu_ps(y, vf);
206     y += 16;
207   }
208   if XNN_UNLIKELY(elements != 0) {
209     // Prepare mask for valid 32-bit elements (depends on elements).
210     elements >>= 2 /* log2(sizeof(float)) */;
211     const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << elements) - UINT32_C(1)));
212 
213     // Load up to 15 inputs at a time.
214     const __m512 vx = _mm512_maskz_loadu_ps(vmask, x);
215 
216     // Compute reduced argument elements := round(x / log(2)).
217     const __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
218 
219     // Compute reduced argument t := x - elements * log(2).
220     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
221     __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);
222     vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
223 
224     // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
225     __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
226     vp = _mm512_fmadd_ps(vp, vt, vc3);
227     vp = _mm512_fmadd_ps(vp, vt, vc2);
228     vp = _mm512_fmadd_ps(vp, vt, vc1);
229     vp = _mm512_fmadd_ps(vp, vt, vc0);
230 
231     // Multiply "extended" floating-point numbers in ("mantissa", "exponent") representation.
232     __m512 vf = _mm512_mul_ps(vp, vscalev);
233     const __m512 ve = _mm512_add_ps(vn, vscalee);
234 
235     // Multiply "mantissa" by the exp2("exponent").
236     vf = _mm512_scalef_ps(vf, ve);
237 
238     // Store up to 15 results at a time.
239     _mm512_mask_storeu_ps(y, vmask, vf);
240   }
241   _mm256_zeroupper();
242 }
243