• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-raddextexp/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 #include <math.h>
12 
13 #include <immintrin.h>
14 
15 #include <xnnpack/common.h>
16 #include <xnnpack/intrinsics-polyfill.h>
17 #include <xnnpack/raddextexp.h>
18 
19 
xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x144(size_t elements,const float * x,float * sum)20 void xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x144(
21     size_t elements,
22     const float* x,
23     float* sum)
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 vminus_inf = _mm512_set1_ps(-INFINITY);
39 
40   __m512 vaccv0 = _mm512_setzero_ps();
41   __m512 vacce0 = vminus_inf;
42   for (; elements >= 144 * sizeof(float); elements -= 144 * sizeof(float)) {
43     // Load 144 (9x16) 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     const __m512 vx8 = _mm512_loadu_ps(x + 128);
53     x += 144;
54 
55     // Compute reduced argument elements := round(x / log(2)).
56     const __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0);
57     const __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0);
58     const __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0);
59     const __m512 vn3 = _mm512_roundscale_ps(_mm512_mul_ps(vx3, vlog2e), 0);
60     const __m512 vn4 = _mm512_roundscale_ps(_mm512_mul_ps(vx4, vlog2e), 0);
61     const __m512 vn5 = _mm512_roundscale_ps(_mm512_mul_ps(vx5, vlog2e), 0);
62     const __m512 vn6 = _mm512_roundscale_ps(_mm512_mul_ps(vx6, vlog2e), 0);
63     const __m512 vn7 = _mm512_roundscale_ps(_mm512_mul_ps(vx7, vlog2e), 0);
64     const __m512 vn8 = _mm512_roundscale_ps(_mm512_mul_ps(vx8, vlog2e), 0);
65 
66     // Compute reduced argument t := x - elements * log(2).
67     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
68     __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_hi, vx0);
69     __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_hi, vx1);
70     __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_hi, vx2);
71     __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_hi, vx3);
72     __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_hi, vx4);
73     __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_hi, vx5);
74     __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_hi, vx6);
75     __m512 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_hi, vx7);
76     __m512 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_hi, vx8);
77 
78     vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0);
79     vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_lo, vt1);
80     vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_lo, vt2);
81     vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_lo, vt3);
82     vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_lo, vt4);
83     vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_lo, vt5);
84     vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_lo, vt6);
85     vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_lo, vt7);
86     vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_lo, vt8);
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     __m512 vp7 = _mm512_fmadd_ps(vc5, vt7, vc4);
97     __m512 vp8 = _mm512_fmadd_ps(vc5, vt8, vc4);
98 
99     vp0 = _mm512_fmadd_ps(vp0, vt0, vc3);
100     vp1 = _mm512_fmadd_ps(vp1, vt1, vc3);
101     vp2 = _mm512_fmadd_ps(vp2, vt2, vc3);
102     vp3 = _mm512_fmadd_ps(vp3, vt3, vc3);
103     vp4 = _mm512_fmadd_ps(vp4, vt4, vc3);
104     vp5 = _mm512_fmadd_ps(vp5, vt5, vc3);
105     vp6 = _mm512_fmadd_ps(vp6, vt6, vc3);
106     vp7 = _mm512_fmadd_ps(vp7, vt7, vc3);
107     vp8 = _mm512_fmadd_ps(vp8, vt8, vc3);
108 
109     vp0 = _mm512_fmadd_ps(vp0, vt0, vc2);
110     vp1 = _mm512_fmadd_ps(vp1, vt1, vc2);
111     vp2 = _mm512_fmadd_ps(vp2, vt2, vc2);
112     vp3 = _mm512_fmadd_ps(vp3, vt3, vc2);
113     vp4 = _mm512_fmadd_ps(vp4, vt4, vc2);
114     vp5 = _mm512_fmadd_ps(vp5, vt5, vc2);
115     vp6 = _mm512_fmadd_ps(vp6, vt6, vc2);
116     vp7 = _mm512_fmadd_ps(vp7, vt7, vc2);
117     vp8 = _mm512_fmadd_ps(vp8, vt8, vc2);
118 
119     vp0 = _mm512_fmadd_ps(vp0, vt0, vc1);
120     vp1 = _mm512_fmadd_ps(vp1, vt1, vc1);
121     vp2 = _mm512_fmadd_ps(vp2, vt2, vc1);
122     vp3 = _mm512_fmadd_ps(vp3, vt3, vc1);
123     vp4 = _mm512_fmadd_ps(vp4, vt4, vc1);
124     vp5 = _mm512_fmadd_ps(vp5, vt5, vc1);
125     vp6 = _mm512_fmadd_ps(vp6, vt6, vc1);
126     vp7 = _mm512_fmadd_ps(vp7, vt7, vc1);
127     vp8 = _mm512_fmadd_ps(vp8, vt8, vc1);
128 
129     vp0 = _mm512_fmadd_ps(vp0, vt0, vc0);
130     vp1 = _mm512_fmadd_ps(vp1, vt1, vc0);
131     vp2 = _mm512_fmadd_ps(vp2, vt2, vc0);
132     vp3 = _mm512_fmadd_ps(vp3, vt3, vc0);
133     vp4 = _mm512_fmadd_ps(vp4, vt4, vc0);
134     vp5 = _mm512_fmadd_ps(vp5, vt5, vc0);
135     vp6 = _mm512_fmadd_ps(vp6, vt6, vc0);
136     vp7 = _mm512_fmadd_ps(vp7, vt7, vc0);
137     vp8 = _mm512_fmadd_ps(vp8, vt8, vc0);
138 
139     // Accumulate "extended" floating-point numbers in ("mantissa", "exponent") representation where
140     //  - vnX is "exponent"
141     //  - vpX is "mantissa"
142     //
143     // exp2(ae) * av + exp2(be) * bv =
144     //   = exp2(max(ae, be)) * exp2(ae - max(ae, be)) * av + exp2(max(ae, be)) * exp2(be - max(ae, be)) * bv
145     //   = exp2(max_e) * (exp2(ae - max_e) * av + exp2(be - max_e) * bv)
146     //   = exp2(max_e) * (exp2(delta_ae) * av + exp2(delta_be) * bv)
147     //
148     // For computational efficiency we add three "extended" floating-point numbers at a time.
149     __m512 vmax_e0 = _mm512_max_ps(vacce0, vn0);
150     vmax_e0 = _mm512_max_ps(vmax_e0, vn1);
151     vmax_e0 = _mm512_max_ps(vmax_e0, vn2);
152     vmax_e0 = _mm512_max_ps(vmax_e0, vn3);
153     vmax_e0 = _mm512_max_ps(vmax_e0, vn4);
154     vmax_e0 = _mm512_max_ps(vmax_e0, vn5);
155     vmax_e0 = _mm512_max_ps(vmax_e0, vn6);
156     vmax_e0 = _mm512_max_ps(vmax_e0, vn7);
157     vmax_e0 = _mm512_max_ps(vmax_e0, vn8);
158 
159     const __m512 vdelta_acce0 = _mm512_sub_ps(vacce0, vmax_e0);
160     const __m512 vdelta_e0 = _mm512_sub_ps(vn0, vmax_e0);
161     const __m512 vdelta_e1 = _mm512_sub_ps(vn1, vmax_e0);
162     const __m512 vdelta_e2 = _mm512_sub_ps(vn2, vmax_e0);
163     const __m512 vdelta_e3 = _mm512_sub_ps(vn3, vmax_e0);
164     const __m512 vdelta_e4 = _mm512_sub_ps(vn4, vmax_e0);
165     const __m512 vdelta_e5 = _mm512_sub_ps(vn5, vmax_e0);
166     const __m512 vdelta_e6 = _mm512_sub_ps(vn6, vmax_e0);
167     const __m512 vdelta_e7 = _mm512_sub_ps(vn7, vmax_e0);
168     const __m512 vdelta_e8 = _mm512_sub_ps(vn8, vmax_e0);
169 
170     // Update accumulated "mantissa" and "exponent" values
171     vaccv0 = _mm512_scalef_ps(vaccv0, vdelta_acce0);
172     vaccv0 = _mm512_add_ps(vaccv0, _mm512_scalef_ps(vp0, vdelta_e0));
173     vaccv0 = _mm512_add_ps(vaccv0, _mm512_scalef_ps(vp1, vdelta_e1));
174     vaccv0 = _mm512_add_ps(vaccv0, _mm512_scalef_ps(vp2, vdelta_e2));
175     vaccv0 = _mm512_add_ps(vaccv0, _mm512_scalef_ps(vp3, vdelta_e3));
176     vaccv0 = _mm512_add_ps(vaccv0, _mm512_scalef_ps(vp4, vdelta_e4));
177     vaccv0 = _mm512_add_ps(vaccv0, _mm512_scalef_ps(vp5, vdelta_e5));
178     vaccv0 = _mm512_add_ps(vaccv0, _mm512_scalef_ps(vp6, vdelta_e6));
179     vaccv0 = _mm512_add_ps(vaccv0, _mm512_scalef_ps(vp7, vdelta_e7));
180     vaccv0 = _mm512_add_ps(vaccv0, _mm512_scalef_ps(vp8, vdelta_e8));
181 
182     vacce0 = vmax_e0;
183   }
184 
185   // Reduce partial sums of "extended" floating-point numbers into a single "extended" SIMD vector of sums.
186   __m512 vaccv = vaccv0;
187   __m512 vacce = vacce0;
188 
189   for (; elements >= 16 * sizeof(float); elements -= 16 * sizeof(float)) {
190     // Load 16 inputs at a time.
191     const __m512 vx = _mm512_loadu_ps(x);
192     x += 16;
193 
194     // Compute reduced argument elements := round(x / log(2)).
195     const __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
196 
197     // Compute reduced argument t := x - elements * log(2).
198     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
199     __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);
200     vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
201 
202     // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
203     __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
204     vp = _mm512_fmadd_ps(vp, vt, vc3);
205     vp = _mm512_fmadd_ps(vp, vt, vc2);
206     vp = _mm512_fmadd_ps(vp, vt, vc1);
207     vp = _mm512_fmadd_ps(vp, vt, vc0);
208 
209     // Accumulate "extended" floating-point numbers in ("mantissa", "exponent") representation.
210     const __m512 vmax_e = _mm512_max_ps(vacce, vn);
211     const __m512 vdelta_acce = _mm512_sub_ps(vacce, vmax_e);
212     const __m512 vdelta_e = _mm512_sub_ps(vn, vmax_e);
213     vaccv = _mm512_scalef_ps(vaccv, vdelta_acce);
214     vaccv = _mm512_add_ps(vaccv, _mm512_scalef_ps(vp, vdelta_e));
215 
216     vacce = vmax_e;
217   }
218   if XNN_UNLIKELY(elements != 0) {
219     // Prepare mask for valid 32-bit elements (depends on elements).
220     elements >>= 2 /* log2(sizeof(float)) */;
221     const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << elements) - UINT32_C(1)));
222 
223     // Load up to 15 inputs at a time.
224     const __m512 vx = _mm512_maskz_loadu_ps(vmask, x);
225 
226     // Compute reduced argument elements := round(x / log(2)).
227     const __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
228 
229     // Compute reduced argument t := x - elements * log(2).
230     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
231     __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);
232     vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
233 
234     // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
235     __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
236     vp = _mm512_fmadd_ps(vp, vt, vc3);
237     vp = _mm512_fmadd_ps(vp, vt, vc2);
238     vp = _mm512_fmadd_ps(vp, vt, vc1);
239     vp = _mm512_fmadd_ps(vp, vt, vc0);
240 
241     // Accumulate "extended" floating-point numbers in ("mantissa", "exponent") representation.
242     const __m512 vmax_e = _mm512_mask_max_ps(vacce, vmask, vacce, vn);
243     const __m512 vdelta_acce = _mm512_sub_ps(vacce, vmax_e);
244     const __m512 vdelta_e = _mm512_sub_ps(vn, vmax_e);
245     vaccv = _mm512_mask_scalef_ps(vaccv, vmask, vaccv, vdelta_acce);
246     vaccv = _mm512_mask_add_ps(vaccv, vmask, vaccv, _mm512_maskz_scalef_ps(vmask, vp, vdelta_e));
247     vacce = vmax_e;
248   }
249 
250   // Reduce partial sums of "extended" floating-point numbers into a single "extended" floating-point sum.
251   const float vmax_acce = _mm512_reduce_max_ps(vacce);
252   const __m512 vdelta_acce = _mm512_sub_ps(vacce, _mm512_set1_ps(vmax_acce));
253 
254   sum[0] = _mm512_reduce_add_ps(_mm512_scalef_ps(vaccv, vdelta_acce));
255   sum[1] = vmax_acce;
256 
257   _mm256_zeroupper();
258 }
259