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_x80(size_t elements,const float * x,float * y,float scale_value,float scale_exp)19 void xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x80(
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 >= 80 * sizeof(float); elements -= 80 * sizeof(float)) {
43 // Load 80 (5x16) 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 x += 80;
50
51 // Compute reduced argument elements := round(x / log(2)).
52 const __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0);
53 const __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0);
54 const __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0);
55 const __m512 vn3 = _mm512_roundscale_ps(_mm512_mul_ps(vx3, vlog2e), 0);
56 const __m512 vn4 = _mm512_roundscale_ps(_mm512_mul_ps(vx4, vlog2e), 0);
57
58 // Compute reduced argument t := x - elements * log(2).
59 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
60 __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_hi, vx0);
61 __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_hi, vx1);
62 __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_hi, vx2);
63 __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_hi, vx3);
64 __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_hi, vx4);
65
66 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0);
67 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_lo, vt1);
68 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_lo, vt2);
69 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_lo, vt3);
70 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_lo, vt4);
71
72 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
73 __m512 vp0 = _mm512_fmadd_ps(vc5, vt0, vc4);
74 __m512 vp1 = _mm512_fmadd_ps(vc5, vt1, vc4);
75 __m512 vp2 = _mm512_fmadd_ps(vc5, vt2, vc4);
76 __m512 vp3 = _mm512_fmadd_ps(vc5, vt3, vc4);
77 __m512 vp4 = _mm512_fmadd_ps(vc5, vt4, vc4);
78
79 vp0 = _mm512_fmadd_ps(vp0, vt0, vc3);
80 vp1 = _mm512_fmadd_ps(vp1, vt1, vc3);
81 vp2 = _mm512_fmadd_ps(vp2, vt2, vc3);
82 vp3 = _mm512_fmadd_ps(vp3, vt3, vc3);
83 vp4 = _mm512_fmadd_ps(vp4, vt4, vc3);
84
85 vp0 = _mm512_fmadd_ps(vp0, vt0, vc2);
86 vp1 = _mm512_fmadd_ps(vp1, vt1, vc2);
87 vp2 = _mm512_fmadd_ps(vp2, vt2, vc2);
88 vp3 = _mm512_fmadd_ps(vp3, vt3, vc2);
89 vp4 = _mm512_fmadd_ps(vp4, vt4, vc2);
90
91 vp0 = _mm512_fmadd_ps(vp0, vt0, vc1);
92 vp1 = _mm512_fmadd_ps(vp1, vt1, vc1);
93 vp2 = _mm512_fmadd_ps(vp2, vt2, vc1);
94 vp3 = _mm512_fmadd_ps(vp3, vt3, vc1);
95 vp4 = _mm512_fmadd_ps(vp4, vt4, vc1);
96
97 vp0 = _mm512_fmadd_ps(vp0, vt0, vc0);
98 vp1 = _mm512_fmadd_ps(vp1, vt1, vc0);
99 vp2 = _mm512_fmadd_ps(vp2, vt2, vc0);
100 vp3 = _mm512_fmadd_ps(vp3, vt3, vc0);
101 vp4 = _mm512_fmadd_ps(vp4, vt4, vc0);
102
103 // Multiply "extended" floating-point numbers in ("mantissa", "exponent") representation where
104 // - vnX is "exponent"
105 // - vpX is "mantissa"
106 //
107 // exp2(ae) * av * exp2(be) * bv =
108 // = exp2(ae + be) * (av * bv)
109 __m512 vf0 = _mm512_mul_ps(vp0, vscalev);
110 __m512 vf1 = _mm512_mul_ps(vp1, vscalev);
111 __m512 vf2 = _mm512_mul_ps(vp2, vscalev);
112 __m512 vf3 = _mm512_mul_ps(vp3, vscalev);
113 __m512 vf4 = _mm512_mul_ps(vp4, vscalev);
114
115 const __m512 ve0 = _mm512_add_ps(vn0, vscalee);
116 const __m512 ve1 = _mm512_add_ps(vn1, vscalee);
117 const __m512 ve2 = _mm512_add_ps(vn2, vscalee);
118 const __m512 ve3 = _mm512_add_ps(vn3, vscalee);
119 const __m512 ve4 = _mm512_add_ps(vn4, vscalee);
120
121 // Multiply "mantissa" by the exp2("exponent").
122 vf0 = _mm512_scalef_ps(vf0, ve0);
123 vf1 = _mm512_scalef_ps(vf1, ve1);
124 vf2 = _mm512_scalef_ps(vf2, ve2);
125 vf3 = _mm512_scalef_ps(vf3, ve3);
126 vf4 = _mm512_scalef_ps(vf4, ve4);
127
128 // Store 128 (8x16) results at a time.
129 _mm512_storeu_ps(y, vf0);
130 _mm512_storeu_ps(y + 0, vf0);
131 _mm512_storeu_ps(y + 16, vf1);
132 _mm512_storeu_ps(y + 32, vf2);
133 _mm512_storeu_ps(y + 48, vf3);
134 _mm512_storeu_ps(y + 64, vf4);
135 y += 80;
136 }
137
138 for (; elements >= 16 * sizeof(float); elements -= 16 * sizeof(float)) {
139 // Load 16 inputs at a time.
140 const __m512 vx = _mm512_loadu_ps(x);
141 x += 16;
142
143 // Compute reduced argument elements := round(x / log(2)).
144 const __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
145
146 // Compute reduced argument t := x - elements * log(2).
147 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
148 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);
149 vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
150
151 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
152 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
153 vp = _mm512_fmadd_ps(vp, vt, vc3);
154 vp = _mm512_fmadd_ps(vp, vt, vc2);
155 vp = _mm512_fmadd_ps(vp, vt, vc1);
156 vp = _mm512_fmadd_ps(vp, vt, vc0);
157
158 // Multiply "extended" floating-point numbers in ("mantissa", "exponent") representation.
159 __m512 vf = _mm512_mul_ps(vp, vscalev);
160 const __m512 ve = _mm512_add_ps(vn, vscalee);
161
162 // Multiply "mantissa" by the exp2("exponent").
163 vf = _mm512_scalef_ps(vf, ve);
164
165 // Store 16 results at a time.
166 _mm512_storeu_ps(y, vf);
167 y += 16;
168 }
169 if XNN_UNLIKELY(elements != 0) {
170 // Prepare mask for valid 32-bit elements (depends on elements).
171 elements >>= 2 /* log2(sizeof(float)) */;
172 const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << elements) - UINT32_C(1)));
173
174 // Load up to 15 inputs at a time.
175 const __m512 vx = _mm512_maskz_loadu_ps(vmask, x);
176
177 // Compute reduced argument elements := round(x / log(2)).
178 const __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
179
180 // Compute reduced argument t := x - elements * log(2).
181 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
182 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);
183 vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
184
185 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
186 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
187 vp = _mm512_fmadd_ps(vp, vt, vc3);
188 vp = _mm512_fmadd_ps(vp, vt, vc2);
189 vp = _mm512_fmadd_ps(vp, vt, vc1);
190 vp = _mm512_fmadd_ps(vp, vt, vc0);
191
192 // Multiply "extended" floating-point numbers in ("mantissa", "exponent") representation.
193 __m512 vf = _mm512_mul_ps(vp, vscalev);
194 const __m512 ve = _mm512_add_ps(vn, vscalee);
195
196 // Multiply "mantissa" by the exp2("exponent").
197 vf = _mm512_scalef_ps(vf, ve);
198
199 // Store up to 15 results at a time.
200 _mm512_mask_storeu_ps(y, vmask, vf);
201 }
202 _mm256_zeroupper();
203 }
204