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_x48(size_t elements,const float * x,float * y,float scale_value,float scale_exp)19 void xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x48(
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 >= 48 * sizeof(float); elements -= 48 * sizeof(float)) {
43 // Load 48 (3x16) 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 x += 48;
48
49 // Compute reduced argument elements := round(x / log(2)).
50 const __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0);
51 const __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0);
52 const __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0);
53
54 // Compute reduced argument t := x - elements * log(2).
55 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
56 __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_hi, vx0);
57 __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_hi, vx1);
58 __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_hi, vx2);
59
60 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0);
61 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_lo, vt1);
62 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_lo, vt2);
63
64 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
65 __m512 vp0 = _mm512_fmadd_ps(vc5, vt0, vc4);
66 __m512 vp1 = _mm512_fmadd_ps(vc5, vt1, vc4);
67 __m512 vp2 = _mm512_fmadd_ps(vc5, vt2, vc4);
68
69 vp0 = _mm512_fmadd_ps(vp0, vt0, vc3);
70 vp1 = _mm512_fmadd_ps(vp1, vt1, vc3);
71 vp2 = _mm512_fmadd_ps(vp2, vt2, vc3);
72
73 vp0 = _mm512_fmadd_ps(vp0, vt0, vc2);
74 vp1 = _mm512_fmadd_ps(vp1, vt1, vc2);
75 vp2 = _mm512_fmadd_ps(vp2, vt2, vc2);
76
77 vp0 = _mm512_fmadd_ps(vp0, vt0, vc1);
78 vp1 = _mm512_fmadd_ps(vp1, vt1, vc1);
79 vp2 = _mm512_fmadd_ps(vp2, vt2, vc1);
80
81 vp0 = _mm512_fmadd_ps(vp0, vt0, vc0);
82 vp1 = _mm512_fmadd_ps(vp1, vt1, vc0);
83 vp2 = _mm512_fmadd_ps(vp2, vt2, vc0);
84
85 // Multiply "extended" floating-point numbers in ("mantissa", "exponent") representation where
86 // - vnX is "exponent"
87 // - vpX is "mantissa"
88 //
89 // exp2(ae) * av * exp2(be) * bv =
90 // = exp2(ae + be) * (av * bv)
91 __m512 vf0 = _mm512_mul_ps(vp0, vscalev);
92 __m512 vf1 = _mm512_mul_ps(vp1, vscalev);
93 __m512 vf2 = _mm512_mul_ps(vp2, vscalev);
94
95 const __m512 ve0 = _mm512_add_ps(vn0, vscalee);
96 const __m512 ve1 = _mm512_add_ps(vn1, vscalee);
97 const __m512 ve2 = _mm512_add_ps(vn2, vscalee);
98
99 // Multiply "mantissa" by the exp2("exponent").
100 vf0 = _mm512_scalef_ps(vf0, ve0);
101 vf1 = _mm512_scalef_ps(vf1, ve1);
102 vf2 = _mm512_scalef_ps(vf2, ve2);
103
104 // Store 128 (8x16) results at a time.
105 _mm512_storeu_ps(y, vf0);
106 _mm512_storeu_ps(y + 0, vf0);
107 _mm512_storeu_ps(y + 16, vf1);
108 _mm512_storeu_ps(y + 32, vf2);
109 y += 48;
110 }
111
112 for (; elements >= 16 * sizeof(float); elements -= 16 * sizeof(float)) {
113 // Load 16 inputs at a time.
114 const __m512 vx = _mm512_loadu_ps(x);
115 x += 16;
116
117 // Compute reduced argument elements := round(x / log(2)).
118 const __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
119
120 // Compute reduced argument t := x - elements * log(2).
121 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
122 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);
123 vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
124
125 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
126 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
127 vp = _mm512_fmadd_ps(vp, vt, vc3);
128 vp = _mm512_fmadd_ps(vp, vt, vc2);
129 vp = _mm512_fmadd_ps(vp, vt, vc1);
130 vp = _mm512_fmadd_ps(vp, vt, vc0);
131
132 // Multiply "extended" floating-point numbers in ("mantissa", "exponent") representation.
133 __m512 vf = _mm512_mul_ps(vp, vscalev);
134 const __m512 ve = _mm512_add_ps(vn, vscalee);
135
136 // Multiply "mantissa" by the exp2("exponent").
137 vf = _mm512_scalef_ps(vf, ve);
138
139 // Store 16 results at a time.
140 _mm512_storeu_ps(y, vf);
141 y += 16;
142 }
143 if XNN_UNLIKELY(elements != 0) {
144 // Prepare mask for valid 32-bit elements (depends on elements).
145 elements >>= 2 /* log2(sizeof(float)) */;
146 const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << elements) - UINT32_C(1)));
147
148 // Load up to 15 inputs at a time.
149 const __m512 vx = _mm512_maskz_loadu_ps(vmask, x);
150
151 // Compute reduced argument elements := round(x / log(2)).
152 const __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
153
154 // Compute reduced argument t := x - elements * log(2).
155 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
156 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);
157 vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
158
159 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
160 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
161 vp = _mm512_fmadd_ps(vp, vt, vc3);
162 vp = _mm512_fmadd_ps(vp, vt, vc2);
163 vp = _mm512_fmadd_ps(vp, vt, vc1);
164 vp = _mm512_fmadd_ps(vp, vt, vc0);
165
166 // Multiply "extended" floating-point numbers in ("mantissa", "exponent") representation.
167 __m512 vf = _mm512_mul_ps(vp, vscalev);
168 const __m512 ve = _mm512_add_ps(vn, vscalee);
169
170 // Multiply "mantissa" by the exp2("exponent").
171 vf = _mm512_scalef_ps(vf, ve);
172
173 // Store up to 15 results at a time.
174 _mm512_mask_storeu_ps(y, vmask, vf);
175 }
176 _mm256_zeroupper();
177 }
178