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_x64(size_t elements,const float * x,float * y,float scale_value,float scale_exp)19 void xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x64(
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 >= 64 * sizeof(float); elements -= 64 * sizeof(float)) {
43 // Load 64 (4x16) 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 x += 64;
49
50 // Compute reduced argument elements := round(x / log(2)).
51 const __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0);
52 const __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0);
53 const __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0);
54 const __m512 vn3 = _mm512_roundscale_ps(_mm512_mul_ps(vx3, vlog2e), 0);
55
56 // Compute reduced argument t := x - elements * log(2).
57 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
58 __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_hi, vx0);
59 __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_hi, vx1);
60 __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_hi, vx2);
61 __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_hi, vx3);
62
63 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0);
64 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_lo, vt1);
65 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_lo, vt2);
66 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_lo, vt3);
67
68 // Compute degree-5 polynomial approxiatmion for exp(t) on [-log(2)/2, log(2)/2].
69 __m512 vp0 = _mm512_fmadd_ps(vc5, vt0, vc4);
70 __m512 vp1 = _mm512_fmadd_ps(vc5, vt1, vc4);
71 __m512 vp2 = _mm512_fmadd_ps(vc5, vt2, vc4);
72 __m512 vp3 = _mm512_fmadd_ps(vc5, vt3, vc4);
73
74 vp0 = _mm512_fmadd_ps(vp0, vt0, vc3);
75 vp1 = _mm512_fmadd_ps(vp1, vt1, vc3);
76 vp2 = _mm512_fmadd_ps(vp2, vt2, vc3);
77 vp3 = _mm512_fmadd_ps(vp3, vt3, vc3);
78
79 vp0 = _mm512_fmadd_ps(vp0, vt0, vc2);
80 vp1 = _mm512_fmadd_ps(vp1, vt1, vc2);
81 vp2 = _mm512_fmadd_ps(vp2, vt2, vc2);
82 vp3 = _mm512_fmadd_ps(vp3, vt3, vc2);
83
84 vp0 = _mm512_fmadd_ps(vp0, vt0, vc1);
85 vp1 = _mm512_fmadd_ps(vp1, vt1, vc1);
86 vp2 = _mm512_fmadd_ps(vp2, vt2, vc1);
87 vp3 = _mm512_fmadd_ps(vp3, vt3, vc1);
88
89 vp0 = _mm512_fmadd_ps(vp0, vt0, vc0);
90 vp1 = _mm512_fmadd_ps(vp1, vt1, vc0);
91 vp2 = _mm512_fmadd_ps(vp2, vt2, vc0);
92 vp3 = _mm512_fmadd_ps(vp3, vt3, vc0);
93
94 // Multiply "extended" floating-point numbers in ("mantissa", "exponent") representation where
95 // - vnX is "exponent"
96 // - vpX is "mantissa"
97 //
98 // exp2(ae) * av * exp2(be) * bv =
99 // = exp2(ae + be) * (av * bv)
100 __m512 vf0 = _mm512_mul_ps(vp0, vscalev);
101 __m512 vf1 = _mm512_mul_ps(vp1, vscalev);
102 __m512 vf2 = _mm512_mul_ps(vp2, vscalev);
103 __m512 vf3 = _mm512_mul_ps(vp3, vscalev);
104
105 const __m512 ve0 = _mm512_add_ps(vn0, vscalee);
106 const __m512 ve1 = _mm512_add_ps(vn1, vscalee);
107 const __m512 ve2 = _mm512_add_ps(vn2, vscalee);
108 const __m512 ve3 = _mm512_add_ps(vn3, vscalee);
109
110 // Multiply "mantissa" by the exp2("exponent").
111 vf0 = _mm512_scalef_ps(vf0, ve0);
112 vf1 = _mm512_scalef_ps(vf1, ve1);
113 vf2 = _mm512_scalef_ps(vf2, ve2);
114 vf3 = _mm512_scalef_ps(vf3, ve3);
115
116 // Store 128 (8x16) results at a time.
117 _mm512_storeu_ps(y, vf0);
118 _mm512_storeu_ps(y + 0, vf0);
119 _mm512_storeu_ps(y + 16, vf1);
120 _mm512_storeu_ps(y + 32, vf2);
121 _mm512_storeu_ps(y + 48, vf3);
122 y += 64;
123 }
124
125 for (; elements >= 16 * sizeof(float); elements -= 16 * sizeof(float)) {
126 // Load 16 inputs at a time.
127 const __m512 vx = _mm512_loadu_ps(x);
128 x += 16;
129
130 // Compute reduced argument elements := round(x / log(2)).
131 const __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
132
133 // Compute reduced argument t := x - elements * log(2).
134 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
135 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);
136 vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
137
138 // Compute degree-5 polynomial approxiatmion for exp(t) on [-log(2)/2, log(2)/2].
139 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
140 vp = _mm512_fmadd_ps(vp, vt, vc3);
141 vp = _mm512_fmadd_ps(vp, vt, vc2);
142 vp = _mm512_fmadd_ps(vp, vt, vc1);
143 vp = _mm512_fmadd_ps(vp, vt, vc0);
144
145 // Multiply "extended" floating-point numbers in ("mantissa", "exponent") representation.
146 __m512 vf = _mm512_mul_ps(vp, vscalev);
147 const __m512 ve = _mm512_add_ps(vn, vscalee);
148
149 // Multiply "mantissa" by the exp2("exponent").
150 vf = _mm512_scalef_ps(vf, ve);
151
152 // Store 16 results at a time.
153 _mm512_storeu_ps(y, vf);
154 y += 16;
155 }
156 if XNN_UNLIKELY(elements != 0) {
157 // Prepare mask for valid 32-bit elements (depends on elements).
158 elements >>= 2 /* log2(sizeof(float)) */;
159 const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << elements) - UINT32_C(1)));
160
161 // Load up to 15 inputs at a time.
162 const __m512 vx = _mm512_maskz_loadu_ps(vmask, x);
163
164 // Compute reduced argument elements := round(x / log(2)).
165 const __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
166
167 // Compute reduced argument t := x - elements * log(2).
168 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
169 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);
170 vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
171
172 // Compute degree-5 polynomial approxiatmion for exp(t) on [-log(2)/2, log(2)/2].
173 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
174 vp = _mm512_fmadd_ps(vp, vt, vc3);
175 vp = _mm512_fmadd_ps(vp, vt, vc2);
176 vp = _mm512_fmadd_ps(vp, vt, vc1);
177 vp = _mm512_fmadd_ps(vp, vt, vc0);
178
179 // Multiply "extended" floating-point numbers in ("mantissa", "exponent") representation.
180 __m512 vf = _mm512_mul_ps(vp, vscalev);
181 const __m512 ve = _mm512_add_ps(vn, vscalee);
182
183 // Multiply "mantissa" by the exp2("exponent").
184 vf = _mm512_scalef_ps(vf, ve);
185
186 // Store up to 15 results at a time.
187 _mm512_mask_storeu_ps(y, vmask, vf);
188 }
189 _mm256_zeroupper();
190 }
191