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_x112(size_t elements,const float * x,float * y,float scale_value,float scale_exp)19 void xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x112(
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 >= 112 * sizeof(float); elements -= 112 * sizeof(float)) {
43 // Load 112 (7x16) 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 x += 112;
52
53 // Compute reduced argument elements := round(x / log(2)).
54 const __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0);
55 const __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0);
56 const __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0);
57 const __m512 vn3 = _mm512_roundscale_ps(_mm512_mul_ps(vx3, vlog2e), 0);
58 const __m512 vn4 = _mm512_roundscale_ps(_mm512_mul_ps(vx4, vlog2e), 0);
59 const __m512 vn5 = _mm512_roundscale_ps(_mm512_mul_ps(vx5, vlog2e), 0);
60 const __m512 vn6 = _mm512_roundscale_ps(_mm512_mul_ps(vx6, vlog2e), 0);
61
62 // Compute reduced argument t := x - elements * log(2).
63 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
64 __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_hi, vx0);
65 __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_hi, vx1);
66 __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_hi, vx2);
67 __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_hi, vx3);
68 __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_hi, vx4);
69 __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_hi, vx5);
70 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_hi, vx6);
71
72 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0);
73 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_lo, vt1);
74 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_lo, vt2);
75 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_lo, vt3);
76 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_lo, vt4);
77 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_lo, vt5);
78 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_lo, vt6);
79
80 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
81 __m512 vp0 = _mm512_fmadd_ps(vc5, vt0, vc4);
82 __m512 vp1 = _mm512_fmadd_ps(vc5, vt1, vc4);
83 __m512 vp2 = _mm512_fmadd_ps(vc5, vt2, vc4);
84 __m512 vp3 = _mm512_fmadd_ps(vc5, vt3, vc4);
85 __m512 vp4 = _mm512_fmadd_ps(vc5, vt4, vc4);
86 __m512 vp5 = _mm512_fmadd_ps(vc5, vt5, vc4);
87 __m512 vp6 = _mm512_fmadd_ps(vc5, vt6, vc4);
88
89 vp0 = _mm512_fmadd_ps(vp0, vt0, vc3);
90 vp1 = _mm512_fmadd_ps(vp1, vt1, vc3);
91 vp2 = _mm512_fmadd_ps(vp2, vt2, vc3);
92 vp3 = _mm512_fmadd_ps(vp3, vt3, vc3);
93 vp4 = _mm512_fmadd_ps(vp4, vt4, vc3);
94 vp5 = _mm512_fmadd_ps(vp5, vt5, vc3);
95 vp6 = _mm512_fmadd_ps(vp6, vt6, vc3);
96
97 vp0 = _mm512_fmadd_ps(vp0, vt0, vc2);
98 vp1 = _mm512_fmadd_ps(vp1, vt1, vc2);
99 vp2 = _mm512_fmadd_ps(vp2, vt2, vc2);
100 vp3 = _mm512_fmadd_ps(vp3, vt3, vc2);
101 vp4 = _mm512_fmadd_ps(vp4, vt4, vc2);
102 vp5 = _mm512_fmadd_ps(vp5, vt5, vc2);
103 vp6 = _mm512_fmadd_ps(vp6, vt6, vc2);
104
105 vp0 = _mm512_fmadd_ps(vp0, vt0, vc1);
106 vp1 = _mm512_fmadd_ps(vp1, vt1, vc1);
107 vp2 = _mm512_fmadd_ps(vp2, vt2, vc1);
108 vp3 = _mm512_fmadd_ps(vp3, vt3, vc1);
109 vp4 = _mm512_fmadd_ps(vp4, vt4, vc1);
110 vp5 = _mm512_fmadd_ps(vp5, vt5, vc1);
111 vp6 = _mm512_fmadd_ps(vp6, vt6, vc1);
112
113 vp0 = _mm512_fmadd_ps(vp0, vt0, vc0);
114 vp1 = _mm512_fmadd_ps(vp1, vt1, vc0);
115 vp2 = _mm512_fmadd_ps(vp2, vt2, vc0);
116 vp3 = _mm512_fmadd_ps(vp3, vt3, vc0);
117 vp4 = _mm512_fmadd_ps(vp4, vt4, vc0);
118 vp5 = _mm512_fmadd_ps(vp5, vt5, vc0);
119 vp6 = _mm512_fmadd_ps(vp6, vt6, vc0);
120
121 // Multiply "extended" floating-point numbers in ("mantissa", "exponent") representation where
122 // - vnX is "exponent"
123 // - vpX is "mantissa"
124 //
125 // exp2(ae) * av * exp2(be) * bv =
126 // = exp2(ae + be) * (av * bv)
127 __m512 vf0 = _mm512_mul_ps(vp0, vscalev);
128 __m512 vf1 = _mm512_mul_ps(vp1, vscalev);
129 __m512 vf2 = _mm512_mul_ps(vp2, vscalev);
130 __m512 vf3 = _mm512_mul_ps(vp3, vscalev);
131 __m512 vf4 = _mm512_mul_ps(vp4, vscalev);
132 __m512 vf5 = _mm512_mul_ps(vp5, vscalev);
133 __m512 vf6 = _mm512_mul_ps(vp6, vscalev);
134
135 const __m512 ve0 = _mm512_add_ps(vn0, vscalee);
136 const __m512 ve1 = _mm512_add_ps(vn1, vscalee);
137 const __m512 ve2 = _mm512_add_ps(vn2, vscalee);
138 const __m512 ve3 = _mm512_add_ps(vn3, vscalee);
139 const __m512 ve4 = _mm512_add_ps(vn4, vscalee);
140 const __m512 ve5 = _mm512_add_ps(vn5, vscalee);
141 const __m512 ve6 = _mm512_add_ps(vn6, vscalee);
142
143 // Multiply "mantissa" by the exp2("exponent").
144 vf0 = _mm512_scalef_ps(vf0, ve0);
145 vf1 = _mm512_scalef_ps(vf1, ve1);
146 vf2 = _mm512_scalef_ps(vf2, ve2);
147 vf3 = _mm512_scalef_ps(vf3, ve3);
148 vf4 = _mm512_scalef_ps(vf4, ve4);
149 vf5 = _mm512_scalef_ps(vf5, ve5);
150 vf6 = _mm512_scalef_ps(vf6, ve6);
151
152 // Store 128 (8x16) results at a time.
153 _mm512_storeu_ps(y, vf0);
154 _mm512_storeu_ps(y + 0, vf0);
155 _mm512_storeu_ps(y + 16, vf1);
156 _mm512_storeu_ps(y + 32, vf2);
157 _mm512_storeu_ps(y + 48, vf3);
158 _mm512_storeu_ps(y + 64, vf4);
159 _mm512_storeu_ps(y + 80, vf5);
160 _mm512_storeu_ps(y + 96, vf6);
161 y += 112;
162 }
163
164 for (; elements >= 16 * sizeof(float); elements -= 16 * sizeof(float)) {
165 // Load 16 inputs at a time.
166 const __m512 vx = _mm512_loadu_ps(x);
167 x += 16;
168
169 // Compute reduced argument elements := round(x / log(2)).
170 const __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
171
172 // Compute reduced argument t := x - elements * log(2).
173 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
174 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);
175 vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
176
177 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
178 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
179 vp = _mm512_fmadd_ps(vp, vt, vc3);
180 vp = _mm512_fmadd_ps(vp, vt, vc2);
181 vp = _mm512_fmadd_ps(vp, vt, vc1);
182 vp = _mm512_fmadd_ps(vp, vt, vc0);
183
184 // Multiply "extended" floating-point numbers in ("mantissa", "exponent") representation.
185 __m512 vf = _mm512_mul_ps(vp, vscalev);
186 const __m512 ve = _mm512_add_ps(vn, vscalee);
187
188 // Multiply "mantissa" by the exp2("exponent").
189 vf = _mm512_scalef_ps(vf, ve);
190
191 // Store 16 results at a time.
192 _mm512_storeu_ps(y, vf);
193 y += 16;
194 }
195 if XNN_UNLIKELY(elements != 0) {
196 // Prepare mask for valid 32-bit elements (depends on elements).
197 elements >>= 2 /* log2(sizeof(float)) */;
198 const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << elements) - UINT32_C(1)));
199
200 // Load up to 15 inputs at a time.
201 const __m512 vx = _mm512_maskz_loadu_ps(vmask, x);
202
203 // Compute reduced argument elements := round(x / log(2)).
204 const __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
205
206 // Compute reduced argument t := x - elements * log(2).
207 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
208 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);
209 vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
210
211 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
212 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
213 vp = _mm512_fmadd_ps(vp, vt, vc3);
214 vp = _mm512_fmadd_ps(vp, vt, vc2);
215 vp = _mm512_fmadd_ps(vp, vt, vc1);
216 vp = _mm512_fmadd_ps(vp, vt, vc0);
217
218 // Multiply "extended" floating-point numbers in ("mantissa", "exponent") representation.
219 __m512 vf = _mm512_mul_ps(vp, vscalev);
220 const __m512 ve = _mm512_add_ps(vn, vscalee);
221
222 // Multiply "mantissa" by the exp2("exponent").
223 vf = _mm512_scalef_ps(vf, ve);
224
225 // Store up to 15 results at a time.
226 _mm512_mask_storeu_ps(y, vmask, vf);
227 }
228 _mm256_zeroupper();
229 }
230