1 // Auto-generated file. Do not edit!
2 // Template: src/f32-vscaleexpminusmax/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/intrinsics-polyfill.h>
15 #include <xnnpack/vscaleexpminusmax.h>
16
17
xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x96(size_t elements,const float * input,float * output,float scale,float max)18 void xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x96(
19 size_t elements,
20 const float* input,
21 float* output,
22 float scale,
23 float max)
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 vscale = _mm512_set1_ps(scale);
39 const __m512 vi_max = _mm512_set1_ps(max);
40
41 for (; elements >= 96 * sizeof(float); elements -= 96 * sizeof(float)) {
42 // Load 96 (6x16) inputs at a time.
43 const __m512 vi0 = _mm512_loadu_ps(input);
44 const __m512 vi1 = _mm512_loadu_ps(input + 16);
45 const __m512 vi2 = _mm512_loadu_ps(input + 32);
46 const __m512 vi3 = _mm512_loadu_ps(input + 48);
47 const __m512 vi4 = _mm512_loadu_ps(input + 64);
48 const __m512 vi5 = _mm512_loadu_ps(input + 80);
49 input += 96;
50
51 // Subtract maximum input x := i - i_max.
52 const __m512 vx0 = _mm512_sub_ps(vi0, vi_max);
53 const __m512 vx1 = _mm512_sub_ps(vi1, vi_max);
54 const __m512 vx2 = _mm512_sub_ps(vi2, vi_max);
55 const __m512 vx3 = _mm512_sub_ps(vi3, vi_max);
56 const __m512 vx4 = _mm512_sub_ps(vi4, vi_max);
57 const __m512 vx5 = _mm512_sub_ps(vi5, vi_max);
58
59 // Compute reduced argument elements := round(x / log(2)).
60 __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0);
61 __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0);
62 __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0);
63 __m512 vn3 = _mm512_roundscale_ps(_mm512_mul_ps(vx3, vlog2e), 0);
64 __m512 vn4 = _mm512_roundscale_ps(_mm512_mul_ps(vx4, vlog2e), 0);
65 __m512 vn5 = _mm512_roundscale_ps(_mm512_mul_ps(vx5, vlog2e), 0);
66
67 // Compute reduced argument t := x - elements * log(2).
68 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
69 __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_hi, vx0);
70 __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_hi, vx1);
71 __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_hi, vx2);
72 __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_hi, vx3);
73 __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_hi, vx4);
74 __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_hi, vx5);
75
76 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0);
77 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_lo, vt1);
78 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_lo, vt2);
79 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_lo, vt3);
80 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_lo, vt4);
81 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_lo, vt5);
82
83 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
84 __m512 vp0 = _mm512_fmadd_ps(vc5, vt0, vc4);
85 __m512 vp1 = _mm512_fmadd_ps(vc5, vt1, vc4);
86 __m512 vp2 = _mm512_fmadd_ps(vc5, vt2, vc4);
87 __m512 vp3 = _mm512_fmadd_ps(vc5, vt3, vc4);
88 __m512 vp4 = _mm512_fmadd_ps(vc5, vt4, vc4);
89 __m512 vp5 = _mm512_fmadd_ps(vc5, vt5, vc4);
90
91 vp0 = _mm512_fmadd_ps(vp0, vt0, vc3);
92 vp1 = _mm512_fmadd_ps(vp1, vt1, vc3);
93 vp2 = _mm512_fmadd_ps(vp2, vt2, vc3);
94 vp3 = _mm512_fmadd_ps(vp3, vt3, vc3);
95 vp4 = _mm512_fmadd_ps(vp4, vt4, vc3);
96 vp5 = _mm512_fmadd_ps(vp5, vt5, vc3);
97
98 vp0 = _mm512_fmadd_ps(vp0, vt0, vc2);
99 vp1 = _mm512_fmadd_ps(vp1, vt1, vc2);
100 vp2 = _mm512_fmadd_ps(vp2, vt2, vc2);
101 vp3 = _mm512_fmadd_ps(vp3, vt3, vc2);
102 vp4 = _mm512_fmadd_ps(vp4, vt4, vc2);
103 vp5 = _mm512_fmadd_ps(vp5, vt5, 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
112 vp0 = _mm512_fmadd_ps(vp0, vt0, vc0);
113 vp1 = _mm512_fmadd_ps(vp1, vt1, vc0);
114 vp2 = _mm512_fmadd_ps(vp2, vt2, vc0);
115 vp3 = _mm512_fmadd_ps(vp3, vt3, vc0);
116 vp4 = _mm512_fmadd_ps(vp4, vt4, vc0);
117 vp5 = _mm512_fmadd_ps(vp5, vt5, vc0);
118
119 // Reconstruct the final f value:
120 // f = 2**elements * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
121 // = 2**elements * p
122 __m512 vf0 = _mm512_scalef_ps(vp0, vn0);
123 __m512 vf1 = _mm512_scalef_ps(vp1, vn1);
124 __m512 vf2 = _mm512_scalef_ps(vp2, vn2);
125 __m512 vf3 = _mm512_scalef_ps(vp3, vn3);
126 __m512 vf4 = _mm512_scalef_ps(vp4, vn4);
127 __m512 vf5 = _mm512_scalef_ps(vp5, vn5);
128
129 // Multiply by scale.
130 vf0 = _mm512_mul_ps(vf0, vscale);
131 vf1 = _mm512_mul_ps(vf1, vscale);
132 vf2 = _mm512_mul_ps(vf2, vscale);
133 vf3 = _mm512_mul_ps(vf3, vscale);
134 vf4 = _mm512_mul_ps(vf4, vscale);
135 vf5 = _mm512_mul_ps(vf5, vscale);
136
137 // Store 96 (6x16) outputs at a time.
138 _mm512_storeu_ps(output, vf0);
139 _mm512_storeu_ps(output + 0, vf0);
140 _mm512_storeu_ps(output + 16, vf1);
141 _mm512_storeu_ps(output + 32, vf2);
142 _mm512_storeu_ps(output + 48, vf3);
143 _mm512_storeu_ps(output + 64, vf4);
144 _mm512_storeu_ps(output + 80, vf5);
145 output += 96;
146 }
147 for (; elements >= 16 * sizeof(float); elements -= 16 * sizeof(float)) {
148 // Load 16 inputs at a time.
149 const __m512 vi = _mm512_loadu_ps(input);
150 input += 16;
151
152 // Subtract maximum input x := i - i_max.
153 const __m512 vx = _mm512_sub_ps(vi, vi_max);
154
155 // Compute reduced argument elements := round(x / log(2)).
156 __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
157
158 // Compute reduced argument t := x - elements * log(2).
159 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
160 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);
161 vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
162
163 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
164 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
165 vp = _mm512_fmadd_ps(vp, vt, vc3);
166 vp = _mm512_fmadd_ps(vp, vt, vc2);
167 vp = _mm512_fmadd_ps(vp, vt, vc1);
168 vp = _mm512_fmadd_ps(vp, vt, vc0);
169
170 // Reconstruct the final f value:
171 // f = 2**elements * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
172 // = 2**elements * p
173 __m512 vf = _mm512_scalef_ps(vp, vn);
174
175 // Multiply by scale.
176 vf = _mm512_mul_ps(vf, vscale);
177
178 // Store 16 outputs at a time.
179 _mm512_storeu_ps(output, vf);
180 output += 16;
181 }
182 if (elements != 0) {
183 // Prepare mask for valid 32-bit elements (depends on elements).
184 elements >>= 2 /* log2(sizeof(float)) */;
185 const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << elements) - UINT32_C(1)));
186
187 // Load up to 15 inputs at a time.
188 const __m512 vi = _mm512_mask_loadu_ps(_mm512_undefined_ps(), vmask, input);
189
190 // Subtract maximum input x := i - i_max.
191 const __m512 vx = _mm512_sub_ps(vi, vi_max);
192
193 // Compute reduced argument elements := round(x / log(2)).
194 __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
195
196 // Compute reduced argument t := x - elements * log(2).
197 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
198 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);
199 vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
200
201 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
202 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
203 vp = _mm512_fmadd_ps(vp, vt, vc3);
204 vp = _mm512_fmadd_ps(vp, vt, vc2);
205 vp = _mm512_fmadd_ps(vp, vt, vc1);
206 vp = _mm512_fmadd_ps(vp, vt, vc0);
207
208 // Reconstruct the final f value:
209 // f = 2**elements * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
210 // = 2**elements * p
211 __m512 vf = _mm512_scalef_ps(vp, vn);
212
213 // Multiply by scale.
214 vf = _mm512_mul_ps(vf, vscale);
215
216 // Store up to 15 outputs at a time.
217 _mm512_mask_storeu_ps(output, vmask, vf);
218 }
219 }
220