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_x64(size_t elements,const float * input,float * output,float scale,float max)18 void xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x64(
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 >= 64 * sizeof(float); elements -= 64 * sizeof(float)) {
42 // Load 64 (4x16) 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 input += 64;
48
49 // Subtract maximum input x := i - i_max.
50 const __m512 vx0 = _mm512_sub_ps(vi0, vi_max);
51 const __m512 vx1 = _mm512_sub_ps(vi1, vi_max);
52 const __m512 vx2 = _mm512_sub_ps(vi2, vi_max);
53 const __m512 vx3 = _mm512_sub_ps(vi3, vi_max);
54
55 // Compute reduced argument elements := round(x / log(2)).
56 __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0);
57 __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0);
58 __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0);
59 __m512 vn3 = _mm512_roundscale_ps(_mm512_mul_ps(vx3, vlog2e), 0);
60
61 // Compute reduced argument t := x - elements * log(2).
62 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
63 __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_hi, vx0);
64 __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_hi, vx1);
65 __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_hi, vx2);
66 __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_hi, vx3);
67
68 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0);
69 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_lo, vt1);
70 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_lo, vt2);
71 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_lo, vt3);
72
73 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
74 __m512 vp0 = _mm512_fmadd_ps(vc5, vt0, vc4);
75 __m512 vp1 = _mm512_fmadd_ps(vc5, vt1, vc4);
76 __m512 vp2 = _mm512_fmadd_ps(vc5, vt2, vc4);
77 __m512 vp3 = _mm512_fmadd_ps(vc5, vt3, 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
84 vp0 = _mm512_fmadd_ps(vp0, vt0, vc2);
85 vp1 = _mm512_fmadd_ps(vp1, vt1, vc2);
86 vp2 = _mm512_fmadd_ps(vp2, vt2, vc2);
87 vp3 = _mm512_fmadd_ps(vp3, vt3, vc2);
88
89 vp0 = _mm512_fmadd_ps(vp0, vt0, vc1);
90 vp1 = _mm512_fmadd_ps(vp1, vt1, vc1);
91 vp2 = _mm512_fmadd_ps(vp2, vt2, vc1);
92 vp3 = _mm512_fmadd_ps(vp3, vt3, vc1);
93
94 vp0 = _mm512_fmadd_ps(vp0, vt0, vc0);
95 vp1 = _mm512_fmadd_ps(vp1, vt1, vc0);
96 vp2 = _mm512_fmadd_ps(vp2, vt2, vc0);
97 vp3 = _mm512_fmadd_ps(vp3, vt3, vc0);
98
99 // Reconstruct the final f value:
100 // f = 2**elements * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
101 // = 2**elements * p
102 __m512 vf0 = _mm512_scalef_ps(vp0, vn0);
103 __m512 vf1 = _mm512_scalef_ps(vp1, vn1);
104 __m512 vf2 = _mm512_scalef_ps(vp2, vn2);
105 __m512 vf3 = _mm512_scalef_ps(vp3, vn3);
106
107 // Multiply by scale.
108 vf0 = _mm512_mul_ps(vf0, vscale);
109 vf1 = _mm512_mul_ps(vf1, vscale);
110 vf2 = _mm512_mul_ps(vf2, vscale);
111 vf3 = _mm512_mul_ps(vf3, vscale);
112
113 // Store 64 (4x16) outputs at a time.
114 _mm512_storeu_ps(output, vf0);
115 _mm512_storeu_ps(output + 0, vf0);
116 _mm512_storeu_ps(output + 16, vf1);
117 _mm512_storeu_ps(output + 32, vf2);
118 _mm512_storeu_ps(output + 48, vf3);
119 output += 64;
120 }
121 for (; elements >= 16 * sizeof(float); elements -= 16 * sizeof(float)) {
122 // Load 16 inputs at a time.
123 const __m512 vi = _mm512_loadu_ps(input);
124 input += 16;
125
126 // Subtract maximum input x := i - i_max.
127 const __m512 vx = _mm512_sub_ps(vi, vi_max);
128
129 // Compute reduced argument elements := round(x / log(2)).
130 __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
131
132 // Compute reduced argument t := x - elements * log(2).
133 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
134 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);
135 vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
136
137 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
138 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
139 vp = _mm512_fmadd_ps(vp, vt, vc3);
140 vp = _mm512_fmadd_ps(vp, vt, vc2);
141 vp = _mm512_fmadd_ps(vp, vt, vc1);
142 vp = _mm512_fmadd_ps(vp, vt, vc0);
143
144 // Reconstruct the final f value:
145 // f = 2**elements * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
146 // = 2**elements * p
147 __m512 vf = _mm512_scalef_ps(vp, vn);
148
149 // Multiply by scale.
150 vf = _mm512_mul_ps(vf, vscale);
151
152 // Store 16 outputs at a time.
153 _mm512_storeu_ps(output, vf);
154 output += 16;
155 }
156 if (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 vi = _mm512_mask_loadu_ps(_mm512_undefined_ps(), vmask, input);
163
164 // Subtract maximum input x := i - i_max.
165 const __m512 vx = _mm512_sub_ps(vi, vi_max);
166
167 // Compute reduced argument elements := round(x / log(2)).
168 __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
169
170 // Compute reduced argument t := x - elements * log(2).
171 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
172 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);
173 vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
174
175 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
176 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
177 vp = _mm512_fmadd_ps(vp, vt, vc3);
178 vp = _mm512_fmadd_ps(vp, vt, vc2);
179 vp = _mm512_fmadd_ps(vp, vt, vc1);
180 vp = _mm512_fmadd_ps(vp, vt, vc0);
181
182 // Reconstruct the final f value:
183 // f = 2**elements * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
184 // = 2**elements * p
185 __m512 vf = _mm512_scalef_ps(vp, vn);
186
187 // Multiply by scale.
188 vf = _mm512_mul_ps(vf, vscale);
189
190 // Store up to 15 outputs at a time.
191 _mm512_mask_storeu_ps(output, vmask, vf);
192 }
193 }
194