Debian Codegraph @

src/f32-vscaleexpminusmax/gen/f32-vscaleexpminusmax-avx2-p5-u8.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/f32-vscaleexpminusmax/avx2-p5.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/vscaleexpminusmax.h"
16
17
18 void xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_u8(
19 size_t batch,
20 const float* input,
21 float* output,
22 float scale,
23 float max)
24 {
25 assert(batch != 0);
26 assert(batch % sizeof(float) == 0);
27 assert(input != NULL);
28 assert(output != NULL);
29
30 static const int32_t mask_table[14] = {-1, -1, -1, -1, -1, -1, -1, 0, 0, 0, 0, 0, 0, 0};
31
32 const __m256 vmagic_bias = _mm256_set1_ps(0x1.8000FEp23f);
33 // The smallest x for which expf(x) is normalized.
34 const __m256 vdenorm_cutoff = _mm256_set1_ps(-0x1.5D589Ep6f);
35 const __m256 vlog2e = _mm256_set1_ps(0x1.715476p+0f);
36 const __m256 vminus_ln2_hi = _mm256_set1_ps(-0x1.62E43p-1f);
37 const __m256 vminus_ln2_lo = _mm256_set1_ps(0x1.05C61p-29f);
38
39 const __m256 vc1 = _mm256_set1_ps(0x1.FFFFF6p-1f);
40 const __m256 vc2 = _mm256_set1_ps(0x1.FFFDC6p-2f);
41 const __m256 vc3 = _mm256_set1_ps(0x1.555A80p-3f);
42 const __m256 vc4 = _mm256_set1_ps(0x1.573A1Ap-5f);
43 const __m256 vc5 = _mm256_set1_ps(0x1.0F9F9Cp-7f);
44
45 const __m256 vscale = _mm256_set1_ps(scale);
46 const __m256 vi_max = _mm256_set1_ps(max);
47
48 for (; batch >= 8 * sizeof(float); batch -= 8 * sizeof(float)) {
49 // Load 8 (1x8) inputs at a time.
50 const __m256 vi0 = _mm256_loadu_ps(input);
51 input += 8;
52
53 // Subtract maximum input x := i - i_max. This implies x <= 0.
54 const __m256 vx0 = _mm256_sub_ps(vi0, vi_max);
55
56 // Compute reduced argument batch := round(x / log(2)).
57 __m256 vn0 = _mm256_fmadd_ps(vx0, vlog2e, vmagic_bias);
58
59 // Create a floating-point number s (scale) such that s == 2**batch for inputs which don't cause underflow, i.e.
60 // -87.33642 <= x <= 0.0, and -126 <= batch <= 0 accordingly.
61 const __m256 vs0 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn0), 23));
62
63 // Subtract the large number back to get final batch := round(x / log(2)).
64 vn0 = _mm256_sub_ps(vn0, vmagic_bias);
65
66 // Compute reduced argument t := x - batch * log(2).
67 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
68 __m256 vt0 = _mm256_fmadd_ps(vn0, vminus_ln2_hi, vx0);
69
70 vt0 = _mm256_fmadd_ps(vn0, vminus_ln2_lo, vt0);
71
72 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
73 __m256 vp0 = _mm256_fmadd_ps(vc5, vt0, vc4);
74
75 vp0 = _mm256_fmadd_ps(vp0, vt0, vc3);
76
77 vp0 = _mm256_fmadd_ps(vp0, vt0, vc2);
78
79 vp0 = _mm256_fmadd_ps(vp0, vt0, vc1);
80
81 // Reconstruct the final f value:
82 // f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
83 // = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
84 // = s + (t * s) * p
85 vt0 = _mm256_mul_ps(vt0, vs0);
86
87 __m256 vf0 = _mm256_fmadd_ps(vt0, vp0, vs0);
88
89 // For inputs below zero cutoff, replace output with +0.0f.
90 // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
91 vf0 = _mm256_andnot_ps(_mm256_cmp_ps(vx0, vdenorm_cutoff, _CMP_LT_OS), vf0);
92
93 // Multiply by scale.
94 vf0 = _mm256_mul_ps(vf0, vscale);
95
96 // Store 8 (1x8) outputs at a time.
97 _mm256_storeu_ps(output, vf0);
98 output += 8;
99 }
100 for (; batch >= 8 * sizeof(float); batch -= 8 * sizeof(float)) {
101 // Load 8 inputs at a time.
102 const __m256 vi = _mm256_loadu_ps(input);
103 input += 8;
104
105 // Subtract maximum input x := i - i_max. This implies x <= 0.
106 const __m256 vx = _mm256_sub_ps(vi, vi_max);
107
108 // Compute reduced argument batch := round(x / log(2)).
109 __m256 vn = _mm256_fmadd_ps(vx, vlog2e, vmagic_bias);
110
111 // Create a floating-point number s (scale) such that s == 2**batch for inputs which don't cause underflow, i.e.
112 // -87.33642 <= x <= 0.0, and -126 <= batch <= 0 accordingly.
113 const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
114
115 // Subtract the large number back to get final batch := round(x / log(2)).
116 vn = _mm256_sub_ps(vn, vmagic_bias);
117
118 // Compute reduced argument t := x - batch * log(2).
119 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
120 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2_hi, vx);
121 vt = _mm256_fmadd_ps(vn, vminus_ln2_lo, vt);
122
123 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
124 __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
125 vp = _mm256_fmadd_ps(vp, vt, vc3);
126 vp = _mm256_fmadd_ps(vp, vt, vc2);
127 vp = _mm256_fmadd_ps(vp, vt, vc1);
128
129 // Reconstruct the final f value:
130 // f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
131 // = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
132 // = s + (t * s) * p
133 vt = _mm256_mul_ps(vt, vs);
134 __m256 vf = _mm256_fmadd_ps(vt, vp, vs);
135
136 // For inputs below zero cutoff, replace output with +0.0f.
137 // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
138 vf = _mm256_andnot_ps(_mm256_cmp_ps(vx, vdenorm_cutoff, _CMP_LT_OS), vf);
139
140 // Multiply by scale.
141 vf = _mm256_mul_ps(vf, vscale);
142
143 // Store 64 (8x8) outputs at a time.
144 _mm256_storeu_ps(output, vf);
145 output += 8;
146 }
147 if (batch != 0) {
148 assert(batch >= 1 * sizeof(float));
149 assert(batch <= 7 * sizeof(float));
150 const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - batch));
151
152 // Load up to 7 inputs at a time.
153 const __m256 vi = _mm256_maskload_ps(input, vmask);
154
155 // Subtract maximum input x := i - i_max. This implies x <= 0.
156 const __m256 vx = _mm256_sub_ps(vi, vi_max);
157
158 // Compute reduced argument batch := round(x / log(2)).
159 __m256 vn = _mm256_fmadd_ps(vx, vlog2e, vmagic_bias);
160
161 // Create a floating-point number s (scale) such that s == 2**batch for inputs which don't cause underflow, i.e.
162 // -87.33642 <= x <= 0.0, and -126 <= batch <= 0 accordingly.
163 const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
164
165 // Subtract the large number back to get final batch := round(x / log(2)).
166 vn = _mm256_sub_ps(vn, vmagic_bias);
167
168 // Compute reduced argument t := x - batch * log(2).
169 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
170 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2_hi, vx);
171 vt = _mm256_fmadd_ps(vn, vminus_ln2_lo, vt);
172
173 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
174 __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
175 vp = _mm256_fmadd_ps(vp, vt, vc3);
176 vp = _mm256_fmadd_ps(vp, vt, vc2);
177 vp = _mm256_fmadd_ps(vp, vt, vc1);
178
179 // Reconstruct the final f value:
180 // f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
181 // = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
182 // = s + (t * s) * p
183 vt = _mm256_mul_ps(vt, vs);
184 __m256 vf = _mm256_fmadd_ps(vt, vp, vs);
185
186 // For inputs below zero cutoff, replace output with +0.0f.
187 // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
188 vf = _mm256_andnot_ps(_mm256_cmp_ps(vx, vdenorm_cutoff, _CMP_LT_OS), vf);
189
190 // Multiply by scale.
191 vf = _mm256_mul_ps(vf, vscale);
192
193 // Store up to 7 outputs at a time.
194 _mm256_maskstore_ps(output, vmask, vf);
195 }
196 }