Debian Codegraph @

src/f32-vscaleexpminusmax/gen/f32-vscaleexpminusmax-avx2-p5-u24.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_u24(
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 >= 24 * sizeof(float); batch -= 24 * sizeof(float)) {
49 // Load 24 (3x8) inputs at a time.
50 const __m256 vi0 = _mm256_loadu_ps(input);
51 const __m256 vi1 = _mm256_loadu_ps(input + 8);
52 const __m256 vi2 = _mm256_loadu_ps(input + 16);
53 input += 24;
54
55 // Subtract maximum input x := i - i_max. This implies x <= 0.
56 const __m256 vx0 = _mm256_sub_ps(vi0, vi_max);
57 const __m256 vx1 = _mm256_sub_ps(vi1, vi_max);
58 const __m256 vx2 = _mm256_sub_ps(vi2, vi_max);
59
60 // Compute reduced argument batch := round(x / log(2)).
61 __m256 vn0 = _mm256_fmadd_ps(vx0, vlog2e, vmagic_bias);
62 __m256 vn1 = _mm256_fmadd_ps(vx1, vlog2e, vmagic_bias);
63 __m256 vn2 = _mm256_fmadd_ps(vx2, vlog2e, vmagic_bias);
64
65 // Create a floating-point number s (scale) such that s == 2**batch for inputs which don't cause underflow, i.e.
66 // -87.33642 <= x <= 0.0, and -126 <= batch <= 0 accordingly.
67 const __m256 vs0 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn0), 23));
68 const __m256 vs1 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn1), 23));
69 const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn2), 23));
70
71 // Subtract the large number back to get final batch := round(x / log(2)).
72 vn0 = _mm256_sub_ps(vn0, vmagic_bias);
73 vn1 = _mm256_sub_ps(vn1, vmagic_bias);
74 vn2 = _mm256_sub_ps(vn2, vmagic_bias);
75
76 // Compute reduced argument t := x - batch * log(2).
77 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
78 __m256 vt0 = _mm256_fmadd_ps(vn0, vminus_ln2_hi, vx0);
79 __m256 vt1 = _mm256_fmadd_ps(vn1, vminus_ln2_hi, vx1);
80 __m256 vt2 = _mm256_fmadd_ps(vn2, vminus_ln2_hi, vx2);
81
82 vt0 = _mm256_fmadd_ps(vn0, vminus_ln2_lo, vt0);
83 vt1 = _mm256_fmadd_ps(vn1, vminus_ln2_lo, vt1);
84 vt2 = _mm256_fmadd_ps(vn2, vminus_ln2_lo, vt2);
85
86 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
87 __m256 vp0 = _mm256_fmadd_ps(vc5, vt0, vc4);
88 __m256 vp1 = _mm256_fmadd_ps(vc5, vt1, vc4);
89 __m256 vp2 = _mm256_fmadd_ps(vc5, vt2, vc4);
90
91 vp0 = _mm256_fmadd_ps(vp0, vt0, vc3);
92 vp1 = _mm256_fmadd_ps(vp1, vt1, vc3);
93 vp2 = _mm256_fmadd_ps(vp2, vt2, vc3);
94
95 vp0 = _mm256_fmadd_ps(vp0, vt0, vc2);
96 vp1 = _mm256_fmadd_ps(vp1, vt1, vc2);
97 vp2 = _mm256_fmadd_ps(vp2, vt2, vc2);
98
99 vp0 = _mm256_fmadd_ps(vp0, vt0, vc1);
100 vp1 = _mm256_fmadd_ps(vp1, vt1, vc1);
101 vp2 = _mm256_fmadd_ps(vp2, vt2, vc1);
102
103 // Reconstruct the final f value:
104 // f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
105 // = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
106 // = s + (t * s) * p
107 vt0 = _mm256_mul_ps(vt0, vs0);
108 vt1 = _mm256_mul_ps(vt1, vs1);
109 vt2 = _mm256_mul_ps(vt2, vs2);
110
111 __m256 vf0 = _mm256_fmadd_ps(vt0, vp0, vs0);
112 __m256 vf1 = _mm256_fmadd_ps(vt1, vp1, vs1);
113 __m256 vf2 = _mm256_fmadd_ps(vt2, vp2, vs2);
114
115 // For inputs below zero cutoff, replace output with +0.0f.
116 // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
117 vf0 = _mm256_andnot_ps(_mm256_cmp_ps(vx0, vdenorm_cutoff, _CMP_LT_OS), vf0);
118 vf1 = _mm256_andnot_ps(_mm256_cmp_ps(vx1, vdenorm_cutoff, _CMP_LT_OS), vf1);
119 vf2 = _mm256_andnot_ps(_mm256_cmp_ps(vx2, vdenorm_cutoff, _CMP_LT_OS), vf2);
120
121 // Multiply by scale.
122 vf0 = _mm256_mul_ps(vf0, vscale);
123 vf1 = _mm256_mul_ps(vf1, vscale);
124 vf2 = _mm256_mul_ps(vf2, vscale);
125
126 // Store 24 (3x8) outputs at a time.
127 _mm256_storeu_ps(output, vf0);
128 _mm256_storeu_ps(output + 8, vf1);
129 _mm256_storeu_ps(output + 16, vf2);
130 output += 24;
131 }
132 for (; batch >= 8 * sizeof(float); batch -= 8 * sizeof(float)) {
133 // Load 8 inputs at a time.
134 const __m256 vi = _mm256_loadu_ps(input);
135 input += 8;
136
137 // Subtract maximum input x := i - i_max. This implies x <= 0.
138 const __m256 vx = _mm256_sub_ps(vi, vi_max);
139
140 // Compute reduced argument batch := round(x / log(2)).
141 __m256 vn = _mm256_fmadd_ps(vx, vlog2e, vmagic_bias);
142
143 // Create a floating-point number s (scale) such that s == 2**batch for inputs which don't cause underflow, i.e.
144 // -87.33642 <= x <= 0.0, and -126 <= batch <= 0 accordingly.
145 const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
146
147 // Subtract the large number back to get final batch := round(x / log(2)).
148 vn = _mm256_sub_ps(vn, vmagic_bias);
149
150 // Compute reduced argument t := x - batch * log(2).
151 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
152 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2_hi, vx);
153 vt = _mm256_fmadd_ps(vn, vminus_ln2_lo, vt);
154
155 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
156 __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
157 vp = _mm256_fmadd_ps(vp, vt, vc3);
158 vp = _mm256_fmadd_ps(vp, vt, vc2);
159 vp = _mm256_fmadd_ps(vp, vt, vc1);
160
161 // Reconstruct the final f value:
162 // f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
163 // = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
164 // = s + (t * s) * p
165 vt = _mm256_mul_ps(vt, vs);
166 __m256 vf = _mm256_fmadd_ps(vt, vp, vs);
167
168 // For inputs below zero cutoff, replace output with +0.0f.
169 // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
170 vf = _mm256_andnot_ps(_mm256_cmp_ps(vx, vdenorm_cutoff, _CMP_LT_OS), vf);
171
172 // Multiply by scale.
173 vf = _mm256_mul_ps(vf, vscale);
174
175 // Store 64 (8x8) outputs at a time.
176 _mm256_storeu_ps(output, vf);
177 output += 8;
178 }
179 if (batch != 0) {
180 assert(batch >= 1 * sizeof(float));
181 assert(batch <= 7 * sizeof(float));
182 const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - batch));
183
184 // Load up to 7 inputs at a time.
185 const __m256 vi = _mm256_maskload_ps(input, vmask);
186
187 // Subtract maximum input x := i - i_max. This implies x <= 0.
188 const __m256 vx = _mm256_sub_ps(vi, vi_max);
189
190 // Compute reduced argument batch := round(x / log(2)).
191 __m256 vn = _mm256_fmadd_ps(vx, vlog2e, vmagic_bias);
192
193 // Create a floating-point number s (scale) such that s == 2**batch for inputs which don't cause underflow, i.e.
194 // -87.33642 <= x <= 0.0, and -126 <= batch <= 0 accordingly.
195 const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
196
197 // Subtract the large number back to get final batch := round(x / log(2)).
198 vn = _mm256_sub_ps(vn, vmagic_bias);
199
200 // Compute reduced argument t := x - batch * log(2).
201 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
202 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2_hi, vx);
203 vt = _mm256_fmadd_ps(vn, vminus_ln2_lo, vt);
204
205 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
206 __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
207 vp = _mm256_fmadd_ps(vp, vt, vc3);
208 vp = _mm256_fmadd_ps(vp, vt, vc2);
209 vp = _mm256_fmadd_ps(vp, vt, vc1);
210
211 // Reconstruct the final f value:
212 // f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
213 // = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
214 // = s + (t * s) * p
215 vt = _mm256_mul_ps(vt, vs);
216 __m256 vf = _mm256_fmadd_ps(vt, vp, vs);
217
218 // For inputs below zero cutoff, replace output with +0.0f.
219 // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
220 vf = _mm256_andnot_ps(_mm256_cmp_ps(vx, vdenorm_cutoff, _CMP_LT_OS), vf);
221
222 // Multiply by scale.
223 vf = _mm256_mul_ps(vf, vscale);
224
225 // Store up to 7 outputs at a time.
226 _mm256_maskstore_ps(output, vmask, vf);
227 }
228 }