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 static const int32_t mask_table[14] = {-1, -1, -1, -1, -1, -1, -1, 0, 0, 0, 0, 0, 0, 0};
19
xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x24(size_t elements,const float * input,float * output,float scale,float max)20 void xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x24(
21 size_t elements,
22 const float* input,
23 float* output,
24 float scale,
25 float max)
26 {
27 assert(elements % sizeof(float) == 0);
28
29 const __m256 vmagic_bias = _mm256_set1_ps(0x1.8000FEp23f);
30 // The smallest x for which expf(x) is normalized.
31 const __m256 vdenorm_cutoff = _mm256_set1_ps(-0x1.5D589Ep6f);
32 const __m256 vlog2e = _mm256_set1_ps(0x1.715476p+0f);
33 const __m256 vminus_ln2_hi = _mm256_set1_ps(-0x1.62E43p-1f);
34 const __m256 vminus_ln2_lo = _mm256_set1_ps(0x1.05C61p-29f);
35
36 const __m256 vc1 = _mm256_set1_ps(0x1.FFFFF6p-1f);
37 const __m256 vc2 = _mm256_set1_ps(0x1.FFFDC6p-2f);
38 const __m256 vc3 = _mm256_set1_ps(0x1.555A80p-3f);
39 const __m256 vc4 = _mm256_set1_ps(0x1.573A1Ap-5f);
40 const __m256 vc5 = _mm256_set1_ps(0x1.0F9F9Cp-7f);
41
42 const __m256 vscale = _mm256_set1_ps(scale);
43 const __m256 vi_max = _mm256_set1_ps(max);
44
45 for (; elements >= 24 * sizeof(float); elements -= 24 * sizeof(float)) {
46 // Load 24 (3x8) inputs at a time.
47 const __m256 vi0 = _mm256_loadu_ps(input);
48 const __m256 vi1 = _mm256_loadu_ps(input + 8);
49 const __m256 vi2 = _mm256_loadu_ps(input + 16);
50 input += 24;
51
52 // Subtract maximum input x := i - i_max. This implies x <= 0.
53 const __m256 vx0 = _mm256_sub_ps(vi0, vi_max);
54 const __m256 vx1 = _mm256_sub_ps(vi1, vi_max);
55 const __m256 vx2 = _mm256_sub_ps(vi2, vi_max);
56
57 // Compute reduced argument elements := round(x / log(2)).
58 __m256 vn0 = _mm256_fmadd_ps(vx0, vlog2e, vmagic_bias);
59 __m256 vn1 = _mm256_fmadd_ps(vx1, vlog2e, vmagic_bias);
60 __m256 vn2 = _mm256_fmadd_ps(vx2, vlog2e, vmagic_bias);
61
62 // Create a floating-point number s (scale) such that s == 2**elements for inputs which don't cause underflow, i.e.
63 // -87.33642 <= x <= 0.0, and -126 <= elements <= 0 accordingly.
64 const __m256 vs0 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn0), 23));
65 const __m256 vs1 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn1), 23));
66 const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn2), 23));
67
68 // Subtract the large number back to get final elements := round(x / log(2)).
69 vn0 = _mm256_sub_ps(vn0, vmagic_bias);
70 vn1 = _mm256_sub_ps(vn1, vmagic_bias);
71 vn2 = _mm256_sub_ps(vn2, vmagic_bias);
72
73 // Compute reduced argument t := x - elements * log(2).
74 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
75 __m256 vt0 = _mm256_fmadd_ps(vn0, vminus_ln2_hi, vx0);
76 __m256 vt1 = _mm256_fmadd_ps(vn1, vminus_ln2_hi, vx1);
77 __m256 vt2 = _mm256_fmadd_ps(vn2, vminus_ln2_hi, vx2);
78
79 vt0 = _mm256_fmadd_ps(vn0, vminus_ln2_lo, vt0);
80 vt1 = _mm256_fmadd_ps(vn1, vminus_ln2_lo, vt1);
81 vt2 = _mm256_fmadd_ps(vn2, vminus_ln2_lo, vt2);
82
83 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
84 __m256 vp0 = _mm256_fmadd_ps(vc5, vt0, vc4);
85 __m256 vp1 = _mm256_fmadd_ps(vc5, vt1, vc4);
86 __m256 vp2 = _mm256_fmadd_ps(vc5, vt2, vc4);
87
88 vp0 = _mm256_fmadd_ps(vp0, vt0, vc3);
89 vp1 = _mm256_fmadd_ps(vp1, vt1, vc3);
90 vp2 = _mm256_fmadd_ps(vp2, vt2, vc3);
91
92 vp0 = _mm256_fmadd_ps(vp0, vt0, vc2);
93 vp1 = _mm256_fmadd_ps(vp1, vt1, vc2);
94 vp2 = _mm256_fmadd_ps(vp2, vt2, vc2);
95
96 vp0 = _mm256_fmadd_ps(vp0, vt0, vc1);
97 vp1 = _mm256_fmadd_ps(vp1, vt1, vc1);
98 vp2 = _mm256_fmadd_ps(vp2, vt2, vc1);
99
100 // Reconstruct the final f value:
101 // f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
102 // = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
103 // = s + (t * s) * p
104 vt0 = _mm256_mul_ps(vt0, vs0);
105 vt1 = _mm256_mul_ps(vt1, vs1);
106 vt2 = _mm256_mul_ps(vt2, vs2);
107
108 __m256 vf0 = _mm256_fmadd_ps(vt0, vp0, vs0);
109 __m256 vf1 = _mm256_fmadd_ps(vt1, vp1, vs1);
110 __m256 vf2 = _mm256_fmadd_ps(vt2, vp2, vs2);
111
112 // For inputs below zero cutoff, replace output with +0.0f.
113 // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
114 vf0 = _mm256_andnot_ps(_mm256_cmp_ps(vx0, vdenorm_cutoff, _CMP_LT_OS), vf0);
115 vf1 = _mm256_andnot_ps(_mm256_cmp_ps(vx1, vdenorm_cutoff, _CMP_LT_OS), vf1);
116 vf2 = _mm256_andnot_ps(_mm256_cmp_ps(vx2, vdenorm_cutoff, _CMP_LT_OS), vf2);
117
118 // Multiply by scale.
119 vf0 = _mm256_mul_ps(vf0, vscale);
120 vf1 = _mm256_mul_ps(vf1, vscale);
121 vf2 = _mm256_mul_ps(vf2, vscale);
122
123 // Store 24 (3x8) outputs at a time.
124 _mm256_storeu_ps(output, vf0);
125 _mm256_storeu_ps(output + 8, vf1);
126 _mm256_storeu_ps(output + 16, vf2);
127 output += 24;
128 }
129 for (; elements >= 8 * sizeof(float); elements -= 8 * sizeof(float)) {
130 // Load 8 inputs at a time.
131 const __m256 vi = _mm256_loadu_ps(input);
132 input += 8;
133
134 // Subtract maximum input x := i - i_max. This implies x <= 0.
135 const __m256 vx = _mm256_sub_ps(vi, vi_max);
136
137 // Compute reduced argument elements := round(x / log(2)).
138 __m256 vn = _mm256_fmadd_ps(vx, vlog2e, vmagic_bias);
139
140 // Create a floating-point number s (scale) such that s == 2**elements for inputs which don't cause underflow, i.e.
141 // -87.33642 <= x <= 0.0, and -126 <= elements <= 0 accordingly.
142 const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
143
144 // Subtract the large number back to get final elements := round(x / log(2)).
145 vn = _mm256_sub_ps(vn, vmagic_bias);
146
147 // Compute reduced argument t := x - elements * log(2).
148 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
149 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2_hi, vx);
150 vt = _mm256_fmadd_ps(vn, vminus_ln2_lo, vt);
151
152 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
153 __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
154 vp = _mm256_fmadd_ps(vp, vt, vc3);
155 vp = _mm256_fmadd_ps(vp, vt, vc2);
156 vp = _mm256_fmadd_ps(vp, vt, vc1);
157
158 // Reconstruct the final f value:
159 // f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
160 // = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
161 // = s + (t * s) * p
162 vt = _mm256_mul_ps(vt, vs);
163 __m256 vf = _mm256_fmadd_ps(vt, vp, vs);
164
165 // For inputs below zero cutoff, replace output with +0.0f.
166 // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
167 vf = _mm256_andnot_ps(_mm256_cmp_ps(vx, vdenorm_cutoff, _CMP_LT_OS), vf);
168
169 // Multiply by scale.
170 vf = _mm256_mul_ps(vf, vscale);
171
172 // Store 64 (8x8) outputs at a time.
173 _mm256_storeu_ps(output, vf);
174 output += 8;
175 }
176 if (elements != 0) {
177 assert(elements >= 1 * sizeof(float));
178 assert(elements <= 7 * sizeof(float));
179 const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - elements));
180
181 // Load up to 7 inputs at a time.
182 const __m256 vi = _mm256_maskload_ps(input, vmask);
183
184 // Subtract maximum input x := i - i_max. This implies x <= 0.
185 const __m256 vx = _mm256_sub_ps(vi, vi_max);
186
187 // Compute reduced argument elements := round(x / log(2)).
188 __m256 vn = _mm256_fmadd_ps(vx, vlog2e, vmagic_bias);
189
190 // Create a floating-point number s (scale) such that s == 2**elements for inputs which don't cause underflow, i.e.
191 // -87.33642 <= x <= 0.0, and -126 <= elements <= 0 accordingly.
192 const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
193
194 // Subtract the large number back to get final elements := round(x / log(2)).
195 vn = _mm256_sub_ps(vn, vmagic_bias);
196
197 // Compute reduced argument t := x - elements * log(2).
198 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
199 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2_hi, vx);
200 vt = _mm256_fmadd_ps(vn, vminus_ln2_lo, vt);
201
202 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
203 __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
204 vp = _mm256_fmadd_ps(vp, vt, vc3);
205 vp = _mm256_fmadd_ps(vp, vt, vc2);
206 vp = _mm256_fmadd_ps(vp, vt, vc1);
207
208 // Reconstruct the final f value:
209 // f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
210 // = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
211 // = s + (t * s) * p
212 vt = _mm256_mul_ps(vt, vs);
213 __m256 vf = _mm256_fmadd_ps(vt, vp, vs);
214
215 // For inputs below zero cutoff, replace output with +0.0f.
216 // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
217 vf = _mm256_andnot_ps(_mm256_cmp_ps(vx, vdenorm_cutoff, _CMP_LT_OS), vf);
218
219 // Multiply by scale.
220 vf = _mm256_mul_ps(vf, vscale);
221
222 // Store up to 7 outputs at a time.
223 _mm256_maskstore_ps(output, vmask, vf);
224 }
225 }
226