Debian Codegraph @

src/f32-vscaleexpminusmax/gen/f32-vscaleexpminusmax-avx2-p5-u32.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_u32(
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 >= 32 * sizeof(float); batch -= 32 * sizeof(float)) {
49 // Load 32 (4x8) 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 const __m256 vi3 = _mm256_loadu_ps(input + 24);
54 input += 32;
55
56 // Subtract maximum input x := i - i_max. This implies x <= 0.
57 const __m256 vx0 = _mm256_sub_ps(vi0, vi_max);
58 const __m256 vx1 = _mm256_sub_ps(vi1, vi_max);
59 const __m256 vx2 = _mm256_sub_ps(vi2, vi_max);
60 const __m256 vx3 = _mm256_sub_ps(vi3, vi_max);
61
62 // Compute reduced argument batch := round(x / log(2)).
63 __m256 vn0 = _mm256_fmadd_ps(vx0, vlog2e, vmagic_bias);
64 __m256 vn1 = _mm256_fmadd_ps(vx1, vlog2e, vmagic_bias);
65 __m256 vn2 = _mm256_fmadd_ps(vx2, vlog2e, vmagic_bias);
66 __m256 vn3 = _mm256_fmadd_ps(vx3, vlog2e, vmagic_bias);
67
68 // Create a floating-point number s (scale) such that s == 2**batch for inputs which don't cause underflow, i.e.
69 // -87.33642 <= x <= 0.0, and -126 <= batch <= 0 accordingly.
70 const __m256 vs0 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn0), 23));
71 const __m256 vs1 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn1), 23));
72 const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn2), 23));
73 const __m256 vs3 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn3), 23));
74
75 // Subtract the large number back to get final batch := round(x / log(2)).
76 vn0 = _mm256_sub_ps(vn0, vmagic_bias);
77 vn1 = _mm256_sub_ps(vn1, vmagic_bias);
78 vn2 = _mm256_sub_ps(vn2, vmagic_bias);
79 vn3 = _mm256_sub_ps(vn3, vmagic_bias);
80
81 // Compute reduced argument t := x - batch * log(2).
82 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
83 __m256 vt0 = _mm256_fmadd_ps(vn0, vminus_ln2_hi, vx0);
84 __m256 vt1 = _mm256_fmadd_ps(vn1, vminus_ln2_hi, vx1);
85 __m256 vt2 = _mm256_fmadd_ps(vn2, vminus_ln2_hi, vx2);
86 __m256 vt3 = _mm256_fmadd_ps(vn3, vminus_ln2_hi, vx3);
87
88 vt0 = _mm256_fmadd_ps(vn0, vminus_ln2_lo, vt0);
89 vt1 = _mm256_fmadd_ps(vn1, vminus_ln2_lo, vt1);
90 vt2 = _mm256_fmadd_ps(vn2, vminus_ln2_lo, vt2);
91 vt3 = _mm256_fmadd_ps(vn3, vminus_ln2_lo, vt3);
92
93 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
94 __m256 vp0 = _mm256_fmadd_ps(vc5, vt0, vc4);
95 __m256 vp1 = _mm256_fmadd_ps(vc5, vt1, vc4);
96 __m256 vp2 = _mm256_fmadd_ps(vc5, vt2, vc4);
97 __m256 vp3 = _mm256_fmadd_ps(vc5, vt3, vc4);
98
99 vp0 = _mm256_fmadd_ps(vp0, vt0, vc3);
100 vp1 = _mm256_fmadd_ps(vp1, vt1, vc3);
101 vp2 = _mm256_fmadd_ps(vp2, vt2, vc3);
102 vp3 = _mm256_fmadd_ps(vp3, vt3, vc3);
103
104 vp0 = _mm256_fmadd_ps(vp0, vt0, vc2);
105 vp1 = _mm256_fmadd_ps(vp1, vt1, vc2);
106 vp2 = _mm256_fmadd_ps(vp2, vt2, vc2);
107 vp3 = _mm256_fmadd_ps(vp3, vt3, vc2);
108
109 vp0 = _mm256_fmadd_ps(vp0, vt0, vc1);
110 vp1 = _mm256_fmadd_ps(vp1, vt1, vc1);
111 vp2 = _mm256_fmadd_ps(vp2, vt2, vc1);
112 vp3 = _mm256_fmadd_ps(vp3, vt3, vc1);
113
114 // Reconstruct the final f value:
115 // f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
116 // = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
117 // = s + (t * s) * p
118 vt0 = _mm256_mul_ps(vt0, vs0);
119 vt1 = _mm256_mul_ps(vt1, vs1);
120 vt2 = _mm256_mul_ps(vt2, vs2);
121 vt3 = _mm256_mul_ps(vt3, vs3);
122
123 __m256 vf0 = _mm256_fmadd_ps(vt0, vp0, vs0);
124 __m256 vf1 = _mm256_fmadd_ps(vt1, vp1, vs1);
125 __m256 vf2 = _mm256_fmadd_ps(vt2, vp2, vs2);
126 __m256 vf3 = _mm256_fmadd_ps(vt3, vp3, vs3);
127
128 // For inputs below zero cutoff, replace output with +0.0f.
129 // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
130 vf0 = _mm256_andnot_ps(_mm256_cmp_ps(vx0, vdenorm_cutoff, _CMP_LT_OS), vf0);
131 vf1 = _mm256_andnot_ps(_mm256_cmp_ps(vx1, vdenorm_cutoff, _CMP_LT_OS), vf1);
132 vf2 = _mm256_andnot_ps(_mm256_cmp_ps(vx2, vdenorm_cutoff, _CMP_LT_OS), vf2);
133 vf3 = _mm256_andnot_ps(_mm256_cmp_ps(vx3, vdenorm_cutoff, _CMP_LT_OS), vf3);
134
135 // Multiply by scale.
136 vf0 = _mm256_mul_ps(vf0, vscale);
137 vf1 = _mm256_mul_ps(vf1, vscale);
138 vf2 = _mm256_mul_ps(vf2, vscale);
139 vf3 = _mm256_mul_ps(vf3, vscale);
140
141 // Store 32 (4x8) outputs at a time.
142 _mm256_storeu_ps(output, vf0);
143 _mm256_storeu_ps(output + 8, vf1);
144 _mm256_storeu_ps(output + 16, vf2);
145 _mm256_storeu_ps(output + 24, vf3);
146 output += 32;
147 }
148 for (; batch >= 8 * sizeof(float); batch -= 8 * sizeof(float)) {
149 // Load 8 inputs at a time.
150 const __m256 vi = _mm256_loadu_ps(input);
151 input += 8;
152
153 // Subtract maximum input x := i - i_max. This implies x <= 0.
154 const __m256 vx = _mm256_sub_ps(vi, vi_max);
155
156 // Compute reduced argument batch := round(x / log(2)).
157 __m256 vn = _mm256_fmadd_ps(vx, vlog2e, vmagic_bias);
158
159 // Create a floating-point number s (scale) such that s == 2**batch for inputs which don't cause underflow, i.e.
160 // -87.33642 <= x <= 0.0, and -126 <= batch <= 0 accordingly.
161 const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
162
163 // Subtract the large number back to get final batch := round(x / log(2)).
164 vn = _mm256_sub_ps(vn, vmagic_bias);
165
166 // Compute reduced argument t := x - batch * log(2).
167 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
168 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2_hi, vx);
169 vt = _mm256_fmadd_ps(vn, vminus_ln2_lo, vt);
170
171 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
172 __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
173 vp = _mm256_fmadd_ps(vp, vt, vc3);
174 vp = _mm256_fmadd_ps(vp, vt, vc2);
175 vp = _mm256_fmadd_ps(vp, vt, vc1);
176
177 // Reconstruct the final f value:
178 // f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
179 // = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
180 // = s + (t * s) * p
181 vt = _mm256_mul_ps(vt, vs);
182 __m256 vf = _mm256_fmadd_ps(vt, vp, vs);
183
184 // For inputs below zero cutoff, replace output with +0.0f.
185 // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
186 vf = _mm256_andnot_ps(_mm256_cmp_ps(vx, vdenorm_cutoff, _CMP_LT_OS), vf);
187
188 // Multiply by scale.
189 vf = _mm256_mul_ps(vf, vscale);
190
191 // Store 64 (8x8) outputs at a time.
192 _mm256_storeu_ps(output, vf);
193 output += 8;
194 }
195 if (batch != 0) {
196 assert(batch >= 1 * sizeof(float));
197 assert(batch <= 7 * sizeof(float));
198 const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - batch));
199
200 // Load up to 7 inputs at a time.
201 const __m256 vi = _mm256_maskload_ps(input, vmask);
202
203 // Subtract maximum input x := i - i_max. This implies x <= 0.
204 const __m256 vx = _mm256_sub_ps(vi, vi_max);
205
206 // Compute reduced argument batch := round(x / log(2)).
207 __m256 vn = _mm256_fmadd_ps(vx, vlog2e, vmagic_bias);
208
209 // Create a floating-point number s (scale) such that s == 2**batch for inputs which don't cause underflow, i.e.
210 // -87.33642 <= x <= 0.0, and -126 <= batch <= 0 accordingly.
211 const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
212
213 // Subtract the large number back to get final batch := round(x / log(2)).
214 vn = _mm256_sub_ps(vn, vmagic_bias);
215
216 // Compute reduced argument t := x - batch * log(2).
217 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
218 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2_hi, vx);
219 vt = _mm256_fmadd_ps(vn, vminus_ln2_lo, vt);
220
221 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
222 __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
223 vp = _mm256_fmadd_ps(vp, vt, vc3);
224 vp = _mm256_fmadd_ps(vp, vt, vc2);
225 vp = _mm256_fmadd_ps(vp, vt, vc1);
226
227 // Reconstruct the final f value:
228 // f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
229 // = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
230 // = s + (t * s) * p
231 vt = _mm256_mul_ps(vt, vs);
232 __m256 vf = _mm256_fmadd_ps(vt, vp, vs);
233
234 // For inputs below zero cutoff, replace output with +0.0f.
235 // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
236 vf = _mm256_andnot_ps(_mm256_cmp_ps(vx, vdenorm_cutoff, _CMP_LT_OS), vf);
237
238 // Multiply by scale.
239 vf = _mm256_mul_ps(vf, vscale);
240
241 // Store up to 7 outputs at a time.
242 _mm256_maskstore_ps(output, vmask, vf);
243 }
244 }