Debian Codegraph @

src/f32-vscaleexpminusmax/gen/f32-vscaleexpminusmax-avx2-p5-u48.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_u48(
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 >= 48 * sizeof(float); batch -= 48 * sizeof(float)) {
49 // Load 48 (6x8) 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 const __m256 vi4 = _mm256_loadu_ps(input + 32);
55 const __m256 vi5 = _mm256_loadu_ps(input + 40);
56 input += 48;
57
58 // Subtract maximum input x := i - i_max. This implies x <= 0.
59 const __m256 vx0 = _mm256_sub_ps(vi0, vi_max);
60 const __m256 vx1 = _mm256_sub_ps(vi1, vi_max);
61 const __m256 vx2 = _mm256_sub_ps(vi2, vi_max);
62 const __m256 vx3 = _mm256_sub_ps(vi3, vi_max);
63 const __m256 vx4 = _mm256_sub_ps(vi4, vi_max);
64 const __m256 vx5 = _mm256_sub_ps(vi5, vi_max);
65
66 // Compute reduced argument batch := round(x / log(2)).
67 __m256 vn0 = _mm256_fmadd_ps(vx0, vlog2e, vmagic_bias);
68 __m256 vn1 = _mm256_fmadd_ps(vx1, vlog2e, vmagic_bias);
69 __m256 vn2 = _mm256_fmadd_ps(vx2, vlog2e, vmagic_bias);
70 __m256 vn3 = _mm256_fmadd_ps(vx3, vlog2e, vmagic_bias);
71 __m256 vn4 = _mm256_fmadd_ps(vx4, vlog2e, vmagic_bias);
72 __m256 vn5 = _mm256_fmadd_ps(vx5, vlog2e, vmagic_bias);
73
74 // Create a floating-point number s (scale) such that s == 2**batch for inputs which don't cause underflow, i.e.
75 // -87.33642 <= x <= 0.0, and -126 <= batch <= 0 accordingly.
76 const __m256 vs0 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn0), 23));
77 const __m256 vs1 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn1), 23));
78 const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn2), 23));
79 const __m256 vs3 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn3), 23));
80 const __m256 vs4 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn4), 23));
81 const __m256 vs5 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn5), 23));
82
83 // Subtract the large number back to get final batch := round(x / log(2)).
84 vn0 = _mm256_sub_ps(vn0, vmagic_bias);
85 vn1 = _mm256_sub_ps(vn1, vmagic_bias);
86 vn2 = _mm256_sub_ps(vn2, vmagic_bias);
87 vn3 = _mm256_sub_ps(vn3, vmagic_bias);
88 vn4 = _mm256_sub_ps(vn4, vmagic_bias);
89 vn5 = _mm256_sub_ps(vn5, vmagic_bias);
90
91 // Compute reduced argument t := x - batch * log(2).
92 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
93 __m256 vt0 = _mm256_fmadd_ps(vn0, vminus_ln2_hi, vx0);
94 __m256 vt1 = _mm256_fmadd_ps(vn1, vminus_ln2_hi, vx1);
95 __m256 vt2 = _mm256_fmadd_ps(vn2, vminus_ln2_hi, vx2);
96 __m256 vt3 = _mm256_fmadd_ps(vn3, vminus_ln2_hi, vx3);
97 __m256 vt4 = _mm256_fmadd_ps(vn4, vminus_ln2_hi, vx4);
98 __m256 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_hi, vx5);
99
100 vt0 = _mm256_fmadd_ps(vn0, vminus_ln2_lo, vt0);
101 vt1 = _mm256_fmadd_ps(vn1, vminus_ln2_lo, vt1);
102 vt2 = _mm256_fmadd_ps(vn2, vminus_ln2_lo, vt2);
103 vt3 = _mm256_fmadd_ps(vn3, vminus_ln2_lo, vt3);
104 vt4 = _mm256_fmadd_ps(vn4, vminus_ln2_lo, vt4);
105 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_lo, vt5);
106
107 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
108 __m256 vp0 = _mm256_fmadd_ps(vc5, vt0, vc4);
109 __m256 vp1 = _mm256_fmadd_ps(vc5, vt1, vc4);
110 __m256 vp2 = _mm256_fmadd_ps(vc5, vt2, vc4);
111 __m256 vp3 = _mm256_fmadd_ps(vc5, vt3, vc4);
112 __m256 vp4 = _mm256_fmadd_ps(vc5, vt4, vc4);
113 __m256 vp5 = _mm256_fmadd_ps(vc5, vt5, vc4);
114
115 vp0 = _mm256_fmadd_ps(vp0, vt0, vc3);
116 vp1 = _mm256_fmadd_ps(vp1, vt1, vc3);
117 vp2 = _mm256_fmadd_ps(vp2, vt2, vc3);
118 vp3 = _mm256_fmadd_ps(vp3, vt3, vc3);
119 vp4 = _mm256_fmadd_ps(vp4, vt4, vc3);
120 vp5 = _mm256_fmadd_ps(vp5, vt5, vc3);
121
122 vp0 = _mm256_fmadd_ps(vp0, vt0, vc2);
123 vp1 = _mm256_fmadd_ps(vp1, vt1, vc2);
124 vp2 = _mm256_fmadd_ps(vp2, vt2, vc2);
125 vp3 = _mm256_fmadd_ps(vp3, vt3, vc2);
126 vp4 = _mm256_fmadd_ps(vp4, vt4, vc2);
127 vp5 = _mm256_fmadd_ps(vp5, vt5, vc2);
128
129 vp0 = _mm256_fmadd_ps(vp0, vt0, vc1);
130 vp1 = _mm256_fmadd_ps(vp1, vt1, vc1);
131 vp2 = _mm256_fmadd_ps(vp2, vt2, vc1);
132 vp3 = _mm256_fmadd_ps(vp3, vt3, vc1);
133 vp4 = _mm256_fmadd_ps(vp4, vt4, vc1);
134 vp5 = _mm256_fmadd_ps(vp5, vt5, vc1);
135
136 // Reconstruct the final f value:
137 // f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
138 // = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
139 // = s + (t * s) * p
140 vt0 = _mm256_mul_ps(vt0, vs0);
141 vt1 = _mm256_mul_ps(vt1, vs1);
142 vt2 = _mm256_mul_ps(vt2, vs2);
143 vt3 = _mm256_mul_ps(vt3, vs3);
144 vt4 = _mm256_mul_ps(vt4, vs4);
145 vt5 = _mm256_mul_ps(vt5, vs5);
146
147 __m256 vf0 = _mm256_fmadd_ps(vt0, vp0, vs0);
148 __m256 vf1 = _mm256_fmadd_ps(vt1, vp1, vs1);
149 __m256 vf2 = _mm256_fmadd_ps(vt2, vp2, vs2);
150 __m256 vf3 = _mm256_fmadd_ps(vt3, vp3, vs3);
151 __m256 vf4 = _mm256_fmadd_ps(vt4, vp4, vs4);
152 __m256 vf5 = _mm256_fmadd_ps(vt5, vp5, vs5);
153
154 // For inputs below zero cutoff, replace output with +0.0f.
155 // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
156 vf0 = _mm256_andnot_ps(_mm256_cmp_ps(vx0, vdenorm_cutoff, _CMP_LT_OS), vf0);
157 vf1 = _mm256_andnot_ps(_mm256_cmp_ps(vx1, vdenorm_cutoff, _CMP_LT_OS), vf1);
158 vf2 = _mm256_andnot_ps(_mm256_cmp_ps(vx2, vdenorm_cutoff, _CMP_LT_OS), vf2);
159 vf3 = _mm256_andnot_ps(_mm256_cmp_ps(vx3, vdenorm_cutoff, _CMP_LT_OS), vf3);
160 vf4 = _mm256_andnot_ps(_mm256_cmp_ps(vx4, vdenorm_cutoff, _CMP_LT_OS), vf4);
161 vf5 = _mm256_andnot_ps(_mm256_cmp_ps(vx5, vdenorm_cutoff, _CMP_LT_OS), vf5);
162
163 // Multiply by scale.
164 vf0 = _mm256_mul_ps(vf0, vscale);
165 vf1 = _mm256_mul_ps(vf1, vscale);
166 vf2 = _mm256_mul_ps(vf2, vscale);
167 vf3 = _mm256_mul_ps(vf3, vscale);
168 vf4 = _mm256_mul_ps(vf4, vscale);
169 vf5 = _mm256_mul_ps(vf5, vscale);
170
171 // Store 48 (6x8) outputs at a time.
172 _mm256_storeu_ps(output, vf0);
173 _mm256_storeu_ps(output + 8, vf1);
174 _mm256_storeu_ps(output + 16, vf2);
175 _mm256_storeu_ps(output + 24, vf3);
176 _mm256_storeu_ps(output + 32, vf4);
177 _mm256_storeu_ps(output + 40, vf5);
178 output += 48;
179 }
180 for (; batch >= 8 * sizeof(float); batch -= 8 * sizeof(float)) {
181 // Load 8 inputs at a time.
182 const __m256 vi = _mm256_loadu_ps(input);
183 input += 8;
184
185 // Subtract maximum input x := i - i_max. This implies x <= 0.
186 const __m256 vx = _mm256_sub_ps(vi, vi_max);
187
188 // Compute reduced argument batch := round(x / log(2)).
189 __m256 vn = _mm256_fmadd_ps(vx, vlog2e, vmagic_bias);
190
191 // Create a floating-point number s (scale) such that s == 2**batch for inputs which don't cause underflow, i.e.
192 // -87.33642 <= x <= 0.0, and -126 <= batch <= 0 accordingly.
193 const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
194
195 // Subtract the large number back to get final batch := round(x / log(2)).
196 vn = _mm256_sub_ps(vn, vmagic_bias);
197
198 // Compute reduced argument t := x - batch * log(2).
199 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
200 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2_hi, vx);
201 vt = _mm256_fmadd_ps(vn, vminus_ln2_lo, vt);
202
203 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
204 __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
205 vp = _mm256_fmadd_ps(vp, vt, vc3);
206 vp = _mm256_fmadd_ps(vp, vt, vc2);
207 vp = _mm256_fmadd_ps(vp, vt, vc1);
208
209 // Reconstruct the final f value:
210 // f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
211 // = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
212 // = s + (t * s) * p
213 vt = _mm256_mul_ps(vt, vs);
214 __m256 vf = _mm256_fmadd_ps(vt, vp, vs);
215
216 // For inputs below zero cutoff, replace output with +0.0f.
217 // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
218 vf = _mm256_andnot_ps(_mm256_cmp_ps(vx, vdenorm_cutoff, _CMP_LT_OS), vf);
219
220 // Multiply by scale.
221 vf = _mm256_mul_ps(vf, vscale);
222
223 // Store 64 (8x8) outputs at a time.
224 _mm256_storeu_ps(output, vf);
225 output += 8;
226 }
227 if (batch != 0) {
228 assert(batch >= 1 * sizeof(float));
229 assert(batch <= 7 * sizeof(float));
230 const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - batch));
231
232 // Load up to 7 inputs at a time.
233 const __m256 vi = _mm256_maskload_ps(input, vmask);
234
235 // Subtract maximum input x := i - i_max. This implies x <= 0.
236 const __m256 vx = _mm256_sub_ps(vi, vi_max);
237
238 // Compute reduced argument batch := round(x / log(2)).
239 __m256 vn = _mm256_fmadd_ps(vx, vlog2e, vmagic_bias);
240
241 // Create a floating-point number s (scale) such that s == 2**batch for inputs which don't cause underflow, i.e.
242 // -87.33642 <= x <= 0.0, and -126 <= batch <= 0 accordingly.
243 const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
244
245 // Subtract the large number back to get final batch := round(x / log(2)).
246 vn = _mm256_sub_ps(vn, vmagic_bias);
247
248 // Compute reduced argument t := x - batch * log(2).
249 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
250 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2_hi, vx);
251 vt = _mm256_fmadd_ps(vn, vminus_ln2_lo, vt);
252
253 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
254 __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
255 vp = _mm256_fmadd_ps(vp, vt, vc3);
256 vp = _mm256_fmadd_ps(vp, vt, vc2);
257 vp = _mm256_fmadd_ps(vp, vt, vc1);
258
259 // Reconstruct the final f value:
260 // f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
261 // = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
262 // = s + (t * s) * p
263 vt = _mm256_mul_ps(vt, vs);
264 __m256 vf = _mm256_fmadd_ps(vt, vp, vs);
265
266 // For inputs below zero cutoff, replace output with +0.0f.
267 // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
268 vf = _mm256_andnot_ps(_mm256_cmp_ps(vx, vdenorm_cutoff, _CMP_LT_OS), vf);
269
270 // Multiply by scale.
271 vf = _mm256_mul_ps(vf, vscale);
272
273 // Store up to 7 outputs at a time.
274 _mm256_maskstore_ps(output, vmask, vf);
275 }
276 }