Debian Codegraph @

src/f32-vscaleexpminusmax/gen/f32-vscaleexpminusmax-avx2-p5-u72.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_u72(
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 >= 72 * sizeof(float); batch -= 72 * sizeof(float)) {
49 // Load 72 (9x8) 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 const __m256 vi6 = _mm256_loadu_ps(input + 48);
57 const __m256 vi7 = _mm256_loadu_ps(input + 56);
58 const __m256 vi8 = _mm256_loadu_ps(input + 64);
59 input += 72;
60
61 // Subtract maximum input x := i - i_max. This implies x <= 0.
62 const __m256 vx0 = _mm256_sub_ps(vi0, vi_max);
63 const __m256 vx1 = _mm256_sub_ps(vi1, vi_max);
64 const __m256 vx2 = _mm256_sub_ps(vi2, vi_max);
65 const __m256 vx3 = _mm256_sub_ps(vi3, vi_max);
66 const __m256 vx4 = _mm256_sub_ps(vi4, vi_max);
67 const __m256 vx5 = _mm256_sub_ps(vi5, vi_max);
68 const __m256 vx6 = _mm256_sub_ps(vi6, vi_max);
69 const __m256 vx7 = _mm256_sub_ps(vi7, vi_max);
70 const __m256 vx8 = _mm256_sub_ps(vi8, vi_max);
71
72 // Compute reduced argument batch := round(x / log(2)).
73 __m256 vn0 = _mm256_fmadd_ps(vx0, vlog2e, vmagic_bias);
74 __m256 vn1 = _mm256_fmadd_ps(vx1, vlog2e, vmagic_bias);
75 __m256 vn2 = _mm256_fmadd_ps(vx2, vlog2e, vmagic_bias);
76 __m256 vn3 = _mm256_fmadd_ps(vx3, vlog2e, vmagic_bias);
77 __m256 vn4 = _mm256_fmadd_ps(vx4, vlog2e, vmagic_bias);
78 __m256 vn5 = _mm256_fmadd_ps(vx5, vlog2e, vmagic_bias);
79 __m256 vn6 = _mm256_fmadd_ps(vx6, vlog2e, vmagic_bias);
80 __m256 vn7 = _mm256_fmadd_ps(vx7, vlog2e, vmagic_bias);
81 __m256 vn8 = _mm256_fmadd_ps(vx8, vlog2e, vmagic_bias);
82
83 // Create a floating-point number s (scale) such that s == 2**batch for inputs which don't cause underflow, i.e.
84 // -87.33642 <= x <= 0.0, and -126 <= batch <= 0 accordingly.
85 const __m256 vs0 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn0), 23));
86 const __m256 vs1 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn1), 23));
87 const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn2), 23));
88 const __m256 vs3 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn3), 23));
89 const __m256 vs4 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn4), 23));
90 const __m256 vs5 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn5), 23));
91 const __m256 vs6 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn6), 23));
92 const __m256 vs7 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn7), 23));
93 const __m256 vs8 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn8), 23));
94
95 // Subtract the large number back to get final batch := round(x / log(2)).
96 vn0 = _mm256_sub_ps(vn0, vmagic_bias);
97 vn1 = _mm256_sub_ps(vn1, vmagic_bias);
98 vn2 = _mm256_sub_ps(vn2, vmagic_bias);
99 vn3 = _mm256_sub_ps(vn3, vmagic_bias);
100 vn4 = _mm256_sub_ps(vn4, vmagic_bias);
101 vn5 = _mm256_sub_ps(vn5, vmagic_bias);
102 vn6 = _mm256_sub_ps(vn6, vmagic_bias);
103 vn7 = _mm256_sub_ps(vn7, vmagic_bias);
104 vn8 = _mm256_sub_ps(vn8, vmagic_bias);
105
106 // Compute reduced argument t := x - batch * log(2).
107 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
108 __m256 vt0 = _mm256_fmadd_ps(vn0, vminus_ln2_hi, vx0);
109 __m256 vt1 = _mm256_fmadd_ps(vn1, vminus_ln2_hi, vx1);
110 __m256 vt2 = _mm256_fmadd_ps(vn2, vminus_ln2_hi, vx2);
111 __m256 vt3 = _mm256_fmadd_ps(vn3, vminus_ln2_hi, vx3);
112 __m256 vt4 = _mm256_fmadd_ps(vn4, vminus_ln2_hi, vx4);
113 __m256 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_hi, vx5);
114 __m256 vt6 = _mm256_fmadd_ps(vn6, vminus_ln2_hi, vx6);
115 __m256 vt7 = _mm256_fmadd_ps(vn7, vminus_ln2_hi, vx7);
116 __m256 vt8 = _mm256_fmadd_ps(vn8, vminus_ln2_hi, vx8);
117
118 vt0 = _mm256_fmadd_ps(vn0, vminus_ln2_lo, vt0);
119 vt1 = _mm256_fmadd_ps(vn1, vminus_ln2_lo, vt1);
120 vt2 = _mm256_fmadd_ps(vn2, vminus_ln2_lo, vt2);
121 vt3 = _mm256_fmadd_ps(vn3, vminus_ln2_lo, vt3);
122 vt4 = _mm256_fmadd_ps(vn4, vminus_ln2_lo, vt4);
123 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_lo, vt5);
124 vt6 = _mm256_fmadd_ps(vn6, vminus_ln2_lo, vt6);
125 vt7 = _mm256_fmadd_ps(vn7, vminus_ln2_lo, vt7);
126 vt8 = _mm256_fmadd_ps(vn8, vminus_ln2_lo, vt8);
127
128 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
129 __m256 vp0 = _mm256_fmadd_ps(vc5, vt0, vc4);
130 __m256 vp1 = _mm256_fmadd_ps(vc5, vt1, vc4);
131 __m256 vp2 = _mm256_fmadd_ps(vc5, vt2, vc4);
132 __m256 vp3 = _mm256_fmadd_ps(vc5, vt3, vc4);
133 __m256 vp4 = _mm256_fmadd_ps(vc5, vt4, vc4);
134 __m256 vp5 = _mm256_fmadd_ps(vc5, vt5, vc4);
135 __m256 vp6 = _mm256_fmadd_ps(vc5, vt6, vc4);
136 __m256 vp7 = _mm256_fmadd_ps(vc5, vt7, vc4);
137 __m256 vp8 = _mm256_fmadd_ps(vc5, vt8, vc4);
138
139 vp0 = _mm256_fmadd_ps(vp0, vt0, vc3);
140 vp1 = _mm256_fmadd_ps(vp1, vt1, vc3);
141 vp2 = _mm256_fmadd_ps(vp2, vt2, vc3);
142 vp3 = _mm256_fmadd_ps(vp3, vt3, vc3);
143 vp4 = _mm256_fmadd_ps(vp4, vt4, vc3);
144 vp5 = _mm256_fmadd_ps(vp5, vt5, vc3);
145 vp6 = _mm256_fmadd_ps(vp6, vt6, vc3);
146 vp7 = _mm256_fmadd_ps(vp7, vt7, vc3);
147 vp8 = _mm256_fmadd_ps(vp8, vt8, vc3);
148
149 vp0 = _mm256_fmadd_ps(vp0, vt0, vc2);
150 vp1 = _mm256_fmadd_ps(vp1, vt1, vc2);
151 vp2 = _mm256_fmadd_ps(vp2, vt2, vc2);
152 vp3 = _mm256_fmadd_ps(vp3, vt3, vc2);
153 vp4 = _mm256_fmadd_ps(vp4, vt4, vc2);
154 vp5 = _mm256_fmadd_ps(vp5, vt5, vc2);
155 vp6 = _mm256_fmadd_ps(vp6, vt6, vc2);
156 vp7 = _mm256_fmadd_ps(vp7, vt7, vc2);
157 vp8 = _mm256_fmadd_ps(vp8, vt8, vc2);
158
159 vp0 = _mm256_fmadd_ps(vp0, vt0, vc1);
160 vp1 = _mm256_fmadd_ps(vp1, vt1, vc1);
161 vp2 = _mm256_fmadd_ps(vp2, vt2, vc1);
162 vp3 = _mm256_fmadd_ps(vp3, vt3, vc1);
163 vp4 = _mm256_fmadd_ps(vp4, vt4, vc1);
164 vp5 = _mm256_fmadd_ps(vp5, vt5, vc1);
165 vp6 = _mm256_fmadd_ps(vp6, vt6, vc1);
166 vp7 = _mm256_fmadd_ps(vp7, vt7, vc1);
167 vp8 = _mm256_fmadd_ps(vp8, vt8, vc1);
168
169 // Reconstruct the final f value:
170 // f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
171 // = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
172 // = s + (t * s) * p
173 vt0 = _mm256_mul_ps(vt0, vs0);
174 vt1 = _mm256_mul_ps(vt1, vs1);
175 vt2 = _mm256_mul_ps(vt2, vs2);
176 vt3 = _mm256_mul_ps(vt3, vs3);
177 vt4 = _mm256_mul_ps(vt4, vs4);
178 vt5 = _mm256_mul_ps(vt5, vs5);
179 vt6 = _mm256_mul_ps(vt6, vs6);
180 vt7 = _mm256_mul_ps(vt7, vs7);
181 vt8 = _mm256_mul_ps(vt8, vs8);
182
183 __m256 vf0 = _mm256_fmadd_ps(vt0, vp0, vs0);
184 __m256 vf1 = _mm256_fmadd_ps(vt1, vp1, vs1);
185 __m256 vf2 = _mm256_fmadd_ps(vt2, vp2, vs2);
186 __m256 vf3 = _mm256_fmadd_ps(vt3, vp3, vs3);
187 __m256 vf4 = _mm256_fmadd_ps(vt4, vp4, vs4);
188 __m256 vf5 = _mm256_fmadd_ps(vt5, vp5, vs5);
189 __m256 vf6 = _mm256_fmadd_ps(vt6, vp6, vs6);
190 __m256 vf7 = _mm256_fmadd_ps(vt7, vp7, vs7);
191 __m256 vf8 = _mm256_fmadd_ps(vt8, vp8, vs8);
192
193 // For inputs below zero cutoff, replace output with +0.0f.
194 // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
195 vf0 = _mm256_andnot_ps(_mm256_cmp_ps(vx0, vdenorm_cutoff, _CMP_LT_OS), vf0);
196 vf1 = _mm256_andnot_ps(_mm256_cmp_ps(vx1, vdenorm_cutoff, _CMP_LT_OS), vf1);
197 vf2 = _mm256_andnot_ps(_mm256_cmp_ps(vx2, vdenorm_cutoff, _CMP_LT_OS), vf2);
198 vf3 = _mm256_andnot_ps(_mm256_cmp_ps(vx3, vdenorm_cutoff, _CMP_LT_OS), vf3);
199 vf4 = _mm256_andnot_ps(_mm256_cmp_ps(vx4, vdenorm_cutoff, _CMP_LT_OS), vf4);
200 vf5 = _mm256_andnot_ps(_mm256_cmp_ps(vx5, vdenorm_cutoff, _CMP_LT_OS), vf5);
201 vf6 = _mm256_andnot_ps(_mm256_cmp_ps(vx6, vdenorm_cutoff, _CMP_LT_OS), vf6);
202 vf7 = _mm256_andnot_ps(_mm256_cmp_ps(vx7, vdenorm_cutoff, _CMP_LT_OS), vf7);
203 vf8 = _mm256_andnot_ps(_mm256_cmp_ps(vx8, vdenorm_cutoff, _CMP_LT_OS), vf8);
204
205 // Multiply by scale.
206 vf0 = _mm256_mul_ps(vf0, vscale);
207 vf1 = _mm256_mul_ps(vf1, vscale);
208 vf2 = _mm256_mul_ps(vf2, vscale);
209 vf3 = _mm256_mul_ps(vf3, vscale);
210 vf4 = _mm256_mul_ps(vf4, vscale);
211 vf5 = _mm256_mul_ps(vf5, vscale);
212 vf6 = _mm256_mul_ps(vf6, vscale);
213 vf7 = _mm256_mul_ps(vf7, vscale);
214 vf8 = _mm256_mul_ps(vf8, vscale);
215
216 // Store 72 (9x8) outputs at a time.
217 _mm256_storeu_ps(output, vf0);
218 _mm256_storeu_ps(output + 8, vf1);
219 _mm256_storeu_ps(output + 16, vf2);
220 _mm256_storeu_ps(output + 24, vf3);
221 _mm256_storeu_ps(output + 32, vf4);
222 _mm256_storeu_ps(output + 40, vf5);
223 _mm256_storeu_ps(output + 48, vf6);
224 _mm256_storeu_ps(output + 56, vf7);
225 _mm256_storeu_ps(output + 64, vf8);
226 output += 72;
227 }
228 for (; batch >= 8 * sizeof(float); batch -= 8 * sizeof(float)) {
229 // Load 8 inputs at a time.
230 const __m256 vi = _mm256_loadu_ps(input);
231 input += 8;
232
233 // Subtract maximum input x := i - i_max. This implies x <= 0.
234 const __m256 vx = _mm256_sub_ps(vi, vi_max);
235
236 // Compute reduced argument batch := round(x / log(2)).
237 __m256 vn = _mm256_fmadd_ps(vx, vlog2e, vmagic_bias);
238
239 // Create a floating-point number s (scale) such that s == 2**batch for inputs which don't cause underflow, i.e.
240 // -87.33642 <= x <= 0.0, and -126 <= batch <= 0 accordingly.
241 const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
242
243 // Subtract the large number back to get final batch := round(x / log(2)).
244 vn = _mm256_sub_ps(vn, vmagic_bias);
245
246 // Compute reduced argument t := x - batch * log(2).
247 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
248 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2_hi, vx);
249 vt = _mm256_fmadd_ps(vn, vminus_ln2_lo, vt);
250
251 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
252 __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
253 vp = _mm256_fmadd_ps(vp, vt, vc3);
254 vp = _mm256_fmadd_ps(vp, vt, vc2);
255 vp = _mm256_fmadd_ps(vp, vt, vc1);
256
257 // Reconstruct the final f value:
258 // f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
259 // = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
260 // = s + (t * s) * p
261 vt = _mm256_mul_ps(vt, vs);
262 __m256 vf = _mm256_fmadd_ps(vt, vp, vs);
263
264 // For inputs below zero cutoff, replace output with +0.0f.
265 // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
266 vf = _mm256_andnot_ps(_mm256_cmp_ps(vx, vdenorm_cutoff, _CMP_LT_OS), vf);
267
268 // Multiply by scale.
269 vf = _mm256_mul_ps(vf, vscale);
270
271 // Store 64 (8x8) outputs at a time.
272 _mm256_storeu_ps(output, vf);
273 output += 8;
274 }
275 if (batch != 0) {
276 assert(batch >= 1 * sizeof(float));
277 assert(batch <= 7 * sizeof(float));
278 const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - batch));
279
280 // Load up to 7 inputs at a time.
281 const __m256 vi = _mm256_maskload_ps(input, vmask);
282
283 // Subtract maximum input x := i - i_max. This implies x <= 0.
284 const __m256 vx = _mm256_sub_ps(vi, vi_max);
285
286 // Compute reduced argument batch := round(x / log(2)).
287 __m256 vn = _mm256_fmadd_ps(vx, vlog2e, vmagic_bias);
288
289 // Create a floating-point number s (scale) such that s == 2**batch for inputs which don't cause underflow, i.e.
290 // -87.33642 <= x <= 0.0, and -126 <= batch <= 0 accordingly.
291 const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
292
293 // Subtract the large number back to get final batch := round(x / log(2)).
294 vn = _mm256_sub_ps(vn, vmagic_bias);
295
296 // Compute reduced argument t := x - batch * log(2).
297 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
298 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2_hi, vx);
299 vt = _mm256_fmadd_ps(vn, vminus_ln2_lo, vt);
300
301 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
302 __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
303 vp = _mm256_fmadd_ps(vp, vt, vc3);
304 vp = _mm256_fmadd_ps(vp, vt, vc2);
305 vp = _mm256_fmadd_ps(vp, vt, vc1);
306
307 // Reconstruct the final f value:
308 // f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
309 // = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
310 // = s + (t * s) * p
311 vt = _mm256_mul_ps(vt, vs);
312 __m256 vf = _mm256_fmadd_ps(vt, vp, vs);
313
314 // For inputs below zero cutoff, replace output with +0.0f.
315 // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
316 vf = _mm256_andnot_ps(_mm256_cmp_ps(vx, vdenorm_cutoff, _CMP_LT_OS), vf);
317
318 // Multiply by scale.
319 vf = _mm256_mul_ps(vf, vscale);
320
321 // Store up to 7 outputs at a time.
322 _mm256_maskstore_ps(output, vmask, vf);
323 }
324 }