Debian Codegraph @

src/f32-vscaleexpminusmax/gen/f32-vscaleexpminusmax-avx2-p5-u64.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_u64(
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 >= 64 * sizeof(float); batch -= 64 * sizeof(float)) {
49 // Load 64 (8x8) 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 input += 64;
59
60 // Subtract maximum input x := i - i_max. This implies x <= 0.
61 const __m256 vx0 = _mm256_sub_ps(vi0, vi_max);
62 const __m256 vx1 = _mm256_sub_ps(vi1, vi_max);
63 const __m256 vx2 = _mm256_sub_ps(vi2, vi_max);
64 const __m256 vx3 = _mm256_sub_ps(vi3, vi_max);
65 const __m256 vx4 = _mm256_sub_ps(vi4, vi_max);
66 const __m256 vx5 = _mm256_sub_ps(vi5, vi_max);
67 const __m256 vx6 = _mm256_sub_ps(vi6, vi_max);
68 const __m256 vx7 = _mm256_sub_ps(vi7, vi_max);
69
70 // Compute reduced argument batch := round(x / log(2)).
71 __m256 vn0 = _mm256_fmadd_ps(vx0, vlog2e, vmagic_bias);
72 __m256 vn1 = _mm256_fmadd_ps(vx1, vlog2e, vmagic_bias);
73 __m256 vn2 = _mm256_fmadd_ps(vx2, vlog2e, vmagic_bias);
74 __m256 vn3 = _mm256_fmadd_ps(vx3, vlog2e, vmagic_bias);
75 __m256 vn4 = _mm256_fmadd_ps(vx4, vlog2e, vmagic_bias);
76 __m256 vn5 = _mm256_fmadd_ps(vx5, vlog2e, vmagic_bias);
77 __m256 vn6 = _mm256_fmadd_ps(vx6, vlog2e, vmagic_bias);
78 __m256 vn7 = _mm256_fmadd_ps(vx7, vlog2e, vmagic_bias);
79
80 // Create a floating-point number s (scale) such that s == 2**batch for inputs which don't cause underflow, i.e.
81 // -87.33642 <= x <= 0.0, and -126 <= batch <= 0 accordingly.
82 const __m256 vs0 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn0), 23));
83 const __m256 vs1 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn1), 23));
84 const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn2), 23));
85 const __m256 vs3 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn3), 23));
86 const __m256 vs4 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn4), 23));
87 const __m256 vs5 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn5), 23));
88 const __m256 vs6 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn6), 23));
89 const __m256 vs7 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn7), 23));
90
91 // Subtract the large number back to get final batch := round(x / log(2)).
92 vn0 = _mm256_sub_ps(vn0, vmagic_bias);
93 vn1 = _mm256_sub_ps(vn1, vmagic_bias);
94 vn2 = _mm256_sub_ps(vn2, vmagic_bias);
95 vn3 = _mm256_sub_ps(vn3, vmagic_bias);
96 vn4 = _mm256_sub_ps(vn4, vmagic_bias);
97 vn5 = _mm256_sub_ps(vn5, vmagic_bias);
98 vn6 = _mm256_sub_ps(vn6, vmagic_bias);
99 vn7 = _mm256_sub_ps(vn7, vmagic_bias);
100
101 // Compute reduced argument t := x - batch * log(2).
102 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
103 __m256 vt0 = _mm256_fmadd_ps(vn0, vminus_ln2_hi, vx0);
104 __m256 vt1 = _mm256_fmadd_ps(vn1, vminus_ln2_hi, vx1);
105 __m256 vt2 = _mm256_fmadd_ps(vn2, vminus_ln2_hi, vx2);
106 __m256 vt3 = _mm256_fmadd_ps(vn3, vminus_ln2_hi, vx3);
107 __m256 vt4 = _mm256_fmadd_ps(vn4, vminus_ln2_hi, vx4);
108 __m256 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_hi, vx5);
109 __m256 vt6 = _mm256_fmadd_ps(vn6, vminus_ln2_hi, vx6);
110 __m256 vt7 = _mm256_fmadd_ps(vn7, vminus_ln2_hi, vx7);
111
112 vt0 = _mm256_fmadd_ps(vn0, vminus_ln2_lo, vt0);
113 vt1 = _mm256_fmadd_ps(vn1, vminus_ln2_lo, vt1);
114 vt2 = _mm256_fmadd_ps(vn2, vminus_ln2_lo, vt2);
115 vt3 = _mm256_fmadd_ps(vn3, vminus_ln2_lo, vt3);
116 vt4 = _mm256_fmadd_ps(vn4, vminus_ln2_lo, vt4);
117 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_lo, vt5);
118 vt6 = _mm256_fmadd_ps(vn6, vminus_ln2_lo, vt6);
119 vt7 = _mm256_fmadd_ps(vn7, vminus_ln2_lo, vt7);
120
121 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
122 __m256 vp0 = _mm256_fmadd_ps(vc5, vt0, vc4);
123 __m256 vp1 = _mm256_fmadd_ps(vc5, vt1, vc4);
124 __m256 vp2 = _mm256_fmadd_ps(vc5, vt2, vc4);
125 __m256 vp3 = _mm256_fmadd_ps(vc5, vt3, vc4);
126 __m256 vp4 = _mm256_fmadd_ps(vc5, vt4, vc4);
127 __m256 vp5 = _mm256_fmadd_ps(vc5, vt5, vc4);
128 __m256 vp6 = _mm256_fmadd_ps(vc5, vt6, vc4);
129 __m256 vp7 = _mm256_fmadd_ps(vc5, vt7, vc4);
130
131 vp0 = _mm256_fmadd_ps(vp0, vt0, vc3);
132 vp1 = _mm256_fmadd_ps(vp1, vt1, vc3);
133 vp2 = _mm256_fmadd_ps(vp2, vt2, vc3);
134 vp3 = _mm256_fmadd_ps(vp3, vt3, vc3);
135 vp4 = _mm256_fmadd_ps(vp4, vt4, vc3);
136 vp5 = _mm256_fmadd_ps(vp5, vt5, vc3);
137 vp6 = _mm256_fmadd_ps(vp6, vt6, vc3);
138 vp7 = _mm256_fmadd_ps(vp7, vt7, vc3);
139
140 vp0 = _mm256_fmadd_ps(vp0, vt0, vc2);
141 vp1 = _mm256_fmadd_ps(vp1, vt1, vc2);
142 vp2 = _mm256_fmadd_ps(vp2, vt2, vc2);
143 vp3 = _mm256_fmadd_ps(vp3, vt3, vc2);
144 vp4 = _mm256_fmadd_ps(vp4, vt4, vc2);
145 vp5 = _mm256_fmadd_ps(vp5, vt5, vc2);
146 vp6 = _mm256_fmadd_ps(vp6, vt6, vc2);
147 vp7 = _mm256_fmadd_ps(vp7, vt7, vc2);
148
149 vp0 = _mm256_fmadd_ps(vp0, vt0, vc1);
150 vp1 = _mm256_fmadd_ps(vp1, vt1, vc1);
151 vp2 = _mm256_fmadd_ps(vp2, vt2, vc1);
152 vp3 = _mm256_fmadd_ps(vp3, vt3, vc1);
153 vp4 = _mm256_fmadd_ps(vp4, vt4, vc1);
154 vp5 = _mm256_fmadd_ps(vp5, vt5, vc1);
155 vp6 = _mm256_fmadd_ps(vp6, vt6, vc1);
156 vp7 = _mm256_fmadd_ps(vp7, vt7, 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 vt0 = _mm256_mul_ps(vt0, vs0);
163 vt1 = _mm256_mul_ps(vt1, vs1);
164 vt2 = _mm256_mul_ps(vt2, vs2);
165 vt3 = _mm256_mul_ps(vt3, vs3);
166 vt4 = _mm256_mul_ps(vt4, vs4);
167 vt5 = _mm256_mul_ps(vt5, vs5);
168 vt6 = _mm256_mul_ps(vt6, vs6);
169 vt7 = _mm256_mul_ps(vt7, vs7);
170
171 __m256 vf0 = _mm256_fmadd_ps(vt0, vp0, vs0);
172 __m256 vf1 = _mm256_fmadd_ps(vt1, vp1, vs1);
173 __m256 vf2 = _mm256_fmadd_ps(vt2, vp2, vs2);
174 __m256 vf3 = _mm256_fmadd_ps(vt3, vp3, vs3);
175 __m256 vf4 = _mm256_fmadd_ps(vt4, vp4, vs4);
176 __m256 vf5 = _mm256_fmadd_ps(vt5, vp5, vs5);
177 __m256 vf6 = _mm256_fmadd_ps(vt6, vp6, vs6);
178 __m256 vf7 = _mm256_fmadd_ps(vt7, vp7, vs7);
179
180 // For inputs below zero cutoff, replace output with +0.0f.
181 // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
182 vf0 = _mm256_andnot_ps(_mm256_cmp_ps(vx0, vdenorm_cutoff, _CMP_LT_OS), vf0);
183 vf1 = _mm256_andnot_ps(_mm256_cmp_ps(vx1, vdenorm_cutoff, _CMP_LT_OS), vf1);
184 vf2 = _mm256_andnot_ps(_mm256_cmp_ps(vx2, vdenorm_cutoff, _CMP_LT_OS), vf2);
185 vf3 = _mm256_andnot_ps(_mm256_cmp_ps(vx3, vdenorm_cutoff, _CMP_LT_OS), vf3);
186 vf4 = _mm256_andnot_ps(_mm256_cmp_ps(vx4, vdenorm_cutoff, _CMP_LT_OS), vf4);
187 vf5 = _mm256_andnot_ps(_mm256_cmp_ps(vx5, vdenorm_cutoff, _CMP_LT_OS), vf5);
188 vf6 = _mm256_andnot_ps(_mm256_cmp_ps(vx6, vdenorm_cutoff, _CMP_LT_OS), vf6);
189 vf7 = _mm256_andnot_ps(_mm256_cmp_ps(vx7, vdenorm_cutoff, _CMP_LT_OS), vf7);
190
191 // Multiply by scale.
192 vf0 = _mm256_mul_ps(vf0, vscale);
193 vf1 = _mm256_mul_ps(vf1, vscale);
194 vf2 = _mm256_mul_ps(vf2, vscale);
195 vf3 = _mm256_mul_ps(vf3, vscale);
196 vf4 = _mm256_mul_ps(vf4, vscale);
197 vf5 = _mm256_mul_ps(vf5, vscale);
198 vf6 = _mm256_mul_ps(vf6, vscale);
199 vf7 = _mm256_mul_ps(vf7, vscale);
200
201 // Store 64 (8x8) outputs at a time.
202 _mm256_storeu_ps(output, vf0);
203 _mm256_storeu_ps(output + 8, vf1);
204 _mm256_storeu_ps(output + 16, vf2);
205 _mm256_storeu_ps(output + 24, vf3);
206 _mm256_storeu_ps(output + 32, vf4);
207 _mm256_storeu_ps(output + 40, vf5);
208 _mm256_storeu_ps(output + 48, vf6);
209 _mm256_storeu_ps(output + 56, vf7);
210 output += 64;
211 }
212 for (; batch >= 8 * sizeof(float); batch -= 8 * sizeof(float)) {
213 // Load 8 inputs at a time.
214 const __m256 vi = _mm256_loadu_ps(input);
215 input += 8;
216
217 // Subtract maximum input x := i - i_max. This implies x <= 0.
218 const __m256 vx = _mm256_sub_ps(vi, vi_max);
219
220 // Compute reduced argument batch := round(x / log(2)).
221 __m256 vn = _mm256_fmadd_ps(vx, vlog2e, vmagic_bias);
222
223 // Create a floating-point number s (scale) such that s == 2**batch for inputs which don't cause underflow, i.e.
224 // -87.33642 <= x <= 0.0, and -126 <= batch <= 0 accordingly.
225 const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
226
227 // Subtract the large number back to get final batch := round(x / log(2)).
228 vn = _mm256_sub_ps(vn, vmagic_bias);
229
230 // Compute reduced argument t := x - batch * log(2).
231 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
232 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2_hi, vx);
233 vt = _mm256_fmadd_ps(vn, vminus_ln2_lo, vt);
234
235 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
236 __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
237 vp = _mm256_fmadd_ps(vp, vt, vc3);
238 vp = _mm256_fmadd_ps(vp, vt, vc2);
239 vp = _mm256_fmadd_ps(vp, vt, vc1);
240
241 // Reconstruct the final f value:
242 // f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
243 // = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
244 // = s + (t * s) * p
245 vt = _mm256_mul_ps(vt, vs);
246 __m256 vf = _mm256_fmadd_ps(vt, vp, vs);
247
248 // For inputs below zero cutoff, replace output with +0.0f.
249 // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
250 vf = _mm256_andnot_ps(_mm256_cmp_ps(vx, vdenorm_cutoff, _CMP_LT_OS), vf);
251
252 // Multiply by scale.
253 vf = _mm256_mul_ps(vf, vscale);
254
255 // Store 64 (8x8) outputs at a time.
256 _mm256_storeu_ps(output, vf);
257 output += 8;
258 }
259 if (batch != 0) {
260 assert(batch >= 1 * sizeof(float));
261 assert(batch <= 7 * sizeof(float));
262 const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - batch));
263
264 // Load up to 7 inputs at a time.
265 const __m256 vi = _mm256_maskload_ps(input, vmask);
266
267 // Subtract maximum input x := i - i_max. This implies x <= 0.
268 const __m256 vx = _mm256_sub_ps(vi, vi_max);
269
270 // Compute reduced argument batch := round(x / log(2)).
271 __m256 vn = _mm256_fmadd_ps(vx, vlog2e, vmagic_bias);
272
273 // Create a floating-point number s (scale) such that s == 2**batch for inputs which don't cause underflow, i.e.
274 // -87.33642 <= x <= 0.0, and -126 <= batch <= 0 accordingly.
275 const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
276
277 // Subtract the large number back to get final batch := round(x / log(2)).
278 vn = _mm256_sub_ps(vn, vmagic_bias);
279
280 // Compute reduced argument t := x - batch * log(2).
281 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
282 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2_hi, vx);
283 vt = _mm256_fmadd_ps(vn, vminus_ln2_lo, vt);
284
285 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
286 __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
287 vp = _mm256_fmadd_ps(vp, vt, vc3);
288 vp = _mm256_fmadd_ps(vp, vt, vc2);
289 vp = _mm256_fmadd_ps(vp, vt, vc1);
290
291 // Reconstruct the final f value:
292 // f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
293 // = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
294 // = s + (t * s) * p
295 vt = _mm256_mul_ps(vt, vs);
296 __m256 vf = _mm256_fmadd_ps(vt, vp, vs);
297
298 // For inputs below zero cutoff, replace output with +0.0f.
299 // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
300 vf = _mm256_andnot_ps(_mm256_cmp_ps(vx, vdenorm_cutoff, _CMP_LT_OS), vf);
301
302 // Multiply by scale.
303 vf = _mm256_mul_ps(vf, vscale);
304
305 // Store up to 7 outputs at a time.
306 _mm256_maskstore_ps(output, vmask, vf);
307 }
308 }