Debian Codegraph @

src/f32-vscaleexpminusmax/gen/f32-vscaleexpminusmax-avx2-p5-u56.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_u56(
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 >= 56 * sizeof(float); batch -= 56 * sizeof(float)) {
49 // Load 56 (7x8) 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 input += 56;
58
59 // Subtract maximum input x := i - i_max. This implies x <= 0.
60 const __m256 vx0 = _mm256_sub_ps(vi0, vi_max);
61 const __m256 vx1 = _mm256_sub_ps(vi1, vi_max);
62 const __m256 vx2 = _mm256_sub_ps(vi2, vi_max);
63 const __m256 vx3 = _mm256_sub_ps(vi3, vi_max);
64 const __m256 vx4 = _mm256_sub_ps(vi4, vi_max);
65 const __m256 vx5 = _mm256_sub_ps(vi5, vi_max);
66 const __m256 vx6 = _mm256_sub_ps(vi6, vi_max);
67
68 // Compute reduced argument batch := round(x / log(2)).
69 __m256 vn0 = _mm256_fmadd_ps(vx0, vlog2e, vmagic_bias);
70 __m256 vn1 = _mm256_fmadd_ps(vx1, vlog2e, vmagic_bias);
71 __m256 vn2 = _mm256_fmadd_ps(vx2, vlog2e, vmagic_bias);
72 __m256 vn3 = _mm256_fmadd_ps(vx3, vlog2e, vmagic_bias);
73 __m256 vn4 = _mm256_fmadd_ps(vx4, vlog2e, vmagic_bias);
74 __m256 vn5 = _mm256_fmadd_ps(vx5, vlog2e, vmagic_bias);
75 __m256 vn6 = _mm256_fmadd_ps(vx6, vlog2e, vmagic_bias);
76
77 // Create a floating-point number s (scale) such that s == 2**batch for inputs which don't cause underflow, i.e.
78 // -87.33642 <= x <= 0.0, and -126 <= batch <= 0 accordingly.
79 const __m256 vs0 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn0), 23));
80 const __m256 vs1 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn1), 23));
81 const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn2), 23));
82 const __m256 vs3 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn3), 23));
83 const __m256 vs4 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn4), 23));
84 const __m256 vs5 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn5), 23));
85 const __m256 vs6 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn6), 23));
86
87 // Subtract the large number back to get final batch := round(x / log(2)).
88 vn0 = _mm256_sub_ps(vn0, vmagic_bias);
89 vn1 = _mm256_sub_ps(vn1, vmagic_bias);
90 vn2 = _mm256_sub_ps(vn2, vmagic_bias);
91 vn3 = _mm256_sub_ps(vn3, vmagic_bias);
92 vn4 = _mm256_sub_ps(vn4, vmagic_bias);
93 vn5 = _mm256_sub_ps(vn5, vmagic_bias);
94 vn6 = _mm256_sub_ps(vn6, vmagic_bias);
95
96 // Compute reduced argument t := x - batch * log(2).
97 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
98 __m256 vt0 = _mm256_fmadd_ps(vn0, vminus_ln2_hi, vx0);
99 __m256 vt1 = _mm256_fmadd_ps(vn1, vminus_ln2_hi, vx1);
100 __m256 vt2 = _mm256_fmadd_ps(vn2, vminus_ln2_hi, vx2);
101 __m256 vt3 = _mm256_fmadd_ps(vn3, vminus_ln2_hi, vx3);
102 __m256 vt4 = _mm256_fmadd_ps(vn4, vminus_ln2_hi, vx4);
103 __m256 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_hi, vx5);
104 __m256 vt6 = _mm256_fmadd_ps(vn6, vminus_ln2_hi, vx6);
105
106 vt0 = _mm256_fmadd_ps(vn0, vminus_ln2_lo, vt0);
107 vt1 = _mm256_fmadd_ps(vn1, vminus_ln2_lo, vt1);
108 vt2 = _mm256_fmadd_ps(vn2, vminus_ln2_lo, vt2);
109 vt3 = _mm256_fmadd_ps(vn3, vminus_ln2_lo, vt3);
110 vt4 = _mm256_fmadd_ps(vn4, vminus_ln2_lo, vt4);
111 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_lo, vt5);
112 vt6 = _mm256_fmadd_ps(vn6, vminus_ln2_lo, vt6);
113
114 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
115 __m256 vp0 = _mm256_fmadd_ps(vc5, vt0, vc4);
116 __m256 vp1 = _mm256_fmadd_ps(vc5, vt1, vc4);
117 __m256 vp2 = _mm256_fmadd_ps(vc5, vt2, vc4);
118 __m256 vp3 = _mm256_fmadd_ps(vc5, vt3, vc4);
119 __m256 vp4 = _mm256_fmadd_ps(vc5, vt4, vc4);
120 __m256 vp5 = _mm256_fmadd_ps(vc5, vt5, vc4);
121 __m256 vp6 = _mm256_fmadd_ps(vc5, vt6, vc4);
122
123 vp0 = _mm256_fmadd_ps(vp0, vt0, vc3);
124 vp1 = _mm256_fmadd_ps(vp1, vt1, vc3);
125 vp2 = _mm256_fmadd_ps(vp2, vt2, vc3);
126 vp3 = _mm256_fmadd_ps(vp3, vt3, vc3);
127 vp4 = _mm256_fmadd_ps(vp4, vt4, vc3);
128 vp5 = _mm256_fmadd_ps(vp5, vt5, vc3);
129 vp6 = _mm256_fmadd_ps(vp6, vt6, vc3);
130
131 vp0 = _mm256_fmadd_ps(vp0, vt0, vc2);
132 vp1 = _mm256_fmadd_ps(vp1, vt1, vc2);
133 vp2 = _mm256_fmadd_ps(vp2, vt2, vc2);
134 vp3 = _mm256_fmadd_ps(vp3, vt3, vc2);
135 vp4 = _mm256_fmadd_ps(vp4, vt4, vc2);
136 vp5 = _mm256_fmadd_ps(vp5, vt5, vc2);
137 vp6 = _mm256_fmadd_ps(vp6, vt6, vc2);
138
139 vp0 = _mm256_fmadd_ps(vp0, vt0, vc1);
140 vp1 = _mm256_fmadd_ps(vp1, vt1, vc1);
141 vp2 = _mm256_fmadd_ps(vp2, vt2, vc1);
142 vp3 = _mm256_fmadd_ps(vp3, vt3, vc1);
143 vp4 = _mm256_fmadd_ps(vp4, vt4, vc1);
144 vp5 = _mm256_fmadd_ps(vp5, vt5, vc1);
145 vp6 = _mm256_fmadd_ps(vp6, vt6, vc1);
146
147 // Reconstruct the final f value:
148 // f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
149 // = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
150 // = s + (t * s) * p
151 vt0 = _mm256_mul_ps(vt0, vs0);
152 vt1 = _mm256_mul_ps(vt1, vs1);
153 vt2 = _mm256_mul_ps(vt2, vs2);
154 vt3 = _mm256_mul_ps(vt3, vs3);
155 vt4 = _mm256_mul_ps(vt4, vs4);
156 vt5 = _mm256_mul_ps(vt5, vs5);
157 vt6 = _mm256_mul_ps(vt6, vs6);
158
159 __m256 vf0 = _mm256_fmadd_ps(vt0, vp0, vs0);
160 __m256 vf1 = _mm256_fmadd_ps(vt1, vp1, vs1);
161 __m256 vf2 = _mm256_fmadd_ps(vt2, vp2, vs2);
162 __m256 vf3 = _mm256_fmadd_ps(vt3, vp3, vs3);
163 __m256 vf4 = _mm256_fmadd_ps(vt4, vp4, vs4);
164 __m256 vf5 = _mm256_fmadd_ps(vt5, vp5, vs5);
165 __m256 vf6 = _mm256_fmadd_ps(vt6, vp6, vs6);
166
167 // For inputs below zero cutoff, replace output with +0.0f.
168 // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
169 vf0 = _mm256_andnot_ps(_mm256_cmp_ps(vx0, vdenorm_cutoff, _CMP_LT_OS), vf0);
170 vf1 = _mm256_andnot_ps(_mm256_cmp_ps(vx1, vdenorm_cutoff, _CMP_LT_OS), vf1);
171 vf2 = _mm256_andnot_ps(_mm256_cmp_ps(vx2, vdenorm_cutoff, _CMP_LT_OS), vf2);
172 vf3 = _mm256_andnot_ps(_mm256_cmp_ps(vx3, vdenorm_cutoff, _CMP_LT_OS), vf3);
173 vf4 = _mm256_andnot_ps(_mm256_cmp_ps(vx4, vdenorm_cutoff, _CMP_LT_OS), vf4);
174 vf5 = _mm256_andnot_ps(_mm256_cmp_ps(vx5, vdenorm_cutoff, _CMP_LT_OS), vf5);
175 vf6 = _mm256_andnot_ps(_mm256_cmp_ps(vx6, vdenorm_cutoff, _CMP_LT_OS), vf6);
176
177 // Multiply by scale.
178 vf0 = _mm256_mul_ps(vf0, vscale);
179 vf1 = _mm256_mul_ps(vf1, vscale);
180 vf2 = _mm256_mul_ps(vf2, vscale);
181 vf3 = _mm256_mul_ps(vf3, vscale);
182 vf4 = _mm256_mul_ps(vf4, vscale);
183 vf5 = _mm256_mul_ps(vf5, vscale);
184 vf6 = _mm256_mul_ps(vf6, vscale);
185
186 // Store 56 (7x8) outputs at a time.
187 _mm256_storeu_ps(output, vf0);
188 _mm256_storeu_ps(output + 8, vf1);
189 _mm256_storeu_ps(output + 16, vf2);
190 _mm256_storeu_ps(output + 24, vf3);
191 _mm256_storeu_ps(output + 32, vf4);
192 _mm256_storeu_ps(output + 40, vf5);
193 _mm256_storeu_ps(output + 48, vf6);
194 output += 56;
195 }
196 for (; batch >= 8 * sizeof(float); batch -= 8 * sizeof(float)) {
197 // Load 8 inputs at a time.
198 const __m256 vi = _mm256_loadu_ps(input);
199 input += 8;
200
201 // Subtract maximum input x := i - i_max. This implies x <= 0.
202 const __m256 vx = _mm256_sub_ps(vi, vi_max);
203
204 // Compute reduced argument batch := round(x / log(2)).
205 __m256 vn = _mm256_fmadd_ps(vx, vlog2e, vmagic_bias);
206
207 // Create a floating-point number s (scale) such that s == 2**batch for inputs which don't cause underflow, i.e.
208 // -87.33642 <= x <= 0.0, and -126 <= batch <= 0 accordingly.
209 const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
210
211 // Subtract the large number back to get final batch := round(x / log(2)).
212 vn = _mm256_sub_ps(vn, vmagic_bias);
213
214 // Compute reduced argument t := x - batch * log(2).
215 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
216 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2_hi, vx);
217 vt = _mm256_fmadd_ps(vn, vminus_ln2_lo, vt);
218
219 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
220 __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
221 vp = _mm256_fmadd_ps(vp, vt, vc3);
222 vp = _mm256_fmadd_ps(vp, vt, vc2);
223 vp = _mm256_fmadd_ps(vp, vt, vc1);
224
225 // Reconstruct the final f value:
226 // f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
227 // = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
228 // = s + (t * s) * p
229 vt = _mm256_mul_ps(vt, vs);
230 __m256 vf = _mm256_fmadd_ps(vt, vp, vs);
231
232 // For inputs below zero cutoff, replace output with +0.0f.
233 // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
234 vf = _mm256_andnot_ps(_mm256_cmp_ps(vx, vdenorm_cutoff, _CMP_LT_OS), vf);
235
236 // Multiply by scale.
237 vf = _mm256_mul_ps(vf, vscale);
238
239 // Store 64 (8x8) outputs at a time.
240 _mm256_storeu_ps(output, vf);
241 output += 8;
242 }
243 if (batch != 0) {
244 assert(batch >= 1 * sizeof(float));
245 assert(batch <= 7 * sizeof(float));
246 const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - batch));
247
248 // Load up to 7 inputs at a time.
249 const __m256 vi = _mm256_maskload_ps(input, vmask);
250
251 // Subtract maximum input x := i - i_max. This implies x <= 0.
252 const __m256 vx = _mm256_sub_ps(vi, vi_max);
253
254 // Compute reduced argument batch := round(x / log(2)).
255 __m256 vn = _mm256_fmadd_ps(vx, vlog2e, vmagic_bias);
256
257 // Create a floating-point number s (scale) such that s == 2**batch for inputs which don't cause underflow, i.e.
258 // -87.33642 <= x <= 0.0, and -126 <= batch <= 0 accordingly.
259 const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
260
261 // Subtract the large number back to get final batch := round(x / log(2)).
262 vn = _mm256_sub_ps(vn, vmagic_bias);
263
264 // Compute reduced argument t := x - batch * log(2).
265 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
266 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2_hi, vx);
267 vt = _mm256_fmadd_ps(vn, vminus_ln2_lo, vt);
268
269 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
270 __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
271 vp = _mm256_fmadd_ps(vp, vt, vc3);
272 vp = _mm256_fmadd_ps(vp, vt, vc2);
273 vp = _mm256_fmadd_ps(vp, vt, vc1);
274
275 // Reconstruct the final f value:
276 // f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
277 // = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
278 // = s + (t * s) * p
279 vt = _mm256_mul_ps(vt, vs);
280 __m256 vf = _mm256_fmadd_ps(vt, vp, vs);
281
282 // For inputs below zero cutoff, replace output with +0.0f.
283 // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
284 vf = _mm256_andnot_ps(_mm256_cmp_ps(vx, vdenorm_cutoff, _CMP_LT_OS), vf);
285
286 // Multiply by scale.
287 vf = _mm256_mul_ps(vf, vscale);
288
289 // Store up to 7 outputs at a time.
290 _mm256_maskstore_ps(output, vmask, vf);
291 }
292 }