• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-raddextexp/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 #include <math.h>
12 
13 #include <immintrin.h>
14 
15 #include <xnnpack/raddextexp.h>
16 
17 
18 static const int32_t mask_table[14] = {-1, -1, -1, -1, -1, -1, -1, 0, 0, 0, 0, 0, 0, 0};
19 
xnn_f32_raddextexp_ukernel__avx2_p5_x72(size_t elements,const float * x,float * sum)20 void xnn_f32_raddextexp_ukernel__avx2_p5_x72(
21     size_t elements,
22     const float* x,
23     float* sum)
24 {
25   assert(elements % sizeof(float) == 0);
26 
27   const __m256 vlog2e = _mm256_set1_ps(0x1.715476p+0f);
28   const __m256 vminus_ln2_hi = _mm256_set1_ps(-0x1.62E43p-1f);
29   const __m256 vminus_ln2_lo = _mm256_set1_ps(0x1.05C61p-29f);
30 
31   // The smallest elements such that 2**elements is considered non-negligible.
32   // For smaller elements, 2**elements is replaced with zero.
33   const __m256 vmin_exponent = _mm256_set1_ps(-127.0f);
34   const __m256 vmagic_bias = _mm256_set1_ps(0x1.8000FEp23f);
35   const __m256 vminus_inf = _mm256_set1_ps(-INFINITY);
36 
37   const __m256 vc0 = _mm256_set1_ps(1.0f);
38   const __m256 vc1 = _mm256_set1_ps(0x1.FFFFF6p-1f);
39   const __m256 vc2 = _mm256_set1_ps(0x1.FFFDC6p-2f);
40   const __m256 vc3 = _mm256_set1_ps(0x1.555A80p-3f);
41   const __m256 vc4 = _mm256_set1_ps(0x1.573A1Ap-5f);
42   const __m256 vc5 = _mm256_set1_ps(0x1.0F9F9Cp-7f);
43 
44   __m256 vaccv0 = _mm256_setzero_ps();
45   __m256 vacce0 = vminus_inf;
46   for (; elements >= 72 * sizeof(float); elements -= 72 * sizeof(float)) {
47     // Load 72 (9x8) inputs at a time.
48     const __m256 vx0 = _mm256_loadu_ps(x);
49     const __m256 vx1 = _mm256_loadu_ps(x + 8);
50     const __m256 vx2 = _mm256_loadu_ps(x + 16);
51     const __m256 vx3 = _mm256_loadu_ps(x + 24);
52     const __m256 vx4 = _mm256_loadu_ps(x + 32);
53     const __m256 vx5 = _mm256_loadu_ps(x + 40);
54     const __m256 vx6 = _mm256_loadu_ps(x + 48);
55     const __m256 vx7 = _mm256_loadu_ps(x + 56);
56     const __m256 vx8 = _mm256_loadu_ps(x + 64);
57     x += 72;
58 
59     // Compute reduced argument elements := round(x / log(2)).
60     const __m256 vn0 = _mm256_round_ps(_mm256_mul_ps(vx0, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
61     const __m256 vn1 = _mm256_round_ps(_mm256_mul_ps(vx1, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
62     const __m256 vn2 = _mm256_round_ps(_mm256_mul_ps(vx2, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
63     const __m256 vn3 = _mm256_round_ps(_mm256_mul_ps(vx3, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
64     const __m256 vn4 = _mm256_round_ps(_mm256_mul_ps(vx4, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
65     const __m256 vn5 = _mm256_round_ps(_mm256_mul_ps(vx5, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
66     const __m256 vn6 = _mm256_round_ps(_mm256_mul_ps(vx6, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
67     const __m256 vn7 = _mm256_round_ps(_mm256_mul_ps(vx7, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
68     const __m256 vn8 = _mm256_round_ps(_mm256_mul_ps(vx8, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
69 
70     // Compute reduced argument t := x - elements * log(2).
71     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
72     __m256 vt0 = _mm256_fmadd_ps(vn0, vminus_ln2_hi, vx0);
73     __m256 vt1 = _mm256_fmadd_ps(vn1, vminus_ln2_hi, vx1);
74     __m256 vt2 = _mm256_fmadd_ps(vn2, vminus_ln2_hi, vx2);
75     __m256 vt3 = _mm256_fmadd_ps(vn3, vminus_ln2_hi, vx3);
76     __m256 vt4 = _mm256_fmadd_ps(vn4, vminus_ln2_hi, vx4);
77     __m256 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_hi, vx5);
78     __m256 vt6 = _mm256_fmadd_ps(vn6, vminus_ln2_hi, vx6);
79     __m256 vt7 = _mm256_fmadd_ps(vn7, vminus_ln2_hi, vx7);
80     __m256 vt8 = _mm256_fmadd_ps(vn8, vminus_ln2_hi, vx8);
81 
82     vt0 = _mm256_fmadd_ps(vn0, vminus_ln2_lo, vt0);
83     vt1 = _mm256_fmadd_ps(vn1, vminus_ln2_lo, vt1);
84     vt2 = _mm256_fmadd_ps(vn2, vminus_ln2_lo, vt2);
85     vt3 = _mm256_fmadd_ps(vn3, vminus_ln2_lo, vt3);
86     vt4 = _mm256_fmadd_ps(vn4, vminus_ln2_lo, vt4);
87     vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_lo, vt5);
88     vt6 = _mm256_fmadd_ps(vn6, vminus_ln2_lo, vt6);
89     vt7 = _mm256_fmadd_ps(vn7, vminus_ln2_lo, vt7);
90     vt8 = _mm256_fmadd_ps(vn8, vminus_ln2_lo, vt8);
91 
92     // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
93     __m256 vp0 = _mm256_fmadd_ps(vc5, vt0, vc4);
94     __m256 vp1 = _mm256_fmadd_ps(vc5, vt1, vc4);
95     __m256 vp2 = _mm256_fmadd_ps(vc5, vt2, vc4);
96     __m256 vp3 = _mm256_fmadd_ps(vc5, vt3, vc4);
97     __m256 vp4 = _mm256_fmadd_ps(vc5, vt4, vc4);
98     __m256 vp5 = _mm256_fmadd_ps(vc5, vt5, vc4);
99     __m256 vp6 = _mm256_fmadd_ps(vc5, vt6, vc4);
100     __m256 vp7 = _mm256_fmadd_ps(vc5, vt7, vc4);
101     __m256 vp8 = _mm256_fmadd_ps(vc5, vt8, vc4);
102 
103     vp0 = _mm256_fmadd_ps(vp0, vt0, vc3);
104     vp1 = _mm256_fmadd_ps(vp1, vt1, vc3);
105     vp2 = _mm256_fmadd_ps(vp2, vt2, vc3);
106     vp3 = _mm256_fmadd_ps(vp3, vt3, vc3);
107     vp4 = _mm256_fmadd_ps(vp4, vt4, vc3);
108     vp5 = _mm256_fmadd_ps(vp5, vt5, vc3);
109     vp6 = _mm256_fmadd_ps(vp6, vt6, vc3);
110     vp7 = _mm256_fmadd_ps(vp7, vt7, vc3);
111     vp8 = _mm256_fmadd_ps(vp8, vt8, vc3);
112 
113     vp0 = _mm256_fmadd_ps(vp0, vt0, vc2);
114     vp1 = _mm256_fmadd_ps(vp1, vt1, vc2);
115     vp2 = _mm256_fmadd_ps(vp2, vt2, vc2);
116     vp3 = _mm256_fmadd_ps(vp3, vt3, vc2);
117     vp4 = _mm256_fmadd_ps(vp4, vt4, vc2);
118     vp5 = _mm256_fmadd_ps(vp5, vt5, vc2);
119     vp6 = _mm256_fmadd_ps(vp6, vt6, vc2);
120     vp7 = _mm256_fmadd_ps(vp7, vt7, vc2);
121     vp8 = _mm256_fmadd_ps(vp8, vt8, vc2);
122 
123     vp0 = _mm256_fmadd_ps(vp0, vt0, vc1);
124     vp1 = _mm256_fmadd_ps(vp1, vt1, vc1);
125     vp2 = _mm256_fmadd_ps(vp2, vt2, vc1);
126     vp3 = _mm256_fmadd_ps(vp3, vt3, vc1);
127     vp4 = _mm256_fmadd_ps(vp4, vt4, vc1);
128     vp5 = _mm256_fmadd_ps(vp5, vt5, vc1);
129     vp6 = _mm256_fmadd_ps(vp6, vt6, vc1);
130     vp7 = _mm256_fmadd_ps(vp7, vt7, vc1);
131     vp8 = _mm256_fmadd_ps(vp8, vt8, vc1);
132 
133     vp0 = _mm256_fmadd_ps(vp0, vt0, vc0);
134     vp1 = _mm256_fmadd_ps(vp1, vt1, vc0);
135     vp2 = _mm256_fmadd_ps(vp2, vt2, vc0);
136     vp3 = _mm256_fmadd_ps(vp3, vt3, vc0);
137     vp4 = _mm256_fmadd_ps(vp4, vt4, vc0);
138     vp5 = _mm256_fmadd_ps(vp5, vt5, vc0);
139     vp6 = _mm256_fmadd_ps(vp6, vt6, vc0);
140     vp7 = _mm256_fmadd_ps(vp7, vt7, vc0);
141     vp8 = _mm256_fmadd_ps(vp8, vt8, vc0);
142 
143     // Accumulate "extended" floating-point numbers in ("mantissa", "exponent") representation where
144     //  - vnX is "exponent"
145     //  - vpX is "mantissa"
146     //
147     // exp2(ae) * av + exp2(be) * bv =
148     //   = exp2(max(ae, be)) * exp2(ae - max(ae, be)) * av + exp2(max(ae, be)) * exp2(be - max(ae, be)) * bv
149     //   = exp2(max_e) * (exp2(ae - max_e) * av + exp2(be - max_e) * bv)
150     //   = exp2(max_e) * (exp2(delta_ae) * av + exp2(delta_be) * bv)
151     //
152     // For computational efficiency we may add several "extended" floating-point numbers at a time.
153     __m256 vmax_e0 = _mm256_max_ps(vacce0, vn0);
154     vmax_e0 = _mm256_max_ps(vmax_e0, vn1);
155     vmax_e0 = _mm256_max_ps(vmax_e0, vn2);
156     vmax_e0 = _mm256_max_ps(vmax_e0, vn3);
157     vmax_e0 = _mm256_max_ps(vmax_e0, vn4);
158     vmax_e0 = _mm256_max_ps(vmax_e0, vn5);
159     vmax_e0 = _mm256_max_ps(vmax_e0, vn6);
160     vmax_e0 = _mm256_max_ps(vmax_e0, vn7);
161     vmax_e0 = _mm256_max_ps(vmax_e0, vn8);
162 
163     // For computational efficiency, replace exp2(delta_e) with 0.0f when delta_e <= -127.0.
164     // This replacement is done in two steps:
165     // 1. Clamp minimum delta_e at -127.0.
166     // 2. Map delta_e to scale factor 0.0 when delta_e == -127.0
167     const __m256 vdelta_acce0 = _mm256_max_ps(_mm256_sub_ps(vacce0, vmax_e0), vmin_exponent);
168     const __m256 vdelta_e0 = _mm256_max_ps(_mm256_sub_ps(vn0, vmax_e0), vmin_exponent);
169     const __m256 vdelta_e1 = _mm256_max_ps(_mm256_sub_ps(vn1, vmax_e0), vmin_exponent);
170     const __m256 vdelta_e2 = _mm256_max_ps(_mm256_sub_ps(vn2, vmax_e0), vmin_exponent);
171     const __m256 vdelta_e3 = _mm256_max_ps(_mm256_sub_ps(vn3, vmax_e0), vmin_exponent);
172     const __m256 vdelta_e4 = _mm256_max_ps(_mm256_sub_ps(vn4, vmax_e0), vmin_exponent);
173     const __m256 vdelta_e5 = _mm256_max_ps(_mm256_sub_ps(vn5, vmax_e0), vmin_exponent);
174     const __m256 vdelta_e6 = _mm256_max_ps(_mm256_sub_ps(vn6, vmax_e0), vmin_exponent);
175     const __m256 vdelta_e7 = _mm256_max_ps(_mm256_sub_ps(vn7, vmax_e0), vmin_exponent);
176     const __m256 vdelta_e8 = _mm256_max_ps(_mm256_sub_ps(vn8, vmax_e0), vmin_exponent);
177 
178     // Convert delta-exponents into scale factors:
179     // - s = exp2(delta_e) when delta_e > -127.0
180     // - s = 0.0 when delta_e <= -127.0
181     //
182     // Note: delta-exponents can not exceed 0.0, thus scale factors can not exceed 1.0.
183     const __m256 vaccs0 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_acce0, vmagic_bias)), 23));
184     const __m256 vs0 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_e0, vmagic_bias)), 23));
185     const __m256 vs1 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_e1, vmagic_bias)), 23));
186     const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_e2, vmagic_bias)), 23));
187     const __m256 vs3 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_e3, vmagic_bias)), 23));
188     const __m256 vs4 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_e4, vmagic_bias)), 23));
189     const __m256 vs5 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_e5, vmagic_bias)), 23));
190     const __m256 vs6 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_e6, vmagic_bias)), 23));
191     const __m256 vs7 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_e7, vmagic_bias)), 23));
192     const __m256 vs8 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_e8, vmagic_bias)), 23));
193 
194     // Update accumulated "mantissa" and "exponent" values
195     vaccv0 = _mm256_mul_ps(vaccv0, vaccs0);
196     vaccv0 = _mm256_fmadd_ps(vp0, vs0, vaccv0);
197     vaccv0 = _mm256_fmadd_ps(vp1, vs1, vaccv0);
198     vaccv0 = _mm256_fmadd_ps(vp2, vs2, vaccv0);
199     vaccv0 = _mm256_fmadd_ps(vp3, vs3, vaccv0);
200     vaccv0 = _mm256_fmadd_ps(vp4, vs4, vaccv0);
201     vaccv0 = _mm256_fmadd_ps(vp5, vs5, vaccv0);
202     vaccv0 = _mm256_fmadd_ps(vp6, vs6, vaccv0);
203     vaccv0 = _mm256_fmadd_ps(vp7, vs7, vaccv0);
204     vaccv0 = _mm256_fmadd_ps(vp8, vs8, vaccv0);
205 
206     vacce0 = vmax_e0;
207   }
208 
209   // Reduce partial sums of "extended" floating-point numbers into a single "extended" SIMD vector of sums.
210   __m256 vaccv = vaccv0;
211   __m256 vacce = vacce0;
212 
213   for (; elements >= 8 * sizeof(float); elements -= 8 * sizeof(float)) {
214     // Load 8 inputs at a time.
215     const __m256 vx = _mm256_loadu_ps(x);
216     x += 8;
217 
218     // Compute reduced argument elements := round(x / log(2)).
219     const __m256 vn = _mm256_round_ps(_mm256_mul_ps(vx, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
220 
221     // Compute reduced argument t := x - elements * log(2).
222     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
223     __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2_hi, vx);
224     vt = _mm256_fmadd_ps(vn, vminus_ln2_lo, vt);
225 
226     // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
227     __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
228     vp = _mm256_fmadd_ps(vp, vt, vc3);
229     vp = _mm256_fmadd_ps(vp, vt, vc2);
230     vp = _mm256_fmadd_ps(vp, vt, vc1);
231     vp = _mm256_fmadd_ps(vp, vt, vc0);
232 
233     // Accumulate "extended" floating-point numbers in ("mantissa", "exponent") representation.
234     const __m256 vmax_e = _mm256_max_ps(vacce, vn);
235 
236     // For computational efficiency, clamp minimum exp2(delta_e) at -127.0. It will be mapped to 0.0 scale factor later.
237     const __m256 vdelta_acce = _mm256_max_ps(_mm256_sub_ps(vacce, vmax_e), vmin_exponent);
238     const __m256 vdelta_e = _mm256_max_ps(_mm256_sub_ps(vn, vmax_e), vmin_exponent);
239 
240     // Convert exponents into scale factors.
241     const __m256 vaccs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_acce, vmagic_bias)), 23));
242     const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_e, vmagic_bias)), 23));
243 
244     // Update accumulated "mantissa" and "exponent" values.
245     vaccv = _mm256_mul_ps(vaccv, vaccs);
246     vaccv = _mm256_fmadd_ps(vp, vs, vaccv);
247 
248     vacce = vmax_e;
249   }
250   if XNN_UNLIKELY(elements != 0) {
251     assert(elements >= 1 * sizeof(float));
252     assert(elements <= 7 * sizeof(float));
253     const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - elements));
254 
255     // Load up to 7 inputs at a time.
256     const __m256 vx = _mm256_maskload_ps(x, vmask);
257 
258     // Compute reduced argument elements := round(x / log(2)).
259     __m256 vn = _mm256_round_ps(_mm256_mul_ps(vx, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
260 
261     // Compute reduced argument t := x - elements * log(2).
262     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
263     __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2_hi, vx);
264     vt = _mm256_fmadd_ps(vn, vminus_ln2_lo, vt);
265 
266     // Correct reduced argument elements for masked out elements.
267     vn = _mm256_blendv_ps(vacce, vn, _mm256_castsi256_ps(vmask));
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     vp = _mm256_fmadd_ps(vp, vt, vc0);
275     vp = _mm256_and_ps(vp, _mm256_castsi256_ps(vmask));
276 
277     // Accumulate "extended" floating-point numbers in ("mantissa", "exponent") representation.
278     const __m256 vmax_e = _mm256_max_ps(vacce, vn);
279 
280     // For computational efficiency, clamp minimum exp2(delta_e) at -127.0. It will be mapped to 0.0 scale factor later.
281     const __m256 vdelta_e = _mm256_max_ps(_mm256_sub_ps(vn, vmax_e), vmin_exponent);
282     const __m256 vdelta_acce = _mm256_max_ps(_mm256_sub_ps(vacce, vmax_e), vmin_exponent);
283 
284     // Convert exponents into scale factors.
285     const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_e, vmagic_bias)), 23));
286     const __m256 vaccs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_acce, vmagic_bias)), 23));
287 
288     // Update accumulated "mantissa" and "exponent" values.
289     vaccv = _mm256_mul_ps(vaccv, vaccs);
290     vaccv = _mm256_fmadd_ps(vp, vs, vaccv);
291 
292     vacce = vmax_e;
293   }
294 
295   // Reduce partial sums of "extended" floating-point numbers into a single "extended" floating-point sum.
296   __m256 vmax_acce = _mm256_max_ps(vacce, _mm256_permute2f128_ps(vacce, vacce, 1));
297   vmax_acce = _mm256_max_ps(vmax_acce, _mm256_shuffle_ps(vmax_acce, vmax_acce, _MM_SHUFFLE(1, 0, 3, 2)));
298   vmax_acce = _mm256_max_ps(vmax_acce, _mm256_shuffle_ps(vmax_acce, vmax_acce, _MM_SHUFFLE(2, 3, 0, 1)));
299   const __m256 vdelta_acce = _mm256_max_ps(_mm256_sub_ps(vacce, vmax_acce), vmin_exponent);
300   const __m256 vaccs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_acce, vmagic_bias)), 23));
301 
302   vaccv = _mm256_mul_ps(vaccv, vaccs);
303   __m128 vaccv_sum = _mm_add_ps(_mm256_castps256_ps128(vaccv), _mm256_extractf128_ps(vaccv, 1));
304   vaccv_sum = _mm_add_ps(vaccv_sum, _mm_movehl_ps(vaccv_sum, vaccv_sum));
305   vaccv_sum = _mm_add_ss(vaccv_sum, _mm_movehdup_ps(vaccv_sum));
306 
307   _mm_store_ss(&sum[0], vaccv_sum);
308   _mm_store_ss(&sum[1], _mm256_castps256_ps128(vmax_acce));
309 
310   _mm256_zeroupper();
311 }
312