xref: /aosp_15_r20/external/XNNPACK/src/f32-vscaleextexp/gen/avx2-p5-x64.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-vscaleextexp/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/vscaleextexp.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_vscaleextexp_ukernel__avx2_p5_x64(size_t elements,const float * x,float * y,float scale_value,float scale_exp)20 void xnn_f32_vscaleextexp_ukernel__avx2_p5_x64(
21     size_t elements,
22     const float* x,
23     float* y,
24     float scale_value,
25     float scale_exp)
26 {
27   assert(elements % sizeof(float) == 0);
28 
29   const __m256 vlog2e = _mm256_set1_ps(0x1.715476p+0f);
30   const __m256 vminus_ln2_hi = _mm256_set1_ps(-0x1.62E43p-1f);
31   const __m256 vminus_ln2_lo = _mm256_set1_ps(0x1.05C61p-29f);
32 
33   // The smallest elements such that 2**elements is considered non-negligible.
34   // For smaller elements, 2**elements is replaced with zero.
35   const __m256 vmin_exponent = _mm256_set1_ps(-127.0f);
36   const __m256 vmagic_bias = _mm256_set1_ps(0x1.8000FEp23f);
37 
38   const __m256 vc0 = _mm256_set1_ps(1.0f);
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 vscalev = _mm256_set1_ps(scale_value);
46   const __m256 vscalee = _mm256_set1_ps(scale_exp);
47 
48   for (; elements >= 64 * sizeof(float); elements -= 64 * sizeof(float)) {
49     // Load 64 (8x8) inputs at a time.
50     const __m256 vx0 = _mm256_loadu_ps(x);
51     const __m256 vx1 = _mm256_loadu_ps(x + 8);
52     const __m256 vx2 = _mm256_loadu_ps(x + 16);
53     const __m256 vx3 = _mm256_loadu_ps(x + 24);
54     const __m256 vx4 = _mm256_loadu_ps(x + 32);
55     const __m256 vx5 = _mm256_loadu_ps(x + 40);
56     const __m256 vx6 = _mm256_loadu_ps(x + 48);
57     const __m256 vx7 = _mm256_loadu_ps(x + 56);
58     x += 64;
59 
60     // Compute reduced argument elements := round(x / log(2)).
61     const __m256 vn0 = _mm256_round_ps(_mm256_mul_ps(vx0, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
62     const __m256 vn1 = _mm256_round_ps(_mm256_mul_ps(vx1, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
63     const __m256 vn2 = _mm256_round_ps(_mm256_mul_ps(vx2, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
64     const __m256 vn3 = _mm256_round_ps(_mm256_mul_ps(vx3, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
65     const __m256 vn4 = _mm256_round_ps(_mm256_mul_ps(vx4, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
66     const __m256 vn5 = _mm256_round_ps(_mm256_mul_ps(vx5, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
67     const __m256 vn6 = _mm256_round_ps(_mm256_mul_ps(vx6, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
68     const __m256 vn7 = _mm256_round_ps(_mm256_mul_ps(vx7, 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 
81     vt0 = _mm256_fmadd_ps(vn0, vminus_ln2_lo, vt0);
82     vt1 = _mm256_fmadd_ps(vn1, vminus_ln2_lo, vt1);
83     vt2 = _mm256_fmadd_ps(vn2, vminus_ln2_lo, vt2);
84     vt3 = _mm256_fmadd_ps(vn3, vminus_ln2_lo, vt3);
85     vt4 = _mm256_fmadd_ps(vn4, vminus_ln2_lo, vt4);
86     vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_lo, vt5);
87     vt6 = _mm256_fmadd_ps(vn6, vminus_ln2_lo, vt6);
88     vt7 = _mm256_fmadd_ps(vn7, vminus_ln2_lo, vt7);
89 
90     // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
91     __m256 vp0 = _mm256_fmadd_ps(vc5, vt0, vc4);
92     __m256 vp1 = _mm256_fmadd_ps(vc5, vt1, vc4);
93     __m256 vp2 = _mm256_fmadd_ps(vc5, vt2, vc4);
94     __m256 vp3 = _mm256_fmadd_ps(vc5, vt3, vc4);
95     __m256 vp4 = _mm256_fmadd_ps(vc5, vt4, vc4);
96     __m256 vp5 = _mm256_fmadd_ps(vc5, vt5, vc4);
97     __m256 vp6 = _mm256_fmadd_ps(vc5, vt6, vc4);
98     __m256 vp7 = _mm256_fmadd_ps(vc5, vt7, vc4);
99 
100     vp0 = _mm256_fmadd_ps(vp0, vt0, vc3);
101     vp1 = _mm256_fmadd_ps(vp1, vt1, vc3);
102     vp2 = _mm256_fmadd_ps(vp2, vt2, vc3);
103     vp3 = _mm256_fmadd_ps(vp3, vt3, vc3);
104     vp4 = _mm256_fmadd_ps(vp4, vt4, vc3);
105     vp5 = _mm256_fmadd_ps(vp5, vt5, vc3);
106     vp6 = _mm256_fmadd_ps(vp6, vt6, vc3);
107     vp7 = _mm256_fmadd_ps(vp7, vt7, vc3);
108 
109     vp0 = _mm256_fmadd_ps(vp0, vt0, vc2);
110     vp1 = _mm256_fmadd_ps(vp1, vt1, vc2);
111     vp2 = _mm256_fmadd_ps(vp2, vt2, vc2);
112     vp3 = _mm256_fmadd_ps(vp3, vt3, vc2);
113     vp4 = _mm256_fmadd_ps(vp4, vt4, vc2);
114     vp5 = _mm256_fmadd_ps(vp5, vt5, vc2);
115     vp6 = _mm256_fmadd_ps(vp6, vt6, vc2);
116     vp7 = _mm256_fmadd_ps(vp7, vt7, vc2);
117 
118     vp0 = _mm256_fmadd_ps(vp0, vt0, vc1);
119     vp1 = _mm256_fmadd_ps(vp1, vt1, vc1);
120     vp2 = _mm256_fmadd_ps(vp2, vt2, vc1);
121     vp3 = _mm256_fmadd_ps(vp3, vt3, vc1);
122     vp4 = _mm256_fmadd_ps(vp4, vt4, vc1);
123     vp5 = _mm256_fmadd_ps(vp5, vt5, vc1);
124     vp6 = _mm256_fmadd_ps(vp6, vt6, vc1);
125     vp7 = _mm256_fmadd_ps(vp7, vt7, vc1);
126 
127     vp0 = _mm256_fmadd_ps(vp0, vt0, vc0);
128     vp1 = _mm256_fmadd_ps(vp1, vt1, vc0);
129     vp2 = _mm256_fmadd_ps(vp2, vt2, vc0);
130     vp3 = _mm256_fmadd_ps(vp3, vt3, vc0);
131     vp4 = _mm256_fmadd_ps(vp4, vt4, vc0);
132     vp5 = _mm256_fmadd_ps(vp5, vt5, vc0);
133     vp6 = _mm256_fmadd_ps(vp6, vt6, vc0);
134     vp7 = _mm256_fmadd_ps(vp7, vt7, vc0);
135 
136     // Multiply "extended" floating-point numbers in ("mantissa", "exponent") representation where
137     //  - vnX is "exponent"
138     //  - vpX is "mantissa"
139     //
140     // exp2(ae) * av * exp2(be) * bv =
141     //   = exp2(ae + be) * (av * bv)
142     __m256 vf0 = _mm256_mul_ps(vp0, vscalev);
143     __m256 vf1 = _mm256_mul_ps(vp1, vscalev);
144     __m256 vf2 = _mm256_mul_ps(vp2, vscalev);
145     __m256 vf3 = _mm256_mul_ps(vp3, vscalev);
146     __m256 vf4 = _mm256_mul_ps(vp4, vscalev);
147     __m256 vf5 = _mm256_mul_ps(vp5, vscalev);
148     __m256 vf6 = _mm256_mul_ps(vp6, vscalev);
149     __m256 vf7 = _mm256_mul_ps(vp7, vscalev);
150 
151     __m256 ve0 = _mm256_add_ps(vn0, vscalee);
152     __m256 ve1 = _mm256_add_ps(vn1, vscalee);
153     __m256 ve2 = _mm256_add_ps(vn2, vscalee);
154     __m256 ve3 = _mm256_add_ps(vn3, vscalee);
155     __m256 ve4 = _mm256_add_ps(vn4, vscalee);
156     __m256 ve5 = _mm256_add_ps(vn5, vscalee);
157     __m256 ve6 = _mm256_add_ps(vn6, vscalee);
158     __m256 ve7 = _mm256_add_ps(vn7, vscalee);
159 
160     // For computational efficiency, replace exp2(e) with 0.0f when e <= -127.0.
161     // This replacement is done in two steps:
162     // 1. Clamp minimum e at -127.0.
163     // 2. Map e to scale factor 0.0 when e == -127.0
164     ve0 = _mm256_max_ps(ve0, vmin_exponent);
165     ve1 = _mm256_max_ps(ve1, vmin_exponent);
166     ve2 = _mm256_max_ps(ve2, vmin_exponent);
167     ve3 = _mm256_max_ps(ve3, vmin_exponent);
168     ve4 = _mm256_max_ps(ve4, vmin_exponent);
169     ve5 = _mm256_max_ps(ve5, vmin_exponent);
170     ve6 = _mm256_max_ps(ve6, vmin_exponent);
171     ve7 = _mm256_max_ps(ve7, vmin_exponent);
172 
173     // Convert exponents into scale factors:
174     // - s = exp2(e) when e > -127.0
175     // - s = 0.0 when e <= -127.0
176     const __m256 vs0 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve0, vmagic_bias)), 23));
177     const __m256 vs1 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve1, vmagic_bias)), 23));
178     const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve2, vmagic_bias)), 23));
179     const __m256 vs3 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve3, vmagic_bias)), 23));
180     const __m256 vs4 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve4, vmagic_bias)), 23));
181     const __m256 vs5 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve5, vmagic_bias)), 23));
182     const __m256 vs6 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve6, vmagic_bias)), 23));
183     const __m256 vs7 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve7, vmagic_bias)), 23));
184 
185     // Multiply "mantissa" by the scale factor.
186     vf0 = _mm256_mul_ps(vf0, vs0);
187     vf1 = _mm256_mul_ps(vf1, vs1);
188     vf2 = _mm256_mul_ps(vf2, vs2);
189     vf3 = _mm256_mul_ps(vf3, vs3);
190     vf4 = _mm256_mul_ps(vf4, vs4);
191     vf5 = _mm256_mul_ps(vf5, vs5);
192     vf6 = _mm256_mul_ps(vf6, vs6);
193     vf7 = _mm256_mul_ps(vf7, vs7);
194 
195     // Store 64 (8x8) outputs at a time.
196     _mm256_storeu_ps(y, vf0);
197     _mm256_storeu_ps(y + 8, vf1);
198     _mm256_storeu_ps(y + 16, vf2);
199     _mm256_storeu_ps(y + 24, vf3);
200     _mm256_storeu_ps(y + 32, vf4);
201     _mm256_storeu_ps(y + 40, vf5);
202     _mm256_storeu_ps(y + 48, vf6);
203     _mm256_storeu_ps(y + 56, vf7);
204     y += 64;
205   }
206 
207   for (; elements >= 8 * sizeof(float); elements -= 8 * sizeof(float)) {
208     // Load 8 inputs at a time.
209     const __m256 vx = _mm256_loadu_ps(x);
210     x += 8;
211 
212     // Compute reduced argument elements := round(x / log(2)).
213     const __m256 vn = _mm256_round_ps(_mm256_mul_ps(vx, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
214 
215     // Compute reduced argument t := x - elements * log(2).
216     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
217     __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2_hi, vx);
218     vt = _mm256_fmadd_ps(vn, vminus_ln2_lo, vt);
219 
220     // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
221     __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
222     vp = _mm256_fmadd_ps(vp, vt, vc3);
223     vp = _mm256_fmadd_ps(vp, vt, vc2);
224     vp = _mm256_fmadd_ps(vp, vt, vc1);
225     vp = _mm256_fmadd_ps(vp, vt, vc0);
226 
227     // Multiply "extended" floating-point numbers in ("mantissa", "exponent") representation.
228     __m256 vf = _mm256_mul_ps(vp, vscalev);
229     __m256 ve = _mm256_add_ps(vn, vscalee);
230 
231     // For computational efficiency, replace exp2(e) with 0.0f when e <= -127.0.
232     ve = _mm256_max_ps(ve, vmin_exponent);
233 
234     // Convert exponents into scale factors.
235     const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve, vmagic_bias)), 23));
236 
237     // Multiply "mantissa" by the scale factor.
238     vf = _mm256_mul_ps(vf, vs);
239 
240     // Store 8 results at a time.
241     _mm256_storeu_ps(y, vf);
242     y += 8;
243   }
244   if XNN_UNLIKELY(elements != 0) {
245     assert(elements >= 1 * sizeof(float));
246     assert(elements <= 7 * sizeof(float));
247     const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - elements));
248 
249     // Load up to 7 inputs at a time.
250     const __m256 vx = _mm256_maskload_ps(x, vmask);
251 
252     // Compute reduced argument elements := round(x / log(2)).
253     const __m256 vn = _mm256_round_ps(_mm256_mul_ps(vx, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
254 
255     // Compute reduced argument t := x - elements * log(2).
256     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
257     __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2_hi, vx);
258     vt = _mm256_fmadd_ps(vn, vminus_ln2_lo, vt);
259 
260     // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
261     __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
262     vp = _mm256_fmadd_ps(vp, vt, vc3);
263     vp = _mm256_fmadd_ps(vp, vt, vc2);
264     vp = _mm256_fmadd_ps(vp, vt, vc1);
265     vp = _mm256_fmadd_ps(vp, vt, vc0);
266 
267     // Multiply "extended" floating-point numbers in ("mantissa", "exponent") representation.
268     __m256 vf = _mm256_mul_ps(vp, vscalev);
269     __m256 ve = _mm256_add_ps(vn, vscalee);
270 
271     // For computational efficiency, replace exp2(e) with 0.0f when e <= -127.0.
272     ve = _mm256_max_ps(ve, vmin_exponent);
273 
274     // Convert exponents into scale factors.
275     const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve, vmagic_bias)), 23));
276 
277     // Multiply "mantissa" by the scale factor.
278     vf = _mm256_mul_ps(vf, vs);
279 
280     // Store up to 7 inputs at a time.
281     _mm256_maskstore_ps(y, vmask, vf);
282   }
283   _mm256_zeroupper();
284 }
285