xref: /aosp_15_r20/external/XNNPACK/src/f32-vscaleextexp/gen/avx2-p5-x48.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_x48(size_t elements,const float * x,float * y,float scale_value,float scale_exp)20 void xnn_f32_vscaleextexp_ukernel__avx2_p5_x48(
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 >= 48 * sizeof(float); elements -= 48 * sizeof(float)) {
49     // Load 48 (6x8) 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     x += 48;
57 
58     // Compute reduced argument elements := round(x / log(2)).
59     const __m256 vn0 = _mm256_round_ps(_mm256_mul_ps(vx0, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
60     const __m256 vn1 = _mm256_round_ps(_mm256_mul_ps(vx1, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
61     const __m256 vn2 = _mm256_round_ps(_mm256_mul_ps(vx2, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
62     const __m256 vn3 = _mm256_round_ps(_mm256_mul_ps(vx3, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
63     const __m256 vn4 = _mm256_round_ps(_mm256_mul_ps(vx4, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
64     const __m256 vn5 = _mm256_round_ps(_mm256_mul_ps(vx5, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
65 
66     // Compute reduced argument t := x - elements * log(2).
67     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
68     __m256 vt0 = _mm256_fmadd_ps(vn0, vminus_ln2_hi, vx0);
69     __m256 vt1 = _mm256_fmadd_ps(vn1, vminus_ln2_hi, vx1);
70     __m256 vt2 = _mm256_fmadd_ps(vn2, vminus_ln2_hi, vx2);
71     __m256 vt3 = _mm256_fmadd_ps(vn3, vminus_ln2_hi, vx3);
72     __m256 vt4 = _mm256_fmadd_ps(vn4, vminus_ln2_hi, vx4);
73     __m256 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_hi, vx5);
74 
75     vt0 = _mm256_fmadd_ps(vn0, vminus_ln2_lo, vt0);
76     vt1 = _mm256_fmadd_ps(vn1, vminus_ln2_lo, vt1);
77     vt2 = _mm256_fmadd_ps(vn2, vminus_ln2_lo, vt2);
78     vt3 = _mm256_fmadd_ps(vn3, vminus_ln2_lo, vt3);
79     vt4 = _mm256_fmadd_ps(vn4, vminus_ln2_lo, vt4);
80     vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_lo, vt5);
81 
82     // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
83     __m256 vp0 = _mm256_fmadd_ps(vc5, vt0, vc4);
84     __m256 vp1 = _mm256_fmadd_ps(vc5, vt1, vc4);
85     __m256 vp2 = _mm256_fmadd_ps(vc5, vt2, vc4);
86     __m256 vp3 = _mm256_fmadd_ps(vc5, vt3, vc4);
87     __m256 vp4 = _mm256_fmadd_ps(vc5, vt4, vc4);
88     __m256 vp5 = _mm256_fmadd_ps(vc5, vt5, vc4);
89 
90     vp0 = _mm256_fmadd_ps(vp0, vt0, vc3);
91     vp1 = _mm256_fmadd_ps(vp1, vt1, vc3);
92     vp2 = _mm256_fmadd_ps(vp2, vt2, vc3);
93     vp3 = _mm256_fmadd_ps(vp3, vt3, vc3);
94     vp4 = _mm256_fmadd_ps(vp4, vt4, vc3);
95     vp5 = _mm256_fmadd_ps(vp5, vt5, vc3);
96 
97     vp0 = _mm256_fmadd_ps(vp0, vt0, vc2);
98     vp1 = _mm256_fmadd_ps(vp1, vt1, vc2);
99     vp2 = _mm256_fmadd_ps(vp2, vt2, vc2);
100     vp3 = _mm256_fmadd_ps(vp3, vt3, vc2);
101     vp4 = _mm256_fmadd_ps(vp4, vt4, vc2);
102     vp5 = _mm256_fmadd_ps(vp5, vt5, vc2);
103 
104     vp0 = _mm256_fmadd_ps(vp0, vt0, vc1);
105     vp1 = _mm256_fmadd_ps(vp1, vt1, vc1);
106     vp2 = _mm256_fmadd_ps(vp2, vt2, vc1);
107     vp3 = _mm256_fmadd_ps(vp3, vt3, vc1);
108     vp4 = _mm256_fmadd_ps(vp4, vt4, vc1);
109     vp5 = _mm256_fmadd_ps(vp5, vt5, vc1);
110 
111     vp0 = _mm256_fmadd_ps(vp0, vt0, vc0);
112     vp1 = _mm256_fmadd_ps(vp1, vt1, vc0);
113     vp2 = _mm256_fmadd_ps(vp2, vt2, vc0);
114     vp3 = _mm256_fmadd_ps(vp3, vt3, vc0);
115     vp4 = _mm256_fmadd_ps(vp4, vt4, vc0);
116     vp5 = _mm256_fmadd_ps(vp5, vt5, vc0);
117 
118     // Multiply "extended" floating-point numbers in ("mantissa", "exponent") representation where
119     //  - vnX is "exponent"
120     //  - vpX is "mantissa"
121     //
122     // exp2(ae) * av * exp2(be) * bv =
123     //   = exp2(ae + be) * (av * bv)
124     __m256 vf0 = _mm256_mul_ps(vp0, vscalev);
125     __m256 vf1 = _mm256_mul_ps(vp1, vscalev);
126     __m256 vf2 = _mm256_mul_ps(vp2, vscalev);
127     __m256 vf3 = _mm256_mul_ps(vp3, vscalev);
128     __m256 vf4 = _mm256_mul_ps(vp4, vscalev);
129     __m256 vf5 = _mm256_mul_ps(vp5, vscalev);
130 
131     __m256 ve0 = _mm256_add_ps(vn0, vscalee);
132     __m256 ve1 = _mm256_add_ps(vn1, vscalee);
133     __m256 ve2 = _mm256_add_ps(vn2, vscalee);
134     __m256 ve3 = _mm256_add_ps(vn3, vscalee);
135     __m256 ve4 = _mm256_add_ps(vn4, vscalee);
136     __m256 ve5 = _mm256_add_ps(vn5, vscalee);
137 
138     // For computational efficiency, replace exp2(e) with 0.0f when e <= -127.0.
139     // This replacement is done in two steps:
140     // 1. Clamp minimum e at -127.0.
141     // 2. Map e to scale factor 0.0 when e == -127.0
142     ve0 = _mm256_max_ps(ve0, vmin_exponent);
143     ve1 = _mm256_max_ps(ve1, vmin_exponent);
144     ve2 = _mm256_max_ps(ve2, vmin_exponent);
145     ve3 = _mm256_max_ps(ve3, vmin_exponent);
146     ve4 = _mm256_max_ps(ve4, vmin_exponent);
147     ve5 = _mm256_max_ps(ve5, vmin_exponent);
148 
149     // Convert exponents into scale factors:
150     // - s = exp2(e) when e > -127.0
151     // - s = 0.0 when e <= -127.0
152     const __m256 vs0 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve0, vmagic_bias)), 23));
153     const __m256 vs1 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve1, vmagic_bias)), 23));
154     const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve2, vmagic_bias)), 23));
155     const __m256 vs3 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve3, vmagic_bias)), 23));
156     const __m256 vs4 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve4, vmagic_bias)), 23));
157     const __m256 vs5 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve5, vmagic_bias)), 23));
158 
159     // Multiply "mantissa" by the scale factor.
160     vf0 = _mm256_mul_ps(vf0, vs0);
161     vf1 = _mm256_mul_ps(vf1, vs1);
162     vf2 = _mm256_mul_ps(vf2, vs2);
163     vf3 = _mm256_mul_ps(vf3, vs3);
164     vf4 = _mm256_mul_ps(vf4, vs4);
165     vf5 = _mm256_mul_ps(vf5, vs5);
166 
167     // Store 48 (6x8) outputs at a time.
168     _mm256_storeu_ps(y, vf0);
169     _mm256_storeu_ps(y + 8, vf1);
170     _mm256_storeu_ps(y + 16, vf2);
171     _mm256_storeu_ps(y + 24, vf3);
172     _mm256_storeu_ps(y + 32, vf4);
173     _mm256_storeu_ps(y + 40, vf5);
174     y += 48;
175   }
176 
177   for (; elements >= 8 * sizeof(float); elements -= 8 * sizeof(float)) {
178     // Load 8 inputs at a time.
179     const __m256 vx = _mm256_loadu_ps(x);
180     x += 8;
181 
182     // Compute reduced argument elements := round(x / log(2)).
183     const __m256 vn = _mm256_round_ps(_mm256_mul_ps(vx, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
184 
185     // Compute reduced argument t := x - elements * log(2).
186     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
187     __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2_hi, vx);
188     vt = _mm256_fmadd_ps(vn, vminus_ln2_lo, vt);
189 
190     // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
191     __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
192     vp = _mm256_fmadd_ps(vp, vt, vc3);
193     vp = _mm256_fmadd_ps(vp, vt, vc2);
194     vp = _mm256_fmadd_ps(vp, vt, vc1);
195     vp = _mm256_fmadd_ps(vp, vt, vc0);
196 
197     // Multiply "extended" floating-point numbers in ("mantissa", "exponent") representation.
198     __m256 vf = _mm256_mul_ps(vp, vscalev);
199     __m256 ve = _mm256_add_ps(vn, vscalee);
200 
201     // For computational efficiency, replace exp2(e) with 0.0f when e <= -127.0.
202     ve = _mm256_max_ps(ve, vmin_exponent);
203 
204     // Convert exponents into scale factors.
205     const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve, vmagic_bias)), 23));
206 
207     // Multiply "mantissa" by the scale factor.
208     vf = _mm256_mul_ps(vf, vs);
209 
210     // Store 8 results at a time.
211     _mm256_storeu_ps(y, vf);
212     y += 8;
213   }
214   if XNN_UNLIKELY(elements != 0) {
215     assert(elements >= 1 * sizeof(float));
216     assert(elements <= 7 * sizeof(float));
217     const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - elements));
218 
219     // Load up to 7 inputs at a time.
220     const __m256 vx = _mm256_maskload_ps(x, vmask);
221 
222     // Compute reduced argument elements := round(x / log(2)).
223     const __m256 vn = _mm256_round_ps(_mm256_mul_ps(vx, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
224 
225     // Compute reduced argument t := x - elements * log(2).
226     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
227     __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2_hi, vx);
228     vt = _mm256_fmadd_ps(vn, vminus_ln2_lo, vt);
229 
230     // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
231     __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
232     vp = _mm256_fmadd_ps(vp, vt, vc3);
233     vp = _mm256_fmadd_ps(vp, vt, vc2);
234     vp = _mm256_fmadd_ps(vp, vt, vc1);
235     vp = _mm256_fmadd_ps(vp, vt, vc0);
236 
237     // Multiply "extended" floating-point numbers in ("mantissa", "exponent") representation.
238     __m256 vf = _mm256_mul_ps(vp, vscalev);
239     __m256 ve = _mm256_add_ps(vn, vscalee);
240 
241     // For computational efficiency, replace exp2(e) with 0.0f when e <= -127.0.
242     ve = _mm256_max_ps(ve, vmin_exponent);
243 
244     // Convert exponents into scale factors.
245     const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve, vmagic_bias)), 23));
246 
247     // Multiply "mantissa" by the scale factor.
248     vf = _mm256_mul_ps(vf, vs);
249 
250     // Store up to 7 inputs at a time.
251     _mm256_maskstore_ps(y, vmask, vf);
252   }
253   _mm256_zeroupper();
254 }
255