xref: /aosp_15_r20/external/XNNPACK/src/f32-vscaleextexp/gen/avx2-p5-x8.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_x8(size_t elements,const float * x,float * y,float scale_value,float scale_exp)20 void xnn_f32_vscaleextexp_ukernel__avx2_p5_x8(
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 >= 8 * sizeof(float); elements -= 8 * sizeof(float)) {
49     // Load 8 (1x8) inputs at a time.
50     const __m256 vx0 = _mm256_loadu_ps(x);
51     x += 8;
52 
53     // Compute reduced argument elements := round(x / log(2)).
54     const __m256 vn0 = _mm256_round_ps(_mm256_mul_ps(vx0, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
55 
56     // Compute reduced argument t := x - elements * log(2).
57     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
58     __m256 vt0 = _mm256_fmadd_ps(vn0, vminus_ln2_hi, vx0);
59 
60     vt0 = _mm256_fmadd_ps(vn0, vminus_ln2_lo, vt0);
61 
62     // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
63     __m256 vp0 = _mm256_fmadd_ps(vc5, vt0, vc4);
64 
65     vp0 = _mm256_fmadd_ps(vp0, vt0, vc3);
66 
67     vp0 = _mm256_fmadd_ps(vp0, vt0, vc2);
68 
69     vp0 = _mm256_fmadd_ps(vp0, vt0, vc1);
70 
71     vp0 = _mm256_fmadd_ps(vp0, vt0, vc0);
72 
73     // Multiply "extended" floating-point numbers in ("mantissa", "exponent") representation where
74     //  - vnX is "exponent"
75     //  - vpX is "mantissa"
76     //
77     // exp2(ae) * av * exp2(be) * bv =
78     //   = exp2(ae + be) * (av * bv)
79     __m256 vf0 = _mm256_mul_ps(vp0, vscalev);
80 
81     __m256 ve0 = _mm256_add_ps(vn0, vscalee);
82 
83     // For computational efficiency, replace exp2(e) with 0.0f when e <= -127.0.
84     // This replacement is done in two steps:
85     // 1. Clamp minimum e at -127.0.
86     // 2. Map e to scale factor 0.0 when e == -127.0
87     ve0 = _mm256_max_ps(ve0, vmin_exponent);
88 
89     // Convert exponents into scale factors:
90     // - s = exp2(e) when e > -127.0
91     // - s = 0.0 when e <= -127.0
92     const __m256 vs0 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve0, vmagic_bias)), 23));
93 
94     // Multiply "mantissa" by the scale factor.
95     vf0 = _mm256_mul_ps(vf0, vs0);
96 
97     // Store 8 (1x8) outputs at a time.
98     _mm256_storeu_ps(y, vf0);
99     y += 8;
100   }
101 
102   for (; elements >= 8 * sizeof(float); elements -= 8 * sizeof(float)) {
103     // Load 8 inputs at a time.
104     const __m256 vx = _mm256_loadu_ps(x);
105     x += 8;
106 
107     // Compute reduced argument elements := round(x / log(2)).
108     const __m256 vn = _mm256_round_ps(_mm256_mul_ps(vx, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
109 
110     // Compute reduced argument t := x - elements * log(2).
111     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
112     __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2_hi, vx);
113     vt = _mm256_fmadd_ps(vn, vminus_ln2_lo, vt);
114 
115     // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
116     __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
117     vp = _mm256_fmadd_ps(vp, vt, vc3);
118     vp = _mm256_fmadd_ps(vp, vt, vc2);
119     vp = _mm256_fmadd_ps(vp, vt, vc1);
120     vp = _mm256_fmadd_ps(vp, vt, vc0);
121 
122     // Multiply "extended" floating-point numbers in ("mantissa", "exponent") representation.
123     __m256 vf = _mm256_mul_ps(vp, vscalev);
124     __m256 ve = _mm256_add_ps(vn, vscalee);
125 
126     // For computational efficiency, replace exp2(e) with 0.0f when e <= -127.0.
127     ve = _mm256_max_ps(ve, vmin_exponent);
128 
129     // Convert exponents into scale factors.
130     const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve, vmagic_bias)), 23));
131 
132     // Multiply "mantissa" by the scale factor.
133     vf = _mm256_mul_ps(vf, vs);
134 
135     // Store 8 results at a time.
136     _mm256_storeu_ps(y, vf);
137     y += 8;
138   }
139   if XNN_UNLIKELY(elements != 0) {
140     assert(elements >= 1 * sizeof(float));
141     assert(elements <= 7 * sizeof(float));
142     const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - elements));
143 
144     // Load up to 7 inputs at a time.
145     const __m256 vx = _mm256_maskload_ps(x, vmask);
146 
147     // Compute reduced argument elements := round(x / log(2)).
148     const __m256 vn = _mm256_round_ps(_mm256_mul_ps(vx, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
149 
150     // Compute reduced argument t := x - elements * log(2).
151     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
152     __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2_hi, vx);
153     vt = _mm256_fmadd_ps(vn, vminus_ln2_lo, vt);
154 
155     // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
156     __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
157     vp = _mm256_fmadd_ps(vp, vt, vc3);
158     vp = _mm256_fmadd_ps(vp, vt, vc2);
159     vp = _mm256_fmadd_ps(vp, vt, vc1);
160     vp = _mm256_fmadd_ps(vp, vt, vc0);
161 
162     // Multiply "extended" floating-point numbers in ("mantissa", "exponent") representation.
163     __m256 vf = _mm256_mul_ps(vp, vscalev);
164     __m256 ve = _mm256_add_ps(vn, vscalee);
165 
166     // For computational efficiency, replace exp2(e) with 0.0f when e <= -127.0.
167     ve = _mm256_max_ps(ve, vmin_exponent);
168 
169     // Convert exponents into scale factors.
170     const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve, vmagic_bias)), 23));
171 
172     // Multiply "mantissa" by the scale factor.
173     vf = _mm256_mul_ps(vf, vs);
174 
175     // Store up to 7 inputs at a time.
176     _mm256_maskstore_ps(y, vmask, vf);
177   }
178   _mm256_zeroupper();
179 }
180