xref: /aosp_15_r20/external/XNNPACK/src/f32-vscaleextexp/gen/avx2-p5-x16.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_x16(size_t elements,const float * x,float * y,float scale_value,float scale_exp)20 void xnn_f32_vscaleextexp_ukernel__avx2_p5_x16(
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 >= 16 * sizeof(float); elements -= 16 * sizeof(float)) {
49     // Load 16 (2x8) inputs at a time.
50     const __m256 vx0 = _mm256_loadu_ps(x);
51     const __m256 vx1 = _mm256_loadu_ps(x + 8);
52     x += 16;
53 
54     // Compute reduced argument elements := round(x / log(2)).
55     const __m256 vn0 = _mm256_round_ps(_mm256_mul_ps(vx0, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
56     const __m256 vn1 = _mm256_round_ps(_mm256_mul_ps(vx1, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
57 
58     // Compute reduced argument t := x - elements * log(2).
59     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
60     __m256 vt0 = _mm256_fmadd_ps(vn0, vminus_ln2_hi, vx0);
61     __m256 vt1 = _mm256_fmadd_ps(vn1, vminus_ln2_hi, vx1);
62 
63     vt0 = _mm256_fmadd_ps(vn0, vminus_ln2_lo, vt0);
64     vt1 = _mm256_fmadd_ps(vn1, vminus_ln2_lo, vt1);
65 
66     // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
67     __m256 vp0 = _mm256_fmadd_ps(vc5, vt0, vc4);
68     __m256 vp1 = _mm256_fmadd_ps(vc5, vt1, vc4);
69 
70     vp0 = _mm256_fmadd_ps(vp0, vt0, vc3);
71     vp1 = _mm256_fmadd_ps(vp1, vt1, vc3);
72 
73     vp0 = _mm256_fmadd_ps(vp0, vt0, vc2);
74     vp1 = _mm256_fmadd_ps(vp1, vt1, vc2);
75 
76     vp0 = _mm256_fmadd_ps(vp0, vt0, vc1);
77     vp1 = _mm256_fmadd_ps(vp1, vt1, vc1);
78 
79     vp0 = _mm256_fmadd_ps(vp0, vt0, vc0);
80     vp1 = _mm256_fmadd_ps(vp1, vt1, vc0);
81 
82     // Multiply "extended" floating-point numbers in ("mantissa", "exponent") representation where
83     //  - vnX is "exponent"
84     //  - vpX is "mantissa"
85     //
86     // exp2(ae) * av * exp2(be) * bv =
87     //   = exp2(ae + be) * (av * bv)
88     __m256 vf0 = _mm256_mul_ps(vp0, vscalev);
89     __m256 vf1 = _mm256_mul_ps(vp1, vscalev);
90 
91     __m256 ve0 = _mm256_add_ps(vn0, vscalee);
92     __m256 ve1 = _mm256_add_ps(vn1, vscalee);
93 
94     // For computational efficiency, replace exp2(e) with 0.0f when e <= -127.0.
95     // This replacement is done in two steps:
96     // 1. Clamp minimum e at -127.0.
97     // 2. Map e to scale factor 0.0 when e == -127.0
98     ve0 = _mm256_max_ps(ve0, vmin_exponent);
99     ve1 = _mm256_max_ps(ve1, vmin_exponent);
100 
101     // Convert exponents into scale factors:
102     // - s = exp2(e) when e > -127.0
103     // - s = 0.0 when e <= -127.0
104     const __m256 vs0 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve0, vmagic_bias)), 23));
105     const __m256 vs1 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve1, vmagic_bias)), 23));
106 
107     // Multiply "mantissa" by the scale factor.
108     vf0 = _mm256_mul_ps(vf0, vs0);
109     vf1 = _mm256_mul_ps(vf1, vs1);
110 
111     // Store 16 (2x8) outputs at a time.
112     _mm256_storeu_ps(y, vf0);
113     _mm256_storeu_ps(y + 8, vf1);
114     y += 16;
115   }
116 
117   for (; elements >= 8 * sizeof(float); elements -= 8 * sizeof(float)) {
118     // Load 8 inputs at a time.
119     const __m256 vx = _mm256_loadu_ps(x);
120     x += 8;
121 
122     // Compute reduced argument elements := round(x / log(2)).
123     const __m256 vn = _mm256_round_ps(_mm256_mul_ps(vx, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
124 
125     // Compute reduced argument t := x - elements * log(2).
126     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
127     __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2_hi, vx);
128     vt = _mm256_fmadd_ps(vn, vminus_ln2_lo, vt);
129 
130     // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
131     __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
132     vp = _mm256_fmadd_ps(vp, vt, vc3);
133     vp = _mm256_fmadd_ps(vp, vt, vc2);
134     vp = _mm256_fmadd_ps(vp, vt, vc1);
135     vp = _mm256_fmadd_ps(vp, vt, vc0);
136 
137     // Multiply "extended" floating-point numbers in ("mantissa", "exponent") representation.
138     __m256 vf = _mm256_mul_ps(vp, vscalev);
139     __m256 ve = _mm256_add_ps(vn, vscalee);
140 
141     // For computational efficiency, replace exp2(e) with 0.0f when e <= -127.0.
142     ve = _mm256_max_ps(ve, vmin_exponent);
143 
144     // Convert exponents into scale factors.
145     const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve, vmagic_bias)), 23));
146 
147     // Multiply "mantissa" by the scale factor.
148     vf = _mm256_mul_ps(vf, vs);
149 
150     // Store 8 results at a time.
151     _mm256_storeu_ps(y, vf);
152     y += 8;
153   }
154   if XNN_UNLIKELY(elements != 0) {
155     assert(elements >= 1 * sizeof(float));
156     assert(elements <= 7 * sizeof(float));
157     const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - elements));
158 
159     // Load up to 7 inputs at a time.
160     const __m256 vx = _mm256_maskload_ps(x, vmask);
161 
162     // Compute reduced argument elements := round(x / log(2)).
163     const __m256 vn = _mm256_round_ps(_mm256_mul_ps(vx, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
164 
165     // Compute reduced argument t := x - elements * log(2).
166     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
167     __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2_hi, vx);
168     vt = _mm256_fmadd_ps(vn, vminus_ln2_lo, vt);
169 
170     // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
171     __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
172     vp = _mm256_fmadd_ps(vp, vt, vc3);
173     vp = _mm256_fmadd_ps(vp, vt, vc2);
174     vp = _mm256_fmadd_ps(vp, vt, vc1);
175     vp = _mm256_fmadd_ps(vp, vt, vc0);
176 
177     // Multiply "extended" floating-point numbers in ("mantissa", "exponent") representation.
178     __m256 vf = _mm256_mul_ps(vp, vscalev);
179     __m256 ve = _mm256_add_ps(vn, vscalee);
180 
181     // For computational efficiency, replace exp2(e) with 0.0f when e <= -127.0.
182     ve = _mm256_max_ps(ve, vmin_exponent);
183 
184     // Convert exponents into scale factors.
185     const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve, vmagic_bias)), 23));
186 
187     // Multiply "mantissa" by the scale factor.
188     vf = _mm256_mul_ps(vf, vs);
189 
190     // Store up to 7 inputs at a time.
191     _mm256_maskstore_ps(y, vmask, vf);
192   }
193   _mm256_zeroupper();
194 }
195