xref: /aosp_15_r20/external/XNNPACK/src/f32-vscaleextexp/gen/avx2-p5-x24.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_x24(size_t elements,const float * x,float * y,float scale_value,float scale_exp)20 void xnn_f32_vscaleextexp_ukernel__avx2_p5_x24(
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 >= 24 * sizeof(float); elements -= 24 * sizeof(float)) {
49     // Load 24 (3x8) 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     x += 24;
54 
55     // Compute reduced argument elements := round(x / log(2)).
56     const __m256 vn0 = _mm256_round_ps(_mm256_mul_ps(vx0, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
57     const __m256 vn1 = _mm256_round_ps(_mm256_mul_ps(vx1, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
58     const __m256 vn2 = _mm256_round_ps(_mm256_mul_ps(vx2, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
59 
60     // Compute reduced argument t := x - elements * log(2).
61     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
62     __m256 vt0 = _mm256_fmadd_ps(vn0, vminus_ln2_hi, vx0);
63     __m256 vt1 = _mm256_fmadd_ps(vn1, vminus_ln2_hi, vx1);
64     __m256 vt2 = _mm256_fmadd_ps(vn2, vminus_ln2_hi, vx2);
65 
66     vt0 = _mm256_fmadd_ps(vn0, vminus_ln2_lo, vt0);
67     vt1 = _mm256_fmadd_ps(vn1, vminus_ln2_lo, vt1);
68     vt2 = _mm256_fmadd_ps(vn2, vminus_ln2_lo, vt2);
69 
70     // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
71     __m256 vp0 = _mm256_fmadd_ps(vc5, vt0, vc4);
72     __m256 vp1 = _mm256_fmadd_ps(vc5, vt1, vc4);
73     __m256 vp2 = _mm256_fmadd_ps(vc5, vt2, vc4);
74 
75     vp0 = _mm256_fmadd_ps(vp0, vt0, vc3);
76     vp1 = _mm256_fmadd_ps(vp1, vt1, vc3);
77     vp2 = _mm256_fmadd_ps(vp2, vt2, vc3);
78 
79     vp0 = _mm256_fmadd_ps(vp0, vt0, vc2);
80     vp1 = _mm256_fmadd_ps(vp1, vt1, vc2);
81     vp2 = _mm256_fmadd_ps(vp2, vt2, vc2);
82 
83     vp0 = _mm256_fmadd_ps(vp0, vt0, vc1);
84     vp1 = _mm256_fmadd_ps(vp1, vt1, vc1);
85     vp2 = _mm256_fmadd_ps(vp2, vt2, vc1);
86 
87     vp0 = _mm256_fmadd_ps(vp0, vt0, vc0);
88     vp1 = _mm256_fmadd_ps(vp1, vt1, vc0);
89     vp2 = _mm256_fmadd_ps(vp2, vt2, vc0);
90 
91     // Multiply "extended" floating-point numbers in ("mantissa", "exponent") representation where
92     //  - vnX is "exponent"
93     //  - vpX is "mantissa"
94     //
95     // exp2(ae) * av * exp2(be) * bv =
96     //   = exp2(ae + be) * (av * bv)
97     __m256 vf0 = _mm256_mul_ps(vp0, vscalev);
98     __m256 vf1 = _mm256_mul_ps(vp1, vscalev);
99     __m256 vf2 = _mm256_mul_ps(vp2, vscalev);
100 
101     __m256 ve0 = _mm256_add_ps(vn0, vscalee);
102     __m256 ve1 = _mm256_add_ps(vn1, vscalee);
103     __m256 ve2 = _mm256_add_ps(vn2, vscalee);
104 
105     // For computational efficiency, replace exp2(e) with 0.0f when e <= -127.0.
106     // This replacement is done in two steps:
107     // 1. Clamp minimum e at -127.0.
108     // 2. Map e to scale factor 0.0 when e == -127.0
109     ve0 = _mm256_max_ps(ve0, vmin_exponent);
110     ve1 = _mm256_max_ps(ve1, vmin_exponent);
111     ve2 = _mm256_max_ps(ve2, vmin_exponent);
112 
113     // Convert exponents into scale factors:
114     // - s = exp2(e) when e > -127.0
115     // - s = 0.0 when e <= -127.0
116     const __m256 vs0 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve0, vmagic_bias)), 23));
117     const __m256 vs1 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve1, vmagic_bias)), 23));
118     const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve2, vmagic_bias)), 23));
119 
120     // Multiply "mantissa" by the scale factor.
121     vf0 = _mm256_mul_ps(vf0, vs0);
122     vf1 = _mm256_mul_ps(vf1, vs1);
123     vf2 = _mm256_mul_ps(vf2, vs2);
124 
125     // Store 24 (3x8) outputs at a time.
126     _mm256_storeu_ps(y, vf0);
127     _mm256_storeu_ps(y + 8, vf1);
128     _mm256_storeu_ps(y + 16, vf2);
129     y += 24;
130   }
131 
132   for (; elements >= 8 * sizeof(float); elements -= 8 * sizeof(float)) {
133     // Load 8 inputs at a time.
134     const __m256 vx = _mm256_loadu_ps(x);
135     x += 8;
136 
137     // Compute reduced argument elements := round(x / log(2)).
138     const __m256 vn = _mm256_round_ps(_mm256_mul_ps(vx, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
139 
140     // Compute reduced argument t := x - elements * log(2).
141     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
142     __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2_hi, vx);
143     vt = _mm256_fmadd_ps(vn, vminus_ln2_lo, vt);
144 
145     // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
146     __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
147     vp = _mm256_fmadd_ps(vp, vt, vc3);
148     vp = _mm256_fmadd_ps(vp, vt, vc2);
149     vp = _mm256_fmadd_ps(vp, vt, vc1);
150     vp = _mm256_fmadd_ps(vp, vt, vc0);
151 
152     // Multiply "extended" floating-point numbers in ("mantissa", "exponent") representation.
153     __m256 vf = _mm256_mul_ps(vp, vscalev);
154     __m256 ve = _mm256_add_ps(vn, vscalee);
155 
156     // For computational efficiency, replace exp2(e) with 0.0f when e <= -127.0.
157     ve = _mm256_max_ps(ve, vmin_exponent);
158 
159     // Convert exponents into scale factors.
160     const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve, vmagic_bias)), 23));
161 
162     // Multiply "mantissa" by the scale factor.
163     vf = _mm256_mul_ps(vf, vs);
164 
165     // Store 8 results at a time.
166     _mm256_storeu_ps(y, vf);
167     y += 8;
168   }
169   if XNN_UNLIKELY(elements != 0) {
170     assert(elements >= 1 * sizeof(float));
171     assert(elements <= 7 * sizeof(float));
172     const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - elements));
173 
174     // Load up to 7 inputs at a time.
175     const __m256 vx = _mm256_maskload_ps(x, vmask);
176 
177     // Compute reduced argument elements := round(x / log(2)).
178     const __m256 vn = _mm256_round_ps(_mm256_mul_ps(vx, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
179 
180     // Compute reduced argument t := x - elements * log(2).
181     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
182     __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2_hi, vx);
183     vt = _mm256_fmadd_ps(vn, vminus_ln2_lo, vt);
184 
185     // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
186     __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
187     vp = _mm256_fmadd_ps(vp, vt, vc3);
188     vp = _mm256_fmadd_ps(vp, vt, vc2);
189     vp = _mm256_fmadd_ps(vp, vt, vc1);
190     vp = _mm256_fmadd_ps(vp, vt, vc0);
191 
192     // Multiply "extended" floating-point numbers in ("mantissa", "exponent") representation.
193     __m256 vf = _mm256_mul_ps(vp, vscalev);
194     __m256 ve = _mm256_add_ps(vn, vscalee);
195 
196     // For computational efficiency, replace exp2(e) with 0.0f when e <= -127.0.
197     ve = _mm256_max_ps(ve, vmin_exponent);
198 
199     // Convert exponents into scale factors.
200     const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve, vmagic_bias)), 23));
201 
202     // Multiply "mantissa" by the scale factor.
203     vf = _mm256_mul_ps(vf, vs);
204 
205     // Store up to 7 inputs at a time.
206     _mm256_maskstore_ps(y, vmask, vf);
207   }
208   _mm256_zeroupper();
209 }
210