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_x32(size_t elements,const float * x,float * y,float scale_value,float scale_exp)20 void xnn_f32_vscaleextexp_ukernel__avx2_p5_x32(
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 >= 32 * sizeof(float); elements -= 32 * sizeof(float)) {
49 // Load 32 (4x8) 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 x += 32;
55
56 // Compute reduced argument elements := round(x / log(2)).
57 const __m256 vn0 = _mm256_round_ps(_mm256_mul_ps(vx0, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
58 const __m256 vn1 = _mm256_round_ps(_mm256_mul_ps(vx1, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
59 const __m256 vn2 = _mm256_round_ps(_mm256_mul_ps(vx2, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
60 const __m256 vn3 = _mm256_round_ps(_mm256_mul_ps(vx3, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
61
62 // Compute reduced argument t := x - elements * log(2).
63 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
64 __m256 vt0 = _mm256_fmadd_ps(vn0, vminus_ln2_hi, vx0);
65 __m256 vt1 = _mm256_fmadd_ps(vn1, vminus_ln2_hi, vx1);
66 __m256 vt2 = _mm256_fmadd_ps(vn2, vminus_ln2_hi, vx2);
67 __m256 vt3 = _mm256_fmadd_ps(vn3, vminus_ln2_hi, vx3);
68
69 vt0 = _mm256_fmadd_ps(vn0, vminus_ln2_lo, vt0);
70 vt1 = _mm256_fmadd_ps(vn1, vminus_ln2_lo, vt1);
71 vt2 = _mm256_fmadd_ps(vn2, vminus_ln2_lo, vt2);
72 vt3 = _mm256_fmadd_ps(vn3, vminus_ln2_lo, vt3);
73
74 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
75 __m256 vp0 = _mm256_fmadd_ps(vc5, vt0, vc4);
76 __m256 vp1 = _mm256_fmadd_ps(vc5, vt1, vc4);
77 __m256 vp2 = _mm256_fmadd_ps(vc5, vt2, vc4);
78 __m256 vp3 = _mm256_fmadd_ps(vc5, vt3, vc4);
79
80 vp0 = _mm256_fmadd_ps(vp0, vt0, vc3);
81 vp1 = _mm256_fmadd_ps(vp1, vt1, vc3);
82 vp2 = _mm256_fmadd_ps(vp2, vt2, vc3);
83 vp3 = _mm256_fmadd_ps(vp3, vt3, vc3);
84
85 vp0 = _mm256_fmadd_ps(vp0, vt0, vc2);
86 vp1 = _mm256_fmadd_ps(vp1, vt1, vc2);
87 vp2 = _mm256_fmadd_ps(vp2, vt2, vc2);
88 vp3 = _mm256_fmadd_ps(vp3, vt3, vc2);
89
90 vp0 = _mm256_fmadd_ps(vp0, vt0, vc1);
91 vp1 = _mm256_fmadd_ps(vp1, vt1, vc1);
92 vp2 = _mm256_fmadd_ps(vp2, vt2, vc1);
93 vp3 = _mm256_fmadd_ps(vp3, vt3, vc1);
94
95 vp0 = _mm256_fmadd_ps(vp0, vt0, vc0);
96 vp1 = _mm256_fmadd_ps(vp1, vt1, vc0);
97 vp2 = _mm256_fmadd_ps(vp2, vt2, vc0);
98 vp3 = _mm256_fmadd_ps(vp3, vt3, vc0);
99
100 // Multiply "extended" floating-point numbers in ("mantissa", "exponent") representation where
101 // - vnX is "exponent"
102 // - vpX is "mantissa"
103 //
104 // exp2(ae) * av * exp2(be) * bv =
105 // = exp2(ae + be) * (av * bv)
106 __m256 vf0 = _mm256_mul_ps(vp0, vscalev);
107 __m256 vf1 = _mm256_mul_ps(vp1, vscalev);
108 __m256 vf2 = _mm256_mul_ps(vp2, vscalev);
109 __m256 vf3 = _mm256_mul_ps(vp3, vscalev);
110
111 __m256 ve0 = _mm256_add_ps(vn0, vscalee);
112 __m256 ve1 = _mm256_add_ps(vn1, vscalee);
113 __m256 ve2 = _mm256_add_ps(vn2, vscalee);
114 __m256 ve3 = _mm256_add_ps(vn3, vscalee);
115
116 // For computational efficiency, replace exp2(e) with 0.0f when e <= -127.0.
117 // This replacement is done in two steps:
118 // 1. Clamp minimum e at -127.0.
119 // 2. Map e to scale factor 0.0 when e == -127.0
120 ve0 = _mm256_max_ps(ve0, vmin_exponent);
121 ve1 = _mm256_max_ps(ve1, vmin_exponent);
122 ve2 = _mm256_max_ps(ve2, vmin_exponent);
123 ve3 = _mm256_max_ps(ve3, vmin_exponent);
124
125 // Convert exponents into scale factors:
126 // - s = exp2(e) when e > -127.0
127 // - s = 0.0 when e <= -127.0
128 const __m256 vs0 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve0, vmagic_bias)), 23));
129 const __m256 vs1 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve1, vmagic_bias)), 23));
130 const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve2, vmagic_bias)), 23));
131 const __m256 vs3 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve3, vmagic_bias)), 23));
132
133 // Multiply "mantissa" by the scale factor.
134 vf0 = _mm256_mul_ps(vf0, vs0);
135 vf1 = _mm256_mul_ps(vf1, vs1);
136 vf2 = _mm256_mul_ps(vf2, vs2);
137 vf3 = _mm256_mul_ps(vf3, vs3);
138
139 // Store 32 (4x8) outputs at a time.
140 _mm256_storeu_ps(y, vf0);
141 _mm256_storeu_ps(y + 8, vf1);
142 _mm256_storeu_ps(y + 16, vf2);
143 _mm256_storeu_ps(y + 24, vf3);
144 y += 32;
145 }
146
147 for (; elements >= 8 * sizeof(float); elements -= 8 * sizeof(float)) {
148 // Load 8 inputs at a time.
149 const __m256 vx = _mm256_loadu_ps(x);
150 x += 8;
151
152 // Compute reduced argument elements := round(x / log(2)).
153 const __m256 vn = _mm256_round_ps(_mm256_mul_ps(vx, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
154
155 // Compute reduced argument t := x - elements * log(2).
156 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
157 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2_hi, vx);
158 vt = _mm256_fmadd_ps(vn, vminus_ln2_lo, vt);
159
160 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
161 __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
162 vp = _mm256_fmadd_ps(vp, vt, vc3);
163 vp = _mm256_fmadd_ps(vp, vt, vc2);
164 vp = _mm256_fmadd_ps(vp, vt, vc1);
165 vp = _mm256_fmadd_ps(vp, vt, vc0);
166
167 // Multiply "extended" floating-point numbers in ("mantissa", "exponent") representation.
168 __m256 vf = _mm256_mul_ps(vp, vscalev);
169 __m256 ve = _mm256_add_ps(vn, vscalee);
170
171 // For computational efficiency, replace exp2(e) with 0.0f when e <= -127.0.
172 ve = _mm256_max_ps(ve, vmin_exponent);
173
174 // Convert exponents into scale factors.
175 const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve, vmagic_bias)), 23));
176
177 // Multiply "mantissa" by the scale factor.
178 vf = _mm256_mul_ps(vf, vs);
179
180 // Store 8 results at a time.
181 _mm256_storeu_ps(y, vf);
182 y += 8;
183 }
184 if XNN_UNLIKELY(elements != 0) {
185 assert(elements >= 1 * sizeof(float));
186 assert(elements <= 7 * sizeof(float));
187 const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - elements));
188
189 // Load up to 7 inputs at a time.
190 const __m256 vx = _mm256_maskload_ps(x, vmask);
191
192 // Compute reduced argument elements := round(x / log(2)).
193 const __m256 vn = _mm256_round_ps(_mm256_mul_ps(vx, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
194
195 // Compute reduced argument t := x - elements * log(2).
196 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
197 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2_hi, vx);
198 vt = _mm256_fmadd_ps(vn, vminus_ln2_lo, vt);
199
200 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
201 __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
202 vp = _mm256_fmadd_ps(vp, vt, vc3);
203 vp = _mm256_fmadd_ps(vp, vt, vc2);
204 vp = _mm256_fmadd_ps(vp, vt, vc1);
205 vp = _mm256_fmadd_ps(vp, vt, vc0);
206
207 // Multiply "extended" floating-point numbers in ("mantissa", "exponent") representation.
208 __m256 vf = _mm256_mul_ps(vp, vscalev);
209 __m256 ve = _mm256_add_ps(vn, vscalee);
210
211 // For computational efficiency, replace exp2(e) with 0.0f when e <= -127.0.
212 ve = _mm256_max_ps(ve, vmin_exponent);
213
214 // Convert exponents into scale factors.
215 const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve, vmagic_bias)), 23));
216
217 // Multiply "mantissa" by the scale factor.
218 vf = _mm256_mul_ps(vf, vs);
219
220 // Store up to 7 inputs at a time.
221 _mm256_maskstore_ps(y, vmask, vf);
222 }
223 _mm256_zeroupper();
224 }
225