xref: /aosp_15_r20/external/XNNPACK/src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-vscaleexpminusmax/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/vscaleexpminusmax.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_vscaleexpminusmax_ukernel__avx2_p5_x8(size_t elements,const float * input,float * output,float scale,float max)20 void xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x8(
21     size_t elements,
22     const float* input,
23     float* output,
24     float scale,
25     float max)
26 {
27   assert(elements % sizeof(float) == 0);
28 
29   const __m256 vmagic_bias = _mm256_set1_ps(0x1.8000FEp23f);
30   // The smallest x for which expf(x) is normalized.
31   const __m256 vdenorm_cutoff = _mm256_set1_ps(-0x1.5D589Ep6f);
32   const __m256 vlog2e = _mm256_set1_ps(0x1.715476p+0f);
33   const __m256 vminus_ln2_hi = _mm256_set1_ps(-0x1.62E43p-1f);
34   const __m256 vminus_ln2_lo = _mm256_set1_ps(0x1.05C61p-29f);
35 
36   const __m256 vc1 = _mm256_set1_ps(0x1.FFFFF6p-1f);
37   const __m256 vc2 = _mm256_set1_ps(0x1.FFFDC6p-2f);
38   const __m256 vc3 = _mm256_set1_ps(0x1.555A80p-3f);
39   const __m256 vc4 = _mm256_set1_ps(0x1.573A1Ap-5f);
40   const __m256 vc5 = _mm256_set1_ps(0x1.0F9F9Cp-7f);
41 
42   const __m256 vscale = _mm256_set1_ps(scale);
43   const __m256 vi_max = _mm256_set1_ps(max);
44 
45   for (; elements >= 8 * sizeof(float); elements -= 8 * sizeof(float)) {
46     // Load 8 (1x8) inputs at a time.
47     const __m256 vi0 = _mm256_loadu_ps(input);
48     input += 8;
49 
50     // Subtract maximum input x := i - i_max. This implies x <= 0.
51     const __m256 vx0 = _mm256_sub_ps(vi0, vi_max);
52 
53     // Compute reduced argument elements := round(x / log(2)).
54     __m256 vn0 = _mm256_fmadd_ps(vx0, vlog2e, vmagic_bias);
55 
56     // Create a floating-point number s (scale) such that s == 2**elements for inputs which don't cause underflow, i.e.
57     // -87.33642 <= x <= 0.0, and -126 <= elements <= 0 accordingly.
58     const __m256 vs0 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn0), 23));
59 
60     // Subtract the large number back to get final elements := round(x / log(2)).
61     vn0 = _mm256_sub_ps(vn0, vmagic_bias);
62 
63     // Compute reduced argument t := x - elements * log(2).
64     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
65     __m256 vt0 = _mm256_fmadd_ps(vn0, vminus_ln2_hi, vx0);
66 
67     vt0 = _mm256_fmadd_ps(vn0, vminus_ln2_lo, vt0);
68 
69     // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
70     __m256 vp0 = _mm256_fmadd_ps(vc5, vt0, vc4);
71 
72     vp0 = _mm256_fmadd_ps(vp0, vt0, vc3);
73 
74     vp0 = _mm256_fmadd_ps(vp0, vt0, vc2);
75 
76     vp0 = _mm256_fmadd_ps(vp0, vt0, vc1);
77 
78     // Reconstruct the final f value:
79     //   f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
80     //     = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
81     //     = s + (t * s) * p
82     vt0 = _mm256_mul_ps(vt0, vs0);
83 
84     __m256 vf0 = _mm256_fmadd_ps(vt0, vp0, vs0);
85 
86     // For inputs below zero cutoff, replace output with +0.0f.
87     // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
88     vf0 = _mm256_andnot_ps(_mm256_cmp_ps(vx0, vdenorm_cutoff, _CMP_LT_OS), vf0);
89 
90     // Multiply by scale.
91     vf0 = _mm256_mul_ps(vf0, vscale);
92 
93     // Store 8 (1x8) outputs at a time.
94     _mm256_storeu_ps(output, vf0);
95     output += 8;
96   }
97   for (; elements >= 8 * sizeof(float); elements -= 8 * sizeof(float)) {
98     // Load 8 inputs at a time.
99     const __m256 vi = _mm256_loadu_ps(input);
100     input += 8;
101 
102     // Subtract maximum input x := i - i_max. This implies x <= 0.
103     const __m256 vx = _mm256_sub_ps(vi, vi_max);
104 
105     // Compute reduced argument elements := round(x / log(2)).
106     __m256 vn = _mm256_fmadd_ps(vx, vlog2e, vmagic_bias);
107 
108     // Create a floating-point number s (scale) such that s == 2**elements for inputs which don't cause underflow, i.e.
109     // -87.33642 <= x <= 0.0, and -126 <= elements <= 0 accordingly.
110     const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
111 
112     // Subtract the large number back to get final elements := round(x / log(2)).
113     vn = _mm256_sub_ps(vn, vmagic_bias);
114 
115     // Compute reduced argument t := x - elements * log(2).
116     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
117     __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2_hi, vx);
118     vt = _mm256_fmadd_ps(vn, vminus_ln2_lo, vt);
119 
120     // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
121     __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
122     vp = _mm256_fmadd_ps(vp, vt, vc3);
123     vp = _mm256_fmadd_ps(vp, vt, vc2);
124     vp = _mm256_fmadd_ps(vp, vt, vc1);
125 
126     // Reconstruct the final f value:
127     //   f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
128     //     = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
129     //     = s + (t * s) * p
130     vt = _mm256_mul_ps(vt, vs);
131     __m256 vf = _mm256_fmadd_ps(vt, vp, vs);
132 
133     // For inputs below zero cutoff, replace output with +0.0f.
134     // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
135     vf = _mm256_andnot_ps(_mm256_cmp_ps(vx, vdenorm_cutoff, _CMP_LT_OS), vf);
136 
137     // Multiply by scale.
138     vf = _mm256_mul_ps(vf, vscale);
139 
140     // Store 64 (8x8) outputs at a time.
141     _mm256_storeu_ps(output, vf);
142     output += 8;
143   }
144   if (elements != 0) {
145     assert(elements >= 1 * sizeof(float));
146     assert(elements <= 7 * sizeof(float));
147     const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - elements));
148 
149     // Load up to 7 inputs at a time.
150     const __m256 vi = _mm256_maskload_ps(input, vmask);
151 
152     // Subtract maximum input x := i - i_max. This implies x <= 0.
153     const __m256 vx = _mm256_sub_ps(vi, vi_max);
154 
155     // Compute reduced argument elements := round(x / log(2)).
156     __m256 vn = _mm256_fmadd_ps(vx, vlog2e, vmagic_bias);
157 
158     // Create a floating-point number s (scale) such that s == 2**elements for inputs which don't cause underflow, i.e.
159     // -87.33642 <= x <= 0.0, and -126 <= elements <= 0 accordingly.
160     const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
161 
162     // Subtract the large number back to get final elements := round(x / log(2)).
163     vn = _mm256_sub_ps(vn, vmagic_bias);
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 
176     // Reconstruct the final f value:
177     //   f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
178     //     = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
179     //     = s + (t * s) * p
180     vt = _mm256_mul_ps(vt, vs);
181     __m256 vf = _mm256_fmadd_ps(vt, vp, vs);
182 
183     // For inputs below zero cutoff, replace output with +0.0f.
184     // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
185     vf = _mm256_andnot_ps(_mm256_cmp_ps(vx, vdenorm_cutoff, _CMP_LT_OS), vf);
186 
187     // Multiply by scale.
188     vf = _mm256_mul_ps(vf, vscale);
189 
190     // Store up to 7 outputs at a time.
191     _mm256_maskstore_ps(output, vmask, vf);
192   }
193 }
194