xref: /aosp_15_r20/external/XNNPACK/src/f32-vscaleexpminusmax/gen/avx2-p5-x32.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_x32(size_t elements,const float * input,float * output,float scale,float max)20 void xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x32(
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 >= 32 * sizeof(float); elements -= 32 * sizeof(float)) {
46     // Load 32 (4x8) inputs at a time.
47     const __m256 vi0 = _mm256_loadu_ps(input);
48     const __m256 vi1 = _mm256_loadu_ps(input + 8);
49     const __m256 vi2 = _mm256_loadu_ps(input + 16);
50     const __m256 vi3 = _mm256_loadu_ps(input + 24);
51     input += 32;
52 
53     // Subtract maximum input x := i - i_max. This implies x <= 0.
54     const __m256 vx0 = _mm256_sub_ps(vi0, vi_max);
55     const __m256 vx1 = _mm256_sub_ps(vi1, vi_max);
56     const __m256 vx2 = _mm256_sub_ps(vi2, vi_max);
57     const __m256 vx3 = _mm256_sub_ps(vi3, vi_max);
58 
59     // Compute reduced argument elements := round(x / log(2)).
60     __m256 vn0 = _mm256_fmadd_ps(vx0, vlog2e, vmagic_bias);
61     __m256 vn1 = _mm256_fmadd_ps(vx1, vlog2e, vmagic_bias);
62     __m256 vn2 = _mm256_fmadd_ps(vx2, vlog2e, vmagic_bias);
63     __m256 vn3 = _mm256_fmadd_ps(vx3, vlog2e, vmagic_bias);
64 
65     // Create a floating-point number s (scale) such that s == 2**elements for inputs which don't cause underflow, i.e.
66     // -87.33642 <= x <= 0.0, and -126 <= elements <= 0 accordingly.
67     const __m256 vs0 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn0), 23));
68     const __m256 vs1 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn1), 23));
69     const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn2), 23));
70     const __m256 vs3 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn3), 23));
71 
72     // Subtract the large number back to get final elements := round(x / log(2)).
73     vn0 = _mm256_sub_ps(vn0, vmagic_bias);
74     vn1 = _mm256_sub_ps(vn1, vmagic_bias);
75     vn2 = _mm256_sub_ps(vn2, vmagic_bias);
76     vn3 = _mm256_sub_ps(vn3, vmagic_bias);
77 
78     // Compute reduced argument t := x - elements * log(2).
79     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
80     __m256 vt0 = _mm256_fmadd_ps(vn0, vminus_ln2_hi, vx0);
81     __m256 vt1 = _mm256_fmadd_ps(vn1, vminus_ln2_hi, vx1);
82     __m256 vt2 = _mm256_fmadd_ps(vn2, vminus_ln2_hi, vx2);
83     __m256 vt3 = _mm256_fmadd_ps(vn3, vminus_ln2_hi, vx3);
84 
85     vt0 = _mm256_fmadd_ps(vn0, vminus_ln2_lo, vt0);
86     vt1 = _mm256_fmadd_ps(vn1, vminus_ln2_lo, vt1);
87     vt2 = _mm256_fmadd_ps(vn2, vminus_ln2_lo, vt2);
88     vt3 = _mm256_fmadd_ps(vn3, vminus_ln2_lo, vt3);
89 
90     // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
91     __m256 vp0 = _mm256_fmadd_ps(vc5, vt0, vc4);
92     __m256 vp1 = _mm256_fmadd_ps(vc5, vt1, vc4);
93     __m256 vp2 = _mm256_fmadd_ps(vc5, vt2, vc4);
94     __m256 vp3 = _mm256_fmadd_ps(vc5, vt3, vc4);
95 
96     vp0 = _mm256_fmadd_ps(vp0, vt0, vc3);
97     vp1 = _mm256_fmadd_ps(vp1, vt1, vc3);
98     vp2 = _mm256_fmadd_ps(vp2, vt2, vc3);
99     vp3 = _mm256_fmadd_ps(vp3, vt3, vc3);
100 
101     vp0 = _mm256_fmadd_ps(vp0, vt0, vc2);
102     vp1 = _mm256_fmadd_ps(vp1, vt1, vc2);
103     vp2 = _mm256_fmadd_ps(vp2, vt2, vc2);
104     vp3 = _mm256_fmadd_ps(vp3, vt3, vc2);
105 
106     vp0 = _mm256_fmadd_ps(vp0, vt0, vc1);
107     vp1 = _mm256_fmadd_ps(vp1, vt1, vc1);
108     vp2 = _mm256_fmadd_ps(vp2, vt2, vc1);
109     vp3 = _mm256_fmadd_ps(vp3, vt3, vc1);
110 
111     // Reconstruct the final f value:
112     //   f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
113     //     = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
114     //     = s + (t * s) * p
115     vt0 = _mm256_mul_ps(vt0, vs0);
116     vt1 = _mm256_mul_ps(vt1, vs1);
117     vt2 = _mm256_mul_ps(vt2, vs2);
118     vt3 = _mm256_mul_ps(vt3, vs3);
119 
120     __m256 vf0 = _mm256_fmadd_ps(vt0, vp0, vs0);
121     __m256 vf1 = _mm256_fmadd_ps(vt1, vp1, vs1);
122     __m256 vf2 = _mm256_fmadd_ps(vt2, vp2, vs2);
123     __m256 vf3 = _mm256_fmadd_ps(vt3, vp3, vs3);
124 
125     // For inputs below zero cutoff, replace output with +0.0f.
126     // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
127     vf0 = _mm256_andnot_ps(_mm256_cmp_ps(vx0, vdenorm_cutoff, _CMP_LT_OS), vf0);
128     vf1 = _mm256_andnot_ps(_mm256_cmp_ps(vx1, vdenorm_cutoff, _CMP_LT_OS), vf1);
129     vf2 = _mm256_andnot_ps(_mm256_cmp_ps(vx2, vdenorm_cutoff, _CMP_LT_OS), vf2);
130     vf3 = _mm256_andnot_ps(_mm256_cmp_ps(vx3, vdenorm_cutoff, _CMP_LT_OS), vf3);
131 
132     // Multiply by scale.
133     vf0 = _mm256_mul_ps(vf0, vscale);
134     vf1 = _mm256_mul_ps(vf1, vscale);
135     vf2 = _mm256_mul_ps(vf2, vscale);
136     vf3 = _mm256_mul_ps(vf3, vscale);
137 
138     // Store 32 (4x8) outputs at a time.
139     _mm256_storeu_ps(output, vf0);
140     _mm256_storeu_ps(output + 8, vf1);
141     _mm256_storeu_ps(output + 16, vf2);
142     _mm256_storeu_ps(output + 24, vf3);
143     output += 32;
144   }
145   for (; elements >= 8 * sizeof(float); elements -= 8 * sizeof(float)) {
146     // Load 8 inputs at a time.
147     const __m256 vi = _mm256_loadu_ps(input);
148     input += 8;
149 
150     // Subtract maximum input x := i - i_max. This implies x <= 0.
151     const __m256 vx = _mm256_sub_ps(vi, vi_max);
152 
153     // Compute reduced argument elements := round(x / log(2)).
154     __m256 vn = _mm256_fmadd_ps(vx, vlog2e, vmagic_bias);
155 
156     // Create a floating-point number s (scale) such that s == 2**elements for inputs which don't cause underflow, i.e.
157     // -87.33642 <= x <= 0.0, and -126 <= elements <= 0 accordingly.
158     const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
159 
160     // Subtract the large number back to get final elements := round(x / log(2)).
161     vn = _mm256_sub_ps(vn, vmagic_bias);
162 
163     // Compute reduced argument t := x - elements * log(2).
164     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
165     __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2_hi, vx);
166     vt = _mm256_fmadd_ps(vn, vminus_ln2_lo, vt);
167 
168     // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
169     __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
170     vp = _mm256_fmadd_ps(vp, vt, vc3);
171     vp = _mm256_fmadd_ps(vp, vt, vc2);
172     vp = _mm256_fmadd_ps(vp, vt, vc1);
173 
174     // Reconstruct the final f value:
175     //   f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
176     //     = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
177     //     = s + (t * s) * p
178     vt = _mm256_mul_ps(vt, vs);
179     __m256 vf = _mm256_fmadd_ps(vt, vp, vs);
180 
181     // For inputs below zero cutoff, replace output with +0.0f.
182     // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
183     vf = _mm256_andnot_ps(_mm256_cmp_ps(vx, vdenorm_cutoff, _CMP_LT_OS), vf);
184 
185     // Multiply by scale.
186     vf = _mm256_mul_ps(vf, vscale);
187 
188     // Store 64 (8x8) outputs at a time.
189     _mm256_storeu_ps(output, vf);
190     output += 8;
191   }
192   if (elements != 0) {
193     assert(elements >= 1 * sizeof(float));
194     assert(elements <= 7 * sizeof(float));
195     const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - elements));
196 
197     // Load up to 7 inputs at a time.
198     const __m256 vi = _mm256_maskload_ps(input, vmask);
199 
200     // Subtract maximum input x := i - i_max. This implies x <= 0.
201     const __m256 vx = _mm256_sub_ps(vi, vi_max);
202 
203     // Compute reduced argument elements := round(x / log(2)).
204     __m256 vn = _mm256_fmadd_ps(vx, vlog2e, vmagic_bias);
205 
206     // Create a floating-point number s (scale) such that s == 2**elements for inputs which don't cause underflow, i.e.
207     // -87.33642 <= x <= 0.0, and -126 <= elements <= 0 accordingly.
208     const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
209 
210     // Subtract the large number back to get final elements := round(x / log(2)).
211     vn = _mm256_sub_ps(vn, vmagic_bias);
212 
213     // Compute reduced argument t := x - elements * log(2).
214     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
215     __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2_hi, vx);
216     vt = _mm256_fmadd_ps(vn, vminus_ln2_lo, vt);
217 
218     // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
219     __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
220     vp = _mm256_fmadd_ps(vp, vt, vc3);
221     vp = _mm256_fmadd_ps(vp, vt, vc2);
222     vp = _mm256_fmadd_ps(vp, vt, vc1);
223 
224     // Reconstruct the final f value:
225     //   f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
226     //     = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
227     //     = s + (t * s) * p
228     vt = _mm256_mul_ps(vt, vs);
229     __m256 vf = _mm256_fmadd_ps(vt, vp, vs);
230 
231     // For inputs below zero cutoff, replace output with +0.0f.
232     // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
233     vf = _mm256_andnot_ps(_mm256_cmp_ps(vx, vdenorm_cutoff, _CMP_LT_OS), vf);
234 
235     // Multiply by scale.
236     vf = _mm256_mul_ps(vf, vscale);
237 
238     // Store up to 7 outputs at a time.
239     _mm256_maskstore_ps(output, vmask, vf);
240   }
241 }
242