xref: /aosp_15_r20/external/XNNPACK/src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-vsqrt/fma3-nr1fma1adj.c.in
3 //   Generator: tools/xngen
4 //
5 // Copyright 2020 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/vunary.h>
16 
17 
xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x24(size_t n,const float * x,float * y,const union xnn_f32_sqrt_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x24(
19     size_t n,
20     const float* x,
21     float* y,
22     const union xnn_f32_sqrt_params params[restrict XNN_MIN_ELEMENTS(1)])
23 {
24   assert(n != 0);
25   assert(n % sizeof(float) == 0);
26 
27   const __m256 vhalf = _mm256_load_ps(params->fma.half);
28   for (; n >= 24 * sizeof(float); n -= 24 * sizeof(float)) {
29     const __m256 vx0 = _mm256_loadu_ps(x);
30     const __m256 vx1 = _mm256_loadu_ps(x + 8);
31     const __m256 vx2 = _mm256_loadu_ps(x + 16);
32     x += 24;
33 
34     const __m256 vrsqrtx0 = _mm256_rsqrt_ps(vx0);
35     const __m256 vrsqrtx1 = _mm256_rsqrt_ps(vx1);
36     const __m256 vrsqrtx2 = _mm256_rsqrt_ps(vx2);
37 
38     __m256 vsqrtx0 = _mm256_mul_ps(vrsqrtx0, vx0);
39     __m256 vhalfrsqrtx0 = _mm256_mul_ps(vrsqrtx0, vhalf);
40     __m256 vsqrtx1 = _mm256_mul_ps(vrsqrtx1, vx1);
41     __m256 vhalfrsqrtx1 = _mm256_mul_ps(vrsqrtx1, vhalf);
42     __m256 vsqrtx2 = _mm256_mul_ps(vrsqrtx2, vx2);
43     __m256 vhalfrsqrtx2 = _mm256_mul_ps(vrsqrtx2, vhalf);
44 
45     const __m256 vresidual0 = _mm256_fnmadd_ps(vsqrtx0, vhalfrsqrtx0, vhalf);
46     const __m256 vresidual1 = _mm256_fnmadd_ps(vsqrtx1, vhalfrsqrtx1, vhalf);
47     const __m256 vresidual2 = _mm256_fnmadd_ps(vsqrtx2, vhalfrsqrtx2, vhalf);
48 
49     vhalfrsqrtx0 = _mm256_fmadd_ps(vhalfrsqrtx0, vresidual0, vhalfrsqrtx0);
50     vsqrtx0 = _mm256_fmadd_ps(vsqrtx0, vresidual0, vsqrtx0);
51     vhalfrsqrtx1 = _mm256_fmadd_ps(vhalfrsqrtx1, vresidual1, vhalfrsqrtx1);
52     vsqrtx1 = _mm256_fmadd_ps(vsqrtx1, vresidual1, vsqrtx1);
53     vhalfrsqrtx2 = _mm256_fmadd_ps(vhalfrsqrtx2, vresidual2, vhalfrsqrtx2);
54     vsqrtx2 = _mm256_fmadd_ps(vsqrtx2, vresidual2, vsqrtx2);
55 
56     const __m256 vadjustment0 = _mm256_fnmadd_ps(vsqrtx0, vsqrtx0, vx0);
57     const __m256 vadjustment1 = _mm256_fnmadd_ps(vsqrtx1, vsqrtx1, vx1);
58     const __m256 vadjustment2 = _mm256_fnmadd_ps(vsqrtx2, vsqrtx2, vx2);
59 
60     const __m256 vy0 = _mm256_fmadd_ps(vhalfrsqrtx0, vadjustment0, vsqrtx0);
61     const __m256 vy1 = _mm256_fmadd_ps(vhalfrsqrtx1, vadjustment1, vsqrtx1);
62     const __m256 vy2 = _mm256_fmadd_ps(vhalfrsqrtx2, vadjustment2, vsqrtx2);
63 
64     _mm256_storeu_ps(y, vy0);
65     _mm256_storeu_ps(y + 8, vy1);
66     _mm256_storeu_ps(y + 16, vy2);
67     y += 24;
68   }
69   for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
70     const __m256 vx = _mm256_loadu_ps(x);
71     x += 8;
72 
73     const __m256 vrsqrtx = _mm256_rsqrt_ps(vx);
74     __m256 vsqrtx = _mm256_mul_ps(vrsqrtx, vx);
75     __m256 vhalfrsqrtx = _mm256_mul_ps(vrsqrtx, vhalf);
76     const __m256 vresidual = _mm256_fnmadd_ps(vsqrtx, vhalfrsqrtx, vhalf);
77     vhalfrsqrtx = _mm256_fmadd_ps(vhalfrsqrtx, vresidual, vhalfrsqrtx);
78     vsqrtx = _mm256_fmadd_ps(vsqrtx, vresidual, vsqrtx);
79     const __m256 vadjustment = _mm256_fnmadd_ps(vsqrtx, vsqrtx, vx);
80     const __m256 vy = _mm256_fmadd_ps(vhalfrsqrtx, vadjustment, vsqrtx);
81 
82     _mm256_storeu_ps(y, vy);
83     y += 8;
84   }
85   if XNN_UNLIKELY(n != 0) {
86     assert(n >= 1 * sizeof(float));
87     assert(n <= 7 * sizeof(float));
88     const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &params->fma.mask_table[7] - n));
89 
90     const __m256 vx = _mm256_maskload_ps(x, vmask);
91 
92     const __m256 vrsqrtx = _mm256_rsqrt_ps(vx);
93     __m256 vsqrtx = _mm256_mul_ps(vrsqrtx, vx);
94     __m256 vhalfrsqrtx = _mm256_mul_ps(vrsqrtx, vhalf);
95     const __m256 vresidual = _mm256_fnmadd_ps(vsqrtx, vhalfrsqrtx, vhalf);
96     vhalfrsqrtx = _mm256_fmadd_ps(vhalfrsqrtx, vresidual, vhalfrsqrtx);
97     vsqrtx = _mm256_fmadd_ps(vsqrtx, vresidual, vsqrtx);
98     const __m256 vadjustment = _mm256_fnmadd_ps(vsqrtx, vsqrtx, vx);
99     const __m256 vy = _mm256_fmadd_ps(vhalfrsqrtx, vadjustment, vsqrtx);
100 
101     __m128 vy_lo = _mm256_castps256_ps128(vy);
102     if (n & (4 * sizeof(float))) {
103       _mm_storeu_ps(y, vy_lo);
104       vy_lo = _mm256_extractf128_ps(vy, 1);
105       y += 4;
106     }
107     if (n & (2 * sizeof(float))) {
108       _mm_storel_pi((__m64*) y, vy_lo);
109       vy_lo = _mm_movehl_ps(vy_lo, vy_lo);
110       y += 2;
111     }
112     if (n & (1 * sizeof(float))) {
113       _mm_store_ss(y, vy_lo);
114     }
115   }
116 }
117