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_x16(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_x16(
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 >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
29 const __m256 vx0 = _mm256_loadu_ps(x);
30 const __m256 vx1 = _mm256_loadu_ps(x + 8);
31 x += 16;
32
33 const __m256 vrsqrtx0 = _mm256_rsqrt_ps(vx0);
34 const __m256 vrsqrtx1 = _mm256_rsqrt_ps(vx1);
35
36 __m256 vsqrtx0 = _mm256_mul_ps(vrsqrtx0, vx0);
37 __m256 vhalfrsqrtx0 = _mm256_mul_ps(vrsqrtx0, vhalf);
38 __m256 vsqrtx1 = _mm256_mul_ps(vrsqrtx1, vx1);
39 __m256 vhalfrsqrtx1 = _mm256_mul_ps(vrsqrtx1, vhalf);
40
41 const __m256 vresidual0 = _mm256_fnmadd_ps(vsqrtx0, vhalfrsqrtx0, vhalf);
42 const __m256 vresidual1 = _mm256_fnmadd_ps(vsqrtx1, vhalfrsqrtx1, vhalf);
43
44 vhalfrsqrtx0 = _mm256_fmadd_ps(vhalfrsqrtx0, vresidual0, vhalfrsqrtx0);
45 vsqrtx0 = _mm256_fmadd_ps(vsqrtx0, vresidual0, vsqrtx0);
46 vhalfrsqrtx1 = _mm256_fmadd_ps(vhalfrsqrtx1, vresidual1, vhalfrsqrtx1);
47 vsqrtx1 = _mm256_fmadd_ps(vsqrtx1, vresidual1, vsqrtx1);
48
49 const __m256 vadjustment0 = _mm256_fnmadd_ps(vsqrtx0, vsqrtx0, vx0);
50 const __m256 vadjustment1 = _mm256_fnmadd_ps(vsqrtx1, vsqrtx1, vx1);
51
52 const __m256 vy0 = _mm256_fmadd_ps(vhalfrsqrtx0, vadjustment0, vsqrtx0);
53 const __m256 vy1 = _mm256_fmadd_ps(vhalfrsqrtx1, vadjustment1, vsqrtx1);
54
55 _mm256_storeu_ps(y, vy0);
56 _mm256_storeu_ps(y + 8, vy1);
57 y += 16;
58 }
59 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
60 const __m256 vx = _mm256_loadu_ps(x);
61 x += 8;
62
63 const __m256 vrsqrtx = _mm256_rsqrt_ps(vx);
64 __m256 vsqrtx = _mm256_mul_ps(vrsqrtx, vx);
65 __m256 vhalfrsqrtx = _mm256_mul_ps(vrsqrtx, vhalf);
66 const __m256 vresidual = _mm256_fnmadd_ps(vsqrtx, vhalfrsqrtx, vhalf);
67 vhalfrsqrtx = _mm256_fmadd_ps(vhalfrsqrtx, vresidual, vhalfrsqrtx);
68 vsqrtx = _mm256_fmadd_ps(vsqrtx, vresidual, vsqrtx);
69 const __m256 vadjustment = _mm256_fnmadd_ps(vsqrtx, vsqrtx, vx);
70 const __m256 vy = _mm256_fmadd_ps(vhalfrsqrtx, vadjustment, vsqrtx);
71
72 _mm256_storeu_ps(y, vy);
73 y += 8;
74 }
75 if XNN_UNLIKELY(n != 0) {
76 assert(n >= 1 * sizeof(float));
77 assert(n <= 7 * sizeof(float));
78 const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) ¶ms->fma.mask_table[7] - n));
79
80 const __m256 vx = _mm256_maskload_ps(x, vmask);
81
82 const __m256 vrsqrtx = _mm256_rsqrt_ps(vx);
83 __m256 vsqrtx = _mm256_mul_ps(vrsqrtx, vx);
84 __m256 vhalfrsqrtx = _mm256_mul_ps(vrsqrtx, vhalf);
85 const __m256 vresidual = _mm256_fnmadd_ps(vsqrtx, vhalfrsqrtx, vhalf);
86 vhalfrsqrtx = _mm256_fmadd_ps(vhalfrsqrtx, vresidual, vhalfrsqrtx);
87 vsqrtx = _mm256_fmadd_ps(vsqrtx, vresidual, vsqrtx);
88 const __m256 vadjustment = _mm256_fnmadd_ps(vsqrtx, vsqrtx, vx);
89 const __m256 vy = _mm256_fmadd_ps(vhalfrsqrtx, vadjustment, vsqrtx);
90
91 __m128 vy_lo = _mm256_castps256_ps128(vy);
92 if (n & (4 * sizeof(float))) {
93 _mm_storeu_ps(y, vy_lo);
94 vy_lo = _mm256_extractf128_ps(vy, 1);
95 y += 4;
96 }
97 if (n & (2 * sizeof(float))) {
98 _mm_storel_pi((__m64*) y, vy_lo);
99 vy_lo = _mm_movehl_ps(vy_lo, vy_lo);
100 y += 2;
101 }
102 if (n & (1 * sizeof(float))) {
103 _mm_store_ss(y, vy_lo);
104 }
105 }
106 }
107