1 // Auto-generated file. Do not edit!
2 // Template: src/f32-vsqrt/avx512f-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/intrinsics-polyfill.h>
16 #include <xnnpack/vunary.h>
17
18
xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x80(size_t n,const float * x,float * y,const union xnn_f32_sqrt_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x80(
20 size_t n,
21 const float* x,
22 float* y,
23 const union xnn_f32_sqrt_params params[restrict XNN_MIN_ELEMENTS(1)])
24 {
25 assert(n != 0);
26 assert(n % sizeof(float) == 0);
27
28 const __m512 vhalf = _mm512_set1_ps(params->fma.half);
29 for (; n >= 80 * sizeof(float); n -= 80 * sizeof(float)) {
30 const __m512 vx0 = _mm512_loadu_ps(x);
31 const __m512 vx1 = _mm512_loadu_ps(x + 16);
32 const __m512 vx2 = _mm512_loadu_ps(x + 32);
33 const __m512 vx3 = _mm512_loadu_ps(x + 48);
34 const __m512 vx4 = _mm512_loadu_ps(x + 64);
35 x += 80;
36
37 const __m512 vrsqrtx0 = _mm512_rsqrt14_ps(vx0);
38 const __m512 vrsqrtx1 = _mm512_rsqrt14_ps(vx1);
39 const __m512 vrsqrtx2 = _mm512_rsqrt14_ps(vx2);
40 const __m512 vrsqrtx3 = _mm512_rsqrt14_ps(vx3);
41 const __m512 vrsqrtx4 = _mm512_rsqrt14_ps(vx4);
42
43 __m512 vsqrtx0 = _mm512_mul_ps(vrsqrtx0, vx0);
44 __m512 vhalfrsqrtx0 = _mm512_mul_ps(vrsqrtx0, vhalf);
45 __m512 vsqrtx1 = _mm512_mul_ps(vrsqrtx1, vx1);
46 __m512 vhalfrsqrtx1 = _mm512_mul_ps(vrsqrtx1, vhalf);
47 __m512 vsqrtx2 = _mm512_mul_ps(vrsqrtx2, vx2);
48 __m512 vhalfrsqrtx2 = _mm512_mul_ps(vrsqrtx2, vhalf);
49 __m512 vsqrtx3 = _mm512_mul_ps(vrsqrtx3, vx3);
50 __m512 vhalfrsqrtx3 = _mm512_mul_ps(vrsqrtx3, vhalf);
51 __m512 vsqrtx4 = _mm512_mul_ps(vrsqrtx4, vx4);
52 __m512 vhalfrsqrtx4 = _mm512_mul_ps(vrsqrtx4, vhalf);
53
54 const __m512 vresidual0 = _mm512_fnmadd_ps(vsqrtx0, vhalfrsqrtx0, vhalf);
55 const __m512 vresidual1 = _mm512_fnmadd_ps(vsqrtx1, vhalfrsqrtx1, vhalf);
56 const __m512 vresidual2 = _mm512_fnmadd_ps(vsqrtx2, vhalfrsqrtx2, vhalf);
57 const __m512 vresidual3 = _mm512_fnmadd_ps(vsqrtx3, vhalfrsqrtx3, vhalf);
58 const __m512 vresidual4 = _mm512_fnmadd_ps(vsqrtx4, vhalfrsqrtx4, vhalf);
59
60 vhalfrsqrtx0 = _mm512_fmadd_ps(vhalfrsqrtx0, vresidual0, vhalfrsqrtx0);
61 vsqrtx0 = _mm512_fmadd_ps(vsqrtx0, vresidual0, vsqrtx0);
62 vhalfrsqrtx1 = _mm512_fmadd_ps(vhalfrsqrtx1, vresidual1, vhalfrsqrtx1);
63 vsqrtx1 = _mm512_fmadd_ps(vsqrtx1, vresidual1, vsqrtx1);
64 vhalfrsqrtx2 = _mm512_fmadd_ps(vhalfrsqrtx2, vresidual2, vhalfrsqrtx2);
65 vsqrtx2 = _mm512_fmadd_ps(vsqrtx2, vresidual2, vsqrtx2);
66 vhalfrsqrtx3 = _mm512_fmadd_ps(vhalfrsqrtx3, vresidual3, vhalfrsqrtx3);
67 vsqrtx3 = _mm512_fmadd_ps(vsqrtx3, vresidual3, vsqrtx3);
68 vhalfrsqrtx4 = _mm512_fmadd_ps(vhalfrsqrtx4, vresidual4, vhalfrsqrtx4);
69 vsqrtx4 = _mm512_fmadd_ps(vsqrtx4, vresidual4, vsqrtx4);
70
71 const __m512 vadjustment0 = _mm512_fnmadd_ps(vsqrtx0, vsqrtx0, vx0);
72 const __m512 vadjustment1 = _mm512_fnmadd_ps(vsqrtx1, vsqrtx1, vx1);
73 const __m512 vadjustment2 = _mm512_fnmadd_ps(vsqrtx2, vsqrtx2, vx2);
74 const __m512 vadjustment3 = _mm512_fnmadd_ps(vsqrtx3, vsqrtx3, vx3);
75 const __m512 vadjustment4 = _mm512_fnmadd_ps(vsqrtx4, vsqrtx4, vx4);
76
77 const __m512 vy0 = _mm512_fmadd_ps(vhalfrsqrtx0, vadjustment0, vsqrtx0);
78 const __m512 vy1 = _mm512_fmadd_ps(vhalfrsqrtx1, vadjustment1, vsqrtx1);
79 const __m512 vy2 = _mm512_fmadd_ps(vhalfrsqrtx2, vadjustment2, vsqrtx2);
80 const __m512 vy3 = _mm512_fmadd_ps(vhalfrsqrtx3, vadjustment3, vsqrtx3);
81 const __m512 vy4 = _mm512_fmadd_ps(vhalfrsqrtx4, vadjustment4, vsqrtx4);
82
83 _mm512_storeu_ps(y, vy0);
84 _mm512_storeu_ps(y + 16, vy1);
85 _mm512_storeu_ps(y + 32, vy2);
86 _mm512_storeu_ps(y + 48, vy3);
87 _mm512_storeu_ps(y + 64, vy4);
88 y += 80;
89 }
90 for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
91 const __m512 vx = _mm512_loadu_ps(x);
92 x += 16;
93
94 const __m512 vrsqrtx = _mm512_rsqrt14_ps(vx);
95 __m512 vsqrtx = _mm512_mul_ps(vrsqrtx, vx);
96 __m512 vhalfrsqrtx = _mm512_mul_ps(vrsqrtx, vhalf);
97 const __m512 vresidual = _mm512_fnmadd_ps(vsqrtx, vhalfrsqrtx, vhalf);
98 vhalfrsqrtx = _mm512_fmadd_ps(vhalfrsqrtx, vresidual, vhalfrsqrtx);
99 vsqrtx = _mm512_fmadd_ps(vsqrtx, vresidual, vsqrtx);
100 const __m512 vadjustment = _mm512_fnmadd_ps(vsqrtx, vsqrtx, vx);
101 const __m512 vy = _mm512_fmadd_ps(vhalfrsqrtx, vadjustment, vsqrtx);
102
103 _mm512_storeu_ps(y, vy);
104 y += 16;
105 }
106 if XNN_UNLIKELY(n != 0) {
107 assert(n >= 1 * sizeof(float));
108 assert(n <= 15 * sizeof(float));
109 // Prepare mask for valid 32-bit elements (depends on n).
110 n >>= 2 /* log2(sizeof(float)) */;
111 const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << n) - UINT32_C(1)));
112
113 const __m512 vx = _mm512_maskz_loadu_ps(vmask, x);
114 const __m512 vrsqrtx = _mm512_rsqrt14_ps(vx);
115 __m512 vsqrtx = _mm512_mul_ps(vrsqrtx, vx);
116 __m512 vhalfrsqrtx = _mm512_mul_ps(vrsqrtx, vhalf);
117 const __m512 vresidual = _mm512_fnmadd_ps(vsqrtx, vhalfrsqrtx, vhalf);
118 vhalfrsqrtx = _mm512_fmadd_ps(vhalfrsqrtx, vresidual, vhalfrsqrtx);
119 vsqrtx = _mm512_fmadd_ps(vsqrtx, vresidual, vsqrtx);
120 const __m512 vadjustment = _mm512_fnmadd_ps(vsqrtx, vsqrtx, vx);
121 const __m512 vy = _mm512_fmadd_ps(vhalfrsqrtx, vadjustment, vsqrtx);
122
123 _mm512_mask_storeu_ps(y, vmask, vy);
124 }
125 }
126