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_x64(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_x64(
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 >= 64 * sizeof(float); n -= 64 * 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 x += 64;
35
36 const __m512 vrsqrtx0 = _mm512_rsqrt14_ps(vx0);
37 const __m512 vrsqrtx1 = _mm512_rsqrt14_ps(vx1);
38 const __m512 vrsqrtx2 = _mm512_rsqrt14_ps(vx2);
39 const __m512 vrsqrtx3 = _mm512_rsqrt14_ps(vx3);
40
41 __m512 vsqrtx0 = _mm512_mul_ps(vrsqrtx0, vx0);
42 __m512 vhalfrsqrtx0 = _mm512_mul_ps(vrsqrtx0, vhalf);
43 __m512 vsqrtx1 = _mm512_mul_ps(vrsqrtx1, vx1);
44 __m512 vhalfrsqrtx1 = _mm512_mul_ps(vrsqrtx1, vhalf);
45 __m512 vsqrtx2 = _mm512_mul_ps(vrsqrtx2, vx2);
46 __m512 vhalfrsqrtx2 = _mm512_mul_ps(vrsqrtx2, vhalf);
47 __m512 vsqrtx3 = _mm512_mul_ps(vrsqrtx3, vx3);
48 __m512 vhalfrsqrtx3 = _mm512_mul_ps(vrsqrtx3, vhalf);
49
50 const __m512 vresidual0 = _mm512_fnmadd_ps(vsqrtx0, vhalfrsqrtx0, vhalf);
51 const __m512 vresidual1 = _mm512_fnmadd_ps(vsqrtx1, vhalfrsqrtx1, vhalf);
52 const __m512 vresidual2 = _mm512_fnmadd_ps(vsqrtx2, vhalfrsqrtx2, vhalf);
53 const __m512 vresidual3 = _mm512_fnmadd_ps(vsqrtx3, vhalfrsqrtx3, vhalf);
54
55 vhalfrsqrtx0 = _mm512_fmadd_ps(vhalfrsqrtx0, vresidual0, vhalfrsqrtx0);
56 vsqrtx0 = _mm512_fmadd_ps(vsqrtx0, vresidual0, vsqrtx0);
57 vhalfrsqrtx1 = _mm512_fmadd_ps(vhalfrsqrtx1, vresidual1, vhalfrsqrtx1);
58 vsqrtx1 = _mm512_fmadd_ps(vsqrtx1, vresidual1, vsqrtx1);
59 vhalfrsqrtx2 = _mm512_fmadd_ps(vhalfrsqrtx2, vresidual2, vhalfrsqrtx2);
60 vsqrtx2 = _mm512_fmadd_ps(vsqrtx2, vresidual2, vsqrtx2);
61 vhalfrsqrtx3 = _mm512_fmadd_ps(vhalfrsqrtx3, vresidual3, vhalfrsqrtx3);
62 vsqrtx3 = _mm512_fmadd_ps(vsqrtx3, vresidual3, vsqrtx3);
63
64 const __m512 vadjustment0 = _mm512_fnmadd_ps(vsqrtx0, vsqrtx0, vx0);
65 const __m512 vadjustment1 = _mm512_fnmadd_ps(vsqrtx1, vsqrtx1, vx1);
66 const __m512 vadjustment2 = _mm512_fnmadd_ps(vsqrtx2, vsqrtx2, vx2);
67 const __m512 vadjustment3 = _mm512_fnmadd_ps(vsqrtx3, vsqrtx3, vx3);
68
69 const __m512 vy0 = _mm512_fmadd_ps(vhalfrsqrtx0, vadjustment0, vsqrtx0);
70 const __m512 vy1 = _mm512_fmadd_ps(vhalfrsqrtx1, vadjustment1, vsqrtx1);
71 const __m512 vy2 = _mm512_fmadd_ps(vhalfrsqrtx2, vadjustment2, vsqrtx2);
72 const __m512 vy3 = _mm512_fmadd_ps(vhalfrsqrtx3, vadjustment3, vsqrtx3);
73
74 _mm512_storeu_ps(y, vy0);
75 _mm512_storeu_ps(y + 16, vy1);
76 _mm512_storeu_ps(y + 32, vy2);
77 _mm512_storeu_ps(y + 48, vy3);
78 y += 64;
79 }
80 for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
81 const __m512 vx = _mm512_loadu_ps(x);
82 x += 16;
83
84 const __m512 vrsqrtx = _mm512_rsqrt14_ps(vx);
85 __m512 vsqrtx = _mm512_mul_ps(vrsqrtx, vx);
86 __m512 vhalfrsqrtx = _mm512_mul_ps(vrsqrtx, vhalf);
87 const __m512 vresidual = _mm512_fnmadd_ps(vsqrtx, vhalfrsqrtx, vhalf);
88 vhalfrsqrtx = _mm512_fmadd_ps(vhalfrsqrtx, vresidual, vhalfrsqrtx);
89 vsqrtx = _mm512_fmadd_ps(vsqrtx, vresidual, vsqrtx);
90 const __m512 vadjustment = _mm512_fnmadd_ps(vsqrtx, vsqrtx, vx);
91 const __m512 vy = _mm512_fmadd_ps(vhalfrsqrtx, vadjustment, vsqrtx);
92
93 _mm512_storeu_ps(y, vy);
94 y += 16;
95 }
96 if XNN_UNLIKELY(n != 0) {
97 assert(n >= 1 * sizeof(float));
98 assert(n <= 15 * sizeof(float));
99 // Prepare mask for valid 32-bit elements (depends on n).
100 n >>= 2 /* log2(sizeof(float)) */;
101 const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << n) - UINT32_C(1)));
102
103 const __m512 vx = _mm512_maskz_loadu_ps(vmask, x);
104 const __m512 vrsqrtx = _mm512_rsqrt14_ps(vx);
105 __m512 vsqrtx = _mm512_mul_ps(vrsqrtx, vx);
106 __m512 vhalfrsqrtx = _mm512_mul_ps(vrsqrtx, vhalf);
107 const __m512 vresidual = _mm512_fnmadd_ps(vsqrtx, vhalfrsqrtx, vhalf);
108 vhalfrsqrtx = _mm512_fmadd_ps(vhalfrsqrtx, vresidual, vhalfrsqrtx);
109 vsqrtx = _mm512_fmadd_ps(vsqrtx, vresidual, vsqrtx);
110 const __m512 vadjustment = _mm512_fnmadd_ps(vsqrtx, vsqrtx, vx);
111 const __m512 vy = _mm512_fmadd_ps(vhalfrsqrtx, vadjustment, vsqrtx);
112
113 _mm512_mask_storeu_ps(y, vmask, vy);
114 }
115 }
116