1// Copyright 2020 Google LLC
2//
3// This source code is licensed under the BSD-style license found in the
4// LICENSE file in the root directory of this source tree.
5
6$assert BATCH_TILE % 16 == 0
7$assert BATCH_TILE >= 16
8$SIMD_TILE = BATCH_TILE // 16
9$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
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
19void xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x${BATCH_TILE}(
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  $if BATCH_TILE > 16:
30    for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) {
31      const __m512 vx${ABC[0]} = _mm512_loadu_ps(x);
32      $for N in range(1, SIMD_TILE):
33        const __m512 vx${ABC[N]} = _mm512_loadu_ps(x + ${N * 16});
34      x += ${BATCH_TILE};
35
36      $for N in range(SIMD_TILE):
37        const __m512 vrsqrtx${ABC[N]} = _mm512_rsqrt14_ps(vx${ABC[N]});
38
39      $for N in range(SIMD_TILE):
40        __m512 vsqrtx${ABC[N]} = _mm512_mul_ps(vrsqrtx${ABC[N]}, vx${ABC[N]});
41        __m512 vhalfrsqrtx${ABC[N]} = _mm512_mul_ps(vrsqrtx${ABC[N]}, vhalf);
42
43      $for N in range(SIMD_TILE):
44        const __m512 vresidual${ABC[N]} = _mm512_fnmadd_ps(vsqrtx${ABC[N]}, vhalfrsqrtx${ABC[N]}, vhalf);
45
46      $for N in range(SIMD_TILE):
47        vhalfrsqrtx${ABC[N]} = _mm512_fmadd_ps(vhalfrsqrtx${ABC[N]}, vresidual${ABC[N]}, vhalfrsqrtx${ABC[N]});
48        vsqrtx${ABC[N]} = _mm512_fmadd_ps(vsqrtx${ABC[N]}, vresidual${ABC[N]}, vsqrtx${ABC[N]});
49
50      $for N in range(SIMD_TILE):
51        const __m512 vadjustment${ABC[N]} = _mm512_fnmadd_ps(vsqrtx${ABC[N]}, vsqrtx${ABC[N]}, vx${ABC[N]});
52
53      $for N in range(SIMD_TILE):
54        const __m512 vy${ABC[N]} = _mm512_fmadd_ps(vhalfrsqrtx${ABC[N]}, vadjustment${ABC[N]}, vsqrtx${ABC[N]});
55
56      _mm512_storeu_ps(y, vy${ABC[0]});
57      $for N in range(1, SIMD_TILE):
58        _mm512_storeu_ps(y + ${N * 16}, vy${ABC[N]});
59      y += ${BATCH_TILE};
60    }
61  for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
62    const __m512 vx = _mm512_loadu_ps(x);
63    x += 16;
64
65    const __m512 vrsqrtx = _mm512_rsqrt14_ps(vx);
66    __m512 vsqrtx = _mm512_mul_ps(vrsqrtx, vx);
67    __m512 vhalfrsqrtx = _mm512_mul_ps(vrsqrtx, vhalf);
68    const __m512 vresidual = _mm512_fnmadd_ps(vsqrtx, vhalfrsqrtx, vhalf);
69    vhalfrsqrtx = _mm512_fmadd_ps(vhalfrsqrtx, vresidual, vhalfrsqrtx);
70    vsqrtx = _mm512_fmadd_ps(vsqrtx, vresidual, vsqrtx);
71    const __m512 vadjustment = _mm512_fnmadd_ps(vsqrtx, vsqrtx, vx);
72    const __m512 vy = _mm512_fmadd_ps(vhalfrsqrtx, vadjustment, vsqrtx);
73
74    _mm512_storeu_ps(y, vy);
75    y += 16;
76  }
77  if XNN_UNLIKELY(n != 0) {
78    assert(n >= 1 * sizeof(float));
79    assert(n <= 15 * sizeof(float));
80    // Prepare mask for valid 32-bit elements (depends on n).
81    n >>= 2 /* log2(sizeof(float)) */;
82    const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << n) - UINT32_C(1)));
83
84    const __m512 vx = _mm512_maskz_loadu_ps(vmask, x);
85    const __m512 vrsqrtx = _mm512_rsqrt14_ps(vx);
86    __m512 vsqrtx = _mm512_mul_ps(vrsqrtx, vx);
87    __m512 vhalfrsqrtx = _mm512_mul_ps(vrsqrtx, vhalf);
88    const __m512 vresidual = _mm512_fnmadd_ps(vsqrtx, vhalfrsqrtx, vhalf);
89    vhalfrsqrtx = _mm512_fmadd_ps(vhalfrsqrtx, vresidual, vhalfrsqrtx);
90    vsqrtx = _mm512_fmadd_ps(vsqrtx, vresidual, vsqrtx);
91    const __m512 vadjustment = _mm512_fnmadd_ps(vsqrtx, vsqrtx, vx);
92    const __m512 vy = _mm512_fmadd_ps(vhalfrsqrtx, vadjustment, vsqrtx);
93
94    _mm512_mask_storeu_ps(y, vmask, vy);
95  }
96}
97