// Copyright 2020 Google LLC
//
// This source code is licensed under the BSD-style license found in the
// LICENSE file in the root directory of this source tree.

$assert BATCH_TILE % 4 == 0
$assert BATCH_TILE >= 4
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
#include <assert.h>
#include <math.h>

#include <arm_neon.h>

#include <xnnpack/common.h>
#include <xnnpack/vunary.h>


void xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x${BATCH_TILE}(
    size_t n,
    const float* x,
    float* y,
    const union xnn_f32_sqrt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
{
  assert(n != 0);
  assert(n % sizeof(float) == 0);

  $if BATCH_TILE > 4:
    const float32x4_t vhalf = vmovq_n_f32(0.5f);
    for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) {
      $for N in range(0, BATCH_TILE, 4):
        const float32x4_t vx${ABC[N:N+4]} = vld1q_f32(x); x += 4;

      $for N in range(0, BATCH_TILE, 4):
        float32x4_t vrsqrtx${ABC[N:N+4]} = vrsqrteq_f32(vx${ABC[N:N+4]});

      $for N in range(0, BATCH_TILE, 4):
        const float32x4_t vrx${ABC[N:N+4]} = vmulq_f32(vrsqrtx${ABC[N:N+4]}, vrsqrtx${ABC[N:N+4]});

      $for N in range(0, BATCH_TILE, 4):
        const float32x4_t vcorrection${ABC[N:N+4]} = vrsqrtsq_f32(vx${ABC[N:N+4]}, vrx${ABC[N:N+4]});

      $for N in range(0, BATCH_TILE, 4):
        vrsqrtx${ABC[N:N+4]} = vmulq_f32(vrsqrtx${ABC[N:N+4]}, vcorrection${ABC[N:N+4]});

      $for N in range(0, BATCH_TILE, 4):
        float32x4_t vsqrtx${ABC[N:N+4]} = vmulq_f32(vrsqrtx${ABC[N:N+4]}, vx${ABC[N:N+4]});
        float32x4_t vhalfrsqrtx${ABC[N:N+4]} = vmulq_f32(vrsqrtx${ABC[N:N+4]}, vhalf);

      $for N in range(0, BATCH_TILE, 4):
        const float32x4_t vresidual${ABC[N:N+4]} = vfmsq_f32(vhalf, vsqrtx${ABC[N:N+4]}, vhalfrsqrtx${ABC[N:N+4]});

      $for N in range(0, BATCH_TILE, 4):
        vhalfrsqrtx${ABC[N:N+4]} = vfmaq_f32(vhalfrsqrtx${ABC[N:N+4]}, vresidual${ABC[N:N+4]}, vhalfrsqrtx${ABC[N:N+4]});
        vsqrtx${ABC[N:N+4]} = vfmaq_f32(vsqrtx${ABC[N:N+4]}, vresidual${ABC[N:N+4]}, vsqrtx${ABC[N:N+4]});

      $for N in range(0, BATCH_TILE, 4):
        const float32x4_t vadjustment${ABC[N:N+4]} = vfmsq_f32(vx${ABC[N:N+4]}, vsqrtx${ABC[N:N+4]}, vsqrtx${ABC[N:N+4]});

      $for N in range(0, BATCH_TILE, 4):
        const float32x4_t vy${ABC[N:N+4]} = vfmaq_f32(vsqrtx${ABC[N:N+4]}, vhalfrsqrtx${ABC[N:N+4]}, vadjustment${ABC[N:N+4]});

      $for N in range(0, BATCH_TILE, 4):
        vst1q_f32(y, vy${ABC[N:N+4]}); y += 4;
    }
  if XNN_UNLIKELY(n != 0) {
    do {
      const float vx = *x++;
      const float vy = sqrtf(vx);
      *y++ = vy;
      n -= sizeof(float);
    } while (n != 0);
  }
}