// 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_nr2fma1adj_x${BATCH_TILE}(
    size_t n,
    const float* x,
    float* y,
    const union xnn_f32_sqrt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(n != 0);
  assert(n % sizeof(float) == 0);

  const float32x4_t vhalf = vmovq_n_f32(0.5f);
  $if BATCH_TILE > 4:
    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;

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

      $for N in range(0, BATCH_TILE, 4):
        float32x4_t vresidual${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]});

      $for N in range(0, BATCH_TILE, 4):
        vresidual${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;
    }
  for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
    const float32x4_t vx = vld1q_f32(x); x += 4;
    const float32x4_t vrsqrtx = vrsqrteq_f32(vx);
    float32x4_t vsqrtx = vmulq_f32(vrsqrtx, vx);
    float32x4_t vhalfrsqrtx = vmulq_f32(vrsqrtx, vhalf);
    float32x4_t vresidual = vfmsq_f32(vhalf, vsqrtx, vhalfrsqrtx);
    vhalfrsqrtx = vfmaq_f32(vhalfrsqrtx, vresidual, vhalfrsqrtx);
    vsqrtx = vfmaq_f32(vsqrtx, vresidual, vsqrtx);
    vresidual = vfmsq_f32(vhalf, vsqrtx, vhalfrsqrtx);
    vhalfrsqrtx = vfmaq_f32(vhalfrsqrtx, vresidual, vhalfrsqrtx);
    vsqrtx = vfmaq_f32(vsqrtx, vresidual, vsqrtx);
    const float32x4_t vadjustment = vfmsq_f32(vx, vsqrtx, vsqrtx);
    const float32x4_t vy = vfmaq_f32(vsqrtx, vhalfrsqrtx, vadjustment);
    vst1q_f32(y, vy); y += 4;
  }
  if XNN_UNLIKELY(n != 0) {
    const float32x4_t vx = vld1q_f32(x);
    const float32x4_t vrsqrtx = vrsqrteq_f32(vx);
    float32x4_t vsqrtx = vmulq_f32(vrsqrtx, vx);
    float32x4_t vhalfrsqrtx = vmulq_f32(vrsqrtx, vhalf);
    float32x4_t vresidual = vfmsq_f32(vhalf, vsqrtx, vhalfrsqrtx);
    vhalfrsqrtx = vfmaq_f32(vhalfrsqrtx, vresidual, vhalfrsqrtx);
    vsqrtx = vfmaq_f32(vsqrtx, vresidual, vsqrtx);
    vresidual = vfmsq_f32(vhalf, vsqrtx, vhalfrsqrtx);
    vhalfrsqrtx = vfmaq_f32(vhalfrsqrtx, vresidual, vhalfrsqrtx);
    vsqrtx = vfmaq_f32(vsqrtx, vresidual, vsqrtx);
    const float32x4_t vadjustment = vfmsq_f32(vx, vsqrtx, vsqrtx);
    const float32x4_t vy = vfmaq_f32(vsqrtx, vhalfrsqrtx, vadjustment);

    float32x2_t vy_lo = vget_low_f32(vy);
    if (n & (2 * sizeof(float))) {
      vst1_f32(y, vy_lo); y += 2;
      vy_lo = vget_high_f32(vy);
    }
    if (n & (1 * sizeof(float))) {
      vst1_lane_f32(y, vy_lo, 0);
    }
  }
}
