// Copyright 2022 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 % 8 == 0
$assert BATCH_TILE >= 8
$SIMD_TILE = BATCH_TILE // 8
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
#include <assert.h>

#include <arm_neon.h>

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


void xnn_f16_vsqrt_ukernel__neonfp16arith_nr1fma1adj_x${BATCH_TILE}(
    size_t batch,
    const void* input,
    void* output,
    const union xnn_f16_sqrt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(batch != 0);
  assert(batch % sizeof(uint16_t) == 0);
  assert(input != NULL);
  assert(output != NULL);

  const uint16x8_t vpositive_infinity = vmovq_n_u16(UINT16_C(0x7C00));
  const float16x8_t vhalf = vreinterpretq_f16_u16(vmovq_n_u16(UINT16_C(0x3800)));  // 0.5h
  const uint16x8_t vexp4_mask = vmovq_n_u16(UINT16_C(0x7800));

  const uint16_t* i = (const uint16_t*) input;
  uint16_t* o = (uint16_t*) output;
  $if BATCH_TILE > 8:
    for (; batch >= ${BATCH_TILE} * sizeof(uint16_t); batch -= ${BATCH_TILE} * sizeof(uint16_t)) {
      $for N in range(SIMD_TILE):
        float16x8_t vi${ABC[N]} = vreinterpretq_f16_u16(vld1q_u16(i)); i += 8;

      $for N in range(SIMD_TILE):
        const float16x8_t vx${N} = vbslq_f16(vexp4_mask, vhalf, vi${N});
        const int16x8_t vexp4i${N} = vreinterpretq_s16_u16(vandq_u16(vreinterpretq_u16_f16(vi${N}), vexp4_mask));

      $for N in range(SIMD_TILE):
        const float16x8_t vrsqrtx${N} = vrsqrteq_f16(vx${N});
        const int16x8_t vpostscale${N} = vhsubq_s16(vexp4i${N}, vreinterpretq_s16_f16(vhalf));

      $for N in range(SIMD_TILE):
        float16x8_t vsqrtx${N} = vmulq_f16(vrsqrtx${N}, vx${N});
        const float16x8_t vhalfrsqrtx${N} = vmulq_f16(vrsqrtx${N}, vhalf);
        uint16x8_t vspecial_mask${N} = vcgeq_u16(vreinterpretq_u16_f16(vi${N}), vpositive_infinity);

      $for N in range(SIMD_TILE):
        const float16x8_t vresidual${N} = vfmsq_f16(vhalf, vsqrtx${N}, vhalfrsqrtx${N});
        const uint16x8_t vzero_mask${N} = vceqq_f16(vi${N}, vreinterpretq_f16_u16(vmovq_n_u16(0)));
        uint16x8_t vspecial_value${N} = vmovq_n_u16(UINT16_C(0x7E00));

      $for N in range(SIMD_TILE):
        vsqrtx${N} = vfmaq_f16(vsqrtx${N}, vresidual${N}, vsqrtx${N});
        vspecial_mask${N} = vorrq_u16(vspecial_mask${N}, vzero_mask${N});
        const uint16x8_t vinfinity_mask${N} = vceqq_u16(vreinterpretq_u16_f16(vi${N}), vpositive_infinity);

      $for N in range(SIMD_TILE):
        const float16x8_t vadjustment${N} = vfmsq_f16(vx${N}, vsqrtx${N}, vsqrtx${N});
        const uint16x8_t vinput_mask${N} = vorrq_u16(vinfinity_mask${N}, vzero_mask${N});

      $for N in range(SIMD_TILE):
        vsqrtx${N} = vfmaq_f16(vsqrtx${N}, vhalfrsqrtx${N}, vadjustment${N});
        vspecial_value${N} = vbslq_u16(vinput_mask${N}, vreinterpretq_u16_f16(vi${N}), vspecial_value${N});

      $for N in range(SIMD_TILE):
        float16x8_t vy${N} = vreinterpretq_f16_s16(vaddq_s16(vreinterpretq_s16_f16(vsqrtx${N}), vpostscale${N}));

      $for N in range(SIMD_TILE):
        vy${N} = vbslq_f16(vspecial_mask${N}, vreinterpretq_f16_u16(vspecial_value${N}), vy${N});

      $for N in range(SIMD_TILE):
        vst1q_u16(o, vreinterpretq_u16_f16(vy${ABC[N]})); o += 8;
    }
  for (; batch >= 8 * sizeof(uint16_t); batch -= 8 * sizeof(uint16_t)) {
    const float16x8_t vi = vreinterpretq_f16_u16(vld1q_u16(i)); i += 8;

    const float16x8_t vx = vbslq_f16(vexp4_mask, vhalf, vi);
    const int16x8_t vexp4i = vreinterpretq_s16_u16(vandq_u16(vreinterpretq_u16_f16(vi), vexp4_mask));

    const float16x8_t vrsqrtx = vrsqrteq_f16(vx);
    const int16x8_t vpostscale = vhsubq_s16(vexp4i, vreinterpretq_s16_f16(vhalf));

    float16x8_t vsqrtx = vmulq_f16(vrsqrtx, vx);
    const float16x8_t vhalfrsqrtx = vmulq_f16(vrsqrtx, vhalf);
    uint16x8_t vspecial_mask = vcgeq_u16(vreinterpretq_u16_f16(vi), vpositive_infinity);

    const float16x8_t vresidual = vfmsq_f16(vhalf, vsqrtx, vhalfrsqrtx);
    const uint16x8_t vzero_mask = vceqq_f16(vi, vreinterpretq_f16_u16(vmovq_n_u16(0)));
    uint16x8_t vspecial_value = vmovq_n_u16(UINT16_C(0x7E00));

    vsqrtx = vfmaq_f16(vsqrtx, vresidual, vsqrtx);
    vspecial_mask = vorrq_u16(vspecial_mask, vzero_mask);
    const uint16x8_t vinfinity_mask = vceqq_u16(vreinterpretq_u16_f16(vi), vpositive_infinity);

    const float16x8_t vadjustment = vfmsq_f16(vx, vsqrtx, vsqrtx);
    const uint16x8_t vinput_mask = vorrq_u16(vinfinity_mask, vzero_mask);

    vsqrtx = vfmaq_f16(vsqrtx, vhalfrsqrtx, vadjustment);
    vspecial_value = vbslq_u16(vinput_mask, vreinterpretq_u16_f16(vi), vspecial_value);

    float16x8_t vy = vreinterpretq_f16_s16(vaddq_s16(vreinterpretq_s16_f16(vsqrtx), vpostscale));

    vy = vbslq_f16(vspecial_mask, vreinterpretq_f16_u16(vspecial_value), vy);

    vst1q_u16(o, vreinterpretq_u16_f16(vy)); o += 8;
  }
  if XNN_UNLIKELY(batch != 0) {
    const float16x8_t vi = vreinterpretq_f16_u16(vld1q_u16(i));

    const float16x8_t vx = vbslq_f16(vexp4_mask, vhalf, vi);
    const int16x8_t vexp4i = vreinterpretq_s16_u16(vandq_u16(vreinterpretq_u16_f16(vi), vexp4_mask));

    const float16x8_t vrsqrtx = vrsqrteq_f16(vx);
    const int16x8_t vpostscale = vhsubq_s16(vexp4i, vreinterpretq_s16_f16(vhalf));

    float16x8_t vsqrtx = vmulq_f16(vrsqrtx, vx);
    const float16x8_t vhalfrsqrtx = vmulq_f16(vrsqrtx, vhalf);
    uint16x8_t vspecial_mask = vcgeq_u16(vreinterpretq_u16_f16(vi), vpositive_infinity);

    const float16x8_t vresidual = vfmsq_f16(vhalf, vsqrtx, vhalfrsqrtx);
    const uint16x8_t vzero_mask = vceqq_f16(vi, vreinterpretq_f16_u16(vmovq_n_u16(0)));
    uint16x8_t vspecial_value = vmovq_n_u16(UINT16_C(0x7E00));

    vsqrtx = vfmaq_f16(vsqrtx, vresidual, vsqrtx);
    vspecial_mask = vorrq_u16(vspecial_mask, vzero_mask);
    const uint16x8_t vinfinity_mask = vceqq_u16(vreinterpretq_u16_f16(vi), vpositive_infinity);

    const float16x8_t vadjustment = vfmsq_f16(vx, vsqrtx, vsqrtx);
    const uint16x8_t vinput_mask = vorrq_u16(vinfinity_mask, vzero_mask);

    vsqrtx = vfmaq_f16(vsqrtx, vhalfrsqrtx, vadjustment);
    vspecial_value = vbslq_u16(vinput_mask, vreinterpretq_u16_f16(vi), vspecial_value);

    float16x8_t vy = vreinterpretq_f16_s16(vaddq_s16(vreinterpretq_s16_f16(vsqrtx), vpostscale));

    vy = vbslq_f16(vspecial_mask, vreinterpretq_f16_u16(vspecial_value), vy);

    float16x4_t vy_lo = vget_low_f16(vy);
    if (batch & (4 * sizeof(uint16_t))) {
      vst1_u16(o, vreinterpret_u16_f16(vy_lo)); o += 4;
      vy_lo = vget_high_f16(vy);
    }
    if (batch & (2 * sizeof(uint16_t))) {
      vst1_lane_u32((void*) o, vreinterpret_u32_f16(vy_lo), 0); o += 2;
      vy_lo = vext_f16(vy_lo, vy_lo, 2);
    }
    if (batch & (1 * sizeof(uint16_t))) {
      vst1_lane_u16(o, vreinterpret_u16_f16(vy_lo), 0);
    }
  }
}