// 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
$assert DIV_ALGO in ["DIV", "NR1FMA", "NR1RECPS"]
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
#include <assert.h>

#include <arm_neon.h>

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


$ISA = "aarch64_neonfp16arith" if DIV_ALGO == "DIV" else "neonfp16arith"
void xnn_f16_vsigmoid_ukernel__${ISA}_rr2_p2_${DIV_ALGO.lower()}_x${BATCH_TILE}(
    size_t batch,
    const void* input,
    void* output,
    const union xnn_f16_sigmoid_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 float16x8_t vmagic_bias = vreinterpretq_f16_u16(vld1q_dup_u16(&params->fp16arith_rr2_p2.magic_bias));
  const float16x8_t vminus_log2e = vreinterpretq_f16_u16(vld1q_dup_u16(&params->fp16arith_rr2_p2.minus_log2e));
  const float16x8_t vln2_hi = vreinterpretq_f16_u16(vld1q_dup_u16(&params->fp16arith_rr2_p2.ln2_hi));
  const float16x8_t vln2_lo = vreinterpretq_f16_u16(vld1q_dup_u16(&params->fp16arith_rr2_p2.ln2_lo));
  const float16x8_t vc2 = vreinterpretq_f16_u16(vld1q_dup_u16(&params->fp16arith_rr2_p2.c2));
  const float16x8_t vc1 = vreinterpretq_f16_u16(vld1q_dup_u16(&params->fp16arith_rr2_p2.c1));
  const float16x8_t vone = vreinterpretq_f16_u16(vmovq_n_u16(UINT16_C(0x3C00)));  // 1.0h
  const float16x8_t vdenorm_cutoff = vreinterpretq_f16_u16(vld1q_dup_u16(&params->fp16arith_rr2_p2.denorm_cutoff));

  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)) {
      $for N in range(SIMD_TILE):
        const float16x8_t vx${ABC[N]} = vreinterpretq_f16_u16(vld1q_u16(i)); i += 8;

      $for N in range(SIMD_TILE):
        const float16x8_t vz${ABC[N]} = vabsq_f16(vx${ABC[N]});

      $for N in range(SIMD_TILE):
        float16x8_t vn${ABC[N]} = vfmaq_f16(vmagic_bias, vz${ABC[N]}, vminus_log2e);

      $for N in range(SIMD_TILE):
        const float16x8_t vs${ABC[N]} = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn${ABC[N]}), 10));

      $for N in range(SIMD_TILE):
        vn${ABC[N]} = vsubq_f16(vn${ABC[N]}, vmagic_bias);

      $for N in range(SIMD_TILE):
        float16x8_t vt${ABC[N]} = vfmaq_f16(vz${ABC[N]}, vn${ABC[N]}, vln2_hi);

      $for N in range(SIMD_TILE):
        vt${ABC[N]} = vfmaq_f16(vt${ABC[N]}, vn${ABC[N]}, vln2_lo);

      $for N in range(SIMD_TILE):
        const float16x8_t vp${ABC[N]} = vfmaq_f16(vc1, vc2, vt${ABC[N]});

      $for N in range(SIMD_TILE):
        vt${ABC[N]} = vmulq_f16(vt${ABC[N]}, vs${ABC[N]});

      $for N in range(SIMD_TILE):
        const float16x8_t ve${ABC[N]} = vfmaq_f16(vs${ABC[N]}, vp${ABC[N]}, vt${ABC[N]});

      $for N in range(SIMD_TILE):
        const float16x8_t vd${ABC[N]} = vaddq_f16(ve${ABC[N]}, vone);

      $if DIV_ALGO == "DIV":
        $for N in range(SIMD_TILE):
          float16x8_t vf${ABC[N]} = vdivq_f16(ve${ABC[N]}, vd${ABC[N]});
      $else:
        $for N in range(SIMD_TILE):
          float16x8_t vr${ABC[N]} = vrecpeq_f16(vd${ABC[N]});

        $if DIV_ALGO == "NR1FMA":
          $for N in range(SIMD_TILE):
            const float16x8_t vadj${ABC[N]} = vfmsq_f16(vone, vr${N}, vd${N});

          $for N in range(SIMD_TILE):
            vr${ABC[N]} = vfmaq_f16(vr${ABC[N]}, vr${ABC[N]}, vadj${ABC[N]});
        $else:
          $for N in range(SIMD_TILE):
            const float16x8_t vadj${ABC[N]} = vrecpsq_f16(vr${ABC[N]}, vd${ABC[N]});

          $for N in range(SIMD_TILE):
            vr${ABC[N]} = vmulq_f16(vr${ABC[N]}, vadj${ABC[N]});

        $for N in range(SIMD_TILE):
          float16x8_t vf${ABC[N]} = vmulq_f16(ve${ABC[N]}, vr${ABC[N]});

      $for N in range(SIMD_TILE):
        vf${ABC[N]} = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf${ABC[N]}), vcagtq_f16(vx${ABC[N]}, vdenorm_cutoff)));

      $for N in range(SIMD_TILE):
        const uint16x8_t vm${ABC[N]} = vcltq_f16(vx${ABC[N]}, vreinterpretq_f16_u16(vmovq_n_u16(0)));

      $for N in range(SIMD_TILE):
        vf${ABC[N]} = vbslq_f16(vm${ABC[N]}, vf${ABC[N]}, vsubq_f16(vone, vf${ABC[N]}));

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

    const float16x8_t vz = vabsq_f16(vx);

    float16x8_t vn = vfmaq_f16(vmagic_bias, vz, vminus_log2e);
    const float16x8_t vs = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn), 10));
    vn = vsubq_f16(vn, vmagic_bias);

    float16x8_t vt = vfmaq_f16(vz, vn, vln2_hi);
    vt = vfmaq_f16(vt, vn, vln2_lo);

    const float16x8_t vp = vfmaq_f16(vc1, vc2, vt);
    vt = vmulq_f16(vt, vs);
    const float16x8_t ve = vfmaq_f16(vs, vp, vt);
    const float16x8_t vd = vaddq_f16(ve, vone);

    $if DIV_ALGO == "DIV":
      float16x8_t vf = vdivq_f16(ve, vd);
    $else:
      float16x8_t vr = vrecpeq_f16(vd);
      $if DIV_ALGO == "NR1FMA":
        const float16x8_t vadj = vfmsq_f16(vone, vr, vd);
        vr = vfmaq_f16(vr, vr, vadj);
      $else:
        const float16x8_t vadj = vrecpsq_f16(vr, vd);
        vr = vmulq_f16(vr, vadj);

      float16x8_t vf = vmulq_f16(ve, vr);
    vf = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf), vcagtq_f16(vx, vdenorm_cutoff)));
    const uint16x8_t vm = vcltq_f16(vx, vreinterpretq_f16_u16(vmovq_n_u16(0)));
    vf = vbslq_f16(vm, vf, vsubq_f16(vone, vf));

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

    const float16x8_t vz = vabsq_f16(vx);

    float16x8_t vn = vfmaq_f16(vmagic_bias, vz, vminus_log2e);
    const float16x8_t vs = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn), 10));
    vn = vsubq_f16(vn, vmagic_bias);

    float16x8_t vt = vfmaq_f16(vz, vn, vln2_hi);
    vt = vfmaq_f16(vt, vn, vln2_lo);

    const float16x8_t vp = vfmaq_f16(vc1, vc2, vt);
    vt = vmulq_f16(vt, vs);
    const float16x8_t ve = vfmaq_f16(vs, vp, vt);
    const float16x8_t vd = vaddq_f16(ve, vone);

    $if DIV_ALGO == "DIV":
      float16x8_t vf = vdivq_f16(ve, vd);
    $else:
      float16x8_t vr = vrecpeq_f16(vd);
      $if DIV_ALGO == "NR1FMA":
        const float16x8_t vadj = vfmsq_f16(vone, vr, vd);
        vr = vfmaq_f16(vr, vr, vadj);
      $else:
        const float16x8_t vadj = vrecpsq_f16(vr, vd);
        vr = vmulq_f16(vr, vadj);

      float16x8_t vf = vmulq_f16(ve, vr);
    vf = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf), vcagtq_f16(vx, vdenorm_cutoff)));
    const uint16x8_t vm = vcltq_f16(vx, vreinterpretq_f16_u16(vmovq_n_u16(0)));
    vf = vbslq_f16(vm, vf, vsubq_f16(vone, vf));

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