// 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
#include <assert.h>

#include <arm_neon.h>

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


void xnn_f16_velu_ukernel__neonfp16arith_rr1_p3_x${BATCH_TILE}(
    size_t batch,
    const void* input,
    void* output,
    const union xnn_f16_elu_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 vprescale = vreinterpretq_f16_u16(vld1q_dup_u16(&params->fp16arith_rr1_p3.prescale));
  const float16x8_t vsat_cutoff = vreinterpretq_f16_u16(vld1q_dup_u16(&params->fp16arith_rr1_p3.sat_cutoff));
  const float16x8_t vmagic_bias = vreinterpretq_f16_u16(vld1q_dup_u16(&params->fp16arith_rr1_p3.magic_bias));
  const float16x8_t vlog2e = vreinterpretq_f16_u16(vld1q_dup_u16(&params->fp16arith_rr1_p3.log2e));
  const float16x8_t vminus_ln2 = vreinterpretq_f16_u16(vld1q_dup_u16(&params->fp16arith_rr1_p3.minus_ln2));
  const float16x8_t vc3 = vreinterpretq_f16_u16(vld1q_dup_u16(&params->fp16arith_rr1_p3.c3));
  const float16x8_t vc2 = vreinterpretq_f16_u16(vld1q_dup_u16(&params->fp16arith_rr1_p3.c2));
  const float16x8_t vminus_alpha = vreinterpretq_f16_u16(vld1q_dup_u16(&params->fp16arith_rr1_p3.minus_alpha));
  const float16x8_t vbeta = vreinterpretq_f16_u16(vld1q_dup_u16(&params->fp16arith_rr1_p3.beta));

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

      $for N in range(SIMD_TILE):
        float16x8_t vz${N} = vmulq_f16(vx${N}, vprescale);

      $for N in range(SIMD_TILE):
        vz${N} = vmaxq_f16(vz${N}, vsat_cutoff);

      $for N in range(SIMD_TILE):
        float16x8_t vn${N} = vfmaq_f16(vmagic_bias, vz${N}, vlog2e);

      $for N in range(SIMD_TILE):
        float16x8_t vs${N} = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn${N}), 10));
        vn${N} = vsubq_f16(vn${N}, vmagic_bias);

      $for N in range(SIMD_TILE):
        float16x8_t vt${N} = vfmaq_f16(vz${N}, vn${N}, vminus_ln2);

      $for N in range(SIMD_TILE):
        float16x8_t vp${N} = vfmaq_f16(vc2, vc3, vt${N});
        vp${N} = vmulq_f16(vp${N}, vt${N});

      $for N in range(SIMD_TILE):
        vt${N} = vmulq_f16(vt${N}, vs${N});
        vs${N} = vfmsq_f16(vminus_alpha, vs${N}, vminus_alpha);

      $for N in range(SIMD_TILE):
        vp${N} = vfmaq_f16(vt${N}, vp${N}, vt${N});

      $for N in range(SIMD_TILE):
        float16x8_t ve${N} = vfmsq_f16(vs${N}, vp${N}, vminus_alpha);
        const uint16x8_t vm${N} = vcltq_s16(vreinterpretq_s16_f16(vx${N}), vmovq_n_s16(0));

      $for N in range(SIMD_TILE):
        vx${N} = vmulq_f16(vx${N}, vbeta);

      $for N in range(SIMD_TILE):
        const float16x8_t vy${N} = vbslq_f16(vm${N}, ve${N}, vx${N});

      $for N in range(SIMD_TILE):
        vst1q_u16(o, vreinterpretq_u16_f16(vy${N})); o += 8;
    }
  for (; batch >= 8 * sizeof(uint16_t); batch -= 8 * sizeof(uint16_t)) {
    float16x8_t vx = vreinterpretq_f16_u16(vld1q_u16(i)); i += 8;
    float16x8_t vz = vmulq_f16(vx, vprescale);
    vz = vmaxq_f16(vz, vsat_cutoff);

    float16x8_t vn = vfmaq_f16(vmagic_bias, vz, vlog2e);
    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, vminus_ln2);

    float16x8_t vp = vfmaq_f16(vc2, vc3, vt);
    vp = vmulq_f16(vp, vt);
    vt = vmulq_f16(vt, vs);
    vs = vfmsq_f16(vminus_alpha, vs, vminus_alpha);
    vp = vfmaq_f16(vt, vp, vt);
    float16x8_t ve = vfmsq_f16(vs, vp, vminus_alpha);

    const uint16x8_t vm = vcltq_s16(vreinterpretq_s16_f16(vx), vmovq_n_s16(0));
    vx = vmulq_f16(vx, vbeta);
    const float16x8_t vy = vbslq_f16(vm, ve, vx);
    vst1q_u16(o, vreinterpretq_u16_f16(vy)); o += 8;
  }
  if XNN_UNLIKELY(batch != 0) {
    float16x8_t vx = vreinterpretq_f16_u16(vld1q_u16(i)); i += 8;
    float16x8_t vz = vmulq_f16(vx, vprescale);
    vz = vmaxq_f16(vz, vsat_cutoff);

    float16x8_t vn = vfmaq_f16(vmagic_bias, vz, vlog2e);
    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, vminus_ln2);

    float16x8_t vp = vfmaq_f16(vc2, vc3, vt);
    vp = vmulq_f16(vp, vt);
    vt = vmulq_f16(vt, vs);
    vs = vfmsq_f16(vminus_alpha, vs, vminus_alpha);
    vp = vfmaq_f16(vt, vp, vt);
    float16x8_t ve = vfmsq_f16(vs, vp, vminus_alpha);

    const uint16x8_t vm = vcltq_s16(vreinterpretq_s16_f16(vx), vmovq_n_s16(0));
    vx = vmulq_f16(vx, vbeta);
    float16x8_t vy = vbslq_f16(vm, ve, vx);
    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);
    }
  }
}