// 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"
$VMULADDQ_F32 = "vfmaq_f32" if FMA else "vmlaq_f32"
#include <assert.h>

#include <arm_neon.h>

#include "xnnpack/vunary.h"
#include "xnnpack/common.h"


void xnn_f32_velu_ukernel__${"neonfma" if FMA else "neon"}_rr${1 if FMA else 2}_p6_u${BATCH_TILE}(
    size_t batch,
    const float* input,
    float* output,
    const struct xnn_f32_elu_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(batch != 0);
  assert(batch % sizeof(float) == 0);
  assert(input != NULL);
  assert(output != NULL);

  const float32x4_t vsat_cutoff = vmovq_n_f32(-0x1.154246p+4f);
  const float32x4_t vmagic_bias = vmovq_n_f32(0x1.8000FEp23f);
  const float32x4_t vlog2e = vmovq_n_f32(0x1.715476p+0f);
  const float32x4_t vc6 = vmovq_n_f32(0x1.6b7338p-10f);
  const float32x4_t vc5 = vmovq_n_f32(0x1.12278Ep-7f);
  const float32x4_t vc4 = vmovq_n_f32(0x1.555716p-5f);
  const float32x4_t vc3 = vmovq_n_f32(0x1.5554B0p-3f);
  const float32x4_t vc2 = vmovq_n_f32(0x1.FFFFFEp-2f);
  const float32x4_t vone = vmovq_n_f32(1.0f);

  XNN_FORCE_REALIZATION(vsat_cutoff);
  XNN_FORCE_REALIZATION(vmagic_bias);
  XNN_FORCE_REALIZATION(vlog2e);
  XNN_FORCE_REALIZATION(vc6);
  XNN_FORCE_REALIZATION(vc5);
  XNN_FORCE_REALIZATION(vc4);
  XNN_FORCE_REALIZATION(vc3);
  XNN_FORCE_REALIZATION(vc2);
  XNN_FORCE_REALIZATION(vone);

  $if FMA:
    const float32x4_t vminus_ln2 = vmovq_n_f32(-0x1.62E430p-1f);
    XNN_FORCE_REALIZATION(vminus_ln2);
  $else:
    const float32x4_t vminus_ln2_hi = vmovq_n_f32(-0x1.62E440p-1f);
    const float32x4_t vminus_ln2_lo = vmovq_n_f32(0x1.0105C6p-21f);
    XNN_FORCE_REALIZATION(vminus_ln2_hi);
    XNN_FORCE_REALIZATION(vminus_ln2_lo);

  const float32x4_t vprescale = vld1q_dup_f32(&params->scalar.prescale);
  const float32x4_t valpha = vld1q_dup_f32(&params->scalar.alpha);
  const float32x4_t vbeta = vld1q_dup_f32(&params->scalar.beta);

  $if BATCH_TILE > 4:
    for (; batch >= ${BATCH_TILE} * sizeof(float); batch -= ${BATCH_TILE} * sizeof(float)) {
      $for N in range(0, BATCH_TILE, 4):
        float32x4_t vx${ABC[N:N+4]} = vld1q_f32(input); input += 4;

      $for N in range(0, BATCH_TILE, 4):
        const float32x4_t vz${ABC[N:N+4]} = vmaxq_f32(vmulq_f32(vx${ABC[N:N+4]}, vprescale), vsat_cutoff);

      $for N in range(0, BATCH_TILE, 4):
        float32x4_t vn${ABC[N:N+4]} = ${VMULADDQ_F32}(vmagic_bias, vz${ABC[N:N+4]}, vlog2e);

      $for N in range(0, BATCH_TILE, 4):
        float32x4_t vs${ABC[N:N+4]} = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn${ABC[N:N+4]}), 23));
        vn${ABC[N:N+4]} = vsubq_f32(vn${ABC[N:N+4]}, vmagic_bias);

      $if FMA:
        $for N in range(0, BATCH_TILE, 4):
          float32x4_t vt${ABC[N:N+4]} = ${VMULADDQ_F32}(vz${ABC[N:N+4]}, vn${ABC[N:N+4]}, vminus_ln2);
      $else:
        $for N in range(0, BATCH_TILE, 4):
          float32x4_t vt${ABC[N:N+4]} = ${VMULADDQ_F32}(vz${ABC[N:N+4]}, vn${ABC[N:N+4]}, vminus_ln2_hi);

        $for N in range(0, BATCH_TILE, 4):
          vt${ABC[N:N+4]} = ${VMULADDQ_F32}(vt${ABC[N:N+4]}, vn${ABC[N:N+4]}, vminus_ln2_lo);

      $for N in range(0, BATCH_TILE, 4):
        float32x4_t vp${ABC[N:N+4]} = ${VMULADDQ_F32}(vc5, vc6, vt${ABC[N:N+4]});

      $for N in range(0, BATCH_TILE, 4):
        vp${ABC[N:N+4]} = ${VMULADDQ_F32}(vc4, vp${ABC[N:N+4]}, vt${ABC[N:N+4]});

      $for N in range(0, BATCH_TILE, 4):
        vp${ABC[N:N+4]} = ${VMULADDQ_F32}(vc3, vp${ABC[N:N+4]}, vt${ABC[N:N+4]});

      $for N in range(0, BATCH_TILE, 4):
        vp${ABC[N:N+4]} = ${VMULADDQ_F32}(vc2, vp${ABC[N:N+4]}, vt${ABC[N:N+4]});

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

      $for N in range(0, BATCH_TILE, 4):
        vt${ABC[N:N+4]} = vmulq_f32(vt${ABC[N:N+4]}, vs${ABC[N:N+4]});
        vs${ABC[N:N+4]} = vsubq_f32(vs${ABC[N:N+4]}, vone);

      $for N in range(0, BATCH_TILE, 4):
        vp${ABC[N:N+4]} = ${VMULADDQ_F32}(vt${ABC[N:N+4]}, vp${ABC[N:N+4]}, vt${ABC[N:N+4]});

      $for N in range(0, BATCH_TILE, 4):
        const float32x4_t ve${ABC[N:N+4]} = vmulq_f32(vaddq_f32(vp${ABC[N:N+4]}, vs${ABC[N:N+4]}), valpha);

      $for N in range(0, BATCH_TILE, 4):
        const uint32x4_t vm${ABC[N:N+4]} = vcltq_f32(vx${ABC[N:N+4]}, vmovq_n_f32(0.0f));
        vx${ABC[N:N+4]} = vmulq_f32(vx${ABC[N:N+4]}, vbeta);

      $for N in range(0, BATCH_TILE, 4):
        const float32x4_t vy${ABC[N:N+4]} = vbslq_f32(vm${ABC[N:N+4]}, ve${ABC[N:N+4]}, vx${ABC[N:N+4]});

      $for N in range(0, BATCH_TILE, 4):
        vst1q_f32(output, vy${ABC[N:N+4]}); output += 4;
    }
  for (; batch >= 4 * sizeof(float); batch -= 4 * sizeof(float)) {
    float32x4_t vx = vld1q_f32(input); input += 4;

    const float32x4_t vz = vmaxq_f32(vmulq_f32(vx, vprescale), vsat_cutoff);

    float32x4_t vn = ${VMULADDQ_F32}(vmagic_bias, vz, vlog2e);
    float32x4_t vs = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn), 23));
    vn = vsubq_f32(vn, vmagic_bias);

    $if FMA:
      float32x4_t vt = ${VMULADDQ_F32}(vz, vn, vminus_ln2);
    $else:
      float32x4_t vt = ${VMULADDQ_F32}(vz, vn, vminus_ln2_hi);
      vt = ${VMULADDQ_F32}(vt, vn, vminus_ln2_lo);

    float32x4_t vp = ${VMULADDQ_F32}(vc5, vc6, vt);
    vp = ${VMULADDQ_F32}(vc4, vp, vt);
    vp = ${VMULADDQ_F32}(vc3, vp, vt);
    vp = ${VMULADDQ_F32}(vc2, vp, vt);
    vp = vmulq_f32(vp, vt);

    vt = vmulq_f32(vt, vs);
    vs = vsubq_f32(vs, vone);
    vp = ${VMULADDQ_F32}(vt, vp, vt);
    const float32x4_t ve = vmulq_f32(vaddq_f32(vp, vs), valpha);

    const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
    vx = vmulq_f32(vx, vbeta);
    const float32x4_t vy = vbslq_f32(vm, ve, vx);

    vst1q_f32(output, vy); output += 4;
  }
  if XNN_UNLIKELY(batch != 0) {
    float32x4_t vx = vld1q_f32(input);

    const float32x4_t vz = vmaxq_f32(vmulq_f32(vx, vprescale), vsat_cutoff);

    float32x4_t vn = ${VMULADDQ_F32}(vmagic_bias, vz, vlog2e);
    float32x4_t vs = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn), 23));
    vn = vsubq_f32(vn, vmagic_bias);

    $if FMA:
      float32x4_t vt = ${VMULADDQ_F32}(vz, vn, vminus_ln2);
    $else:
      float32x4_t vt = ${VMULADDQ_F32}(vz, vn, vminus_ln2_hi);
      vt = ${VMULADDQ_F32}(vt, vn, vminus_ln2_lo);

    float32x4_t vp = ${VMULADDQ_F32}(vc5, vc6, vt);
    vp = ${VMULADDQ_F32}(vc4, vp, vt);
    vp = ${VMULADDQ_F32}(vc3, vp, vt);
    vp = ${VMULADDQ_F32}(vc2, vp, vt);
    vp = vmulq_f32(vp, vt);

    vt = vmulq_f32(vt, vs);
    vs = vsubq_f32(vs, vone);
    vp = ${VMULADDQ_F32}(vt, vp, vt);
    const float32x4_t ve = vmulq_f32(vaddq_f32(vp, vs), valpha);

    const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
    vx = vmulq_f32(vx, vbeta);
    const float32x4_t vy = vbslq_f32(vm, ve, vx);

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