// 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>


extern XNN_INTERNAL const int32_t xnn_table_exp2minus_k_over_16[16];

void xnn_f32_velu_ukernel__${"neonfma" if FMA else "neon"}_rr${1 if FMA else 2}_lut16_p3_x${BATCH_TILE}(
    size_t n,
    const float* x,
    float* y,
    const union xnn_f32_elu_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
{
  assert(n != 0);
  assert(n % sizeof(float) == 0);
  assert(x != NULL);
  assert(y != NULL);

  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);

  const float32x4_t vsat_cutoff = vmovq_n_f32(-0x1.154246p+4f);
  const float32x4_t vmagic_bias = vmovq_n_f32(0x1.800000p19f);
  const float32x4_t vlog2e = vmovq_n_f32(0x1.715476p+0f);
  const int32x4_t vindex_mask = vmovq_n_s32(0xF);
  $if FMA:
    const float32x4_t vminus_ln2 = vmovq_n_f32(-0x1.62E430p-1f);
  $else:
    const float32x4_t vminus_ln2_hi = vmovq_n_f32(-0x1.62E400p-1f);
    const float32x4_t vminus_ln2_lo = vmovq_n_f32(-0x1.7F7D1Cp-20f);
  const float32x4_t vc3 = vmovq_n_f32(0x1.55561Cp-3f);
  const float32x4_t vc2 = vmovq_n_f32(0x1.0001ECp-1f);
  const float32x4_t vone = vmovq_n_f32(1.0f);

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

      $for N in range(0, BATCH_TILE, 4):
        const uint64x2_t vidx${ABC[N:N+4]} = vreinterpretq_u64_s32(vshlq_n_s32(vandq_s32(vreinterpretq_s32_f32(vn${ABC[N:N+4]}), vindex_mask), 2));
        const int32x4_t ven${ABC[N:N+4]} = vshlq_n_s32(vreinterpretq_s32_f32(vn${ABC[N:N+4]}), 19);

      $for N in range(0, BATCH_TILE, 4):
        const uint64_t vidx${ABC[N:N+2]} = vgetq_lane_u64(vidx${ABC[N:N+4]}, 0);
        const uint64_t vidx${ABC[N+2:N+4]} = vgetq_lane_u64(vidx${ABC[N:N+4]}, 1);
        int32x2_t vl${ABC[N:N+2]} = vld1_dup_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx${ABC[N:N+2]}));
        int32x2_t vl${ABC[N+2:N+4]} = vld1_dup_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx${ABC[N+2:N+4]}));
        vl${ABC[N:N+2]} = vld1_lane_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx${ABC[N:N+2]} >> 32)), vl${ABC[N:N+2]}, 1);
        vl${ABC[N+2:N+4]} = vld1_lane_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx${ABC[N+2:N+4]} >> 32)), vl${ABC[N+2:N+4]}, 1);
        const int32x4_t vl${ABC[N:N+4]} = vcombine_s32(vl${ABC[N:N+2]}, vl${ABC[N+2:N+4]});

      $for N in range(0, BATCH_TILE, 4):
        vn${ABC[N:N+4]} = vsubq_f32(vn${ABC[N:N+4]}, vmagic_bias);
        float32x4_t vs${ABC[N:N+4]} = vreinterpretq_f32_s32(vaddq_s32(vl${ABC[N:N+4]}, ven${ABC[N:N+4]}));

      $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]} = ${VMULADDQ_F32}(vc2, vc3, vt${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]});

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

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

    float32x4_t vn = ${VMULADDQ_F32}(vmagic_bias, vz, vlog2e);
    const uint64x2_t vidx = vreinterpretq_u64_s32(vshlq_n_s32(vandq_s32(vreinterpretq_s32_f32(vn), vindex_mask), 2));
    const int32x4_t ven = vshlq_n_s32(vreinterpretq_s32_f32(vn), 19);

    const uint64_t vidx_lo = vgetq_lane_u64(vidx, 0);
    const uint64_t vidx_hi = vgetq_lane_u64(vidx, 1);
    int32x2_t vl_lo = vld1_dup_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_lo));
    int32x2_t vl_hi = vld1_dup_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_hi));
    vl_lo = vld1_lane_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_lo >> 32)), vl_lo, 1);
    vl_hi = vld1_lane_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_hi >> 32)), vl_hi, 1);

    vn = vsubq_f32(vn, vmagic_bias);
    const int32x4_t vl = vcombine_s32(vl_lo, vl_hi);

    $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 vs = vreinterpretq_f32_s32(vaddq_s32(vl, ven));

    float32x4_t vp = ${VMULADDQ_F32}(vc2, vc3, 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(y, vy); y += 4;
  }
  if XNN_UNLIKELY(n != 0) {
    float32x4_t vx = vld1q_f32(x);

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

    float32x4_t vn = ${VMULADDQ_F32}(vmagic_bias, vz, vlog2e);
    const uint64x2_t vidx = vreinterpretq_u64_s32(vshlq_n_s32(vandq_s32(vreinterpretq_s32_f32(vn), vindex_mask), 2));
    const int32x4_t ven = vshlq_n_s32(vreinterpretq_s32_f32(vn), 19);

    const uint64_t vidx_lo = vgetq_lane_u64(vidx, 0);
    const uint64_t vidx_hi = vgetq_lane_u64(vidx, 1);
    int32x2_t vl_lo = vld1_dup_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_lo));
    int32x2_t vl_hi = vld1_dup_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_hi));
    vl_lo = vld1_lane_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_lo >> 32)), vl_lo, 1);
    vl_hi = vld1_lane_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_hi >> 32)), vl_hi, 1);

    vn = vsubq_f32(vn, vmagic_bias);
    const int32x4_t vl = vcombine_s32(vl_lo, vl_hi);

    $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 vs = vreinterpretq_f32_s32(vaddq_s32(vl, ven));

    float32x4_t vp = ${VMULADDQ_F32}(vc2, vc3, 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 (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);
    }
  }
}