// 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/raddstoreexpminusmax.h"


void xnn_f16_raddstoreexpminusmax_ukernel__neonfp16arith_rr2_p2_u${BATCH_TILE}${"" if ACCUMULATORS == 1 else "_acc%d" % ACCUMULATORS}(
    size_t batch,
    const xnn_float16* input,
    const xnn_float16* max,
    xnn_float16* output,
    xnn_float16* sum,
    const void* params) XNN_OOB_READS
{
  assert(batch != 0);
  assert(batch % sizeof(uint16_t) == 0);
  assert(input != NULL);
  assert(max != NULL);
  assert(output != NULL);
  assert(sum != NULL);

  const float16x8_t vlog2e = vreinterpretq_f16_u16(vmovq_n_u16(UINT16_C(0x3DC5)));  // 0x1.714p+0h
  const float16x8_t vmagic_bias = vreinterpretq_f16_u16(vmovq_n_u16(UINT16_C(0x660F)));  // 0x1.83Cp+10h
  const float16x8_t vminus_ln2_hi = vreinterpretq_f16_u16(vmovq_n_u16(UINT16_C(0xB98C)));  // -0x1.630p-1h
  const float16x8_t vminus_ln2_lo = vreinterpretq_f16_u16(vmovq_n_u16(UINT16_C(0x0AF4)));  // 0x1.BD0p-13h
  const float16x8_t vc2 = vreinterpretq_f16_u16(vmovq_n_u16(UINT16_C(0x37F9)));  // 0x1.FE4p-2h
  const float16x8_t vc1 = vreinterpretq_f16_u16(vmovq_n_u16(UINT16_C(0x3C0E)));  // 0x1.038p+0h
  const float16x8_t vdenorm_cutoff = vreinterpretq_f16_u16(vmovq_n_u16(UINT16_C(0xC8DA)));  // -0x1.368p+3h

  XNN_FORCE_REALIZATION(vlog2e);
  XNN_FORCE_REALIZATION(vmagic_bias);
  XNN_FORCE_REALIZATION(vminus_ln2_hi);
  XNN_FORCE_REALIZATION(vminus_ln2_lo);
  XNN_FORCE_REALIZATION(vc2);
  XNN_FORCE_REALIZATION(vc1);
  XNN_FORCE_REALIZATION(vdenorm_cutoff);

  const float16x8_t vi_max = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*)max));

  const uint16_t* i = (const uint16_t*) input;
  uint16_t* o = (uint16_t*) output;
  $if BATCH_TILE > 8:
    $for K in range(ACCUMULATORS):
      float16x8_t vacc${K} = vreinterpretq_f16_u16(vmovq_n_u16(0));
    for (; batch >= ${BATCH_TILE} * sizeof(uint16_t); batch -= ${BATCH_TILE} * sizeof(uint16_t)) {
      $for N in range(SIMD_TILE):
        const float16x8_t vi${ABC[N]} = vreinterpretq_f16_u16(vld1q_u16(i)); i += 8;

      $for N in range(SIMD_TILE):
        const float16x8_t vx${ABC[N]} = vsubq_f16(vi${ABC[N]}, vi_max);

      $for N in range(SIMD_TILE):
        float16x8_t vn${ABC[N]} = vfmaq_f16(vmagic_bias, vx${ABC[N]}, vlog2e);

      $for N in range(SIMD_TILE):
        const float16x8_t vs${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(vx${ABC[N]}, vn${ABC[N]}, vminus_ln2_hi);

      $for N in range(SIMD_TILE):
        vt${ABC[N]} = vfmaq_f16(vt${ABC[N]}, vn${ABC[N]}, vminus_ln2_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]});

      $for N in range(SIMD_TILE):
        float16x8_t vf${ABC[N]} = vfmaq_f16(vs${ABC[N]}, vp${ABC[N]}, vt${ABC[N]});
        const uint16x8_t vm${ABC[N]} = vcltq_f16(vx${ABC[N]}, vdenorm_cutoff);

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

      $for N in range(SIMD_TILE):
        vst1q_u16(o, vreinterpretq_u16_f16(vf${ABC[N]})); o += 8;

      $for N in range(SIMD_TILE):
        vacc${N % ACCUMULATORS} = vaddq_f16(vacc${N % ACCUMULATORS}, vf${ABC[N]});
    }
    $if ACCUMULATORS > 1:
      $ACC_SLICE = 1
      $while ACC_SLICE < ACCUMULATORS:
        $for A in range(0, ACCUMULATORS, ACC_SLICE * 2):
          $if A + ACC_SLICE < ACCUMULATORS:
            vacc${A} = vaddq_f16(vacc${A}, vacc${A + ACC_SLICE});
        $ACC_SLICE *= 2

    float16x8_t vacc = vacc0;
  $else:
    float16x8_t vacc = vreinterpretq_f16_u16(vmovq_n_u16(0));
  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 = vsubq_f16(vi, vi_max);

    float16x8_t vn = vfmaq_f16(vmagic_bias, vx, vlog2e);
    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(vx, vn, vminus_ln2_hi);
    vt = vfmaq_f16(vt, vn, vminus_ln2_lo);

    const float16x8_t vp = vfmaq_f16(vc1, vc2, vt);
    vt = vmulq_f16(vt, vs);

    float16x8_t vf = vfmaq_f16(vs, vp, vt);
    const uint16x8_t vm = vcltq_f16(vx, vdenorm_cutoff);
    vf = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf), vm));

    vst1q_u16(o, vreinterpretq_u16_f16(vf)); o += 8;

    vacc = vaddq_f16(vacc, vf);
  }
  float16x4_t vacc_lo = vadd_f16(vget_low_f16(vacc), vget_high_f16(vacc));
  if (batch != 0) {
    assert(batch >= 1 * sizeof(uint16_t));
    assert(batch <= 7 * sizeof(uint16_t));
    const float16x8_t vi = vreinterpretq_f16_u16(vld1q_u16(i));

    const float16x8_t vx = vsubq_f16(vi, vi_max);

    float16x8_t vn = vfmaq_f16(vmagic_bias, vx, vlog2e);
    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(vx, vn, vminus_ln2_hi);
    vt = vfmaq_f16(vt, vn, vminus_ln2_lo);

    const float16x8_t vp = vfmaq_f16(vc1, vc2, vt);
    vt = vmulq_f16(vt, vs);

    float16x8_t vf = vfmaq_f16(vs, vp, vt);
    const uint16x8_t vm = vcltq_f16(vx, vdenorm_cutoff);
    vf = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf), vm));

    float16x4_t vf_lo = vget_low_f16(vf);
    if (batch & (4 * sizeof(uint16_t))) {
      vst1_u16(o, vreinterpret_u16_f16(vf_lo)); o += 4;
      vacc_lo = vadd_f16(vacc_lo, vf_lo);
      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;
      vacc_lo = vadd_f16(vacc_lo, vreinterpret_f16_u64(vshl_n_u64(vreinterpret_u64_f16(vf_lo), 32)));
      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);
      vacc_lo = vadd_f16(vacc_lo, vreinterpret_f16_u64(vshl_n_u64(vreinterpret_u64_f16(vf_lo), 48)));
    }
  }
  vacc_lo = vpadd_f16(vacc_lo, vacc_lo);
  vacc_lo = vpadd_f16(vacc_lo, vacc_lo);
  vst1_lane_u16(sum, vreinterpret_u16_f16(vacc_lo), 0);
}