// 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
$SIMD_TILE = BATCH_TILE // 4
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
$VMULADDQ_F32 = "vfmaq_f32" if FMA else "vmlaq_f32"
#include <assert.h>

#include <arm_neon.h>

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


$PARAMS_STRUCT = "neonfma_rr1_p5" if FMA else "neon_rr2_p5"
void xnn_f32_raddstoreexpminusmax_ukernel__${"neonfma" if FMA else "neon"}_rr${1 if FMA else 2}_p5_u${BATCH_TILE}${"" if ACCUMULATORS == 1 else "_acc%d" % ACCUMULATORS}(
    size_t batch,
    const float* input,
    const float* max,
    float* output,
    float* sum,
    const void* params) XNN_OOB_READS
{
  assert(batch != 0);
  assert(batch % sizeof(float) == 0);
  assert(input != NULL);
  assert(max != NULL);
  assert(output != NULL);
  assert(sum != NULL);

  const float32x4_t vlog2e = vmovq_n_f32(0x1.715476p+0f);
  const float32x4_t vmagic_bias = vmovq_n_f32(0x1.8000FEp23f);
  const float32x4_t vc5 = vmovq_n_f32(0x1.0F9F9Cp-7f);
  const float32x4_t vc4 = vmovq_n_f32(0x1.573A1Ap-5f);
  const float32x4_t vc3 = vmovq_n_f32(0x1.555A80p-3f);
  const float32x4_t vc2 = vmovq_n_f32(0x1.FFFDC6p-2f);
  const float32x4_t vc1 = vmovq_n_f32(0x1.FFFFF6p-1f);
  const float32x4_t vdenorm_cutoff = vmovq_n_f32(-0x1.5D589Ep6f);

  XNN_FORCE_REALIZATION(vlog2e);
  XNN_FORCE_REALIZATION(vmagic_bias);
  XNN_FORCE_REALIZATION(vc5);
  XNN_FORCE_REALIZATION(vc4);
  XNN_FORCE_REALIZATION(vc3);
  XNN_FORCE_REALIZATION(vc2);
  XNN_FORCE_REALIZATION(vc1);
  XNN_FORCE_REALIZATION(vdenorm_cutoff);

  $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(-0x1.62E400p-1f);
    const float32x4_t vminus_ln2_lo = vmovq_n_f32(-0x1.7F7D1Cp-20f);
    XNN_FORCE_REALIZATION(vminus_ln2_hi);
    XNN_FORCE_REALIZATION(vminus_ln2_lo);

  const float32x4_t vi_max = vld1q_dup_f32(max);

  $if BATCH_TILE > 4:
    $for K in range(ACCUMULATORS):
      float32x4_t vacc${K} = vmovq_n_f32(0.0f);
    for (; batch >= ${BATCH_TILE} * sizeof(float); batch -= ${BATCH_TILE} * sizeof(float)) {
      $for N in range(0, BATCH_TILE, 4):
        const float32x4_t vi${ABC[N:N+4]} = vld1q_f32(input); input += 4;

      $for N in range(0, BATCH_TILE, 4):
        const float32x4_t vx${ABC[N:N+4]} = vsubq_f32(vi${ABC[N:N+4]}, vi_max);

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

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

      $for N in range(0, BATCH_TILE, 4):
        vn${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]} = ${VMULADDQ_F32}(vx${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]} = ${VMULADDQ_F32}(vx${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]} = ${VMULADDQ_F32}(vc4, vc5, vt${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]} = ${VMULADDQ_F32}(vc1, vp${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]});

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

      $for N in range(0, BATCH_TILE, 4):
        vf${ABC[N:N+4]} = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf${ABC[N:N+4]}), vcltq_f32(vx${ABC[N:N+4]}, vdenorm_cutoff)));

      $for N in range(0, BATCH_TILE, 4):
        vst1q_f32(output, vf${ABC[N:N+4]}); output += 4;

      $for N in range(0, BATCH_TILE, 4):
        vacc${N % ACCUMULATORS} = vaddq_f32(vacc${N % ACCUMULATORS}, vf${ABC[N:N+4]});
    }
    $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_f32(vacc${A}, vacc${A + ACC_SLICE});
        $ACC_SLICE *= 2

    float32x4_t vacc = vacc0;
  $else:
    float32x4_t vacc = vmovq_n_f32(0.0f);
  for (; batch >= 4 * sizeof(float); batch -= 4 * sizeof(float)) {
    const float32x4_t vi = vld1q_f32(input); input += 4;

    const float32x4_t vx = vsubq_f32(vi, vi_max);

    float32x4_t vn = ${VMULADDQ_F32}(vmagic_bias, vx, vlog2e);

    const 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}(vx, vn, vminus_ln2);
    $else:
      float32x4_t vt = ${VMULADDQ_F32}(vx, vn, vminus_ln2_hi);
      vt = ${VMULADDQ_F32}(vt, vn, vminus_ln2_lo);

    float32x4_t vp = ${VMULADDQ_F32}(vc4, vc5, vt);
    vp = ${VMULADDQ_F32}(vc3, vp, vt);
    vp = ${VMULADDQ_F32}(vc2, vp, vt);
    vp = ${VMULADDQ_F32}(vc1, vp, vt);

    vt = vmulq_f32(vt, vs);
    float32x4_t vf = ${VMULADDQ_F32}(vs, vp, vt);

    vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcltq_f32(vx, vdenorm_cutoff)));

    vst1q_f32(output, vf); output += 4;

    vacc = vaddq_f32(vacc, vf);
  }
#if XNN_ARCH_ARM64
  float vacc_lo = vaddvq_f32(vacc);
#else
  float32x2_t vacc_lo = vadd_f32(vget_high_f32(vacc), vget_low_f32(vacc));
#endif
  if (batch != 0) {
    assert(batch >= 1 * sizeof(float));
    assert(batch <= 3 * sizeof(float));
    const float32x4_t vi = vld1q_f32(input); input += 4;

    const float32x4_t vx = vsubq_f32(vi, vi_max);

    float32x4_t vn = ${VMULADDQ_F32}(vmagic_bias, vx, vlog2e);

    const 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}(vx, vn, vminus_ln2);
    $else:
      float32x4_t vt = ${VMULADDQ_F32}(vx, vn, vminus_ln2_hi);
      vt = ${VMULADDQ_F32}(vt, vn, vminus_ln2_lo);

    float32x4_t vp = ${VMULADDQ_F32}(vc4, vc5, vt);
    vp = ${VMULADDQ_F32}(vc3, vp, vt);
    vp = ${VMULADDQ_F32}(vc2, vp, vt);
    vp = ${VMULADDQ_F32}(vc1, vp, vt);

    vt = vmulq_f32(vt, vs);
    float32x4_t vf = ${VMULADDQ_F32}(vs, vp, vt);

    vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcltq_f32(vx, vdenorm_cutoff)));

    float32x2_t vf_lo = vget_low_f32(vf);
    if (batch & (2 * sizeof(float))) {
      vst1_f32(output, vf_lo); output += 2;

      #if XNN_ARCH_ARM64
        vacc_lo += vaddv_f32(vf_lo);
      #else
        vacc_lo = vadd_f32(vacc_lo, vf_lo);
      #endif

      vf_lo = vget_high_f32(vf);
    }
    if (batch & (1 * sizeof(float))) {
      vst1_lane_f32(output, vf_lo, 0);

      #if XNN_ARCH_ARM64
        vacc_lo += vget_lane_f32(vf_lo, 0);
      #else
        vacc_lo = vadd_f32(vacc_lo, vreinterpret_f32_u64(vshl_n_u64(vreinterpret_u64_f32(vf_lo), 32)));
      #endif
    }
  }
#if XNN_ARCH_ARM64
  *sum = vacc_lo;
#else
  vst1_lane_f32(sum, vpadd_f32(vacc_lo, vacc_lo), 0);
#endif
}