// 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
$assert ARCH in ["ARM", "X86", "RELAXED"]
$assert not FMA or ARCH == "RELAXED"
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
#include <assert.h>

#include <wasm_simd128.h>

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


$WASM_F32X4_MAX={"ARM": "wasm_f32x4_max", "X86": "wasm_f32x4_pmax", "RELAXED": "wasm_f32x4_relaxed_max"}[ARCH]
$WASM_V32X4_LANESELECT="wasm_i32x4_relaxed_laneselect" if ARCH == "RELAXED" else "wasm_v128_bitselect"
$ISA = "wasmsimd" if ARCH != "RELAXED" else "wasmrelaxedsimd"
$ARCH_SUFFIX = "" if ARCH == "RELAXED" and not FMA else "_" + ("fma" if FMA else ARCH.lower())
void xnn_f32_velu_ukernel__${ISA}${ARCH_SUFFIX}_rr2_p6_u${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 v128_t vsat_cutoff = wasm_f32x4_const_splat(-0x1.154246p+4f);
  const v128_t vmagic_bias = wasm_f32x4_const_splat(0x1.8000FEp23f);
  const v128_t vlog2e = wasm_f32x4_const_splat(0x1.715476p+0f);
  const v128_t vminus_ln2_hi = wasm_f32x4_const_splat(-0x1.62E440p-1f);
  const v128_t vminus_ln2_lo = wasm_f32x4_const_splat(0x1.0105C6p-21f);
  const v128_t vc6 = wasm_f32x4_const_splat(0x1.6b7338p-10f);
  const v128_t vc5 = wasm_f32x4_const_splat(0x1.12278Ep-7f);
  const v128_t vc4 = wasm_f32x4_const_splat(0x1.555716p-5f);
  const v128_t vc3 = wasm_f32x4_const_splat(0x1.5554B0p-3f);
  const v128_t vc2 = wasm_f32x4_const_splat(0x1.FFFFFEp-2f);
  const v128_t vone = wasm_f32x4_const_splat(1.0f);

  XNN_FORCE_REALIZATION(vsat_cutoff);
  XNN_FORCE_REALIZATION(vmagic_bias);
  XNN_FORCE_REALIZATION(vlog2e);
  XNN_FORCE_REALIZATION(vminus_ln2_hi);
  XNN_FORCE_REALIZATION(vminus_ln2_lo);
  XNN_FORCE_REALIZATION(vc6);
  XNN_FORCE_REALIZATION(vc5);
  XNN_FORCE_REALIZATION(vc4);
  XNN_FORCE_REALIZATION(vc3);
  XNN_FORCE_REALIZATION(vc2);
  XNN_FORCE_REALIZATION(vone);

  const v128_t vprescale = wasm_v128_load32_splat(&params->scalar.prescale);
  const v128_t valpha = wasm_v128_load32_splat(&params->scalar.alpha);
  const v128_t vbeta = wasm_v128_load32_splat(&params->scalar.beta);

  $if BATCH_TILE > 4:
    for (; batch >= ${BATCH_TILE} * sizeof(float); batch -= ${BATCH_TILE} * sizeof(float)) {
      v128_t vx${ABC[0:4]} = wasm_v128_load(input);
      $for N in range(4, BATCH_TILE, 4):
        v128_t vx${ABC[N:N+4]} = wasm_v128_load(input + ${N});
      input += ${BATCH_TILE};

      $for N in range(0, BATCH_TILE, 4):
        const v128_t vz${ABC[N:N+4]} = ${WASM_F32X4_MAX}(vsat_cutoff, wasm_f32x4_mul(vx${ABC[N:N+4]}, vprescale));

      $for N in range(0, BATCH_TILE, 4):
        $if FMA:
          v128_t vn${ABC[N:N+4]} = wasm_f32x4_relaxed_madd(vz${ABC[N:N+4]}, vlog2e, vmagic_bias);
        $else:
          v128_t vn${ABC[N:N+4]} = wasm_f32x4_add(wasm_f32x4_mul(vz${ABC[N:N+4]}, vlog2e), vmagic_bias);

      $for N in range(0, BATCH_TILE, 4):
        v128_t vs${ABC[N:N+4]} = wasm_i32x4_shl(vn${ABC[N:N+4]}, 23);

      $for N in range(0, BATCH_TILE, 4):
        vn${ABC[N:N+4]} = wasm_f32x4_sub(vn${ABC[N:N+4]}, vmagic_bias);

      $for N in range(0, BATCH_TILE, 4):
        $if FMA:
          v128_t vt${ABC[N:N+4]} = wasm_f32x4_relaxed_madd(vn${ABC[N:N+4]}, vminus_ln2_hi, vz${ABC[N:N+4]});
        $else:
          v128_t vt${ABC[N:N+4]} = wasm_f32x4_add(wasm_f32x4_mul(vn${ABC[N:N+4]}, vminus_ln2_hi), vz${ABC[N:N+4]});

      $for N in range(0, BATCH_TILE, 4):
        $if FMA:
          vt${ABC[N:N+4]} = wasm_f32x4_relaxed_madd(vn${ABC[N:N+4]}, vminus_ln2_lo, vt${ABC[N:N+4]});
        $else:
          vt${ABC[N:N+4]} = wasm_f32x4_add(wasm_f32x4_mul(vn${ABC[N:N+4]}, vminus_ln2_lo), vt${ABC[N:N+4]});

      $for N in range(0, BATCH_TILE, 4):
        $if FMA:
          v128_t vp${ABC[N:N+4]} = wasm_f32x4_relaxed_madd(vc6, vt${ABC[N:N+4]}, vc5);
        $else:
          v128_t vp${ABC[N:N+4]} = wasm_f32x4_add(wasm_f32x4_mul(vc6, vt${ABC[N:N+4]}), vc5);

      $for N in range(0, BATCH_TILE, 4):
        $if FMA:
          vp${ABC[N:N+4]} = wasm_f32x4_relaxed_madd(vp${ABC[N:N+4]}, vt${ABC[N:N+4]}, vc4);
        $else:
          vp${ABC[N:N+4]} = wasm_f32x4_add(wasm_f32x4_mul(vp${ABC[N:N+4]}, vt${ABC[N:N+4]}), vc4);

      $for N in range(0, BATCH_TILE, 4):
        $if FMA:
          vp${ABC[N:N+4]} = wasm_f32x4_relaxed_madd(vp${ABC[N:N+4]}, vt${ABC[N:N+4]}, vc3);
        $else:
          vp${ABC[N:N+4]} = wasm_f32x4_add(wasm_f32x4_mul(vp${ABC[N:N+4]}, vt${ABC[N:N+4]}), vc3);

      $for N in range(0, BATCH_TILE, 4):
        $if FMA:
          vp${ABC[N:N+4]} = wasm_f32x4_relaxed_madd(vp${ABC[N:N+4]}, vt${ABC[N:N+4]}, vc2);
        $else:
          vp${ABC[N:N+4]} = wasm_f32x4_add(wasm_f32x4_mul(vp${ABC[N:N+4]}, vt${ABC[N:N+4]}), vc2);

      $for N in range(0, BATCH_TILE, 4):
        vp${ABC[N:N+4]} = wasm_f32x4_mul(vp${ABC[N:N+4]}, vt${ABC[N:N+4]});

      $for N in range(0, BATCH_TILE, 4):
        vt${ABC[N:N+4]} = wasm_f32x4_mul(vt${ABC[N:N+4]}, vs${ABC[N:N+4]});
        vs${ABC[N:N+4]} = wasm_f32x4_sub(vs${ABC[N:N+4]}, vone);

      $for N in range(0, BATCH_TILE, 4):
        $if FMA:
          vp${ABC[N:N+4]} = wasm_f32x4_relaxed_madd(vp${ABC[N:N+4]}, vt${ABC[N:N+4]}, vt${ABC[N:N+4]});
        $else:
          vp${ABC[N:N+4]} = wasm_f32x4_add(wasm_f32x4_mul(vp${ABC[N:N+4]}, vt${ABC[N:N+4]}), vt${ABC[N:N+4]});

      $for N in range(0, BATCH_TILE, 4):
        const v128_t ve${ABC[N:N+4]} = wasm_f32x4_mul(wasm_f32x4_add(vp${ABC[N:N+4]}, vs${ABC[N:N+4]}), valpha);

      $for N in range(0, BATCH_TILE, 4):
        const v128_t vsignm${ABC[N:N+4]} = wasm_i32x4_shr(vx${ABC[N:N+4]}, 31);
        vx${ABC[N:N+4]} = wasm_f32x4_mul(vx${ABC[N:N+4]}, vbeta);

      $for N in range(0, BATCH_TILE, 4):
        const v128_t vy${ABC[N:N+4]} = ${WASM_V32X4_LANESELECT}(ve${ABC[N:N+4]}, vx${ABC[N:N+4]}, vsignm${ABC[N:N+4]});

      wasm_v128_store(output, vy${ABC[0:4]});
      $for N in range(4, BATCH_TILE, 4):
        wasm_v128_store(output + ${N}, vy${ABC[N:N+4]});
      output += ${BATCH_TILE};
    }
  for (; batch >= 4 * sizeof(float); batch -= 4 * sizeof(float)) {
    v128_t vx = wasm_v128_load(input);
    input += 4;

    const v128_t vz = ${WASM_F32X4_MAX}(vsat_cutoff, wasm_f32x4_mul(vx, vprescale));

    $if FMA:
      v128_t vn = wasm_f32x4_relaxed_madd(vz, vlog2e, vmagic_bias);
    $else:
      v128_t vn = wasm_f32x4_add(wasm_f32x4_mul(vz, vlog2e), vmagic_bias);
    v128_t vs = wasm_i32x4_shl(vn, 23);
    vn = wasm_f32x4_sub(vn, vmagic_bias);

    $if FMA:
      v128_t vt = wasm_f32x4_relaxed_madd(vn, vminus_ln2_hi, vz);
      vt = wasm_f32x4_relaxed_madd(vn, vminus_ln2_lo, vt);
    $else:
      v128_t vt = wasm_f32x4_add(wasm_f32x4_mul(vn, vminus_ln2_hi), vz);
      vt = wasm_f32x4_add(wasm_f32x4_mul(vn, vminus_ln2_lo), vt);

    $if FMA:
      v128_t vp = wasm_f32x4_relaxed_madd(vc6, vt, vc5);
      vp = wasm_f32x4_relaxed_madd(vp, vt, vc4);
      vp = wasm_f32x4_relaxed_madd(vp, vt, vc3);
      vp = wasm_f32x4_relaxed_madd(vp, vt, vc2);
    $else:
      v128_t vp = wasm_f32x4_add(wasm_f32x4_mul(vc6, vt), vc5);
      vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vc4);
      vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vc3);
      vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vc2);
    vp = wasm_f32x4_mul(vp, vt);

    vt = wasm_f32x4_mul(vt, vs);
    vs = wasm_f32x4_sub(vs, vone);
    $if FMA:
      vp = wasm_f32x4_relaxed_madd(vp, vt, vt);
    $else:
      vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vt);
    const v128_t ve = wasm_f32x4_mul(wasm_f32x4_add(vp, vs), valpha);

    const v128_t vsignm = wasm_i32x4_shr(vx, 31);
    vx = wasm_f32x4_mul(vx, vbeta);
    const v128_t vy = ${WASM_V32X4_LANESELECT}(ve, vx, vsignm);

    wasm_v128_store(output, vy);
    output += 4;
  }
  if XNN_UNLIKELY(batch != 0) {
    v128_t vx = wasm_v128_load(input);

    const v128_t vz = ${WASM_F32X4_MAX}(wasm_f32x4_mul(vx, vprescale), vsat_cutoff);

    $if FMA:
      v128_t vn = wasm_f32x4_relaxed_madd(vz, vlog2e, vmagic_bias);
    $else:
      v128_t vn = wasm_f32x4_add(wasm_f32x4_mul(vz, vlog2e), vmagic_bias);
    v128_t vs = wasm_i32x4_shl(vn, 23);
    vn = wasm_f32x4_sub(vn, vmagic_bias);

    $if FMA:
      v128_t vt = wasm_f32x4_relaxed_madd(vn, vminus_ln2_hi, vz);
      vt = wasm_f32x4_relaxed_madd(vn, vminus_ln2_lo, vt);
    $else:
      v128_t vt = wasm_f32x4_add(wasm_f32x4_mul(vn, vminus_ln2_hi), vz);
      vt = wasm_f32x4_add(wasm_f32x4_mul(vn, vminus_ln2_lo), vt);

    $if FMA:
      v128_t vp = wasm_f32x4_relaxed_madd(vc6, vt, vc5);
      vp = wasm_f32x4_relaxed_madd(vp, vt, vc4);
      vp = wasm_f32x4_relaxed_madd(vp, vt, vc3);
      vp = wasm_f32x4_relaxed_madd(vp, vt, vc2);
    $else:
      v128_t vp = wasm_f32x4_add(wasm_f32x4_mul(vc6, vt), vc5);
      vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vc4);
      vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vc3);
      vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vc2);
    vp = wasm_f32x4_mul(vp, vt);

    vt = wasm_f32x4_mul(vt, vs);
    vs = wasm_f32x4_sub(vs, vone);
    $if FMA:
      vp = wasm_f32x4_relaxed_madd(vp, vt, vt);
    $else:
      vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vt);
    const v128_t ve = wasm_f32x4_mul(wasm_f32x4_add(vp, vs), valpha);

    const v128_t vsignm = wasm_i32x4_shr(vx, 31);
    vx = wasm_f32x4_mul(vx, vbeta);
    v128_t vy = ${WASM_V32X4_LANESELECT}(ve, vx, vsignm);

    if (batch & (2 * sizeof(float))) {
      wasm_v128_store64_lane(output, vy, 0);
      vy = wasm_v64x2_shuffle(vy, vy, 1, 1);
      output += 2;
    }
    if (batch & (1 * sizeof(float))) {
      wasm_v128_store32_lane(output, vy, 0);
    }
  }
}