// 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 not FMA or RELAXED
$assert not BLENDVPS or RELAXED
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
#include <assert.h>

#include <wasm_simd128.h>

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


$WASM_V32X4_LANESELECT = "wasm_i32x4_relaxed_laneselect" if RELAXED else "wasm_v128_bitselect"
$ISA = "wasmblendvps" if BLENDVPS else "wasmrelaxedsimd" if RELAXED else "wasmsimd"
$ARCH_SUFFIX = "_fma" if FMA else ""
void xnn_f32_vsigmoid_ukernel__${ISA}${ARCH_SUFFIX}_rr2_p5_div_u${BATCH_TILE}(
    size_t batch,
    const float* input,
    float* output,
    const struct xnn_f32_default_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 vmagic_bias = wasm_f32x4_const_splat(0x1.8000FEp23f);
  const v128_t vminus_log2e = wasm_f32x4_const_splat(-0x1.715476p+0f);
  const v128_t vln2_hi = wasm_f32x4_const_splat(0x1.62E400p-1f);
  const v128_t vln2_lo = wasm_f32x4_const_splat(0x1.7F7D1Cp-20f);
  const v128_t vc5 = wasm_f32x4_const_splat(-0x1.0F9F9Cp-7f);
  const v128_t vc4 = wasm_f32x4_const_splat(0x1.573A1Ap-5f);
  const v128_t vc3 = wasm_f32x4_const_splat(-0x1.555A80p-3f);
  const v128_t vc2 = wasm_f32x4_const_splat(0x1.FFFDC6p-2f);
  const v128_t vc1 = wasm_f32x4_const_splat(-0x1.FFFFF6p-1f);
  const v128_t vone = wasm_f32x4_const_splat(1.0f);
  const v128_t vdenorm_cutoff = wasm_f32x4_const_splat(0x1.5D589Ep+6f);

  XNN_FORCE_REALIZATION(vmagic_bias);
  XNN_FORCE_REALIZATION(vminus_log2e);
  XNN_FORCE_REALIZATION(vln2_hi);
  XNN_FORCE_REALIZATION(vln2_lo);
  XNN_FORCE_REALIZATION(vc5);
  XNN_FORCE_REALIZATION(vc4);
  XNN_FORCE_REALIZATION(vc3);
  XNN_FORCE_REALIZATION(vc2);
  XNN_FORCE_REALIZATION(vc1);
  // XNN_FORCE_REALIZATION(vone);
  XNN_FORCE_REALIZATION(vdenorm_cutoff);

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

      $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]}, vminus_log2e, vmagic_bias);
        $else:
          v128_t vn${ABC[N:N+4]} = wasm_f32x4_add(wasm_f32x4_mul(vz${ABC[N:N+4]}, vminus_log2e), vmagic_bias);

      $for N in range(0, BATCH_TILE, 4):
        const v128_t vs${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${ABC[N:N+4]}, vln2_hi, vz${ABC[N:N+4]});
        $else:
          v128_t vt${ABC[N:N+4]} = wasm_f32x4_add(wasm_f32x4_mul(vn${ABC[N:N+4]}, vln2_hi), vz${ABC[N:N+4]});

      $for N in range(0, BATCH_TILE, 4):
        $if FMA:
          vt${ABC[N:N+4]} = wasm_f32x4_relaxed_madd(vn${ABC[N:N+4]}, vln2_lo, vt${ABC[N:N+4]});
        $else:
          vt${ABC[N:N+4]} = wasm_f32x4_add(wasm_f32x4_mul(vn${ABC[N:N+4]}, vln2_lo), vt${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(vt${ABC[N:N+4]}, vc5, vc4);
        $else:
          v128_t vp${ABC[N:N+4]} = wasm_f32x4_add(wasm_f32x4_mul(vt${ABC[N:N+4]}, vc5), vc4);

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

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

      $for N in range(0, BATCH_TILE, 4):
        $if FMA:
          vp${ABC[N:N+4]} = wasm_f32x4_relaxed_madd(vt${ABC[N:N+4]}, vp${ABC[N:N+4]}, vc1);
        $else:
          vp${ABC[N:N+4]} = wasm_f32x4_add(wasm_f32x4_mul(vt${ABC[N:N+4]}, vp${ABC[N:N+4]}), vc1);

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

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

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

      $for N in range(0, BATCH_TILE, 4):
        v128_t vf${ABC[N:N+4]} = wasm_f32x4_div(ve${ABC[N:N+4]}, vd${ABC[N:N+4]});

      $for N in range(0, BATCH_TILE, 4):
        vf${ABC[N:N+4]} = wasm_v128_andnot(vf${ABC[N:N+4]}, wasm_f32x4_gt(vz${ABC[N:N+4]}, vdenorm_cutoff));

      $for N in range(0, BATCH_TILE, 4):
        const v128_t vcf${ABC[N:N+4]} = wasm_f32x4_sub(vone, vf${ABC[N:N+4]});

      $for N in range(0, BATCH_TILE, 4):
        $if BLENDVPS:
          vf${ABC[N:N+4]} = ${WASM_V32X4_LANESELECT}(vf${ABC[N:N+4]}, vcf${ABC[N:N+4]}, vx${ABC[N:N+4]});
        $else:
          vf${ABC[N:N+4]} = ${WASM_V32X4_LANESELECT}(vf${ABC[N:N+4]}, vcf${ABC[N:N+4]}, wasm_i32x4_shr(vx${ABC[N:N+4]}, 31));

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

    const v128_t vz = wasm_f32x4_abs(vx);

    $if FMA:
      v128_t vn = wasm_f32x4_relaxed_madd(vz, vminus_log2e, vmagic_bias);
    $else:
      v128_t vn = wasm_f32x4_add(wasm_f32x4_mul(vz, vminus_log2e), vmagic_bias);
    const 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, vln2_hi, vz);
      vt = wasm_f32x4_relaxed_madd(vn, vln2_lo, vt);
    $else:
      v128_t vt = wasm_f32x4_add(wasm_f32x4_mul(vn, vln2_hi), vz);
      vt = wasm_f32x4_add(wasm_f32x4_mul(vn, vln2_lo), vt);

    $if FMA:
      v128_t vp = wasm_f32x4_relaxed_madd(vt, vc5, vc4);
      vp = wasm_f32x4_relaxed_madd(vt, vp, vc3);
      vp = wasm_f32x4_relaxed_madd(vt, vp, vc2);
      vp = wasm_f32x4_relaxed_madd(vt, vp, vc1);
    $else:
      v128_t vp = wasm_f32x4_add(wasm_f32x4_mul(vt, vc5), vc4);
      vp = wasm_f32x4_add(wasm_f32x4_mul(vt, vp), vc3);
      vp = wasm_f32x4_add(wasm_f32x4_mul(vt, vp), vc2);
      vp = wasm_f32x4_add(wasm_f32x4_mul(vt, vp), vc1);

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

    v128_t vf = wasm_f32x4_div(ve, vd);
    vf = wasm_v128_andnot(vf, wasm_f32x4_gt(vz, vdenorm_cutoff));
    const v128_t vcf = wasm_f32x4_sub(vone, vf);
    $if BLENDVPS:
      vf = ${WASM_V32X4_LANESELECT}(vf, vcf, vx);
    $else:
      vf = ${WASM_V32X4_LANESELECT}(vf, vcf, wasm_i32x4_shr(vx, 31));

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

    const v128_t vz = wasm_f32x4_abs(vx);

    $if FMA:
      v128_t vn = wasm_f32x4_relaxed_madd(vz, vminus_log2e, vmagic_bias);
    $else:
      v128_t vn = wasm_f32x4_add(wasm_f32x4_mul(vz, vminus_log2e), vmagic_bias);
    const 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, vln2_hi, vz);
      vt = wasm_f32x4_relaxed_madd(vn, vln2_lo, vt);
    $else:
      v128_t vt = wasm_f32x4_add(wasm_f32x4_mul(vn, vln2_hi), vz);
      vt = wasm_f32x4_add(wasm_f32x4_mul(vn, vln2_lo), vt);

    $if FMA:
      v128_t vp = wasm_f32x4_relaxed_madd(vt, vc5, vc4);
      vp = wasm_f32x4_relaxed_madd(vt, vp, vc3);
      vp = wasm_f32x4_relaxed_madd(vt, vp, vc2);
      vp = wasm_f32x4_relaxed_madd(vt, vp, vc1);
    $else:
      v128_t vp = wasm_f32x4_add(wasm_f32x4_mul(vt, vc5), vc4);
      vp = wasm_f32x4_add(wasm_f32x4_mul(vt, vp), vc3);
      vp = wasm_f32x4_add(wasm_f32x4_mul(vt, vp), vc2);
      vp = wasm_f32x4_add(wasm_f32x4_mul(vt, vp), vc1);

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

    v128_t vf = wasm_f32x4_div(ve, vd);
    vf = wasm_v128_andnot(vf, wasm_f32x4_gt(vz, vdenorm_cutoff));
    const v128_t vcf = wasm_f32x4_sub(vone, vf);
    $if BLENDVPS:
      vf = ${WASM_V32X4_LANESELECT}(vf, vcf, vx);
    $else:
      vf = ${WASM_V32X4_LANESELECT}(vf, vcf, wasm_i32x4_shr(vx, 31));

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