// 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"
#include <assert.h>

#include <wasm_simd128.h>

#include <xnnpack/vunary.h>
#include <xnnpack/common.h>


extern XNN_INTERNAL const float xnn_table_exp2minus_k_over_16[16];

void xnn_f32_velu_ukernel__wasmsimd_${"x86" if X86 else "arm"}_rr2_lut16_p3_x${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 v128_t vprescale = wasm_v32x4_load_splat(&params->scalar.prescale);
  const v128_t valpha = wasm_v32x4_load_splat(&params->scalar.alpha);
  const v128_t vbeta = wasm_v32x4_load_splat(&params->scalar.beta);

  const v128_t vsat_cutoff = wasm_f32x4_splat(-0x1.154246p+4f);
  const v128_t vmagic_bias = wasm_f32x4_splat(0x1.800000p19f);
  const v128_t vlog2e = wasm_f32x4_splat(0x1.715476p+0f);
  const v128_t vindex_mask = wasm_i32x4_splat(0xF);
  const v128_t vminus_ln2_hi = wasm_f32x4_splat(-0x1.62E400p-1f);
  const v128_t vminus_ln2_lo = wasm_f32x4_splat(-0x1.7F7D1Cp-20f);
  const v128_t vc3 = wasm_f32x4_splat(0x1.55561Cp-3f);
  const v128_t vc2 = wasm_f32x4_splat(0x1.0001ECp-1f);
  const v128_t vone = wasm_f32x4_splat(1.0f);

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

      $for N in range(0, BATCH_TILE, 4):
        $if X86:
          const v128_t vz${ABC[N:N+4]} = wasm_f32x4_mul(vx${ABC[N:N+4]}, vprescale);
        $else:
          const v128_t vz${ABC[N:N+4]} = wasm_f32x4_max(wasm_f32x4_mul(vx${ABC[N:N+4]}, vprescale), vsat_cutoff);

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

      $for N in range(0, BATCH_TILE, 4):
        const v128_t vidx${ABC[N:N+4]} = wasm_i32x4_shl(wasm_v128_and(vn${ABC[N:N+4]}, vindex_mask), 2);
        const v128_t ven${ABC[N:N+4]} = wasm_i32x4_shl(vn${ABC[N:N+4]}, 19);

      $for N in range(0, BATCH_TILE, 4):
        const uint64_t vidx${ABC[N:N+2]} = wasm_i64x2_extract_lane(vidx${ABC[N:N+4]}, 0);
        const uint64_t vidx${ABC[N+2:N+4]} = wasm_i64x2_extract_lane(vidx${ABC[N:N+4]}, 1);
        const float vl${ABC[N]}   = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx${ABC[N:N+2]}));
        const float vl${ABC[N+1]} = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx${ABC[N:N+2]} >> 32)));
        const float vl${ABC[N+2]} = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx${ABC[N+2:N+4]}));
        const float vl${ABC[N+3]} = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx${ABC[N+2:N+4]} >> 32)));
        const v128_t vl${ABC[N:N+4]} = wasm_f32x4_make(vl${ABC[N]}, vl${ABC[N+1]}, vl${ABC[N+2]}, vl${ABC[N+3]});

      $for N in range(0, BATCH_TILE, 4):
        vn${ABC[N:N+4]} = wasm_f32x4_sub(vn${ABC[N:N+4]}, vmagic_bias);
        v128_t vs${ABC[N:N+4]} = wasm_i32x4_add(vl${ABC[N:N+4]}, ven${ABC[N:N+4]});

      $for N in range(0, BATCH_TILE, 4):
        v128_t vt${ABC[N:N+4]} = wasm_f32x4_add(wasm_f32x4_mul(vn${ABC[N:N+4]}, vminus_ln2_hi), vz${ABC[N:N+4]});
        $if X86:
          const v128_t vsatm${ABC[N:N+4]} = wasm_f32x4_le(vz${ABC[N:N+4]}, vsat_cutoff);

      $for N in range(0, BATCH_TILE, 4):
        vt${ABC[N:N+4]} = wasm_f32x4_add(wasm_f32x4_mul(vn${ABC[N:N+4]}, vminus_ln2_lo), vt${ABC[N:N+4]});
        $if X86:
          vs${ABC[N:N+4]} = wasm_v128_andnot(vs${ABC[N:N+4]}, vsatm${ABC[N:N+4]});

      $for N in range(0, BATCH_TILE, 4):
        $if X86:
          vt${ABC[N:N+4]} = wasm_v128_andnot(vt${ABC[N:N+4]}, vsatm${ABC[N:N+4]});
        v128_t vp${ABC[N:N+4]} = wasm_f32x4_add(wasm_f32x4_mul(vc3, vt${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);

      $for N in range(0, BATCH_TILE, 4):
        vp${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_v128_bitselect(ve${ABC[N:N+4]}, vx${ABC[N:N+4]}, vsignm${ABC[N:N+4]});

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

    $if X86:
      const v128_t vz = wasm_f32x4_mul(vx, vprescale);
    $else:
      const v128_t vz = wasm_f32x4_max(wasm_f32x4_mul(vx, vprescale), vsat_cutoff);

    v128_t vn = wasm_f32x4_add(wasm_f32x4_mul(vz, vlog2e), vmagic_bias);
    const v128_t vidx = wasm_i32x4_shl(wasm_v128_and(vn, vindex_mask), 2);
    const v128_t ven = wasm_i32x4_shl(vn, 19);

    const uint64_t vidx_lo = wasm_i64x2_extract_lane(vidx, 0);
    const uint64_t vidx_hi = wasm_i64x2_extract_lane(vidx, 1);
    const float vl0 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_lo));
    const float vl1 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_lo >> 32)));
    const float vl2 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_hi));
    const float vl3 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_hi >> 32)));
    const v128_t vl = wasm_f32x4_make(vl0, vl1, vl2, vl3);

    v128_t vs = wasm_i32x4_add(vl, ven);
    vn = wasm_f32x4_sub(vn, vmagic_bias);

    v128_t vt = wasm_f32x4_add(wasm_f32x4_mul(vn, vminus_ln2_hi), vz);
    $if X86:
      const v128_t vsatm = wasm_f32x4_le(vz, vsat_cutoff);
    vt = wasm_f32x4_add(wasm_f32x4_mul(vn, vminus_ln2_lo), vt);
    $if X86:
      vs = wasm_v128_andnot(vs, vsatm);
      vt = wasm_v128_andnot(vt, vsatm);

    v128_t vp = wasm_f32x4_add(wasm_f32x4_mul(vc3, vt), vc2);
    vp = wasm_f32x4_mul(vp, vt);

    vt = wasm_f32x4_mul(vt, vs);
    vs = wasm_f32x4_sub(vs, vone);
    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_v128_bitselect(ve, vx, vsignm);

    wasm_v128_store(y, vy);
    y += 4;
  }
  if XNN_UNLIKELY(n != 0) {
    v128_t vx = wasm_v128_load(x);

    $if X86:
      const v128_t vz = wasm_f32x4_mul(vx, vprescale);
    $else:
      const v128_t vz = wasm_f32x4_max(wasm_f32x4_mul(vx, vprescale), vsat_cutoff);

    v128_t vn = wasm_f32x4_add(wasm_f32x4_mul(vz, vlog2e), vmagic_bias);
    const v128_t vidx = wasm_i32x4_shl(wasm_v128_and(vn, vindex_mask), 2);
    const v128_t ven = wasm_i32x4_shl(vn, 19);

    const uint64_t vidx_lo = wasm_i64x2_extract_lane(vidx, 0);
    const uint64_t vidx_hi = wasm_i64x2_extract_lane(vidx, 1);
    const float vl0 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_lo));
    const float vl1 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_lo >> 32)));
    const float vl2 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_hi));
    const float vl3 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_hi >> 32)));
    const v128_t vl = wasm_f32x4_make(vl0, vl1, vl2, vl3);

    v128_t vs = wasm_i32x4_add(vl, ven);
    vn = wasm_f32x4_sub(vn, vmagic_bias);

    v128_t vt = wasm_f32x4_add(wasm_f32x4_mul(vn, vminus_ln2_hi), vz);
    $if X86:
      const v128_t vsatm = wasm_f32x4_le(vz, vsat_cutoff);
    vt = wasm_f32x4_add(wasm_f32x4_mul(vn, vminus_ln2_lo), vt);
    $if X86:
      vs = wasm_v128_andnot(vs, vsatm);
      vt = wasm_v128_andnot(vt, vsatm);

    v128_t vp = wasm_f32x4_add(wasm_f32x4_mul(vc3, vt), vc2);
    vp = wasm_f32x4_mul(vp, vt);

    vt = wasm_f32x4_mul(vt, vs);
    vs = wasm_f32x4_sub(vs, vone);
    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_v128_bitselect(ve, vx, vsignm);

    if (n & (2 * sizeof(float))) {
      *((double*) y) = wasm_f64x2_extract_lane(vy, 0);
      vy = wasm_v32x4_shuffle(vy, vy, 2, 3, 2, 3);
      y += 2;
    }
    if (n & (1 * sizeof(float))) {
      *y = wasm_f32x4_extract_lane(vy, 0);
    }
  }
}