// 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"


extern XNN_INTERNAL const float xnn_table_exp2minus_k_over_16[16];

$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}${ARCH_SUFFIX}_rr2_lut16_p3_u${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(0x1.800000p19f);
  const v128_t vlog2e = wasm_f32x4_const_splat(0x1.715476p+0f);
  const v128_t vindex_mask = wasm_u32x4_const_splat(UINT32_C(0xF));
  const v128_t vminus_ln2_hi = wasm_f32x4_const_splat(-0x1.62E400p-1f);
  const v128_t vminus_ln2_lo = wasm_f32x4_const_splat(-0x1.7F7D1Cp-20f);
  const v128_t vc3 = wasm_f32x4_const_splat(0x1.55561Cp-3f);
  const v128_t vc2 = wasm_f32x4_const_splat(0x1.0001ECp-1f);
  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(vindex_mask);
  XNN_FORCE_REALIZATION(vminus_ln2_hi);
  XNN_FORCE_REALIZATION(vminus_ln2_lo);
  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));

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

      $for N in range(0, BATCH_TILE, 4):
        const v128_t vidx${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);

      $for N in range(0, BATCH_TILE, 4):
        const uint32_t vidx${ABC[N]} = wasm_u32x4_extract_lane(vidx${ABC[N:N+4]}, 0);
        v128_t vl${ABC[N:N+4]} = wasm_v128_load32_zero((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx${ABC[N]}));

      $for L in range(1, 4):
        $for N in range(0, BATCH_TILE, 4):
          const uint32_t vidx${ABC[N+L]} = wasm_u32x4_extract_lane(vidx${ABC[N:N+4]}, ${L});
          vl${ABC[N:N+4]} = wasm_v128_load32_lane((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx${ABC[N+L]}), vl${ABC[N:N+4]}, ${L});

      $for N in range(0, BATCH_TILE, 4):
        vn${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(vz${ABC[N:N+4]}, wasm_f32x4_mul(vn${ABC[N:N+4]}, vminus_ln2_hi));

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

      $for N in range(0, BATCH_TILE, 4):
        v128_t vp${ABC[N:N+4]} = wasm_f32x4_add(vc2, wasm_f32x4_mul(vc3, vt${ABC[N:N+4]}));

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

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

      $for N in range(0, BATCH_TILE, 4):
        const v128_t vsignm${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));

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

    const uint32_t vidx0 = wasm_u32x4_extract_lane(vidx, 0);
    v128_t vl = wasm_v128_load32_zero((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx0));

    const uint32_t vidx1 = wasm_u32x4_extract_lane(vidx, 1);
    vl = wasm_v128_load32_lane((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx1), vl, 1);

    const uint32_t vidx2 = wasm_u32x4_extract_lane(vidx, 2);
    vl = wasm_v128_load32_lane((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx2), vl, 2);

    const uint32_t vidx3 = wasm_u32x4_extract_lane(vidx, 3);
    vl = wasm_v128_load32_lane((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx3), vl, 3);

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

    v128_t vt = wasm_f32x4_add(vz, wasm_f32x4_mul(vn, vminus_ln2_hi));
    vt = wasm_f32x4_add(vt, wasm_f32x4_mul(vn, vminus_ln2_lo));

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

    vt = wasm_f32x4_mul(vt, vs);
    vs = wasm_f32x4_sub(vs, vone);
    vp = wasm_f32x4_add(vt, wasm_f32x4_mul(vp, 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}(vsat_cutoff, wasm_f32x4_mul(vx, vprescale));

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

    const uint32_t vidx0 = wasm_u32x4_extract_lane(vidx, 0);
    v128_t vl = wasm_v128_load32_zero((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx0));

    const uint32_t vidx1 = wasm_u32x4_extract_lane(vidx, 1);
    vl = wasm_v128_load32_lane((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx1), vl, 1);

    const uint32_t vidx2 = wasm_u32x4_extract_lane(vidx, 2);
    vl = wasm_v128_load32_lane((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx2), vl, 2);

    const uint32_t vidx3 = wasm_u32x4_extract_lane(vidx, 3);
    vl = wasm_v128_load32_lane((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx3), vl, 3);

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

    v128_t vt = wasm_f32x4_add(vz, wasm_f32x4_mul(vn, vminus_ln2_hi));
    vt = wasm_f32x4_add(vt, wasm_f32x4_mul(vn, vminus_ln2_lo));

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

    vt = wasm_f32x4_mul(vt, vs);
    vs = wasm_f32x4_sub(vs, vone);
    vp = wasm_f32x4_add(vt, wasm_f32x4_mul(vp, 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);
    }
  }
}