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


void xnn_f32_hswish_ukernel__wasmsimd_x${BATCH_TILE}(
    size_t n,
    const float* x,
    float* y,
    const union xnn_f32_hswish_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
{
  assert(n != 0);
  assert(n % sizeof(float) == 0);

  const v128_t vsixth = wasm_v32x4_load_splat(&params->scalar.sixth);
  const v128_t vthree = wasm_v32x4_load_splat(&params->scalar.three);
  const v128_t vsix = wasm_v32x4_load_splat(&params->scalar.six);
  const v128_t vzero = wasm_f32x4_splat(0.0f);

  $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};

      $for N in range(0, BATCH_TILE, 4):
        v128_t vacc${ABC[N:N+4]} = wasm_f32x4_add(vx${ABC[N:N+4]}, vthree);
        vx${ABC[N:N+4]} = wasm_f32x4_mul(vx${ABC[N:N+4]}, vsixth);

      $for N in range(0, BATCH_TILE, 4):
        vacc${ABC[N:N+4]} = wasm_i32x4_max(vacc${ABC[N:N+4]}, vzero);

      $for N in range(0, BATCH_TILE, 4):
        vacc${ABC[N:N+4]} = wasm_i32x4_min(vacc${ABC[N:N+4]}, vsix);

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

      wasm_v128_store(y, vacc${ABC[0:4]});
      $for N in range(4, BATCH_TILE, 4):
        wasm_v128_store(y + ${N}, vacc${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;

    v128_t vacc = wasm_f32x4_add(vx, vthree);
    vx = wasm_f32x4_mul(vx, vsixth);
    vacc = wasm_i32x4_max(vacc, vzero);
    vacc = wasm_i32x4_min(vacc, vsix);
    vacc = wasm_f32x4_mul(vacc, vx);

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

    v128_t vacc = wasm_f32x4_add(vx, vthree);
    vx = wasm_f32x4_mul(vx, vsixth);
    vacc = wasm_i32x4_max(vacc, vzero);
    vacc = wasm_i32x4_min(vacc, vsix);
    vacc = wasm_f32x4_mul(vacc, vx);

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