// 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 SSE in [1, 2, 4]
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
$SSE_HEADER = {1: "xmmintrin.h", 2: "emmintrin.h", 4: "smmintrin.h"}[SSE]
#include <assert.h>

#include <${SSE_HEADER}>

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


$ISA = {1: "sse", 2: "sse2", 4: "sse41"}[SSE]
void xnn_f32_vlrelu_ukernel__${ISA}_x${BATCH_TILE}(
    size_t n,
    const float* x,
    float* y,
    const union xnn_f32_lrelu_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(n != 0);
  assert(n % sizeof(float) == 0);

  const __m128 vslope = _mm_load_ps(params->sse.slope);
  $if SSE == 1:
    const __m128 vzero = _mm_setzero_ps();
  for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) {
    $if SSE == 1:
      __m128 vx${ABC[0:4]} = _mm_loadu_ps(x);
      $for N in range(4, BATCH_TILE, 4):
        __m128 vx${ABC[N:N+4]} = _mm_loadu_ps(x + ${N});
    $else:
      const __m128 vx${ABC[0:4]} = _mm_loadu_ps(x);
      $for N in range(4, BATCH_TILE, 4):
        const __m128 vx${ABC[N:N+4]} = _mm_loadu_ps(x + ${N});
    x += ${BATCH_TILE};

    $for N in range(0, BATCH_TILE, 4):
      $if SSE == 1:
        __m128 vacc${ABC[N:N+4]} = _mm_max_ps(_mm_setzero_ps(), vx${ABC[N:N+4]});
        vx${ABC[N:N+4]} = _mm_min_ps(vx${ABC[N:N+4]}, vzero);
      $else:
        __m128 vacc${ABC[N:N+4]} = _mm_mul_ps(vx${ABC[N:N+4]}, vslope);
        $if SSE == 2:
          const __m128 vmask${ABC[N:N+4]} = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx${ABC[N:N+4]})));

    $for N in range(0, BATCH_TILE, 4):
      $if SSE == 1:
        vacc${ABC[N:N+4]} = _mm_add_ps(vacc${ABC[N:N+4]}, _mm_mul_ps(vx${ABC[N:N+4]}, vslope));
      $elif SSE == 2:
        vacc${ABC[N:N+4]} = _mm_or_ps(_mm_and_ps(vacc${ABC[N:N+4]}, vmask${ABC[N:N+4]}), _mm_andnot_ps(vmask${ABC[N:N+4]}, vx${ABC[N:N+4]}));
      $elif SSE == 4:
        vacc${ABC[N:N+4]} = _mm_blendv_ps(vx${ABC[N:N+4]}, vacc${ABC[N:N+4]}, vx${ABC[N:N+4]});

    _mm_storeu_ps(y, vacc${ABC[0:4]});
    $for N in range(4, BATCH_TILE, 4):
      _mm_storeu_ps(y + ${N}, vacc${ABC[N:N+4]});
    y += ${BATCH_TILE};
  }
  $if BATCH_TILE > 4:
    for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
      $if SSE == 1:
        __m128 vx = _mm_loadu_ps(x);
      $else:
        const __m128 vx = _mm_loadu_ps(x);
      x += 4;

      $if SSE == 1:
        __m128 vacc = _mm_max_ps(_mm_setzero_ps(), vx);
        vx = _mm_min_ps(vx, vzero);
        vacc = _mm_add_ps(vacc, _mm_mul_ps(vx, vslope));
      $else:
        __m128 vacc = _mm_mul_ps(vx, vslope);
        $if SSE == 2:
          const __m128 vmask = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx)));
          vacc = _mm_or_ps(_mm_and_ps(vacc, vmask), _mm_andnot_ps(vmask, vx));
        $elif SSE == 4:
          vacc = _mm_blendv_ps(vx, vacc, vx);

      _mm_storeu_ps(y, vacc);
      y += 4;
    }
  if XNN_UNLIKELY(n != 0) {
    $if SSE == 1:
      __m128 vx = _mm_loadu_ps(x);

      __m128 vacc = _mm_max_ps(_mm_setzero_ps(), vx);
      vx = _mm_min_ps(vx, vzero);
      vacc = _mm_add_ps(vacc, _mm_mul_ps(vx, vslope));
    $else:
      const __m128 vx = _mm_loadu_ps(x);

      __m128 vacc = _mm_mul_ps(vx, vslope);
      $if SSE == 2:
        const __m128 vmask = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx)));
        vacc = _mm_or_ps(_mm_and_ps(vacc, vmask), _mm_andnot_ps(vmask, vx));
      $elif SSE == 4:
        vacc = _mm_blendv_ps(vx, vacc, vx);

    if (n & (2 * sizeof(float))) {
      _mm_storel_pi((__m64*) y, vacc);
      vacc = _mm_movehl_ps(vacc, vacc);
      y += 2;
    }
    if (n & (1 * sizeof(float))) {
      _mm_store_ss(y, vacc);
    }
  }
}