// Copyright 2022 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 >= 16
$assert BATCH_TILE % 16 == 0
$SIMD_TILE = BATCH_TILE // 16
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
#include <assert.h>

#include <emmintrin.h>

#include <xnnpack/common.h>
#include <xnnpack/vlrelu.h>
#include <xnnpack/unaligned.h>


$XINT8_T = {"QS8": "int8_t", "QU8": "uint8_t"}[DATATYPE]
$_MM_PACKXS_EPI16 = {"QS8": "_mm_packs_epi16", "QU8": "_mm_packus_epi16"}[DATATYPE]
void xnn_${DATATYPE.lower()}_vlrelu_ukernel__sse2_x${BATCH_TILE}(
    size_t n,
    const ${XINT8_T}* x,
    ${XINT8_T}* y,
    const union xnn_${DATATYPE.lower()}_lrelu_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(n != 0);
  assert(n % sizeof(${XINT8_T}) == 0);
  assert(x != NULL);
  assert(y != NULL);

  const __m128i vinput_zero_point = _mm_load_si128((const __m128i*) params->sse2.input_zero_point);
  const __m128i vmultiplier_diff = _mm_load_si128((const __m128i*) params->sse2.multiplier_diff);
  const __m128i vmultiplier_base = _mm_load_si128((const __m128i*) params->sse2.multiplier_base);
  const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
  const __m128i vzero = _mm_setzero_si128();
  $if BATCH_TILE > 16:
    for (; n >= ${BATCH_TILE} * sizeof(${XINT8_T}); n -= ${BATCH_TILE} * sizeof(${XINT8_T})) {
      const __m128i vx${ABC[0]} = _mm_loadu_si128((const __m128i*) x);
      $for N in range(1, SIMD_TILE):
        const __m128i vx${ABC[N]} = _mm_loadu_si128((const __m128i*) (x + ${N * 16}));
      x += ${BATCH_TILE};

      $for N in range(SIMD_TILE):
        $if DATATYPE == "QU8":
          __m128i vextx${ABC[2*N]} = _mm_unpacklo_epi8(vx${ABC[N]}, vzero);
          __m128i vextx${ABC[2*N+1]} = _mm_unpackhi_epi8(vx${ABC[N]}, vzero);
        $else:
          const __m128i vm${ABC[N]} = _mm_cmpgt_epi8(_mm_setzero_si128(), vx${ABC[N]});
          __m128i vextx${ABC[2*N]} = _mm_unpacklo_epi8(vx${ABC[N]}, vm${ABC[N]});
          __m128i vextx${ABC[2*N+1]} = _mm_unpackhi_epi8(vx${ABC[N]}, vm${ABC[N]});

      $for N in range(2*SIMD_TILE):
        __m128i vmultiplier${ABC[N]} = _mm_cmpgt_epi16(vextx${ABC[N]}, vinput_zero_point);
        vextx${ABC[N]} = _mm_sub_epi16(vinput_zero_point, vextx${ABC[N]});

      $for N in range(2*SIMD_TILE):
        vmultiplier${ABC[N]} = _mm_and_si128(vmultiplier${ABC[N]}, vmultiplier_diff);

      $for N in range(2*SIMD_TILE):
        vmultiplier${ABC[N]} = _mm_xor_si128(vmultiplier${ABC[N]}, vmultiplier_base);

      $for N in range(2*SIMD_TILE):
        __m128i vprodlo${ABC[N]} = _mm_mullo_epi16(vextx${ABC[N]}, vmultiplier${ABC[N]});

      $for N in range(2*SIMD_TILE):
        vprodlo${ABC[N]} = _mm_srli_epi16(vprodlo${ABC[N]}, 7);
        __m128i vprodhi${ABC[N]} = _mm_mulhi_epi16(vextx${ABC[N]}, vmultiplier${ABC[N]});

      $for N in range(2*SIMD_TILE):
        vprodhi${ABC[N]} = _mm_slli_epi16(vprodhi${ABC[N]}, 8);
        vprodlo${ABC[N]} = _mm_avg_epu16(vprodlo${ABC[N]}, vzero);

      $for N in range(2*SIMD_TILE):
        __m128i vacc${ABC[N]} = _mm_add_epi16(vprodlo${ABC[N]}, vprodhi${ABC[N]});

      $for N in range(2*SIMD_TILE):
        vacc${ABC[N]} = _mm_adds_epi16(vacc${ABC[N]}, voutput_zero_point);

      $for N in range(SIMD_TILE):
        const __m128i vy${ABC[N]} = ${_MM_PACKXS_EPI16}(vacc${ABC[2*N]}, vacc${ABC[2*N+1]});

      _mm_storeu_si128((__m128i*) y, vy${ABC[0]});
      $for N in range(1, SIMD_TILE):
        _mm_storeu_si128((__m128i*) (y + ${N * 16}), vy${ABC[N]});
      y += ${BATCH_TILE};
    }
  for (; n >= 16 * sizeof(${XINT8_T}); n -= 16 * sizeof(${XINT8_T})) {
    const __m128i vx = _mm_loadu_si128((const __m128i*) x);
    x += 16;

    $if DATATYPE == "QU8":
      __m128i vextx0 = _mm_unpacklo_epi8(vx, vzero);
      __m128i vextx1 = _mm_unpackhi_epi8(vx, vzero);
    $else:
      const __m128i vm = _mm_cmpgt_epi8(_mm_setzero_si128(), vx);
      __m128i vextx0 = _mm_unpacklo_epi8(vx, vm);
      __m128i vextx1 = _mm_unpackhi_epi8(vx, vm);

    __m128i vmultiplier0 = _mm_cmpgt_epi16(vextx0, vinput_zero_point);
    __m128i vmultiplier1 = _mm_cmpgt_epi16(vextx1, vinput_zero_point);
    vextx0 = _mm_sub_epi16(vinput_zero_point, vextx0);
    vextx1 = _mm_sub_epi16(vinput_zero_point, vextx1);

    vmultiplier0 = _mm_and_si128(vmultiplier0, vmultiplier_diff);
    vmultiplier1 = _mm_and_si128(vmultiplier1, vmultiplier_diff);

    vmultiplier0 = _mm_xor_si128(vmultiplier0, vmultiplier_base);
    vmultiplier1 = _mm_xor_si128(vmultiplier1, vmultiplier_base);

    __m128i vprodlo0 = _mm_mullo_epi16(vextx0, vmultiplier0);
    __m128i vprodlo1 = _mm_mullo_epi16(vextx1, vmultiplier1);

    vprodlo0 = _mm_srli_epi16(vprodlo0, 7);
    vprodlo1 = _mm_srli_epi16(vprodlo1, 7);
    __m128i vprodhi0 = _mm_mulhi_epi16(vextx0, vmultiplier0);
    __m128i vprodhi1 = _mm_mulhi_epi16(vextx1, vmultiplier1);

    vprodhi0 = _mm_slli_epi16(vprodhi0, 8);
    vprodhi1 = _mm_slli_epi16(vprodhi1, 8);
    vprodlo0 = _mm_avg_epu16(vprodlo0, vzero);
    vprodlo1 = _mm_avg_epu16(vprodlo1, vzero);

    __m128i vacc0 = _mm_add_epi16(vprodlo0, vprodhi0);
    __m128i vacc1 = _mm_add_epi16(vprodlo1, vprodhi1);

    vacc0 = _mm_adds_epi16(vacc0, voutput_zero_point);
    vacc1 = _mm_adds_epi16(vacc1, voutput_zero_point);

    const __m128i vy = ${_MM_PACKXS_EPI16}(vacc0, vacc1);
    _mm_storeu_si128((__m128i*) y, vy);
    y += 16;
  }
  if XNN_UNLIKELY(n != 0) {
    assert(n >= 1 * sizeof(${XINT8_T}));
    assert(n <= 15 * sizeof(${XINT8_T}));

    const __m128i vx = _mm_loadu_si128((const __m128i*) x);

    $if DATATYPE == "QU8":
      __m128i vextx0 = _mm_unpacklo_epi8(vx, vzero);
      __m128i vextx1 = _mm_unpackhi_epi8(vx, vzero);
    $else:
      const __m128i vm = _mm_cmpgt_epi8(_mm_setzero_si128(), vx);
      __m128i vextx0 = _mm_unpacklo_epi8(vx, vm);
      __m128i vextx1 = _mm_unpackhi_epi8(vx, vm);

    __m128i vmultiplier0 = _mm_cmpgt_epi16(vextx0, vinput_zero_point);
    __m128i vmultiplier1 = _mm_cmpgt_epi16(vextx1, vinput_zero_point);
    vextx0 = _mm_sub_epi16(vinput_zero_point, vextx0);
    vextx1 = _mm_sub_epi16(vinput_zero_point, vextx1);

    vmultiplier0 = _mm_and_si128(vmultiplier0, vmultiplier_diff);
    vmultiplier1 = _mm_and_si128(vmultiplier1, vmultiplier_diff);

    vmultiplier0 = _mm_xor_si128(vmultiplier0, vmultiplier_base);
    vmultiplier1 = _mm_xor_si128(vmultiplier1, vmultiplier_base);

    __m128i vprodlo0 = _mm_mullo_epi16(vextx0, vmultiplier0);
    __m128i vprodlo1 = _mm_mullo_epi16(vextx1, vmultiplier1);

    vprodlo0 = _mm_srli_epi16(vprodlo0, 7);
    vprodlo1 = _mm_srli_epi16(vprodlo1, 7);
    __m128i vprodhi0 = _mm_mulhi_epi16(vextx0, vmultiplier0);
    __m128i vprodhi1 = _mm_mulhi_epi16(vextx1, vmultiplier1);

    vprodhi0 = _mm_slli_epi16(vprodhi0, 8);
    vprodhi1 = _mm_slli_epi16(vprodhi1, 8);
    vprodlo0 = _mm_avg_epu16(vprodlo0, vzero);
    vprodlo1 = _mm_avg_epu16(vprodlo1, vzero);

    __m128i vacc0 = _mm_add_epi16(vprodlo0, vprodhi0);
    __m128i vacc1 = _mm_add_epi16(vprodlo1, vprodhi1);

    vacc0 = _mm_adds_epi16(vacc0, voutput_zero_point);
    vacc1 = _mm_adds_epi16(vacc1, voutput_zero_point);

    __m128i vy = ${_MM_PACKXS_EPI16}(vacc0, vacc1);
    if (n & (8 * sizeof(${XINT8_T}))) {
      _mm_storel_epi64((__m128i*) y, vy);
      vy = _mm_unpackhi_epi64(vy, vy);
      y += 8;
    }
    if (n & (4 * sizeof(${XINT8_T}))) {
      unaligned_store_u32(y, (uint32_t) _mm_cvtsi128_si32(vy));
      vy = _mm_srli_epi64(vy, 32);
      y += 4;
    }
    uint32_t vy0 = (uint32_t) _mm_cvtsi128_si32(vy);
    if (n & (2 * sizeof(${XINT8_T}))) {
      unaligned_store_u16(y, (uint16_t) vy0);
      vy0 >>= 16;
      y += 2;
    }
    if (n & (1 * sizeof(${XINT8_T}))) {
      *y = (${XINT8_T}) vy0;
    }
  }
}