// 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.

$SSE_HEADER = {2: "emmintrin.h", 3: "tmmintrin.h", 4: "smmintrin.h"}[SSE]
$assert CHANNEL_TILE % 8 == 0
$assert CHANNEL_TILE >= 8
$assert ROW_TILE >= 2
$assert ROW_SUBTILE >= 2
$assert ROW_SUBTILE <= ROW_TILE
$assert ACCUMULATORS >= 1
$assert ROW_TILE >= ACCUMULATORS * 2
$assert ROW_SUBTILE >= ACCUMULATORS * 2
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
#include <assert.h>

#include <${SSE_HEADER}>

#include <xnnpack/gavgpool.h>
#include <xnnpack/math.h>


$ISA = {2: "sse2", 3: "ssse3", 4: "sse41"}[SSE]
void xnn_qs8_gavgpool_minmax_ukernel_${ROW_TILE}p${ROW_SUBTILE}x__${ISA}_c${CHANNEL_TILE}${"" if ACCUMULATORS == 1 else "_acc%d" % ACCUMULATORS}(
    size_t rows,
    size_t channels,
    const int8_t* input,
    size_t input_stride,
    const int8_t* zero,
    int32_t* buffer,
    int8_t* output,
    const union xnn_qs8_avgpool_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
{
  assert(rows > ${ROW_TILE});
  assert(channels != 0);

  const int8_t* i0 = input;
  $for M in range(1, ROW_TILE):
    const int8_t* i${M} = (const int8_t*) ((uintptr_t) i${M-1} + input_stride);
  $if CHANNEL_TILE <= 16:
    const size_t input_increment = ${ROW_TILE} * input_stride - round_up_po2(channels, ${CHANNEL_TILE});
  $else:
    const size_t input_increment = ${ROW_TILE} * input_stride - round_up_po2(channels, 8);

  const __m128i vbias = _mm_load_si128((const __m128i*) params->sse2.bias);
  int32_t* b = buffer;
  size_t c = channels;
  for (; ${"c >= %d" % CHANNEL_TILE if CHANNEL_TILE > 16 else "c != 0"}; ${("c -= %d" if CHANNEL_TILE > 16 else "c = doz(c, %d)") % CHANNEL_TILE}) {
    $for M in range(ROW_TILE):
      $if SSE >= 4:
        const __m128i vxi${M}x${ABC[0:8]} = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i${M}));
        $for C in range(8, CHANNEL_TILE, 8):
          const __m128i vxi${M}x${ABC[C:C+8]} = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i${M} + ${C})));
      $else:
        const __m128i vi${M}x${ABC[0:8]} = _mm_loadl_epi64((const __m128i*) i${M});
        $for C in range(8, CHANNEL_TILE, 8):
          const __m128i vi${M}x${ABC[C:C+8]} = _mm_loadl_epi64((const __m128i*) (i${M} + ${C}));
      i${M} += ${CHANNEL_TILE};

    $if SSE < 4:
      $for M in range(ROW_TILE):
        $for C in range(0, CHANNEL_TILE, 8):
          const __m128i vxi${M}x${ABC[C:C+8]} = _mm_unpacklo_epi8(vi${M}x${ABC[C:C+8]}, _mm_cmpgt_epi8(_mm_setzero_si128(), vi${M}x${ABC[C:C+8]}));

    $for A in range(ACCUMULATORS):
      $for C in range(0, CHANNEL_TILE, 8):
        __m128i vacc${A}x${ABC[C:C+8]} = _mm_add_epi16(vxi${A*2}x${ABC[C:C+8]}, vxi${A*2+1}x${ABC[C:C+8]});

    $for M in range(ACCUMULATORS * 2, ROW_TILE):
      $for C in range(0, CHANNEL_TILE, 8):
        vacc${M % ACCUMULATORS}x${ABC[C:C+8]} = _mm_add_epi16(vacc${M % ACCUMULATORS}x${ABC[C:C+8]}, vxi${M}x${ABC[C:C+8]});

    $if ACCUMULATORS > 1:
      // Add up all accumulators to vacc0x${ABC[0:CHANNEL_TILE]}
      $ACC_SLICE = 1
      $while ACC_SLICE < ACCUMULATORS:
        $for A in range(0, ACCUMULATORS, ACC_SLICE * 2):
          $if A + ACC_SLICE < ACCUMULATORS:
            $for C in range(0, CHANNEL_TILE, 8):
              vacc${A}x${ABC[C:C+8]} = _mm_add_epi16(vacc${A}x${ABC[C:C+8]}, vacc${A + ACC_SLICE}x${ABC[C:C+8]});
        $ACC_SLICE *= 2

    $for C in range(0, CHANNEL_TILE, 8):
      $if SSE >= 4:
        const __m128i vacc${ABC[C:C+4]} = _mm_add_epi32(vbias, _mm_cvtepi16_epi32(vacc0x${ABC[C:C+8]}));
        const __m128i vacc${ABC[C+4:C+8]} = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vacc0x${ABC[C:C+8]}, _mm_cmpgt_epi16(_mm_setzero_si128(), vacc0x${ABC[C:C+8]})));
      $else:
        const __m128i vsgnacc0x${ABC[C:C+8]} = _mm_cmpgt_epi16(_mm_setzero_si128(), vacc0x${ABC[C:C+8]});
        const __m128i vacc${ABC[C:C+4]} = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vacc0x${ABC[C:C+8]}, vsgnacc0x${ABC[C:C+8]}));
        const __m128i vacc${ABC[C+4:C+8]} = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vacc0x${ABC[C:C+8]}, vsgnacc0x${ABC[C:C+8]}));

    _mm_store_si128((__m128i*) b, vacc${ABC[0:4]});
    $for C in range(4, CHANNEL_TILE, 4):
      _mm_store_si128((__m128i*) (b + ${C}), vacc${ABC[C:C+4]});
    b += ${CHANNEL_TILE};
  }
  $if CHANNEL_TILE > 16:
    if XNN_UNLIKELY(c != 0) {
      do {
        $for M in range(ROW_TILE):
          $if SSE >= 4:
            const __m128i vxi${M}x${ABC[0:8]} = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i${M}));
          $else:
            const __m128i vi${M}x${ABC[0:8]} = _mm_loadl_epi64((const __m128i*) i${M});
          i${M} += 8;

        $if SSE < 4:
          $for M in range(ROW_TILE):
            const __m128i vxi${M}x${ABC[0:8]} = _mm_unpacklo_epi8(vi${M}x${ABC[0:8]}, _mm_cmpgt_epi8(_mm_setzero_si128(), vi${M}x${ABC[0:8]}));

        $for A in range(ACCUMULATORS):
          __m128i vacc${A}x${ABC[0:8]} = _mm_add_epi16(vxi${A*2}x${ABC[0:8]}, vxi${A*2+1}x${ABC[0:8]});

        $for M in range(ACCUMULATORS * 2, ROW_TILE):
          vacc${M % ACCUMULATORS}x${ABC[0:8]} = _mm_add_epi16(vacc${M % ACCUMULATORS}x${ABC[0:8]}, vxi${M}x${ABC[0:8]});

        $if ACCUMULATORS > 1:
          // Add up all accumulators to vacc0x${ABC[0:8]}
          $ACC_SLICE = 1
          $while ACC_SLICE < ACCUMULATORS:
            $for A in range(0, ACCUMULATORS, ACC_SLICE * 2):
              $if A + ACC_SLICE < ACCUMULATORS:
                vacc${A}x${ABC[0:8]} = _mm_add_epi16(vacc${A}x${ABC[0:8]}, vacc${A + ACC_SLICE}x${ABC[0:8]});
            $ACC_SLICE *= 2

        $if SSE >= 4:
          const __m128i vacc${ABC[0:4]} = _mm_add_epi32(vbias, _mm_cvtepi16_epi32(vacc0x${ABC[0:8]}));
          const __m128i vacc${ABC[4:8]} = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vacc0x${ABC[0:8]}, _mm_cmpgt_epi16(_mm_setzero_si128(), vacc0x${ABC[0:8]})));
        $else:
          const __m128i vsgnacc0x${ABC[0:8]} = _mm_cmpgt_epi16(_mm_setzero_si128(), vacc0x${ABC[0:8]});
          const __m128i vacc${ABC[0:4]} = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vacc0x${ABC[0:8]}, vsgnacc0x${ABC[0:8]}));
          const __m128i vacc${ABC[4:8]} = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vacc0x${ABC[0:8]}, vsgnacc0x${ABC[0:8]}));

        _mm_store_si128((__m128i*) b, vacc${ABC[0:4]});
        _mm_store_si128((__m128i*) (b + 4), vacc${ABC[4:8]});
        b += 8;

        c = doz(c, 8);
      } while (c != 0);
    }

  for (rows -= ${ROW_TILE}; rows > ${ROW_SUBTILE}; rows -= ${ROW_SUBTILE}) {
    $for M in range(ROW_SUBTILE):
      i${M} = (const int8_t*) ((uintptr_t) i${M + ROW_TILE - ROW_SUBTILE} + input_increment);

    int32_t* b = buffer;
    size_t c = channels;
    for (; ${"c >= %d" % CHANNEL_TILE if CHANNEL_TILE > 16 else "c != 0"}; ${("c -= %d" if CHANNEL_TILE > 16 else "c = doz(c, %d)") % CHANNEL_TILE}) {
      $for M in range(ROW_SUBTILE):
        $if SSE >= 4:
          const __m128i vxi${M}x${ABC[0:8]} = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i${M}));
          $for C in range(8, CHANNEL_TILE, 8):
            const __m128i vxi${M}x${ABC[C:C+8]} = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i${M} + ${C})));
        $else:
          const __m128i vi${M}x${ABC[0:8]} = _mm_loadl_epi64((const __m128i*) i${M});
          $for C in range(8, CHANNEL_TILE, 8):
            const __m128i vi${M}x${ABC[C:C+8]} = _mm_loadl_epi64((const __m128i*) (i${M} + ${C}));
        i${M} += ${CHANNEL_TILE};

      $if SSE < 4:
        $for M in range(ROW_SUBTILE):
          $for C in range(0, CHANNEL_TILE, 8):
            const __m128i vxi${M}x${ABC[C:C+8]} = _mm_unpacklo_epi8(vi${M}x${ABC[C:C+8]}, _mm_cmpgt_epi8(_mm_setzero_si128(), vi${M}x${ABC[C:C+8]}));

      $for A in range(ACCUMULATORS):
        $for C in range(0, CHANNEL_TILE, 8):
          __m128i vacc${A}x${ABC[C:C+8]} = _mm_add_epi16(vxi${A*2}x${ABC[C:C+8]}, vxi${A*2+1}x${ABC[C:C+8]});

      $for M in range(ACCUMULATORS * 2, ROW_SUBTILE):
        $for C in range(0, CHANNEL_TILE, 8):
          vacc${M % ACCUMULATORS}x${ABC[C:C+8]} = _mm_add_epi16(vacc${M % ACCUMULATORS}x${ABC[C:C+8]}, vxi${M}x${ABC[C:C+8]});

      $if ACCUMULATORS > 1:
        // Add up all accumulators to vacc0x${ABC[0:CHANNEL_TILE]}
        $ACC_SLICE = 1
        $while ACC_SLICE < ACCUMULATORS:
          $for A in range(0, ACCUMULATORS, ACC_SLICE * 2):
            $if A + ACC_SLICE < ACCUMULATORS:
              $for C in range(0, CHANNEL_TILE, 8):
                vacc${A}x${ABC[C:C+8]} = _mm_add_epi16(vacc${A}x${ABC[C:C+8]}, vacc${A + ACC_SLICE}x${ABC[C:C+8]});
          $ACC_SLICE *= 2

      $for C in range(0, CHANNEL_TILE, 8):
        $if SSE >= 4:
          const __m128i vacc${ABC[C:C+4]} = _mm_add_epi32(_mm_cvtepi16_epi32(vacc0x${ABC[C:C+8]}), _mm_load_si128((const __m128i*) (b + ${C})));
          const __m128i vacc${ABC[C+4:C+8]} = _mm_add_epi32(_mm_unpackhi_epi16(vacc0x${ABC[C:C+8]}, _mm_cmpgt_epi16(_mm_setzero_si128(), vacc0x${ABC[C:C+8]})), _mm_load_si128((const __m128i*) (b + ${C+4})));
        $else:
          const __m128i vsgnacc0x${ABC[C:C+8]} = _mm_cmpgt_epi16(_mm_setzero_si128(), vacc0x${ABC[C:C+8]});
          const __m128i vacc${ABC[C:C+4]} = _mm_add_epi32(_mm_unpacklo_epi16(vacc0x${ABC[C:C+8]}, vsgnacc0x${ABC[C:C+8]}), _mm_load_si128((const __m128i*) (b + ${C})));
          const __m128i vacc${ABC[C+4:C+8]} = _mm_add_epi32(_mm_unpackhi_epi16(vacc0x${ABC[C:C+8]}, vsgnacc0x${ABC[C:C+8]}), _mm_load_si128((const __m128i*) (b + ${C+4})));

      _mm_store_si128((__m128i*) b, vacc${ABC[0:4]});
      $for C in range(4, CHANNEL_TILE, 4):
        _mm_store_si128((__m128i*) (b + ${C}), vacc${ABC[C:C+4]});
      b += ${CHANNEL_TILE};
    }
    $if CHANNEL_TILE > 16:
      if XNN_UNLIKELY(c != 0) {
        do {
          $for M in range(ROW_SUBTILE):
            $if SSE >= 4:
              const __m128i vxi${M}x${ABC[0:8]} = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i${M}));
            $else:
              const __m128i vi${M}x${ABC[0:8]} = _mm_loadl_epi64((const __m128i*) i${M});
            i${M} += 8;

          $if SSE < 4:
            $for M in range(ROW_SUBTILE):
              const __m128i vxi${M}x${ABC[0:8]} = _mm_unpacklo_epi8(vi${M}x${ABC[0:8]}, _mm_cmpgt_epi8(_mm_setzero_si128(), vi${M}x${ABC[0:8]}));

          $for A in range(ACCUMULATORS):
            __m128i vacc${A}x${ABC[0:8]} = _mm_add_epi16(vxi${A*2}x${ABC[0:8]}, vxi${A*2+1}x${ABC[0:8]});

          $for M in range(ACCUMULATORS * 2, ROW_SUBTILE):
            vacc${M % ACCUMULATORS}x${ABC[0:8]} = _mm_add_epi16(vacc${M % ACCUMULATORS}x${ABC[0:8]}, vxi${M}x${ABC[0:8]});

          $if ACCUMULATORS > 1:
            // Add up all accumulators to vacc0x${ABC[0:8]}
            $ACC_SLICE = 1
            $while ACC_SLICE < ACCUMULATORS:
              $for A in range(0, ACCUMULATORS, ACC_SLICE * 2):
                $if A + ACC_SLICE < ACCUMULATORS:
                  vacc${A}x${ABC[0:8]} = _mm_add_epi16(vacc${A}x${ABC[0:8]}, vacc${A + ACC_SLICE}x${ABC[0:8]});
              $ACC_SLICE *= 2

          $if SSE >= 4:
            const __m128i vacc${ABC[0:4]} = _mm_add_epi32(_mm_cvtepi16_epi32(vacc0x${ABC[0:8]}), _mm_load_si128((const __m128i*) b));
            const __m128i vacc${ABC[4:8]} = _mm_add_epi32(_mm_unpackhi_epi16(vacc0x${ABC[0:8]}, _mm_cmpgt_epi16(_mm_setzero_si128(), vacc0x${ABC[0:8]})), _mm_load_si128((const __m128i*) (b + 4)));
          $else:
            const __m128i vsgnacc0x${ABC[0:8]} = _mm_cmpgt_epi16(_mm_setzero_si128(), vacc0x${ABC[0:8]});
            const __m128i vacc${ABC[0:4]} = _mm_add_epi32(_mm_unpacklo_epi16(vacc0x${ABC[0:8]}, vsgnacc0x${ABC[0:8]}), _mm_load_si128((const __m128i*) b));
            const __m128i vacc${ABC[4:8]} = _mm_add_epi32(_mm_unpackhi_epi16(vacc0x${ABC[0:8]}, vsgnacc0x${ABC[0:8]}), _mm_load_si128((const __m128i*) (b + 4)));

          _mm_store_si128((__m128i*) b, vacc${ABC[0:4]});
          _mm_store_si128((__m128i*) (b + 4), vacc${ABC[4:8]});
          b += 8;

          c = doz(c, 8);
        } while (c != 0);
      }
  }

  i0 = (const int8_t*) ((uintptr_t) i${ROW_TILE - ROW_SUBTILE} + input_increment);
  $for M in range(1, ROW_SUBTILE):
    i${M} = (const int8_t*) ((uintptr_t) i${M + ROW_TILE - ROW_SUBTILE} + input_increment);
    $if M % 2 == 1:
      if XNN_UNPREDICTABLE(rows < ${M+1}) {
        i${M} = zero;
      }
    $else:
      if XNN_UNPREDICTABLE(rows <= ${M}) {
        i${M} = zero;
      }

  const __m128i vmultiplier = _mm_load_si128((const __m128i*) params->sse2.multiplier);
  const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse2.rounding);
  const __m128i vshift = _mm_loadl_epi64((const __m128i*) params->sse2.shift);
  while (channels >= ${CHANNEL_TILE}) {
    $for M in range(ROW_SUBTILE):
      $if SSE >= 4:
        const __m128i vxi${M}x${ABC[0:8]} = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i${M}));
        $for C in range(8, CHANNEL_TILE, 8):
          const __m128i vxi${M}x${ABC[C:C+8]} = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i${M} + ${C})));
      $else:
        const __m128i vi${M}x${ABC[0:8]} = _mm_loadl_epi64((const __m128i*) i${M});
        $for C in range(8, CHANNEL_TILE, 8):
          const __m128i vi${M}x${ABC[C:C+8]} = _mm_loadl_epi64((const __m128i*) (i${M} + ${C}));
      i${M} += ${CHANNEL_TILE};

    $if SSE < 4:
      $for M in range(ROW_SUBTILE):
        $for C in range(0, CHANNEL_TILE, 8):
          const __m128i vxi${M}x${ABC[C:C+8]} = _mm_unpacklo_epi8(vi${M}x${ABC[C:C+8]}, _mm_cmpgt_epi8(_mm_setzero_si128(), vi${M}x${ABC[C:C+8]}));

    $for A in range(ACCUMULATORS):
      $for C in range(0, CHANNEL_TILE, 8):
        __m128i vacc${A}x${ABC[C:C+8]} = _mm_add_epi16(vxi${A*2}x${ABC[C:C+8]}, vxi${A*2+1}x${ABC[C:C+8]});

    $for M in range(ACCUMULATORS * 2, ROW_SUBTILE):
      $for C in range(0, CHANNEL_TILE, 8):
        vacc${M % ACCUMULATORS}x${ABC[C:C+8]} = _mm_add_epi16(vacc${M % ACCUMULATORS}x${ABC[C:C+8]}, vxi${M}x${ABC[C:C+8]});

    $if ACCUMULATORS > 1:
      // Add up all accumulators to vacc0x${ABC[0:CHANNEL_TILE]}
      $ACC_SLICE = 1
      $while ACC_SLICE < ACCUMULATORS:
        $for A in range(0, ACCUMULATORS, ACC_SLICE * 2):
          $if A + ACC_SLICE < ACCUMULATORS:
            $for C in range(0, CHANNEL_TILE, 8):
              vacc${A}x${ABC[C:C+8]} = _mm_add_epi16(vacc${A}x${ABC[C:C+8]}, vacc${A + ACC_SLICE}x${ABC[C:C+8]});
        $ACC_SLICE *= 2

    $for C in range(0, CHANNEL_TILE, 8):
      $if SSE >= 4:
        const __m128i vacc${ABC[C:C+4]} = _mm_add_epi32(_mm_cvtepi16_epi32(vacc0x${ABC[C:C+8]}), _mm_load_si128((const __m128i*) (buffer + ${C})));
        const __m128i vacc${ABC[C+4:C+8]} = _mm_add_epi32(_mm_unpackhi_epi16(vacc0x${ABC[C:C+8]}, _mm_cmpgt_epi16(_mm_setzero_si128(), vacc0x${ABC[C:C+8]})), _mm_load_si128((const __m128i*) (buffer + ${C+4})));
      $else:
        const __m128i vsgnacc0x${ABC[C:C+8]} = _mm_cmpgt_epi16(_mm_setzero_si128(), vacc0x${ABC[C:C+8]});
        const __m128i vacc${ABC[C:C+4]} = _mm_add_epi32(_mm_unpacklo_epi16(vacc0x${ABC[C:C+8]}, vsgnacc0x${ABC[C:C+8]}), _mm_load_si128((const __m128i*) (buffer + ${C})));
        const __m128i vacc${ABC[C+4:C+8]} = _mm_add_epi32(_mm_unpackhi_epi16(vacc0x${ABC[C:C+8]}, vsgnacc0x${ABC[C:C+8]}), _mm_load_si128((const __m128i*) (buffer + ${C+4})));
    buffer += ${CHANNEL_TILE};

    $if SSE >= 3:
      $for C in range(0, CHANNEL_TILE, 4):
        const __m128i vabsacc${ABC[C:C+4]} = _mm_abs_epi32(vacc${ABC[C:C+4]});
    $else:
      $for C in range(0, CHANNEL_TILE, 4):
        const __m128i vsgnacc${ABC[C:C+4]} = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc${ABC[C:C+4]});

      $for C in range(0, CHANNEL_TILE, 4):
        const __m128i vabsacc${ABC[C:C+4]} = _mm_sub_epi32(_mm_xor_si128(vacc${ABC[C:C+4]}, vsgnacc${ABC[C:C+4]}), vsgnacc${ABC[C:C+4]});

    $for C in range(0, CHANNEL_TILE, 4):
      const __m128i vabsacc${ABC[C+1:C+4:2]} = _mm_shuffle_epi32(vabsacc${ABC[C:C+4]}, _MM_SHUFFLE(3, 3, 1, 1));

    $for C in range(0, CHANNEL_TILE, 4):
      const __m128i vabsprod${ABC[C:C+4:2]} = _mm_mul_epu32(vabsacc${ABC[C:C+4]}, vmultiplier);
      const __m128i vabsprod${ABC[C+1:C+4:2]} = _mm_mul_epu32(vabsacc${ABC[C+1:C+4:2]}, vmultiplier);

    $for C in range(0, CHANNEL_TILE, 4):
      const __m128i vabsout${ABC[C:C+4:2]} = _mm_srl_epi64(_mm_add_epi64(vabsprod${ABC[C:C+4:2]}, vrounding), vshift);
      const __m128i vabsout${ABC[C+1:C+4:2]} = _mm_srl_epi64(_mm_add_epi64(vabsprod${ABC[C+1:C+4:2]}, vrounding), vshift);

    $if SSE >= 4:
      $for C in range(0, CHANNEL_TILE, 4):
        const __m128i vabsout${ABC[C:C+4]} = _mm_blend_epi16(vabsout${ABC[C:C+4:2]}, _mm_shuffle_epi32(vabsout${ABC[C+1:C+4:2]}, _MM_SHUFFLE(2, 2, 0, 0)), 0xCC);
    $else:
      $for C in range(0, CHANNEL_TILE, 4):
        const __m128i vabsout${ABC[C:C+4:2]}${ABC[C+1:C+4:2]} = _mm_castps_si128(
            _mm_shuffle_ps(_mm_castsi128_ps(vabsout${ABC[C:C+4:2]}), _mm_castsi128_ps(vabsout${ABC[C+1:C+4:2]}), _MM_SHUFFLE(2, 0, 2, 0)));

      $for C in range(0, CHANNEL_TILE, 4):
        const __m128i vabsout${ABC[C:C+4]} = _mm_shuffle_epi32(vabsout${ABC[C:C+4:2]}${ABC[C+1:C+4:2]}, _MM_SHUFFLE(3, 1, 2, 0));

    $if SSE >= 3:
      $for C in range(0, CHANNEL_TILE, 4):
        const __m128i vout${ABC[C:C+4]} = _mm_sign_epi32(vabsout${ABC[C:C+4]}, vacc${ABC[C:C+4]});
    $else:
      $for C in range(0, CHANNEL_TILE, 4):
        const __m128i vout${ABC[C:C+4]} = _mm_sub_epi32(_mm_xor_si128(vabsout${ABC[C:C+4]}, vsgnacc${ABC[C:C+4]}), vsgnacc${ABC[C:C+4]});

    const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
    $for C in range(0, CHANNEL_TILE, 8):
      __m128i vout${ABC[C:C+8]} = _mm_adds_epi16(_mm_packs_epi32(vout${ABC[C:C+4]}, vout${ABC[C+4:C+8]}), voutput_zero_point);

    const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
    const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
    $for C in range(0, CHANNEL_TILE, 8):
      vout${ABC[C:C+8]} = _mm_min_epi16(_mm_max_epi16(vout${ABC[C:C+8]}, voutput_min), voutput_max);

    $for C in range(0, CHANNEL_TILE, 16):
      $if C + 8 < CHANNEL_TILE:
        __m128i vout${ABC[C:C+16]} = _mm_packs_epi16(vout${ABC[C:C+8]}, vout${ABC[C+8:C+16]});
      $else:
        __m128i vout${ABC[C:C+8]}${ABC[C:C+8]} = _mm_packs_epi16(vout${ABC[C:C+8]}, vout${ABC[C:C+8]});

    $if CHANNEL_TILE > 8:
      _mm_storeu_si128((__m128i*) output, vout${ABC[0:16]});
    $else:
      _mm_storel_epi64((__m128i*) output, vout${ABC[0:8]}${ABC[0:8]});
    $for C in range(16, CHANNEL_TILE, 16):
      $if C + 8 < CHANNEL_TILE:
        _mm_storeu_si128((__m128i*) (output + ${C}), vout${ABC[C:C+16]});
      $else:
        _mm_storel_epi64((__m128i*) (output + ${C}), vout${ABC[C:C+8]}${ABC[C:C+8]});
    output += ${CHANNEL_TILE};

    channels -= ${CHANNEL_TILE};
  }
  if XNN_UNLIKELY(channels != 0) {
    ${"do " if CHANNEL_TILE > 8 else ""}{
      $for M in range(ROW_SUBTILE):
        $if SSE >= 4:
          const __m128i vxi${M}x${ABC[0:8]} = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i${M}));
        $else:
          const __m128i vi${M}x${ABC[0:8]} = _mm_loadl_epi64((const __m128i*) i${M});
        i${M} += 8;

      $if SSE < 4:
        $for M in range(ROW_SUBTILE):
          const __m128i vxi${M}x${ABC[0:8]} = _mm_unpacklo_epi8(vi${M}x${ABC[0:8]}, _mm_cmpgt_epi8(_mm_setzero_si128(), vi${M}x${ABC[0:8]}));

      $for A in range(ACCUMULATORS):
        __m128i vacc${A}x${ABC[0:8]} = _mm_add_epi16(vxi${A*2}x${ABC[0:8]}, vxi${A*2+1}x${ABC[0:8]});

      $for M in range(ACCUMULATORS * 2, ROW_SUBTILE):
        vacc${M % ACCUMULATORS}x${ABC[0:8]} = _mm_add_epi16(vacc${M % ACCUMULATORS}x${ABC[0:8]}, vxi${M}x${ABC[0:8]});

      $if ACCUMULATORS > 1:
        // Add up all accumulators to vacc0x${ABC[0:8]}
        $ACC_SLICE = 1
        $while ACC_SLICE < ACCUMULATORS:
          $for A in range(0, ACCUMULATORS, ACC_SLICE * 2):
            $if A + ACC_SLICE < ACCUMULATORS:
              vacc${A}x${ABC[0:8]} = _mm_add_epi16(vacc${A}x${ABC[0:8]}, vacc${A + ACC_SLICE}x${ABC[0:8]});
          $ACC_SLICE *= 2

      $if SSE >= 4:
        const __m128i vacc${ABC[0:4]} = _mm_add_epi32(_mm_cvtepi16_epi32(vacc0x${ABC[0:8]}), _mm_load_si128((const __m128i*) buffer));
        const __m128i vacc${ABC[4:8]} = _mm_add_epi32(_mm_unpackhi_epi16(vacc0x${ABC[0:8]}, _mm_cmpgt_epi16(_mm_setzero_si128(), vacc0x${ABC[0:8]})), _mm_load_si128((const __m128i*) (buffer + 4)));
      $else:
        const __m128i vsgnacc0x${ABC[0:8]} = _mm_cmpgt_epi16(_mm_setzero_si128(), vacc0x${ABC[0:8]});
        const __m128i vacc${ABC[0:4]} = _mm_add_epi32(_mm_unpacklo_epi16(vacc0x${ABC[0:8]}, vsgnacc0x${ABC[0:8]}), _mm_load_si128((const __m128i*) buffer));
        const __m128i vacc${ABC[4:8]} = _mm_add_epi32(_mm_unpackhi_epi16(vacc0x${ABC[0:8]}, vsgnacc0x${ABC[0:8]}), _mm_load_si128((const __m128i*) (buffer + 4)));
      buffer += 8;

      $if SSE >= 3:
        const __m128i vabsacc${ABC[0:4]} = _mm_abs_epi32(vacc${ABC[0:4]});
        const __m128i vabsacc${ABC[4:8]} = _mm_abs_epi32(vacc${ABC[4:8]});
      $else:
        const __m128i vsgnacc${ABC[0:4]} = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc${ABC[0:4]});
        const __m128i vsgnacc${ABC[4:8]} = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc${ABC[4:8]});

        const __m128i vabsacc${ABC[0:4]} = _mm_sub_epi32(_mm_xor_si128(vacc${ABC[0:4]}, vsgnacc${ABC[0:4]}), vsgnacc${ABC[0:4]});
        const __m128i vabsacc${ABC[4:8]} = _mm_sub_epi32(_mm_xor_si128(vacc${ABC[4:8]}, vsgnacc${ABC[4:8]}), vsgnacc${ABC[4:8]});

      const __m128i vabsacc${ABC[1:4:2]} = _mm_shuffle_epi32(vabsacc${ABC[0:4]}, _MM_SHUFFLE(3, 3, 1, 1));
      const __m128i vabsacc${ABC[5:8:2]} = _mm_shuffle_epi32(vabsacc${ABC[4:8]}, _MM_SHUFFLE(3, 3, 1, 1));

      const __m128i vabsprod${ABC[0:4:2]} = _mm_mul_epu32(vabsacc${ABC[0:4]}, vmultiplier);
      const __m128i vabsprod${ABC[1:4:2]} = _mm_mul_epu32(vabsacc${ABC[1:4:2]}, vmultiplier);
      const __m128i vabsprod${ABC[4:8:2]} = _mm_mul_epu32(vabsacc${ABC[4:8]}, vmultiplier);
      const __m128i vabsprod${ABC[5:8:2]} = _mm_mul_epu32(vabsacc${ABC[5:8:2]}, vmultiplier);

      const __m128i vabsout${ABC[0:4:2]} = _mm_srl_epi64(_mm_add_epi64(vabsprod${ABC[0:4:2]}, vrounding), vshift);
      const __m128i vabsout${ABC[1:4:2]} = _mm_srl_epi64(_mm_add_epi64(vabsprod${ABC[1:4:2]}, vrounding), vshift);
      const __m128i vabsout${ABC[4:8:2]} = _mm_srl_epi64(_mm_add_epi64(vabsprod${ABC[4:8:2]}, vrounding), vshift);
      const __m128i vabsout${ABC[5:8:2]} = _mm_srl_epi64(_mm_add_epi64(vabsprod${ABC[5:8:2]}, vrounding), vshift);

      $if SSE >= 4:
        const __m128i vabsout${ABC[0:4]} = _mm_blend_epi16(vabsout${ABC[0:4:2]}, _mm_shuffle_epi32(vabsout${ABC[1:4:2]}, _MM_SHUFFLE(2, 2, 0, 0)), 0xCC);
        const __m128i vabsout${ABC[4:8]} = _mm_blend_epi16(vabsout${ABC[4:8:2]}, _mm_shuffle_epi32(vabsout${ABC[5:8:2]}, _MM_SHUFFLE(2, 2, 0, 0)), 0xCC);
      $else:
        const __m128i vabsout${ABC[0:4:2]}${ABC[1:4:2]} = _mm_castps_si128(
            _mm_shuffle_ps(_mm_castsi128_ps(vabsout${ABC[0:4:2]}), _mm_castsi128_ps(vabsout${ABC[1:4:2]}), _MM_SHUFFLE(2, 0, 2, 0)));
        const __m128i vabsout${ABC[4:8:2]}${ABC[5:8:2]} = _mm_castps_si128(
            _mm_shuffle_ps(_mm_castsi128_ps(vabsout${ABC[4:8:2]}), _mm_castsi128_ps(vabsout${ABC[5:8:2]}), _MM_SHUFFLE(2, 0, 2, 0)));

        const __m128i vabsout${ABC[0:4]} = _mm_shuffle_epi32(vabsout${ABC[0:4:2]}${ABC[1:4:2]}, _MM_SHUFFLE(3, 1, 2, 0));
        const __m128i vabsout${ABC[4:8]} = _mm_shuffle_epi32(vabsout${ABC[4:8:2]}${ABC[5:8:2]}, _MM_SHUFFLE(3, 1, 2, 0));

      $if SSE >= 3:
        const __m128i vout${ABC[0:4]} = _mm_sign_epi32(vabsout${ABC[0:4]}, vacc${ABC[0:4]});
        const __m128i vout${ABC[4:8]} = _mm_sign_epi32(vabsout${ABC[4:8]}, vacc${ABC[4:8]});
      $else:
        const __m128i vout${ABC[0:4]} = _mm_sub_epi32(_mm_xor_si128(vabsout${ABC[0:4]}, vsgnacc${ABC[0:4]}), vsgnacc${ABC[0:4]});
        const __m128i vout${ABC[4:8]} = _mm_sub_epi32(_mm_xor_si128(vabsout${ABC[4:8]}, vsgnacc${ABC[4:8]}), vsgnacc${ABC[4:8]});

      const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
      __m128i vout${ABC[0:8]} = _mm_adds_epi16(_mm_packs_epi32(vout${ABC[0:4]}, vout${ABC[4:8]}), voutput_zero_point);

      const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
      const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
      vout${ABC[0:8]} = _mm_min_epi16(_mm_max_epi16(vout${ABC[0:8]}, voutput_min), voutput_max);

      __m128i vout${ABC[0:8]}${ABC[0:8]} = _mm_packs_epi16(vout${ABC[0:8]}, vout${ABC[0:8]});

      $if CHANNEL_TILE > 8:
        if XNN_LIKELY(channels >= 8) {
          _mm_storel_epi64((__m128i*) output, vout${ABC[0:8]}${ABC[0:8]});
          output += 8;
          channels -= 8;
        } else {
          if (channels & 4) {
            *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout${ABC[0:8]}${ABC[0:8]});
            vout${ABC[0:8]}${ABC[0:8]} = _mm_srli_epi64(vout${ABC[0:8]}${ABC[0:8]}, 32);
            output += 4;
          }
          if (channels & 2) {
            *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout${ABC[0:8]}${ABC[0:8]}, 0);
            vout${ABC[0:8]}${ABC[0:8]} = _mm_srli_epi32(vout${ABC[0:8]}${ABC[0:8]}, 16);
            output += 2;
          }
          if (channels & 1) {
            $if SSE >= 4:
              *output = (int8_t) _mm_extract_epi8(vout${ABC[0:8]}${ABC[0:8]}, 0);
            $else:
              *output = (int32_t) _mm_cvtsi128_si32(vout${ABC[0:8]}${ABC[0:8]});
            output += 1;
          }
          channels = 0;
        }
      $else:
        if (channels & 4) {
          *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout${ABC[0:8]}${ABC[0:8]});
          vout${ABC[0:8]}${ABC[0:8]} = _mm_srli_epi64(vout${ABC[0:8]}${ABC[0:8]}, 32);
          output += 4;
        }
        if (channels & 2) {
          *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout${ABC[0:8]}${ABC[0:8]}, 0);
          vout${ABC[0:8]}${ABC[0:8]} = _mm_srli_epi32(vout${ABC[0:8]}${ABC[0:8]}, 16);
          output += 2;
        }
        if (channels & 1) {
          $if SSE >= 4:
            *output = (int8_t) _mm_extract_epi8(vout${ABC[0:8]}${ABC[0:8]}, 0);
          $else:
            *output = (int32_t) _mm_cvtsi128_si32(vout${ABC[0:8]}${ABC[0:8]});
        }
    }${" while (channels != 0);" if CHANNEL_TILE > 8 else ""}
  }
}