// 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.

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

#include <${SSE_HEADER}>

#include <xnnpack/dwconv.h>


$ISA = {2: "sse2", 3: "ssse3", 4: "sse41"}[SSE]
void xnn_qs8_dwconv_minmax_ukernel_up${CHANNEL_TILE}x${KERNEL_TILE}__${ISA}_mul16(
    size_t channels,
    size_t output_width,
    const int8_t** input,
    const void* weights,
    int8_t* output,
    size_t input_stride,
    size_t output_increment,
    size_t input_offset,
    const int8_t* zero,
    const union xnn_qs8_gemm_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
{
  assert(channels != 0);
  assert(output_width != 0);

  do {
    $for K in range(KERNEL_TILE):
      const int8_t* i${K} = input[${K}];
      assert(i${K} != NULL);
      if XNN_UNPREDICTABLE(i${K} != zero) {
        i${K} = (const int8_t*) ((uintptr_t) i${K} + input_offset);
      }
    input = (const int8_t**) ((uintptr_t) input + input_stride);

    size_t c = channels;
    const void* w = weights;
    for (; c >= ${CHANNEL_TILE}; c -= ${CHANNEL_TILE}) {
      __m128i vacc${ABC[0:4]} = _mm_loadu_si128((const __m128i*) w);
      $for C in range(4, CHANNEL_TILE, 4):
        __m128i vacc${ABC[C:C+4]} = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + ${C} * sizeof(int32_t)));

      $for K in range(KERNEL_TILE):

        $for C in range(0, CHANNEL_TILE, 8):
          $if C == 0:
            const __m128i vi${K}x${ABC[0:8]} = _mm_loadl_epi64((const __m128i*) i${K});
          $else:
            const __m128i vi${K}x${ABC[C:C+8]} = _mm_loadl_epi64((const __m128i*) (i${K} + ${C}));
          $if SSE >= 4:
            const __m128i vxi${K}x${ABC[C:C+8]} = _mm_cvtepi8_epi16(vi${K}x${ABC[C:C+8]});
          const __m128i vk${K}x${ABC[C:C+8]} = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + ${CHANNEL_TILE} * sizeof(int32_t) + ${K * CHANNEL_TILE + C} * sizeof(int8_t)));
          $if SSE >= 4:
            const __m128i vxk${K}x${ABC[C:C+8]} = _mm_cvtepi8_epi16(vk${K}x${ABC[C:C+8]});
        i${K} += ${CHANNEL_TILE};

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

        $for C in range(0, CHANNEL_TILE, 8):
          const __m128i vp${K}x${ABC[C:C+8]}lo = _mm_mullo_epi16(vxi${K}x${ABC[C:C+8]}, vxk${K}x${ABC[C:C+8]});
          const __m128i vp${K}x${ABC[C:C+8]}hi = _mm_mulhi_epi16(vxi${K}x${ABC[C:C+8]}, vxk${K}x${ABC[C:C+8]});

        $for C in range(0, CHANNEL_TILE, 8):
          vacc${ABC[C:C+4]} = _mm_add_epi32(vacc${ABC[C:C+4]}, _mm_unpacklo_epi16(vp${K}x${ABC[C:C+8]}lo, vp${K}x${ABC[C:C+8]}hi));
          vacc${ABC[C+4:C+8]} = _mm_add_epi32(vacc${ABC[C+4:C+8]}, _mm_unpackhi_epi16(vp${K}x${ABC[C:C+8]}lo, vp${K}x${ABC[C:C+8]}hi));

      w = (const void*) ((uintptr_t) w + ${CHANNEL_TILE} * sizeof(int32_t) + ${KERNEL_TILE * CHANNEL_TILE} * sizeof(int8_t));

      const __m128i vmultiplier = _mm_load_si128((const __m128i*) params->sse2.multiplier);
      const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse2.rounding);

      $if SSE >= 4:
        $for C in range(0, CHANNEL_TILE, 4):
          const __m128i vacc${ABC[C+1:C+4:2]} = _mm_shuffle_epi32(vacc${ABC[C:C+4]}, _MM_SHUFFLE(3, 3, 1, 1));
          const __m128i vprod${ABC[C:C+4:2]} = _mm_add_epi64(_mm_mul_epi32(vacc${ABC[C:C+4]}, vmultiplier), vrounding);
          const __m128i vprod${ABC[C+1:C+4:2]} = _mm_add_epi64(_mm_mul_epi32(vacc${ABC[C+1:C+4:2]}, vmultiplier), vrounding);

        $for C in range(0, CHANNEL_TILE, 4):
          const __m128i vq31prod${ABC[C:C+4:2]} = _mm_srli_epi64(vprod${ABC[C:C+4:2]}, 31);
          const __m128i vq31prod${ABC[C+1:C+4:2]} = _mm_add_epi64(vprod${ABC[C+1:C+4:2]}, vprod${ABC[C+1:C+4:2]});

        $for C in range(0, CHANNEL_TILE, 4):
          const __m128i vq31prod${ABC[C:C+4]} = _mm_blend_epi16(vq31prod${ABC[C:C+4:2]}, vq31prod${ABC[C+1:C+4:2]}, 0xCC);
      $else:
        $for C in range(0, CHANNEL_TILE, 4):
          const __m128i vnmask${ABC[C:C+4]} = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc${ABC[C:C+4]});

        $for C in range(0, CHANNEL_TILE, 4):
          $if SSE >= 3:
            const __m128i vabsacc${ABC[C:C+4]} = _mm_abs_epi32(vacc${ABC[C:C+4]});
          $else:
            const __m128i vabsacc${ABC[C:C+4]} = _mm_sub_epi32(_mm_xor_si128(vacc${ABC[C:C+4]}, vnmask${ABC[C:C+4]}), vnmask${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${ABC[C:C+4]}, _MM_SHUFFLE(3, 3, 1, 1));
          const __m128i vabsprod${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);

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

        $for C in range(0, CHANNEL_TILE, 4):
          const __m128i vprod${ABC[C:C+4:2]} = _mm_sub_epi64(_mm_xor_si128(vabsprod${ABC[C:C+4:2]}, vnmask${ABC[C:C+4:2]}), vnmask${ABC[C:C+4:2]});
          const __m128i vprod${ABC[C+1:C+4:2]} = _mm_sub_epi64(_mm_xor_si128(vabsprod${ABC[C+1:C+4:2]}, vnmask${ABC[C+1:C+4:2]}), vnmask${ABC[C+1:C+4:2]});

        $for C in range(0, CHANNEL_TILE, 4):
          const __m128i vq31prod${ABC[C:C+4:2]} = _mm_srli_epi64(_mm_add_epi64(vprod${ABC[C:C+4:2]}, vrounding), 31);
          const __m128i vq31prod${ABC[C+1:C+4:2]} = _mm_srli_epi64(_mm_add_epi64(vprod${ABC[C+1:C+4:2]}, vrounding), 31);

        $for C in range(0, CHANNEL_TILE, 4):
          const __m128i vq31prod${ABC[C:C+4:2]}${ABC[C+1:C+4:2]} = _mm_castps_si128(_mm_shuffle_ps(
              _mm_castsi128_ps(vq31prod${ABC[C:C+4:2]}), _mm_castsi128_ps(vq31prod${ABC[C+1:C+4:2]}), _MM_SHUFFLE(2, 0, 2, 0)));

        $for C in range(0, CHANNEL_TILE, 4):
          const __m128i vq31prod${ABC[C:C+4]} = _mm_shuffle_epi32(vq31prod${ABC[C:C+4:2]}${ABC[C+1:C+4:2]}, _MM_SHUFFLE(3, 1, 2, 0));

      const __m128i vremainder_mask = _mm_load_si128((const __m128i*) params->sse2.remainder_mask);
      $for C in range(0, CHANNEL_TILE, 4):
        const __m128i vrem${ABC[C:C+4]} =
          _mm_add_epi32(_mm_and_si128(vq31prod${ABC[C:C+4]}, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod${ABC[C:C+4]}));

      const __m128i vremainder_threshold = _mm_load_si128((const __m128i*) params->sse2.remainder_threshold);
      const __m128i vshift = _mm_load_si128((const __m128i*) params->sse2.shift);
      $for C in range(0, CHANNEL_TILE, 4):
        vacc${ABC[C:C+4]} =
          _mm_sub_epi32(_mm_sra_epi32(vq31prod${ABC[C:C+4]}, vshift), _mm_cmpgt_epi32(vrem${ABC[C:C+4]}, vremainder_threshold));

      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(vacc${ABC[C:C+4]}, vacc${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};
    }
    if XNN_UNLIKELY(c != 0) {
      $if CHANNEL_TILE > 8:
        const int8_t* k = (const int8_t*) ((uintptr_t) w + ${CHANNEL_TILE} * sizeof(int32_t));
      ${"do " if CHANNEL_TILE > 8 else ""}{
        __m128i vacc${ABC[0:4]} = _mm_loadu_si128((const __m128i*) w);
        __m128i vacc${ABC[4:8]} = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + 4 * sizeof(int32_t)));

        $for K in range(KERNEL_TILE):

          const __m128i vi${K}x${ABC[0:8]} = _mm_loadl_epi64((const __m128i*) i${K});
          $if SSE >= 4:
            const __m128i vxi${K}x${ABC[0:8]} = _mm_cvtepi8_epi16(vi${K}x${ABC[0:8]});
          $if CHANNEL_TILE > 8:
            $if K == 0:
              const __m128i vk${K}x${ABC[0:8]} = _mm_loadl_epi64((const __m128i*) k);
            $else:
              const __m128i vk${K}x${ABC[0:8]} = _mm_loadl_epi64((const __m128i*) (k + ${K * CHANNEL_TILE}));
          $else:
            const __m128i vk${K}x${ABC[0:8]} = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + ${CHANNEL_TILE} * sizeof(int32_t) + ${K * CHANNEL_TILE} * sizeof(int8_t)));
          $if SSE >= 4:
            const __m128i vxk${K}x${ABC[0:8]} = _mm_cvtepi8_epi16(vk${K}x${ABC[0:8]});
          $if CHANNEL_TILE > 8:
            i${K} += 8;

          $if SSE < 4:
            const __m128i vxi${K}x${ABC[0:8]} = _mm_unpacklo_epi8(vi${K}x${ABC[0:8]}, _mm_cmpgt_epi8(_mm_setzero_si128(), vi${K}x${ABC[0:8]}));
            const __m128i vxk${K}x${ABC[0:8]} = _mm_unpacklo_epi8(vk${K}x${ABC[0:8]}, _mm_cmpgt_epi8(_mm_setzero_si128(), vk${K}x${ABC[0:8]}));

          const __m128i vp${K}x${ABC[0:8]}lo = _mm_mullo_epi16(vxi${K}x${ABC[0:8]}, vxk${K}x${ABC[0:8]});
          const __m128i vp${K}x${ABC[0:8]}hi = _mm_mulhi_epi16(vxi${K}x${ABC[0:8]}, vxk${K}x${ABC[0:8]});

          vacc${ABC[0:4]} = _mm_add_epi32(vacc${ABC[0:4]}, _mm_unpacklo_epi16(vp${K}x${ABC[0:8]}lo, vp${K}x${ABC[0:8]}hi));
          vacc${ABC[4:8]} = _mm_add_epi32(vacc${ABC[4:8]}, _mm_unpackhi_epi16(vp${K}x${ABC[0:8]}lo, vp${K}x${ABC[0:8]}hi));

        $if CHANNEL_TILE > 8:
          w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t));
          k += 8;

        const __m128i vmultiplier = _mm_load_si128((const __m128i*) params->sse2.multiplier);
        const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse2.rounding);

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

          const __m128i vprod${ABC[0:4:2]} = _mm_add_epi64(_mm_mul_epi32(vacc${ABC[0:4]}, vmultiplier), vrounding);
          const __m128i vprod${ABC[4:8:2]} = _mm_add_epi64(_mm_mul_epi32(vacc${ABC[4:8]}, vmultiplier), vrounding);

          const __m128i vprod${ABC[1:4:2]} = _mm_add_epi64(_mm_mul_epi32(vacc${ABC[1:4:2]}, vmultiplier), vrounding);
          const __m128i vprod${ABC[5:8:2]} = _mm_add_epi64(_mm_mul_epi32(vacc${ABC[5:8:2]}, vmultiplier), vrounding);

          const __m128i vq31prod${ABC[0:4:2]} = _mm_srli_epi64(vprod${ABC[0:4:2]}, 31);
          const __m128i vq31prod${ABC[1:4:2]} = _mm_add_epi64(vprod${ABC[1:4:2]}, vprod${ABC[1:4:2]});
          const __m128i vq31prod${ABC[4:8:2]} = _mm_srli_epi64(vprod${ABC[4:8:2]}, 31);
          const __m128i vq31prod${ABC[5:8:2]} = _mm_add_epi64(vprod${ABC[5:8:2]}, vprod${ABC[5:8:2]});

          const __m128i vq31prod${ABC[0:4]} = _mm_blend_epi16(vq31prod${ABC[0:4:2]}, vq31prod${ABC[1:4:2]}, 0xCC);
          const __m128i vq31prod${ABC[4:8]} = _mm_blend_epi16(vq31prod${ABC[4:8:2]}, vq31prod${ABC[5:8:2]}, 0xCC);
        $else:
          const __m128i vnmask${ABC[0:4]} = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc${ABC[0:4]});
          const __m128i vnmask${ABC[4:8]} = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc${ABC[4:8]});

          $if SSE >= 3:
            const __m128i vabsacc${ABC[0:4]} = _mm_abs_epi32(vacc${ABC[0:4]});
            const __m128i vabsacc${ABC[4:8]} = _mm_abs_epi32(vacc${ABC[4:8]});
          $else:
            const __m128i vabsacc${ABC[0:4]} = _mm_sub_epi32(_mm_xor_si128(vacc${ABC[0:4]}, vnmask${ABC[0:4]}), vnmask${ABC[0:4]});
            const __m128i vabsacc${ABC[4:8]} = _mm_sub_epi32(_mm_xor_si128(vacc${ABC[4:8]}, vnmask${ABC[4:8]}), vnmask${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 vnmask${ABC[0:4:2]} = _mm_shuffle_epi32(vnmask${ABC[0:4]}, _MM_SHUFFLE(2, 2, 0, 0));
          const __m128i vnmask${ABC[1:4:2]} = _mm_shuffle_epi32(vnmask${ABC[0:4]}, _MM_SHUFFLE(3, 3, 1, 1));
          const __m128i vnmask${ABC[4:8:2]} = _mm_shuffle_epi32(vnmask${ABC[4:8]}, _MM_SHUFFLE(2, 2, 0, 0));
          const __m128i vnmask${ABC[5:8:2]} = _mm_shuffle_epi32(vnmask${ABC[4:8]}, _MM_SHUFFLE(3, 3, 1, 1));

          const __m128i vprod${ABC[0:4:2]} = _mm_sub_epi64(_mm_xor_si128(vabsprod${ABC[0:4:2]}, vnmask${ABC[0:4:2]}), vnmask${ABC[0:4:2]});
          const __m128i vprod${ABC[1:4:2]} = _mm_sub_epi64(_mm_xor_si128(vabsprod${ABC[1:4:2]}, vnmask${ABC[1:4:2]}), vnmask${ABC[1:4:2]});
          const __m128i vprod${ABC[4:8:2]} = _mm_sub_epi64(_mm_xor_si128(vabsprod${ABC[4:8:2]}, vnmask${ABC[4:8:2]}), vnmask${ABC[4:8:2]});
          const __m128i vprod${ABC[5:8:2]} = _mm_sub_epi64(_mm_xor_si128(vabsprod${ABC[5:8:2]}, vnmask${ABC[5:8:2]}), vnmask${ABC[5:8:2]});

          const __m128i vq31prod${ABC[0:4:2]} = _mm_srli_epi64(_mm_add_epi64(vprod${ABC[0:4:2]}, vrounding), 31);
          const __m128i vq31prod${ABC[1:4:2]} = _mm_srli_epi64(_mm_add_epi64(vprod${ABC[1:4:2]}, vrounding), 31);
          const __m128i vq31prod${ABC[4:8:2]} = _mm_srli_epi64(_mm_add_epi64(vprod${ABC[4:8:2]}, vrounding), 31);
          const __m128i vq31prod${ABC[5:8:2]} = _mm_srli_epi64(_mm_add_epi64(vprod${ABC[5:8:2]}, vrounding), 31);

          const __m128i vq31prod${ABC[0:4:2]}${ABC[1:4:2]} = _mm_castps_si128(_mm_shuffle_ps(
              _mm_castsi128_ps(vq31prod${ABC[0:4:2]}), _mm_castsi128_ps(vq31prod${ABC[1:4:2]}), _MM_SHUFFLE(2, 0, 2, 0)));
          const __m128i vq31prod${ABC[4:8:2]}${ABC[5:8:2]} = _mm_castps_si128(_mm_shuffle_ps(
              _mm_castsi128_ps(vq31prod${ABC[4:8:2]}), _mm_castsi128_ps(vq31prod${ABC[5:8:2]}), _MM_SHUFFLE(2, 0, 2, 0)));

          const __m128i vq31prod${ABC[0:4]} = _mm_shuffle_epi32(vq31prod${ABC[0:4:2]}${ABC[1:4:2]}, _MM_SHUFFLE(3, 1, 2, 0));
          const __m128i vq31prod${ABC[4:8]} = _mm_shuffle_epi32(vq31prod${ABC[4:8:2]}${ABC[5:8:2]}, _MM_SHUFFLE(3, 1, 2, 0));

        const __m128i vremainder_mask = _mm_load_si128((const __m128i*) params->sse2.remainder_mask);
        const __m128i vrem${ABC[0:4]} =
          _mm_add_epi32(_mm_and_si128(vq31prod${ABC[0:4]}, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod${ABC[0:4]}));
        const __m128i vrem${ABC[4:8]} =
          _mm_add_epi32(_mm_and_si128(vq31prod${ABC[4:8]}, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod${ABC[4:8]}));

        const __m128i vremainder_threshold = _mm_load_si128((const __m128i*) params->sse2.remainder_threshold);
        const __m128i vshift = _mm_load_si128((const __m128i*) params->sse2.shift);
        vacc${ABC[0:4]} =
          _mm_sub_epi32(_mm_sra_epi32(vq31prod${ABC[0:4]}, vshift), _mm_cmpgt_epi32(vrem${ABC[0:4]}, vremainder_threshold));
        vacc${ABC[4:8]} =
          _mm_sub_epi32(_mm_sra_epi32(vq31prod${ABC[4:8]}, vshift), _mm_cmpgt_epi32(vrem${ABC[4:8]}, vremainder_threshold));

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

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

    output = (int8_t*) ((uintptr_t) output + output_increment);
  } while (--output_width != 0);
}