xref: /aosp_15_r20/external/XNNPACK/src/qu8-dwconv/gen/up8x25-minmax-fp32-sse2-mul16.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)

// Auto-generated file. Do not edit!
//   Template: src/qs8-dwconv/unipass-sse-mul16.c.in
//   Generator: tools/xngen
//
// 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.

#include <assert.h>

#include <emmintrin.h>

#include <xnnpack/dwconv.h>
#include <xnnpack/unaligned.h>


void xnn_qu8_dwconv_minmax_fp32_ukernel_up8x25__sse2_mul16(
    size_t channels,
    size_t output_width,
    const uint8_t** input,
    const void* weights,
    uint8_t* output,
    size_t input_stride,
    size_t output_increment,
    size_t input_offset,
    const uint8_t* zero,
    const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(channels != 0);
  assert(output_width != 0);

  do {
    const uint8_t* i0 = input[0];
    assert(i0 != NULL);
    if XNN_UNPREDICTABLE(i0 != zero) {
      i0 = (const uint8_t*) ((uintptr_t) i0 + input_offset);
    }
    const uint8_t* i1 = input[1];
    assert(i1 != NULL);
    if XNN_UNPREDICTABLE(i1 != zero) {
      i1 = (const uint8_t*) ((uintptr_t) i1 + input_offset);
    }
    const uint8_t* i2 = input[2];
    assert(i2 != NULL);
    if XNN_UNPREDICTABLE(i2 != zero) {
      i2 = (const uint8_t*) ((uintptr_t) i2 + input_offset);
    }
    const uint8_t* i3 = input[3];
    assert(i3 != NULL);
    if XNN_UNPREDICTABLE(i3 != zero) {
      i3 = (const uint8_t*) ((uintptr_t) i3 + input_offset);
    }
    const uint8_t* i4 = input[4];
    assert(i4 != NULL);
    if XNN_UNPREDICTABLE(i4 != zero) {
      i4 = (const uint8_t*) ((uintptr_t) i4 + input_offset);
    }
    const uint8_t* i5 = input[5];
    assert(i5 != NULL);
    if XNN_UNPREDICTABLE(i5 != zero) {
      i5 = (const uint8_t*) ((uintptr_t) i5 + input_offset);
    }
    const uint8_t* i6 = input[6];
    assert(i6 != NULL);
    if XNN_UNPREDICTABLE(i6 != zero) {
      i6 = (const uint8_t*) ((uintptr_t) i6 + input_offset);
    }
    const uint8_t* i7 = input[7];
    assert(i7 != NULL);
    if XNN_UNPREDICTABLE(i7 != zero) {
      i7 = (const uint8_t*) ((uintptr_t) i7 + input_offset);
    }
    const uint8_t* i8 = input[8];
    assert(i8 != NULL);
    if XNN_UNPREDICTABLE(i8 != zero) {
      i8 = (const uint8_t*) ((uintptr_t) i8 + input_offset);
    }
    const uint8_t* i9 = input[9];
    assert(i9 != NULL);
    if XNN_UNPREDICTABLE(i9 != zero) {
      i9 = (const uint8_t*) ((uintptr_t) i9 + input_offset);
    }
    const uint8_t* i10 = input[10];
    assert(i10 != NULL);
    if XNN_UNPREDICTABLE(i10 != zero) {
      i10 = (const uint8_t*) ((uintptr_t) i10 + input_offset);
    }
    const uint8_t* i11 = input[11];
    assert(i11 != NULL);
    if XNN_UNPREDICTABLE(i11 != zero) {
      i11 = (const uint8_t*) ((uintptr_t) i11 + input_offset);
    }
    const uint8_t* i12 = input[12];
    assert(i12 != NULL);
    if XNN_UNPREDICTABLE(i12 != zero) {
      i12 = (const uint8_t*) ((uintptr_t) i12 + input_offset);
    }
    const uint8_t* i13 = input[13];
    assert(i13 != NULL);
    if XNN_UNPREDICTABLE(i13 != zero) {
      i13 = (const uint8_t*) ((uintptr_t) i13 + input_offset);
    }
    const uint8_t* i14 = input[14];
    assert(i14 != NULL);
    if XNN_UNPREDICTABLE(i14 != zero) {
      i14 = (const uint8_t*) ((uintptr_t) i14 + input_offset);
    }
    const uint8_t* i15 = input[15];
    assert(i15 != NULL);
    if XNN_UNPREDICTABLE(i15 != zero) {
      i15 = (const uint8_t*) ((uintptr_t) i15 + input_offset);
    }
    const uint8_t* i16 = input[16];
    assert(i16 != NULL);
    if XNN_UNPREDICTABLE(i16 != zero) {
      i16 = (const uint8_t*) ((uintptr_t) i16 + input_offset);
    }
    const uint8_t* i17 = input[17];
    assert(i17 != NULL);
    if XNN_UNPREDICTABLE(i17 != zero) {
      i17 = (const uint8_t*) ((uintptr_t) i17 + input_offset);
    }
    const uint8_t* i18 = input[18];
    assert(i18 != NULL);
    if XNN_UNPREDICTABLE(i18 != zero) {
      i18 = (const uint8_t*) ((uintptr_t) i18 + input_offset);
    }
    const uint8_t* i19 = input[19];
    assert(i19 != NULL);
    if XNN_UNPREDICTABLE(i19 != zero) {
      i19 = (const uint8_t*) ((uintptr_t) i19 + input_offset);
    }
    const uint8_t* i20 = input[20];
    assert(i20 != NULL);
    if XNN_UNPREDICTABLE(i20 != zero) {
      i20 = (const uint8_t*) ((uintptr_t) i20 + input_offset);
    }
    const uint8_t* i21 = input[21];
    assert(i21 != NULL);
    if XNN_UNPREDICTABLE(i21 != zero) {
      i21 = (const uint8_t*) ((uintptr_t) i21 + input_offset);
    }
    const uint8_t* i22 = input[22];
    assert(i22 != NULL);
    if XNN_UNPREDICTABLE(i22 != zero) {
      i22 = (const uint8_t*) ((uintptr_t) i22 + input_offset);
    }
    const uint8_t* i23 = input[23];
    assert(i23 != NULL);
    if XNN_UNPREDICTABLE(i23 != zero) {
      i23 = (const uint8_t*) ((uintptr_t) i23 + input_offset);
    }
    const uint8_t* i24 = input[24];
    assert(i24 != NULL);
    if XNN_UNPREDICTABLE(i24 != zero) {
      i24 = (const uint8_t*) ((uintptr_t) i24 + input_offset);
    }
    input = (const uint8_t**) ((uintptr_t) input + input_stride);

    size_t c = channels;
    const void* w = weights;
    const __m128i vk_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.kernel_zero_point);
    for (; c >= 8; c -= 8) {
      __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
      __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));


      const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
      const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(uint8_t)));
      i0 += 8;

      const __m128i vzero = _mm_setzero_si128();
      const __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, vzero);
      const __m128i vxk0x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk0x01234567, vzero), vk_zero_point);

      const __m128i vprod0x01234567lo = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
      const __m128i vprod0x01234567hi = _mm_mulhi_epi16(vxi0x01234567, vxk0x01234567);

      vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod0x01234567lo, vprod0x01234567hi));
      vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod0x01234567lo, vprod0x01234567hi));

      const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
      const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(uint8_t)));
      i1 += 8;

      const __m128i vxi1x01234567 = _mm_unpacklo_epi8(vi1x01234567, vzero);
      const __m128i vxk1x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk1x01234567, vzero), vk_zero_point);

      const __m128i vprod1x01234567lo = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
      const __m128i vprod1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567);

      vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod1x01234567lo, vprod1x01234567hi));
      vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod1x01234567lo, vprod1x01234567hi));

      const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
      const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(uint8_t)));
      i2 += 8;

      const __m128i vxi2x01234567 = _mm_unpacklo_epi8(vi2x01234567, vzero);
      const __m128i vxk2x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk2x01234567, vzero), vk_zero_point);

      const __m128i vprod2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
      const __m128i vprod2x01234567hi = _mm_mulhi_epi16(vxi2x01234567, vxk2x01234567);

      vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod2x01234567lo, vprod2x01234567hi));
      vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod2x01234567lo, vprod2x01234567hi));

      const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
      const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(uint8_t)));
      i3 += 8;

      const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, vzero);
      const __m128i vxk3x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk3x01234567, vzero), vk_zero_point);

      const __m128i vprod3x01234567lo = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
      const __m128i vprod3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567);

      vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod3x01234567lo, vprod3x01234567hi));
      vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod3x01234567lo, vprod3x01234567hi));

      const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
      const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(uint8_t)));
      i4 += 8;

      const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, vzero);
      const __m128i vxk4x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk4x01234567, vzero), vk_zero_point);

      const __m128i vprod4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
      const __m128i vprod4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567);

      vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod4x01234567lo, vprod4x01234567hi));
      vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod4x01234567lo, vprod4x01234567hi));

      const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
      const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(uint8_t)));
      i5 += 8;

      const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, vzero);
      const __m128i vxk5x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk5x01234567, vzero), vk_zero_point);

      const __m128i vprod5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
      const __m128i vprod5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567);

      vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod5x01234567lo, vprod5x01234567hi));
      vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod5x01234567lo, vprod5x01234567hi));

      const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
      const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(uint8_t)));
      i6 += 8;

      const __m128i vxi6x01234567 = _mm_unpacklo_epi8(vi6x01234567, vzero);
      const __m128i vxk6x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk6x01234567, vzero), vk_zero_point);

      const __m128i vprod6x01234567lo = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
      const __m128i vprod6x01234567hi = _mm_mulhi_epi16(vxi6x01234567, vxk6x01234567);

      vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod6x01234567lo, vprod6x01234567hi));
      vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod6x01234567lo, vprod6x01234567hi));

      const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
      const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(uint8_t)));
      i7 += 8;

      const __m128i vxi7x01234567 = _mm_unpacklo_epi8(vi7x01234567, vzero);
      const __m128i vxk7x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk7x01234567, vzero), vk_zero_point);

      const __m128i vprod7x01234567lo = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
      const __m128i vprod7x01234567hi = _mm_mulhi_epi16(vxi7x01234567, vxk7x01234567);

      vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod7x01234567lo, vprod7x01234567hi));
      vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod7x01234567lo, vprod7x01234567hi));

      const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
      const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(uint8_t)));
      i8 += 8;

      const __m128i vxi8x01234567 = _mm_unpacklo_epi8(vi8x01234567, vzero);
      const __m128i vxk8x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk8x01234567, vzero), vk_zero_point);

      const __m128i vprod8x01234567lo = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
      const __m128i vprod8x01234567hi = _mm_mulhi_epi16(vxi8x01234567, vxk8x01234567);

      vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod8x01234567lo, vprod8x01234567hi));
      vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod8x01234567lo, vprod8x01234567hi));

      const __m128i vi9x01234567 = _mm_loadl_epi64((const __m128i*) i9);
      const __m128i vk9x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(uint8_t)));
      i9 += 8;

      const __m128i vxi9x01234567 = _mm_unpacklo_epi8(vi9x01234567, vzero);
      const __m128i vxk9x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk9x01234567, vzero), vk_zero_point);

      const __m128i vprod9x01234567lo = _mm_mullo_epi16(vxi9x01234567, vxk9x01234567);
      const __m128i vprod9x01234567hi = _mm_mulhi_epi16(vxi9x01234567, vxk9x01234567);

      vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod9x01234567lo, vprod9x01234567hi));
      vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod9x01234567lo, vprod9x01234567hi));

      const __m128i vi10x01234567 = _mm_loadl_epi64((const __m128i*) i10);
      const __m128i vk10x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 80 * sizeof(uint8_t)));
      i10 += 8;

      const __m128i vxi10x01234567 = _mm_unpacklo_epi8(vi10x01234567, vzero);
      const __m128i vxk10x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk10x01234567, vzero), vk_zero_point);

      const __m128i vprod10x01234567lo = _mm_mullo_epi16(vxi10x01234567, vxk10x01234567);
      const __m128i vprod10x01234567hi = _mm_mulhi_epi16(vxi10x01234567, vxk10x01234567);

      vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod10x01234567lo, vprod10x01234567hi));
      vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod10x01234567lo, vprod10x01234567hi));

      const __m128i vi11x01234567 = _mm_loadl_epi64((const __m128i*) i11);
      const __m128i vk11x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 88 * sizeof(uint8_t)));
      i11 += 8;

      const __m128i vxi11x01234567 = _mm_unpacklo_epi8(vi11x01234567, vzero);
      const __m128i vxk11x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk11x01234567, vzero), vk_zero_point);

      const __m128i vprod11x01234567lo = _mm_mullo_epi16(vxi11x01234567, vxk11x01234567);
      const __m128i vprod11x01234567hi = _mm_mulhi_epi16(vxi11x01234567, vxk11x01234567);

      vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod11x01234567lo, vprod11x01234567hi));
      vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod11x01234567lo, vprod11x01234567hi));

      const __m128i vi12x01234567 = _mm_loadl_epi64((const __m128i*) i12);
      const __m128i vk12x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 96 * sizeof(uint8_t)));
      i12 += 8;

      const __m128i vxi12x01234567 = _mm_unpacklo_epi8(vi12x01234567, vzero);
      const __m128i vxk12x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk12x01234567, vzero), vk_zero_point);

      const __m128i vprod12x01234567lo = _mm_mullo_epi16(vxi12x01234567, vxk12x01234567);
      const __m128i vprod12x01234567hi = _mm_mulhi_epi16(vxi12x01234567, vxk12x01234567);

      vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod12x01234567lo, vprod12x01234567hi));
      vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod12x01234567lo, vprod12x01234567hi));

      const __m128i vi13x01234567 = _mm_loadl_epi64((const __m128i*) i13);
      const __m128i vk13x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 104 * sizeof(uint8_t)));
      i13 += 8;

      const __m128i vxi13x01234567 = _mm_unpacklo_epi8(vi13x01234567, vzero);
      const __m128i vxk13x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk13x01234567, vzero), vk_zero_point);

      const __m128i vprod13x01234567lo = _mm_mullo_epi16(vxi13x01234567, vxk13x01234567);
      const __m128i vprod13x01234567hi = _mm_mulhi_epi16(vxi13x01234567, vxk13x01234567);

      vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod13x01234567lo, vprod13x01234567hi));
      vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod13x01234567lo, vprod13x01234567hi));

      const __m128i vi14x01234567 = _mm_loadl_epi64((const __m128i*) i14);
      const __m128i vk14x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 112 * sizeof(uint8_t)));
      i14 += 8;

      const __m128i vxi14x01234567 = _mm_unpacklo_epi8(vi14x01234567, vzero);
      const __m128i vxk14x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk14x01234567, vzero), vk_zero_point);

      const __m128i vprod14x01234567lo = _mm_mullo_epi16(vxi14x01234567, vxk14x01234567);
      const __m128i vprod14x01234567hi = _mm_mulhi_epi16(vxi14x01234567, vxk14x01234567);

      vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod14x01234567lo, vprod14x01234567hi));
      vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod14x01234567lo, vprod14x01234567hi));

      const __m128i vi15x01234567 = _mm_loadl_epi64((const __m128i*) i15);
      const __m128i vk15x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 120 * sizeof(uint8_t)));
      i15 += 8;

      const __m128i vxi15x01234567 = _mm_unpacklo_epi8(vi15x01234567, vzero);
      const __m128i vxk15x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk15x01234567, vzero), vk_zero_point);

      const __m128i vprod15x01234567lo = _mm_mullo_epi16(vxi15x01234567, vxk15x01234567);
      const __m128i vprod15x01234567hi = _mm_mulhi_epi16(vxi15x01234567, vxk15x01234567);

      vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod15x01234567lo, vprod15x01234567hi));
      vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod15x01234567lo, vprod15x01234567hi));

      const __m128i vi16x01234567 = _mm_loadl_epi64((const __m128i*) i16);
      const __m128i vk16x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 128 * sizeof(uint8_t)));
      i16 += 8;

      const __m128i vxi16x01234567 = _mm_unpacklo_epi8(vi16x01234567, vzero);
      const __m128i vxk16x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk16x01234567, vzero), vk_zero_point);

      const __m128i vprod16x01234567lo = _mm_mullo_epi16(vxi16x01234567, vxk16x01234567);
      const __m128i vprod16x01234567hi = _mm_mulhi_epi16(vxi16x01234567, vxk16x01234567);

      vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod16x01234567lo, vprod16x01234567hi));
      vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod16x01234567lo, vprod16x01234567hi));

      const __m128i vi17x01234567 = _mm_loadl_epi64((const __m128i*) i17);
      const __m128i vk17x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 136 * sizeof(uint8_t)));
      i17 += 8;

      const __m128i vxi17x01234567 = _mm_unpacklo_epi8(vi17x01234567, vzero);
      const __m128i vxk17x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk17x01234567, vzero), vk_zero_point);

      const __m128i vprod17x01234567lo = _mm_mullo_epi16(vxi17x01234567, vxk17x01234567);
      const __m128i vprod17x01234567hi = _mm_mulhi_epi16(vxi17x01234567, vxk17x01234567);

      vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod17x01234567lo, vprod17x01234567hi));
      vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod17x01234567lo, vprod17x01234567hi));

      const __m128i vi18x01234567 = _mm_loadl_epi64((const __m128i*) i18);
      const __m128i vk18x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 144 * sizeof(uint8_t)));
      i18 += 8;

      const __m128i vxi18x01234567 = _mm_unpacklo_epi8(vi18x01234567, vzero);
      const __m128i vxk18x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk18x01234567, vzero), vk_zero_point);

      const __m128i vprod18x01234567lo = _mm_mullo_epi16(vxi18x01234567, vxk18x01234567);
      const __m128i vprod18x01234567hi = _mm_mulhi_epi16(vxi18x01234567, vxk18x01234567);

      vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod18x01234567lo, vprod18x01234567hi));
      vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod18x01234567lo, vprod18x01234567hi));

      const __m128i vi19x01234567 = _mm_loadl_epi64((const __m128i*) i19);
      const __m128i vk19x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 152 * sizeof(uint8_t)));
      i19 += 8;

      const __m128i vxi19x01234567 = _mm_unpacklo_epi8(vi19x01234567, vzero);
      const __m128i vxk19x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk19x01234567, vzero), vk_zero_point);

      const __m128i vprod19x01234567lo = _mm_mullo_epi16(vxi19x01234567, vxk19x01234567);
      const __m128i vprod19x01234567hi = _mm_mulhi_epi16(vxi19x01234567, vxk19x01234567);

      vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod19x01234567lo, vprod19x01234567hi));
      vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod19x01234567lo, vprod19x01234567hi));

      const __m128i vi20x01234567 = _mm_loadl_epi64((const __m128i*) i20);
      const __m128i vk20x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 160 * sizeof(uint8_t)));
      i20 += 8;

      const __m128i vxi20x01234567 = _mm_unpacklo_epi8(vi20x01234567, vzero);
      const __m128i vxk20x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk20x01234567, vzero), vk_zero_point);

      const __m128i vprod20x01234567lo = _mm_mullo_epi16(vxi20x01234567, vxk20x01234567);
      const __m128i vprod20x01234567hi = _mm_mulhi_epi16(vxi20x01234567, vxk20x01234567);

      vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod20x01234567lo, vprod20x01234567hi));
      vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod20x01234567lo, vprod20x01234567hi));

      const __m128i vi21x01234567 = _mm_loadl_epi64((const __m128i*) i21);
      const __m128i vk21x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 168 * sizeof(uint8_t)));
      i21 += 8;

      const __m128i vxi21x01234567 = _mm_unpacklo_epi8(vi21x01234567, vzero);
      const __m128i vxk21x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk21x01234567, vzero), vk_zero_point);

      const __m128i vprod21x01234567lo = _mm_mullo_epi16(vxi21x01234567, vxk21x01234567);
      const __m128i vprod21x01234567hi = _mm_mulhi_epi16(vxi21x01234567, vxk21x01234567);

      vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod21x01234567lo, vprod21x01234567hi));
      vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod21x01234567lo, vprod21x01234567hi));

      const __m128i vi22x01234567 = _mm_loadl_epi64((const __m128i*) i22);
      const __m128i vk22x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 176 * sizeof(uint8_t)));
      i22 += 8;

      const __m128i vxi22x01234567 = _mm_unpacklo_epi8(vi22x01234567, vzero);
      const __m128i vxk22x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk22x01234567, vzero), vk_zero_point);

      const __m128i vprod22x01234567lo = _mm_mullo_epi16(vxi22x01234567, vxk22x01234567);
      const __m128i vprod22x01234567hi = _mm_mulhi_epi16(vxi22x01234567, vxk22x01234567);

      vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod22x01234567lo, vprod22x01234567hi));
      vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod22x01234567lo, vprod22x01234567hi));

      const __m128i vi23x01234567 = _mm_loadl_epi64((const __m128i*) i23);
      const __m128i vk23x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 184 * sizeof(uint8_t)));
      i23 += 8;

      const __m128i vxi23x01234567 = _mm_unpacklo_epi8(vi23x01234567, vzero);
      const __m128i vxk23x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk23x01234567, vzero), vk_zero_point);

      const __m128i vprod23x01234567lo = _mm_mullo_epi16(vxi23x01234567, vxk23x01234567);
      const __m128i vprod23x01234567hi = _mm_mulhi_epi16(vxi23x01234567, vxk23x01234567);

      vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod23x01234567lo, vprod23x01234567hi));
      vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod23x01234567lo, vprod23x01234567hi));

      const __m128i vi24x01234567 = _mm_loadl_epi64((const __m128i*) i24);
      const __m128i vk24x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 192 * sizeof(uint8_t)));
      i24 += 8;

      const __m128i vxi24x01234567 = _mm_unpacklo_epi8(vi24x01234567, vzero);
      const __m128i vxk24x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk24x01234567, vzero), vk_zero_point);

      const __m128i vprod24x01234567lo = _mm_mullo_epi16(vxi24x01234567, vxk24x01234567);
      const __m128i vprod24x01234567hi = _mm_mulhi_epi16(vxi24x01234567, vxk24x01234567);

      vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod24x01234567lo, vprod24x01234567hi));
      vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod24x01234567lo, vprod24x01234567hi));

      w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 200 * sizeof(uint8_t));

      __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
      __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);

      const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
      vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
      vscaled4567 = _mm_mul_ps(vscaled4567, vscale);

      const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
      vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
      vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);

      vacc0123 = _mm_cvtps_epi32(vscaled0123);
      vacc4567 = _mm_cvtps_epi32(vscaled4567);

      const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
      __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);

      __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);

      const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min);
      vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min);

      _mm_storel_epi64((__m128i*) output, vout0123456701234567);
      output += 8;
    }
    if XNN_UNLIKELY(c != 0) {
      {
        __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
        __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));


        const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
        const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(uint8_t)));

        const __m128i vzero = _mm_setzero_si128();
        const __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, vzero);
        const __m128i vxk0x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk0x01234567, vzero), vk_zero_point);

        const __m128i vprod0x01234567lo = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
        const __m128i vprod0x01234567hi = _mm_mulhi_epi16(vxi0x01234567, vxk0x01234567);

        vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod0x01234567lo, vprod0x01234567hi));
        vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod0x01234567lo, vprod0x01234567hi));

        const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
        const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(uint8_t)));

        const __m128i vxi1x01234567 = _mm_unpacklo_epi8(vi1x01234567, vzero);
        const __m128i vxk1x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk1x01234567, vzero), vk_zero_point);

        const __m128i vprod1x01234567lo = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
        const __m128i vprod1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567);

        vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod1x01234567lo, vprod1x01234567hi));
        vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod1x01234567lo, vprod1x01234567hi));

        const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
        const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(uint8_t)));

        const __m128i vxi2x01234567 = _mm_unpacklo_epi8(vi2x01234567, vzero);
        const __m128i vxk2x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk2x01234567, vzero), vk_zero_point);

        const __m128i vprod2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
        const __m128i vprod2x01234567hi = _mm_mulhi_epi16(vxi2x01234567, vxk2x01234567);

        vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod2x01234567lo, vprod2x01234567hi));
        vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod2x01234567lo, vprod2x01234567hi));

        const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
        const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(uint8_t)));

        const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, vzero);
        const __m128i vxk3x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk3x01234567, vzero), vk_zero_point);

        const __m128i vprod3x01234567lo = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
        const __m128i vprod3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567);

        vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod3x01234567lo, vprod3x01234567hi));
        vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod3x01234567lo, vprod3x01234567hi));

        const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
        const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(uint8_t)));

        const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, vzero);
        const __m128i vxk4x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk4x01234567, vzero), vk_zero_point);

        const __m128i vprod4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
        const __m128i vprod4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567);

        vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod4x01234567lo, vprod4x01234567hi));
        vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod4x01234567lo, vprod4x01234567hi));

        const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
        const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(uint8_t)));

        const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, vzero);
        const __m128i vxk5x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk5x01234567, vzero), vk_zero_point);

        const __m128i vprod5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
        const __m128i vprod5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567);

        vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod5x01234567lo, vprod5x01234567hi));
        vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod5x01234567lo, vprod5x01234567hi));

        const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
        const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(uint8_t)));

        const __m128i vxi6x01234567 = _mm_unpacklo_epi8(vi6x01234567, vzero);
        const __m128i vxk6x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk6x01234567, vzero), vk_zero_point);

        const __m128i vprod6x01234567lo = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
        const __m128i vprod6x01234567hi = _mm_mulhi_epi16(vxi6x01234567, vxk6x01234567);

        vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod6x01234567lo, vprod6x01234567hi));
        vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod6x01234567lo, vprod6x01234567hi));

        const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
        const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(uint8_t)));

        const __m128i vxi7x01234567 = _mm_unpacklo_epi8(vi7x01234567, vzero);
        const __m128i vxk7x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk7x01234567, vzero), vk_zero_point);

        const __m128i vprod7x01234567lo = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
        const __m128i vprod7x01234567hi = _mm_mulhi_epi16(vxi7x01234567, vxk7x01234567);

        vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod7x01234567lo, vprod7x01234567hi));
        vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod7x01234567lo, vprod7x01234567hi));

        const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
        const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(uint8_t)));

        const __m128i vxi8x01234567 = _mm_unpacklo_epi8(vi8x01234567, vzero);
        const __m128i vxk8x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk8x01234567, vzero), vk_zero_point);

        const __m128i vprod8x01234567lo = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
        const __m128i vprod8x01234567hi = _mm_mulhi_epi16(vxi8x01234567, vxk8x01234567);

        vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod8x01234567lo, vprod8x01234567hi));
        vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod8x01234567lo, vprod8x01234567hi));

        const __m128i vi9x01234567 = _mm_loadl_epi64((const __m128i*) i9);
        const __m128i vk9x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(uint8_t)));

        const __m128i vxi9x01234567 = _mm_unpacklo_epi8(vi9x01234567, vzero);
        const __m128i vxk9x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk9x01234567, vzero), vk_zero_point);

        const __m128i vprod9x01234567lo = _mm_mullo_epi16(vxi9x01234567, vxk9x01234567);
        const __m128i vprod9x01234567hi = _mm_mulhi_epi16(vxi9x01234567, vxk9x01234567);

        vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod9x01234567lo, vprod9x01234567hi));
        vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod9x01234567lo, vprod9x01234567hi));

        const __m128i vi10x01234567 = _mm_loadl_epi64((const __m128i*) i10);
        const __m128i vk10x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 80 * sizeof(uint8_t)));

        const __m128i vxi10x01234567 = _mm_unpacklo_epi8(vi10x01234567, vzero);
        const __m128i vxk10x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk10x01234567, vzero), vk_zero_point);

        const __m128i vprod10x01234567lo = _mm_mullo_epi16(vxi10x01234567, vxk10x01234567);
        const __m128i vprod10x01234567hi = _mm_mulhi_epi16(vxi10x01234567, vxk10x01234567);

        vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod10x01234567lo, vprod10x01234567hi));
        vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod10x01234567lo, vprod10x01234567hi));

        const __m128i vi11x01234567 = _mm_loadl_epi64((const __m128i*) i11);
        const __m128i vk11x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 88 * sizeof(uint8_t)));

        const __m128i vxi11x01234567 = _mm_unpacklo_epi8(vi11x01234567, vzero);
        const __m128i vxk11x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk11x01234567, vzero), vk_zero_point);

        const __m128i vprod11x01234567lo = _mm_mullo_epi16(vxi11x01234567, vxk11x01234567);
        const __m128i vprod11x01234567hi = _mm_mulhi_epi16(vxi11x01234567, vxk11x01234567);

        vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod11x01234567lo, vprod11x01234567hi));
        vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod11x01234567lo, vprod11x01234567hi));

        const __m128i vi12x01234567 = _mm_loadl_epi64((const __m128i*) i12);
        const __m128i vk12x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 96 * sizeof(uint8_t)));

        const __m128i vxi12x01234567 = _mm_unpacklo_epi8(vi12x01234567, vzero);
        const __m128i vxk12x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk12x01234567, vzero), vk_zero_point);

        const __m128i vprod12x01234567lo = _mm_mullo_epi16(vxi12x01234567, vxk12x01234567);
        const __m128i vprod12x01234567hi = _mm_mulhi_epi16(vxi12x01234567, vxk12x01234567);

        vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod12x01234567lo, vprod12x01234567hi));
        vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod12x01234567lo, vprod12x01234567hi));

        const __m128i vi13x01234567 = _mm_loadl_epi64((const __m128i*) i13);
        const __m128i vk13x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 104 * sizeof(uint8_t)));

        const __m128i vxi13x01234567 = _mm_unpacklo_epi8(vi13x01234567, vzero);
        const __m128i vxk13x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk13x01234567, vzero), vk_zero_point);

        const __m128i vprod13x01234567lo = _mm_mullo_epi16(vxi13x01234567, vxk13x01234567);
        const __m128i vprod13x01234567hi = _mm_mulhi_epi16(vxi13x01234567, vxk13x01234567);

        vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod13x01234567lo, vprod13x01234567hi));
        vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod13x01234567lo, vprod13x01234567hi));

        const __m128i vi14x01234567 = _mm_loadl_epi64((const __m128i*) i14);
        const __m128i vk14x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 112 * sizeof(uint8_t)));

        const __m128i vxi14x01234567 = _mm_unpacklo_epi8(vi14x01234567, vzero);
        const __m128i vxk14x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk14x01234567, vzero), vk_zero_point);

        const __m128i vprod14x01234567lo = _mm_mullo_epi16(vxi14x01234567, vxk14x01234567);
        const __m128i vprod14x01234567hi = _mm_mulhi_epi16(vxi14x01234567, vxk14x01234567);

        vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod14x01234567lo, vprod14x01234567hi));
        vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod14x01234567lo, vprod14x01234567hi));

        const __m128i vi15x01234567 = _mm_loadl_epi64((const __m128i*) i15);
        const __m128i vk15x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 120 * sizeof(uint8_t)));

        const __m128i vxi15x01234567 = _mm_unpacklo_epi8(vi15x01234567, vzero);
        const __m128i vxk15x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk15x01234567, vzero), vk_zero_point);

        const __m128i vprod15x01234567lo = _mm_mullo_epi16(vxi15x01234567, vxk15x01234567);
        const __m128i vprod15x01234567hi = _mm_mulhi_epi16(vxi15x01234567, vxk15x01234567);

        vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod15x01234567lo, vprod15x01234567hi));
        vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod15x01234567lo, vprod15x01234567hi));

        const __m128i vi16x01234567 = _mm_loadl_epi64((const __m128i*) i16);
        const __m128i vk16x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 128 * sizeof(uint8_t)));

        const __m128i vxi16x01234567 = _mm_unpacklo_epi8(vi16x01234567, vzero);
        const __m128i vxk16x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk16x01234567, vzero), vk_zero_point);

        const __m128i vprod16x01234567lo = _mm_mullo_epi16(vxi16x01234567, vxk16x01234567);
        const __m128i vprod16x01234567hi = _mm_mulhi_epi16(vxi16x01234567, vxk16x01234567);

        vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod16x01234567lo, vprod16x01234567hi));
        vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod16x01234567lo, vprod16x01234567hi));

        const __m128i vi17x01234567 = _mm_loadl_epi64((const __m128i*) i17);
        const __m128i vk17x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 136 * sizeof(uint8_t)));

        const __m128i vxi17x01234567 = _mm_unpacklo_epi8(vi17x01234567, vzero);
        const __m128i vxk17x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk17x01234567, vzero), vk_zero_point);

        const __m128i vprod17x01234567lo = _mm_mullo_epi16(vxi17x01234567, vxk17x01234567);
        const __m128i vprod17x01234567hi = _mm_mulhi_epi16(vxi17x01234567, vxk17x01234567);

        vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod17x01234567lo, vprod17x01234567hi));
        vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod17x01234567lo, vprod17x01234567hi));

        const __m128i vi18x01234567 = _mm_loadl_epi64((const __m128i*) i18);
        const __m128i vk18x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 144 * sizeof(uint8_t)));

        const __m128i vxi18x01234567 = _mm_unpacklo_epi8(vi18x01234567, vzero);
        const __m128i vxk18x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk18x01234567, vzero), vk_zero_point);

        const __m128i vprod18x01234567lo = _mm_mullo_epi16(vxi18x01234567, vxk18x01234567);
        const __m128i vprod18x01234567hi = _mm_mulhi_epi16(vxi18x01234567, vxk18x01234567);

        vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod18x01234567lo, vprod18x01234567hi));
        vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod18x01234567lo, vprod18x01234567hi));

        const __m128i vi19x01234567 = _mm_loadl_epi64((const __m128i*) i19);
        const __m128i vk19x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 152 * sizeof(uint8_t)));

        const __m128i vxi19x01234567 = _mm_unpacklo_epi8(vi19x01234567, vzero);
        const __m128i vxk19x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk19x01234567, vzero), vk_zero_point);

        const __m128i vprod19x01234567lo = _mm_mullo_epi16(vxi19x01234567, vxk19x01234567);
        const __m128i vprod19x01234567hi = _mm_mulhi_epi16(vxi19x01234567, vxk19x01234567);

        vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod19x01234567lo, vprod19x01234567hi));
        vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod19x01234567lo, vprod19x01234567hi));

        const __m128i vi20x01234567 = _mm_loadl_epi64((const __m128i*) i20);
        const __m128i vk20x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 160 * sizeof(uint8_t)));

        const __m128i vxi20x01234567 = _mm_unpacklo_epi8(vi20x01234567, vzero);
        const __m128i vxk20x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk20x01234567, vzero), vk_zero_point);

        const __m128i vprod20x01234567lo = _mm_mullo_epi16(vxi20x01234567, vxk20x01234567);
        const __m128i vprod20x01234567hi = _mm_mulhi_epi16(vxi20x01234567, vxk20x01234567);

        vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod20x01234567lo, vprod20x01234567hi));
        vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod20x01234567lo, vprod20x01234567hi));

        const __m128i vi21x01234567 = _mm_loadl_epi64((const __m128i*) i21);
        const __m128i vk21x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 168 * sizeof(uint8_t)));

        const __m128i vxi21x01234567 = _mm_unpacklo_epi8(vi21x01234567, vzero);
        const __m128i vxk21x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk21x01234567, vzero), vk_zero_point);

        const __m128i vprod21x01234567lo = _mm_mullo_epi16(vxi21x01234567, vxk21x01234567);
        const __m128i vprod21x01234567hi = _mm_mulhi_epi16(vxi21x01234567, vxk21x01234567);

        vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod21x01234567lo, vprod21x01234567hi));
        vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod21x01234567lo, vprod21x01234567hi));

        const __m128i vi22x01234567 = _mm_loadl_epi64((const __m128i*) i22);
        const __m128i vk22x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 176 * sizeof(uint8_t)));

        const __m128i vxi22x01234567 = _mm_unpacklo_epi8(vi22x01234567, vzero);
        const __m128i vxk22x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk22x01234567, vzero), vk_zero_point);

        const __m128i vprod22x01234567lo = _mm_mullo_epi16(vxi22x01234567, vxk22x01234567);
        const __m128i vprod22x01234567hi = _mm_mulhi_epi16(vxi22x01234567, vxk22x01234567);

        vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod22x01234567lo, vprod22x01234567hi));
        vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod22x01234567lo, vprod22x01234567hi));

        const __m128i vi23x01234567 = _mm_loadl_epi64((const __m128i*) i23);
        const __m128i vk23x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 184 * sizeof(uint8_t)));

        const __m128i vxi23x01234567 = _mm_unpacklo_epi8(vi23x01234567, vzero);
        const __m128i vxk23x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk23x01234567, vzero), vk_zero_point);

        const __m128i vprod23x01234567lo = _mm_mullo_epi16(vxi23x01234567, vxk23x01234567);
        const __m128i vprod23x01234567hi = _mm_mulhi_epi16(vxi23x01234567, vxk23x01234567);

        vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod23x01234567lo, vprod23x01234567hi));
        vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod23x01234567lo, vprod23x01234567hi));

        const __m128i vi24x01234567 = _mm_loadl_epi64((const __m128i*) i24);
        const __m128i vk24x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 192 * sizeof(uint8_t)));

        const __m128i vxi24x01234567 = _mm_unpacklo_epi8(vi24x01234567, vzero);
        const __m128i vxk24x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk24x01234567, vzero), vk_zero_point);

        const __m128i vprod24x01234567lo = _mm_mullo_epi16(vxi24x01234567, vxk24x01234567);
        const __m128i vprod24x01234567hi = _mm_mulhi_epi16(vxi24x01234567, vxk24x01234567);

        vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod24x01234567lo, vprod24x01234567hi));
        vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod24x01234567lo, vprod24x01234567hi));


        __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
        __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);

        const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
        vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
        vscaled4567 = _mm_mul_ps(vscaled4567, vscale);

        const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
        vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
        vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);

        vacc0123 = _mm_cvtps_epi32(vscaled0123);
        vacc4567 = _mm_cvtps_epi32(vscaled4567);


        const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
        __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);

        __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);

        vout0123456701234567 = _mm_max_epu8(vout0123456701234567, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));

        if (c & 4) {
          unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
          vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
          output += 4;
        }
        if (c & 2) {
          unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
          vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
          output += 2;
        }
        if (c & 1) {
          *output = (uint8_t) _mm_cvtsi128_si32(vout0123456701234567);
          output += 1;
        }
      }
    }

    output = (uint8_t*) ((uintptr_t) output + output_increment);
  } while (--output_width != 0);
}