src/qu8-igemm/gen/4x4c2-minmax-fp32-avx-ld64.c - platform/external/XNNPACK - Gitiles

 // Auto-generated file. Do not edit!
 //   Template: src/qs8-igemm/MRx4c2-sse.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 <smmintrin.h>

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


 void xnn_qu8_igemm_minmax_fp32_ukernel_4x4c2__avx_ld64(
     size_t mr,
     size_t nc,
     size_t kc,
     size_t ks,
     const uint8_t** restrict a,
     const void* restrict w,
     uint8_t* restrict c,
     size_t cm_stride,
     size_t cn_stride,
     size_t a_offset,
     const uint8_t* zero,
     const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
 {
   assert(mr != 0);
   assert(mr <= 4);
   assert(nc != 0);
   assert(kc != 0);
   assert(ks != 0);
   assert(ks % (4 * sizeof(void*)) == 0);
   assert(a_offset % sizeof(uint8_t) == 0);
   assert(a != NULL);
   assert(w != NULL);
   assert(c != NULL);

   kc = round_up_po2(kc, 2);
   uint8_t* c0 = c;
   uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
   if XNN_UNPREDICTABLE(mr < 2) {
     c1 = c0;
   }
   uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
   if XNN_UNPREDICTABLE(mr <= 2) {
     c2 = c1;
   }
   uint8_t* c3 = (uint8_t*) ((uintptr_t) c2 + cm_stride);
   if XNN_UNPREDICTABLE(mr != 4) {
     c3 = c2;
   }

   do {
     __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
     __m128i vacc1x0123 = vacc0x0123;
     __m128i vacc2x0123 = vacc0x0123;
     __m128i vacc3x0123 = vacc0x0123;
     w = (const void*) ((const int32_t*) w + 4);

     size_t p = ks;
     do {
       const uint8_t* restrict a0 = a[0];
       if XNN_UNPREDICTABLE(a0 != zero) {
         a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
       }
       const uint8_t* restrict a1 = a[1];
       if XNN_UNPREDICTABLE(a1 != zero) {
         a1 = (const uint8_t*) ((uintptr_t) a1 + a_offset);
       }
       const uint8_t* restrict a2 = a[2];
       if XNN_UNPREDICTABLE(a2 != zero) {
         a2 = (const uint8_t*) ((uintptr_t) a2 + a_offset);
       }
       const uint8_t* restrict a3 = a[3];
       if XNN_UNPREDICTABLE(a3 != zero) {
         a3 = (const uint8_t*) ((uintptr_t) a3 + a_offset);
       }
       a += 4;

       size_t k = kc;
       const __m128i vb_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.kernel_zero_point);
       while (k >= 8 * sizeof(uint8_t)) {
         const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
         const __m128i vxa0 = _mm_cvtepu8_epi16(va0);
         a0 += 8;
         const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
         const __m128i vxa1 = _mm_cvtepu8_epi16(va1);
         a1 += 8;
         const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
         const __m128i vxa2 = _mm_cvtepu8_epi16(va2);
         a2 += 8;
         const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
         const __m128i vxa3 = _mm_cvtepu8_epi16(va3);
         a3 += 8;

         const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
         const __m128i vxb0 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb0), vb_zero_point);

         vacc0x0123 = _mm_add_epi32(vacc0x0123,
           _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
         vacc1x0123 = _mm_add_epi32(vacc1x0123,
           _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
         vacc2x0123 = _mm_add_epi32(vacc2x0123,
           _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
         vacc3x0123 = _mm_add_epi32(vacc3x0123,
           _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
         const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 8));
         const __m128i vxb1 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb1), vb_zero_point);

         vacc0x0123 = _mm_add_epi32(vacc0x0123,
           _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
         vacc1x0123 = _mm_add_epi32(vacc1x0123,
           _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
         vacc2x0123 = _mm_add_epi32(vacc2x0123,
           _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
         vacc3x0123 = _mm_add_epi32(vacc3x0123,
           _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
         const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 16));
         const __m128i vxb2 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb2), vb_zero_point);

         vacc0x0123 = _mm_add_epi32(vacc0x0123,
           _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
         vacc1x0123 = _mm_add_epi32(vacc1x0123,
           _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
         vacc2x0123 = _mm_add_epi32(vacc2x0123,
           _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
         vacc3x0123 = _mm_add_epi32(vacc3x0123,
           _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
         const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 24));
         const __m128i vxb3 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb3), vb_zero_point);

         vacc0x0123 = _mm_add_epi32(vacc0x0123,
           _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
         vacc1x0123 = _mm_add_epi32(vacc1x0123,
           _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
         vacc2x0123 = _mm_add_epi32(vacc2x0123,
           _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
         vacc3x0123 = _mm_add_epi32(vacc3x0123,
           _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));

         w = (const void*) ((const uint8_t*) w + 32);
         k -= 8 * sizeof(uint8_t);
       }
       if (k != 0) {
         const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
         const __m128i vxa0 = _mm_cvtepu8_epi16(va0);
         a0 = (const uint8_t*) ((uintptr_t) a0 + k);
         const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
         const __m128i vxa1 = _mm_cvtepu8_epi16(va1);
         a1 = (const uint8_t*) ((uintptr_t) a1 + k);
         const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
         const __m128i vxa2 = _mm_cvtepu8_epi16(va2);
         a2 = (const uint8_t*) ((uintptr_t) a2 + k);
         const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
         const __m128i vxa3 = _mm_cvtepu8_epi16(va3);
         a3 = (const uint8_t*) ((uintptr_t) a3 + k);

         const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
         w = (const void*) ((const uint8_t*) w + 8);
         const __m128i vxb0 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb0), vb_zero_point);

         vacc0x0123 = _mm_add_epi32(vacc0x0123,
           _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
         vacc1x0123 = _mm_add_epi32(vacc1x0123,
           _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
         vacc2x0123 = _mm_add_epi32(vacc2x0123,
           _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
         vacc3x0123 = _mm_add_epi32(vacc3x0123,
           _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));

         if (k > 2 * sizeof(uint8_t)) {
           const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
           w = (const void*) ((const uint8_t*) w + 8);
           const __m128i vxb1 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb1), vb_zero_point);

           vacc0x0123 = _mm_add_epi32(vacc0x0123,
             _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
           vacc1x0123 = _mm_add_epi32(vacc1x0123,
             _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
           vacc2x0123 = _mm_add_epi32(vacc2x0123,
             _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
           vacc3x0123 = _mm_add_epi32(vacc3x0123,
             _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));

           if (k > 4 * sizeof(uint8_t)) {
             const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
             w = (const void*) ((const uint8_t*) w + 8);
             const __m128i vxb2 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb2), vb_zero_point);

             vacc0x0123 = _mm_add_epi32(vacc0x0123,
               _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
             vacc1x0123 = _mm_add_epi32(vacc1x0123,
               _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
             vacc2x0123 = _mm_add_epi32(vacc2x0123,
               _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
             vacc3x0123 = _mm_add_epi32(vacc3x0123,
               _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
           }
         }
       }
       p -= 4 * sizeof(void*);
     } while (p != 0);

     __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
     __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
     __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
     __m128 vscaled3x0123 = _mm_cvtepi32_ps(vacc3x0123);

     const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
     vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
     vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale);
     vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale);
     vscaled3x0123 = _mm_mul_ps(vscaled3x0123, vscale);

     const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
     vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
     vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
     vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point);
     vscaled3x0123 = _mm_min_ps(vscaled3x0123, voutput_max_less_zero_point);

     vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
     vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
     vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
     vacc3x0123 = _mm_cvtps_epi32(vscaled3x0123);

     const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
     __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
     __m128i vacc23x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc3x0123), voutput_zero_point);

     __m128i vout = _mm_packus_epi16(vacc01x0123, vacc23x0123);

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

     if (nc >= 4) {
       *((uint32_t*) c3) = (uint32_t) _mm_extract_epi32(vout, 3);
       c3 = (uint8_t*) ((uintptr_t) c3 + cn_stride);
       *((uint32_t*) c2) = (uint32_t) _mm_extract_epi32(vout, 2);
       c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
       *((uint32_t*) c1) = (uint32_t) _mm_extract_epi32(vout, 1);
       c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
       *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
       c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);

       a = (const uint8_t**restrict) ((uintptr_t) a - ks);

       nc -= 4;
     } else {
       if (nc & 2) {
         *((uint16_t*) c3) = (uint16_t) _mm_extract_epi16(vout, 6);
         c3 += 2;
         *((uint16_t*) c2) = (uint16_t) _mm_extract_epi16(vout, 4);
         c2 += 2;
         *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
         c1 += 2;
         *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
         c0 += 2;
         vout = _mm_srli_epi32(vout, 16);
       }
       if (nc & 1) {
         *c3 = (uint8_t) _mm_extract_epi8(vout, 12);
         *c2 = (uint8_t) _mm_extract_epi8(vout, 8);
         *c1 = (uint8_t) _mm_extract_epi8(vout, 4);
         *c0 = (uint8_t) _mm_extract_epi8(vout, 0);
       }

       nc = 0;
     }
   } while (nc != 0);
 }
	// Auto-generated file. Do not edit!
	// Template: src/qs8-igemm/MRx4c2-sse.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 <smmintrin.h>

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


	void xnn_qu8_igemm_minmax_fp32_ukernel_4x4c2__avx_ld64(
	size_t mr,
	size_t nc,
	size_t kc,
	size_t ks,
	const uint8_t** restrict a,
	const void* restrict w,
	uint8_t* restrict c,
	size_t cm_stride,
	size_t cn_stride,
	size_t a_offset,
	const uint8_t* zero,
	const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
	{
	assert(mr != 0);
	assert(mr <= 4);
	assert(nc != 0);
	assert(kc != 0);
	assert(ks != 0);
	assert(ks % (4 * sizeof(void*)) == 0);
	assert(a_offset % sizeof(uint8_t) == 0);
	assert(a != NULL);
	assert(w != NULL);
	assert(c != NULL);

	kc = round_up_po2(kc, 2);
	uint8_t* c0 = c;
	uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
	if XNN_UNPREDICTABLE(mr < 2) {
	c1 = c0;
	}
	uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
	if XNN_UNPREDICTABLE(mr <= 2) {
	c2 = c1;
	}
	uint8_t* c3 = (uint8_t*) ((uintptr_t) c2 + cm_stride);
	if XNN_UNPREDICTABLE(mr != 4) {
	c3 = c2;
	}

	do {
	__m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
	__m128i vacc1x0123 = vacc0x0123;
	__m128i vacc2x0123 = vacc0x0123;
	__m128i vacc3x0123 = vacc0x0123;
	w = (const void) ((const int32_t) w + 4);

	size_t p = ks;
	do {
	const uint8_t* restrict a0 = a[0];
	if XNN_UNPREDICTABLE(a0 != zero) {
	a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
	}
	const uint8_t* restrict a1 = a[1];
	if XNN_UNPREDICTABLE(a1 != zero) {
	a1 = (const uint8_t*) ((uintptr_t) a1 + a_offset);
	}
	const uint8_t* restrict a2 = a[2];
	if XNN_UNPREDICTABLE(a2 != zero) {
	a2 = (const uint8_t*) ((uintptr_t) a2 + a_offset);
	}
	const uint8_t* restrict a3 = a[3];
	if XNN_UNPREDICTABLE(a3 != zero) {
	a3 = (const uint8_t*) ((uintptr_t) a3 + a_offset);
	}
	a += 4;

	size_t k = kc;
	const __m128i vb_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.kernel_zero_point);
	while (k >= 8 * sizeof(uint8_t)) {
	const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
	const __m128i vxa0 = _mm_cvtepu8_epi16(va0);
	a0 += 8;
	const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
	const __m128i vxa1 = _mm_cvtepu8_epi16(va1);
	a1 += 8;
	const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
	const __m128i vxa2 = _mm_cvtepu8_epi16(va2);
	a2 += 8;
	const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
	const __m128i vxa3 = _mm_cvtepu8_epi16(va3);
	a3 += 8;

	const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
	const __m128i vxb0 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb0), vb_zero_point);

	vacc0x0123 = _mm_add_epi32(vacc0x0123,
	_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
	vacc1x0123 = _mm_add_epi32(vacc1x0123,
	_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
	vacc2x0123 = _mm_add_epi32(vacc2x0123,
	_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
	vacc3x0123 = _mm_add_epi32(vacc3x0123,
	_mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
	const __m128i vb1 = _mm_loadl_epi64((const __m128i) ((const uint8_t) w + 8));
	const __m128i vxb1 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb1), vb_zero_point);

	vacc0x0123 = _mm_add_epi32(vacc0x0123,
	_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
	vacc1x0123 = _mm_add_epi32(vacc1x0123,
	_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
	vacc2x0123 = _mm_add_epi32(vacc2x0123,
	_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
	vacc3x0123 = _mm_add_epi32(vacc3x0123,
	_mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
	const __m128i vb2 = _mm_loadl_epi64((const __m128i) ((const uint8_t) w + 16));
	const __m128i vxb2 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb2), vb_zero_point);

	vacc0x0123 = _mm_add_epi32(vacc0x0123,
	_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
	vacc1x0123 = _mm_add_epi32(vacc1x0123,
	_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
	vacc2x0123 = _mm_add_epi32(vacc2x0123,
	_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
	vacc3x0123 = _mm_add_epi32(vacc3x0123,
	_mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
	const __m128i vb3 = _mm_loadl_epi64((const __m128i) ((const uint8_t) w + 24));
	const __m128i vxb3 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb3), vb_zero_point);

	vacc0x0123 = _mm_add_epi32(vacc0x0123,
	_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
	vacc1x0123 = _mm_add_epi32(vacc1x0123,
	_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
	vacc2x0123 = _mm_add_epi32(vacc2x0123,
	_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
	vacc3x0123 = _mm_add_epi32(vacc3x0123,
	_mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));

	w = (const void) ((const uint8_t) w + 32);
	k -= 8 * sizeof(uint8_t);
	}
	if (k != 0) {
	const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
	const __m128i vxa0 = _mm_cvtepu8_epi16(va0);
	a0 = (const uint8_t*) ((uintptr_t) a0 + k);
	const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
	const __m128i vxa1 = _mm_cvtepu8_epi16(va1);
	a1 = (const uint8_t*) ((uintptr_t) a1 + k);
	const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
	const __m128i vxa2 = _mm_cvtepu8_epi16(va2);
	a2 = (const uint8_t*) ((uintptr_t) a2 + k);
	const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
	const __m128i vxa3 = _mm_cvtepu8_epi16(va3);
	a3 = (const uint8_t*) ((uintptr_t) a3 + k);

	const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
	w = (const void) ((const uint8_t) w + 8);
	const __m128i vxb0 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb0), vb_zero_point);

	vacc0x0123 = _mm_add_epi32(vacc0x0123,
	_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
	vacc1x0123 = _mm_add_epi32(vacc1x0123,
	_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
	vacc2x0123 = _mm_add_epi32(vacc2x0123,
	_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
	vacc3x0123 = _mm_add_epi32(vacc3x0123,
	_mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));

	if (k > 2 * sizeof(uint8_t)) {
	const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
	w = (const void) ((const uint8_t) w + 8);
	const __m128i vxb1 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb1), vb_zero_point);

	vacc0x0123 = _mm_add_epi32(vacc0x0123,
	_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
	vacc1x0123 = _mm_add_epi32(vacc1x0123,
	_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
	vacc2x0123 = _mm_add_epi32(vacc2x0123,
	_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
	vacc3x0123 = _mm_add_epi32(vacc3x0123,
	_mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));

	if (k > 4 * sizeof(uint8_t)) {
	const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
	w = (const void) ((const uint8_t) w + 8);
	const __m128i vxb2 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb2), vb_zero_point);

	vacc0x0123 = _mm_add_epi32(vacc0x0123,
	_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
	vacc1x0123 = _mm_add_epi32(vacc1x0123,
	_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
	vacc2x0123 = _mm_add_epi32(vacc2x0123,
	_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
	vacc3x0123 = _mm_add_epi32(vacc3x0123,
	_mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
	}
	}
	}
	p -= 4 * sizeof(void*);
	} while (p != 0);

	__m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
	__m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
	__m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
	__m128 vscaled3x0123 = _mm_cvtepi32_ps(vacc3x0123);

	const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
	vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
	vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale);
	vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale);
	vscaled3x0123 = _mm_mul_ps(vscaled3x0123, vscale);

	const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
	vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
	vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
	vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point);
	vscaled3x0123 = _mm_min_ps(vscaled3x0123, voutput_max_less_zero_point);

	vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
	vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
	vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
	vacc3x0123 = _mm_cvtps_epi32(vscaled3x0123);

	const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
	__m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
	__m128i vacc23x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc3x0123), voutput_zero_point);

	__m128i vout = _mm_packus_epi16(vacc01x0123, vacc23x0123);

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

	if (nc >= 4) {
	((uint32_t) c3) = (uint32_t) _mm_extract_epi32(vout, 3);
	c3 = (uint8_t*) ((uintptr_t) c3 + cn_stride);
	((uint32_t) c2) = (uint32_t) _mm_extract_epi32(vout, 2);
	c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
	((uint32_t) c1) = (uint32_t) _mm_extract_epi32(vout, 1);
	c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
	((uint32_t) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
	c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);

	a = (const uint8_t**restrict) ((uintptr_t) a - ks);

	nc -= 4;
	} else {
	if (nc & 2) {
	((uint16_t) c3) = (uint16_t) _mm_extract_epi16(vout, 6);
	c3 += 2;
	((uint16_t) c2) = (uint16_t) _mm_extract_epi16(vout, 4);
	c2 += 2;
	((uint16_t) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
	c1 += 2;
	((uint16_t) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
	c0 += 2;
	vout = _mm_srli_epi32(vout, 16);
	}
	if (nc & 1) {
	*c3 = (uint8_t) _mm_extract_epi8(vout, 12);
	*c2 = (uint8_t) _mm_extract_epi8(vout, 8);
	*c1 = (uint8_t) _mm_extract_epi8(vout, 4);
	*c0 = (uint8_t) _mm_extract_epi8(vout, 0);
	}

	nc = 0;
	}
	} while (nc != 0);
	}