src/qs8-gemm/gen/1x16c4s2-minmax-rndnu-neon-mull-padal.c - platform/external/XNNPACK - Gitiles

 // Auto-generated file. Do not edit!
 //   Template: src/qs8-gemm/c4-neon-mull-padal-shuffle.c.in
 //   Generator: tools/xngen
 //
 // Copyright 2021 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 <arm_neon.h>

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


 void xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4s2__neon_mull_padal(
     size_t mr,
     size_t nc,
     size_t kc,
     const int8_t* restrict a,
     size_t a_stride,
     const void* restrict w,
     int8_t* restrict c,
     size_t cm_stride,
     size_t cn_stride,
     const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
 {
   assert(mr != 0);
   assert(mr <= 1);
   assert(nc != 0);
   assert(kc != 0);
   assert(kc % sizeof(int8_t) == 0);
   assert(a != NULL);
   assert(w != NULL);
   assert(c != NULL);

   const int8_t* a0 = a;
   int8_t* c0 = c;

   do {
     int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc0x89 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc0xAB = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc0xCD = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc0xEF = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));

     size_t k = kc;


     while (k >= 8 * sizeof(int8_t)) {
       int8x8_t va0x0 = vld1_s8(a0); a0 += 8;

       const int8x8_t vb01c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
       const int8x8_t vb23c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
       const int8x8_t vb45c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
       const int8x8_t vb67c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
       const int8x8_t vb89c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
       const int8x8_t vbABc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
       const int8x8_t vbCDc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
       const int8x8_t vbEFc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
       const int8x8_t vb01c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
       const int8x8_t vb23c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
       const int8x8_t vb45c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
       const int8x8_t vb67c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
       const int8x8_t vb89c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
       const int8x8_t vbABc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
       const int8x8_t vbCDc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
       const int8x8_t vbEFc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));

       int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0x0);
       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
       int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0x0);
       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
       int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0x0);
       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
       int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0x0);
       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
       int16x8_t vprod0x89c0 = vmull_s8(vb89c0x0, va0x0);
       vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
       int16x8_t vprod0xABc0 = vmull_s8(vbABc0x0, va0x0);
       vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
       int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0x0, va0x0);
       vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
       int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0x0, va0x0);
       vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
       va0x0 = vext_s8(va0x0, va0x0, 4);
       int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0x0);
       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
       int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0x0);
       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
       int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0x0);
       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
       int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0x0);
       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
       int16x8_t vprod0x89c1 = vmull_s8(vb89c1x0, va0x0);
       vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c1);
       int16x8_t vprod0xABc1 = vmull_s8(vbABc1x0, va0x0);
       vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc1);
       int16x8_t vprod0xCDc1 = vmull_s8(vbCDc1x0, va0x0);
       vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc1);
       int16x8_t vprod0xEFc1 = vmull_s8(vbEFc1x0, va0x0);
       vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc1);

       k -= 8 * sizeof(int8_t);
     }

     if XNN_UNLIKELY(k != 0) {
       const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);

       const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
       const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
       const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
       const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
       const int8x8_t vb89c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
       const int8x8_t vbABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
       const int8x8_t vbCDc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
       const int8x8_t vbEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));

       const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
       const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
       const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
       const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
       const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
       const int16x8_t vprod0x89c0 = vmull_s8(vb89c0, va0c0);
       vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
       const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0);
       vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
       const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0);
       vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
       const int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0, va0c0);
       vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);

       if (k > 4 * sizeof(int8_t)) {
         const int8x8_t vb01c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
         const int8x8_t vb23c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
         const int8x8_t vb45c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
         const int8x8_t vb67c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
         const int8x8_t vb89c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
         const int8x8_t vbABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
         const int8x8_t vbCDc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
         const int8x8_t vbEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));

         const int8x8_t va0c1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 1));
         const int16x8_t vprod0x01c1 = vmull_s8(vb01c1, va0c1);
         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
         const int16x8_t vprod0x23c1 = vmull_s8(vb23c1, va0c1);
         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
         const int16x8_t vprod0x45c1 = vmull_s8(vb45c1, va0c1);
         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
         const int16x8_t vprod0x67c1 = vmull_s8(vb67c1, va0c1);
         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
         const int16x8_t vprod0x89c1 = vmull_s8(vb89c1, va0c1);
         vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c1);
         const int16x8_t vprod0xABc1 = vmull_s8(vbABc1, va0c1);
         vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc1);
         const int16x8_t vprod0xCDc1 = vmull_s8(vbCDc1, va0c1);
         vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc1);
         const int16x8_t vprod0xEFc1 = vmull_s8(vbEFc1, va0c1);
         vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc1);
       }
     }

 #if XNN_ARCH_ARM64
     int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
     int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
     int32x4_t vacc0x89AB = vpaddq_s32(vacc0x89, vacc0xAB);
     int32x4_t vacc0xCDEF = vpaddq_s32(vacc0xCD, vacc0xEF);
 #else
     const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
     const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
     int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
     const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
     const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
     int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
     const int32x2_t vsum0x89 = vpadd_s32(vget_low_s32(vacc0x89), vget_high_s32(vacc0x89));
     const int32x2_t vsum0xAB = vpadd_s32(vget_low_s32(vacc0xAB), vget_high_s32(vacc0xAB));
     int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB);
     const int32x2_t vsum0xCD = vpadd_s32(vget_low_s32(vacc0xCD), vget_high_s32(vacc0xCD));
     const int32x2_t vsum0xEF = vpadd_s32(vget_low_s32(vacc0xEF), vget_high_s32(vacc0xEF));
     int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF);
 #endif

     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);

     vacc0x0123 = vshlq_s32(vacc0x0123, vright_pre_shift);
     vacc0x4567 = vshlq_s32(vacc0x4567, vright_pre_shift);
     vacc0x89AB = vshlq_s32(vacc0x89AB, vright_pre_shift);
     vacc0xCDEF = vshlq_s32(vacc0xCDEF, vright_pre_shift);

     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
     vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
     vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);

     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
     vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);

     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
 #if XNN_ARCH_ARM64
     const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
     const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), voutput_zero_point);

     int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
 #else
     const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
     const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF)), voutput_zero_point);

     int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
 #endif
     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);

     vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);

     vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);

     if (nc >= 16) {
       vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);

       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);

       a0 = (const int8_t*) ((uintptr_t) a0 - kc);

       nc -= 16;
     } else {
       // Final case where not all of the 16 columns fit in the destination.
       int8x8_t vout0x01234567 = vget_low_s8(vout0x0123456789ABCDEF);
       if (nc & 8) {
         vst1_s8(c0, vout0x01234567); c0 += 8;
         vout0x01234567 = vget_high_s8(vout0x0123456789ABCDEF);
       }
       if (nc & 4) {
         vst1_lane_u32(__builtin_assume_aligned(c0, 1), vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
         vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
       }
       if (nc & 2) {
         vst1_lane_u16(__builtin_assume_aligned(c0, 1), vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
         vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
       }
       if (nc & 1) {
         vst1_lane_s8(c0, vout0x01234567, 0);
       }

       nc = 0;
     }
   } while (nc != 0);
 }
	// Auto-generated file. Do not edit!
	// Template: src/qs8-gemm/c4-neon-mull-padal-shuffle.c.in
	// Generator: tools/xngen
	//
	// Copyright 2021 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 <arm_neon.h>

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


	void xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4s2__neon_mull_padal(
	size_t mr,
	size_t nc,
	size_t kc,
	const int8_t* restrict a,
	size_t a_stride,
	const void* restrict w,
	int8_t* restrict c,
	size_t cm_stride,
	size_t cn_stride,
	const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
	{
	assert(mr != 0);
	assert(mr <= 1);
	assert(nc != 0);
	assert(kc != 0);
	assert(kc % sizeof(int8_t) == 0);
	assert(a != NULL);
	assert(w != NULL);
	assert(c != NULL);

	const int8_t* a0 = a;
	int8_t* c0 = c;

	do {
	int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void) ((uintptr_t) w + 2 sizeof(int32_t));
	int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void) ((uintptr_t) w + 2 sizeof(int32_t));
	int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void) ((uintptr_t) w + 2 sizeof(int32_t));
	int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void) ((uintptr_t) w + 2 sizeof(int32_t));
	int32x4_t vacc0x89 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void) ((uintptr_t) w + 2 sizeof(int32_t));
	int32x4_t vacc0xAB = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void) ((uintptr_t) w + 2 sizeof(int32_t));
	int32x4_t vacc0xCD = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void) ((uintptr_t) w + 2 sizeof(int32_t));
	int32x4_t vacc0xEF = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void) ((uintptr_t) w + 2 sizeof(int32_t));

	size_t k = kc;


	while (k >= 8 * sizeof(int8_t)) {
	int8x8_t va0x0 = vld1_s8(a0); a0 += 8;

	const int8x8_t vb01c0x0 = vld1_s8(w); w = (const void) ((uintptr_t) w + 8 sizeof(int8_t));
	const int8x8_t vb23c0x0 = vld1_s8(w); w = (const void) ((uintptr_t) w + 8 sizeof(int8_t));
	const int8x8_t vb45c0x0 = vld1_s8(w); w = (const void) ((uintptr_t) w + 8 sizeof(int8_t));
	const int8x8_t vb67c0x0 = vld1_s8(w); w = (const void) ((uintptr_t) w + 8 sizeof(int8_t));
	const int8x8_t vb89c0x0 = vld1_s8(w); w = (const void) ((uintptr_t) w + 8 sizeof(int8_t));
	const int8x8_t vbABc0x0 = vld1_s8(w); w = (const void) ((uintptr_t) w + 8 sizeof(int8_t));
	const int8x8_t vbCDc0x0 = vld1_s8(w); w = (const void) ((uintptr_t) w + 8 sizeof(int8_t));
	const int8x8_t vbEFc0x0 = vld1_s8(w); w = (const void) ((uintptr_t) w + 8 sizeof(int8_t));
	const int8x8_t vb01c1x0 = vld1_s8(w); w = (const void) ((uintptr_t) w + 8 sizeof(int8_t));
	const int8x8_t vb23c1x0 = vld1_s8(w); w = (const void) ((uintptr_t) w + 8 sizeof(int8_t));
	const int8x8_t vb45c1x0 = vld1_s8(w); w = (const void) ((uintptr_t) w + 8 sizeof(int8_t));
	const int8x8_t vb67c1x0 = vld1_s8(w); w = (const void) ((uintptr_t) w + 8 sizeof(int8_t));
	const int8x8_t vb89c1x0 = vld1_s8(w); w = (const void) ((uintptr_t) w + 8 sizeof(int8_t));
	const int8x8_t vbABc1x0 = vld1_s8(w); w = (const void) ((uintptr_t) w + 8 sizeof(int8_t));
	const int8x8_t vbCDc1x0 = vld1_s8(w); w = (const void) ((uintptr_t) w + 8 sizeof(int8_t));
	const int8x8_t vbEFc1x0 = vld1_s8(w); w = (const void) ((uintptr_t) w + 8 sizeof(int8_t));

	int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0x0);
	vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
	int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0x0);
	vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
	int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0x0);
	vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
	int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0x0);
	vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
	int16x8_t vprod0x89c0 = vmull_s8(vb89c0x0, va0x0);
	vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
	int16x8_t vprod0xABc0 = vmull_s8(vbABc0x0, va0x0);
	vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
	int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0x0, va0x0);
	vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
	int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0x0, va0x0);
	vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
	va0x0 = vext_s8(va0x0, va0x0, 4);
	int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0x0);
	vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
	int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0x0);
	vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
	int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0x0);
	vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
	int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0x0);
	vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
	int16x8_t vprod0x89c1 = vmull_s8(vb89c1x0, va0x0);
	vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c1);
	int16x8_t vprod0xABc1 = vmull_s8(vbABc1x0, va0x0);
	vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc1);
	int16x8_t vprod0xCDc1 = vmull_s8(vbCDc1x0, va0x0);
	vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc1);
	int16x8_t vprod0xEFc1 = vmull_s8(vbEFc1x0, va0x0);
	vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc1);

	k -= 8 * sizeof(int8_t);
	}

	if XNN_UNLIKELY(k != 0) {
	const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);

	const int8x8_t vb01c0 = vld1_s8(w); w = (const void) ((uintptr_t) w + 8 sizeof(int8_t));
	const int8x8_t vb23c0 = vld1_s8(w); w = (const void) ((uintptr_t) w + 8 sizeof(int8_t));
	const int8x8_t vb45c0 = vld1_s8(w); w = (const void) ((uintptr_t) w + 8 sizeof(int8_t));
	const int8x8_t vb67c0 = vld1_s8(w); w = (const void) ((uintptr_t) w + 8 sizeof(int8_t));
	const int8x8_t vb89c0 = vld1_s8(w); w = (const void) ((uintptr_t) w + 8 sizeof(int8_t));
	const int8x8_t vbABc0 = vld1_s8(w); w = (const void) ((uintptr_t) w + 8 sizeof(int8_t));
	const int8x8_t vbCDc0 = vld1_s8(w); w = (const void) ((uintptr_t) w + 8 sizeof(int8_t));
	const int8x8_t vbEFc0 = vld1_s8(w); w = (const void) ((uintptr_t) w + 8 sizeof(int8_t));

	const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
	const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
	vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
	const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
	vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
	const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
	vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
	const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
	vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
	const int16x8_t vprod0x89c0 = vmull_s8(vb89c0, va0c0);
	vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
	const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0);
	vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
	const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0);
	vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
	const int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0, va0c0);
	vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);

	if (k > 4 * sizeof(int8_t)) {
	const int8x8_t vb01c1 = vld1_s8(w); w = (const void) ((uintptr_t) w + 8 sizeof(int8_t));
	const int8x8_t vb23c1 = vld1_s8(w); w = (const void) ((uintptr_t) w + 8 sizeof(int8_t));
	const int8x8_t vb45c1 = vld1_s8(w); w = (const void) ((uintptr_t) w + 8 sizeof(int8_t));
	const int8x8_t vb67c1 = vld1_s8(w); w = (const void) ((uintptr_t) w + 8 sizeof(int8_t));
	const int8x8_t vb89c1 = vld1_s8(w); w = (const void) ((uintptr_t) w + 8 sizeof(int8_t));
	const int8x8_t vbABc1 = vld1_s8(w); w = (const void) ((uintptr_t) w + 8 sizeof(int8_t));
	const int8x8_t vbCDc1 = vld1_s8(w); w = (const void) ((uintptr_t) w + 8 sizeof(int8_t));
	const int8x8_t vbEFc1 = vld1_s8(w); w = (const void) ((uintptr_t) w + 8 sizeof(int8_t));

	const int8x8_t va0c1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 1));
	const int16x8_t vprod0x01c1 = vmull_s8(vb01c1, va0c1);
	vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
	const int16x8_t vprod0x23c1 = vmull_s8(vb23c1, va0c1);
	vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
	const int16x8_t vprod0x45c1 = vmull_s8(vb45c1, va0c1);
	vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
	const int16x8_t vprod0x67c1 = vmull_s8(vb67c1, va0c1);
	vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
	const int16x8_t vprod0x89c1 = vmull_s8(vb89c1, va0c1);
	vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c1);
	const int16x8_t vprod0xABc1 = vmull_s8(vbABc1, va0c1);
	vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc1);
	const int16x8_t vprod0xCDc1 = vmull_s8(vbCDc1, va0c1);
	vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc1);
	const int16x8_t vprod0xEFc1 = vmull_s8(vbEFc1, va0c1);
	vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc1);
	}
	}

	#if XNN_ARCH_ARM64
	int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
	int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
	int32x4_t vacc0x89AB = vpaddq_s32(vacc0x89, vacc0xAB);
	int32x4_t vacc0xCDEF = vpaddq_s32(vacc0xCD, vacc0xEF);
	#else
	const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
	const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
	int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
	const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
	const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
	int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
	const int32x2_t vsum0x89 = vpadd_s32(vget_low_s32(vacc0x89), vget_high_s32(vacc0x89));
	const int32x2_t vsum0xAB = vpadd_s32(vget_low_s32(vacc0xAB), vget_high_s32(vacc0xAB));
	int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB);
	const int32x2_t vsum0xCD = vpadd_s32(vget_low_s32(vacc0xCD), vget_high_s32(vacc0xCD));
	const int32x2_t vsum0xEF = vpadd_s32(vget_low_s32(vacc0xEF), vget_high_s32(vacc0xEF));
	int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF);
	#endif

	const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
	const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
	const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);

	vacc0x0123 = vshlq_s32(vacc0x0123, vright_pre_shift);
	vacc0x4567 = vshlq_s32(vacc0x4567, vright_pre_shift);
	vacc0x89AB = vshlq_s32(vacc0x89AB, vright_pre_shift);
	vacc0xCDEF = vshlq_s32(vacc0xCDEF, vright_pre_shift);

	vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
	vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
	vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
	vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);

	vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
	vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
	vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
	vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);

	const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
	#if XNN_ARCH_ARM64
	const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
	const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), voutput_zero_point);

	int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
	#else
	const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
	const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF)), voutput_zero_point);

	int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
	#endif
	const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
	const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);

	vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);

	vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);

	if (nc >= 16) {
	vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);

	c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);

	a0 = (const int8_t*) ((uintptr_t) a0 - kc);

	nc -= 16;
	} else {
	// Final case where not all of the 16 columns fit in the destination.
	int8x8_t vout0x01234567 = vget_low_s8(vout0x0123456789ABCDEF);
	if (nc & 8) {
	vst1_s8(c0, vout0x01234567); c0 += 8;
	vout0x01234567 = vget_high_s8(vout0x0123456789ABCDEF);
	}
	if (nc & 4) {
	vst1_lane_u32(__builtin_assume_aligned(c0, 1), vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
	vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
	}
	if (nc & 2) {
	vst1_lane_u16(__builtin_assume_aligned(c0, 1), vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
	vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
	}
	if (nc & 1) {
	vst1_lane_s8(c0, vout0x01234567, 0);
	}

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