src/qu8-igemm/c4-neondot.c.in - platform/external/XNNPACK - Gitiles

 // Copyright 2020 Google LLC
 //
 // This source code is licensed under the BSD-style license found in the
 // LICENSE file in the root directory of this source tree.

 $ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
 $assert NR % 8 == 0
 $assert 8 <= NR <= 32
 $assert REQUANTIZATION in ["FP32", "RNDNU"]
 #include <assert.h>

 #include <arm_neon.h>

 #include <xnnpack/igemm.h>
 $if REQUANTIZATION == "FP32":
   #include <xnnpack/intrinsics-polyfill.h>
 #include <xnnpack/math.h>


 $PARAMS_STRUCT = "fp32_neonv8" if REQUANTIZATION == "FP32" else REQUANTIZATION.lower() + "_neon"
 void xnn_qu8_igemm_minmax_${REQUANTIZATION.lower()}_ukernel_${MR}x${NR}c4__neondot(
     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 <= ${MR});
   assert(nc != 0);
   assert(kc != 0);
   assert(ks != 0);
   assert(ks % (${MR} * sizeof(void*)) == 0);
   assert(a_offset % sizeof(uint8_t) == 0);
   assert(a != NULL);
   assert(w != NULL);
   assert(c != NULL);

   kc = round_up_po2(kc, 4 * sizeof(uint8_t));
   uint8_t* c0 = c;
   $for M in range(1, MR):
     uint8_t* c${M} = (uint8_t*) ((uintptr_t) c${M-1} + cm_stride);
     $if M % 2 == 0:
       if XNN_UNPREDICTABLE(mr <= ${M}) {
         c${M} = c${M-1};
       }
     $elif M + 1 == MR:
       if XNN_UNPREDICTABLE(mr != ${M+1}) {
         c${M} = c${M-1};
       }
     $else:
       if XNN_UNPREDICTABLE(mr < ${M+1}) {
         c${M} = c${M-1};
       }

   const uint8x8_t va_zero_point = vld1_dup_u8(&params->${PARAMS_STRUCT}.kernel_zero_point[0]);

   do {
     // Initialize accumulators with bias. ${NR} bias values are loaded from the
     // weight matrix, at the start of the group of ${NR} columns.
     $for N in range(0, NR, 4):
       uint32x4_t vpacc0x${ABC[N:N+4]} = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
     $for M in range(1, MR):
       $for N in range(0, NR, 4):
         uint32x4_t vpacc${M}x${ABC[N:N+4]} = vpacc0x${ABC[N:N+4]};
     $for M in range(0, MR):
       uint32x2_t vnacc${M} = vmov_n_u32(0);

     size_t p = ks;
     do {
       $for M in range(MR):
         const uint8_t* restrict a${M} = a[${M}];
         if XNN_UNPREDICTABLE(a${M} != zero) {
           a${M} = (const uint8_t*) ((uintptr_t) a${M} + a_offset);
         }
       a += ${MR};

       // Inner accumulation loop along the ${NR} columns.
       size_t k = kc;
       // 2x partial unrolled loop to load 8 bytes at a time.
       while (k >= 8 * sizeof(uint8_t)) {
         // Load a ${MR}x8 block of activations.
         $for M in range(MR):
           const uint8x8_t va${M}x01234567 = vld1_u8(a${M}); a${M} += 8;

         // Load a 8x${NR} block of weights.
         $for K in range(0, 8, 4):
           $for N in range(0, NR, 4):
             const uint8x16_t vb${ABC[K:K+4]}x${ABC[N:N+4]} = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);

         // Multiply-accumulate: ${MR}x8 * 8x${NR} --> ${MR}x${NR}.
         $for M in range(MR):
           vnacc${M} = vdot_u32(vnacc${M}, va_zero_point, va${M}x01234567);
           $for K in range(0, 8, 4):
             $for N in range(0, NR, 4):
               vpacc${M}x${ABC[N:N+4]} = vdotq_lane_u32(vpacc${M}x${ABC[N:N+4]}, vb${ABC[K:K+4]}x${ABC[N:N+4]}, va${M}x01234567, ${K//4});

         k -= 8 * sizeof(uint8_t);
       }
       // Handle up to 4 final positions of `k`
       if XNN_UNLIKELY(k != 0) {
         // Load a ${MR}x4 block of activations.
         $for M in range(MR):
           const uint8x8_t va${M}x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a${M}, vmov_n_u32(0), 0)); a${M} += 4;

         // Load a 4x${NR} block of weights.
         $for N in range(0, NR, 4):
           const uint8x16_t vb0123x${ABC[N:N+4]} = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);

         // Multiply-accumulate: ${MR}x4 * 4x${NR} --> ${MR}x${NR}.
         $for M in range(MR):
           vnacc${M} = vdot_u32(vnacc${M}, va_zero_point, va${M}x01234567);
           $for N in range(0, NR, 4):
             vpacc${M}x${ABC[N:N+4]} = vdotq_lane_u32(vpacc${M}x${ABC[N:N+4]}, vb0123x${ABC[N:N+4]}, va${M}x01234567, 0);
       }
       p -= ${MR} * sizeof(void*);
     } while (p != 0);

     // Subtract zero point from accumulators.
     $for M in range(0, MR):
       vnacc${M} = vpadd_u32(vnacc${M}, vnacc${M});
       const uint32x4_t vnacc${M}x0123 = vcombine_u32(vnacc${M}, vnacc${M});
       $for N in range(0, NR, 4):
         int32x4_t vacc${M}x${ABC[N:N+4]} = vreinterpretq_s32_u32(vsubq_u32(vpacc${M}x${ABC[N:N+4]}, vnacc${M}x0123));

     $if REQUANTIZATION == "RNDNU":
       const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->${PARAMS_STRUCT}.right_pre_shift);
       const int32x4_t vmultiplier = vld1q_dup_s32(&params->${PARAMS_STRUCT}.multiplier);
       const int32x4_t vright_post_shift = vld1q_dup_s32(&params->${PARAMS_STRUCT}.right_post_shift);

       $for M in range(MR):
         $for N in range(0, NR, 4):
           vacc${M}x${ABC[N:N+4]} = vshlq_s32(vacc${M}x${ABC[N:N+4]}, vright_pre_shift);

       $for M in range(MR):
         $for N in range(0, NR, 4):
           vacc${M}x${ABC[N:N+4]} = vqdmulhq_s32(vacc${M}x${ABC[N:N+4]}, vmultiplier);

       $for M in range(MR):
         $for N in range(0, NR, 4):
           vacc${M}x${ABC[N:N+4]} = vrshlq_s32(vacc${M}x${ABC[N:N+4]}, vright_post_shift);
     $elif REQUANTIZATION == "FP32":
       $for M in range(MR):
         $for N in range(0, NR, 4):
           float32x4_t vfpacc${M}x${ABC[N:N+4]} = vcvtq_f32_s32(vacc${M}x${ABC[N:N+4]});

       const float32x4_t vscale = vld1q_dup_f32(&params->${PARAMS_STRUCT}.scale);
       $for M in range(MR):
         $for N in range(0, NR, 4):
           vfpacc${M}x${ABC[N:N+4]} = vmulq_f32(vfpacc${M}x${ABC[N:N+4]}, vscale);

       $for M in range(MR):
         $for N in range(0, NR, 4):
           vacc${M}x${ABC[N:N+4]} = vcvtnq_s32_f32(vfpacc${M}x${ABC[N:N+4]});

     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->${PARAMS_STRUCT}.output_zero_point);
 #if XNN_ARCH_ARM64
     $for M in range(MR):
       $for N in range(0, NR, 8):
         const int16x8_t vacc${M}x${ABC[N:N+8]} = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc${M}x${ABC[N:N+4]}), vacc${M}x${ABC[N+4:N+8]}), voutput_zero_point);

     $for M in range(MR):
       $for N in range(0, NR, 16):
         $if N + 8 < NR:
           uint8x16_t vout${M}x${ABC[N:N+16]} = vqmovun_high_s16(vqmovun_s16(vacc${M}x${ABC[N:N+8]}), vacc${M}x${ABC[N+8:N+16]});
         $elif M % 2 == 1:
           uint8x16_t vout${M-1}x${ABC[N:N+8]}_${M}x${ABC[N:N+8]} = vqmovun_high_s16(vqmovun_s16(vacc${M-1}x${ABC[N:N+8]}), vacc${M}x${ABC[N:N+8]});
         $elif M + 1 == MR:
           uint8x8_t vout${M}x${ABC[N:N+8]} = vqmovun_s16(vacc${M}x${ABC[N:N+8]});
 #else
     $for M in range(MR):
       $for N in range(0, NR, 8):
         const int16x8_t vacc${M}x${ABC[N:N+8]} = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc${M}x${ABC[N:N+4]}), vqmovn_s32(vacc${M}x${ABC[N+4:N+8]})), voutput_zero_point);

     $for M in range(MR):
       $for N in range(0, NR, 16):
         $if N + 8 < NR:
           uint8x16_t vout${M}x${ABC[N:N+16]} = vcombine_u8(vqmovun_s16(vacc${M}x${ABC[N:N+8]}), vqmovun_s16(vacc${M}x${ABC[N+8:N+16]}));
         $elif M % 2 == 1:
           uint8x16_t vout${M-1}x${ABC[N:N+8]}_${M}x${ABC[N:N+8]} = vcombine_u8(vqmovun_s16(vacc${M-1}x${ABC[N:N+8]}), vqmovun_s16(vacc${M}x${ABC[N:N+8]}));
         $elif M + 1 == MR:
           uint8x8_t vout${M}x${ABC[N:N+8]} = vqmovun_s16(vacc${M}x${ABC[N:N+8]});
 #endif
     $if NR == 8 and MR == 1:
       const uint8x8_t voutput_min = vld1_dup_u8(&params->${PARAMS_STRUCT}.output_min);
       const uint8x8_t voutput_max = vld1_dup_u8(&params->${PARAMS_STRUCT}.output_max);
     $else:
       const uint8x16_t voutput_min = vld1q_dup_u8(&params->${PARAMS_STRUCT}.output_min);
       const uint8x16_t voutput_max = vld1q_dup_u8(&params->${PARAMS_STRUCT}.output_max);

     $for M in range(MR):
       $for N in range(0, NR, 16):
         $if N + 8 < NR:
           vout${M}x${ABC[N:N+16]} = vmaxq_u8(vout${M}x${ABC[N:N+16]}, voutput_min);
         $elif M % 2 == 1:
           vout${M-1}x${ABC[N:N+8]}_${M}x${ABC[N:N+8]} = vmaxq_u8(vout${M-1}x${ABC[N:N+8]}_${M}x${ABC[N:N+8]}, voutput_min);
         $elif M + 1 == MR:
           $if NR == 8 and MR == 1:
             vout${M}x${ABC[N:N+8]} = vmax_u8(vout${M}x${ABC[N:N+8]}, voutput_min);
           $else:
             vout${M}x${ABC[N:N+8]} = vmax_u8(vout${M}x${ABC[N:N+8]}, vget_low_u8(voutput_min));

     $for M in range(MR):
       $for N in range(0, NR, 16):
         $if N + 8 < NR:
           vout${M}x${ABC[N:N+16]} = vminq_u8(vout${M}x${ABC[N:N+16]}, voutput_max);
         $elif M % 2 == 1:
           vout${M-1}x${ABC[N:N+8]}_${M}x${ABC[N:N+8]} = vminq_u8(vout${M-1}x${ABC[N:N+8]}_${M}x${ABC[N:N+8]}, voutput_max);
         $elif M + 1 == MR:
           $if NR == 8 and MR == 1:
             vout${M}x${ABC[N:N+8]} = vmin_u8(vout${M}x${ABC[N:N+8]}, voutput_max);
           $else:
             vout${M}x${ABC[N:N+8]} = vmin_u8(vout${M}x${ABC[N:N+8]}, vget_low_u8(voutput_max));

     if (nc >= ${NR}) {
       $for M in reversed(range(MR)):
         $for N in range(0, NR, 16):
           $if N + 8 < NR:
             vst1q_u8(c${M} + ${N}, vout${M}x${ABC[N:N+16]});
           $elif M % 2 == 1:
             vst1_u8(c${M} + ${N}, vget_high_u8(vout${M-1}x${ABC[N:N+8]}_${M}x${ABC[N:N+8]}));
             vst1_u8(c${M-1} + ${N}, vget_low_u8(vout${M-1}x${ABC[N:N+8]}_${M}x${ABC[N:N+8]}));
           $elif M + 1 == MR:
             vst1_u8(c${M} + ${N}, vout${M}x${ABC[N:N+8]});

       $for M in reversed(range(MR)):
         c${M} = (uint8_t*) ((uintptr_t) c${M} + cn_stride);

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

       nc -= ${NR};
     } else {
       $if NR == 32:
         if (nc & 16) {
           $for M in reversed(range(MR)):
             vst1q_u8(c${M}, vout${M}x${ABC[0:16]});  c${M} += 16;

           $for M in reversed(range(MR)):
             vout${M}x${ABC[0:16]} = vout${M}x${ABC[16:32]};
         }
       $if NR >= 16:
         $for M in reversed(range(MR)):
           $if M % 2 == 1:
             uint8x16_t vout${M-1}x01234567_${M}x01234567 = vcombine_u8(vget_low_u8(vout${M-1}x0123456789ABCDEF), vget_low_u8(vout${M}x0123456789ABCDEF));
           $elif M + 1 == MR:
             uint8x8_t vout${M}x01234567 = vget_low_u8(vout${M}x0123456789ABCDEF);
         if (nc & 8) {
           $for M in reversed(range(MR)):
             $if M % 2 == 1:
               vst1_u8(c${M}, vget_high_u8(vout${M-1}x01234567_${M}x01234567)); c${M} += 8;
               vst1_u8(c${M-1}, vget_low_u8(vout${M-1}x01234567_${M}x01234567)); c${M-1} += 8;
             $elif M + 1 == MR:
               vst1_u8(c${M}, vout${M}x01234567); c${M} += 8;  // This line
           $for M in reversed(range(MR)):
             $if M % 2 == 1:
               vout${M-1}x01234567_${M}x01234567 = vcombine_u8(vget_high_u8(vout${M-1}x0123456789ABCDEF), vget_high_u8(vout${M}x0123456789ABCDEF));
             $elif M + 1 == MR:
               vout${M}x01234567 = vget_high_u8(vout${M}x0123456789ABCDEF);
         }
       if (nc & 4) {
         $for M in reversed(range(MR)):
           $if M % 2 == 1:
             vst1q_lane_u32((void*) c${M}, vreinterpretq_u32_u8(vout${M-1}x01234567_${M}x01234567), 2); c${M} += 4;
             vst1q_lane_u32((void*) c${M-1}, vreinterpretq_u32_u8(vout${M-1}x01234567_${M}x01234567), 0); c${M-1} += 4;
           $elif M + 1 == MR:
             vst1_lane_u32((void*) c${M}, vreinterpret_u32_u8(vout${M}x01234567), 0); c${M} += 4;
         $for M in reversed(range(MR)):
           $if M % 2 == 1:
             vout${M-1}x01234567_${M}x01234567 = vextq_u8(vout${M-1}x01234567_${M}x01234567, vout${M-1}x01234567_${M}x01234567, 4);
           $elif M + 1 == MR:
             vout${M}x01234567 = vext_u8(vout${M}x01234567, vout${M}x01234567, 4);
       }
       if (nc & 2) {
         $for M in reversed(range(MR)):
           $if M % 2 == 1:
             vst1q_lane_u16((void*) c${M}, vreinterpretq_u16_u8(vout${M-1}x01234567_${M}x01234567), 4); c${M} += 2;
             vst1q_lane_u16((void*) c${M-1}, vreinterpretq_u16_u8(vout${M-1}x01234567_${M}x01234567), 0); c${M-1} += 2;
           $elif M + 1 == MR:
             vst1_lane_u16((void*) c${M}, vreinterpret_u16_u8(vout${M}x01234567), 0); c${M} += 2;
         $for M in reversed(range(MR)):
           $if M % 2 == 1:
             vout${M-1}x01234567_${M}x01234567 = vextq_u8(vout${M-1}x01234567_${M}x01234567, vout${M-1}x01234567_${M}x01234567, 2);
           $elif M + 1 == MR:
             vout${M}x01234567 = vext_u8(vout${M}x01234567, vout${M}x01234567, 2);
       }
       if (nc & 1) {
         $for M in reversed(range(MR)):
           $if M % 2 == 1:
             vst1q_lane_u8(c${M}, vout${M-1}x01234567_${M}x01234567, 8);
             vst1q_lane_u8(c${M-1}, vout${M-1}x01234567_${M}x01234567, 0);
           $elif M + 1 == MR:
             vst1_lane_u8(c${M}, vout${M}x01234567, 0);
       }

       nc = 0;
     }
   } while (nc != 0);
 }
	// Copyright 2020 Google LLC
	//
	// This source code is licensed under the BSD-style license found in the
	// LICENSE file in the root directory of this source tree.

	$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
	$assert NR % 8 == 0
	$assert 8 <= NR <= 32
	$assert REQUANTIZATION in ["FP32", "RNDNU"]
	#include <assert.h>

	#include <arm_neon.h>

	#include <xnnpack/igemm.h>
	$if REQUANTIZATION == "FP32":
	#include <xnnpack/intrinsics-polyfill.h>
	#include <xnnpack/math.h>


	$PARAMS_STRUCT = "fp32_neonv8" if REQUANTIZATION == "FP32" else REQUANTIZATION.lower() + "_neon"
	void xnn_qu8_igemm_minmax_${REQUANTIZATION.lower()}_ukernel_${MR}x${NR}c4__neondot(
	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 <= ${MR});
	assert(nc != 0);
	assert(kc != 0);
	assert(ks != 0);
	assert(ks % (${MR} * sizeof(void*)) == 0);
	assert(a_offset % sizeof(uint8_t) == 0);
	assert(a != NULL);
	assert(w != NULL);
	assert(c != NULL);

	kc = round_up_po2(kc, 4 * sizeof(uint8_t));
	uint8_t* c0 = c;
	$for M in range(1, MR):
	uint8_t* c${M} = (uint8_t*) ((uintptr_t) c${M-1} + cm_stride);
	$if M % 2 == 0:
	if XNN_UNPREDICTABLE(mr <= ${M}) {
	c${M} = c${M-1};
	}
	$elif M + 1 == MR:
	if XNN_UNPREDICTABLE(mr != ${M+1}) {
	c${M} = c${M-1};
	}
	$else:
	if XNN_UNPREDICTABLE(mr < ${M+1}) {
	c${M} = c${M-1};
	}

	const uint8x8_t va_zero_point = vld1_dup_u8(&params->${PARAMS_STRUCT}.kernel_zero_point[0]);

	do {
	// Initialize accumulators with bias. ${NR} bias values are loaded from the
	// weight matrix, at the start of the group of ${NR} columns.
	$for N in range(0, NR, 4):
	uint32x4_t vpacc0x${ABC[N:N+4]} = vld1q_u32(w); w = (const void) ((const uint32_t) w + 4);
	$for M in range(1, MR):
	$for N in range(0, NR, 4):
	uint32x4_t vpacc${M}x${ABC[N:N+4]} = vpacc0x${ABC[N:N+4]};
	$for M in range(0, MR):
	uint32x2_t vnacc${M} = vmov_n_u32(0);

	size_t p = ks;
	do {
	$for M in range(MR):
	const uint8_t* restrict a${M} = a[${M}];
	if XNN_UNPREDICTABLE(a${M} != zero) {
	a${M} = (const uint8_t*) ((uintptr_t) a${M} + a_offset);
	}
	a += ${MR};

	// Inner accumulation loop along the ${NR} columns.
	size_t k = kc;
	// 2x partial unrolled loop to load 8 bytes at a time.
	while (k >= 8 * sizeof(uint8_t)) {
	// Load a ${MR}x8 block of activations.
	$for M in range(MR):
	const uint8x8_t va${M}x01234567 = vld1_u8(a${M}); a${M} += 8;

	// Load a 8x${NR} block of weights.
	$for K in range(0, 8, 4):
	$for N in range(0, NR, 4):
	const uint8x16_t vb${ABC[K:K+4]}x${ABC[N:N+4]} = vld1q_u8(w); w = (const void) ((const uint8_t) w + 16);

	// Multiply-accumulate: ${MR}x8 * 8x${NR} --> ${MR}x${NR}.
	$for M in range(MR):
	vnacc${M} = vdot_u32(vnacc${M}, va_zero_point, va${M}x01234567);
	$for K in range(0, 8, 4):
	$for N in range(0, NR, 4):
	vpacc${M}x${ABC[N:N+4]} = vdotq_lane_u32(vpacc${M}x${ABC[N:N+4]}, vb${ABC[K:K+4]}x${ABC[N:N+4]}, va${M}x01234567, ${K//4});

	k -= 8 * sizeof(uint8_t);
	}
	// Handle up to 4 final positions of `k`
	if XNN_UNLIKELY(k != 0) {
	// Load a ${MR}x4 block of activations.
	$for M in range(MR):
	const uint8x8_t va${M}x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a${M}, vmov_n_u32(0), 0)); a${M} += 4;

	// Load a 4x${NR} block of weights.
	$for N in range(0, NR, 4):
	const uint8x16_t vb0123x${ABC[N:N+4]} = vld1q_u8(w); w = (const void) ((const uint8_t) w + 16);

	// Multiply-accumulate: ${MR}x4 * 4x${NR} --> ${MR}x${NR}.
	$for M in range(MR):
	vnacc${M} = vdot_u32(vnacc${M}, va_zero_point, va${M}x01234567);
	$for N in range(0, NR, 4):
	vpacc${M}x${ABC[N:N+4]} = vdotq_lane_u32(vpacc${M}x${ABC[N:N+4]}, vb0123x${ABC[N:N+4]}, va${M}x01234567, 0);
	}
	p -= ${MR} * sizeof(void*);
	} while (p != 0);

	// Subtract zero point from accumulators.
	$for M in range(0, MR):
	vnacc${M} = vpadd_u32(vnacc${M}, vnacc${M});
	const uint32x4_t vnacc${M}x0123 = vcombine_u32(vnacc${M}, vnacc${M});
	$for N in range(0, NR, 4):
	int32x4_t vacc${M}x${ABC[N:N+4]} = vreinterpretq_s32_u32(vsubq_u32(vpacc${M}x${ABC[N:N+4]}, vnacc${M}x0123));

	$if REQUANTIZATION == "RNDNU":
	const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->${PARAMS_STRUCT}.right_pre_shift);
	const int32x4_t vmultiplier = vld1q_dup_s32(&params->${PARAMS_STRUCT}.multiplier);
	const int32x4_t vright_post_shift = vld1q_dup_s32(&params->${PARAMS_STRUCT}.right_post_shift);

	$for M in range(MR):
	$for N in range(0, NR, 4):
	vacc${M}x${ABC[N:N+4]} = vshlq_s32(vacc${M}x${ABC[N:N+4]}, vright_pre_shift);

	$for M in range(MR):
	$for N in range(0, NR, 4):
	vacc${M}x${ABC[N:N+4]} = vqdmulhq_s32(vacc${M}x${ABC[N:N+4]}, vmultiplier);

	$for M in range(MR):
	$for N in range(0, NR, 4):
	vacc${M}x${ABC[N:N+4]} = vrshlq_s32(vacc${M}x${ABC[N:N+4]}, vright_post_shift);
	$elif REQUANTIZATION == "FP32":
	$for M in range(MR):
	$for N in range(0, NR, 4):
	float32x4_t vfpacc${M}x${ABC[N:N+4]} = vcvtq_f32_s32(vacc${M}x${ABC[N:N+4]});

	const float32x4_t vscale = vld1q_dup_f32(&params->${PARAMS_STRUCT}.scale);
	$for M in range(MR):
	$for N in range(0, NR, 4):
	vfpacc${M}x${ABC[N:N+4]} = vmulq_f32(vfpacc${M}x${ABC[N:N+4]}, vscale);

	$for M in range(MR):
	$for N in range(0, NR, 4):
	vacc${M}x${ABC[N:N+4]} = vcvtnq_s32_f32(vfpacc${M}x${ABC[N:N+4]});

	const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->${PARAMS_STRUCT}.output_zero_point);
	#if XNN_ARCH_ARM64
	$for M in range(MR):
	$for N in range(0, NR, 8):
	const int16x8_t vacc${M}x${ABC[N:N+8]} = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc${M}x${ABC[N:N+4]}), vacc${M}x${ABC[N+4:N+8]}), voutput_zero_point);

	$for M in range(MR):
	$for N in range(0, NR, 16):
	$if N + 8 < NR:
	uint8x16_t vout${M}x${ABC[N:N+16]} = vqmovun_high_s16(vqmovun_s16(vacc${M}x${ABC[N:N+8]}), vacc${M}x${ABC[N+8:N+16]});
	$elif M % 2 == 1:
	uint8x16_t vout${M-1}x${ABC[N:N+8]}_${M}x${ABC[N:N+8]} = vqmovun_high_s16(vqmovun_s16(vacc${M-1}x${ABC[N:N+8]}), vacc${M}x${ABC[N:N+8]});
	$elif M + 1 == MR:
	uint8x8_t vout${M}x${ABC[N:N+8]} = vqmovun_s16(vacc${M}x${ABC[N:N+8]});
	#else
	$for M in range(MR):
	$for N in range(0, NR, 8):
	const int16x8_t vacc${M}x${ABC[N:N+8]} = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc${M}x${ABC[N:N+4]}), vqmovn_s32(vacc${M}x${ABC[N+4:N+8]})), voutput_zero_point);

	$for M in range(MR):
	$for N in range(0, NR, 16):
	$if N + 8 < NR:
	uint8x16_t vout${M}x${ABC[N:N+16]} = vcombine_u8(vqmovun_s16(vacc${M}x${ABC[N:N+8]}), vqmovun_s16(vacc${M}x${ABC[N+8:N+16]}));
	$elif M % 2 == 1:
	uint8x16_t vout${M-1}x${ABC[N:N+8]}_${M}x${ABC[N:N+8]} = vcombine_u8(vqmovun_s16(vacc${M-1}x${ABC[N:N+8]}), vqmovun_s16(vacc${M}x${ABC[N:N+8]}));
	$elif M + 1 == MR:
	uint8x8_t vout${M}x${ABC[N:N+8]} = vqmovun_s16(vacc${M}x${ABC[N:N+8]});
	#endif
	$if NR == 8 and MR == 1:
	const uint8x8_t voutput_min = vld1_dup_u8(&params->${PARAMS_STRUCT}.output_min);
	const uint8x8_t voutput_max = vld1_dup_u8(&params->${PARAMS_STRUCT}.output_max);
	$else:
	const uint8x16_t voutput_min = vld1q_dup_u8(&params->${PARAMS_STRUCT}.output_min);
	const uint8x16_t voutput_max = vld1q_dup_u8(&params->${PARAMS_STRUCT}.output_max);

	$for M in range(MR):
	$for N in range(0, NR, 16):
	$if N + 8 < NR:
	vout${M}x${ABC[N:N+16]} = vmaxq_u8(vout${M}x${ABC[N:N+16]}, voutput_min);
	$elif M % 2 == 1:
	vout${M-1}x${ABC[N:N+8]}_${M}x${ABC[N:N+8]} = vmaxq_u8(vout${M-1}x${ABC[N:N+8]}_${M}x${ABC[N:N+8]}, voutput_min);
	$elif M + 1 == MR:
	$if NR == 8 and MR == 1:
	vout${M}x${ABC[N:N+8]} = vmax_u8(vout${M}x${ABC[N:N+8]}, voutput_min);
	$else:
	vout${M}x${ABC[N:N+8]} = vmax_u8(vout${M}x${ABC[N:N+8]}, vget_low_u8(voutput_min));

	$for M in range(MR):
	$for N in range(0, NR, 16):
	$if N + 8 < NR:
	vout${M}x${ABC[N:N+16]} = vminq_u8(vout${M}x${ABC[N:N+16]}, voutput_max);
	$elif M % 2 == 1:
	vout${M-1}x${ABC[N:N+8]}_${M}x${ABC[N:N+8]} = vminq_u8(vout${M-1}x${ABC[N:N+8]}_${M}x${ABC[N:N+8]}, voutput_max);
	$elif M + 1 == MR:
	$if NR == 8 and MR == 1:
	vout${M}x${ABC[N:N+8]} = vmin_u8(vout${M}x${ABC[N:N+8]}, voutput_max);
	$else:
	vout${M}x${ABC[N:N+8]} = vmin_u8(vout${M}x${ABC[N:N+8]}, vget_low_u8(voutput_max));

	if (nc >= ${NR}) {
	$for M in reversed(range(MR)):
	$for N in range(0, NR, 16):
	$if N + 8 < NR:
	vst1q_u8(c${M} + ${N}, vout${M}x${ABC[N:N+16]});
	$elif M % 2 == 1:
	vst1_u8(c${M} + ${N}, vget_high_u8(vout${M-1}x${ABC[N:N+8]}_${M}x${ABC[N:N+8]}));
	vst1_u8(c${M-1} + ${N}, vget_low_u8(vout${M-1}x${ABC[N:N+8]}_${M}x${ABC[N:N+8]}));
	$elif M + 1 == MR:
	vst1_u8(c${M} + ${N}, vout${M}x${ABC[N:N+8]});

	$for M in reversed(range(MR)):
	c${M} = (uint8_t*) ((uintptr_t) c${M} + cn_stride);

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

	nc -= ${NR};
	} else {
	$if NR == 32:
	if (nc & 16) {
	$for M in reversed(range(MR)):
	vst1q_u8(c${M}, vout${M}x${ABC[0:16]}); c${M} += 16;

	$for M in reversed(range(MR)):
	vout${M}x${ABC[0:16]} = vout${M}x${ABC[16:32]};
	}
	$if NR >= 16:
	$for M in reversed(range(MR)):
	$if M % 2 == 1:
	uint8x16_t vout${M-1}x01234567_${M}x01234567 = vcombine_u8(vget_low_u8(vout${M-1}x0123456789ABCDEF), vget_low_u8(vout${M}x0123456789ABCDEF));
	$elif M + 1 == MR:
	uint8x8_t vout${M}x01234567 = vget_low_u8(vout${M}x0123456789ABCDEF);
	if (nc & 8) {
	$for M in reversed(range(MR)):
	$if M % 2 == 1:
	vst1_u8(c${M}, vget_high_u8(vout${M-1}x01234567_${M}x01234567)); c${M} += 8;
	vst1_u8(c${M-1}, vget_low_u8(vout${M-1}x01234567_${M}x01234567)); c${M-1} += 8;
	$elif M + 1 == MR:
	vst1_u8(c${M}, vout${M}x01234567); c${M} += 8; // This line
	$for M in reversed(range(MR)):
	$if M % 2 == 1:
	vout${M-1}x01234567_${M}x01234567 = vcombine_u8(vget_high_u8(vout${M-1}x0123456789ABCDEF), vget_high_u8(vout${M}x0123456789ABCDEF));
	$elif M + 1 == MR:
	vout${M}x01234567 = vget_high_u8(vout${M}x0123456789ABCDEF);
	}
	if (nc & 4) {
	$for M in reversed(range(MR)):
	$if M % 2 == 1:
	vst1q_lane_u32((void*) c${M}, vreinterpretq_u32_u8(vout${M-1}x01234567_${M}x01234567), 2); c${M} += 4;
	vst1q_lane_u32((void*) c${M-1}, vreinterpretq_u32_u8(vout${M-1}x01234567_${M}x01234567), 0); c${M-1} += 4;
	$elif M + 1 == MR:
	vst1_lane_u32((void*) c${M}, vreinterpret_u32_u8(vout${M}x01234567), 0); c${M} += 4;
	$for M in reversed(range(MR)):
	$if M % 2 == 1:
	vout${M-1}x01234567_${M}x01234567 = vextq_u8(vout${M-1}x01234567_${M}x01234567, vout${M-1}x01234567_${M}x01234567, 4);
	$elif M + 1 == MR:
	vout${M}x01234567 = vext_u8(vout${M}x01234567, vout${M}x01234567, 4);
	}
	if (nc & 2) {
	$for M in reversed(range(MR)):
	$if M % 2 == 1:
	vst1q_lane_u16((void*) c${M}, vreinterpretq_u16_u8(vout${M-1}x01234567_${M}x01234567), 4); c${M} += 2;
	vst1q_lane_u16((void*) c${M-1}, vreinterpretq_u16_u8(vout${M-1}x01234567_${M}x01234567), 0); c${M-1} += 2;
	$elif M + 1 == MR:
	vst1_lane_u16((void*) c${M}, vreinterpret_u16_u8(vout${M}x01234567), 0); c${M} += 2;
	$for M in reversed(range(MR)):
	$if M % 2 == 1:
	vout${M-1}x01234567_${M}x01234567 = vextq_u8(vout${M-1}x01234567_${M}x01234567, vout${M-1}x01234567_${M}x01234567, 2);
	$elif M + 1 == MR:
	vout${M}x01234567 = vext_u8(vout${M}x01234567, vout${M}x01234567, 2);
	}
	if (nc & 1) {
	$for M in reversed(range(MR)):
	$if M % 2 == 1:
	vst1q_lane_u8(c${M}, vout${M-1}x01234567_${M}x01234567, 8);
	vst1q_lane_u8(c${M-1}, vout${M-1}x01234567_${M}x01234567, 0);
	$elif M + 1 == MR:
	vst1_lane_u8(c${M}, vout${M}x01234567, 0);
	}

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