// 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 <= 16
$assert REQUANTIZATION in ["FP32", "RNDNU", "RNDNU16"] or not REQUANTIZATION
$assert DATATYPE in ["QC8", "QD8", "QS8", "QU8"]
$assert DATATYPE != "QC8" or REQUANTIZATION == "FP32"
$assert not DATATYPE == "QD8" or not REQUANTIZATION

#include <assert.h>

#include <arm_neon.h>

#include "xnnpack/common.h"
#include "xnnpack/igemm.h"
$if REQUANTIZATION == "FP32" and ARMV8:
  #include "xnnpack/intrinsics-polyfill.h"
$if PREFETCH:
  #include "xnnpack/prefetch.h"


$DATATYPE_SPEC = {"QC8": "qs8_qc8w", "QS8": "qs8", "QU8": "qu8", "QD8": "qd8_f32_qc8w"}[DATATYPE]
$PARAMS_STRUCT = REQUANTIZATION.lower() + "_" + (("neonv8" if REQUANTIZATION == "FP32" and ARMV8 else "neon") if REQUANTIZATION != "RNDNU16" else "scalar")
$PARAMS_TYPE = {"QC8": "union xnn_qs8_qc8w_conv_minmax_params", "QD8": "union xnn_f32_minmax_params", "QS8": "union xnn_qs8_conv_minmax_params", "QU8": "union xnn_qu8_conv_minmax_params"}[DATATYPE]
$REQUANTIZATION_SPEC = "_" + REQUANTIZATION.lower() if REQUANTIZATION else ""
$SCALING_PARAMS = "xnn_qd8_quantization_params" if DATATYPE == "QD8" else ""
$XINT8_T = "uint8_t" if DATATYPE == "QU8" else "int8_t"
$OUT_T = XINT8_T if not DATATYPE in ["QD8", "QC4"] else "float"
$XINT8X8_T = "uint8x8_t" if DATATYPE == "QU8" else "int8x8_t"
$XINT8X16_T = "uint8x16_t" if DATATYPE == "QU8" else "int8x16_t"
$VGET_LOW_X8 = "vget_low_u8" if DATATYPE == "QU8" else "vget_low_s8"
$VGET_HIGH_X8 = "vget_high_u8" if DATATYPE == "QU8" else "vget_high_s8"
$VCOMBINE_X8 = "vcombine_u8" if DATATYPE == "QU8" else "vcombine_s8"
$VREINTERPRET_U32_X8 = "vreinterpret_u32_u8" if DATATYPE == "QU8" else "vreinterpret_u32_s8"
$VREINTERPRETQ_U32_X8 = "vreinterpretq_u32_u8" if DATATYPE == "QU8" else "vreinterpretq_u32_s8"
$VREINTERPRET_U16_X8 = "vreinterpret_u16_u8" if DATATYPE == "QU8" else "vreinterpret_u16_s8"
$VREINTERPRETQ_U16_X8 = "vreinterpretq_u16_u8" if DATATYPE == "QU8" else "vreinterpretq_u16_s8"
$VLD1_X8 = "vld1_u8" if DATATYPE == "QU8" else "vld1_s8"
$VLD1_DUP_X8 = "vld1_dup_u8" if DATATYPE == "QU8" else "vld1_dup_s8"
$VLD1Q_DUP_X8 = "vld1q_dup_u8" if DATATYPE == "QU8" else "vld1q_dup_s8"
$VST1_X8 = "vst1_u8" if DATATYPE == "QU8" else "vst1_s8"
$VST1Q_X8 = "vst1q_u8" if DATATYPE == "QU8" else "vst1q_s8"
$VST1_LANE_X8 = "vst1_lane_u8" if DATATYPE == "QU8" else "vst1_lane_s8"
$VST1Q_LANE_X8 = "vst1q_lane_u8" if DATATYPE == "QU8" else "vst1q_lane_s8"
$VMIN_X8 = "vmin_u8" if DATATYPE == "QU8" else "vmin_s8"
$VMAX_X8 = "vmax_u8" if DATATYPE == "QU8" else "vmax_s8"
$VMINQ_X8 = "vminq_u8" if DATATYPE == "QU8" else "vminq_s8"
$VMAXQ_X8 = "vmaxq_u8" if DATATYPE == "QU8" else "vmaxq_s8"
$VEXT_X8 = "vext_u8" if DATATYPE == "QU8" else "vext_s8"
$VEXTQ_X8 = "vextq_u8" if DATATYPE == "QU8" else "vextq_s8"
$VQMOVXN_S16 = "vqmovun_s16" if DATATYPE == "QU8" else "vqmovn_s16"
$VQMOVXN_HIGH_S16 = "vqmovun_high_s16" if DATATYPE == "QU8" else "vqmovn_high_s16"
$ISA = "neonv8" if ARMV8 else "neon"
void xnn_${DATATYPE_SPEC}_igemm_minmax${REQUANTIZATION_SPEC}_ukernel_${MR}x${NR}__${ISA}_mlal_lane${"_prfm" if PREFETCH else ""}(
    size_t mr,
    size_t nc,
    size_t kc,
    size_t ks,
    const ${XINT8_T}** restrict a,
    const void* restrict w,
    ${OUT_T}* restrict c,
    size_t cm_stride,
    size_t cn_stride,
    size_t a_offset,
    const ${XINT8_T}* zero,
    $if DATATYPE in ["QD8"]:
      const int8_t* zero_data,
    $if DATATYPE in ["QD8", "QC4"]:
      const ${PARAMS_TYPE} params[restrict XNN_MIN_ELEMENTS(1)],
      const struct ${SCALING_PARAMS} quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
    $else:
      const ${PARAMS_TYPE} 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(${XINT8_T}) == 0);
  assert(a != NULL);
  assert(w != NULL);
  assert(c != NULL);

  ${OUT_T}* c0 = c;
  $for M in range(1, MR):
    ${OUT_T}* c${M} = (${OUT_T}*) ((uintptr_t) c${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};
      }

  $if DATATYPE == "QU8":
    const uint8x8_t vb_zero_point = vld1_dup_u8(&params->${PARAMS_STRUCT}.kernel_zero_point);
  do {
    $if DATATYPE == "QD8":
      $for N in range(0, NR, 4):
        int32x4_t vksum${ABC[N:N+4]} = vld1q_s32(w); w = (const int32_t*) w + 4;
      const int32x4_t vinput_zero_point = vld1q_dup_s32(&quantization_params->zero_point);
      $for N in range(0, NR, 4):
        int32x4_t vacc0x${ABC[N:N+4]} = vmulq_s32(vksum${ABC[N:N+4]}, vinput_zero_point);
      $for M in range(1, MR):
        $for N in range(0, NR, 4):
          int32x4_t vacc${M}x${ABC[N:N+4]} = vacc0x${ABC[N:N+4]};
    $else:
      $for N in range(0, NR, 4):
        int32x4_t vacc0x${ABC[N:N+4]} = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
      $for M in range(1, MR):
        $for N in range(0, NR, 4):
          int32x4_t vacc${M}x${ABC[N:N+4]} = vacc0x${ABC[N:N+4]};

    size_t p = ks;
    do {
      $for M in range(MR):
        const ${XINT8_T}* restrict a${M} = a[${M}];
        if XNN_UNPREDICTABLE(a${M} != zero) {
          a${M} = (const ${XINT8_T}*) ((uintptr_t) a${M} + a_offset);
        $if DATATYPE == "QD8":
          } else {
            a${M} = zero_data;
        }
      a += ${MR};

      size_t k = kc;
      while (k >= 8 * sizeof(${XINT8_T})) {
        $for M in range(MR):
          const ${XINT8X8_T} va${M} = ${VLD1_X8}(a${M}); a${M} += 8;
          $if DATATYPE == "QU8":
            const int16x8_t vxa${M} = vreinterpretq_s16_u16(vmovl_u8(va${M}));
          $else:
            const int16x8_t vxa${M} = vmovl_s8(va${M});

        $for K in range(4):
          $for N in range(0, NR, 8):
            const ${XINT8X8_T} vb${ABC[N:N+8]}c${K} = ${VLD1_X8}(w); w = (const ${XINT8_T}*) w + 8;
            $if DATATYPE == "QU8":
              const int16x8_t vxb${ABC[N:N+8]}c${K} = vreinterpretq_s16_u16(vsubl_u8(vb${ABC[N:N+8]}c${K}, vb_zero_point));
            $else:
              const int16x8_t vxb${ABC[N:N+8]}c${K} = vmovl_s8(vb${ABC[N:N+8]}c${K});

            $for M in range(MR):
              vacc${M}x${ABC[N:N+4]} = vmlal_lane_s16(vacc${M}x${ABC[N:N+4]}, vget_low_s16(vxb${ABC[N:N+8]}c${K}), vget_low_s16(vxa${M}), ${K});
              vacc${M}x${ABC[N+4:N+8]} = vmlal_lane_s16(vacc${M}x${ABC[N+4:N+8]}, vget_high_s16(vxb${ABC[N:N+8]}c${K}), vget_low_s16(vxa${M}), ${K});

        $if PREFETCH:
          $for N in range(0, NR, 8):
            xnn_prefetch_to_l1((const ${XINT8_T}*) w + ${N * 8 + 448});

        $for K in range(4, 8):
          $for N in range(0, NR, 8):
            const ${XINT8X8_T} vb${ABC[N:N+8]}c${K} = ${VLD1_X8}(w); w = (const ${XINT8_T}*) w + 8;
            $if DATATYPE == "QU8":
              const int16x8_t vxb${ABC[N:N+8]}c${K} = vreinterpretq_s16_u16(vsubl_u8(vb${ABC[N:N+8]}c${K}, vb_zero_point));
            $else:
              const int16x8_t vxb${ABC[N:N+8]}c${K} = vmovl_s8(vb${ABC[N:N+8]}c${K});

            $for M in range(MR):
              vacc${M}x${ABC[N:N+4]} = vmlal_lane_s16(vacc${M}x${ABC[N:N+4]}, vget_low_s16(vxb${ABC[N:N+8]}c${K}), vget_high_s16(vxa${M}), ${K-4});
              vacc${M}x${ABC[N+4:N+8]} = vmlal_lane_s16(vacc${M}x${ABC[N+4:N+8]}, vget_high_s16(vxb${ABC[N:N+8]}c${K}), vget_high_s16(vxa${M}), ${K-4});

        k -= 8 * sizeof(${XINT8_T});
      }
      if XNN_UNLIKELY(k != 0) {
        $for M in range(MR):
          const ${XINT8X8_T} va${M} = ${VLD1_X8}(a${M}); a${M} = (const ${XINT8_T}*) ((uintptr_t) a${M} + k);
          $if DATATYPE == "QU8":
            const int16x8_t vxa${M} = vreinterpretq_s16_u16(vmovl_u8(va${M}));
          $else:
            const int16x8_t vxa${M} = vmovl_s8(va${M});

        $for N in range(0, NR, 8):
          const ${XINT8X8_T} vb${ABC[N:N+8]}c0 = ${VLD1_X8}(w); w = (const ${XINT8_T}*) w + 8;
          $if DATATYPE == "QU8":
            const int16x8_t vxb${ABC[N:N+8]}c0 = vreinterpretq_s16_u16(vsubl_u8(vb${ABC[N:N+8]}c0, vb_zero_point));
          $else:
            const int16x8_t vxb${ABC[N:N+8]}c0 = vmovl_s8(vb${ABC[N:N+8]}c0);

        $for M in range(MR):
          $for N in range(0, NR, 8):
            vacc${M}x${ABC[N:N+4]} = vmlal_lane_s16(vacc${M}x${ABC[N:N+4]}, vget_low_s16(vxb${ABC[N:N+8]}c0), vget_low_s16(vxa${M}), 0);
            vacc${M}x${ABC[N+4:N+8]} = vmlal_lane_s16(vacc${M}x${ABC[N+4:N+8]}, vget_high_s16(vxb${ABC[N:N+8]}c0), vget_low_s16(vxa${M}), 0);

        if (k >= 2 * sizeof(${XINT8_T})) {
          $for N in range(0, NR, 8):
            const ${XINT8X8_T} vb${ABC[N:N+8]}c1 = ${VLD1_X8}(w); w = (const ${XINT8_T}*) w + 8;
            $if DATATYPE == "QU8":
              const int16x8_t vxb${ABC[N:N+8]}c1 = vreinterpretq_s16_u16(vsubl_u8(vb${ABC[N:N+8]}c1, vb_zero_point));
            $else:
              const int16x8_t vxb${ABC[N:N+8]}c1 = vmovl_s8(vb${ABC[N:N+8]}c1);

          $for M in range(MR):
            $for N in range(0, NR, 8):
              vacc${M}x${ABC[N:N+4]} = vmlal_lane_s16(vacc${M}x${ABC[N:N+4]}, vget_low_s16(vxb${ABC[N:N+8]}c1), vget_low_s16(vxa${M}), 1);
              vacc${M}x${ABC[N+4:N+8]} = vmlal_lane_s16(vacc${M}x${ABC[N+4:N+8]}, vget_high_s16(vxb${ABC[N:N+8]}c1), vget_low_s16(vxa${M}), 1);

          if (k > 2 * sizeof(${XINT8_T})) {
            $for N in range(0, NR, 8):
              const ${XINT8X8_T} vb${ABC[N:N+8]}c2 = ${VLD1_X8}(w); w = (const ${XINT8_T}*) w + 8;
              $if DATATYPE == "QU8":
                const int16x8_t vxb${ABC[N:N+8]}c2 = vreinterpretq_s16_u16(vsubl_u8(vb${ABC[N:N+8]}c2, vb_zero_point));
              $else:
                const int16x8_t vxb${ABC[N:N+8]}c2 = vmovl_s8(vb${ABC[N:N+8]}c2);

            $for M in range(MR):
              $for N in range(0, NR, 8):
                vacc${M}x${ABC[N:N+4]} = vmlal_lane_s16(vacc${M}x${ABC[N:N+4]}, vget_low_s16(vxb${ABC[N:N+8]}c2), vget_low_s16(vxa${M}), 2);
                vacc${M}x${ABC[N+4:N+8]} = vmlal_lane_s16(vacc${M}x${ABC[N+4:N+8]}, vget_high_s16(vxb${ABC[N:N+8]}c2), vget_low_s16(vxa${M}), 2);

            if (k >= 4 * sizeof(${XINT8_T})) {
              $for N in range(0, NR, 8):
                const ${XINT8X8_T} vb${ABC[N:N+8]}c3 = ${VLD1_X8}(w); w = (const ${XINT8_T}*) w + 8;
                $if DATATYPE == "QU8":
                  const int16x8_t vxb${ABC[N:N+8]}c3 = vreinterpretq_s16_u16(vsubl_u8(vb${ABC[N:N+8]}c3, vb_zero_point));
                $else:
                  const int16x8_t vxb${ABC[N:N+8]}c3 = vmovl_s8(vb${ABC[N:N+8]}c3);

              $for M in range(MR):
                $for N in range(0, NR, 8):
                  vacc${M}x${ABC[N:N+4]} = vmlal_lane_s16(vacc${M}x${ABC[N:N+4]}, vget_low_s16(vxb${ABC[N:N+8]}c3), vget_low_s16(vxa${M}), 3);
                  vacc${M}x${ABC[N+4:N+8]} = vmlal_lane_s16(vacc${M}x${ABC[N+4:N+8]}, vget_high_s16(vxb${ABC[N:N+8]}c3), vget_low_s16(vxa${M}), 3);

              if (k > 4 * sizeof(${XINT8_T})) {
                $for N in range(0, NR, 8):
                  const ${XINT8X8_T} vb${ABC[N:N+8]}c4 = ${VLD1_X8}(w); w = (const ${XINT8_T}*) w + 8;
                  $if DATATYPE == "QU8":
                    const int16x8_t vxb${ABC[N:N+8]}c4 = vreinterpretq_s16_u16(vsubl_u8(vb${ABC[N:N+8]}c4, vb_zero_point));
                  $else:
                    const int16x8_t vxb${ABC[N:N+8]}c4 = vmovl_s8(vb${ABC[N:N+8]}c4);

                $for M in range(MR):
                  $for N in range(0, NR, 8):
                    vacc${M}x${ABC[N:N+4]} = vmlal_lane_s16(vacc${M}x${ABC[N:N+4]}, vget_low_s16(vxb${ABC[N:N+8]}c4), vget_high_s16(vxa${M}), 0);
                    vacc${M}x${ABC[N+4:N+8]} = vmlal_lane_s16(vacc${M}x${ABC[N+4:N+8]}, vget_high_s16(vxb${ABC[N:N+8]}c4), vget_high_s16(vxa${M}), 0);

                if (k >= 6 * sizeof(${XINT8_T})) {
                  $for N in range(0, NR, 8):
                    const ${XINT8X8_T} vb${ABC[N:N+8]}c5 = ${VLD1_X8}(w); w = (const ${XINT8_T}*) w + 8;
                    $if DATATYPE == "QU8":
                      const int16x8_t vxb${ABC[N:N+8]}c5 = vreinterpretq_s16_u16(vsubl_u8(vb${ABC[N:N+8]}c5, vb_zero_point));
                    $else:
                      const int16x8_t vxb${ABC[N:N+8]}c5 = vmovl_s8(vb${ABC[N:N+8]}c5);

                  $for M in range(MR):
                    $for N in range(0, NR, 8):
                      vacc${M}x${ABC[N:N+4]} = vmlal_lane_s16(vacc${M}x${ABC[N:N+4]}, vget_low_s16(vxb${ABC[N:N+8]}c5), vget_high_s16(vxa${M}), 1);
                      vacc${M}x${ABC[N+4:N+8]} = vmlal_lane_s16(vacc${M}x${ABC[N+4:N+8]}, vget_high_s16(vxb${ABC[N:N+8]}c5), vget_high_s16(vxa${M}), 1);

                  if (k > 6 * sizeof(${XINT8_T})) {
                    $for N in range(0, NR, 8):
                      const ${XINT8X8_T} vb${ABC[N:N+8]}c6 = ${VLD1_X8}(w); w = (const ${XINT8_T}*) w + 8;
                      $if DATATYPE == "QU8":
                        const int16x8_t vxb${ABC[N:N+8]}c6 = vreinterpretq_s16_u16(vsubl_u8(vb${ABC[N:N+8]}c6, vb_zero_point));
                      $else:
                        const int16x8_t vxb${ABC[N:N+8]}c6 = vmovl_s8(vb${ABC[N:N+8]}c6);

                    $for M in range(MR):
                      $for N in range(0, NR, 8):
                        vacc${M}x${ABC[N:N+4]} = vmlal_lane_s16(vacc${M}x${ABC[N:N+4]}, vget_low_s16(vxb${ABC[N:N+8]}c6), vget_high_s16(vxa${M}), 2);
                        vacc${M}x${ABC[N+4:N+8]} = vmlal_lane_s16(vacc${M}x${ABC[N+4:N+8]}, vget_high_s16(vxb${ABC[N:N+8]}c6), vget_high_s16(vxa${M}), 2);
                  }
                }
              }
            }
          }
        }
      }
      p -= ${MR} * sizeof(void*);
    } while (p != 0);

    $if DATATYPE == "QD8":
      $for M in range(0, MR):
        $for N in range(0, NR, 4):
          float32x4_t vout${M}x${ABC[N:N+4]} = vcvtq_f32_s32(vacc${M}x${ABC[N:N+4]});

      const float32x4_t vinput_scale = vld1q_dup_f32(&quantization_params->inv_scale);
      $for M in range(MR):
        $for N in range(0, NR, 4):
          vout${M}x${ABC[N:N+4]} = vmulq_f32(vout${M}x${ABC[N:N+4]}, vinput_scale);

      $for N in range(0, NR, 4):
        const float32x4_t vfilter_output_scale${ABC[N:N+4]} = vld1q_f32(w); w = (const float*) w + 4;

      $for N in range(0, NR, 4):
        const float32x4_t vbias${ABC[N:N+4]} = vld1q_f32(w); w = (const float*) w + 4;
        #if XNN_ARCH_ARM64
          $for M in range(MR):
            vout${M}x${ABC[N:N+4]} = vfmaq_f32(vbias${ABC[N:N+4]}, vout${M}x${ABC[N:N+4]}, vfilter_output_scale${ABC[N:N+4]});
        #else
          $for M in range(MR):
            vout${M}x${ABC[N:N+4]} = vmlaq_f32(vbias${ABC[N:N+4]}, vout${M}x${ABC[N:N+4]}, vfilter_output_scale${ABC[N:N+4]});
        #endif

      const float32x4_t voutput_min = vld1q_dup_f32(&params->scalar.min);
      $for M in range(0, MR):
        $for N in range(0, NR, 4):
          vout${M}x${ABC[N:N+4]} = vmaxq_f32(vout${M}x${ABC[N:N+4]}, voutput_min);

      const float32x4_t voutput_max = vld1q_dup_f32(&params->scalar.max);
      $for M in range(0, MR):
        $for N in range(0, NR, 4):
          vout${M}x${ABC[N:N+4]} = vminq_f32(vout${M}x${ABC[N:N+4]}, voutput_max);

      if XNN_LIKELY(nc >= ${NR}) {
        $for M in reversed(range(MR)):
          $for N in range(0, NR, 4):
            vst1q_f32(&c${M}[${N}], vout${M}x${ABC[N:N+4]});

        a = (const ${XINT8_T}**restrict) ((uintptr_t) a - ks);

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

        nc -= ${NR};
      } else {
        $for LOG2N in reversed(range(NR.bit_length())):
          $if NR != 1 << LOG2N:
            $if LOG2N == 1:
              $for M in reversed(range(MR)):
                float32x2_t vout${M}x${ABC[N:N+2]} = vget_low_f32(vout${M}x${ABC[N:N+4]});
            if (nc & ${1 << LOG2N}) {
              $if LOG2N > 1:
                $for N in range(0, 1 << LOG2N, 4):
                  $for M in reversed(range(MR)):
                    vst1q_f32(c${M}, vout${M}x${ABC[N:N+4]}); c${M} += 4;
                    vout${M}x${ABC[N:N+4]} = vout${M}x${ABC[N+(1 << LOG2N):N+(1 << LOG2N)+4]};
              $elif LOG2N == 1:
                $for M in reversed(range(MR)):
                  vst1_f32(c${M}, vout${M}x${ABC[N:N+2]}); c${M} += 2;
                $for M in reversed(range(MR)):
                  vout${M}x${ABC[N:N+2]} = vget_high_f32(vout${M}x${ABC[N:N+4]});
              $elif LOG2N == 0:
                $for M in reversed(range(MR)):
                  vst1_lane_f32(c${M}, vout${M}x${ABC[N:N+2]}, 0);
            }
        nc = 0;
      }
    $# QC8 or QS8
    $else:
      // Post-accumulation work
      $if REQUANTIZATION == "RNDNU":
        const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->${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${ABC[N:N+4]} = vqshlq_s32(vacc${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);
      $elif REQUANTIZATION == "RNDNU16":
        const int32x4_t vleft_pre_shift = vld1q_dup_s32(&params->${PARAMS_STRUCT}.left_pre_shift);
        const int16x8_t vmultiplier = vld1q_dup_s16(&params->${PARAMS_STRUCT}.multiplier);
        $for M in range(MR):
          $for N in range(0, NR, 4):
            vacc${M}x${ABC[N:N+4]} = vqrshlq_s32(vacc${M}x${ABC[N:N+4]}, vleft_pre_shift);

        $for M in range(MR):
          $for N in range(0, NR, 8):
            int16x8_t vacc_preshifted${M}x${ABC[N:N+8]} = vcombine_s16(vqmovn_s32(vacc${M}x${ABC[N:N+4]}), vqmovn_s32(vacc${M}x${ABC[N+4:N+8]}));
            int16x8_t vacc${M}x${ABC[N:N+8]} = vqrdmulhq_s16(vacc_preshifted${M}x${ABC[N:N+8]}, vmultiplier);
      $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]});

        $if DATATYPE == "QC8":
          $for N in range(0, NR, 4):
            const float32x4_t vscale${ABC[N:N+4]} = vld1q_f32((const float*) w); w = (const float*) w + 4;
            $for M in range(MR):
              vfpacc${M}x${ABC[N:N+4]} = vmulq_f32(vfpacc${M}x${ABC[N:N+4]}, vscale${ABC[N:N+4]});
        $else:
          const float32x4_t vscale = vld1q_dup_f32(&params->${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);

        $if ARMV8:
          $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]});
        $else:
          const float32x4_t vmagic_bias = vld1q_dup_f32(&params->${PARAMS_STRUCT}.magic_bias);
          $for M in range(MR):
            $for N in range(0, NR, 4):
              vacc${M}x${ABC[N:N+4]} = vreinterpretq_s32_f32(vaddq_f32(vfpacc${M}x${ABC[N:N+4]}, vmagic_bias));

          const int32x4_t vmagic_bias_less_output_zero_point = vld1q_dup_s32(&params->${PARAMS_STRUCT}.magic_bias_less_output_zero_point);
          $for M in range(MR):
            $for N in range(0, NR, 4):
              vacc${M}x${ABC[N:N+4]} = vqsubq_s32(vacc${M}x${ABC[N:N+4]}, vmagic_bias_less_output_zero_point);

      $if REQUANTIZATION != "FP32" or ARMV8:
        const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->${PARAMS_STRUCT}.output_zero_point);
      #if XNN_ARCH_ARM64
        $if REQUANTIZATION != "RNDNU16":
          $for M in range(MR):
            $for N in range(0, NR, 8):
              int16x8_t vacc${M}x${ABC[N:N+8]} = vqmovn_high_s32(vqmovn_s32(vacc${M}x${ABC[N:N+4]}), vacc${M}x${ABC[N+4:N+8]});

        $if REQUANTIZATION != "FP32" or ARMV8:
          $for M in range(MR):
            $for N in range(0, NR, 8):
              vacc${M}x${ABC[N:N+8]} = vqaddq_s16(vacc${M}x${ABC[N:N+8]}, voutput_zero_point);

        $for M in range(MR):
          $for N in range(0, NR, 16):
            $if N + 8 < NR:
              ${XINT8X16_T} vout${M}x${ABC[N:N+16]} = ${VQMOVXN_HIGH_S16}(${VQMOVXN_S16}(vacc${M}x${ABC[N:N+8]}), vacc${M}x${ABC[N+8:N+16]});
            $elif M % 2 == 1:
              ${XINT8X16_T} vout${M-1}x${ABC[N:N+8]}_${M}x${ABC[N:N+8]} = ${VQMOVXN_HIGH_S16}(${VQMOVXN_S16}(vacc${M-1}x${ABC[N:N+8]}), vacc${M}x${ABC[N:N+8]});
            $elif M + 1 == MR:
              ${XINT8X8_T} vout${M}x${ABC[N:N+8]} = ${VQMOVXN_S16}(vacc${M}x${ABC[N:N+8]});
      #else
        $if REQUANTIZATION != "RNDNU16":
          $for M in range(MR):
            $for N in range(0, NR, 8):
              int16x8_t vacc${M}x${ABC[N:N+8]} = vcombine_s16(vqmovn_s32(vacc${M}x${ABC[N:N+4]}), vqmovn_s32(vacc${M}x${ABC[N+4:N+8]}));

        $if REQUANTIZATION != "FP32" or ARMV8:
          $for M in range(MR):
            $for N in range(0, NR, 8):
              vacc${M}x${ABC[N:N+8]} = vqaddq_s16(vacc${M}x${ABC[N:N+8]}, voutput_zero_point);

        $for M in range(MR):
          $for N in range(0, NR, 16):
            $if N + 8 < NR:
              ${XINT8X16_T} vout${M}x${ABC[N:N+16]} = ${VCOMBINE_X8}(${VQMOVXN_S16}(vacc${M}x${ABC[N:N+8]}), ${VQMOVXN_S16}(vacc${M}x${ABC[N+8:N+16]}));
            $elif M % 2 == 1:
              ${XINT8X16_T} vout${M-1}x${ABC[N:N+8]}_${M}x${ABC[N:N+8]} = ${VCOMBINE_X8}(${VQMOVXN_S16}(vacc${M-1}x${ABC[N:N+8]}), ${VQMOVXN_S16}(vacc${M}x${ABC[N:N+8]}));
            $elif M + 1 == MR:
              ${XINT8X8_T} vout${M}x${ABC[N:N+8]} = ${VQMOVXN_S16}(vacc${M}x${ABC[N:N+8]});
      #endif

      $if NR == 8 and MR == 1:
        const ${XINT8X8_T} voutput_min = ${VLD1_DUP_X8}(&params->${PARAMS_STRUCT}.output_min);
      $else:
        const ${XINT8X16_T} voutput_min = ${VLD1Q_DUP_X8}(&params->${PARAMS_STRUCT}.output_min);
      $for M in range(MR):
        $for N in range(0, NR, 16):
          $if N + 8 < NR:
            vout${M}x${ABC[N:N+16]} = ${VMAXQ_X8}(vout${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]} = ${VMAXQ_X8}(vout${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]} = ${VMAX_X8}(vout${M}x${ABC[N:N+8]}, voutput_min);
            $else:
              vout${M}x${ABC[N:N+8]} = ${VMAX_X8}(vout${M}x${ABC[N:N+8]}, ${VGET_LOW_X8}(voutput_min));

      $if NR == 8 and MR == 1:
        const ${XINT8X8_T} voutput_max = ${VLD1_DUP_X8}(&params->${PARAMS_STRUCT}.output_max);
      $else:
        const ${XINT8X16_T} voutput_max = ${VLD1Q_DUP_X8}(&params->${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]} = ${VMINQ_X8}(vout${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]} = ${VMINQ_X8}(vout${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]} = ${VMIN_X8}(vout${M}x${ABC[N:N+8]}, voutput_max);
            $else:
              vout${M}x${ABC[N:N+8]} = ${VMIN_X8}(vout${M}x${ABC[N:N+8]}, ${VGET_LOW_X8}(voutput_max));

      if (nc >= ${NR}) {
        $for M in reversed(range(MR)):
          $for N in range(0, NR, 16):
            $if N + 8 < NR:
              ${VST1Q_X8}(c${M} + ${N}, vout${M}x${ABC[N:N+16]});
            $elif M % 2 == 1:
              ${VST1_X8}(c${M} + ${N}, ${VGET_HIGH_X8}(vout${M-1}x${ABC[N:N+8]}_${M}x${ABC[N:N+8]}));
              ${VST1_X8}(c${M-1} + ${N}, ${VGET_LOW_X8}(vout${M-1}x${ABC[N:N+8]}_${M}x${ABC[N:N+8]}));
            $elif M + 1 == MR:
              ${VST1_X8}(c${M} + ${N}, vout${M}x${ABC[N:N+8]});

        $for M in reversed(range(MR)):
          c${M} = (${XINT8_T}*) ((uintptr_t) c${M} + cn_stride);

        a = (const ${XINT8_T}**restrict) ((uintptr_t) a - ks);

        nc -= ${NR};
      } else {
        $if NR == 16:
          $for M in reversed(range(MR)):
            $if M % 2 == 1:
              ${XINT8X16_T} vout${M-1}x01234567_${M}x01234567 = ${VCOMBINE_X8}(${VGET_LOW_X8}(vout${M-1}x0123456789ABCDEF), ${VGET_LOW_X8}(vout${M}x0123456789ABCDEF));
            $elif M + 1 == MR:
              ${XINT8X8_T} vout${M}x01234567 = ${VGET_LOW_X8}(vout${M}x0123456789ABCDEF);
          if (nc & 8) {
            $for M in reversed(range(MR)):
              $if M % 2 == 1:
                ${VST1_X8}(c${M}, ${VGET_HIGH_X8}(vout${M-1}x01234567_${M}x01234567)); c${M} += 8;
                ${VST1_X8}(c${M-1}, ${VGET_LOW_X8}(vout${M-1}x01234567_${M}x01234567)); c${M-1} += 8;
              $elif M + 1 == MR:
                ${VST1_X8}(c${M}, vout${M}x01234567); c${M} += 8;
            $for M in reversed(range(MR)):
              $if M % 2 == 1:
                vout${M-1}x01234567_${M}x01234567 = ${VCOMBINE_X8}(${VGET_HIGH_X8}(vout${M-1}x0123456789ABCDEF), ${VGET_HIGH_X8}(vout${M}x0123456789ABCDEF));
              $elif M + 1 == MR:
                vout${M}x01234567 = ${VGET_HIGH_X8}(vout${M}x0123456789ABCDEF);
          }
        if (nc & 4) {
          $for M in reversed(range(MR)):
            $if M % 2 == 1:
              vst1q_lane_u32((void*) c${M}, ${VREINTERPRETQ_U32_X8}(vout${M-1}x01234567_${M}x01234567), 2); c${M} += 4;
              vst1q_lane_u32((void*) c${M-1}, ${VREINTERPRETQ_U32_X8}(vout${M-1}x01234567_${M}x01234567), 0); c${M-1} += 4;
            $elif M + 1 == MR:
              vst1_lane_u32((void*) c${M}, ${VREINTERPRET_U32_X8}(vout${M}x01234567), 0); c${M} += 4;
          $for M in reversed(range(MR)):
            $if M % 2 == 1:
              vout${M-1}x01234567_${M}x01234567 = ${VEXTQ_X8}(vout${M-1}x01234567_${M}x01234567, vout${M-1}x01234567_${M}x01234567, 4);
            $elif M + 1 == MR:
              vout${M}x01234567 = ${VEXT_X8}(vout${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}, ${VREINTERPRETQ_U16_X8}(vout${M-1}x01234567_${M}x01234567), 4); c${M} += 2;
              vst1q_lane_u16((void*) c${M-1}, ${VREINTERPRETQ_U16_X8}(vout${M-1}x01234567_${M}x01234567), 0); c${M-1} += 2;
            $elif M + 1 == MR:
              vst1_lane_u16((void*) c${M}, ${VREINTERPRET_U16_X8}(vout${M}x01234567), 0); c${M} += 2;
          $for M in reversed(range(MR)):
            $if M % 2 == 1:
              vout${M-1}x01234567_${M}x01234567 = ${VEXTQ_X8}(vout${M-1}x01234567_${M}x01234567, vout${M-1}x01234567_${M}x01234567, 2);
            $elif M + 1 == MR:
              vout${M}x01234567 = ${VEXT_X8}(vout${M}x01234567, vout${M}x01234567, 2);
        }
        if (nc & 1) {
          $for M in reversed(range(MR)):
            $if M % 2 == 1:
              ${VST1Q_LANE_X8}(c${M}, vout${M-1}x01234567_${M}x01234567, 8);
              ${VST1Q_LANE_X8}(c${M-1}, vout${M-1}x01234567_${M}x01234567, 0);
            $elif M + 1 == MR:
              ${VST1_LANE_X8}(c${M}, vout${M}x01234567, 0);
        }

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