// 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.

$assert DATATYPE in ["QS8", "QU8"]
$assert CHANNEL_TILE % 8 == 0
$assert CHANNEL_TILE >= 8
$assert ROW_TILE >= 3
$assert ROW_SUBTILE >= 3
$assert ROW_SUBTILE <= ROW_TILE
$assert REQUANTIZATION in ["FP32", "RNDNU"]
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
#include <assert.h>

#include <arm_neon.h>

#include <xnnpack/gavgpool.h>
$if ARMV8:
  #include <xnnpack/intrinsics-polyfill.h>
#include <xnnpack/math.h>


$PARAMS_STRUCT = REQUANTIZATION.lower() + "_" + ("neonv8" if ARMV8 else "neon")
$XINT8_T = {"QS8": "int8_t", "QU8": "uint8_t"}[DATATYPE]
$XINT8X8_T = {"QS8": "int8x8_t", "QU8": "uint8x8_t"}[DATATYPE]
$XINT8X16_T = {"QS8": "int8x16_t", "QU8": "uint8x16_t"}[DATATYPE]
$XINT16X8_T = {"QS8": "int16x8_t", "QU8": "uint16x8_t"}[DATATYPE]
$VLD1_X8 = {"QS8": "vld1_s8", "QU8": "vld1_u8"}[DATATYPE]
$VLD1_DUP_X8 = {"QS8": "vld1_dup_s8", "QU8": "vld1_dup_u8"}[DATATYPE]
$VLD1Q_DUP_X8 = {"QS8": "vld1q_dup_s8", "QU8": "vld1q_dup_u8"}[DATATYPE]
$VST1_X8 = {"QS8": "vst1_s8", "QU8": "vst1_u8"}[DATATYPE]
$VST1Q_X8 = {"QS8": "vst1q_s8", "QU8": "vst1q_u8"}[DATATYPE]
$VST1_LANE_X8 = {"QS8": "vst1_lane_s8", "QU8": "vst1_lane_u8"}[DATATYPE]
$VADDL_X8 = {"QS8": "vaddl_s8", "QU8": "vaddl_u8"}[DATATYPE]
$VADDW_X8 = {"QS8": "vaddw_s8", "QU8": "vaddw_u8"}[DATATYPE]
$VMIN_X8 = {"QS8": "vmin_s8", "QU8": "vmin_u8"}[DATATYPE]
$VMINQ_X8 = {"QS8": "vminq_s8", "QU8": "vminq_u8"}[DATATYPE]
$VMAX_X8 = {"QS8": "vmax_s8", "QU8": "vmax_u8"}[DATATYPE]
$VMAXQ_X8 = {"QS8": "vmaxq_s8", "QU8": "vmaxq_u8"}[DATATYPE]
$VEXT_X8 = {"QS8": "vext_s8", "QU8": "vext_u8"}[DATATYPE]
$VQMOVXN_S16 = {"QS8": "vqmovn_s16", "QU8": "vqmovun_s16"}[DATATYPE]
$VQMOVXN_HIGH_S16 = {"QS8": "vqmovn_high_s16", "QU8": "vqmovun_high_s16"}[DATATYPE]
$VGET_LOW_X8 = {"QS8": "vget_low_s8", "QU8": "vget_low_u8"}[DATATYPE]
$VCOMBINE_X8 = {"QS8": "vcombine_s8", "QU8": "vcombine_u8"}[DATATYPE]
$VREINTERPRET_U32_X8 = {"QS8": "vreinterpret_u32_s8", "QU8": "vreinterpret_u32_u8"}[DATATYPE]
$VREINTERPRET_U16_X8 = {"QS8": "vreinterpret_u16_s8", "QU8": "vreinterpret_u16_u8"}[DATATYPE]
$ISA = "neonv8" if ARMV8 else "neon"
void xnn_${DATATYPE.lower()}_gavgpool_minmax_${REQUANTIZATION.lower()}_ukernel_${ROW_TILE}p${ROW_SUBTILE}x__${ISA}_c${CHANNEL_TILE}(
    size_t rows,
    size_t channels,
    const ${XINT8_T}* input,
    size_t input_stride,
    const ${XINT8_T}* zero,
    int32_t* buffer,
    ${XINT8_T}* output,
    const union xnn_${DATATYPE.lower()}_avgpool_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(rows > ${ROW_TILE});
  assert(channels != 0);

  const ${XINT8_T}* i0 = input;
  $for M in range(1, ROW_TILE):
    const ${XINT8_T}* i${M} = (const ${XINT8_T}*) ((uintptr_t) i${M-1} + input_stride);
  $if CHANNEL_TILE <= 16:
    const size_t input_increment = ${ROW_TILE} * input_stride - round_up_po2(channels, ${CHANNEL_TILE}) * sizeof(${XINT8_T});
  $else:
    const size_t input_increment = ${ROW_TILE} * input_stride - round_up_po2(channels, 8) * sizeof(${XINT8_T});

  const int32x4_t vinit_bias = vld1q_dup_s32(&params->${PARAMS_STRUCT}.init_bias);
  int32_t* b = buffer;
  size_t c = channels;
  for (; ${"c >= %d" % CHANNEL_TILE if CHANNEL_TILE > 16 else "c != 0"}; ${("c -= %d" if CHANNEL_TILE > 16 else "c = doz(c, %d)") % CHANNEL_TILE}) {
    $for M in range(2):
      $for C in range(0, CHANNEL_TILE, 8):
        const ${XINT8X8_T} vi${M}x${ABC[C:C+8]} = ${VLD1_X8}(i${M}); i${M} += 8;

    $for C in range(0, CHANNEL_TILE, 8):
      const ${XINT8X8_T} vi2x${ABC[C:C+8]} = ${VLD1_X8}(i2); i2 += 8;
      ${XINT16X8_T} vsum${ABC[C:C+8]} = ${VADDL_X8}(vi0x${ABC[C:C+8]}, vi1x${ABC[C:C+8]});

    $for M in range(2, ROW_TILE):
      $for C in range(0, CHANNEL_TILE, 8):
        $if M + 1 != ROW_TILE:
          const ${XINT8X8_T} vi${M+1}x${ABC[C:C+8]} = ${VLD1_X8}(i${M+1}); i${M+1} += 8;
        vsum${ABC[C:C+8]} = ${VADDW_X8}(vsum${ABC[C:C+8]}, vi${M}x${ABC[C:C+8]});

    $for C in range(0, CHANNEL_TILE, 8):
      $if DATATYPE == "QS8":
        const int32x4_t vacc${ABC[C:C+4]} = vaddw_s16(vinit_bias, vget_low_s16(vsum${ABC[C:C+8]}));
        const int32x4_t vacc${ABC[C+4:C+8]} = vaddw_s16(vinit_bias, vget_high_s16(vsum${ABC[C:C+8]}));
      $else:
        const int32x4_t vacc${ABC[C:C+4]} = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_low_u16(vsum${ABC[C:C+8]})));
        const int32x4_t vacc${ABC[C+4:C+8]} = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_high_u16(vsum${ABC[C:C+8]})));

    $for C in range(0, CHANNEL_TILE, 4):
      vst1q_s32(b, vacc${ABC[C:C+4]}); b += 4;
  }
  $if CHANNEL_TILE > 16:
    if XNN_UNLIKELY(c != 0) {
      do {
        $for M in range(3):
          const ${XINT8X8_T} vi${M}x${ABC[0:8]} = ${VLD1_X8}(i${M}); i${M} += 8;
        ${XINT16X8_T} vsum${ABC[0:8]} = ${VADDL_X8}(vi0x${ABC[0:8]}, vi1x${ABC[0:8]});

        $for M in range(2, ROW_TILE):
          $if M + 1 != ROW_TILE:
            const ${XINT8X8_T} vi${M+1}x${ABC[0:8]} = ${VLD1_X8}(i${M+1}); i${M+1} += 8;
          vsum${ABC[0:8]} = ${VADDW_X8}(vsum${ABC[0:8]}, vi${M}x${ABC[0:8]});

        $if DATATYPE == "QS8":
          const int32x4_t vacc${ABC[0:4]} = vaddw_s16(vinit_bias, vget_low_s16(vsum${ABC[0:8]}));
          const int32x4_t vacc${ABC[4:8]} = vaddw_s16(vinit_bias, vget_high_s16(vsum${ABC[0:8]}));
        $else:
          const int32x4_t vacc${ABC[0:4]} = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_low_u16(vsum${ABC[0:8]})));
          const int32x4_t vacc${ABC[4:8]} = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_high_u16(vsum${ABC[0:8]})));

        vst1q_s32(b, vacc${ABC[0:4]}); b += 4;
        vst1q_s32(b, vacc${ABC[4:8]}); b += 4;

        c = doz(c, 8);
      } while (c != 0);
    }

  for (rows -= ${ROW_TILE}; rows > ${ROW_SUBTILE}; rows -= ${ROW_SUBTILE}) {
    $for M in range(ROW_SUBTILE):
      i${M} = (const ${XINT8_T}*) ((uintptr_t) i${M + ROW_TILE - ROW_SUBTILE} + input_increment);

    int32_t* b = buffer;
    size_t c = channels;
    for (; ${"c >= %d" % CHANNEL_TILE if CHANNEL_TILE > 16 else "c != 0"}; ${("c -= %d" if CHANNEL_TILE > 16 else "c = doz(c, %d)") % CHANNEL_TILE}) {
      $for M in range(2):
        $for C in range(0, CHANNEL_TILE, 8):
          const ${XINT8X8_T} vi${M}x${ABC[C:C+8]} = ${VLD1_X8}(i${M}); i${M} += 8;

      $for C in range(0, CHANNEL_TILE, 8):
        const ${XINT8X8_T} vi2x${ABC[C:C+8]} = ${VLD1_X8}(i2); i2 += 8;
        ${XINT16X8_T} vsum${ABC[C:C+8]} = ${VADDL_X8}(vi0x${ABC[C:C+8]}, vi1x${ABC[C:C+8]});

      $for M in range(2, ROW_TILE):
        $for C in range(0, CHANNEL_TILE, 8):
          $if M + 1 != ROW_TILE:
            const ${XINT8X8_T} vi${M+1}x${ABC[C:C+8]} = ${VLD1_X8}(i${M+1}); i${M+1} += 8;
          $else:
            $if C == 0:
              int32x4_t vacc${ABC[C:C+4]} = vld1q_s32(b);
            $else:
              int32x4_t vacc${ABC[C:C+4]} = vld1q_s32(b + ${C});
            int32x4_t vacc${ABC[C+4:C+8]} = vld1q_s32(b + ${C+4});
          vsum${ABC[C:C+8]} = ${VADDW_X8}(vsum${ABC[C:C+8]}, vi${M}x${ABC[C:C+8]});

      $for C in range(0, CHANNEL_TILE, 8):
        $if DATATYPE == "QS8":
          vacc${ABC[C:C+4]} = vaddw_s16(vacc${ABC[C:C+4]}, vget_low_s16(vsum${ABC[C:C+8]}));
          vacc${ABC[C+4:C+8]} = vaddw_s16(vacc${ABC[C+4:C+8]}, vget_high_s16(vsum${ABC[C:C+8]}));
        $else:
          vacc${ABC[C:C+4]} = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc${ABC[C:C+4]}), vget_low_u16(vsum${ABC[C:C+8]})));
          vacc${ABC[C+4:C+8]} = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc${ABC[C+4:C+8]}), vget_high_u16(vsum${ABC[C:C+8]})));

      $for C in range(0, CHANNEL_TILE, 4):
        vst1q_s32(b, vacc${ABC[C:C+4]}); b += 4;
    }
    $if CHANNEL_TILE > 16:
      if XNN_UNLIKELY(c != 0) {
        do {
          $for M in range(3):
            const ${XINT8X8_T} vi${M}x${ABC[0:8]} = ${VLD1_X8}(i${M}); i${M} += 8;
          ${XINT16X8_T} vsum${ABC[0:8]} = ${VADDL_X8}(vi0x${ABC[0:8]}, vi1x${ABC[0:8]});

          $for M in range(2, ROW_TILE):
            $if M + 1 != ROW_TILE:
              const ${XINT8X8_T} vi${M+1}x${ABC[0:8]} = ${VLD1_X8}(i${M+1}); i${M+1} += 8;
            $else:
              int32x4_t vacc${ABC[0:4]} = vld1q_s32(b);
              int32x4_t vacc${ABC[4:8]} = vld1q_s32(b + 4);
            vsum${ABC[0:8]} = ${VADDW_X8}(vsum${ABC[0:8]}, vi${M}x${ABC[0:8]});

          $if DATATYPE == "QS8":
            vacc${ABC[0:4]} = vaddw_s16(vacc${ABC[0:4]}, vget_low_s16(vsum${ABC[0:8]}));
            vacc${ABC[4:8]} = vaddw_s16(vacc${ABC[4:8]}, vget_high_s16(vsum${ABC[0:8]}));
          $else:
            vacc${ABC[0:4]} = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc${ABC[0:4]}), vget_low_u16(vsum${ABC[0:8]})));
            vacc${ABC[4:8]} = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc${ABC[4:8]}), vget_high_u16(vsum${ABC[0:8]})));

          vst1q_s32(b, vacc${ABC[0:4]}); b += 4;
          vst1q_s32(b, vacc${ABC[4:8]}); b += 4;

          c = doz(c, 8);
        } while (c != 0);
      }
  }

  i0 = (const ${XINT8_T}*) ((uintptr_t) i${ROW_TILE - ROW_SUBTILE} + input_increment);
  $for M in range(1, ROW_SUBTILE):
    i${M} = (const ${XINT8_T}*) ((uintptr_t) i${M + ROW_TILE - ROW_SUBTILE} + input_increment);
    $if M % 2 == 1:
      if XNN_UNPREDICTABLE(rows < ${M+1}) {
        i${M} = zero;
      }
    $else:
      if XNN_UNPREDICTABLE(rows <= ${M}) {
        i${M} = zero;
      }

  $if REQUANTIZATION == "FP32":
    const float32x4_t vscale = vld1q_dup_f32(&params->${PARAMS_STRUCT}.scale);
    $if ARMV8:
      const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->fp32_neonv8.output_zero_point);
    $else:
      const float32x4_t vmagic_bias = vld1q_dup_f32(&params->fp32_neon.magic_bias);
      const int32x4_t vmagic_bias_less_output_zero_point = vld1q_dup_s32(&params->fp32_neon.magic_bias_less_output_zero_point);
  $elif REQUANTIZATION == "RNDNU":
    const int32x4_t vleft_pre_shift = vld1q_dup_s32(&params->rndnu_neon.left_pre_shift);
    const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
    const int32x4_t vleft_post_shift = vld1q_dup_s32(&params->rndnu_neon.left_post_shift);
    const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
  $if CHANNEL_TILE > 8:
    const ${XINT8X16_T} voutput_min = ${VLD1Q_DUP_X8}(&params->${PARAMS_STRUCT}.output_min);
    const ${XINT8X16_T} voutput_max = ${VLD1Q_DUP_X8}(&params->${PARAMS_STRUCT}.output_max);
  $else:
    const ${XINT8X8_T} voutput_min = ${VLD1_DUP_X8}(&params->${PARAMS_STRUCT}.output_min);
    const ${XINT8X8_T} voutput_max = ${VLD1_DUP_X8}(&params->${PARAMS_STRUCT}.output_max);
  for (; channels >= ${CHANNEL_TILE}; channels -= ${CHANNEL_TILE}) {
    $for M in range(2):
      $for C in range(0, CHANNEL_TILE, 8):
        const ${XINT8X8_T} vi${M}x${ABC[C:C+8]} = ${VLD1_X8}(i${M}); i${M} += 8;

    $for C in range(0, CHANNEL_TILE, 8):
      const ${XINT8X8_T} vi2x${ABC[C:C+8]} = ${VLD1_X8}(i2); i2 += 8;
      ${XINT16X8_T} vsum${ABC[C:C+8]} = ${VADDL_X8}(vi0x${ABC[C:C+8]}, vi1x${ABC[C:C+8]});

    $for M in range(2, ROW_TILE):
      $for C in range(0, CHANNEL_TILE, 8):
        $if M + 1 != ROW_TILE:
          const ${XINT8X8_T} vi${M+1}x${ABC[C:C+8]} = ${VLD1_X8}(i${M+1}); i${M+1} += 8;
        $else:
          int32x4_t vacc${ABC[C:C+4]} = vld1q_s32(buffer); buffer += 4;
          int32x4_t vacc${ABC[C+4:C+8]} = vld1q_s32(buffer); buffer += 4;
        vsum${ABC[C:C+8]} = ${VADDW_X8}(vsum${ABC[C:C+8]}, vi${M}x${ABC[C:C+8]});

    $for C in range(0, CHANNEL_TILE, 8):
      $if DATATYPE == "QS8":
        vacc${ABC[C:C+4]} = vaddw_s16(vacc${ABC[C:C+4]}, vget_low_s16(vsum${ABC[C:C+8]}));
        vacc${ABC[C+4:C+8]} = vaddw_s16(vacc${ABC[C+4:C+8]}, vget_high_s16(vsum${ABC[C:C+8]}));
      $else:
        vacc${ABC[C:C+4]} = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc${ABC[C:C+4]}), vget_low_u16(vsum${ABC[C:C+8]})));
        vacc${ABC[C+4:C+8]} = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc${ABC[C+4:C+8]}), vget_high_u16(vsum${ABC[C:C+8]})));

    $if REQUANTIZATION == "FP32":
      $for C in range(0, CHANNEL_TILE, 4):
        float32x4_t vfpacc${ABC[C:C+4]} = vcvtq_f32_s32(vacc${ABC[C:C+4]});

      $for C in range(0, CHANNEL_TILE, 4):
        vfpacc${ABC[C:C+4]} = vmulq_f32(vfpacc${ABC[C:C+4]}, vscale);

      $if ARMV8:
        $for C in range(0, CHANNEL_TILE, 4):
          vacc${ABC[C:C+4]} = vcvtnq_s32_f32(vfpacc${ABC[C:C+4]});
      $else:
        $for C in range(0, CHANNEL_TILE, 4):
          vacc${ABC[C:C+4]} = vreinterpretq_s32_f32(vaddq_f32(vfpacc${ABC[C:C+4]}, vmagic_bias));

        $for C in range(0, CHANNEL_TILE, 4):
          vacc${ABC[C:C+4]} = vqsubq_s32(vacc${ABC[C:C+4]}, vmagic_bias_less_output_zero_point);
    $elif REQUANTIZATION == "RNDNU":
      $for C in range(0, CHANNEL_TILE, 4):
        vacc${ABC[C:C+4]} = vqshlq_s32(vacc${ABC[C:C+4]}, vleft_pre_shift);

      $for C in range(0, CHANNEL_TILE, 4):
        vacc${ABC[C:C+4]} = vqdmulhq_s32(vacc${ABC[C:C+4]}, vmultiplier);

      $for C in range(0, CHANNEL_TILE, 4):
        vacc${ABC[C:C+4]} = vrshlq_s32(vacc${ABC[C:C+4]}, vleft_post_shift);

    #if XNN_ARCH_ARM64
      $for C in range(0, CHANNEL_TILE, 8):
        int16x8_t vacc${ABC[C:C+8]} = vqmovn_high_s32(vqmovn_s32(vacc${ABC[C:C+4]}), vacc${ABC[C+4:C+8]});
    #else  // !XNN_ARCH_ARM64
      $for C in range(0, CHANNEL_TILE, 8):
        int16x8_t vacc${ABC[C:C+8]} = vcombine_s16(vqmovn_s32(vacc${ABC[C:C+4]}), vqmovn_s32(vacc${ABC[C+4:C+8]}));
    #endif  // !XNN_ARCH_ARM64

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

    #if XNN_ARCH_ARM64
      $for C in range(0, CHANNEL_TILE, 16):
        $if C + 8 < CHANNEL_TILE:
          ${XINT8X16_T} vout${ABC[C:C+16]} = ${VQMOVXN_HIGH_S16}(${VQMOVXN_S16}(vacc${ABC[C:C+8]}), vacc${ABC[C+8:C+16]});
        $else:
          ${XINT8X8_T} vout${ABC[C:C+8]} = ${VQMOVXN_S16}(vacc${ABC[C:C+8]});
    #else  // !XNN_ARCH_ARM64
      $for C in range(0, CHANNEL_TILE, 16):
        $if C + 8 < CHANNEL_TILE:
          ${XINT8X16_T} vout${ABC[C:C+16]} = ${VCOMBINE_X8}(${VQMOVXN_S16}(vacc${ABC[C:C+8]}), ${VQMOVXN_S16}(vacc${ABC[C+8:C+16]}));
        $else:
          ${XINT8X8_T} vout${ABC[C:C+8]} = ${VQMOVXN_S16}(vacc${ABC[C:C+8]});
    #endif  // !XNN_ARCH_ARM64

    $for C in range(0, CHANNEL_TILE, 16):
      $if C + 8 < CHANNEL_TILE:
        vout${ABC[C:C+16]} = ${VMAXQ_X8}(vout${ABC[C:C+16]}, voutput_min);
      $elif CHANNEL_TILE > 8:
        vout${ABC[C:C+8]} = ${VMAX_X8}(vout${ABC[C:C+8]}, ${VGET_LOW_X8}(voutput_min));
      $else:
        vout${ABC[C:C+8]} = ${VMAX_X8}(vout${ABC[C:C+8]}, voutput_min);

    $for C in range(0, CHANNEL_TILE, 16):
      $if C + 8 < CHANNEL_TILE:
        vout${ABC[C:C+16]} = ${VMINQ_X8}(vout${ABC[C:C+16]}, voutput_max);
      $elif CHANNEL_TILE > 8:
        vout${ABC[C:C+8]} = ${VMIN_X8}(vout${ABC[C:C+8]}, ${VGET_LOW_X8}(voutput_max));
      $else:
        vout${ABC[C:C+8]} = ${VMIN_X8}(vout${ABC[C:C+8]}, voutput_max);

    $for C in range(0, CHANNEL_TILE, 16):
      $if C + 8 < CHANNEL_TILE:
        ${VST1Q_X8}(output, vout${ABC[C:C+16]}); output += 16;
      $else:
        ${VST1_X8}(output, vout${ABC[C:C+8]}); output += 8;
  }
  if XNN_UNLIKELY(channels != 0) {
    ${"do " if CHANNEL_TILE > 8 else ""}{
      $for M in range(3):
        $if CHANNEL_TILE > 8:
          const ${XINT8X8_T} vi${M}x${ABC[0:8]} = ${VLD1_X8}(i${M}); i${M} += 8;
        $else:
          const ${XINT8X8_T} vi${M}x${ABC[0:8]} = ${VLD1_X8}(i${M});
      ${XINT16X8_T} vsum${ABC[0:8]} = ${VADDL_X8}(vi0x${ABC[0:8]}, vi1x${ABC[0:8]});

      $for M in range(2, ROW_TILE):
        $if M + 1 != ROW_TILE:
          $if CHANNEL_TILE > 8:
            const ${XINT8X8_T} vi${M+1}x${ABC[0:8]} = ${VLD1_X8}(i${M+1}); i${M+1} += 8;
          $else:
            const ${XINT8X8_T} vi${M+1}x${ABC[0:8]} = ${VLD1_X8}(i${M+1});
        $else:
          int32x4_t vacc${ABC[0:4]} = vld1q_s32(buffer); buffer += 4;
          int32x4_t vacc${ABC[4:8]} = vld1q_s32(buffer); buffer += 4;
        vsum${ABC[0:8]} = ${VADDW_X8}(vsum${ABC[0:8]}, vi${M}x${ABC[0:8]});

      $if DATATYPE == "QS8":
        vacc${ABC[0:4]} = vaddw_s16(vacc${ABC[0:4]}, vget_low_s16(vsum${ABC[0:8]}));
        vacc${ABC[4:8]} = vaddw_s16(vacc${ABC[4:8]}, vget_high_s16(vsum${ABC[0:8]}));
      $else:
        vacc${ABC[0:4]} = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc${ABC[0:4]}), vget_low_u16(vsum${ABC[0:8]})));
        vacc${ABC[4:8]} = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc${ABC[4:8]}), vget_high_u16(vsum${ABC[0:8]})));

      $if REQUANTIZATION == "FP32":
        float32x4_t vfpacc${ABC[0:4]} = vcvtq_f32_s32(vacc${ABC[0:4]});
        float32x4_t vfpacc${ABC[4:8]} = vcvtq_f32_s32(vacc${ABC[4:8]});

        vfpacc${ABC[0:4]} = vmulq_f32(vfpacc${ABC[0:4]}, vscale);
        vfpacc${ABC[4:8]} = vmulq_f32(vfpacc${ABC[4:8]}, vscale);

        $if ARMV8:
          vacc${ABC[0:4]} = vcvtnq_s32_f32(vfpacc${ABC[0:4]});
          vacc${ABC[4:8]} = vcvtnq_s32_f32(vfpacc${ABC[4:8]});
        $else:
          vacc${ABC[0:4]} = vreinterpretq_s32_f32(vaddq_f32(vfpacc${ABC[0:4]}, vmagic_bias));
          vacc${ABC[4:8]} = vreinterpretq_s32_f32(vaddq_f32(vfpacc${ABC[4:8]}, vmagic_bias));

          vacc${ABC[0:4]} = vqsubq_s32(vacc${ABC[0:4]}, vmagic_bias_less_output_zero_point);
          vacc${ABC[4:8]} = vqsubq_s32(vacc${ABC[4:8]}, vmagic_bias_less_output_zero_point);
      $elif REQUANTIZATION == "RNDNU":
        vacc${ABC[0:4]} = vqshlq_s32(vacc${ABC[0:4]}, vleft_pre_shift);
        vacc${ABC[4:8]} = vqshlq_s32(vacc${ABC[4:8]}, vleft_pre_shift);

        vacc${ABC[0:4]} = vqdmulhq_s32(vacc${ABC[0:4]}, vmultiplier);
        vacc${ABC[4:8]} = vqdmulhq_s32(vacc${ABC[4:8]}, vmultiplier);

        vacc${ABC[0:4]} = vrshlq_s32(vacc${ABC[0:4]}, vleft_post_shift);
        vacc${ABC[4:8]} = vrshlq_s32(vacc${ABC[4:8]}, vleft_post_shift);

      #if XNN_ARCH_ARM64
        int16x8_t vacc${ABC[0:8]} = vqmovn_high_s32(vqmovn_s32(vacc${ABC[0:4]}), vacc${ABC[4:8]});
      #else
        int16x8_t vacc${ABC[0:8]} = vcombine_s16(vqmovn_s32(vacc${ABC[0:4]}), vqmovn_s32(vacc${ABC[4:8]}));
      #endif
      $if REQUANTIZATION != "FP32" or ARMV8:
        vacc${ABC[0:8]} = vqaddq_s16(vacc${ABC[0:8]}, voutput_zero_point);

      ${XINT8X8_T} vout${ABC[0:8]} = ${VQMOVXN_S16}(vacc${ABC[0:8]});
      $if CHANNEL_TILE > 8:
        vout${ABC[0:8]} = ${VMAX_X8}(vout${ABC[0:8]}, ${VGET_LOW_X8}(voutput_min));
        vout${ABC[0:8]} = ${VMIN_X8}(vout${ABC[0:8]}, ${VGET_LOW_X8}(voutput_max));

        if XNN_LIKELY(channels >= 8) {
          ${VST1_X8}(output, vout${ABC[0:8]}); output += 8;
          channels -= 8;
        } else {
          if (channels & 4) {
            vst1_lane_u32((void*) output, ${VREINTERPRET_U32_X8}(vout${ABC[0:8]}), 0); output += 4;
            vout${ABC[0:8]} = ${VEXT_X8}(vout${ABC[0:8]}, vout${ABC[0:8]}, 4);
          }
          if (channels & 2) {
            vst1_lane_u16((void*) output, ${VREINTERPRET_U16_X8}(vout${ABC[0:8]}), 0); output += 2;
            vout${ABC[0:8]} = ${VEXT_X8}(vout${ABC[0:8]}, vout${ABC[0:8]}, 2);
          }
          if (channels & 1) {
            ${VST1_LANE_X8}(output, vout${ABC[0:8]}, 0); output += 1;
          }
          channels = 0;
        }
      $else:
        vout${ABC[0:8]} = ${VMAX_X8}(vout${ABC[0:8]}, voutput_min);
        vout${ABC[0:8]} = ${VMIN_X8}(vout${ABC[0:8]}, voutput_max);

        if (channels & 4) {
          vst1_lane_u32((void*) output, ${VREINTERPRET_U32_X8}(vout${ABC[0:8]}), 0); output += 4;
          vout${ABC[0:8]} = ${VEXT_X8}(vout${ABC[0:8]}, vout${ABC[0:8]}, 4);
        }
        if (channels & 2) {
          vst1_lane_u16((void*) output, ${VREINTERPRET_U16_X8}(vout${ABC[0:8]}), 0); output += 2;
          vout${ABC[0:8]} = ${VEXT_X8}(vout${ABC[0:8]}, vout${ABC[0:8]}, 2);
        }
        if (channels & 1) {
          ${VST1_LANE_X8}(output, vout${ABC[0:8]}, 0);
        }
    }${" while (channels != 0);" if CHANNEL_TILE > 8 else ""}
  }
}