xref: /aosp_15_r20/external/XNNPACK/src/qs8-vmul/neon.c.in (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)

// 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.

$assert DATATYPE in ["QS8", "QU8"]
$assert REQUANTIZATION in ["FP32", "RNDNU"]
$assert BATCH_TILE % (16 if LD128 else 8) == 0
$assert BATCH_TILE >= 8
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
#include <assert.h>

#include <arm_neon.h>

$if REQUANTIZATION == "FP32" and ARMV8:
  #include <xnnpack/intrinsics-polyfill.h>
#include <xnnpack/vmul.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]
$VLD1_X8 = {"QS8": "vld1_s8", "QU8": "vld1_u8"}[DATATYPE]
$VLD1Q_X8 = {"QS8": "vld1q_s8", "QU8": "vld1q_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_LANE_X8 = {"QS8": "vst1_lane_s8", "QU8": "vst1_lane_u8"}[DATATYPE]
$VST1_X8 = {"QS8": "vst1_s8", "QU8": "vst1_u8"}[DATATYPE]
$VST1Q_X8 = {"QS8": "vst1q_s8", "QU8": "vst1q_u8"}[DATATYPE]
$VMIN_X8 = {"QS8": "vmin_s8", "QU8": "vmin_u8"}[DATATYPE]
$VMAX_X8 = {"QS8": "vmax_s8", "QU8": "vmax_u8"}[DATATYPE]
$VMINQ_X8 = {"QS8": "vminq_s8", "QU8": "vminq_u8"}[DATATYPE]
$VMAXQ_X8 = {"QS8": "vmaxq_s8", "QU8": "vmaxq_u8"}[DATATYPE]
$VQMOVXN_S16 = {"QS8": "vqmovn_s16", "QU8": "vqmovun_s16"}[DATATYPE]
$VQMOVXN_HIGH_S16 = {"QS8": "vqmovn_high_s16", "QU8": "vqmovun_high_s16"}[DATATYPE]
$VEXT_X8 = {"QS8": "vext_s8", "QU8": "vext_u8"}[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]
$VREINTERPRET_U16_X8 = {"QS8": "vreinterpret_u16_s8", "QU8": "vreinterpret_u16_u8"}[DATATYPE]
void xnn_${DATATYPE.lower()}_vmul_minmax_${REQUANTIZATION.lower()}_ukernel__${"neonv8" if ARMV8 else "neon"}_${"ld128" if LD128 else "ld64"}_x${BATCH_TILE}(
    size_t n,
    const ${XINT8_T}* input_a,
    const ${XINT8_T}* input_b,
    ${XINT8_T}* output,
    const union xnn_${DATATYPE.lower()}_mul_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  $if LD128:
    #if XNN_ARCH_ARM64
      const ${XINT8X16_T} va_zero_point = ${VLD1Q_DUP_X8}(params->${PARAMS_STRUCT}.a_zero_point);
      const ${XINT8X16_T} vb_zero_point = ${VLD1Q_DUP_X8}(params->${PARAMS_STRUCT}.b_zero_point);
    #else
      const ${XINT8X8_T} va_zero_point = ${VLD1_DUP_X8}(params->${PARAMS_STRUCT}.a_zero_point);
      const ${XINT8X8_T} vb_zero_point = ${VLD1_DUP_X8}(params->${PARAMS_STRUCT}.b_zero_point);
    #endif
  $else:
    const ${XINT8X8_T} va_zero_point = ${VLD1_DUP_X8}(params->${PARAMS_STRUCT}.a_zero_point);
    const ${XINT8X8_T} vb_zero_point = ${VLD1_DUP_X8}(params->${PARAMS_STRUCT}.b_zero_point);
  $if REQUANTIZATION == "FP32":
    const float32x4_t vscale = vld1q_dup_f32(&params->${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 BATCH_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 (; n >= ${BATCH_TILE} * sizeof(${XINT8_T}); n -= ${BATCH_TILE} * sizeof(${XINT8_T})) {
    $if LD128:
      $for N in range(0, BATCH_TILE, 16):
        const ${XINT8X16_T} va${ABC[N:N+16]} = ${VLD1Q_X8}(input_a); input_a += 16;
        const ${XINT8X16_T} vb${ABC[N:N+16]} = ${VLD1Q_X8}(input_b); input_b += 16;

      #if XNN_ARCH_ARM64
        $for N in range(0, BATCH_TILE, 16):
          $if DATATYPE == "QU8":
            const int16x8_t vxa${ABC[N:N+8]} = vreinterpretq_s16_u16(vsubl_u8(vget_low_u8(va${ABC[N:N+16]}), vget_low_u8(va_zero_point)));
            const int16x8_t vxa${ABC[N+8:N+16]} = vreinterpretq_s16_u16(vsubl_high_u8(va${ABC[N:N+16]}, va_zero_point));
            const int16x8_t vxb${ABC[N:N+8]} = vreinterpretq_s16_u16(vsubl_u8(vget_low_u8(vb${ABC[N:N+16]}), vget_low_u8(vb_zero_point)));
            const int16x8_t vxb${ABC[N+8:N+16]} = vreinterpretq_s16_u16(vsubl_high_u8(vb${ABC[N:N+16]}, vb_zero_point));
          $else:
            const int16x8_t vxa${ABC[N:N+8]} = vsubl_s8(vget_low_s8(va${ABC[N:N+16]}), vget_low_s8(va_zero_point));
            const int16x8_t vxa${ABC[N+8:N+16]} = vsubl_high_s8(va${ABC[N:N+16]}, va_zero_point);
            const int16x8_t vxb${ABC[N:N+8]} = vsubl_s8(vget_low_s8(vb${ABC[N:N+16]}), vget_low_s8(vb_zero_point));
            const int16x8_t vxb${ABC[N+8:N+16]} = vsubl_high_s8(vb${ABC[N:N+16]}, vb_zero_point);
      #else  // !XNN_ARCH_ARM64
        $for N in range(0, BATCH_TILE, 16):
          $if DATATYPE == "QU8":
            const int16x8_t vxa${ABC[N:N+8]} = vreinterpretq_s16_u16(vsubl_u8(vget_low_u8(va${ABC[N:N+16]}), va_zero_point));
            const int16x8_t vxa${ABC[N+8:N+16]} = vreinterpretq_s16_u16(vsubl_u8(vget_high_u8(va${ABC[N:N+16]}), va_zero_point));
            const int16x8_t vxb${ABC[N:N+8]} = vreinterpretq_s16_u16(vsubl_u8(vget_low_u8(vb${ABC[N:N+16]}), vb_zero_point));
            const int16x8_t vxb${ABC[N+8:N+16]} = vreinterpretq_s16_u16(vsubl_u8(vget_high_u8(vb${ABC[N:N+16]}), vb_zero_point));
          $else:
            const int16x8_t vxa${ABC[N:N+8]} = vsubl_s8(vget_low_s8(va${ABC[N:N+16]}), va_zero_point);
            const int16x8_t vxa${ABC[N+8:N+16]} = vsubl_s8(vget_high_s8(va${ABC[N:N+16]}), va_zero_point);
            const int16x8_t vxb${ABC[N:N+8]} = vsubl_s8(vget_low_s8(vb${ABC[N:N+16]}), vb_zero_point);
            const int16x8_t vxb${ABC[N+8:N+16]} = vsubl_s8(vget_high_s8(vb${ABC[N:N+16]}), vb_zero_point);
      #endif  // XNN_ARCH_ARM64
    $else:
      $for N in range(0, BATCH_TILE, 8):
        const ${XINT8X8_T} va${ABC[N:N+8]} = ${VLD1_X8}(input_a); input_a += 8;
        const ${XINT8X8_T} vb${ABC[N:N+8]} = ${VLD1_X8}(input_b); input_b += 8;

      $for N in range(0, BATCH_TILE, 8):
        $if DATATYPE == "QU8":
          const int16x8_t vxa${ABC[N:N+8]} = vreinterpretq_s16_u16(vsubl_u8(va${ABC[N:N+8]}, va_zero_point));
          const int16x8_t vxb${ABC[N:N+8]} = vreinterpretq_s16_u16(vsubl_u8(vb${ABC[N:N+8]}, vb_zero_point));
        $else:
          const int16x8_t vxa${ABC[N:N+8]} = vsubl_s8(va${ABC[N:N+8]}, va_zero_point);
          const int16x8_t vxb${ABC[N:N+8]} = vsubl_s8(vb${ABC[N:N+8]}, vb_zero_point);

    $for N in range(0, BATCH_TILE, 8):
      int32x4_t vacc${ABC[N:N+4]} = vmull_s16(vget_low_s16(vxa${ABC[N:N+8]}), vget_low_s16(vxb${ABC[N:N+8]}));
      int32x4_t vacc${ABC[N+4:N+8]} = vmull_s16(vget_high_s16(vxa${ABC[N:N+8]}), vget_high_s16(vxb${ABC[N:N+8]}));

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

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

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

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

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

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

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

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

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

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

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

    $for N in range(0, BATCH_TILE, 16):
      $if N + 8 < BATCH_TILE:
        ${VST1Q_X8}(output, vout${ABC[N:N+16]}); output += 16;
      $else:
        ${VST1_X8}(output, vout${ABC[N:N+8]}); output += 8;
  }
  if XNN_UNLIKELY(n != 0) {
    ${"do " if BATCH_TILE > 8 else ""}{
      $if BATCH_TILE > 8:
        const ${XINT8X8_T} va${ABC[0:8]} = ${VLD1_X8}(input_a); input_a += 8;
        const ${XINT8X8_T} vb${ABC[0:8]} = ${VLD1_X8}(input_b); input_b += 8;
      $else:
        const ${XINT8X8_T} va${ABC[0:8]} = ${VLD1_X8}(input_a);
        const ${XINT8X8_T} vb${ABC[0:8]} = ${VLD1_X8}(input_b);

      $if LD128:
        $if DATATYPE == "QU8":
          #if XNN_ARCH_ARM64
            const int16x8_t vxa${ABC[0:8]} = vreinterpretq_s16_u16(vsubl_u8(va${ABC[0:8]}, vget_low_u8(va_zero_point)));
            const int16x8_t vxb${ABC[0:8]} = vreinterpretq_s16_u16(vsubl_u8(vb${ABC[0:8]}, vget_low_u8(vb_zero_point)));
          #else  // !XNN_ARCH_ARM64
            const int16x8_t vxa${ABC[0:8]} = vreinterpretq_s16_u16(vsubl_u8(va${ABC[0:8]}, va_zero_point));
            const int16x8_t vxb${ABC[0:8]} = vreinterpretq_s16_u16(vsubl_u8(vb${ABC[0:8]}, vb_zero_point));
          #endif  // XNN_ARCH_ARM64
        $else:
          #if XNN_ARCH_ARM64
            const int16x8_t vxa${ABC[0:8]} = vsubl_s8(va${ABC[0:8]}, vget_low_s8(va_zero_point));
            const int16x8_t vxb${ABC[0:8]} = vsubl_s8(vb${ABC[0:8]}, vget_low_s8(vb_zero_point));
          #else  // !XNN_ARCH_ARM64
            const int16x8_t vxa${ABC[0:8]} = vsubl_s8(va${ABC[0:8]}, va_zero_point);
            const int16x8_t vxb${ABC[0:8]} = vsubl_s8(vb${ABC[0:8]}, vb_zero_point);
          #endif  // XNN_ARCH_ARM64
      $else:
        $if DATATYPE == "QU8":
          const int16x8_t vxa${ABC[0:8]} = vreinterpretq_s16_u16(vsubl_u8(va${ABC[0:8]}, va_zero_point));
          const int16x8_t vxb${ABC[0:8]} = vreinterpretq_s16_u16(vsubl_u8(vb${ABC[0:8]}, vb_zero_point));
        $else:
          const int16x8_t vxa${ABC[0:8]} = vsubl_s8(va${ABC[0:8]}, va_zero_point);
          const int16x8_t vxb${ABC[0:8]} = vsubl_s8(vb${ABC[0:8]}, vb_zero_point);

      int32x4_t vacc${ABC[0:4]} = vmull_s16(vget_low_s16(vxa${ABC[0:8]}), vget_low_s16(vxb${ABC[0:8]}));
      int32x4_t vacc${ABC[4:8]} = vmull_s16(vget_high_s16(vxa${ABC[0:8]}), vget_high_s16(vxb${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]});

      $if BATCH_TILE > 8:
        vout${ABC[0:8]} = ${VMAX_X8}(vout${ABC[0:8]}, ${VGET_LOW_X8}(voutput_min));
        vout${ABC[0:8]} = ${VMIN_X8}(vout${ABC[0:8]}, ${VGET_LOW_X8}(voutput_max));
        if XNN_LIKELY(n >= (8 * sizeof(${XINT8_T}))) {
          ${VST1_X8}(output, vout${ABC[0:8]}); output += 8;
          n -= 8 * sizeof(${XINT8_T});
        } else {
          if (n & (4 * sizeof(${XINT8_T}))) {
            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${ABC[0:8]}, 4);
          }
          if (n & (2 * sizeof(${XINT8_T}))) {
            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${ABC[0:8]}, 2);
          }
          if (n & (1 * sizeof(${XINT8_T}))) {
            ${VST1_LANE_X8}(output, vout${ABC[0:8]}, 0);
          }
          n = 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 (n & (4 * sizeof(${XINT8_T}))) {
          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${ABC[0:8]}, 4);
        }
        if (n & (2 * sizeof(${XINT8_T}))) {
          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${ABC[0:8]}, 2);
        }
        if (n & (1 * sizeof(${XINT8_T}))) {
          ${VST1_LANE_X8}(output, vout${ABC[0:8]}, 0);
        }
    }${" while (n != 0);" if BATCH_TILE > 8 else ""}
  }
}