src/qs8-gavgpool/multipass-wasmsimd.c.in

// 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 == "FP32"
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
#include <assert.h>

#include <wasm_simd128.h>

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


$XINT8_T = {"QS8": "int8_t", "QU8": "uint8_t"}[DATATYPE]
$WASM_X16X8_LOAD8X8 = {"QS8": "wasm_i16x8_load8x8", "QU8": "wasm_u16x8_load8x8"}[DATATYPE]
$WASM_X32X4_EXTEND_LOW_X16X8 = {"QS8": "wasm_i32x4_extend_low_i16x8", "QU8": "wasm_u32x4_extend_low_u16x8"}[DATATYPE]
$WASM_X32X4_EXTEND_HIGH_X16X8 = {"QS8": "wasm_i32x4_extend_high_i16x8", "QU8": "wasm_u32x4_extend_high_u16x8"}[DATATYPE]
$WASM_X8X16_NARROW_I16X8 = {"QS8": "wasm_i8x16_narrow_i16x8", "QU8": "wasm_u8x16_narrow_i16x8"}[DATATYPE]
$WASM_X8X16_MIN = {"QS8": "wasm_i8x16_min", "QU8": "wasm_u8x16_min"}[DATATYPE]
void xnn_${DATATYPE.lower()}_gavgpool_minmax_fp32_ukernel_${ROW_TILE}p${ROW_SUBTILE}x__wasmsimd_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 v128_t vinit_bias = wasm_v128_load64_splat(params->fp32_wasmsimd.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}) {
    $for M in range(2):
      const v128_t vxi${M}x${ABC[0:8]} = ${WASM_X16X8_LOAD8X8}(i${M});
      $for C in range(8, CHANNEL_TILE, 8):
        const v128_t vxi${M}x${ABC[C:C+8]} = ${WASM_X16X8_LOAD8X8}(i${M} + ${C});
      i${M} += ${CHANNEL_TILE};

    v128_t vacc${ABC[0:8]} = wasm_i16x8_add(vxi0x${ABC[0:8]}, vxi1x${ABC[0:8]});
    const v128_t vxi2x${ABC[0:8]} = ${WASM_X16X8_LOAD8X8}(i2);
    $for C in range(8, CHANNEL_TILE, 8):
      v128_t vacc${ABC[C:C+8]} = wasm_i16x8_add(vxi0x${ABC[C:C+8]}, vxi1x${ABC[C:C+8]});
      const v128_t vxi2x${ABC[C:C+8]} = ${WASM_X16X8_LOAD8X8}(i2 + ${C});
    i2 += ${CHANNEL_TILE};

    $for M in range(3, ROW_TILE):
      vacc${ABC[0:8]} = wasm_i16x8_add(vacc${ABC[0:8]}, vxi${M-1}x${ABC[0:8]});
      const v128_t vxi${M}x${ABC[0:8]} = ${WASM_X16X8_LOAD8X8}(i${M});
      $for C in range(8, CHANNEL_TILE, 8):
        vacc${ABC[C:C+8]} = wasm_i16x8_add(vacc${ABC[C:C+8]}, vxi${M-1}x${ABC[C:C+8]});
        const v128_t vxi${M}x${ABC[C:C+8]} = ${WASM_X16X8_LOAD8X8}(i${M} + ${C});
      i${M} += ${CHANNEL_TILE};

    $for C in range(0, CHANNEL_TILE, 8):
      vacc${ABC[C:C+8]} = wasm_i16x8_add(vacc${ABC[C:C+8]}, vxi${ROW_TILE-1}x${ABC[C:C+8]});

    $for C in range(0, CHANNEL_TILE, 8):
      const v128_t vacc${ABC[C:C+4]} = wasm_i32x4_add(vinit_bias, ${WASM_X32X4_EXTEND_LOW_X16X8}(vacc${ABC[C:C+8]}));
      const v128_t vacc${ABC[C+4:C+8]} = wasm_i32x4_add(vinit_bias, ${WASM_X32X4_EXTEND_HIGH_X16X8}(vacc${ABC[C:C+8]}));

    wasm_v128_store(b, vacc${ABC[0:4]});
    $for C in range(4, CHANNEL_TILE, 4):
      wasm_v128_store(b + ${C}, vacc${ABC[C:C+4]});
    b += ${CHANNEL_TILE};
  }
  $if CHANNEL_TILE > 16:
    if XNN_UNLIKELY(c != 0) {
      do {
        $for M in range(2):
          const v128_t vxi${M}x${ABC[0:8]} = ${WASM_X16X8_LOAD8X8}(i${M});
          i${M} += 8;

        v128_t vacc${ABC[0:8]} = wasm_i16x8_add(vxi0x${ABC[0:8]}, vxi1x${ABC[0:8]});
        const v128_t vxi2x${ABC[0:8]} = ${WASM_X16X8_LOAD8X8}(i2);
        i2 += 8;

        $for M in range(3, ROW_TILE):
          vacc${ABC[0:8]} = wasm_i16x8_add(vacc${ABC[0:8]}, vxi${M-1}x${ABC[0:8]});
          const v128_t vxi${M}x${ABC[0:8]} = ${WASM_X16X8_LOAD8X8}(i${M});
          i${M} += 8;

        vacc${ABC[0:8]} = wasm_i16x8_add(vacc${ABC[0:8]}, vxi${ROW_SUBTILE-1}x${ABC[0:8]});

        const v128_t vacc${ABC[0:4]} = wasm_i32x4_add(vinit_bias, ${WASM_X32X4_EXTEND_LOW_X16X8}(vacc${ABC[0:8]}));
        const v128_t vacc${ABC[4:8]} = wasm_i32x4_add(vinit_bias, ${WASM_X32X4_EXTEND_HIGH_X16X8}(vacc${ABC[0:8]}));

        wasm_v128_store(b, vacc${ABC[0:4]});
        wasm_v128_store(b + 4, vacc${ABC[4:8]});
        b += 8;

        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}) {
      $for M in range(2):
        const v128_t vxi${M}x${ABC[0:8]} = ${WASM_X16X8_LOAD8X8}(i${M});
        $for C in range(8, CHANNEL_TILE, 8):
          const v128_t vxi${M}x${ABC[C:C+8]} = ${WASM_X16X8_LOAD8X8}(i${M} + ${C});
        i${M} += ${CHANNEL_TILE};

      v128_t vacc${ABC[0:8]} = wasm_i16x8_add(vxi0x${ABC[0:8]}, vxi1x${ABC[0:8]});
      const v128_t vxi2x${ABC[0:8]} = ${WASM_X16X8_LOAD8X8}(i2);
      $for C in range(8, CHANNEL_TILE, 8):
        v128_t vacc${ABC[C:C+8]} = wasm_i16x8_add(vxi0x${ABC[C:C+8]}, vxi1x${ABC[C:C+8]});
        const v128_t vxi2x${ABC[C:C+8]} = ${WASM_X16X8_LOAD8X8}(i2 + ${C});
      i2 += ${CHANNEL_TILE};

      $for M in range(3, ROW_SUBTILE):
        vacc${ABC[0:8]} = wasm_i16x8_add(vacc${ABC[0:8]}, vxi${M-1}x${ABC[0:8]});
        const v128_t vxi${M}x${ABC[0:8]} = ${WASM_X16X8_LOAD8X8}(i${M});
        $for C in range(8, CHANNEL_TILE, 8):
          vacc${ABC[C:C+8]} = wasm_i16x8_add(vacc${ABC[C:C+8]}, vxi${M-1}x${ABC[C:C+8]});
          const v128_t vxi${M}x${ABC[C:C+8]} = ${WASM_X16X8_LOAD8X8}(i${M} + ${C});
        i${M} += ${CHANNEL_TILE};

      $for C in range(0, CHANNEL_TILE, 8):
        vacc${ABC[C:C+8]} = wasm_i16x8_add(vacc${ABC[C:C+8]}, vxi${ROW_SUBTILE-1}x${ABC[C:C+8]});

      v128_t vacc${ABC[0:4]} = wasm_v128_load(b);
      $for C in range(4, CHANNEL_TILE, 4):
        v128_t vacc${ABC[C:C+4]} = wasm_v128_load(b + ${C});

      $for C in range(0, CHANNEL_TILE, 8):
        vacc${ABC[C:C+4]} = wasm_i32x4_add(vacc${ABC[C:C+4]}, ${WASM_X32X4_EXTEND_LOW_X16X8}(vacc${ABC[C:C+8]}));
        vacc${ABC[C+4:C+8]} = wasm_i32x4_add(vacc${ABC[C+4:C+8]}, ${WASM_X32X4_EXTEND_HIGH_X16X8}(vacc${ABC[C:C+8]}));

      wasm_v128_store(b, vacc${ABC[0:4]});
      $for C in range(4, CHANNEL_TILE, 4):
        wasm_v128_store(b + ${C}, vacc${ABC[C:C+4]});
      b += ${CHANNEL_TILE};
    }
    $if CHANNEL_TILE > 16:
      if XNN_UNLIKELY(c != 0) {
        do {
          $for M in range(2):
            const v128_t vxi${M}x${ABC[0:8]} = ${WASM_X16X8_LOAD8X8}(i${M});
            i${M} += 8;

          v128_t vacc${ABC[0:8]} = wasm_i16x8_add(vxi0x${ABC[0:8]}, vxi1x${ABC[0:8]});
          const v128_t vxi2x${ABC[0:8]} = ${WASM_X16X8_LOAD8X8}(i2);
          i2 += 8;

          $for M in range(3, ROW_SUBTILE):
            vacc${ABC[0:8]} = wasm_i16x8_add(vacc${ABC[0:8]}, vxi${M-1}x${ABC[0:8]});
            const v128_t vxi${M}x${ABC[0:8]} = ${WASM_X16X8_LOAD8X8}(i${M});
            i${M} += 8;

          vacc${ABC[0:8]} = wasm_i16x8_add(vacc${ABC[0:8]}, vxi${ROW_SUBTILE-1}x${ABC[0:8]});

          v128_t vacc${ABC[0:4]} = wasm_v128_load(b);
          v128_t vacc${ABC[4:8]} = wasm_v128_load(b + 4);

          vacc${ABC[0:4]} = wasm_i32x4_add(vacc${ABC[0:4]}, ${WASM_X32X4_EXTEND_LOW_X16X8}(vacc${ABC[0:8]}));
          vacc${ABC[4:8]} = wasm_i32x4_add(vacc${ABC[4:8]}, ${WASM_X32X4_EXTEND_HIGH_X16X8}(vacc${ABC[0:8]}));

          wasm_v128_store(b, vacc${ABC[0:4]});
          wasm_v128_store(b + 4, vacc${ABC[4:8]});
          b += 8;

          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;
      }

  const v128_t vscale = wasm_v128_load64_splat(params->fp32_wasmsimd.scale);
  const v128_t vmagic_bias = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias);
  const v128_t vmagic_min = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_min);
  const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias_less_output_zero_point);
  const v128_t voutput_max = wasm_v128_load64_splat(params->fp32_wasmsimd.output_max);
  for (; channels >= ${CHANNEL_TILE}; channels -= ${CHANNEL_TILE}) {
    $for M in range(2):
      const v128_t vxi${M}x${ABC[0:8]} = ${WASM_X16X8_LOAD8X8}(i${M});
      $for C in range(8, CHANNEL_TILE, 8):
        const v128_t vxi${M}x${ABC[C:C+8]} = ${WASM_X16X8_LOAD8X8}(i${M} + ${C});
      i${M} += ${CHANNEL_TILE};

    v128_t vacc${ABC[0:8]} = wasm_i16x8_add(vxi0x${ABC[0:8]}, vxi1x${ABC[0:8]});
    const v128_t vxi2x${ABC[0:8]} = ${WASM_X16X8_LOAD8X8}(i2);
    $for C in range(8, CHANNEL_TILE, 8):
      v128_t vacc${ABC[C:C+8]} = wasm_i16x8_add(vxi0x${ABC[C:C+8]}, vxi1x${ABC[C:C+8]});
      const v128_t vxi2x${ABC[C:C+8]} = ${WASM_X16X8_LOAD8X8}(i2 + ${C});
    i2 += ${CHANNEL_TILE};

    $for M in range(3, ROW_SUBTILE):
      vacc${ABC[0:8]} = wasm_i16x8_add(vacc${ABC[0:8]}, vxi${M-1}x${ABC[0:8]});
      const v128_t vxi${M}x${ABC[0:8]} = ${WASM_X16X8_LOAD8X8}(i${M});
      $for C in range(8, CHANNEL_TILE, 8):
        vacc${ABC[C:C+8]} = wasm_i16x8_add(vacc${ABC[C:C+8]}, vxi${M-1}x${ABC[C:C+8]});
        const v128_t vxi${M}x${ABC[C:C+8]} = ${WASM_X16X8_LOAD8X8}(i${M} + ${C});
      i${M} += ${CHANNEL_TILE};

    $for C in range(0, CHANNEL_TILE, 8):
      vacc${ABC[C:C+8]} = wasm_i16x8_add(vacc${ABC[C:C+8]}, vxi${ROW_SUBTILE-1}x${ABC[C:C+8]});

    v128_t vacc${ABC[0:4]} = wasm_v128_load(buffer);
    $for C in range(4, CHANNEL_TILE, 4):
      v128_t vacc${ABC[C:C+4]} = wasm_v128_load(buffer + ${C});
    buffer += ${CHANNEL_TILE};

    $for C in range(0, CHANNEL_TILE, 8):
      vacc${ABC[C:C+4]} = wasm_i32x4_add(vacc${ABC[C:C+4]}, ${WASM_X32X4_EXTEND_LOW_X16X8}(vacc${ABC[C:C+8]}));
      vacc${ABC[C+4:C+8]} = wasm_i32x4_add(vacc${ABC[C+4:C+8]}, ${WASM_X32X4_EXTEND_HIGH_X16X8}(vacc${ABC[C:C+8]}));

    $for C in range(0, CHANNEL_TILE, 4):
      vacc${ABC[C:C+4]} = wasm_f32x4_convert_i32x4(vacc${ABC[C:C+4]});

    $for C in range(0, CHANNEL_TILE, 4):
      vacc${ABC[C:C+4]} = wasm_f32x4_mul(vacc${ABC[C:C+4]}, vscale);

    $for C in range(0, CHANNEL_TILE, 4):
      vacc${ABC[C:C+4]} = wasm_f32x4_add(vacc${ABC[C:C+4]}, vmagic_bias);

    $for C in range(0, CHANNEL_TILE, 4):
      vacc${ABC[C:C+4]} = wasm_i32x4_max(vacc${ABC[C:C+4]}, vmagic_min);

    $for C in range(0, CHANNEL_TILE, 4):
      vacc${ABC[C:C+4]} = wasm_i32x4_sub(vacc${ABC[C:C+4]}, vmagic_bias_less_output_zero_point);

    $for C in range(0, CHANNEL_TILE, 8):
      v128_t vout${ABC[C:C+8]} = wasm_i16x8_narrow_i32x4(vacc${ABC[C:C+4]}, vacc${ABC[C+4:C+8]});

    $for C in range(0, CHANNEL_TILE, 16):
      $if C + 8 < CHANNEL_TILE:
        v128_t vout${ABC[C:C+16]} = ${WASM_X8X16_NARROW_I16X8}(vout${ABC[C:C+8]}, vout${ABC[C+8:C+16]});
      $else:
        v128_t vout${ABC[C:C+8]}${ABC[C:C+8]} = ${WASM_X8X16_NARROW_I16X8}(vout${ABC[C:C+8]}, vout${ABC[C:C+8]});

    $for C in range(0, CHANNEL_TILE, 16):
      $if C + 8 < CHANNEL_TILE:
        vout${ABC[C:C+16]} = ${WASM_X8X16_MIN}(vout${ABC[C:C+16]}, voutput_max);
      $else:
        vout${ABC[C:C+8]}${ABC[C:C+8]} = ${WASM_X8X16_MIN}(vout${ABC[C:C+8]}${ABC[C:C+8]}, voutput_max);

    $if CHANNEL_TILE > 8:
      wasm_v128_store(output, vout${ABC[0:16]});
    $else:
      *((double*) output) = wasm_f64x2_extract_lane(vout${ABC[0:8]}${ABC[0:8]}, 0);
    $for C in range(16, CHANNEL_TILE, 16):
      $if C + 8 < CHANNEL_TILE:
        wasm_v128_store(output + ${C}, vout${ABC[C:C+16]});
      $else:
        *((double*) (output + ${C})) = wasm_f64x2_extract_lane(vout${ABC[C:C+8]}${ABC[C:C+8]}, 0);
    output += ${CHANNEL_TILE};
  }
  if XNN_UNLIKELY(channels != 0) {
    ${"do " if CHANNEL_TILE > 8 else ""}{
      $for M in range(2):
        const v128_t vxi${M}x${ABC[0:8]} = ${WASM_X16X8_LOAD8X8}(i${M});
        i${M} += 8;

      v128_t vacc${ABC[0:8]} = wasm_i16x8_add(vxi0x${ABC[0:8]}, vxi1x${ABC[0:8]});
      const v128_t vxi2x${ABC[0:8]} = ${WASM_X16X8_LOAD8X8}(i2);
      i2 += 8;

      $for M in range(3, ROW_SUBTILE):
        vacc${ABC[0:8]} = wasm_i16x8_add(vacc${ABC[0:8]}, vxi${M-1}x${ABC[0:8]});
        const v128_t vxi${M}x${ABC[0:8]} = ${WASM_X16X8_LOAD8X8}(i${M});
        i${M} += 8;

      vacc${ABC[0:8]} = wasm_i16x8_add(vacc${ABC[0:8]}, vxi${ROW_SUBTILE-1}x${ABC[0:8]});

      v128_t vacc${ABC[0:4]} = wasm_v128_load(buffer);
      v128_t vacc${ABC[4:8]} = wasm_v128_load(buffer + 4);
      buffer += 8;

      vacc${ABC[0:4]} = wasm_i32x4_add(vacc${ABC[0:4]}, ${WASM_X32X4_EXTEND_LOW_X16X8}(vacc${ABC[0:8]}));
      vacc${ABC[4:8]} = wasm_i32x4_add(vacc${ABC[4:8]}, ${WASM_X32X4_EXTEND_HIGH_X16X8}(vacc${ABC[0:8]}));

      vacc${ABC[0:4]} = wasm_f32x4_convert_i32x4(vacc${ABC[0:4]});
      vacc${ABC[4:8]} = wasm_f32x4_convert_i32x4(vacc${ABC[4:8]});

      vacc${ABC[0:4]} = wasm_f32x4_mul(vacc${ABC[0:4]}, vscale);
      vacc${ABC[4:8]} = wasm_f32x4_mul(vacc${ABC[4:8]}, vscale);

      vacc${ABC[0:4]} = wasm_f32x4_add(vacc${ABC[0:4]}, vmagic_bias);
      vacc${ABC[4:8]} = wasm_f32x4_add(vacc${ABC[4:8]}, vmagic_bias);

      vacc${ABC[0:4]} = wasm_i32x4_max(vacc${ABC[0:4]}, vmagic_min);
      vacc${ABC[4:8]} = wasm_i32x4_max(vacc${ABC[4:8]}, vmagic_min);

      vacc${ABC[0:4]} = wasm_i32x4_sub(vacc${ABC[0:4]}, vmagic_bias_less_output_zero_point);
      vacc${ABC[4:8]} = wasm_i32x4_sub(vacc${ABC[4:8]}, vmagic_bias_less_output_zero_point);

      const v128_t vout${ABC[0:8]} = wasm_i16x8_narrow_i32x4(vacc${ABC[0:4]}, vacc${ABC[4:8]});
      v128_t vout${ABC[0:8]}${ABC[0:8]} = ${WASM_X8X16_NARROW_I16X8}(vout${ABC[0:8]}, vout${ABC[0:8]});
      vout${ABC[0:8]}${ABC[0:8]} = ${WASM_X8X16_MIN}(vout${ABC[0:8]}${ABC[0:8]}, voutput_max);

      $if CHANNEL_TILE > 8:
        if XNN_LIKELY(channels >= 8) {
          *((double*) output) = wasm_f64x2_extract_lane(vout${ABC[0:8]}${ABC[0:8]}, 0);
          output += 8;
          channels -= 8;
        } else {
          if (channels & 4) {
            *((float*) output) = wasm_f32x4_extract_lane(vout${ABC[0:8]}${ABC[0:8]}, 0);
            vout${ABC[0:8]}${ABC[0:8]} = wasm_u64x2_shr(vout${ABC[0:8]}${ABC[0:8]}, 32);
            output += 4;
          }
          uint32_t vout${ABC[0:4]} = wasm_i32x4_extract_lane(vout${ABC[0:8]}${ABC[0:8]}, 0);
          if (channels & 2) {
            *((uint16_t*) output) = (uint16_t) vout${ABC[0:4]};
            vout${ABC[0:4]} >>= 16;
            output += 2;
          }
          if (channels & 1) {
            *output = (${XINT8_T}) vout${ABC[0:4]};
            output += 1;
          }
          channels = 0;
        }
      $else:
        if (channels & 4) {
          *((float*) output) = wasm_f32x4_extract_lane(vout${ABC[0:8]}${ABC[0:8]}, 0);
          vout${ABC[0:8]}${ABC[0:8]} = wasm_u64x2_shr(vout${ABC[0:8]}${ABC[0:8]}, 32);
          output += 4;
        }
        uint32_t vout${ABC[0:4]} = wasm_i32x4_extract_lane(vout${ABC[0:8]}${ABC[0:8]}, 0);
        if (channels & 2) {
          *((uint16_t*) output) = (uint16_t) vout${ABC[0:4]};
          vout${ABC[0:4]} >>= 16;
          output += 2;
        }
        if (channels & 1) {
          *output = (${XINT8_T}) vout${ABC[0:4]};
        }
    }${" while (channels != 0);" if CHANNEL_TILE > 8 else ""}
  }
}