// 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 CHANNEL_TILE % 8 == 0
$assert CHANNEL_TILE >= 8
$assert ROW_TILE >= 2
$assert ROW_SUBTILE >= 2
$assert ROW_SUBTILE <= ROW_TILE
$assert ACCUMULATORS >= 1
$assert ROW_TILE >= ACCUMULATORS * 2
$assert ROW_SUBTILE >= ACCUMULATORS * 2
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
#include <assert.h>

#include <wasm_simd128.h>

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


void xnn_qs8_gavgpool_minmax_ukernel_${ROW_TILE}p${ROW_SUBTILE}x__wasmsimd_c${CHANNEL_TILE}${"" if ACCUMULATORS == 1 else "_acc%d" % ACCUMULATORS}(
    size_t rows,
    size_t channels,
    const int8_t* input,
    size_t input_stride,
    const int8_t* zero,
    int32_t* buffer,
    int8_t* output,
    const union xnn_qs8_avgpool_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
{
  assert(rows > ${ROW_TILE});
  assert(channels != 0);

  const int8_t* i0 = input;
  $for M in range(1, ROW_TILE):
    const int8_t* i${M} = (const int8_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});
  $else:
    const size_t input_increment = ${ROW_TILE} * input_stride - round_up_po2(channels, 8);

  const v128_t vbias = wasm_v128_load(params->wasmsimd.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}) {
    $for M in range(ROW_TILE):
      const v128_t vxi${M}x${ABC[0:8]} = wasm_i16x8_load_8x8(i${M});
      $for C in range(8, CHANNEL_TILE, 8):
        const v128_t vxi${M}x${ABC[C:C+8]} = wasm_i16x8_load_8x8(i${M} + ${C});
      i${M} += ${CHANNEL_TILE};

    $for A in range(ACCUMULATORS):
      $for C in range(0, CHANNEL_TILE, 8):
        v128_t vacc${A}x${ABC[C:C+8]} = wasm_i16x8_add(vxi${A*2}x${ABC[C:C+8]}, vxi${A*2+1}x${ABC[C:C+8]});

    $for M in range(ACCUMULATORS * 2, ROW_TILE):
      $for C in range(0, CHANNEL_TILE, 8):
        vacc${M % ACCUMULATORS}x${ABC[C:C+8]} = wasm_i16x8_add(vacc${M % ACCUMULATORS}x${ABC[C:C+8]}, vxi${M}x${ABC[C:C+8]});

    $if ACCUMULATORS > 1:
      // Add up all accumulators to vacc0x${ABC[0:CHANNEL_TILE]}
      $ACC_SLICE = 1
      $while ACC_SLICE < ACCUMULATORS:
        $for A in range(0, ACCUMULATORS, ACC_SLICE * 2):
          $if A + ACC_SLICE < ACCUMULATORS:
            $for C in range(0, CHANNEL_TILE, 8):
              vacc${A}x${ABC[C:C+8]} = wasm_i16x8_add(vacc${A}x${ABC[C:C+8]}, vacc${A + ACC_SLICE}x${ABC[C:C+8]});
        $ACC_SLICE *= 2

    $for C in range(0, CHANNEL_TILE, 8):
      const v128_t vacc${ABC[C:C+4]} = wasm_i32x4_add(vbias, wasm_i32x4_widen_low_i16x8(vacc0x${ABC[C:C+8]}));
      const v128_t vacc${ABC[C+4:C+8]} = wasm_i32x4_add(vbias, wasm_i32x4_widen_high_i16x8(vacc0x${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 {
        $for M in range(ROW_TILE):
          const v128_t vxi${M}x${ABC[0:8]} = wasm_i16x8_load_8x8(i${M});
          i${M} += 8;

        $for A in range(ACCUMULATORS):
          v128_t vacc${A}x${ABC[0:8]} = wasm_i16x8_add(vxi${A*2}x${ABC[0:8]}, vxi${A*2+1}x${ABC[0:8]});

        $for M in range(ACCUMULATORS * 2, ROW_TILE):
          vacc${M % ACCUMULATORS}x${ABC[0:8]} = wasm_i16x8_add(vacc${M % ACCUMULATORS}x${ABC[0:8]}, vxi${M}x${ABC[0:8]});

        $if ACCUMULATORS > 1:
          // Add up all accumulators to vacc0x${ABC[0:8]}
          $ACC_SLICE = 1
          $while ACC_SLICE < ACCUMULATORS:
            $for A in range(0, ACCUMULATORS, ACC_SLICE * 2):
              $if A + ACC_SLICE < ACCUMULATORS:
                vacc${A}x${ABC[0:8]} = wasm_i16x8_add(vacc${A}x${ABC[0:8]}, vacc${A + ACC_SLICE}x${ABC[0:8]});
            $ACC_SLICE *= 2

        const v128_t vacc${ABC[0:4]} = wasm_i32x4_add(vbias, wasm_i32x4_widen_low_i16x8(vacc0x${ABC[0:8]}));
        const v128_t vacc${ABC[4:8]} = wasm_i32x4_add(vbias, wasm_i32x4_widen_high_i16x8(vacc0x${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 int8_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}) {
      $for M in range(ROW_SUBTILE):
        const v128_t vxi${M}x${ABC[0:8]} = wasm_i16x8_load_8x8(i${M});
        $for C in range(8, CHANNEL_TILE, 8):
          const v128_t vxi${M}x${ABC[C:C+8]} = wasm_i16x8_load_8x8(i${M} + ${C});
        i${M} += ${CHANNEL_TILE};

      $for A in range(ACCUMULATORS):
        $for C in range(0, CHANNEL_TILE, 8):
          v128_t vacc${A}x${ABC[C:C+8]} = wasm_i16x8_add(vxi${A*2}x${ABC[C:C+8]}, vxi${A*2+1}x${ABC[C:C+8]});

      $for M in range(ACCUMULATORS * 2, ROW_SUBTILE):
        $for C in range(0, CHANNEL_TILE, 8):
          vacc${M % ACCUMULATORS}x${ABC[C:C+8]} = wasm_i16x8_add(vacc${M % ACCUMULATORS}x${ABC[C:C+8]}, vxi${M}x${ABC[C:C+8]});

      $if ACCUMULATORS > 1:
        // Add up all accumulators to vacc0x${ABC[0:CHANNEL_TILE]}
        $ACC_SLICE = 1
        $while ACC_SLICE < ACCUMULATORS:
          $for A in range(0, ACCUMULATORS, ACC_SLICE * 2):
            $if A + ACC_SLICE < ACCUMULATORS:
              $for C in range(0, CHANNEL_TILE, 8):
                vacc${A}x${ABC[C:C+8]} = wasm_i16x8_add(vacc${A}x${ABC[C:C+8]}, vacc${A + ACC_SLICE}x${ABC[C:C+8]});
          $ACC_SLICE *= 2

      $for C in range(0, CHANNEL_TILE, 8):
        const v128_t vacc${ABC[C:C+4]} = wasm_i32x4_add(wasm_i32x4_widen_low_i16x8(vacc0x${ABC[C:C+8]}), wasm_v128_load(b + ${C}));
        const v128_t vacc${ABC[C+4:C+8]} = wasm_i32x4_add(wasm_i32x4_widen_high_i16x8(vacc0x${ABC[C:C+8]}), wasm_v128_load(b + ${C+4}));

      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 {
          $for M in range(ROW_SUBTILE):
            const v128_t vxi${M}x${ABC[0:8]} = wasm_i16x8_load_8x8(i${M});
            i${M} += 8;

          $for A in range(ACCUMULATORS):
            v128_t vacc${A}x${ABC[0:8]} = wasm_i16x8_add(vxi${A*2}x${ABC[0:8]}, vxi${A*2+1}x${ABC[0:8]});

          $for M in range(ACCUMULATORS * 2, ROW_SUBTILE):
            vacc${M % ACCUMULATORS}x${ABC[0:8]} = wasm_i16x8_add(vacc${M % ACCUMULATORS}x${ABC[0:8]}, vxi${M}x${ABC[0:8]});

          $if ACCUMULATORS > 1:
            // Add up all accumulators to vacc0x${ABC[0:8]}
            $ACC_SLICE = 1
            $while ACC_SLICE < ACCUMULATORS:
              $for A in range(0, ACCUMULATORS, ACC_SLICE * 2):
                $if A + ACC_SLICE < ACCUMULATORS:
                  vacc${A}x${ABC[0:8]} = wasm_i16x8_add(vacc${A}x${ABC[0:8]}, vacc${A + ACC_SLICE}x${ABC[0:8]});
              $ACC_SLICE *= 2

          const v128_t vacc${ABC[0:4]} = wasm_i32x4_add(wasm_i32x4_widen_low_i16x8(vacc0x${ABC[0:8]}), wasm_v128_load(b));
          const v128_t vacc${ABC[4:8]} = wasm_i32x4_add(wasm_i32x4_widen_high_i16x8(vacc0x${ABC[0:8]}), wasm_v128_load(b + 4));

          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 int8_t*) ((uintptr_t) i${ROW_TILE - ROW_SUBTILE} + input_increment);
  $for M in range(1, ROW_SUBTILE):
    i${M} = (const int8_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 vmultiplier = wasm_v128_load(params->wasmsimd.multiplier);
  const v128_t vrounding = wasm_v128_load(params->wasmsimd.rounding);
  const int32_t vshift = params->wasmsimd.shift;
  const v128_t vzero = wasm_f64x2_splat(0.0);
  while (channels >= ${CHANNEL_TILE}) {
    $for M in range(ROW_SUBTILE):
      const v128_t vxi${M}x${ABC[0:8]} = wasm_i16x8_load_8x8(i${M});
      $for C in range(8, CHANNEL_TILE, 8):
        const v128_t vxi${M}x${ABC[C:C+8]} = wasm_i16x8_load_8x8(i${M} + ${C});
      i${M} += ${CHANNEL_TILE};

    $for A in range(ACCUMULATORS):
      $for C in range(0, CHANNEL_TILE, 8):
        v128_t vacc${A}x${ABC[C:C+8]} = wasm_i16x8_add(vxi${A*2}x${ABC[C:C+8]}, vxi${A*2+1}x${ABC[C:C+8]});

    $for M in range(ACCUMULATORS * 2, ROW_SUBTILE):
      $for C in range(0, CHANNEL_TILE, 8):
        vacc${M % ACCUMULATORS}x${ABC[C:C+8]} = wasm_i16x8_add(vacc${M % ACCUMULATORS}x${ABC[C:C+8]}, vxi${M}x${ABC[C:C+8]});

    $if ACCUMULATORS > 1:
      // Add up all accumulators to vacc0x${ABC[0:CHANNEL_TILE]}
      $ACC_SLICE = 1
      $while ACC_SLICE < ACCUMULATORS:
        $for A in range(0, ACCUMULATORS, ACC_SLICE * 2):
          $if A + ACC_SLICE < ACCUMULATORS:
            $for C in range(0, CHANNEL_TILE, 8):
              vacc${A}x${ABC[C:C+8]} = wasm_i16x8_add(vacc${A}x${ABC[C:C+8]}, vacc${A + ACC_SLICE}x${ABC[C:C+8]});
        $ACC_SLICE *= 2

    $for C in range(0, CHANNEL_TILE, 8):
      const v128_t vacc${ABC[C:C+4]} = wasm_i32x4_add(wasm_i32x4_widen_low_i16x8(vacc0x${ABC[C:C+8]}), wasm_v128_load(buffer + ${C}));
      const v128_t vacc${ABC[C+4:C+8]} = wasm_i32x4_add(wasm_i32x4_widen_high_i16x8(vacc0x${ABC[C:C+8]}), wasm_v128_load(buffer + ${C+4}));
    buffer += ${CHANNEL_TILE};

    $for C in range(0, CHANNEL_TILE, 4):
      const v128_t vabsacc${ABC[C:C+4]} = wasm_i32x4_abs(vacc${ABC[C:C+4]});

    $for C in range(0, CHANNEL_TILE, 4):
      const v128_t vsgnacc${ABC[C:C+4]} = wasm_i32x4_gt(vabsacc${ABC[C:C+4]}, vacc${ABC[C:C+4]});

    $for C in range(0, CHANNEL_TILE, 4):
      const v128_t vabsacc${ABC[C:C+2]} = wasm_v32x4_shuffle(vabsacc${ABC[C:C+4]}, vzero, 0, 4, 1, 5);
      const v128_t vabsacc${ABC[C+2:C+4]} = wasm_v32x4_shuffle(vabsacc${ABC[C:C+4]}, vzero, 2, 6, 3, 7);

    $for C in range(0, CHANNEL_TILE, 2):
      const v128_t vabsprod${ABC[C:C+2]} = wasm_i64x2_mul(vabsacc${ABC[C:C+2]}, vmultiplier);

    $for C in range(0, CHANNEL_TILE, 2):
      const v128_t vabsout${ABC[C:C+2]} = wasm_u64x2_shr(wasm_i64x2_add(vabsprod${ABC[C:C+2]}, vrounding), vshift);

    $for C in range(0, CHANNEL_TILE, 4):
      const v128_t vabsout${ABC[C:C+4]} = wasm_v32x4_shuffle(vabsout${ABC[C:C+2]}, vabsout${ABC[C+2:C+4]}, 0, 2, 4, 6);

    $for C in range(0, CHANNEL_TILE, 4):
      const v128_t vout${ABC[C:C+4]} = wasm_i32x4_sub(wasm_v128_xor(vabsout${ABC[C:C+4]}, vsgnacc${ABC[C:C+4]}), vsgnacc${ABC[C:C+4]});

    const v128_t voutput_zero_point = wasm_v128_load(params->wasmsimd.output_zero_point);
    $for C in range(0, CHANNEL_TILE, 8):
      v128_t vout${ABC[C:C+8]} = wasm_i16x8_add_saturate(wasm_i16x8_narrow_i32x4(vout${ABC[C:C+4]}, vout${ABC[C+4:C+8]}), voutput_zero_point);

    const v128_t voutput_min = wasm_v128_load(params->wasmsimd.output_min);
    const v128_t voutput_max = wasm_v128_load(params->wasmsimd.output_max);
    $for C in range(0, CHANNEL_TILE, 16):
      $if C + 8 < CHANNEL_TILE:
        v128_t vout${ABC[C:C+16]} = wasm_i8x16_min(wasm_i8x16_max(wasm_i8x16_narrow_i16x8(vout${ABC[C:C+8]}, vout${ABC[C+8:C+16]}), voutput_min), voutput_max);
      $else:
        v128_t vout${ABC[C:C+8]}${ABC[C:C+8]} = wasm_i8x16_min(wasm_i8x16_max(wasm_i8x16_narrow_i16x8(vout${ABC[C:C+8]}, vout${ABC[C:C+8]}), voutput_min), 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};

    channels -= ${CHANNEL_TILE};
  }
  if XNN_UNLIKELY(channels != 0) {
    ${"do " if CHANNEL_TILE > 8 else ""}{
      $for M in range(ROW_SUBTILE):
        const v128_t vxi${M}x${ABC[0:8]} = wasm_i16x8_load_8x8(i${M});
        i${M} += 8;

      $for A in range(ACCUMULATORS):
        v128_t vacc${A}x${ABC[0:8]} = wasm_i16x8_add(vxi${A*2}x${ABC[0:8]}, vxi${A*2+1}x${ABC[0:8]});

      $for M in range(ACCUMULATORS * 2, ROW_SUBTILE):
        vacc${M % ACCUMULATORS}x${ABC[0:8]} = wasm_i16x8_add(vacc${M % ACCUMULATORS}x${ABC[0:8]}, vxi${M}x${ABC[0:8]});

      $if ACCUMULATORS > 1:
        // Add up all accumulators to vacc0x${ABC[0:8]}
        $ACC_SLICE = 1
        $while ACC_SLICE < ACCUMULATORS:
          $for A in range(0, ACCUMULATORS, ACC_SLICE * 2):
            $if A + ACC_SLICE < ACCUMULATORS:
              vacc${A}x${ABC[0:8]} = wasm_i16x8_add(vacc${A}x${ABC[0:8]}, vacc${A + ACC_SLICE}x${ABC[0:8]});
          $ACC_SLICE *= 2

      const v128_t vacc${ABC[0:4]} = wasm_i32x4_add(wasm_i32x4_widen_low_i16x8(vacc0x${ABC[0:8]}), wasm_v128_load(buffer));
      const v128_t vacc${ABC[4:8]} = wasm_i32x4_add(wasm_i32x4_widen_high_i16x8(vacc0x${ABC[0:8]}), wasm_v128_load(buffer + 4));
      buffer += 8;

      const v128_t vabsacc${ABC[0:4]} = wasm_i32x4_abs(vacc${ABC[0:4]});
      const v128_t vabsacc${ABC[4:8]} = wasm_i32x4_abs(vacc${ABC[4:8]});

      const v128_t vsgnacc${ABC[0:4]} = wasm_i32x4_gt(vabsacc${ABC[0:4]}, vacc${ABC[0:4]});
      const v128_t vsgnacc${ABC[4:8]} = wasm_i32x4_gt(vabsacc${ABC[4:8]}, vacc${ABC[4:8]});

      const v128_t vabsacc${ABC[0:2]} = wasm_v32x4_shuffle(vabsacc${ABC[0:4]}, vzero, 0, 4, 1, 5);
      const v128_t vabsacc${ABC[2:4]} = wasm_v32x4_shuffle(vabsacc${ABC[0:4]}, vzero, 2, 6, 3, 7);
      const v128_t vabsacc${ABC[4:6]} = wasm_v32x4_shuffle(vabsacc${ABC[4:8]}, vzero, 0, 4, 1, 5);
      const v128_t vabsacc${ABC[6:8]} = wasm_v32x4_shuffle(vabsacc${ABC[4:8]}, vzero, 2, 6, 3, 7);

      const v128_t vabsprod${ABC[0:2]} = wasm_i64x2_mul(vabsacc${ABC[0:2]}, vmultiplier);
      const v128_t vabsprod${ABC[2:4]} = wasm_i64x2_mul(vabsacc${ABC[2:4]}, vmultiplier);
      const v128_t vabsprod${ABC[4:6]} = wasm_i64x2_mul(vabsacc${ABC[4:6]}, vmultiplier);
      const v128_t vabsprod${ABC[6:8]} = wasm_i64x2_mul(vabsacc${ABC[6:8]}, vmultiplier);

      const v128_t vabsout${ABC[0:2]} = wasm_u64x2_shr(wasm_i64x2_add(vabsprod${ABC[0:2]}, vrounding), vshift);
      const v128_t vabsout${ABC[2:4]} = wasm_u64x2_shr(wasm_i64x2_add(vabsprod${ABC[2:4]}, vrounding), vshift);
      const v128_t vabsout${ABC[4:6]} = wasm_u64x2_shr(wasm_i64x2_add(vabsprod${ABC[4:6]}, vrounding), vshift);
      const v128_t vabsout${ABC[6:8]} = wasm_u64x2_shr(wasm_i64x2_add(vabsprod${ABC[6:8]}, vrounding), vshift);

      const v128_t vabsout${ABC[0:4]} = wasm_v32x4_shuffle(vabsout${ABC[0:2]}, vabsout${ABC[2:4]}, 0, 2, 4, 6);
      const v128_t vabsout${ABC[4:8]} = wasm_v32x4_shuffle(vabsout${ABC[4:6]}, vabsout${ABC[6:8]}, 0, 2, 4, 6);

      const v128_t vout${ABC[0:4]} = wasm_i32x4_sub(wasm_v128_xor(vabsout${ABC[0:4]}, vsgnacc${ABC[0:4]}), vsgnacc${ABC[0:4]});
      const v128_t vout${ABC[4:8]} = wasm_i32x4_sub(wasm_v128_xor(vabsout${ABC[4:8]}, vsgnacc${ABC[4:8]}), vsgnacc${ABC[4:8]});

      const v128_t voutput_zero_point = wasm_v128_load(params->wasmsimd.output_zero_point);
      const v128_t vout${ABC[0:8]} = wasm_i16x8_add_saturate(wasm_i16x8_narrow_i32x4(vout${ABC[0:4]}, vout${ABC[4:8]}), voutput_zero_point);

      const v128_t voutput_min = wasm_v128_load(params->wasmsimd.output_min);
      const v128_t voutput_max = wasm_v128_load(params->wasmsimd.output_max);
      v128_t vout${ABC[0:8]}${ABC[0:8]} = wasm_i8x16_min(wasm_i8x16_max(wasm_i8x16_narrow_i16x8(vout${ABC[0:8]}, vout${ABC[0:8]}), voutput_min), 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]}${ABC[0:8]}, 32);
            output += 4;
          }
          if (channels & 2) {
            *((uint16_t*) output) = (uint16_t) wasm_i16x8_extract_lane(vout${ABC[0:8]}${ABC[0:8]}, 0);
            vout${ABC[0:8]}${ABC[0:8]} = wasm_u32x4_shr(vout${ABC[0:8]}${ABC[0:8]}, 16);
            output += 2;
          }
          if (channels & 1) {
            *output = (int8_t) wasm_i8x16_extract_lane(vout${ABC[0:8]}${ABC[0:8]}, 0);
            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]}${ABC[0:8]}, 32);
          output += 4;
        }
        if (channels & 2) {
          *((uint16_t*) output) = (uint16_t) wasm_i16x8_extract_lane(vout${ABC[0:8]}${ABC[0:8]}, 0);
          vout${ABC[0:8]}${ABC[0:8]} = wasm_u32x4_shr(vout${ABC[0:8]}${ABC[0:8]}, 16);
          output += 2;
        }
        if (channels & 1) {
          *output = (int8_t) wasm_i8x16_extract_lane(vout${ABC[0:8]}${ABC[0:8]}, 0);
        }
    }${" while (channels != 0);" if CHANNEL_TILE > 8 else ""}
  }
}