/*
 * Copyright (c) 2021 Arm Limited.
 *
 * SPDX-License-Identifier: MIT
 *
 * Permission is hereby granted, free of charge, to any person obtaining a copy
 * of this software and associated documentation files (the "Software"), to
 * deal in the Software without restriction, including without limitation the
 * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
 * sell copies of the Software, and to permit persons to whom the Software is
 * furnished to do so, subject to the following conditions:
 *
 * The above copyright notice and this permission notice shall be included in all
 * copies or substantial portions of the Software.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
 * SOFTWARE.
 */

#include "pooling.hpp"
#include <cstddef>
#include <cstdint>
#include <cstring>
#include <cmath>


#if defined(__aarch64__)

namespace arm_conv {
namespace pooling {

namespace {
  struct RescaleParams
  {
    int32_t multiplier, shift;
  };

  constexpr RescaleParams rescale_params[8] = {
    {0x40000000, -0},  // 1/2
    {0x55555555, -1},  // 1/3
    {0x40000000, -1},  // 1/4
    {0x66666666, -2},  // 1/5
    {0x55555555, -2},  // 1/6
    {0x49249249, -2},  // 1/7
    {0x40000000, -2},  // 1/8
    {0x71c71c72, -3},  // 1/9
  };
}

void a64_s8q_nhwc_avg_generic_depthfirst_impl(
  const uint64_t window_cells,
  const uint64_t n_valid_cells,
  uint64_t n_channels,
  const int8_t *const *const inptrs,
  int8_t *outptr,
  const Requantize32 &qp
)
{
  if (n_valid_cells == 1 && window_cells == 1)
  {
    // In this case, simply copy from the input to the output
    std::memcpy(outptr, *inptrs, n_channels);
    return;
  }

  // Compute (or look up) the rescale values
  int32_t shift_value = 0, rescale_value = 0;
  if (2 <= window_cells && window_cells <= 9)
  {
    auto &params = rescale_params[window_cells - 2];
    rescale_value = params.multiplier;
    shift_value = params.shift;
  }
  else
  {
    auto f_rescale_value = 1.0f / static_cast<float>(window_cells);

    shift_value = 0;
    while (f_rescale_value < 0.5f)
    {
      shift_value--;
      f_rescale_value *= 2.0f;
    }

    rescale_value = static_cast<int32_t>(round(f_rescale_value * static_cast<float>(1ll << 31)));
    if (static_cast<int64_t>(rescale_value) == (1ll << 31))
    {
      shift_value++;
      rescale_value >>= 1;
    }
  }

  // Combine together the rescale value for the requantization and the scaling
  // factor for the average pool.
  const int32_t shift = qp.per_layer_left_shift - qp.per_layer_right_shift + shift_value;
  const int32_t left_shift = shift > 0 ? shift : 0;
  const int32_t right_shift = shift <= 0 ? shift : 0;

  int32_t combined_rescale_value = 0;
  __asm__ __volatile__ (
      "mov v16.s[0], %w[per_layer_mul]\n"
      "mov v17.s[0], %w[rescale_value]\n"
      "sqrdmulh s18, s16, s17\n"
      "mov %w[combined_rescale_value], v18.s[0]\n"
    : [combined_rescale_value] "=r" (combined_rescale_value)
    : [per_layer_mul] "r" (qp.per_layer_mul), [rescale_value] "r" (rescale_value)
    : "v16", "v17", "v18"
  );

  __asm__ __volatile__(
    "mov x26, #0x0\n"
    "mov x25, #0x10\n" // cntb _, ALL, #1
    "mov x24, #0x20\n" // cntb _, ALL, #2
    "mov x23, #0x30\n" // cntb _, ALL, #3
    "cmp %x[n_channels], #0x40\n"
    "blt 7f\n"
    "1:"  // 4-vectors of channels
    "movi v15.4s, #0x0\n"
    "mov x19, %x[inptrs]\n"
    "movi v14.4s, #0x0\n"
    "lsr x22, %x[n_valid_cells], #0x1\n"
    "movi v13.4s, #0x0\n"
    "movi v12.4s, #0x0\n"
    "movi v11.4s, #0x0\n"
    "movi v10.4s, #0x0\n"
    "movi v9.4s, #0x0\n"
    "movi v8.4s, #0x0\n"
    "movi v7.4s, #0x0\n"
    "movi v6.4s, #0x0\n"
    "movi v5.4s, #0x0\n"
    "movi v4.4s, #0x0\n"
    "movi v3.4s, #0x0\n"
    "movi v2.4s, #0x0\n"
    "movi v1.4s, #0x0\n"
    "movi v0.4s, #0x0\n"
    "cbz x22, 4f\n"
    "ldp x21, x20, [x19, #0x0]\n"
    "add x19, x19, #0x10\n"
    "subs x22, x22, #0x1\n"
    "ldr q31, [x21, x26]\n"
    "ldr q30, [x20, x26]\n"
    "ldr q29, [x21, x25]\n"
    "ldr q28, [x20, x25]\n"
    "ldr q27, [x21, x24]\n"
    "ldr q26, [x20, x24]\n"
    "ldr q25, [x21, x23]\n"
    "ldr q24, [x20, x23]\n"
    "beq 3f\n"
    "2:"  // 4-vectors of channels: 2 inputs loop
    "saddl v23.8h, v31.8b, v30.8b\n"
    "ldp x21, x20, [x19, #0x0]\n"
    "add x19, x19, #0x10\n"
    "saddl2 v22.8h, v31.16b, v30.16b\n"
    "subs x22, x22, #0x1\n"
    "saddl v21.8h, v29.8b, v28.8b\n"
    "ldr q31, [x21, x26]\n"
    "saddl2 v20.8h, v29.16b, v28.16b\n"
    "saddl v19.8h, v27.8b, v26.8b\n"
    "ldr q30, [x20, x26]\n"
    "saddl2 v18.8h, v27.16b, v26.16b\n"
    "ldr q29, [x21, x25]\n"
    "saddl v17.8h, v25.8b, v24.8b\n"
    "ldr q28, [x20, x25]\n"
    "saddl2 v16.8h, v25.16b, v24.16b\n"
    "ldr q27, [x21, x24]\n"
    "saddw v15.4s, v15.4s, v23.4h\n"
    "ldr q26, [x20, x24]\n"
    "saddw2 v14.4s, v14.4s, v23.8h\n"
    "ldr q25, [x21, x23]\n"
    "saddw v13.4s, v13.4s, v22.4h\n"
    "ldr q24, [x20, x23]\n"
    "saddw2 v12.4s, v12.4s, v22.8h\n"
    "saddw v11.4s, v11.4s, v21.4h\n"
    "saddw2 v10.4s, v10.4s, v21.8h\n"
    "saddw v9.4s, v9.4s, v20.4h\n"
    "saddw2 v8.4s, v8.4s, v20.8h\n"
    "saddw v7.4s, v7.4s, v19.4h\n"
    "saddw2 v6.4s, v6.4s, v19.8h\n"
    "saddw v5.4s, v5.4s, v18.4h\n"
    "saddw2 v4.4s, v4.4s, v18.8h\n"
    "saddw v3.4s, v3.4s, v17.4h\n"
    "saddw2 v2.4s, v2.4s, v17.8h\n"
    "saddw v1.4s, v1.4s, v16.4h\n"
    "saddw2 v0.4s, v0.4s, v16.8h\n"
    "bgt 2b\n"
    "3:"  // 4-vectors of channels: 2 inputs tail
    "saddl v23.8h, v31.8b, v30.8b\n"
    "saddl2 v22.8h, v31.16b, v30.16b\n"
    "saddl v21.8h, v29.8b, v28.8b\n"
    "saddl2 v20.8h, v29.16b, v28.16b\n"
    "saddl v19.8h, v27.8b, v26.8b\n"
    "saddl2 v18.8h, v27.16b, v26.16b\n"
    "saddl v17.8h, v25.8b, v24.8b\n"
    "saddl2 v16.8h, v25.16b, v24.16b\n"
    "saddw v15.4s, v15.4s, v23.4h\n"
    "saddw2 v14.4s, v14.4s, v23.8h\n"
    "saddw v13.4s, v13.4s, v22.4h\n"
    "saddw2 v12.4s, v12.4s, v22.8h\n"
    "saddw v11.4s, v11.4s, v21.4h\n"
    "saddw2 v10.4s, v10.4s, v21.8h\n"
    "saddw v9.4s, v9.4s, v20.4h\n"
    "saddw2 v8.4s, v8.4s, v20.8h\n"
    "saddw v7.4s, v7.4s, v19.4h\n"
    "saddw2 v6.4s, v6.4s, v19.8h\n"
    "saddw v5.4s, v5.4s, v18.4h\n"
    "saddw2 v4.4s, v4.4s, v18.8h\n"
    "saddw v3.4s, v3.4s, v17.4h\n"
    "saddw2 v2.4s, v2.4s, v17.8h\n"
    "saddw v1.4s, v1.4s, v16.4h\n"
    "saddw2 v0.4s, v0.4s, v16.8h\n"
    "4:"  // 4-vectors of channels: After loop
    "ands x20, %x[n_valid_cells], #0x1\n"
    "beq 6f\n"
    "5:"  // 4-vectors of channels: Single input loop
    "ldr x21, [x19], #0x8\n"
    "subs x20, x20, #0x1\n"
    "ldr q31, [x21, x26]\n"
    "sxtl v16.8h, v31.8b\n"
    "ldr q29, [x21, x25]\n"
    "sxtl2 v22.8h, v31.16b\n"
    "ldr q27, [x21, x24]\n"
    "ldr q25, [x21, x23]\n"
    "saddw v15.4s, v15.4s, v16.4h\n"
    "saddw2 v14.4s, v14.4s, v16.8h\n"
    "sxtl v21.8h, v29.8b\n"
    "sxtl2 v20.8h, v29.16b\n"
    "sxtl v19.8h, v27.8b\n"
    "sxtl2 v18.8h, v27.16b\n"
    "sxtl v17.8h, v25.8b\n"
    "sxtl2 v16.8h, v25.16b\n"
    "saddw v13.4s, v13.4s, v22.4h\n"
    "saddw2 v12.4s, v12.4s, v22.8h\n"
    "saddw v11.4s, v11.4s, v21.4h\n"
    "saddw2 v10.4s, v10.4s, v21.8h\n"
    "saddw v9.4s, v9.4s, v20.4h\n"
    "saddw2 v8.4s, v8.4s, v20.8h\n"
    "saddw v7.4s, v7.4s, v19.4h\n"
    "saddw2 v6.4s, v6.4s, v19.8h\n"
    "saddw v5.4s, v5.4s, v18.4h\n"
    "saddw2 v4.4s, v4.4s, v18.8h\n"
    "saddw v3.4s, v3.4s, v17.4h\n"
    "saddw2 v2.4s, v2.4s, v17.8h\n"
    "saddw v1.4s, v1.4s, v16.4h\n"
    "saddw2 v0.4s, v0.4s, v16.8h\n"
    "bgt 5b\n"
    "6:"  // 4-vectors of channels: Single input loop: End
    "movi v20.4s, #0x7f\n"
    "ld1r { v19.4s }, [%x[combined_rescale_value]]\n"
    "sub %x[n_channels], %x[n_channels], #0x40\n"
    "ld1r { v18.4s }, [%x[left_shift]]\n"
    "srshl v15.4s, v15.4s, v18.4s\n"
    "ld1r { v17.4s }, [%x[right_shift]]\n"
    "not v16.16b, v20.16b\n"
    "srshl v14.4s, v14.4s, v18.4s\n"
    "cmp %x[n_channels], #0x40\n"
    "srshl v13.4s, v13.4s, v18.4s\n"
    "srshl v12.4s, v12.4s, v18.4s\n"
    "srshl v11.4s, v11.4s, v18.4s\n"
    "sqrdmulh v15.4s, v15.4s, v19.4s\n"
    "sqrdmulh v14.4s, v14.4s, v19.4s\n"
    "sqrdmulh v13.4s, v13.4s, v19.4s\n"
    "sqrdmulh v12.4s, v12.4s, v19.4s\n"
    "srshl v15.4s, v15.4s, v17.4s\n"
    "srshl v14.4s, v14.4s, v17.4s\n"
    "srshl v13.4s, v13.4s, v17.4s\n"
    "srshl v12.4s, v12.4s, v17.4s\n"
    "sqrdmulh v11.4s, v11.4s, v19.4s\n"
    "srshl v10.4s, v10.4s, v18.4s\n"
    "srshl v9.4s, v9.4s, v18.4s\n"
    "srshl v8.4s, v8.4s, v18.4s\n"
    "srshl v11.4s, v11.4s, v17.4s\n"
    "sqrdmulh v10.4s, v10.4s, v19.4s\n"
    "sqrdmulh v9.4s, v9.4s, v19.4s\n"
    "sqrdmulh v8.4s, v8.4s, v19.4s\n"
    "srshl v7.4s, v7.4s, v18.4s\n"
    "srshl v10.4s, v10.4s, v17.4s\n"
    "srshl v9.4s, v9.4s, v17.4s\n"
    "srshl v8.4s, v8.4s, v17.4s\n"
    "sqrdmulh v7.4s, v7.4s, v19.4s\n"
    "srshl v6.4s, v6.4s, v18.4s\n"
    "srshl v5.4s, v5.4s, v18.4s\n"
    "srshl v4.4s, v4.4s, v18.4s\n"
    "srshl v7.4s, v7.4s, v17.4s\n"
    "sqrdmulh v6.4s, v6.4s, v19.4s\n"
    "sqrdmulh v5.4s, v5.4s, v19.4s\n"
    "sqrdmulh v4.4s, v4.4s, v19.4s\n"
    "srshl v3.4s, v3.4s, v18.4s\n"
    "srshl v6.4s, v6.4s, v17.4s\n"
    "srshl v5.4s, v5.4s, v17.4s\n"
    "srshl v4.4s, v4.4s, v17.4s\n"
    "sqrdmulh v3.4s, v3.4s, v19.4s\n"
    "srshl v2.4s, v2.4s, v18.4s\n"
    "srshl v1.4s, v1.4s, v18.4s\n"
    "srshl v0.4s, v0.4s, v18.4s\n"
    "srshl v3.4s, v3.4s, v17.4s\n"
    "sqrdmulh v2.4s, v2.4s, v19.4s\n"
    "sqrdmulh v1.4s, v1.4s, v19.4s\n"
    "sqrdmulh v0.4s, v0.4s, v19.4s\n"
    "smax v15.4s, v15.4s, v16.4s\n"
    "srshl v2.4s, v2.4s, v17.4s\n"
    "srshl v1.4s, v1.4s, v17.4s\n"
    "srshl v0.4s, v0.4s, v17.4s\n"
    "smin v15.4s, v15.4s, v20.4s\n"
    "smax v14.4s, v14.4s, v16.4s\n"
    "smax v13.4s, v13.4s, v16.4s\n"
    "smax v12.4s, v12.4s, v16.4s\n"
    "smin v14.4s, v14.4s, v20.4s\n"
    "smin v13.4s, v13.4s, v20.4s\n"
    "smin v12.4s, v12.4s, v20.4s\n"
    "smax v11.4s, v11.4s, v16.4s\n"
    "smax v10.4s, v10.4s, v16.4s\n"
    "smax v9.4s, v9.4s, v16.4s\n"
    "smin v11.4s, v11.4s, v20.4s\n"
    "smin v10.4s, v10.4s, v20.4s\n"
    "smin v9.4s, v9.4s, v20.4s\n"
    "smax v8.4s, v8.4s, v16.4s\n"
    "smax v7.4s, v7.4s, v16.4s\n"
    "smax v6.4s, v6.4s, v16.4s\n"
    "smin v8.4s, v8.4s, v20.4s\n"
    "smin v7.4s, v7.4s, v20.4s\n"
    "smin v6.4s, v6.4s, v20.4s\n"
    "smax v5.4s, v5.4s, v16.4s\n"
    "smax v4.4s, v4.4s, v16.4s\n"
    "smax v3.4s, v3.4s, v16.4s\n"
    "smin v5.4s, v5.4s, v20.4s\n"
    "smin v4.4s, v4.4s, v20.4s\n"
    "smin v3.4s, v3.4s, v20.4s\n"
    "smax v2.4s, v2.4s, v16.4s\n"
    "smax v1.4s, v1.4s, v16.4s\n"
    "smax v0.4s, v0.4s, v16.4s\n"
    "smin v2.4s, v2.4s, v20.4s\n"
    "smin v1.4s, v1.4s, v20.4s\n"
    "smin v0.4s, v0.4s, v20.4s\n"
    "uzp1 v23.16b, v15.16b, v14.16b\n"
    "uzp1 v16.16b, v13.16b, v12.16b\n"
    "uzp1 v22.16b, v11.16b, v10.16b\n"
    "uzp1 v18.16b, v9.16b, v8.16b\n"
    "uzp1 v21.16b, v7.16b, v6.16b\n"
    "uzp1 v17.16b, v5.16b, v4.16b\n"
    "uzp1 v20.16b, v3.16b, v2.16b\n"
    "uzp1 v19.16b, v1.16b, v0.16b\n"
    "uzp1 v16.16b, v23.16b, v16.16b\n"
    "str q16, [%x[outptr], x26]\n"
    "uzp1 v18.16b, v22.16b, v18.16b\n"
    "uzp1 v17.16b, v21.16b, v17.16b\n"
    "add x26, x26, #0x40\n"
    "uzp1 v16.16b, v20.16b, v19.16b\n"
    "str q18, [%x[outptr], x25]\n"
    "str q17, [%x[outptr], x24]\n"
    "str q16, [%x[outptr], x23]\n"
    "add x25, x25, #0x40\n"
    "add x24, x24, #0x40\n"
    "add x23, x23, #0x40\n"
    "bge 1b\n"
    "cbz %x[n_channels], 43f\n"
    "7:"  // Single vector of channels
    "cmp %x[n_channels], #0x10\n"
    "blt 14f\n"
    "8:"  // Single vector of channels: Loop
    "movi v15.4s, #0x0\n"
    "mov x19, %x[inptrs]\n"
    "movi v14.4s, #0x0\n"
    "lsr x22, %x[n_valid_cells], #0x1\n"
    "movi v13.4s, #0x0\n"
    "movi v12.4s, #0x0\n"
    "cbz x22, 11f\n"
    "ldp x21, x20, [x19, #0x0]\n"
    "add x19, x19, #0x10\n"
    "subs x22, x22, #0x1\n"
    "ldr q31, [x21, x26]\n"
    "ldr q30, [x20, x26]\n"
    "beq 10f\n"
    "9:"  // Single vector of channels: Loop: 2 inputs loop
    "saddl v23.8h, v31.8b, v30.8b\n"
    "ldp x21, x20, [x19, #0x0]\n"
    "add x19, x19, #0x10\n"
    "saddl2 v22.8h, v31.16b, v30.16b\n"
    "subs x22, x22, #0x1\n"
    "saddw v15.4s, v15.4s, v23.4h\n"
    "ldr q31, [x21, x26]\n"
    "saddw2 v14.4s, v14.4s, v23.8h\n"
    "ldr q30, [x20, x26]\n"
    "saddw v13.4s, v13.4s, v22.4h\n"
    "saddw2 v12.4s, v12.4s, v22.8h\n"
    "bgt 9b\n"
    "10:"  // Single vector of channels: Loop: 2 inputs tail
    "saddl v23.8h, v31.8b, v30.8b\n"
    "saddl2 v22.8h, v31.16b, v30.16b\n"
    "saddw v15.4s, v15.4s, v23.4h\n"
    "saddw2 v14.4s, v14.4s, v23.8h\n"
    "saddw v13.4s, v13.4s, v22.4h\n"
    "saddw2 v12.4s, v12.4s, v22.8h\n"
    "11:"  // Single vector of channels: Loop: After loop
    "ands x20, %x[n_valid_cells], #0x1\n"
    "beq 13f\n"
    "12:"  // Single vector of channels: Loop: Single input loop
    "ldr x21, [x19], #0x8\n"
    "subs x20, x20, #0x1\n"
    "ldr q31, [x21, x26]\n"
    "sxtl v16.8h, v31.8b\n"
    "sxtl2 v22.8h, v31.16b\n"
    "saddw v15.4s, v15.4s, v16.4h\n"
    "saddw2 v14.4s, v14.4s, v16.8h\n"
    "saddw v13.4s, v13.4s, v22.4h\n"
    "saddw2 v12.4s, v12.4s, v22.8h\n"
    "bgt 12b\n"
    "13:"  // Single vector of channels: Loop: Single input loop: End
    "movi v20.4s, #0x7f\n"
    "ld1r { v19.4s }, [%x[combined_rescale_value]]\n"
    "sub %x[n_channels], %x[n_channels], #0x10\n"
    "ld1r { v18.4s }, [%x[left_shift]]\n"
    "srshl v15.4s, v15.4s, v18.4s\n"
    "ld1r { v17.4s }, [%x[right_shift]]\n"
    "not v16.16b, v20.16b\n"
    "srshl v14.4s, v14.4s, v18.4s\n"
    "cmp %x[n_channels], #0x10\n"
    "srshl v13.4s, v13.4s, v18.4s\n"
    "srshl v12.4s, v12.4s, v18.4s\n"
    "sqrdmulh v15.4s, v15.4s, v19.4s\n"
    "sqrdmulh v14.4s, v14.4s, v19.4s\n"
    "sqrdmulh v13.4s, v13.4s, v19.4s\n"
    "sqrdmulh v12.4s, v12.4s, v19.4s\n"
    "srshl v15.4s, v15.4s, v17.4s\n"
    "srshl v14.4s, v14.4s, v17.4s\n"
    "srshl v13.4s, v13.4s, v17.4s\n"
    "srshl v12.4s, v12.4s, v17.4s\n"
    "smax v15.4s, v15.4s, v16.4s\n"
    "smax v14.4s, v14.4s, v16.4s\n"
    "smax v13.4s, v13.4s, v16.4s\n"
    "smax v12.4s, v12.4s, v16.4s\n"
    "smin v15.4s, v15.4s, v20.4s\n"
    "smin v14.4s, v14.4s, v20.4s\n"
    "smin v13.4s, v13.4s, v20.4s\n"
    "smin v12.4s, v12.4s, v20.4s\n"
    "uzp1 v23.16b, v15.16b, v14.16b\n"
    "uzp1 v16.16b, v13.16b, v12.16b\n"
    "uzp1 v16.16b, v23.16b, v16.16b\n"
    "str q16, [%x[outptr], x26]\n"
    "add x26, x26, #0x10\n"
    "bge 8b\n"
    "cbz %x[n_channels], 43f\n"
    "14:"  // Oddments
    "movi v15.4s, #0x0\n"
    "add %x[outptr], %x[outptr], x26\n"
    "movi v14.4s, #0x0\n"
    "mov x19, %x[inptrs]\n"
    "movi v13.4s, #0x0\n"
    "lsr x22, %x[n_valid_cells], #0x1\n"
    "movi v12.4s, #0x0\n"
    "cbz x22, 24f\n"
    "15:"  // Oddments: 2 inputs loop
    "movi v31.16b, #0x0\n"
    "ldp x21, x20, [x19, #0x0]\n"
    "add x19, x19, #0x10\n"
    "movi v30.16b, #0x0\n"
    "add x21, x21, x26\n"
    "add x20, x20, x26\n"
    "tbz %x[n_channels], #3, 19f\n"
    "ldr d31, [x21], #0x8\n"
    "ldr d30, [x20], #0x8\n"
    "tbz %x[n_channels], #2, 17f\n"
    "ld1 { v31.s }[2], [x21], #0x4\n"
    "ld1 { v30.s }[2], [x20], #0x4\n"
    "tbz %x[n_channels], #1, 16f\n"
    "ld1 { v31.h }[6], [x21], #0x2\n"
    "ld1 { v30.h }[6], [x20], #0x2\n"
    "tbz %x[n_channels], #0, 23f\n"
    "ld1 { v31.b }[14], [x21], #0x1\n"
    "ld1 { v30.b }[14], [x20], #0x1\n"
    "b 23f\n"
    "16:"  // Oddments: 2 inputs loop: Load: Bit 3: Bit 2: Bit 1: Unset
    "tbz %x[n_channels], #0, 23f\n"
    "ld1 { v31.b }[12], [x21], #0x1\n"
    "ld1 { v30.b }[12], [x20], #0x1\n"
    "b 23f\n"
    "17:"  // Oddments: 2 inputs loop: Load: Bit 3: Bit 2: Unset
    "tbz %x[n_channels], #1, 18f\n"
    "ld1 { v31.h }[4], [x21], #0x2\n"
    "ld1 { v30.h }[4], [x20], #0x2\n"
    "tbz %x[n_channels], #0, 23f\n"
    "ld1 { v31.b }[10], [x21], #0x1\n"
    "ld1 { v30.b }[10], [x20], #0x1\n"
    "b 23f\n"
    "18:"  // Oddments: 2 inputs loop: Load: Bit 3: Bit 2: Unset: Bit 1: Unset
    "tbz %x[n_channels], #0, 23f\n"
    "ld1 { v31.b }[8], [x21], #0x1\n"
    "ld1 { v30.b }[8], [x20], #0x1\n"
    "b 23f\n"
    "19:"  // Oddments: 2 inputs loop: Load: Bit 3: Unset
    "tbz %x[n_channels], #2, 21f\n"
    "ldr s31, [x21], #0x4\n"
    "ldr s30, [x20], #0x4\n"
    "tbz %x[n_channels], #1, 20f\n"
    "ld1 { v31.h }[2], [x21], #0x2\n"
    "ld1 { v30.h }[2], [x20], #0x2\n"
    "tbz %x[n_channels], #0, 23f\n"
    "ld1 { v31.b }[6], [x21], #0x1\n"
    "ld1 { v30.b }[6], [x20], #0x1\n"
    "b 23f\n"
    "20:"  // Oddments: 2 inputs loop: Load: Bit 3: Unset: Bit 2: Bit 1: Unset
    "tbz %x[n_channels], #0, 23f\n"
    "ld1 { v31.b }[4], [x21], #0x1\n"
    "ld1 { v30.b }[4], [x20], #0x1\n"
    "b 23f\n"
    "21:"  // Oddments: 2 inputs loop: Load: Bit 3: Unset: Bit 2: Unset
    "tbz %x[n_channels], #1, 22f\n"
    "ldr h31, [x21], #0x2\n"
    "ldr h30, [x20], #0x2\n"
    "tbz %x[n_channels], #0, 23f\n"
    "ld1 { v31.b }[2], [x21], #0x1\n"
    "ld1 { v30.b }[2], [x20], #0x1\n"
    "b 23f\n"
    "22:"  // Oddments: 2 inputs loop: Load: Bit 3: Unset: Bit 2: Unset: Bit 1: Unset
    "tbz %x[n_channels], #0, 23f\n"
    "ldr b31, [x21], #0x1\n"
    "ldr b30, [x20], #0x1\n"
    "23:"  // Oddments: 2 inputs loop: Load: Bit 3: End
    "saddl v23.8h, v31.8b, v30.8b\n"
    "subs x22, x22, #0x1\n"
    "saddl2 v22.8h, v31.16b, v30.16b\n"
    "saddw v15.4s, v15.4s, v23.4h\n"
    "saddw2 v14.4s, v14.4s, v23.8h\n"
    "saddw v13.4s, v13.4s, v22.4h\n"
    "saddw2 v12.4s, v12.4s, v22.8h\n"
    "bgt 15b\n"
    "24:"  // Oddments: After loop
    "ands x20, %x[n_valid_cells], #0x1\n"
    "beq 34f\n"
    "25:"  // Oddments: Single input loop
    "movi v31.16b, #0x0\n"
    "ldr x21, [x19], #0x8\n"
    "add x21, x21, x26\n"
    "tbz %x[n_channels], #3, 29f\n"
    "ldr d31, [x21], #0x8\n"
    "tbz %x[n_channels], #2, 27f\n"
    "ld1 { v31.s }[2], [x21], #0x4\n"
    "tbz %x[n_channels], #1, 26f\n"
    "ld1 { v31.h }[6], [x21], #0x2\n"
    "tbz %x[n_channels], #0, 33f\n"
    "ld1 { v31.b }[14], [x21], #0x1\n"
    "b 33f\n"
    "26:"  // Oddments: Single input loop: Load: Bit 3: Bit 2: Bit 1: Unset
    "tbz %x[n_channels], #0, 33f\n"
    "ld1 { v31.b }[12], [x21], #0x1\n"
    "b 33f\n"
    "27:"  // Oddments: Single input loop: Load: Bit 3: Bit 2: Unset
    "tbz %x[n_channels], #1, 28f\n"
    "ld1 { v31.h }[4], [x21], #0x2\n"
    "tbz %x[n_channels], #0, 33f\n"
    "ld1 { v31.b }[10], [x21], #0x1\n"
    "b 33f\n"
    "28:"  // Oddments: Single input loop: Load: Bit 3: Bit 2: Unset: Bit 1: Unset
    "tbz %x[n_channels], #0, 33f\n"
    "ld1 { v31.b }[8], [x21], #0x1\n"
    "b 33f\n"
    "29:"  // Oddments: Single input loop: Load: Bit 3: Unset
    "tbz %x[n_channels], #2, 31f\n"
    "ldr s31, [x21], #0x4\n"
    "tbz %x[n_channels], #1, 30f\n"
    "ld1 { v31.h }[2], [x21], #0x2\n"
    "tbz %x[n_channels], #0, 33f\n"
    "ld1 { v31.b }[6], [x21], #0x1\n"
    "b 33f\n"
    "30:"  // Oddments: Single input loop: Load: Bit 3: Unset: Bit 2: Bit 1: Unset
    "tbz %x[n_channels], #0, 33f\n"
    "ld1 { v31.b }[4], [x21], #0x1\n"
    "b 33f\n"
    "31:"  // Oddments: Single input loop: Load: Bit 3: Unset: Bit 2: Unset
    "tbz %x[n_channels], #1, 32f\n"
    "ldr h31, [x21], #0x2\n"
    "tbz %x[n_channels], #0, 33f\n"
    "ld1 { v31.b }[2], [x21], #0x1\n"
    "b 33f\n"
    "32:"  // Oddments: Single input loop: Load: Bit 3: Unset: Bit 2: Unset: Bit 1: Unset
    "tbz %x[n_channels], #0, 33f\n"
    "ldr b31, [x21], #0x1\n"
    "33:"  // Oddments: Single input loop: Load: Bit 3: End
    "sxtl v16.8h, v31.8b\n"
    "subs x20, x20, #0x1\n"
    "sxtl2 v22.8h, v31.16b\n"
    "saddw v15.4s, v15.4s, v16.4h\n"
    "saddw2 v14.4s, v14.4s, v16.8h\n"
    "saddw v13.4s, v13.4s, v22.4h\n"
    "saddw2 v12.4s, v12.4s, v22.8h\n"
    "bgt 25b\n"
    "34:"  // Oddments: Single input loop: End
    "movi v20.4s, #0x7f\n"
    "ld1r { v19.4s }, [%x[combined_rescale_value]]\n"
    "not v16.16b, v20.16b\n"
    "ld1r { v18.4s }, [%x[left_shift]]\n"
    "srshl v15.4s, v15.4s, v18.4s\n"
    "ld1r { v17.4s }, [%x[right_shift]]\n"
    "srshl v14.4s, v14.4s, v18.4s\n"
    "srshl v13.4s, v13.4s, v18.4s\n"
    "srshl v12.4s, v12.4s, v18.4s\n"
    "sqrdmulh v15.4s, v15.4s, v19.4s\n"
    "sqrdmulh v14.4s, v14.4s, v19.4s\n"
    "sqrdmulh v13.4s, v13.4s, v19.4s\n"
    "sqrdmulh v12.4s, v12.4s, v19.4s\n"
    "srshl v15.4s, v15.4s, v17.4s\n"
    "srshl v14.4s, v14.4s, v17.4s\n"
    "srshl v13.4s, v13.4s, v17.4s\n"
    "srshl v12.4s, v12.4s, v17.4s\n"
    "smax v15.4s, v15.4s, v16.4s\n"
    "smax v14.4s, v14.4s, v16.4s\n"
    "smax v13.4s, v13.4s, v16.4s\n"
    "smax v12.4s, v12.4s, v16.4s\n"
    "smin v15.4s, v15.4s, v20.4s\n"
    "smin v14.4s, v14.4s, v20.4s\n"
    "smin v13.4s, v13.4s, v20.4s\n"
    "smin v12.4s, v12.4s, v20.4s\n"
    "uzp1 v23.16b, v15.16b, v14.16b\n"
    "uzp1 v16.16b, v13.16b, v12.16b\n"
    "uzp1 v16.16b, v23.16b, v16.16b\n"
    "tbz %x[n_channels], #3, 38f\n"
    "st1 { v16.d }[0], [%x[outptr]], #0x8\n"
    "tbz %x[n_channels], #2, 36f\n"
    "st1 { v16.s }[2], [%x[outptr]], #0x4\n"
    "tbz %x[n_channels], #1, 35f\n"
    "st1 { v16.h }[6], [%x[outptr]], #0x2\n"
    "tbz %x[n_channels], #0, 42f\n"
    "st1 { v16.b }[14], [%x[outptr]], #0x1\n"
    "b 42f\n"
    "35:"  // Oddments: Store: Bit 3: Bit 2: Bit 1: Unset
    "tbz %x[n_channels], #0, 42f\n"
    "st1 { v16.b }[12], [%x[outptr]], #0x1\n"
    "b 42f\n"
    "36:"  // Oddments: Store: Bit 3: Bit 2: Unset
    "tbz %x[n_channels], #1, 37f\n"
    "st1 { v16.h }[4], [%x[outptr]], #0x2\n"
    "tbz %x[n_channels], #0, 42f\n"
    "st1 { v16.b }[10], [%x[outptr]], #0x1\n"
    "b 42f\n"
    "37:"  // Oddments: Store: Bit 3: Bit 2: Unset: Bit 1: Unset
    "tbz %x[n_channels], #0, 42f\n"
    "st1 { v16.b }[8], [%x[outptr]], #0x1\n"
    "b 42f\n"
    "38:"  // Oddments: Store: Bit 3: Unset
    "tbz %x[n_channels], #2, 40f\n"
    "st1 { v16.s }[0], [%x[outptr]], #0x4\n"
    "tbz %x[n_channels], #1, 39f\n"
    "st1 { v16.h }[2], [%x[outptr]], #0x2\n"
    "tbz %x[n_channels], #0, 42f\n"
    "st1 { v16.b }[6], [%x[outptr]], #0x1\n"
    "b 42f\n"
    "39:"  // Oddments: Store: Bit 3: Unset: Bit 2: Bit 1: Unset
    "tbz %x[n_channels], #0, 42f\n"
    "st1 { v16.b }[4], [%x[outptr]], #0x1\n"
    "b 42f\n"
    "40:"  // Oddments: Store: Bit 3: Unset: Bit 2: Unset
    "tbz %x[n_channels], #1, 41f\n"
    "st1 { v16.h }[0], [%x[outptr]], #0x2\n"
    "tbz %x[n_channels], #0, 42f\n"
    "st1 { v16.b }[2], [%x[outptr]], #0x1\n"
    "b 42f\n"
    "41:"  // Oddments: Store: Bit 3: Unset: Bit 2: Unset: Bit 1: Unset
    "tbz %x[n_channels], #0, 42f\n"
    "st1 { v16.b }[0], [%x[outptr]], #0x1\n"
    "42:"  // Oddments: Store: Bit 3: End

    "43:"  // End

    : [n_channels] "+&r" (n_channels), [outptr] "+&r" (outptr)
    : [combined_rescale_value] "r" (&combined_rescale_value), [inptrs] "r" (inptrs), [left_shift] "r" (&left_shift), [n_valid_cells] "r" (n_valid_cells), [right_shift] "r" (&right_shift)
    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "x26"
  );
}

}  // namespace pooling
}  // namespace arm_conv

#endif  // defined(__aarch64__)