path: root/src/core/NEON/kernels/arm_conv/pooling/kernels/a64_u8q_nhwc_max_generic_depthfirst/generic.cpp

/*
 * 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>

#if defined(__aarch64__)

namespace arm_conv {
namespace pooling {


void a64_u8q_nhwc_max_generic_depthfirst_impl(
  const uint64_t,
  const uint64_t n_valid_cells,
  uint64_t n_channels,
  const uint8_t *const *const inptrs,
  uint8_t *outptr,
  const Requantize32 &qp
)
{
  __asm__ __volatile__(
    "mov x28, #0x0\n"
    "mov x27, #0x10\n" // cntb _, ALL, #1
    "mov x26, #0x20\n" // cntb _, ALL, #2
    "mov x25, #0x30\n" // cntb _, ALL, #3
    "cmp %x[n_channels], #0x40\n"
    "blt 7f\n"
    "1:"  // 4-vectors of channels
    "movi v4.16b, #0x0\n"
    "mov x19, %x[inptrs]\n"
    "movi v8.16b, #0x0\n"
    "lsr x24, %x[n_valid_cells], #0x2\n"
    "movi v7.16b, #0x0\n"
    "movi v6.16b, #0x0\n"
    "cbz x24, 4f\n"
    "ldp x23, x22, [x19, #0x0]\n"
    "ldp x21, x20, [x19, #0x10]\n"
    "add x19, x19, #0x20\n"
    "subs x24, x24, #0x1\n"
    "ldr q3, [x23, x28]\n"
    "ldr q2, [x22, x28]\n"
    "ldr q1, [x21, x28]\n"
    "ldr q0, [x20, x28]\n"
    "ldr q31, [x23, x27]\n"
    "ldr q30, [x22, x27]\n"
    "ldr q29, [x21, x27]\n"
    "ldr q28, [x20, x27]\n"
    "ldr q27, [x23, x26]\n"
    "ldr q21, [x22, x26]\n"
    "ldr q26, [x21, x26]\n"
    "ldr q17, [x20, x26]\n"
    "ldr q25, [x23, x25]\n"
    "ldr q20, [x22, x25]\n"
    "ldr q24, [x21, x25]\n"
    "ldr q16, [x20, x25]\n"
    "beq 3f\n"
    "2:"  // 4-vectors of channels: 4 inputs loop
    "umax v23.16b, v3.16b, v2.16b\n"
    "ldp x23, x22, [x19, #0x0]\n"
    "subs x24, x24, #0x1\n"
    "umax v19.16b, v1.16b, v0.16b\n"
    "ldp x21, x20, [x19, #0x10]\n"
    "add x19, x19, #0x20\n"
    "umax v22.16b, v31.16b, v30.16b\n"
    "ldr q3, [x23, x28]\n"
    "umax v18.16b, v29.16b, v28.16b\n"
    "umax v21.16b, v27.16b, v21.16b\n"
    "ldr q2, [x22, x28]\n"
    "umax v17.16b, v26.16b, v17.16b\n"
    "ldr q1, [x21, x28]\n"
    "umax v20.16b, v25.16b, v20.16b\n"
    "ldr q0, [x20, x28]\n"
    "umax v16.16b, v24.16b, v16.16b\n"
    "ldr q31, [x23, x27]\n"
    "umax v19.16b, v23.16b, v19.16b\n"
    "ldr q30, [x22, x27]\n"
    "umax v18.16b, v22.16b, v18.16b\n"
    "ldr q29, [x21, x27]\n"
    "umax v17.16b, v21.16b, v17.16b\n"
    "ldr q28, [x20, x27]\n"
    "umax v16.16b, v20.16b, v16.16b\n"
    "ldr q27, [x23, x26]\n"
    "umax v4.16b, v4.16b, v19.16b\n"
    "ldr q21, [x22, x26]\n"
    "umax v8.16b, v8.16b, v18.16b\n"
    "ldr q26, [x21, x26]\n"
    "umax v7.16b, v7.16b, v17.16b\n"
    "ldr q17, [x20, x26]\n"
    "umax v6.16b, v6.16b, v16.16b\n"
    "ldr q25, [x23, x25]\n"
    "ldr q20, [x22, x25]\n"
    "ldr q24, [x21, x25]\n"
    "ldr q16, [x20, x25]\n"
    "bgt 2b\n"
    "3:"  // 4-vectors of channels: 4 inputs tail
    "umax v23.16b, v3.16b, v2.16b\n"
    "umax v19.16b, v1.16b, v0.16b\n"
    "umax v22.16b, v31.16b, v30.16b\n"
    "umax v18.16b, v29.16b, v28.16b\n"
    "umax v21.16b, v27.16b, v21.16b\n"
    "umax v17.16b, v26.16b, v17.16b\n"
    "umax v20.16b, v25.16b, v20.16b\n"
    "umax v16.16b, v24.16b, v16.16b\n"
    "umax v19.16b, v23.16b, v19.16b\n"
    "umax v18.16b, v22.16b, v18.16b\n"
    "umax v17.16b, v21.16b, v17.16b\n"
    "umax v16.16b, v20.16b, v16.16b\n"
    "umax v4.16b, v4.16b, v19.16b\n"
    "umax v8.16b, v8.16b, v18.16b\n"
    "umax v7.16b, v7.16b, v17.16b\n"
    "umax v6.16b, v6.16b, v16.16b\n"
    "4:"  // 4-vectors of channels: After loop
    "ands x20, %x[n_valid_cells], #0x3\n"
    "beq 6f\n"
    "5:"  // 4-vectors of channels: Single input loop
    "ldr x23, [x19], #0x8\n"
    "subs x20, x20, #0x1\n"
    "ldr q3, [x23, x28]\n"
    "umax v4.16b, v4.16b, v3.16b\n"
    "ldr q31, [x23, x27]\n"
    "ldr q27, [x23, x26]\n"
    "umax v8.16b, v8.16b, v31.16b\n"
    "ldr q25, [x23, x25]\n"
    "umax v7.16b, v7.16b, v27.16b\n"
    "umax v6.16b, v6.16b, v25.16b\n"
    "bgt 5b\n"
    "6:"  // 4-vectors of channels: Single input loop: End
    "uxtl v17.8h, v4.8b\n"
    "add x19, %x[quant_params], %[offsetof_qp_input_offset]\n"
    "ld1r { v5.4s }, [x19]\n"
    "uxtl2 v16.8h, v4.16b\n"
    "add x19, %x[quant_params], %[offsetof_qp_per_layer_mul]\n"
    "uxtl v21.8h, v8.8b\n"
    "ld1r { v4.4s }, [x19]\n"
    "add x19, %x[quant_params], %[offsetof_qp_per_layer_left_shift]\n"
    "uxtl2 v20.8h, v8.16b\n"
    "ld1r { v3.4s }, [x19]\n"
    "add x19, %x[quant_params], %[offsetof_qp_per_layer_right_shift]\n"
    "uxtl v19.8h, v7.8b\n"
    "ld1r { v2.4s }, [x19]\n"
    "add x19, %x[quant_params], %[offsetof_qp_output_offset]\n"
    "uxtl2 v24.8h, v7.16b\n"
    "ld1r { v1.4s }, [x19]\n"
    "sub %x[n_channels], %x[n_channels], #0x40\n"
    "uxtl v0.8h, v6.8b\n"
    "cmp %x[n_channels], #0x40\n"
    "uxtl2 v31.8h, v6.16b\n"
    "neg v5.4s, v5.4s\n"
    "movi v30.4s, #0x0\n"
    "movi v29.4s, #0xff\n"
    "saddw v23.4s, v5.4s, v17.4h\n"
    "saddw2 v18.4s, v5.4s, v17.8h\n"
    "saddw v17.4s, v5.4s, v16.4h\n"
    "saddw2 v16.4s, v5.4s, v16.8h\n"
    "saddw v22.4s, v5.4s, v21.4h\n"
    "saddw2 v21.4s, v5.4s, v21.8h\n"
    "saddw v28.4s, v5.4s, v20.4h\n"
    "saddw2 v20.4s, v5.4s, v20.8h\n"
    "saddw v27.4s, v5.4s, v19.4h\n"
    "saddw2 v19.4s, v5.4s, v19.8h\n"
    "srshl v23.4s, v23.4s, v3.4s\n"
    "srshl v18.4s, v18.4s, v3.4s\n"
    "srshl v17.4s, v17.4s, v3.4s\n"
    "srshl v16.4s, v16.4s, v3.4s\n"
    "sqrdmulh v23.4s, v23.4s, v4.4s\n"
    "sqrdmulh v18.4s, v18.4s, v4.4s\n"
    "sqrdmulh v17.4s, v17.4s, v4.4s\n"
    "sqrdmulh v16.4s, v16.4s, v4.4s\n"
    "srshl v23.4s, v23.4s, v2.4s\n"
    "srshl v18.4s, v18.4s, v2.4s\n"
    "srshl v17.4s, v17.4s, v2.4s\n"
    "srshl v16.4s, v16.4s, v2.4s\n"
    "srshl v22.4s, v22.4s, v3.4s\n"
    "srshl v21.4s, v21.4s, v3.4s\n"
    "srshl v28.4s, v28.4s, v3.4s\n"
    "srshl v20.4s, v20.4s, v3.4s\n"
    "sqrdmulh v22.4s, v22.4s, v4.4s\n"
    "sqrdmulh v21.4s, v21.4s, v4.4s\n"
    "sqrdmulh v28.4s, v28.4s, v4.4s\n"
    "sqrdmulh v20.4s, v20.4s, v4.4s\n"
    "srshl v22.4s, v22.4s, v2.4s\n"
    "srshl v21.4s, v21.4s, v2.4s\n"
    "srshl v28.4s, v28.4s, v2.4s\n"
    "srshl v20.4s, v20.4s, v2.4s\n"
    "srshl v27.4s, v27.4s, v3.4s\n"
    "srshl v19.4s, v19.4s, v3.4s\n"
    "add v23.4s, v23.4s, v1.4s\n"
    "add v18.4s, v18.4s, v1.4s\n"
    "sqrdmulh v27.4s, v27.4s, v4.4s\n"
    "sqrdmulh v19.4s, v19.4s, v4.4s\n"
    "add v17.4s, v17.4s, v1.4s\n"
    "add v16.4s, v16.4s, v1.4s\n"
    "srshl v27.4s, v27.4s, v2.4s\n"
    "srshl v19.4s, v19.4s, v2.4s\n"
    "add v22.4s, v22.4s, v1.4s\n"
    "add v21.4s, v21.4s, v1.4s\n"
    "add v28.4s, v28.4s, v1.4s\n"
    "add v20.4s, v20.4s, v1.4s\n"
    "add v27.4s, v27.4s, v1.4s\n"
    "add v19.4s, v19.4s, v1.4s\n"
    "smax v23.4s, v23.4s, v30.4s\n"
    "smax v18.4s, v18.4s, v30.4s\n"
    "smax v17.4s, v17.4s, v30.4s\n"
    "smin v23.4s, v23.4s, v29.4s\n"
    "smin v18.4s, v18.4s, v29.4s\n"
    "smin v17.4s, v17.4s, v29.4s\n"
    "smax v16.4s, v16.4s, v30.4s\n"
    "smax v22.4s, v22.4s, v30.4s\n"
    "smax v21.4s, v21.4s, v30.4s\n"
    "smin v16.4s, v16.4s, v29.4s\n"
    "smin v22.4s, v22.4s, v29.4s\n"
    "smin v21.4s, v21.4s, v29.4s\n"
    "smax v28.4s, v28.4s, v30.4s\n"
    "smax v20.4s, v20.4s, v30.4s\n"
    "smax v27.4s, v27.4s, v30.4s\n"
    "smin v28.4s, v28.4s, v29.4s\n"
    "smin v20.4s, v20.4s, v29.4s\n"
    "smin v27.4s, v27.4s, v29.4s\n"
    "smax v19.4s, v19.4s, v30.4s\n"
    "uzp1 v26.16b, v23.16b, v18.16b\n"
    "saddw v25.4s, v5.4s, v24.4h\n"
    "saddw2 v18.4s, v5.4s, v24.8h\n"
    "smin v19.4s, v19.4s, v29.4s\n"
    "srshl v25.4s, v25.4s, v3.4s\n"
    "srshl v18.4s, v18.4s, v3.4s\n"
    "uzp1 v24.16b, v17.16b, v16.16b\n"
    "saddw v17.4s, v5.4s, v0.4h\n"
    "saddw2 v16.4s, v5.4s, v0.8h\n"
    "sqrdmulh v25.4s, v25.4s, v4.4s\n"
    "sqrdmulh v18.4s, v18.4s, v4.4s\n"
    "srshl v17.4s, v17.4s, v3.4s\n"
    "srshl v16.4s, v16.4s, v3.4s\n"
    "srshl v25.4s, v25.4s, v2.4s\n"
    "srshl v18.4s, v18.4s, v2.4s\n"
    "sqrdmulh v17.4s, v17.4s, v4.4s\n"
    "sqrdmulh v16.4s, v16.4s, v4.4s\n"
    "add v25.4s, v25.4s, v1.4s\n"
    "add v18.4s, v18.4s, v1.4s\n"
    "srshl v17.4s, v17.4s, v2.4s\n"
    "srshl v16.4s, v16.4s, v2.4s\n"
    "smax v25.4s, v25.4s, v30.4s\n"
    "smax v18.4s, v18.4s, v30.4s\n"
    "add v17.4s, v17.4s, v1.4s\n"
    "add v16.4s, v16.4s, v1.4s\n"
    "smin v25.4s, v25.4s, v29.4s\n"
    "smin v18.4s, v18.4s, v29.4s\n"
    "smax v17.4s, v17.4s, v30.4s\n"
    "smax v16.4s, v16.4s, v30.4s\n"
    "uzp1 v23.16b, v22.16b, v21.16b\n"
    "saddw v22.4s, v5.4s, v31.4h\n"
    "saddw2 v21.4s, v5.4s, v31.8h\n"
    "smin v17.4s, v17.4s, v29.4s\n"
    "srshl v22.4s, v22.4s, v3.4s\n"
    "srshl v21.4s, v21.4s, v3.4s\n"
    "smin v16.4s, v16.4s, v29.4s\n"
    "uzp1 v20.16b, v28.16b, v20.16b\n"
    "sqrdmulh v22.4s, v22.4s, v4.4s\n"
    "sqrdmulh v21.4s, v21.4s, v4.4s\n"
    "uzp1 v19.16b, v27.16b, v19.16b\n"
    "uzp1 v18.16b, v25.16b, v18.16b\n"
    "srshl v22.4s, v22.4s, v2.4s\n"
    "srshl v21.4s, v21.4s, v2.4s\n"
    "uzp1 v17.16b, v17.16b, v16.16b\n"
    "uzp1 v16.16b, v26.16b, v24.16b\n"
    "str q16, [%x[outptr], x28]\n"
    "add v22.4s, v22.4s, v1.4s\n"
    "add x28, x28, #0x40\n"
    "add v21.4s, v21.4s, v1.4s\n"
    "uzp1 v16.16b, v23.16b, v20.16b\n"
    "str q16, [%x[outptr], x27]\n"
    "smax v22.4s, v22.4s, v30.4s\n"
    "add x27, x27, #0x40\n"
    "smax v21.4s, v21.4s, v30.4s\n"
    "uzp1 v16.16b, v19.16b, v18.16b\n"
    "str q16, [%x[outptr], x26]\n"
    "smin v22.4s, v22.4s, v29.4s\n"
    "add x26, x26, #0x40\n"
    "smin v21.4s, v21.4s, v29.4s\n"
    "uzp1 v16.16b, v22.16b, v21.16b\n"
    "uzp1 v16.16b, v17.16b, v16.16b\n"
    "str q16, [%x[outptr], x25]\n"
    "add x25, x25, #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 v4.16b, #0x0\n"
    "mov x19, %x[inptrs]\n"
    "lsr x24, %x[n_valid_cells], #0x2\n"
    "cbz x24, 11f\n"
    "ldp x23, x22, [x19, #0x0]\n"
    "ldp x21, x20, [x19, #0x10]\n"
    "add x19, x19, #0x20\n"
    "subs x24, x24, #0x1\n"
    "ldr q3, [x23, x28]\n"
    "ldr q2, [x22, x28]\n"
    "ldr q1, [x21, x28]\n"
    "ldr q0, [x20, x28]\n"
    "beq 10f\n"
    "9:"  // Single vector of channels: Loop: 4 inputs loop
    "umax v23.16b, v3.16b, v2.16b\n"
    "ldp x23, x22, [x19, #0x0]\n"
    "subs x24, x24, #0x1\n"
    "umax v19.16b, v1.16b, v0.16b\n"
    "ldp x21, x20, [x19, #0x10]\n"
    "add x19, x19, #0x20\n"
    "umax v19.16b, v23.16b, v19.16b\n"
    "ldr q3, [x23, x28]\n"
    "ldr q2, [x22, x28]\n"
    "umax v4.16b, v4.16b, v19.16b\n"
    "ldr q1, [x21, x28]\n"
    "ldr q0, [x20, x28]\n"
    "bgt 9b\n"
    "10:"  // Single vector of channels: Loop: 4 inputs tail
    "umax v23.16b, v3.16b, v2.16b\n"
    "umax v19.16b, v1.16b, v0.16b\n"
    "umax v19.16b, v23.16b, v19.16b\n"
    "umax v4.16b, v4.16b, v19.16b\n"
    "11:"  // Single vector of channels: Loop: After loop
    "ands x20, %x[n_valid_cells], #0x3\n"
    "beq 13f\n"
    "12:"  // Single vector of channels: Loop: Single input loop
    "ldr x23, [x19], #0x8\n"
    "subs x20, x20, #0x1\n"
    "ldr q3, [x23, x28]\n"
    "umax v4.16b, v4.16b, v3.16b\n"
    "bgt 12b\n"
    "13:"  // Single vector of channels: Loop: Single input loop: End
    "uxtl v17.8h, v4.8b\n"
    "add x19, %x[quant_params], %[offsetof_qp_input_offset]\n"
    "ld1r { v5.4s }, [x19]\n"
    "uxtl2 v16.8h, v4.16b\n"
    "add x19, %x[quant_params], %[offsetof_qp_per_layer_mul]\n"
    "movi v30.4s, #0x0\n"
    "ld1r { v4.4s }, [x19]\n"
    "add x19, %x[quant_params], %[offsetof_qp_per_layer_left_shift]\n"
    "movi v29.4s, #0xff\n"
    "ld1r { v3.4s }, [x19]\n"
    "add x19, %x[quant_params], %[offsetof_qp_per_layer_right_shift]\n"
    "neg v5.4s, v5.4s\n"
    "ld1r { v2.4s }, [x19]\n"
    "add x19, %x[quant_params], %[offsetof_qp_output_offset]\n"
    "saddw v23.4s, v5.4s, v17.4h\n"
    "ld1r { v1.4s }, [x19]\n"
    "sub %x[n_channels], %x[n_channels], #0x10\n"
    "saddw2 v18.4s, v5.4s, v17.8h\n"
    "cmp %x[n_channels], #0x10\n"
    "saddw v17.4s, v5.4s, v16.4h\n"
    "saddw2 v16.4s, v5.4s, v16.8h\n"
    "srshl v23.4s, v23.4s, v3.4s\n"
    "srshl v18.4s, v18.4s, v3.4s\n"
    "srshl v17.4s, v17.4s, v3.4s\n"
    "srshl v16.4s, v16.4s, v3.4s\n"
    "sqrdmulh v23.4s, v23.4s, v4.4s\n"
    "sqrdmulh v18.4s, v18.4s, v4.4s\n"
    "sqrdmulh v17.4s, v17.4s, v4.4s\n"
    "sqrdmulh v16.4s, v16.4s, v4.4s\n"
    "srshl v23.4s, v23.4s, v2.4s\n"
    "srshl v18.4s, v18.4s, v2.4s\n"
    "srshl v17.4s, v17.4s, v2.4s\n"
    "srshl v16.4s, v16.4s, v2.4s\n"
    "add v23.4s, v23.4s, v1.4s\n"
    "add v18.4s, v18.4s, v1.4s\n"
    "add v17.4s, v17.4s, v1.4s\n"
    "add v16.4s, v16.4s, v1.4s\n"
    "smax v23.4s, v23.4s, v30.4s\n"
    "smax v18.4s, v18.4s, v30.4s\n"
    "smax v17.4s, v17.4s, v30.4s\n"
    "smin v23.4s, v23.4s, v29.4s\n"
    "smin v18.4s, v18.4s, v29.4s\n"
    "smin v17.4s, v17.4s, v29.4s\n"
    "smax v16.4s, v16.4s, v30.4s\n"
    "uzp1 v26.16b, v23.16b, v18.16b\n"
    "smin v16.4s, v16.4s, v29.4s\n"
    "uzp1 v24.16b, v17.16b, v16.16b\n"
    "uzp1 v16.16b, v26.16b, v24.16b\n"
    "str q16, [%x[outptr], x28]\n"
    "add x28, x28, #0x10\n"
    "bge 8b\n"
    "cbz %x[n_channels], 43f\n"
    "14:"  // Oddments
    "movi v4.16b, #0x0\n"
    "add %x[outptr], %x[outptr], x28\n"
    "mov x19, %x[inptrs]\n"
    "lsr x24, %x[n_valid_cells], #0x2\n"
    "cbz x24, 24f\n"
    "15:"  // Oddments: 4 inputs loop
    "movi v3.16b, #0x0\n"
    "ldp x23, x22, [x19, #0x0]\n"
    "add x23, x23, x28\n"
    "movi v2.16b, #0x0\n"
    "ldp x21, x20, [x19, #0x10]\n"
    "movi v1.16b, #0x0\n"
    "add x19, x19, #0x20\n"
    "movi v0.16b, #0x0\n"
    "add x22, x22, x28\n"
    "add x21, x21, x28\n"
    "add x20, x20, x28\n"
    "tbz %x[n_channels], #3, 19f\n"
    "ldr d3, [x23], #0x8\n"
    "ldr d2, [x22], #0x8\n"
    "ldr d1, [x21], #0x8\n"
    "ldr d0, [x20], #0x8\n"
    "tbz %x[n_channels], #2, 17f\n"
    "ld1 { v3.s }[2], [x23], #0x4\n"
    "ld1 { v2.s }[2], [x22], #0x4\n"
    "ld1 { v1.s }[2], [x21], #0x4\n"
    "ld1 { v0.s }[2], [x20], #0x4\n"
    "tbz %x[n_channels], #1, 16f\n"
    "ld1 { v3.h }[6], [x23], #0x2\n"
    "ld1 { v2.h }[6], [x22], #0x2\n"
    "ld1 { v1.h }[6], [x21], #0x2\n"
    "ld1 { v0.h }[6], [x20], #0x2\n"
    "tbz %x[n_channels], #0, 23f\n"
    "ld1 { v3.b }[14], [x23], #0x1\n"
    "ld1 { v2.b }[14], [x22], #0x1\n"
    "ld1 { v1.b }[14], [x21], #0x1\n"
    "ld1 { v0.b }[14], [x20], #0x1\n"
    "b 23f\n"
    "16:"  // Oddments: 4 inputs loop: Load: Bit 3: Bit 2: Bit 1: Unset
    "tbz %x[n_channels], #0, 23f\n"
    "ld1 { v3.b }[12], [x23], #0x1\n"
    "ld1 { v2.b }[12], [x22], #0x1\n"
    "ld1 { v1.b }[12], [x21], #0x1\n"
    "ld1 { v0.b }[12], [x20], #0x1\n"
    "b 23f\n"
    "17:"  // Oddments: 4 inputs loop: Load: Bit 3: Bit 2: Unset
    "tbz %x[n_channels], #1, 18f\n"
    "ld1 { v3.h }[4], [x23], #0x2\n"
    "ld1 { v2.h }[4], [x22], #0x2\n"
    "ld1 { v1.h }[4], [x21], #0x2\n"
    "ld1 { v0.h }[4], [x20], #0x2\n"
    "tbz %x[n_channels], #0, 23f\n"
    "ld1 { v3.b }[10], [x23], #0x1\n"
    "ld1 { v2.b }[10], [x22], #0x1\n"
    "ld1 { v1.b }[10], [x21], #0x1\n"
    "ld1 { v0.b }[10], [x20], #0x1\n"
    "b 23f\n"
    "18:"  // Oddments: 4 inputs loop: Load: Bit 3: Bit 2: Unset: Bit 1: Unset
    "tbz %x[n_channels], #0, 23f\n"
    "ld1 { v3.b }[8], [x23], #0x1\n"
    "ld1 { v2.b }[8], [x22], #0x1\n"
    "ld1 { v1.b }[8], [x21], #0x1\n"
    "ld1 { v0.b }[8], [x20], #0x1\n"
    "b 23f\n"
    "19:"  // Oddments: 4 inputs loop: Load: Bit 3: Unset
    "tbz %x[n_channels], #2, 21f\n"
    "ldr s3, [x23], #0x4\n"
    "ldr s2, [x22], #0x4\n"
    "ldr s1, [x21], #0x4\n"
    "ldr s0, [x20], #0x4\n"
    "tbz %x[n_channels], #1, 20f\n"
    "ld1 { v3.h }[2], [x23], #0x2\n"
    "ld1 { v2.h }[2], [x22], #0x2\n"
    "ld1 { v1.h }[2], [x21], #0x2\n"
    "ld1 { v0.h }[2], [x20], #0x2\n"
    "tbz %x[n_channels], #0, 23f\n"
    "ld1 { v3.b }[6], [x23], #0x1\n"
    "ld1 { v2.b }[6], [x22], #0x1\n"
    "ld1 { v1.b }[6], [x21], #0x1\n"
    "ld1 { v0.b }[6], [x20], #0x1\n"
    "b 23f\n"
    "20:"  // Oddments: 4 inputs loop: Load: Bit 3: Unset: Bit 2: Bit 1: Unset
    "tbz %x[n_channels], #0, 23f\n"
    "ld1 { v3.b }[4], [x23], #0x1\n"
    "ld1 { v2.b }[4], [x22], #0x1\n"
    "ld1 { v1.b }[4], [x21], #0x1\n"
    "ld1 { v0.b }[4], [x20], #0x1\n"
    "b 23f\n"
    "21:"  // Oddments: 4 inputs loop: Load: Bit 3: Unset: Bit 2: Unset
    "tbz %x[n_channels], #1, 22f\n"
    "ldr h3, [x23], #0x2\n"
    "ldr h2, [x22], #0x2\n"
    "ldr h1, [x21], #0x2\n"
    "ldr h0, [x20], #0x2\n"
    "tbz %x[n_channels], #0, 23f\n"
    "ld1 { v3.b }[2], [x23], #0x1\n"
    "ld1 { v2.b }[2], [x22], #0x1\n"
    "ld1 { v1.b }[2], [x21], #0x1\n"
    "ld1 { v0.b }[2], [x20], #0x1\n"
    "b 23f\n"
    "22:"  // Oddments: 4 inputs loop: Load: Bit 3: Unset: Bit 2: Unset: Bit 1: Unset
    "tbz %x[n_channels], #0, 23f\n"
    "ldr b3, [x23], #0x1\n"
    "ldr b2, [x22], #0x1\n"
    "ldr b1, [x21], #0x1\n"
    "ldr b0, [x20], #0x1\n"
    "23:"  // Oddments: 4 inputs loop: Load: Bit 3: End
    "umax v23.16b, v3.16b, v2.16b\n"
    "subs x24, x24, #0x1\n"
    "umax v19.16b, v1.16b, v0.16b\n"
    "umax v19.16b, v23.16b, v19.16b\n"
    "umax v4.16b, v4.16b, v19.16b\n"
    "bgt 15b\n"
    "24:"  // Oddments: After loop
    "ands x20, %x[n_valid_cells], #0x3\n"
    "beq 34f\n"
    "25:"  // Oddments: Single input loop
    "movi v3.16b, #0x0\n"
    "ldr x23, [x19], #0x8\n"
    "add x23, x23, x28\n"
    "tbz %x[n_channels], #3, 29f\n"
    "ldr d3, [x23], #0x8\n"
    "tbz %x[n_channels], #2, 27f\n"
    "ld1 { v3.s }[2], [x23], #0x4\n"
    "tbz %x[n_channels], #1, 26f\n"
    "ld1 { v3.h }[6], [x23], #0x2\n"
    "tbz %x[n_channels], #0, 33f\n"
    "ld1 { v3.b }[14], [x23], #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 { v3.b }[12], [x23], #0x1\n"
    "b 33f\n"
    "27:"  // Oddments: Single input loop: Load: Bit 3: Bit 2: Unset
    "tbz %x[n_channels], #1, 28f\n"
    "ld1 { v3.h }[4], [x23], #0x2\n"
    "tbz %x[n_channels], #0, 33f\n"
    "ld1 { v3.b }[10], [x23], #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 { v3.b }[8], [x23], #0x1\n"
    "b 33f\n"
    "29:"  // Oddments: Single input loop: Load: Bit 3: Unset
    "tbz %x[n_channels], #2, 31f\n"
    "ldr s3, [x23], #0x4\n"
    "tbz %x[n_channels], #1, 30f\n"
    "ld1 { v3.h }[2], [x23], #0x2\n"
    "tbz %x[n_channels], #0, 33f\n"
    "ld1 { v3.b }[6], [x23], #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 { v3.b }[4], [x23], #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 h3, [x23], #0x2\n"
    "tbz %x[n_channels], #0, 33f\n"
    "ld1 { v3.b }[2], [x23], #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 b3, [x23], #0x1\n"
    "33:"  // Oddments: Single input loop: Load: Bit 3: End
    "umax v4.16b, v4.16b, v3.16b\n"
    "subs x20, x20, #0x1\n"
    "bgt 25b\n"
    "34:"  // Oddments: Single input loop: End
    "uxtl v17.8h, v4.8b\n"
    "add x19, %x[quant_params], %[offsetof_qp_input_offset]\n"
    "ld1r { v5.4s }, [x19]\n"
    "uxtl2 v16.8h, v4.16b\n"
    "add x19, %x[quant_params], %[offsetof_qp_per_layer_mul]\n"
    "movi v30.4s, #0x0\n"
    "ld1r { v4.4s }, [x19]\n"
    "add x19, %x[quant_params], %[offsetof_qp_per_layer_left_shift]\n"
    "movi v29.4s, #0xff\n"
    "ld1r { v3.4s }, [x19]\n"
    "add x19, %x[quant_params], %[offsetof_qp_per_layer_right_shift]\n"
    "neg v5.4s, v5.4s\n"
    "ld1r { v2.4s }, [x19]\n"
    "add x19, %x[quant_params], %[offsetof_qp_output_offset]\n"
    "saddw v23.4s, v5.4s, v17.4h\n"
    "ld1r { v1.4s }, [x19]\n"
    "saddw2 v18.4s, v5.4s, v17.8h\n"
    "saddw v17.4s, v5.4s, v16.4h\n"
    "saddw2 v16.4s, v5.4s, v16.8h\n"
    "srshl v23.4s, v23.4s, v3.4s\n"
    "srshl v18.4s, v18.4s, v3.4s\n"
    "srshl v17.4s, v17.4s, v3.4s\n"
    "srshl v16.4s, v16.4s, v3.4s\n"
    "sqrdmulh v23.4s, v23.4s, v4.4s\n"
    "sqrdmulh v18.4s, v18.4s, v4.4s\n"
    "sqrdmulh v17.4s, v17.4s, v4.4s\n"
    "sqrdmulh v16.4s, v16.4s, v4.4s\n"
    "srshl v23.4s, v23.4s, v2.4s\n"
    "srshl v18.4s, v18.4s, v2.4s\n"
    "srshl v17.4s, v17.4s, v2.4s\n"
    "srshl v16.4s, v16.4s, v2.4s\n"
    "add v23.4s, v23.4s, v1.4s\n"
    "add v18.4s, v18.4s, v1.4s\n"
    "add v17.4s, v17.4s, v1.4s\n"
    "add v16.4s, v16.4s, v1.4s\n"
    "smax v23.4s, v23.4s, v30.4s\n"
    "smax v18.4s, v18.4s, v30.4s\n"
    "smax v17.4s, v17.4s, v30.4s\n"
    "smin v23.4s, v23.4s, v29.4s\n"
    "smin v18.4s, v18.4s, v29.4s\n"
    "smin v17.4s, v17.4s, v29.4s\n"
    "smax v16.4s, v16.4s, v30.4s\n"
    "uzp1 v26.16b, v23.16b, v18.16b\n"
    "smin v16.4s, v16.4s, v29.4s\n"
    "uzp1 v24.16b, v17.16b, v16.16b\n"
    "uzp1 v16.16b, v26.16b, v24.16b\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)
    : [inptrs] "r" (inptrs), [n_valid_cells] "r" (n_valid_cells), [offsetof_qp_input_offset] "I" (offsetof(Requantize32, input_offset)), [offsetof_qp_output_offset] "I" (offsetof(Requantize32, output_offset)), [offsetof_qp_per_layer_left_shift] "I" (offsetof(Requantize32, per_layer_left_shift)), [offsetof_qp_per_layer_mul] "I" (offsetof(Requantize32, per_layer_mul)), [offsetof_qp_per_layer_right_shift] "I" (offsetof(Requantize32, per_layer_right_shift)), [quant_params] "r" (&qp)
    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
  );
}

}  // namespace pooling
}  // namespace arm_conv

#endif  // defined(__aarch64__)