// 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 ROW_TILE >= 1
$assert ACCUMULATORS >= 1
#include <assert.h>

#include <arm_neon.h>

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


void xnn_f16_dwconv2d_chw_ukernel_5x5s2p2__neonfp16arith_${ROW_TILE}x4${"_acc%d" % ACCUMULATORS if ACCUMULATORS > 1 else ""}(
    size_t input_height,
    size_t input_width,
    const void* input,
    const void* weights,
    const void* zero,
    void* output,
    uint32_t padding_top,
    const union xnn_f16_chw_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(input_height != 0);
  assert(input_width != 0);
  assert(input_width % sizeof(__fp16) == 0);
  assert(padding_top >= 1);
  assert(padding_top <= 2);

  const uint16x4_t vmask_even = vld1_u16(params->neonfp16arith.mask_even);
  const uint16x4_t vmask_odd = vld1_u16(params->neonfp16arith.mask_odd);
  const float16x4_t vmax = vld1_dup_f16(&params->neonfp16arith.max);
  const float16x4_t vmin = vld1_dup_f16(&params->neonfp16arith.min);

  const __fp16* w0 = (const __fp16*)weights;
  const float16x8_t vw01234567 = vld1q_f16(w0);
  const float16x8_t vw89ABCDEF = vld1q_f16(w0 + 8);
  const float16x8_t vwGHIJKLMN = vld1q_f16(w0 + 16);
  const float16x4_t vwOP = vreinterpret_f16_u32(vld1_lane_u32((const void*)(w0 + 24), vmov_n_u32(0), 0));

  const uint32_t padding_top_less_1 = padding_top - 1;
  const size_t input_decrement = round_up_po2(input_width, 8 * sizeof(__fp16));

  const __fp16* i0 = zero;
  const __fp16* i1 = (const __fp16*) ((uintptr_t) input - ((-padding_top_less_1) & input_width));
  const __fp16* i2 = (const __fp16*) ((uintptr_t) i1 + input_width);
  if XNN_UNPREDICTABLE(padding_top_less_1 != 0) {
    i1 = zero;
  }
  $for M in range(3, 3 + 2 * ROW_TILE):
    const __fp16* i${M} = (const __fp16*) ((uintptr_t) i${M-1} + input_width);

  $if ROW_TILE > 1:
    const size_t output_width = round_down_po2((input_width + (2 /* padding */ - 3 /* kernel size */ + 2 /* subsampling */) * sizeof(__fp16)) / 2, sizeof(__fp16));

  __fp16* o0 = output;
  $for M in range(1, ROW_TILE):
    __fp16* o${M} = (__fp16*) ((uintptr_t) o${M-1} + output_width);

  size_t padded_input_height = input_height + (padding_top_less_1 + 1) + 2 /* padding bottom */;
  size_t output_height = (padded_input_height - 5 /* kernel size */ + 2 /* subsampling */) / 2;
  do {
    $for M in range(3, 3 + 2 * ROW_TILE):
      if XNN_UNPREDICTABLE(padded_input_height < ${3 + M}) {
        i${M} = zero;
        $if M % 2 == 0 and M <= 2 * ROW_TILE + 1:
          o${M // 2 - 1} = o${M // 2 - 2};
      }

    $for M in range(3 + 2 * ROW_TILE):
      float16x4_t vi${M}x0246 = vmov_n_f16(0.0f);

    $for M in range(3 + 2 * ROW_TILE):
      float16x4_t vi${M}x1357 = vmov_n_f16(0.0f);

    $for M in range(3 + 2 * ROW_TILE):
      float16x4x2_t vi${M}x8ACE9BDF = vld2_f16(i${M}); i${M} += 8;

    size_t w = input_width;
    for (; w > 8 * sizeof(__fp16); w -= 8 * sizeof(__fp16)) {
      $for M in range(ROW_TILE):
        float16x4_t vo${M}p0 = vdup_laneq_f16(vw01234567, 0);

      $for M in range(ROW_TILE):
        $if ACCUMULATORS > 1:
          float16x4_t vo${M}p1 = vmul_laneq_f16(vi${2*M}x8ACE9BDF.val[0], vw01234567, 3);
        $else:
          vo${M}p0 = vfma_laneq_f16(vo${M}p0, vi${2*M}x8ACE9BDF.val[0], vw01234567, 3);

      $for M in range(ROW_TILE):
        $if ACCUMULATORS > 2:
          float16x4_t vo${M}p2 = vmul_laneq_f16(vi${2*M+1}x8ACE9BDF.val[0], vw89ABCDEF, 0);
        $else:
          vo${M}p0 = vfma_laneq_f16(vo${M}p0, vi${2*M+1}x8ACE9BDF.val[0], vw89ABCDEF, 0);

      $for M in range(ROW_TILE):
        $if ACCUMULATORS > 3:
          float16x4_t vo${M}p3 = vmul_laneq_f16(vi${2*M+2}x8ACE9BDF.val[0], vw89ABCDEF, 5);
        $else:
          vo${M}p${4 % ACCUMULATORS} = vfma_laneq_f16(vo${M}p${4 % ACCUMULATORS}, vi${2*M+2}x8ACE9BDF.val[0], vw89ABCDEF, 5);

      $for M in range(ROW_TILE):
        $if ACCUMULATORS > 4:
          float16x4_t vo${M}p4 = vmul_laneq_f16(vi${2*M+3}x8ACE9BDF.val[0], vwGHIJKLMN, 2);
        $else:
          vo${M}p${5 % ACCUMULATORS} = vfma_laneq_f16(vo${M}p${5 % ACCUMULATORS}, vi${2*M+3}x8ACE9BDF.val[0], vwGHIJKLMN, 2);

      $for M in range(ROW_TILE):
        $if ACCUMULATORS > 5:
          vo${M}p5 = vmul_laneq_f16(vi${2*M+4}x8ACE9BDF.val[0], vwGHIJKLMN, 7);
        $else:
          vo${M}p${6 % ACCUMULATORS} = vfma_laneq_f16(vo${M}p${6 % ACCUMULATORS}, vi${2*M+4}x8ACE9BDF.val[0], vwGHIJKLMN, 7);

      $for M in range(ROW_TILE):
        vo${M}p${7 % ACCUMULATORS} = vfma_laneq_f16(vo${M}p${7 % ACCUMULATORS}, vi${2*M}x8ACE9BDF.val[1], vw01234567, 4);

      $for M in range(ROW_TILE):
        vo${M}p${8 % ACCUMULATORS} = vfma_laneq_f16(vo${M}p${8 % ACCUMULATORS}, vi${2*M+1}x8ACE9BDF.val[1], vw89ABCDEF, 1);

      $for M in range(ROW_TILE):
        vo${M}p${9 % ACCUMULATORS} = vfma_laneq_f16(vo${M}p${9 % ACCUMULATORS}, vi${2*M+2}x8ACE9BDF.val[1], vw89ABCDEF, 6);

      $for M in range(ROW_TILE):
        vo${M}p${10 % ACCUMULATORS} = vfma_laneq_f16(vo${M}p${10 % ACCUMULATORS}, vi${2*M+3}x8ACE9BDF.val[1], vwGHIJKLMN, 3);

      $for M in range(ROW_TILE):
        vo${M}p${11 % ACCUMULATORS} = vfma_lane_f16(vo${M}p${11 % ACCUMULATORS}, vi${2*M+4}x8ACE9BDF.val[1], vwOP, 0);

      $for M in range(3 + 2 * ROW_TILE):
        const float16x4_t vi${M}x68AC = vext_f16(vi${M}x0246, vi${M}x8ACE9BDF.val[0], 3);
        vi${M}x0246 = vi${M}x8ACE9BDF.val[0];

      $for M in range(ROW_TILE):
        vo${M}p${12 % ACCUMULATORS} = vfma_laneq_f16(vo${M}p${12 % ACCUMULATORS}, vi${2*M}x68AC, vw01234567, 1);

      $for M in range(ROW_TILE):
        vo${M}p${13 % ACCUMULATORS} = vfma_laneq_f16(vo${M}p${13 % ACCUMULATORS}, vi${2*M+1}x68AC, vw01234567, 6);

      $for M in range(ROW_TILE):
        vo${M}p${14 % ACCUMULATORS} = vfma_laneq_f16(vo${M}p${14 % ACCUMULATORS}, vi${2*M+2}x68AC, vw89ABCDEF, 3);

      $for M in range(ROW_TILE):
        vo${M}p${15 % ACCUMULATORS} = vfma_laneq_f16(vo${M}p${15 % ACCUMULATORS}, vi${2*M+3}x68AC, vwGHIJKLMN, 0);

      $for M in range(ROW_TILE):
        vo${M}p${16 % ACCUMULATORS} = vfma_laneq_f16(vo${M}p${16 % ACCUMULATORS}, vi${2*M+4}x68AC, vwGHIJKLMN, 5);

      $for M in range(3 + 2 * ROW_TILE):
        const float16x4_t vi${M}x79BD = vext_f16(vi${M}x1357, vi${M}x8ACE9BDF.val[1], 3);
        vi${M}x1357 = vi${M}x8ACE9BDF.val[1];

      $for M in range(3 + 2 * ROW_TILE):
        const float16x4x2_t vi${M}xGIKMHJLN = vld2_f16(i${M}); i${M} += 8;

      $for M in range(ROW_TILE):
        vo${M}p${17 % ACCUMULATORS} = vfma_laneq_f16(vo${M}p${17 % ACCUMULATORS}, vi${2*M}x79BD, vw01234567, 2);

      $for M in range(ROW_TILE):
        vo${M}p${18 % ACCUMULATORS} = vfma_laneq_f16(vo${M}p${18 % ACCUMULATORS}, vi${2*M+1}x79BD, vw01234567, 7);

      $for M in range(ROW_TILE):
        vo${M}p${19 % ACCUMULATORS} = vfma_laneq_f16(vo${M}p${19 % ACCUMULATORS}, vi${2*M+2}x79BD, vw89ABCDEF, 4);

      $for M in range(ROW_TILE):
        vo${M}p${20 % ACCUMULATORS} = vfma_laneq_f16(vo${M}p${20 % ACCUMULATORS}, vi${2*M+3}x79BD, vwGHIJKLMN, 1);

      $for M in range(ROW_TILE):
        vo${M}p${21 % ACCUMULATORS} = vfma_laneq_f16(vo${M}p${21 % ACCUMULATORS}, vi${2*M+4}x79BD, vwGHIJKLMN, 6);

      $for M in range(3 + 2 * ROW_TILE):
        const float16x4_t vi${M}xACEG = vext_f16(vi${M}x8ACE9BDF.val[0], vi${M}xGIKMHJLN.val[0], 1);
        vi${M}x8ACE9BDF = vi${M}xGIKMHJLN;

      $for M in range(ROW_TILE):
        vo${M}p${22 % ACCUMULATORS} = vfma_laneq_f16(vo${M}p${22 % ACCUMULATORS}, vi${2*M}xACEG, vw01234567, 5);

      $for M in range(ROW_TILE):
        vo${M}p${23 % ACCUMULATORS} = vfma_laneq_f16(vo${M}p${23 % ACCUMULATORS}, vi${2*M+1}xACEG, vw89ABCDEF, 2);

      $for M in range(ROW_TILE):
        vo${M}p${24 % ACCUMULATORS} = vfma_laneq_f16(vo${M}p${24 % ACCUMULATORS}, vi${2*M+2}xACEG, vw89ABCDEF, 7);

      $for M in range(ROW_TILE):
        vo${M}p${25 % ACCUMULATORS} = vfma_laneq_f16(vo${M}p${25 % ACCUMULATORS}, vi${2*M+3}xACEG, vwGHIJKLMN, 4);

      $for M in range(ROW_TILE):
        vo${M}p${26 % ACCUMULATORS} = vfma_lane_f16(vo${M}p${26 % ACCUMULATORS}, vi${2*M+4}xACEG, vwOP, 1);

      $if ACCUMULATORS > 1:
        $ACC_SLICE = 1
        $while ACC_SLICE < ACCUMULATORS:
          $for A in range(0, ACCUMULATORS, ACC_SLICE * 2):
            $if A + ACC_SLICE < ACCUMULATORS:
              $for M in range(ROW_TILE):
                vo${M}p${A} = vadd_f16(vo${M}p${A}, vo${M}p${A + ACC_SLICE});
          $ACC_SLICE *= 2

      $for M in range(ROW_TILE):
        float16x4_t vo${M} = vmax_f16(vo${M}p0, vmin);

      $for M in range(ROW_TILE):
        vo${M} = vmin_f16(vo${M}, vmax);

      $for M in reversed(range(ROW_TILE)):
        vst1_f16(o${M}, vo${M}); o${M} += 4;
    }
    // Last block has 1-8 pixels to process.
    assert(w <= 8 * sizeof(__fp16));
    assert(w >= 1 * sizeof(__fp16));
    {
      $for M in range(ROW_TILE):
        float16x4_t vo${M}p0 = vdup_laneq_f16(vw01234567, 0);

      $for M in range(3 + 2 * ROW_TILE):
        const float16x4_t vi${M}x8ACE = vreinterpret_f16_u16(vand_u16(vmask_even, vreinterpret_u16_f16(vi${M}x8ACE9BDF.val[0])));

      $for M in range(3 + 2 * ROW_TILE):
        const float16x4_t vi${M}x9BDF = vreinterpret_f16_u16(vand_u16(vmask_odd, vreinterpret_u16_f16(vi${M}x8ACE9BDF.val[1])));

      $for M in range(ROW_TILE):
        $if ACCUMULATORS > 1:
          float16x4_t vo${M}p1 = vmul_laneq_f16(vi${2*M}x8ACE, vw01234567, 3);
        $else:
          vo${M}p0 = vfma_laneq_f16(vo${M}p0, vi${2*M}x8ACE, vw01234567, 3);

      $for M in range(ROW_TILE):
        $if ACCUMULATORS > 2:
          float16x4_t vo${M}p2 = vmul_laneq_f16(vi${2*M+1}x8ACE, vw89ABCDEF, 0);
        $else:
          vo${M}p0 = vfma_laneq_f16(vo${M}p0, vi${2*M+1}x8ACE, vw89ABCDEF, 0);

      $for M in range(ROW_TILE):
        $if ACCUMULATORS > 3:
          float16x4_t vo${M}p3 = vmul_laneq_f16(vi${2*M+2}x8ACE, vw89ABCDEF, 5);
        $else:
          vo${M}p${4 % ACCUMULATORS} = vfma_laneq_f16(vo${M}p${4 % ACCUMULATORS}, vi${2*M+2}x8ACE, vw89ABCDEF, 5);

      $for M in range(ROW_TILE):
        $if ACCUMULATORS > 4:
          float16x4_t vo${M}p4 = vmul_laneq_f16(vi${2*M+3}x8ACE, vwGHIJKLMN, 2);
        $else:
          vo${M}p${5 % ACCUMULATORS} = vfma_laneq_f16(vo${M}p${5 % ACCUMULATORS}, vi${2*M+3}x8ACE, vwGHIJKLMN, 2);

      $for M in range(ROW_TILE):
        $if ACCUMULATORS > 5:
          vo${M}p5 = vmul_laneq_f16(vi${2*M+4}x8ACE, vwGHIJKLMN, 7);
        $else:
          vo${M}p${6 % ACCUMULATORS} = vfma_laneq_f16(vo${M}p${6 % ACCUMULATORS}, vi${2*M+4}x8ACE, vwGHIJKLMN, 7);

      $for M in range(ROW_TILE):
        vo${M}p${7 % ACCUMULATORS} = vfma_laneq_f16(vo${M}p${7 % ACCUMULATORS}, vi${2*M}x9BDF, vw01234567, 4);

      $for M in range(ROW_TILE):
        vo${M}p${8 % ACCUMULATORS} = vfma_laneq_f16(vo${M}p${8 % ACCUMULATORS}, vi${2*M+1}x9BDF, vw89ABCDEF, 1);

      $for M in range(ROW_TILE):
        vo${M}p${9 % ACCUMULATORS} = vfma_laneq_f16(vo${M}p${9 % ACCUMULATORS}, vi${2*M+2}x9BDF, vw89ABCDEF, 6);

      $for M in range(ROW_TILE):
        vo${M}p${10 % ACCUMULATORS} = vfma_laneq_f16(vo${M}p${10 % ACCUMULATORS}, vi${2*M+3}x9BDF, vwGHIJKLMN, 3);

      $for M in range(ROW_TILE):
        vo${M}p${11 % ACCUMULATORS} = vfma_lane_f16(vo${M}p${11 % ACCUMULATORS}, vi${2*M+4}x9BDF, vwOP, 0);

      $for M in range(3 + 2 * ROW_TILE):
        const float16x4_t vi${M}x68AC = vext_f16(vi${M}x0246, vi${M}x8ACE, 3);

      $for M in range(ROW_TILE):
        vo${M}p${12 % ACCUMULATORS} = vfma_laneq_f16(vo${M}p${12 % ACCUMULATORS}, vi${2*M}x68AC, vw01234567, 1);

      $for M in range(ROW_TILE):
        vo${M}p${13 % ACCUMULATORS} = vfma_laneq_f16(vo${M}p${13 % ACCUMULATORS}, vi${2*M+1}x68AC, vw01234567, 6);

      $for M in range(ROW_TILE):
        vo${M}p${14 % ACCUMULATORS} = vfma_laneq_f16(vo${M}p${14 % ACCUMULATORS}, vi${2*M+2}x68AC, vw89ABCDEF, 3);

      $for M in range(ROW_TILE):
        vo${M}p${15 % ACCUMULATORS} = vfma_laneq_f16(vo${M}p${15 % ACCUMULATORS}, vi${2*M+3}x68AC, vwGHIJKLMN, 0);

      $for M in range(ROW_TILE):
        vo${M}p${16 % ACCUMULATORS} = vfma_laneq_f16(vo${M}p${16 % ACCUMULATORS}, vi${2*M+4}x68AC, vwGHIJKLMN, 5);

      $for M in range(3 + 2 * ROW_TILE):
        const float16x4_t vi${M}x79BD = vext_f16(vi${M}x1357, vi${M}x9BDF, 3);

      $for M in range(ROW_TILE):
        vo${M}p${17 % ACCUMULATORS} = vfma_laneq_f16(vo${M}p${17 % ACCUMULATORS}, vi${2*M}x79BD, vw01234567, 2);

      $for M in range(ROW_TILE):
        vo${M}p${18 % ACCUMULATORS} = vfma_laneq_f16(vo${M}p${18 % ACCUMULATORS}, vi${2*M+1}x79BD, vw01234567, 7);

      $for M in range(ROW_TILE):
        vo${M}p${19 % ACCUMULATORS} = vfma_laneq_f16(vo${M}p${19 % ACCUMULATORS}, vi${2*M+2}x79BD, vw89ABCDEF, 4);

      $for M in range(ROW_TILE):
        vo${M}p${20 % ACCUMULATORS} = vfma_laneq_f16(vo${M}p${20 % ACCUMULATORS}, vi${2*M+3}x79BD, vwGHIJKLMN, 1);

      $for M in range(ROW_TILE):
        vo${M}p${21 % ACCUMULATORS} = vfma_laneq_f16(vo${M}p${21 % ACCUMULATORS}, vi${2*M+4}x79BD, vwGHIJKLMN, 6);

      const float16x4_t vzero = vmov_n_f16(0.0f);
      $for M in range(3 + 2 * ROW_TILE):
        const float16x4_t vi${M}xACEG = vext_f16(vi${M}x8ACE, vzero, 1);

      $for M in range(ROW_TILE):
        vo${M}p${22 % ACCUMULATORS} = vfma_laneq_f16(vo${M}p${22 % ACCUMULATORS}, vi${2*M}xACEG, vw01234567, 5);

      $for M in range(ROW_TILE):
        vo${M}p${23 % ACCUMULATORS} = vfma_laneq_f16(vo${M}p${23 % ACCUMULATORS}, vi${2*M+1}xACEG, vw89ABCDEF, 2);

      $for M in range(ROW_TILE):
        vo${M}p${24 % ACCUMULATORS} = vfma_laneq_f16(vo${M}p${24 % ACCUMULATORS}, vi${2*M+2}xACEG, vw89ABCDEF, 7);

      $for M in range(ROW_TILE):
        vo${M}p${25 % ACCUMULATORS} = vfma_laneq_f16(vo${M}p${25 % ACCUMULATORS}, vi${2*M+3}xACEG, vwGHIJKLMN, 4);

      $for M in range(ROW_TILE):
        vo${M}p${26 % ACCUMULATORS} = vfma_lane_f16(vo${M}p${26 % ACCUMULATORS}, vi${2*M+4}xACEG, vwOP, 1);

      $if ACCUMULATORS > 1:
        $ACC_SLICE = 1
        $while ACC_SLICE < ACCUMULATORS:
          $for A in range(0, ACCUMULATORS, ACC_SLICE * 2):
            $if A + ACC_SLICE < ACCUMULATORS:
              $for M in range(ROW_TILE):
                vo${M}p${A} = vadd_f16(vo${M}p${A}, vo${M}p${A + ACC_SLICE});
          $ACC_SLICE *= 2

      $for M in range(ROW_TILE):
        float16x4_t vo${M} = vmax_f16(vo${M}p0, vmin);

      $for M in range(ROW_TILE):
        vo${M} = vmin_f16(vo${M}, vmax);

      size_t w_tmp = (w + 1 * sizeof(__fp16)) / (2 * sizeof(__fp16));
      if XNN_LIKELY(w_tmp >= 4) {
        $for M in reversed(range(ROW_TILE)):
          vst1_f16(o${M}, vo${M}); o${M} += 4;
      } else {
        if (w_tmp & 2) {
          $for M in reversed(range(ROW_TILE)):
            vst1_lane_u32((void*) o${M}, vreinterpret_u32_f16(vo${M}), 0); o${M} += 2;

          $for M in range(ROW_TILE):
            vo${M} = vext_f16(vo${M}, vo${M}, 2);
        }
        if (w_tmp & 1) {
          $for M in reversed(range(ROW_TILE)):
            vst1_lane_f16(o${M}, vo${M}, 0); o${M} += 1;
        }
      }
    }

    i0 = (const __fp16*) ((uintptr_t) i${2 * ROW_TILE} - input_decrement);
    i1 = (const __fp16*) ((uintptr_t) i${2 * ROW_TILE + 1} - input_decrement);
    i2 = (const __fp16*) ((uintptr_t) i${2 * ROW_TILE + 2} - input_decrement);
    $for M in range(3, 3 + 2 * ROW_TILE):
      i${M} = (const __fp16*) ((uintptr_t) i${M-1} + input_width);

    $if ROW_TILE > 1:
      o0 = o${ROW_TILE - 1};
      $for M in range(1, ROW_TILE):
        o${M} = (__fp16*) ((uintptr_t) o${M-1} + output_width);

    $if ROW_TILE > 1:
      output_height = doz(output_height, ${ROW_TILE});
      padded_input_height = doz(padded_input_height, ${ROW_TILE * 2});
    $else:
      output_height -= 1;
      padded_input_height -= 2;
  } while (output_height != 0);
}