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


void xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_${ROW_TILE}x1${"_acc%d" % ACCUMULATORS if ACCUMULATORS > 1 else ""}(
    size_t input_height,
    size_t input_width,
    const float* input,
    const float* weights,
    const float* zero,
    float* output,
    uint32_t padding_top,
    const union xnn_f32_chw_params params[restrict XNN_MIN_ELEMENTS(1)])
{
  assert(input_height != 0);
  assert(input_width != 0);
  assert(input_width % sizeof(float) == 0);
  assert(padding_top >= 0);
  assert(padding_top <= 1);

  const float vmin = params->scalar.min;
  const float vmax = params->scalar.max;

  const float vbias = weights[0];
  const float vk00 = weights[1];
  const float vk01 = weights[2];
  const float vk02 = weights[3];
  const float vk10 = weights[4];
  const float vk11 = weights[5];
  const float vk12 = weights[6];
  const float vk20 = weights[7];
  const float vk21 = weights[8];
  const float vk22 = weights[9];

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

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

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

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

    $for M in range(1 + 2 * ROW_TILE):
      float vi${M}x0 = 0.0f;

    size_t w = input_width;
    for (; w >= 2 * sizeof(float); w -= 2 * sizeof(float)) {
      $for M in range(1 + 2 * ROW_TILE):
        const float vi${M}x1 = i${M}[0];

      $for K in range(3):
        $for M in range(ROW_TILE):
          $if K == 0:
            float vo${M}p0 = vbias + vi${2*M+K}x0 * vk${K}0;
          $elif K < ACCUMULATORS:
            float vo${M}p${K} = vi${2*M+K}x0 * vk${K}0;
          $else:
            vo${M}p${K % ACCUMULATORS} += vi${2*M+K}x0 * vk${K}0;

      $for M in range(1 + 2 * ROW_TILE):
        const float vi${M}x2 = i${M}[1];
        i${M} += 2;

      $for K in range(3):
        $for M in range(ROW_TILE):
          $if K + 3 < ACCUMULATORS:
            float vo${M}p${K+3} = vi${2*M+K}x1 * vk${K}1;
          $else:
            vo${M}p${(K+3) % ACCUMULATORS} += vi${2*M+K}x1 * vk${K}1;

      $for M in range(1 + 2 * ROW_TILE):
        vi${M}x0 = vi${M}x2;

      $for K in range(3):
        $for M in range(ROW_TILE):
          vo${M}p${(K+6) % ACCUMULATORS} += vi${2*M+K}x2 * vk${K}2;

      $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${A} += vo${M}p${A + ACC_SLICE};
          $ACC_SLICE *= 2

      $for M in range(ROW_TILE):
        float vo${M} = math_max_f32(vo${M}p0, vmin);

      $for M in range(ROW_TILE):
        vo${M} = math_min_f32(vo${M}, vmax);

      $for M in reversed(range(ROW_TILE)):
        *o${M}++ = vo${M};
    }
    // Potentially process the last pixel.
    assert(w <= 1 * sizeof(float));
    if (w != 0) {
      $for M in range(1 + 2 * ROW_TILE):
        const float vi${M}x1 = *i${M}++;

      $for K in range(3):
        $for M in range(ROW_TILE):
          $if K == 0:
            float vo${M}p0 = vbias + vi${2*M+K}x0 * vk${K}0;
          $elif K < ACCUMULATORS:
            float vo${M}p${K} = vi${2*M+K}x0 * vk${K}0;
          $else:
            vo${M}p${K % ACCUMULATORS} += vi${2*M+K}x0 * vk${K}0;

      $for K in range(3):
        $for M in range(ROW_TILE):
          $if K + 3 < ACCUMULATORS:
            float vo${M}p${K+3} = vi${2*M+K}x1 * vk${K}1;
          $else:
            vo${M}p${(K+3) % ACCUMULATORS} += vi${2*M+K}x1 * vk${K}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${A} += vo${M}p${A + ACC_SLICE};
          $ACC_SLICE *= 2

      $for M in range(ROW_TILE):
        float vo${M} = math_max_f32(vo${M}p0, vmin);

      $for M in range(ROW_TILE):
        vo${M} = math_min_f32(vo${M}, vmax);

      $for M in reversed(range(ROW_TILE)):
        *o${M}++ = vo${M};
    }

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