// Copyright 2019 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 >= 1
$assert PIXEL_TILE == 1
$ABC = "0123456789ABCDEFGHIJKLMN"
#include <assert.h>

#include <xnnpack/bilinear.h>


void xnn_f32_bilinear_ukernel__scalar_c${CHANNEL_TILE}${"" if PIXEL_TILE == 1 else "x%d" % PIXEL_TILE}(
    size_t output_pixels,
    size_t channels,
    const float**restrict input,
    size_t input_offset,
    const float*restrict weights,
    float*restrict output,
    size_t output_increment)
{
  assert(output_pixels != 0);
  assert(channels != 0);
  assert(channels % sizeof(float) == 0);

  do {
    const float* i0 = (const float*) ((uintptr_t) input[0] + input_offset);
    const float* i1 = (const float*) ((uintptr_t) input[1] + input_offset);
    const float* i2 = (const float*) ((uintptr_t) input[2] + input_offset);
    const float* i3 = (const float*) ((uintptr_t) input[3] + input_offset);
    input += 4;

    const float valphah = weights[0];
    const float valphav = weights[1];
    weights += 2;

    size_t c = channels;
    $if CHANNEL_TILE > 1:
      for (; c >= ${CHANNEL_TILE} * sizeof(float); c -= ${CHANNEL_TILE} * sizeof(float)) {
        $for C in range(CHANNEL_TILE):
          const float vtl${ABC[C]} = i0[${C}];
          const float vtr${ABC[C]} = i1[${C}];
          const float vbl${ABC[C]} = i2[${C}];
          const float vbr${ABC[C]} = i3[${C}];
        i0 += ${CHANNEL_TILE};
        i1 += ${CHANNEL_TILE};
        i2 += ${CHANNEL_TILE};
        i3 += ${CHANNEL_TILE};

        $for C in range(CHANNEL_TILE):
          const float vtd${ABC[C]} = vtr${ABC[C]} - vtl${ABC[C]};
          const float vbd${ABC[C]} = vbr${ABC[C]} - vbl${ABC[C]};

        $for C in range(CHANNEL_TILE):
          const float vt${ABC[C]} = vtl${ABC[C]} + vtd${ABC[C]} * valphah;
          const float vb${ABC[C]} = vbl${ABC[C]} + vbd${ABC[C]} * valphah;

        $for C in range(CHANNEL_TILE):
          const float vd${ABC[C]} = vb${ABC[C]} - vt${ABC[C]};

        $for C in range(CHANNEL_TILE):
          const float vo${ABC[C]} = vt${ABC[C]} + vd${ABC[C]} * valphav;

        $for C in range(CHANNEL_TILE):
          output[${C}] = vo${ABC[C]};
        output += ${CHANNEL_TILE};
      }
      for (; c >= sizeof(float); c -= sizeof(float)) {
        const float vtl = *i0++;
        const float vtr = *i1++;
        const float vbl = *i2++;
        const float vbr = *i3++;

        const float vtd = vtr - vtl;
        const float vbd = vbr - vbl;

        const float vt = vtl + vtd * valphah;
        const float vb = vbl + vbd * valphah;

        const float vd = vb - vt;

        const float vo = vt + vd * valphav;

        *output++ = vo;
      }
    $else:
      do {
        const float vtl = *i0++;
        const float vtr = *i1++;
        const float vbl = *i2++;
        const float vbr = *i3++;

        const float vtd = vtr - vtl;
        const float vbd = vbr - vbl;

        const float vt = vtl + vtd * valphah;
        const float vb = vbl + vbd * valphah;

        const float vd = vb - vt;

        const float vo = vt + vd * valphav;

        *output++ = vo;

        c -= sizeof(float);
      } while (c != 0);

    output = (float*) ((uintptr_t) output + output_increment);
  } while (--output_pixels != 0);
}