// Copyright 2021 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 = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
#include <assert.h>

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


$XINT8_T = {"S8": "int8_t", "U8": "uint8_t"}[DATATYPE]
void xnn_${DATATYPE.lower()}_ibilinear_ukernel__scalar_c${CHANNEL_TILE}${"" if PIXEL_TILE == 1 else "x%d" % PIXEL_TILE}(
    size_t output_pixels,
    size_t channels,
    const ${XINT8_T}**restrict input,
    size_t input_offset,
    const int16_t*restrict weights,
    ${XINT8_T}*restrict output,
    size_t output_increment)
{
  assert(output_pixels != 0);
  assert(channels != 0);

  do {
    const ${XINT8_T}* i0 = (const ${XINT8_T}*) ((uintptr_t) input[0] + input_offset);
    const ${XINT8_T}* i1 = (const ${XINT8_T}*) ((uintptr_t) input[1] + input_offset);
    const ${XINT8_T}* i2 = (const ${XINT8_T}*) ((uintptr_t) input[2] + input_offset);
    const ${XINT8_T}* i3 = (const ${XINT8_T}*) ((uintptr_t) input[3] + input_offset);
    input += 4;

    const int32_t valphah = (int32_t) (uint32_t) (uint16_t) weights[0];
    const int32_t valphav = (int32_t) (uint32_t) (uint16_t) weights[1];
    weights += 2;

    const int32_t vrounding = INT32_C(0x00200000);

    size_t c = channels;
    $if CHANNEL_TILE > 1:
      for (; c >= ${CHANNEL_TILE} * sizeof(${XINT8_T}); c -= ${CHANNEL_TILE} * sizeof(${XINT8_T})) {
        $for C in range(CHANNEL_TILE):
          const int32_t vtl${ABC[C]} = (int32_t) i0[${C}];
          const int32_t vtr${ABC[C]} = (int32_t) i1[${C}];
          const int32_t vbl${ABC[C]} = (int32_t) i2[${C}];
          const int32_t vbr${ABC[C]} = (int32_t) i3[${C}];
        i0 += ${CHANNEL_TILE};
        i1 += ${CHANNEL_TILE};
        i2 += ${CHANNEL_TILE};
        i3 += ${CHANNEL_TILE};

        $for C in range(CHANNEL_TILE):
          const int32_t vtd${ABC[C]} = vtr${ABC[C]} - vtl${ABC[C]};
          const int32_t vbd${ABC[C]} = vbr${ABC[C]} - vbl${ABC[C]};

        $for C in range(CHANNEL_TILE):
          const int32_t vt${ABC[C]} = (int32_t) ((uint32_t) vtl${ABC[C]} << 11) + vtd${ABC[C]} * valphah;
          const int32_t vb${ABC[C]} = (int32_t) ((uint32_t) vbl${ABC[C]} << 11) + vbd${ABC[C]} * valphah;

        $for C in range(CHANNEL_TILE):
          const int32_t vd${ABC[C]} = vb${ABC[C]} - vt${ABC[C]};

        $for C in range(CHANNEL_TILE):
          const int32_t vacc${ABC[C]} = (int32_t) ((uint32_t) vt${ABC[C]} << 11) + vd${ABC[C]} * valphav;

        $for C in range(CHANNEL_TILE):
          const int32_t vo${ABC[C]} = math_asr_s32(vacc${ABC[C]} + vrounding, 22);

        $for C in range(CHANNEL_TILE):
          output[${C}] = (${XINT8_T}) vo${ABC[C]};
        output += ${CHANNEL_TILE};
      }
      for (; c >= sizeof(${XINT8_T}); c -= sizeof(${XINT8_T})) {
        const int32_t vtl = (int32_t) *i0++;
        const int32_t vtr = (int32_t) *i1++;
        const int32_t vbl = (int32_t) *i2++;
        const int32_t vbr = (int32_t) *i3++;

        const int32_t vtd = vtr - vtl;
        const int32_t vbd = vbr - vbl;

        const int32_t vt = (int32_t) ((uint32_t) vtl << 11) + vtd * valphah;
        const int32_t vb = (int32_t) ((uint32_t) vbl << 11) + vbd * valphah;

        const int32_t vd = vb - vt;

        const int32_t vacc = (int32_t) ((uint32_t) vt << 11) + vd * valphav;

        const int32_t vo = math_asr_s32(vacc + vrounding, 22);

        *output++ = vo;
      }
    $else:
      do {
        const int32_t vtl = (int32_t) *i0++;
        const int32_t vtr = (int32_t) *i1++;
        const int32_t vbl = (int32_t) *i2++;
        const int32_t vbr = (int32_t) *i3++;

        const int32_t vtd = vtr - vtl;
        const int32_t vbd = vbr - vbl;

        const int32_t vt = (int32_t) ((uint32_t) vtl << 11) + vtd * valphah;
        const int32_t vb = (int32_t) ((uint32_t) vbl << 11) + vbd * valphah;

        const int32_t vd = vb - vt;

        const int32_t vacc = (int32_t) ((uint32_t) vt << 11) + vd * valphav;

        const int32_t vo = math_asr_s32(vacc + vrounding, 22);

        *output++ = vo;

        c -= sizeof(${XINT8_T});
      } while (c != 0);

    output = (${XINT8_T}*) ((uintptr_t) output + output_increment);
  } while (--output_pixels != 0);
}