// 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 CHANNEL_TILE % 4 == 0
$assert CHANNEL_TILE >= 4
$assert PIXEL_TILE == 1
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
#include <assert.h>

#include <wasm_simd128.h>

#include <xnnpack/ibilinear.h>


void xnn_f32_ibilinear_ukernel__wasmsimd_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) XNN_OOB_READS
{
  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 v128_t valphah = wasm_v128_load32_splat(weights);
    const v128_t valphav = wasm_v128_load32_splat(weights + 1);
    weights += 2;

    size_t c = channels;
    $if CHANNEL_TILE > 4:
      for (; c >= ${CHANNEL_TILE} * sizeof(float); c -= ${CHANNEL_TILE} * sizeof(float)) {
        const v128_t vtl${ABC[0:4]} = wasm_v128_load(i0);
        const v128_t vtr${ABC[0:4]} = wasm_v128_load(i1);
        const v128_t vbl${ABC[0:4]} = wasm_v128_load(i2);
        const v128_t vbr${ABC[0:4]} = wasm_v128_load(i3);
        $for C in range(4, CHANNEL_TILE, 4):
          const v128_t vtl${ABC[C:C+4]} = wasm_v128_load(i0 + ${C});
          const v128_t vtr${ABC[C:C+4]} = wasm_v128_load(i1 + ${C});
          const v128_t vbl${ABC[C:C+4]} = wasm_v128_load(i2 + ${C});
          const v128_t vbr${ABC[C:C+4]} = wasm_v128_load(i3 + ${C});
        i0 += ${CHANNEL_TILE};
        i1 += ${CHANNEL_TILE};
        i2 += ${CHANNEL_TILE};
        i3 += ${CHANNEL_TILE};

        $for C in range(0, CHANNEL_TILE, 4):
          const v128_t vtd${ABC[C:C+4]} = wasm_f32x4_sub(vtr${ABC[C:C+4]}, vtl${ABC[C:C+4]});
          const v128_t vbd${ABC[C:C+4]} = wasm_f32x4_sub(vbr${ABC[C:C+4]}, vbl${ABC[C:C+4]});

        $for C in range(0, CHANNEL_TILE, 4):
          const v128_t vt${ABC[C:C+4]} = wasm_f32x4_add(vtl${ABC[C:C+4]}, wasm_f32x4_mul(vtd${ABC[C:C+4]}, valphah));
          const v128_t vb${ABC[C:C+4]} = wasm_f32x4_add(vbl${ABC[C:C+4]}, wasm_f32x4_mul(vbd${ABC[C:C+4]}, valphah));

        $for C in range(0, CHANNEL_TILE, 4):
          const v128_t vd${ABC[C:C+4]} = wasm_f32x4_sub(vb${ABC[C:C+4]}, vt${ABC[C:C+4]});

        $for C in range(0, CHANNEL_TILE, 4):
          const v128_t vo${ABC[C:C+4]} = wasm_f32x4_add(vt${ABC[C:C+4]}, wasm_f32x4_mul(vd${ABC[C:C+4]}, valphav));

        wasm_v128_store(output, vo${ABC[0:4]});
        $for C in range(4, CHANNEL_TILE, 4):
          wasm_v128_store(output + ${C}, vo${ABC[C:C+4]});
        output += ${CHANNEL_TILE};
      }
    for (; c >= 4 * sizeof(float); c -= 4 * sizeof(float)) {
      const v128_t vtl = wasm_v128_load(i0);
      const v128_t vtr = wasm_v128_load(i1);
      const v128_t vbl = wasm_v128_load(i2);
      const v128_t vbr = wasm_v128_load(i3);
      i0 += 4;
      i1 += 4;
      i2 += 4;
      i3 += 4;

      const v128_t vtd = wasm_f32x4_sub(vtr, vtl);
      const v128_t vbd = wasm_f32x4_sub(vbr, vbl);
      const v128_t vt = wasm_f32x4_add(vtl, wasm_f32x4_mul(vtd, valphah));
      const v128_t vb = wasm_f32x4_add(vbl, wasm_f32x4_mul(vbd, valphah));
      const v128_t vd = wasm_f32x4_sub(vb, vt);
      const v128_t vo = wasm_f32x4_add(vt, wasm_f32x4_mul(vd, valphav));

      wasm_v128_store(output, vo);
      output += 4;
    }
    if XNN_UNLIKELY(c != 0) {
      const v128_t vtl = wasm_v128_load(i0);
      const v128_t vtr = wasm_v128_load(i1);
      const v128_t vbl = wasm_v128_load(i2);
      const v128_t vbr = wasm_v128_load(i3);

      const v128_t vtd = wasm_f32x4_sub(vtr, vtl);
      const v128_t vbd = wasm_f32x4_sub(vbr, vbl);
      const v128_t vt = wasm_f32x4_add(vtl, wasm_f32x4_mul(vtd, valphah));
      const v128_t vb = wasm_f32x4_add(vbl, wasm_f32x4_mul(vbd, valphah));
      const v128_t vd = wasm_f32x4_sub(vb, vt);
      v128_t vo = wasm_f32x4_add(vt, wasm_f32x4_mul(vd, valphav));

      if (c & (2 * sizeof(float))) {
        *((double*) output) = wasm_f64x2_extract_lane(vo, 0);
        vo = wasm_v32x4_shuffle(vo, vo, 2, 3, 2, 3);
        output += 2;
      }
      if (c & (1 * sizeof(float))) {
        *output++ = wasm_f32x4_extract_lane(vo, 0);
      }
    }

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