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

#include <wasm_simd128.h>

#include <xnnpack/ibilinear.h>


void xnn_f32_ibilinear_chw_ukernel__wasmsimd_p${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 input_increment) XNN_DISABLE_TSAN
{
  assert(output_pixels != 0);
  assert(channels != 0);
  assert(input_increment % sizeof(float) == 0);

  do {
    const float** i = input;
    const float* w = weights;
    size_t p = output_pixels;
    $if PIXEL_TILE > 4:
      for (; p >= ${PIXEL_TILE}; p -= ${PIXEL_TILE}) {
        $for P in range(PIXEL_TILE):
          const float* itl${ABC[P]} = (const float*) ((uintptr_t) i[${2 * P}] + input_offset);
          const float* ibl${ABC[P]} = (const float*) ((uintptr_t) i[${2 * P + 1}] + input_offset);
        i += 2 * ${PIXEL_TILE};

        $for P in range(0, PIXEL_TILE, 4):
          const v128_t vw${ABC[P:P+4]}p0 = wasm_v128_load(w + ${2 * P});
          const v128_t vw${ABC[P:P+4]}p1 = wasm_v128_load(w + ${2 * P + 4});
        w += 2 * ${PIXEL_TILE};

        $for P in range(0, PIXEL_TILE, 2):
          const v128_t vtltr${ABC[P]} = wasm_v64x2_load_splat(itl${ABC[P]});
          const v128_t vblbr${ABC[P]} = wasm_v64x2_load_splat(ibl${ABC[P]});
          const double vtltr${ABC[P+1]} = *((const double*) itl${ABC[P+1]});
          const double vblbr${ABC[P+1]} = *((const double*) ibl${ABC[P+1]});

        $for P in range(0, PIXEL_TILE, 4):
          const v128_t valphah${ABC[P:P+4]} = wasm_v32x4_shuffle(vw${ABC[P:P+4]}p0, vw${ABC[P:P+4]}p1, 0, 2, 4, 6);
          const v128_t valphav${ABC[P:P+4]} = wasm_v32x4_shuffle(vw${ABC[P:P+4]}p0, vw${ABC[P:P+4]}p1, 1, 3, 5, 7);

        $for P in range(0, PIXEL_TILE, 2):
          const v128_t vtltr${ABC[P:P+2]} = wasm_f64x2_replace_lane(vtltr${ABC[P]}, 1, vtltr${ABC[P+1]});
          const v128_t vblbr${ABC[P:P+2]} = wasm_f64x2_replace_lane(vblbr${ABC[P]}, 1, vblbr${ABC[P+1]});

        $for P in range(0, PIXEL_TILE, 2):
          const v128_t vldrd${ABC[P:P+2]} = wasm_f32x4_sub(vblbr${ABC[P:P+2]}, vtltr${ABC[P:P+2]});

        $for P in range(0, PIXEL_TILE, 4):
          const v128_t vld${ABC[P:P+4]} = wasm_v32x4_shuffle(vldrd${ABC[P:P+2]}, vldrd${ABC[P+2:P+4]}, 0, 2, 4, 6);
          const v128_t vrd${ABC[P:P+4]} = wasm_v32x4_shuffle(vldrd${ABC[P:P+2]}, vldrd${ABC[P+2:P+4]}, 1, 3, 5, 7);

        $for P in range(0, PIXEL_TILE, 4):
          const v128_t vtl${ABC[P:P+4]} = wasm_v32x4_shuffle(vtltr${ABC[P:P+2]}, vtltr${ABC[P+2:P+4]}, 0, 2, 4, 6);
          const v128_t vtr${ABC[P:P+4]} = wasm_v32x4_shuffle(vtltr${ABC[P:P+2]}, vtltr${ABC[P+2:P+4]}, 1, 3, 5, 7);

        $for P in range(0, PIXEL_TILE, 4):
          const v128_t vl${ABC[P:P+4]} = wasm_f32x4_add(vtl${ABC[P:P+4]}, wasm_f32x4_mul(vld${ABC[P:P+4]}, valphav${ABC[P:P+4]}));
          const v128_t vr${ABC[P:P+4]} = wasm_f32x4_add(vtr${ABC[P:P+4]}, wasm_f32x4_mul(vrd${ABC[P:P+4]}, valphav${ABC[P:P+4]}));

        $for P in range(0, PIXEL_TILE, 4):
          const v128_t vd${ABC[P:P+4]} = wasm_f32x4_sub(vr${ABC[P:P+4]}, vl${ABC[P:P+4]});

        $for P in range(0, PIXEL_TILE, 4):
          const v128_t vo${ABC[P:P+4]} = wasm_f32x4_add(vl${ABC[P:P+4]}, wasm_f32x4_mul(vd${ABC[P:P+4]}, valphah${ABC[P:P+4]}));

        $for P in range(0, PIXEL_TILE, 4):
          wasm_v128_store(output + ${P}, vo${ABC[P:P+4]});
        output += ${PIXEL_TILE};
      }

    for (; p >= 4; p -= 4) {
      $for P in range(4):
        const float* itl${P} = (const float*) ((uintptr_t) i[${2 * P}] + input_offset);
        const float* ibl${P} = (const float*) ((uintptr_t) i[${2 * P + 1}] + input_offset);
      i += 8;

      const v128_t vw0 = wasm_v128_load(w);
      const v128_t vw1 = wasm_v128_load(w + 4);
      w += 8;

      $for P in range(0, 4, 2):
        const v128_t vtltr${ABC[P]} = wasm_v64x2_load_splat(itl${P});
        const v128_t vblbr${ABC[P]} = wasm_v64x2_load_splat(ibl${P});
        const double vtltr${ABC[P+1]} = *((const double*) itl${P+1});
        const double vblbr${ABC[P+1]} = *((const double*) ibl${P+1});

      const v128_t valphah = wasm_v32x4_shuffle(vw0, vw1, 0, 2, 4, 6);
      const v128_t valphav = wasm_v32x4_shuffle(vw0, vw1, 1, 3, 5, 7);

      $for P in range(0, 4, 2):
        const v128_t vtltr${ABC[P:P+2]} = wasm_f64x2_replace_lane(vtltr${ABC[P]}, 1, vtltr${ABC[P+1]});
        const v128_t vblbr${ABC[P:P+2]} = wasm_f64x2_replace_lane(vblbr${ABC[P]}, 1, vblbr${ABC[P+1]});

      $for P in range(0, 4, 2):
        const v128_t vldrd${ABC[P:P+2]} = wasm_f32x4_sub(vblbr${ABC[P:P+2]}, vtltr${ABC[P:P+2]});

      const v128_t vld = wasm_v32x4_shuffle(vldrd01, vldrd23, 0, 2, 4, 6);
      const v128_t vrd = wasm_v32x4_shuffle(vldrd01, vldrd23, 1, 3, 5, 7);

      const v128_t vtl = wasm_v32x4_shuffle(vtltr01, vtltr23, 0, 2, 4, 6);
      const v128_t vtr = wasm_v32x4_shuffle(vtltr01, vtltr23, 1, 3, 5, 7);

      const v128_t vl = wasm_f32x4_add(vtl, wasm_f32x4_mul(vld, valphav));
      const v128_t vr = wasm_f32x4_add(vtr, wasm_f32x4_mul(vrd, valphav));

      const v128_t vd = wasm_f32x4_sub(vr, vl);
      const v128_t vo = wasm_f32x4_add(vl, wasm_f32x4_mul(vd, valphah));

      wasm_v128_store(output, vo);
      output += 4;
    }

    if XNN_UNLIKELY(p != 0) {
      if (p & 2) {
        const v128_t vw = wasm_v128_load(w);
        w += 4;

        const v128_t valphah = wasm_v32x4_shuffle(vw, vw, 0, 2, 0, 2);
        const v128_t valphav = wasm_v32x4_shuffle(vw, vw, 1, 3, 1, 3);

        $for P in range(2):
          const float* itl${P} = (const float*) ((uintptr_t) i[${2 * P}] + input_offset);
          const float* ibl${P} = (const float*) ((uintptr_t) i[${2 * P + 1}] + input_offset);
        i += 4;

        const v128_t vtltr = wasm_f64x2_replace_lane(wasm_v64x2_load_splat(itl0), 1, *((const double*) itl1));
        const v128_t vblbr = wasm_f64x2_replace_lane(wasm_v64x2_load_splat(ibl0), 1, *((const double*) ibl1));

        const v128_t vldrd = wasm_f32x4_sub(vblbr, vtltr);
        const v128_t vld = wasm_v32x4_shuffle(vldrd, vldrd, 0, 2, 0, 2);
        const v128_t vrd = wasm_v32x4_shuffle(vldrd, vldrd, 1, 3, 1, 3);

        const v128_t vtl = wasm_v32x4_shuffle(vtltr, vtltr, 0, 2, 0, 2);
        const v128_t vtr = wasm_v32x4_shuffle(vtltr, vtltr, 1, 3, 1, 3);

        const v128_t vl = wasm_f32x4_add(vtl, wasm_f32x4_mul(vld, valphav));
        const v128_t vr = wasm_f32x4_add(vtr, wasm_f32x4_mul(vrd, valphav));

        const v128_t vd = wasm_f32x4_sub(vr, vl);
        const v128_t vo = wasm_f32x4_add(vl, wasm_f32x4_mul(vd, valphah));

        *((double*) output) = wasm_f64x2_extract_lane(vo, 0);
        output += 2;
      }

      if (p & 1) {
        // We are computing the following formula:
        //   result = (1 - alpha_h) * (1 - alpha_v) * top_left +
        //                 alpha_h  * (1 - alpha_v) * top_right +
        //            (1 - alpha_h) *      alpha_v  * bottom_left +
        //                 alpha_h  *      alpha_v  * bottom_right.
        //
        // Rearranging gives
        //   result =    left + alpha_h * (right        - left),
        // where
        //   left =  top_left + alpha_v * (bottom_left  - top_left),
        //  right = top_right + alpha_v * (bottom_right - top_right).

        const float alphah = *w;
        const v128_t valphav = wasm_v32x4_load_splat(w + 1);
        w += 2;

        const float* itl = (const float*) ((uintptr_t) i[0] + input_offset);
        const float* ibl = (const float*) ((uintptr_t) i[1] + input_offset);
        i += 2;

        const v128_t vtltr = wasm_v64x2_load_splat(itl);
        const v128_t vblbr = wasm_v64x2_load_splat(ibl);

        // Compute at once
        //    left_diff = bottom_left  - top_left
        //   right_diff = bottom_right - top_right
        const v128_t vldrd = wasm_f32x4_sub(vblbr, vtltr);
        const v128_t vlr = wasm_f32x4_add(vtltr, wasm_f32x4_mul(vldrd, valphav));

        // Extract them and compute the result.
        const float l = wasm_f32x4_extract_lane(vlr, 0);
        const float r = wasm_f32x4_extract_lane(vlr, 1);

        *output++ = l + alphah * (r - l);
      }
    }

    input_offset += input_increment;
  } while (--channels != 0);
}