// 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 BATCH_TILE % 8 == 0
$assert BATCH_TILE >= 8
$SIMD_TILE = BATCH_TILE // 8
#include <assert.h>

#include <wasm_simd128.h>

#include <xnnpack/common.h>
#include <xnnpack/vcvt.h>


$WASM_V16X8_LANESELECT = "wasm_i16x8_relaxed_laneselect" if RELAXED else "wasm_v128_bitselect"
$ISA = "wasmrelaxedsimd" if RELAXED else "wasmsimd"
void xnn_f32_f16_vcvt_ukernel__${ISA}_x${BATCH_TILE}(
    size_t batch,
    const float* input,
    void* output,
    const union xnn_f32_f16_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(batch != 0);
  assert(batch % sizeof(float) == 0);
  assert(input != NULL);
  assert(output != NULL);

  const v128_t vexp_bias = wasm_v128_load64_splat(params->wasmsimd.exp_bias);
  const v128_t vscale_to_inf = wasm_v128_load64_splat(params->wasmsimd.scale_to_inf);
  const v128_t vexpw_max = wasm_v128_load64_splat(params->wasmsimd.expw_max);
  const v128_t vscale_to_zero = wasm_v128_load64_splat(params->wasmsimd.scale_to_zero);
  const v128_t vbias_min = wasm_v128_load64_splat(params->wasmsimd.bias_min);
  const v128_t vmanth_mask = wasm_v128_load64_splat(params->wasmsimd.manth_mask);
  const v128_t vexph_mask = wasm_v128_load64_splat(params->wasmsimd.exph_mask);
  const v128_t vnanh = wasm_v128_load64_splat(params->wasmsimd.nanh);

  uint16_t* o = (uint16_t*) output;
  $if BATCH_TILE > 8:
    for (; batch >= ${BATCH_TILE} * sizeof(float); batch -= ${BATCH_TILE} * sizeof(float)) {
      const v128_t vx0 = wasm_v128_load(input);
      $for N in range(1, 2*SIMD_TILE):
        const v128_t vx${N} = wasm_v128_load(input + ${N * 4});
      input += ${BATCH_TILE};

      $for N in range(2*SIMD_TILE):
        const v128_t vabsx${N} = wasm_f32x4_abs(vx${N});

      $for N in range(2*SIMD_TILE):
        const v128_t vsignx${N} = wasm_v128_xor(vx${N}, vabsx${N});

      $for N in range(2*SIMD_TILE):
        v128_t vbias${N} = wasm_i32x4_add(vabsx${N}, vexp_bias);

      $for N in range(2*SIMD_TILE):
        v128_t vf${N} = wasm_f32x4_mul(vabsx${N}, vscale_to_inf);

      $for N in range(2*SIMD_TILE):
        const v128_t vnanmaskw${N} = wasm_i32x4_gt(vabsx${N}, vexpw_max);

      $for N in range(2*SIMD_TILE):
        vbias${N} = wasm_v128_and(vbias${N}, vexpw_max);

      $for N in range(2*SIMD_TILE):
        vf${N} = wasm_f32x4_mul(vf${N}, vscale_to_zero);

      $for N in range(SIMD_TILE):
        const v128_t vnanmaskh${N} = wasm_i16x8_narrow_i32x4(vnanmaskw${2*N}, vnanmaskw${2*N+1});

      $for N in range(SIMD_TILE):
        const v128_t vsignh${N} = wasm_i16x8_narrow_i32x4(vsignx${2*N}, vsignx${2*N+1});

      $for N in range(2*SIMD_TILE):
        vbias${N} = wasm_i16x8_max(vbias${N}, vbias_min);

      $for N in range(2*SIMD_TILE):
        vf${N} = wasm_f32x4_add(vf${N}, vbias${N});

      $for N in range(2*SIMD_TILE):
        v128_t vexpw${N} = wasm_i32x4_shr(vf${N}, 13);

      $for N in range(2*SIMD_TILE):
        const v128_t vmantw${N} = wasm_v128_and(vf${N}, vmanth_mask);

      $for N in range(2*SIMD_TILE):
        vexpw${N} = wasm_v128_and(vexpw${N}, vexph_mask);

      $for N in range(2*SIMD_TILE):
        const v128_t vnonsignw${N} = wasm_i32x4_add(vmantw${N}, vexpw${N});

      $for N in range(SIMD_TILE):
        const v128_t vnonsignh${N} = wasm_i16x8_narrow_i32x4(vnonsignw${2*N}, vnonsignw${2*N+1});

      $for N in range(SIMD_TILE):
        const v128_t vabsh${N} = ${WASM_V16X8_LANESELECT}(vnanh, vnonsignh${N}, vnanmaskh${N});

      $for N in range(SIMD_TILE):
        const v128_t vh${N} = wasm_v128_or(vabsh${N}, vsignh${N});

      wasm_v128_store(o, vh0);
      $for N in range(1, SIMD_TILE):
        wasm_v128_store(o + ${N * 8}, vh${N});
      o += ${BATCH_TILE};
    }
  for (; batch >= 8 * sizeof(float); batch -= 8 * sizeof(float)) {
    const v128_t vx_lo = wasm_v128_load(input);
    const v128_t vx_hi = wasm_v128_load(input + 4);
    input += 8;

    const v128_t vabsx_lo = wasm_f32x4_abs(vx_lo);
    const v128_t vabsx_hi = wasm_f32x4_abs(vx_hi);

    const v128_t vsignx_lo = wasm_v128_xor(vx_lo, vabsx_lo);
    const v128_t vsignx_hi = wasm_v128_xor(vx_hi, vabsx_hi);
    v128_t vbias_lo = wasm_i32x4_add(vabsx_lo, vexp_bias);
    v128_t vbias_hi = wasm_i32x4_add(vabsx_hi, vexp_bias);
    v128_t vf_lo = wasm_f32x4_mul(vabsx_lo, vscale_to_inf);
    v128_t vf_hi = wasm_f32x4_mul(vabsx_hi, vscale_to_inf);
    const v128_t vnanmaskw_lo = wasm_i32x4_gt(vabsx_lo, vexpw_max);
    const v128_t vnanmaskw_hi = wasm_i32x4_gt(vabsx_hi, vexpw_max);

    vbias_lo = wasm_v128_and(vbias_lo, vexpw_max);
    vbias_hi = wasm_v128_and(vbias_hi, vexpw_max);
    vf_lo = wasm_f32x4_mul(vf_lo, vscale_to_zero);
    vf_hi = wasm_f32x4_mul(vf_hi, vscale_to_zero);
    const v128_t vnanmaskh = wasm_i16x8_narrow_i32x4(vnanmaskw_lo, vnanmaskw_hi);
    const v128_t vsignh = wasm_i16x8_narrow_i32x4(vsignx_lo, vsignx_hi);

    vbias_lo = wasm_i16x8_max(vbias_lo, vbias_min);
    vbias_hi = wasm_i16x8_max(vbias_hi, vbias_min);

    vf_lo = wasm_f32x4_add(vf_lo, vbias_lo);
    vf_hi = wasm_f32x4_add(vf_hi, vbias_hi);

    v128_t vexpw_lo = wasm_i32x4_shr(vf_lo, 13);
    v128_t vexpw_hi = wasm_i32x4_shr(vf_hi, 13);
    const v128_t vmantw_lo = wasm_v128_and(vf_lo, vmanth_mask);
    const v128_t vmantw_hi = wasm_v128_and(vf_hi, vmanth_mask);

    vexpw_lo = wasm_v128_and(vexpw_lo, vexph_mask);
    vexpw_hi = wasm_v128_and(vexpw_hi, vexph_mask);

    const v128_t vnonsignw_lo = wasm_i32x4_add(vmantw_lo, vexpw_lo);
    const v128_t vnonsignw_hi = wasm_i32x4_add(vmantw_hi, vexpw_hi);

    const v128_t vnonsignh = wasm_i16x8_narrow_i32x4(vnonsignw_lo, vnonsignw_hi);

    const v128_t vabsh = ${WASM_V16X8_LANESELECT}(vnanh, vnonsignh, vnanmaskh);

    const v128_t vh = wasm_v128_or(vabsh, vsignh);

    wasm_v128_store(o, vh);
    o += 8;
  }
  if XNN_UNPREDICTABLE(batch != 0) {
    const v128_t vx_lo = wasm_v128_load(input);
    const float* input_hi = (const float*) ((uintptr_t) input + (batch & (4 * sizeof(float))));
    const v128_t vx_hi = wasm_v128_load(input_hi);

    const v128_t vabsx_lo = wasm_f32x4_abs(vx_lo);
    const v128_t vabsx_hi = wasm_f32x4_abs(vx_hi);

    const v128_t vsignx_lo = wasm_v128_xor(vx_lo, vabsx_lo);
    const v128_t vsignx_hi = wasm_v128_xor(vx_hi, vabsx_hi);
    v128_t vbias_lo = wasm_i32x4_add(vabsx_lo, vexp_bias);
    v128_t vbias_hi = wasm_i32x4_add(vabsx_hi, vexp_bias);
    v128_t vf_lo = wasm_f32x4_mul(vabsx_lo, vscale_to_inf);
    v128_t vf_hi = wasm_f32x4_mul(vabsx_hi, vscale_to_inf);
    const v128_t vnanmaskw_lo = wasm_i32x4_gt(vabsx_lo, vexpw_max);
    const v128_t vnanmaskw_hi = wasm_i32x4_gt(vabsx_hi, vexpw_max);

    vbias_lo = wasm_v128_and(vbias_lo, vexpw_max);
    vbias_hi = wasm_v128_and(vbias_hi, vexpw_max);
    vf_lo = wasm_f32x4_mul(vf_lo, vscale_to_zero);
    vf_hi = wasm_f32x4_mul(vf_hi, vscale_to_zero);
    const v128_t vnanmaskh = wasm_i16x8_narrow_i32x4(vnanmaskw_lo, vnanmaskw_hi);
    const v128_t vsignh = wasm_i16x8_narrow_i32x4(vsignx_lo, vsignx_hi);

    vbias_lo = wasm_i16x8_max(vbias_lo, vbias_min);
    vbias_hi = wasm_i16x8_max(vbias_hi, vbias_min);

    vf_lo = wasm_f32x4_add(vf_lo, vbias_lo);
    vf_hi = wasm_f32x4_add(vf_hi, vbias_hi);

    v128_t vexpw_lo = wasm_i32x4_shr(vf_lo, 13);
    v128_t vexpw_hi = wasm_i32x4_shr(vf_hi, 13);
    const v128_t vmantw_lo = wasm_v128_and(vf_lo, vmanth_mask);
    const v128_t vmantw_hi = wasm_v128_and(vf_hi, vmanth_mask);

    vexpw_lo = wasm_v128_and(vexpw_lo, vexph_mask);
    vexpw_hi = wasm_v128_and(vexpw_hi, vexph_mask);

    const v128_t vnonsignw_lo = wasm_i32x4_add(vmantw_lo, vexpw_lo);
    const v128_t vnonsignw_hi = wasm_i32x4_add(vmantw_hi, vexpw_hi);

    const v128_t vnonsignh = wasm_i16x8_narrow_i32x4(vnonsignw_lo, vnonsignw_hi);

    const v128_t vabsh = ${WASM_V16X8_LANESELECT}(vnanh, vnonsignh, vnanmaskh);

    v128_t vh = wasm_v128_or(vabsh, vsignh);

    if (batch & (4 * sizeof(float))) {
      wasm_v128_store64_lane(o, vh, 0);
      vh = wasm_v64x2_shuffle(vh, vh, 1, 1);
      o += 4;
    }
    if (batch & (2 * sizeof(float))) {
      wasm_v128_store32_lane(o, vh, 0);
      vh = wasm_i64x2_shr(vh, 32);
      o += 2;
    }
    if (batch & (1 * sizeof(float))) {
      wasm_v128_store16_lane(o, vh, 0);
    }
  }
}