// 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 >= 1
#include <assert.h>

#include "xnnpack/common.h"
#include "xnnpack/math.h"
#include "xnnpack/vcvt.h"


void xnn_f32_f16_vcvt_ukernel__scalar_bitcast_u${BATCH_TILE}(
    size_t batch,
    const float* input,
    xnn_float16* output,
    const void* params)
{
  assert(batch != 0);
  assert(batch % sizeof(float) == 0);
  assert(input != NULL);
  assert(output != NULL);

  const uint32_t vnonsign_mask = UINT32_C(0x7FFFFFFF);
  const uint32_t vexp_bias = UINT32_C(0x07800000);
  const float vscale_to_inf = 0x1.0p+112f;
  const uint32_t vexpw_max = UINT32_C(0x7F800000);
  const float vscale_to_zero = 0x1.0p-110f;
  const uint32_t vbias_min = UINT32_C(0x40000000);
  const uint16_t vexph_mask = UINT16_C(0x7C00);
  const uint16_t vmanth_mask = UINT16_C(0x0FFF);
  const uint16_t vnanh = UINT16_C(0x7E00);

  const uint32_t* i = (const uint32_t*) input;
  uint16_t* o = (uint16_t*) output;
  $if BATCH_TILE == 1:
    do {
      const uint32_t vw = *i++;

      const uint32_t vnonsignw = vw & vnonsign_mask;

      float vf = uint32_as_float(vnonsignw);
      const uint32_t vsignw = vw ^ vnonsignw;
      uint32_t vbias = vnonsignw + vexp_bias;

      vf *= vscale_to_inf;
      vbias &= vexpw_max;

      vf *= vscale_to_zero;
      vbias = math_max_u32(vbias, vbias_min);

      vf += uint32_as_float(vbias);

      const uint32_t vbits = float_as_uint32(vf);

      const uint16_t vexph = (uint16_t) (vbits >> 13) & vexph_mask;
      const uint16_t vmanth = (uint16_t) vbits & vmanth_mask;
      const uint16_t vsignh = (uint16_t) (vsignw >> 16);

      uint16_t vh = vexph + vmanth;
      if XNN_UNPREDICTABLE(vnonsignw > vexpw_max) {
        vh = vnanh;
      }
      vh |= vsignh;

      *o++ = vh;

      batch -= sizeof(float);
    } while (batch != 0);
  $else:
    for (; batch >= ${BATCH_TILE} * sizeof(float); batch -= ${BATCH_TILE} * sizeof(float)) {
      $for N in range(BATCH_TILE):
        const uint32_t vw${N} = i[${N}];
      i += ${BATCH_TILE};

      $for N in range(BATCH_TILE):
        const uint32_t vnonsignw${N} = vw${N} & vnonsign_mask;

      $for N in range(BATCH_TILE):
        float vf${N} = uint32_as_float(vnonsignw${N});
      $for N in range(BATCH_TILE):
        const uint32_t vsignw${N} = vw${N} ^ vnonsignw${N};
      $for N in range(BATCH_TILE):
        uint32_t vbias${N} = vnonsignw${N} + vexp_bias;

      $for N in range(BATCH_TILE):
        vf${N} *= vscale_to_inf;
      $for N in range(BATCH_TILE):
        vbias${N} &= vexpw_max;

      $for N in range(BATCH_TILE):
        vf${N} *= vscale_to_zero;
      $for N in range(BATCH_TILE):
        vbias${N} = math_max_u32(vbias${N}, vbias_min);

      $for N in range(BATCH_TILE):
        vf${N} += uint32_as_float(vbias${N});

      $for N in range(BATCH_TILE):
        const uint32_t vbits${N} = float_as_uint32(vf${N});

      $for N in range(BATCH_TILE):
        const uint16_t vexph${N} = (uint16_t) (vbits${N} >> 13) & vexph_mask;
      $for N in range(BATCH_TILE):
        const uint16_t vmanth${N} = (uint16_t) vbits${N} & vmanth_mask;
      $for N in range(BATCH_TILE):
        const uint16_t vsignh${N} = (uint16_t) (vsignw${N} >> 16);

      $for N in range(BATCH_TILE):
        uint16_t vh${N} = vexph${N} + vmanth${N};
      $for N in range(BATCH_TILE):
        if XNN_UNPREDICTABLE(vnonsignw${N} > vexpw_max) {
          vh${N} = vnanh;
        }
      $for N in range(BATCH_TILE):
        vh${N} |= vsignh${N};

      $for N in range(BATCH_TILE):
        o[${N}] = vh${N};
      o += ${BATCH_TILE};
    }
    $if BATCH_TILE == 2:
      if XNN_UNLIKELY(batch != 0) {
        const uint32_t vw = *i;

        const uint32_t vnonsignw = vw & vnonsign_mask;

        float vf = uint32_as_float(vnonsignw);
        const uint32_t vsignw = vw ^ vnonsignw;
        uint32_t vbias = vnonsignw + vexp_bias;

        vf *= vscale_to_inf;
        vbias &= vexpw_max;

        vf *= vscale_to_zero;
        vbias = math_max_u32(vbias, vbias_min);

        vf += uint32_as_float(vbias);

        const uint32_t vbits = float_as_uint32(vf);

        const uint16_t vexph = (uint16_t) (vbits >> 13) & vexph_mask;
        const uint16_t vmanth = (uint16_t) vbits & vmanth_mask;
        const uint16_t vsignh = (uint16_t) (vsignw >> 16);

        uint16_t vh = vexph + vmanth;
        if XNN_UNPREDICTABLE(vnonsignw > vexpw_max) {
          vh = vnanh;
        }
        vh |= vsignh;

        *o = vh;
      }
    $else:
      if XNN_UNLIKELY(batch != 0) {
        do {
          const uint32_t vw = *i++;

          const uint32_t vnonsignw = vw & vnonsign_mask;

          float vf = uint32_as_float(vnonsignw);
          const uint32_t vsignw = vw ^ vnonsignw;
          uint32_t vbias = vnonsignw + vexp_bias;

          vf *= vscale_to_inf;
          vbias &= vexpw_max;

          vf *= vscale_to_zero;
          vbias = math_max_u32(vbias, vbias_min);

          vf += uint32_as_float(vbias);

          const uint32_t vbits = float_as_uint32(vf);

          const uint16_t vexph = (uint16_t) (vbits >> 13) & vexph_mask;
          const uint16_t vmanth = (uint16_t) vbits & vmanth_mask;
          const uint16_t vsignh = (uint16_t) (vsignw >> 16);

          uint16_t vh = vexph + vmanth;
          if XNN_UNPREDICTABLE(vnonsignw > vexpw_max) {
            vh = vnanh;
          }
          vh |= vsignh;

          *o++ = vh;

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