// 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
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
#include <assert.h>

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


void xnn_f16_f32_vcvt_ukernel__scalar_x${BATCH_TILE}(
    size_t batch,
    const void* input,
    float* output,
    const union xnn_f16_f32_cvt_params params[restrict XNN_MIN_ELEMENTS(1)])
{
  assert(batch != 0);
  assert(batch % sizeof(uint16_t) == 0);
  assert(input != NULL);
  assert(output != NULL);

  const uint32_t vsign_mask = params->scalar.sign_mask;
  const uint32_t vexp_offset = params->scalar.exp_offset;
  const float vexp_scale = params->scalar.exp_scale;
  const uint32_t vmagic_mask = params->scalar.magic_mask;
  const float vmagic_bias = params->scalar.magic_bias;
  const uint32_t vdenorm_cutoff = params->scalar.denorm_cutoff;

  const uint16_t* i = (const uint16_t*) input;
  uint32_t* o = (uint32_t*) output;
  $if BATCH_TILE > 1:
    for (; batch >= ${BATCH_TILE} * sizeof(uint16_t); batch -= ${BATCH_TILE} * sizeof(uint16_t)) {
      $for N in range(BATCH_TILE):
        const uint16_t vh${N} = i[${N}];
      i += ${BATCH_TILE};

      $for N in range(BATCH_TILE):
        const uint32_t vw${N} = (uint32_t) vh${N} << 16;

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

      $for N in range(BATCH_TILE):
        const uint32_t v2w${N} = vw${N} + vw${N};

      $for N in range(BATCH_TILE):
        const uint32_t vnorm${N} = float_as_uint32(uint32_as_float((v2w${N} >> 4) + vexp_offset) * vexp_scale);

      $for N in range(BATCH_TILE):
        const uint32_t vdenorm${N} = float_as_uint32(uint32_as_float((v2w${N} >> 17) | vmagic_mask) - vmagic_bias);

      $for N in range(BATCH_TILE):
        const uint32_t vf${N} = vsign${N} | (XNN_UNPREDICTABLE(v2w${N} < vdenorm_cutoff) ? vdenorm${N} : vnorm${N});

      $for N in range(BATCH_TILE):
        o[${N}] = vf${N};
      o += ${BATCH_TILE};
    }
  $if BATCH_TILE == 1:
    do {
      const uint16_t vh = *i++;

      const uint32_t vw = (uint32_t) vh << 16;
      const uint32_t vsign = vw & vsign_mask;
      const uint32_t v2w = vw + vw;
      const uint32_t vnorm = float_as_uint32(uint32_as_float((v2w >> 4) + vexp_offset) * vexp_scale);
      const uint32_t vdenorm = float_as_uint32(uint32_as_float((v2w >> 17) | vmagic_mask) - vmagic_bias);
      const uint32_t vf = vsign | (XNN_UNPREDICTABLE(v2w < vdenorm_cutoff) ? vdenorm : vnorm);

      *o++ = vf;

      batch -= sizeof(uint16_t);
    } while (batch != 0);
  $elif BATCH_TILE == 2:
    if XNN_UNLIKELY(batch != 0) {
      const uint16_t vh = *i;

      const uint32_t vw = (uint32_t) vh << 16;
      const uint32_t vsign = vw & vsign_mask;
      const uint32_t v2w = vw + vw;
      const uint32_t vnorm = float_as_uint32(uint32_as_float((v2w >> 4) + vexp_offset) * vexp_scale);
      const uint32_t vdenorm = float_as_uint32(uint32_as_float((v2w >> 17) | vmagic_mask) - vmagic_bias);
      const uint32_t vf = vsign | (XNN_UNPREDICTABLE(v2w < vdenorm_cutoff) ? vdenorm : vnorm);

      *o = vf;
    }
  $else:
    if XNN_UNLIKELY(batch != 0) {
      do {
        const uint16_t vh = *i++;

        const uint32_t vw = (uint32_t) vh << 16;
        const uint32_t vsign = vw & vsign_mask;
        const uint32_t v2w = vw + vw;
        const uint32_t vnorm = float_as_uint32(uint32_as_float((v2w >> 4) + vexp_offset) * vexp_scale);
        const uint32_t vdenorm = float_as_uint32(uint32_as_float((v2w >> 17) | vmagic_mask) - vmagic_bias);
        const uint32_t vf = vsign | (XNN_UNPREDICTABLE(v2w < vdenorm_cutoff) ? vdenorm : vnorm);

        *o++ = vf;

        batch -= sizeof(uint16_t);
      } while (batch != 0);
    }
}