// 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>


$XINT8_T = {"QS8": "int8_t", "QU8": "uint8_t"}[DATATYPE]
void xnn_${DATATYPE.lower()}_f32_vcvt_ukernel__scalar_x${BATCH_TILE}(
    size_t n,
    const ${XINT8_T}* x,
    float* y,
    const union xnn_${DATATYPE.lower()}_f32_cvt_params params[restrict XNN_MIN_ELEMENTS(1)])
{
  assert(n != 0);
  assert(n % sizeof(${XINT8_T}) == 0);
  assert(x != NULL);
  assert(y != NULL);

  const int32_t vzero_point = params->scalar.zero_point;
  const float vscale = params->scalar.scale;

  $if BATCH_TILE > 1:
    for (; n >= ${BATCH_TILE} * sizeof(${XINT8_T}); n -= ${BATCH_TILE} * sizeof(${XINT8_T})) {
      $for N in range(BATCH_TILE):
        int32_t vx${N} = (int32_t) x[${N}];
      x += ${BATCH_TILE};

      $for N in range(BATCH_TILE):
        vx${N} -= vzero_point;

      $for N in range(BATCH_TILE):
        float vy${N} = (float) vx${N};

      $for N in range(BATCH_TILE):
        vy${N} *= vscale;

      $for N in range(BATCH_TILE):
        y[${N}] = vy${N};
      y += ${BATCH_TILE};
    }
  $if BATCH_TILE == 1:
    do {
      int32_t vx = *x++;
      vx -= vzero_point;

      float vy = (float) vx;
      vy *= vscale;
      *y++ = vy;

      n -= sizeof(${XINT8_T});
    } while (n != 0);
  $elif BATCH_TILE == 2:
    if XNN_UNLIKELY(n != 0) {
      int32_t vx = *x;
      vx -= vzero_point;

      float vy = (float) vx;
      vy *= vscale;
      *y = vy;
    }
  $else:
    if XNN_UNLIKELY(n != 0) {
      do {
        int32_t vx = *x++;
        vx -= vzero_point;

        float vy = (float) vx;
        vy *= vscale;
        *y++ = vy;

        n -= sizeof(${XINT8_T});
      } while (n != 0);
    }
}