// 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_u${BATCH_TILE}(
    size_t batch,
    const ${XINT8_T}* input,
    float* output,
    const struct xnn_${DATATYPE.lower()}_f32_cvt_params params[restrict XNN_MIN_ELEMENTS(1)])
{
  assert(batch != 0);
  assert(batch % sizeof(${XINT8_T}) == 0);
  assert(input != NULL);
  assert(output != NULL);

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

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

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

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

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

      float vy = (float) vx;
      vy *= vscale;
      *output = vy;
    }
  $else:
    if XNN_UNLIKELY(batch != 0) {
      do {
        int32_t vx = *input++;
        vx -= vzero_point;

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

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