// Copyright 2023 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
$assert OP in ["MAX", "MIN", "MINMAX"]
$EMIT_MIN = "MIN" in OP
$EMIT_MAX = "MAX" in OP
#include <assert.h>

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


$ACC_SUFFIX = "" if ACCUMULATORS == 1 else "_acc%d" % ACCUMULATORS
$MAX = "math_max_s16"
$MIN = "math_min_s16"
$ISA = "wasm" if WASM else "scalar"
$MAX_POS = 1 if OP == "MINMAX" else 0
void xnn_f16_r${OP.lower()}_ukernel__scalar_u${BATCH_TILE}${ACC_SUFFIX}(
    size_t batch,
    const xnn_float16* input,
    xnn_float16* output,
    const struct xnn_f16_default_params params[restrict XNN_MIN_ELEMENTS(1)])
{
  assert(batch != 0);
  assert(batch % sizeof(uint16_t) == 0);
  assert(input != NULL);
  assert(output != NULL);

  const uint16_t* i = (const uint16_t*) input;
  uint16_t* o = (uint16_t*) output;

  int16_t vt = math_signcomplement_f16(*i);
  $for A in range(ACCUMULATORS):
    $if EMIT_MIN:
      int16_t vmin${A} = vt;
    $if EMIT_MAX:
      int16_t vmax${A} = vt;
  $if BATCH_TILE == 1:
    do {
      vt = math_signcomplement_f16(*i++);
      $if EMIT_MIN:
        vmin0 = ${MIN}(vmin0, vt);
      $if EMIT_MAX:
        vmax0 = ${MAX}(vmax0, vt);
      batch -= sizeof(uint16_t);
    } while (batch != 0);
  $else:
    for (; batch >= ${BATCH_TILE} * sizeof(uint16_t); batch -= ${BATCH_TILE} * sizeof(uint16_t)) {
      $for N in range(BATCH_TILE):
        const int16_t vt${N} = math_signcomplement_f16(i[${N}]);
      i += ${BATCH_TILE};

      $for N in range(BATCH_TILE):
        $if EMIT_MIN:
          vmin${N % ACCUMULATORS} = ${MIN}(vmin${N % ACCUMULATORS}, vt${N});
        $if EMIT_MAX:
          vmax${N % ACCUMULATORS} = ${MAX}(vmax${N % ACCUMULATORS}, vt${N});
    }
    $if ACCUMULATORS > 1:
      $ACC_SLICE = 1
      $while ACC_SLICE < ACCUMULATORS:
        $for A in range(0, ACCUMULATORS, ACC_SLICE * 2):
          $if A + ACC_SLICE < ACCUMULATORS:
            $if EMIT_MIN:
              vmin${A} = ${MIN}(vmin${A}, vmin${A + ACC_SLICE});
            $if EMIT_MAX:
              vmax${A} = ${MAX}(vmax${A}, vmax${A + ACC_SLICE});
        $ACC_SLICE *= 2

    if XNN_UNLIKELY(batch != 0) {
      $if BATCH_TILE == 2:
        vt = math_signcomplement_f16(*i);
        $if EMIT_MIN:
          vmin0 = ${MIN}(vmin0, vt);
        $if EMIT_MAX:
          vmax0 = ${MAX}(vmax0, vt);
      $else:
        do {
          vt = math_signcomplement_f16(*i++);
          $if EMIT_MIN:
            vmin0 = ${MIN}(vmin0, vt);
          $if EMIT_MAX:
            vmax0 = ${MAX}(vmax0, vt);
          batch -= sizeof(uint16_t);
        } while (batch != 0);
    }
  $if EMIT_MIN:
    o[0] = (uint16_t) math_signcomplement_f16((uint16_t) vmin0);
  $if EMIT_MAX:
    o[${MAX_POS}] = (uint16_t) math_signcomplement_f16((uint16_t) vmax0);
}