// 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 % 4 == 0
$assert BATCH_TILE >= 4
$SIMD_TILE = BATCH_TILE // 4
$assert ACCUMULATORS <= SIMD_TILE
$assert OP in ["MAX", "MIN", "MINMAX"]
#include <assert.h>

#include <arm_neon.h>

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


$ACC_SUFFIX = "" if ACCUMULATORS == 1 else "_acc%d" % ACCUMULATORS
$EMIT_MIN = "MIN" in OP
$EMIT_MAX = "MAX" in OP
void xnn_f32_r${OP.lower()}_ukernel__neon_u${BATCH_TILE}${ACC_SUFFIX}(
    size_t batch,
    const float* input,
    float* output,
    const struct xnn_f32_default_params params[restrict XNN_MIN_ELEMENTS(1)])
{
  assert(batch != 0);
  assert(batch % sizeof(float) == 0);
  assert(input != NULL);
  assert(output != NULL);

  $if EMIT_MIN:
    float32x4_t vmin0 = vld1q_dup_f32(input);
    $if EMIT_MAX:
      float32x4_t vmax0 = vmin0;
  $elif EMIT_MAX:
    float32x4_t vmax0 = vld1q_dup_f32(input);
  $for A in range(1, ACCUMULATORS):
    $if EMIT_MIN:
      float32x4_t vmin${A} = vmin0;
    $if EMIT_MAX:
      float32x4_t vmax${A} = vmax0;
  $if BATCH_TILE > 4:
    for (; batch >= ${BATCH_TILE} * sizeof(float); batch -= ${BATCH_TILE} * sizeof(float)) {
      $for N in range(SIMD_TILE):
        const float32x4_t vt${N} = vld1q_f32(input); input += 4;

      $for N in range(SIMD_TILE):
        $if $EMIT_MIN:
          vmin${N % ACCUMULATORS} = vminq_f32(vmin${N % ACCUMULATORS}, vt${N});
        $if $EMIT_MAX:
          vmax${N % ACCUMULATORS} = vmaxq_f32(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} = vminq_f32(vmin${A}, vmin${A + ACC_SLICE});
            $if $EMIT_MAX:
              vmax${A} = vmaxq_f32(vmax${A}, vmax${A + ACC_SLICE});
        $ACC_SLICE *= 2
  for (; batch >= 4 * sizeof(float); batch -= 4 * sizeof(float)) {
    const float32x4_t vt = vld1q_f32(input); input += 4;
    $if $EMIT_MIN:
      vmin0 = vminq_f32(vmin0, vt);
    $if $EMIT_MAX:
      vmax0 = vmaxq_f32(vmax0, vt);
  }
  $if $EMIT_MIN:
    float32x2_t vmin = vmin_f32(vget_low_f32(vmin0), vget_high_f32(vmin0));
  $if $EMIT_MAX:
    float32x2_t vmax = vmax_f32(vget_low_f32(vmax0), vget_high_f32(vmax0));
  if XNN_UNLIKELY(batch & (2 * sizeof(float))) {
    const float32x2_t vt = vld1_f32(input); input += 2;
    $if $EMIT_MIN:
      vmin = vmin_f32(vmin, vt);
    $if $EMIT_MAX:
      vmax = vmax_f32(vmax, vt);
  }
  $if $EMIT_MIN:
    vmin = vpmin_f32(vmin, vmin);
  $if $EMIT_MAX:
    vmax = vpmax_f32(vmax, vmax);
  if XNN_UNLIKELY(batch & (1 * sizeof(float))) {
    const float32x2_t vt = vld1_dup_f32(input);
    $if $EMIT_MIN:
      vmin = vmin_f32(vmin, vt);
    $if $EMIT_MAX:
      vmax = vmax_f32(vmax, vt);
  }
  $if $EMIT_MIN:
    vst1_lane_f32(output, vmin, 0);
    $if $EMIT_MAX:
      vst1_lane_f32(output + 1, vmax, 0);
  $elif $EMIT_MAX:
    vst1_lane_f32(output, vmax, 0);
}