// 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 <xmmintrin.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__sse_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:
    __m128 vmin0 = _mm_load_ss(input);
    vmin0 = _mm_shuffle_ps(vmin0, vmin0, _MM_SHUFFLE(0, 0, 0, 0));
    $if EMIT_MAX:
      __m128 vmax0 = vmin0;
  $elif EMIT_MAX:
    __m128 vmax0 = _mm_load_ss(input);
    vmax0 = _mm_shuffle_ps(vmax0, vmax0, _MM_SHUFFLE(0, 0, 0, 0));
  $for A in range(1, ACCUMULATORS):
    $if EMIT_MIN:
      __m128 vmin${A} = vmin0;
    $if EMIT_MAX:
      __m128 vmax${A} = vmax0;
  $if BATCH_TILE > 4:
    for (; batch >= ${BATCH_TILE} * sizeof(float); batch -= ${BATCH_TILE} * sizeof(float)) {
      const __m128 vt0 = _mm_loadu_ps(input);
      $for N in range(1, SIMD_TILE):
        const __m128 vt${N} = _mm_loadu_ps(input + ${N * 4});
      input += ${BATCH_TILE};

      $for N in range(SIMD_TILE):
        $if EMIT_MIN:
          vmin${N % ACCUMULATORS} = _mm_min_ps(vmin${N % ACCUMULATORS}, vt${N});
        $if EMIT_MAX:
          vmax${N % ACCUMULATORS} = _mm_max_ps(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} = _mm_min_ps(vmin${A}, vmin${A + ACC_SLICE});
            $if EMIT_MAX:
              vmax${A} = _mm_max_ps(vmax${A}, vmax${A + ACC_SLICE});
        $ACC_SLICE *= 2
  for (; batch >= 4 * sizeof(float); batch -= 4 * sizeof(float)) {
    const __m128 vt = _mm_loadu_ps(input);
    input += 4;

    $if EMIT_MIN:
      vmin0 = _mm_min_ps(vmin0, vt);
    $if EMIT_MAX:
      vmax0 = _mm_max_ps(vmax0, vt);
  }
  if XNN_UNLIKELY(batch != 0) {
    do {
      const __m128 vt = _mm_load_ss(input);
      input += 1;
      $if EMIT_MIN:
        vmin0 = _mm_min_ss(vmin0, vt);
      $if EMIT_MAX:
        vmax0 = _mm_max_ss(vmax0, vt);
      batch -= sizeof(float);
    } while (batch != 0);
  }
  $if EMIT_MIN:
    vmin0 = _mm_min_ps(vmin0, _mm_movehl_ps(vmin0, vmin0));
  $if EMIT_MAX:
    vmax0 = _mm_max_ps(vmax0, _mm_movehl_ps(vmax0, vmax0));
  $if EMIT_MIN:
    vmin0 = _mm_min_ss(vmin0, _mm_shuffle_ps(vmin0, vmin0, _MM_SHUFFLE(1, 1, 1, 1)));
  $if EMIT_MAX:
    vmax0 = _mm_max_ss(vmax0, _mm_shuffle_ps(vmax0, vmax0, _MM_SHUFFLE(1, 1, 1, 1)));
  $if EMIT_MIN:
    _mm_store_ss(output, vmin0);
    $if EMIT_MAX:
      _mm_store_ss(output + 1, vmax0);
  $elif EMIT_MAX:
    _mm_store_ss(output, vmax0);
}