// 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 % 8 == 0
$assert BATCH_TILE >= 8
$SIMD_TILE = BATCH_TILE // 8
$assert ACCUMULATORS <= SIMD_TILE
$assert OP in ["MAX", "MIN", "MINMAX"]
#include <assert.h>

#include <immintrin.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__avx_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);

  static const int32_t mask_table[14] = {-1, -1, -1, -1, -1, -1, -1, 0, 0, 0, 0, 0, 0, 0};

  $if EMIT_MIN:
    __m256 vmin0 = _mm256_broadcast_ss(input);
    $if EMIT_MAX:
      __m256 vmax0 = vmin0;
  $elif EMIT_MAX:
    __m256 vmax0 = _mm256_broadcast_ss(input);
  $for A in range(1, ACCUMULATORS):
    $if EMIT_MIN:
      __m256 vmin${A} = vmin0;
    $if EMIT_MAX:
      __m256 vmax${A} = vmax0;
  $if BATCH_TILE > 8:
    for (; batch >= ${BATCH_TILE} * sizeof(float); batch -= ${BATCH_TILE} * sizeof(float)) {
      const __m256 vt0 = _mm256_loadu_ps(input);
      $for N in range(1, SIMD_TILE):
        const __m256 vt${N} = _mm256_loadu_ps(input + ${N * 8});
      input += ${BATCH_TILE};

      $for N in range(SIMD_TILE):
        $if EMIT_MIN:
          vmin${N % ACCUMULATORS} = _mm256_min_ps(vmin${N % ACCUMULATORS}, vt${N});
        $if EMIT_MAX:
          vmax${N % ACCUMULATORS} = _mm256_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} = _mm256_min_ps(vmin${A}, vmin${A + ACC_SLICE});
            $if EMIT_MAX:
              vmax${A} = _mm256_max_ps(vmax${A}, vmax${A + ACC_SLICE});
        $ACC_SLICE *= 2
  for (; batch >= 8 * sizeof(float); batch -= 8 * sizeof(float)) {
    const __m256 vt = _mm256_loadu_ps(input);
    input += 8;

    $if EMIT_MIN:
      vmin0 = _mm256_min_ps(vmin0, vt);
    $if EMIT_MAX:
      vmax0 = _mm256_max_ps(vmax0, vt);
  }
  if XNN_UNLIKELY(batch != 0) {
    assert(batch >= 1 * sizeof(float));
    assert(batch <= 7 * sizeof(float));
    const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - batch));

    const __m256 vt = _mm256_maskload_ps(input, vmask);

    $if EMIT_MIN:
      vmin0 = _mm256_blendv_ps(vmin0, _mm256_min_ps(vmin0, vt), _mm256_castsi256_ps(vmask));
    $if EMIT_MAX:
      vmax0 = _mm256_blendv_ps(vmax0, _mm256_max_ps(vmax0, vt), _mm256_castsi256_ps(vmask));
  }
  $if EMIT_MIN:
    __m128 vmin = _mm_min_ps(_mm256_castps256_ps128(vmin0), _mm256_extractf128_ps(vmin0, 1));
  $if EMIT_MAX:
    __m128 vmax = _mm_max_ps(_mm256_castps256_ps128(vmax0), _mm256_extractf128_ps(vmax0, 1));
  $if EMIT_MIN:
    vmin = _mm_min_ps(vmin, _mm_movehl_ps(vmin, vmin));
  $if EMIT_MAX:
    vmax = _mm_max_ps(vmax, _mm_movehl_ps(vmax, vmax));
  $if EMIT_MIN:
    vmin = _mm_min_ss(vmin, _mm_movehdup_ps(vmin));
  $if EMIT_MAX:
    vmax = _mm_max_ss(vmax, _mm_movehdup_ps(vmax));
  $if EMIT_MIN:
    _mm_store_ss(output, vmin);
    $if EMIT_MAX:
      _mm_store_ss(output + 1, vmax);
  $elif EMIT_MAX:
    _mm_store_ss(output, vmax);
}