// 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
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
#include <assert.h>

#include <immintrin.h>

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


$ACC_SUFFIX = "" if ACCUMULATORS == 1 else "_acc%d" % ACCUMULATORS
void xnn_f32_rsum_ukernel__avx_u${BATCH_TILE}${ACC_SUFFIX}(
    size_t batch,
    const float* input,
    float* output,
    const struct xnn_f32_scale_params params[restrict XNN_MIN_ELEMENTS(1)])
{
  static const int32_t mask_table[14] = {-1, -1, -1, -1, -1, -1, -1, 0, 0, 0, 0, 0, 0, 0};

  assert(batch != 0);
  assert(batch % sizeof(float) == 0);
  assert(input != NULL);
  assert(output != NULL);

  $for A in range(ACCUMULATORS):
    __m256 vacc${A} = _mm256_setzero_ps();
  $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};

      $for N in range(SIMD_TILE):
        vacc${N % ACCUMULATORS} = _mm256_add_ps(vacc${N % ACCUMULATORS}, vt${N});
    }
    $if ACCUMULATORS > 1:
      $ACC_SLICE = 1
      $while ACC_SLICE < ACCUMULATORS:
        $for A in range(0, ACCUMULATORS, ACC_SLICE * 2):
          $if A + ACC_SLICE < ACCUMULATORS:
            vacc${A} = _mm256_add_ps(vacc${A}, vacc${A + ACC_SLICE});
        $ACC_SLICE *= 2
  for (; batch >= 8 * sizeof(float); batch -= 8 * sizeof(float)) {
    const __m256 vt = _mm256_loadu_ps(input);
    input += 8;

    vacc0 = _mm256_add_ps(vacc0, 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);
    vacc0 = _mm256_add_ps(vacc0, vt);
  }
  __m128 vacc = _mm_add_ps(_mm256_castps256_ps128(vacc0), _mm256_extractf128_ps(vacc0, 1));
  vacc = _mm_add_ps(vacc, _mm_movehl_ps(vacc, vacc));
  vacc = _mm_add_ss(vacc, _mm_movehdup_ps(vacc));
  vacc = _mm_mul_ss(vacc, _mm_load_ss(&params->scalar.scale));
  *output += _mm_cvtss_f32(vacc);
}