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

#include <xmmintrin.h>

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


$ACC_SUFFIX = "" if ACCUMULATORS == 1 else "_acc%d" % ACCUMULATORS
void xnn_f32_rsum_ukernel__sse_u${BATCH_TILE}${ACC_SUFFIX}(
    size_t batch,
    const float* input,
    float* output,
    const struct xnn_f32_scale_params params[restrict XNN_MIN_ELEMENTS(1)])
{
  assert(batch != 0);
  assert(batch % sizeof(float) == 0);
  assert(input != NULL);
  assert(output != NULL);

  $for A in range(ACCUMULATORS):
    __m128 vacc${A} = _mm_setzero_ps();
  $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};

      $for N in range(SIMD_TILE):
        vacc${N % ACCUMULATORS} = _mm_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} = _mm_add_ps(vacc${A}, vacc${A + ACC_SLICE});
        $ACC_SLICE *= 2
  for (; batch >= 4 * sizeof(float); batch -= 4 * sizeof(float)) {
    const __m128 vt = _mm_loadu_ps(input);
    input += 4;

    vacc0 = _mm_add_ps(vacc0, vt);
  }
  vacc0 = _mm_add_ps(vacc0, _mm_movehl_ps(vacc0, vacc0));
  if XNN_UNLIKELY(batch != 0) {
    do {
      const __m128 vt = _mm_load_ss(input);
      input += 1;
      vacc0 = _mm_add_ss(vacc0, vt);
      batch -= sizeof(float);
    } while (batch != 0);
  }
  vacc0 = _mm_add_ss(vacc0, _mm_shuffle_ps(vacc0, vacc0, _MM_SHUFFLE(1, 1, 1, 1)));
  vacc0 = _mm_mul_ss(vacc0, _mm_load_ss(&params->scalar.scale));
  *output += _mm_cvtss_f32(vacc0);
}