// 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 <arm_neon.h>

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


$ACC_SUFFIX = "" if ACCUMULATORS == 1 else "_acc%d" % ACCUMULATORS
void xnn_f32_rsum_ukernel__neon_u${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);

  $for A in range(ACCUMULATORS):
    float32x4_t vacc${A} = vmovq_n_f32(0.0f);
  $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;

      $for N in range(SIMD_TILE):
        vacc${N % ACCUMULATORS} = vaddq_f32(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} = vaddq_f32(vacc${A}, vacc${A + ACC_SLICE});
        $ACC_SLICE *= 2
  for (; batch >= 4 * sizeof(float); batch -= 4 * sizeof(float)) {
    const float32x4_t vt = vld1q_f32(input); input += 4;
    vacc0 = vaddq_f32(vacc0, vt);
  }
  const float32x2_t vscale = vld1_dup_f32(&params->scalar.scale);
  float32x2_t vacc = vadd_f32(vget_low_f32(vacc0), vget_high_f32(vacc0));
  if XNN_UNLIKELY(batch & (2 * sizeof(float))) {
    const float32x2_t vt = vld1_f32(input); input += 2;
    vacc = vadd_f32(vacc, vt);
  }
  vacc = vpadd_f32(vacc, vacc);
  if XNN_UNLIKELY(batch & (1 * sizeof(float))) {
    const float32x2_t vt = vld1_dup_f32(input);
    vacc = vadd_f32(vacc, vt);
  }
  vacc = vmul_f32(vacc, vscale);
  *output += vget_lane_f32(vacc, 0);
}