// 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 <arm_neon.h>

#include <xnnpack/common.h>
#include <xnnpack/reduce.h>


$ACC_SUFFIX = "" if ACCUMULATORS == 1 else "_acc%d" % ACCUMULATORS
void xnn_f16_rsum_ukernel__neonfp16arith_x${BATCH_TILE}${ACC_SUFFIX}(
    size_t batch,
    const void* input,
    void* output,
    const union xnn_f16_scale_params params[restrict XNN_MIN_ELEMENTS(1)])
{
  assert(batch != 0);
  assert(batch % sizeof(uint16_t) == 0);
  assert(input != NULL);
  assert(output != NULL);

  const uint16_t* i = (const uint16_t*) input;
  uint16_t* o = (uint16_t*) output;
  $for A in range(ACCUMULATORS):
    float16x8_t vacc${A} = vreinterpretq_f16_u16(vmovq_n_u16(0));
  $if BATCH_TILE > 8:
    for (; batch >= ${BATCH_TILE} * sizeof(uint16_t); batch -= ${BATCH_TILE} * sizeof(uint16_t)) {
      $for N in range(SIMD_TILE):
        const float16x8_t vt${N} = vreinterpretq_f16_u16(vld1q_u16(i)); i += 8;

      $for N in range(SIMD_TILE):
        vacc${N % ACCUMULATORS} = vaddq_f16(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_f16(vacc${A}, vacc${A + ACC_SLICE});
        $ACC_SLICE *= 2
  for (; batch >= 8 * sizeof(uint16_t); batch -= 8 * sizeof(uint16_t)) {
    const float16x8_t vt = vreinterpretq_f16_u16(vld1q_u16(i)); i += 8;
    vacc0 = vaddq_f16(vacc0, vt);
  }
  const float16x4_t vscale = vreinterpret_f16_u16(vld1_dup_u16(&params->fp16arith.scale));
  float16x4_t vacc = vadd_f16(vget_low_f16(vacc0), vget_high_f16(vacc0));
  if XNN_UNLIKELY(batch & (4 * sizeof(uint16_t))) {
    const float16x4_t vt = vreinterpret_f16_u16(vld1_u16(i)); i += 4;
    vacc = vadd_f16(vacc, vt);
  }
  vacc = vpadd_f16(vacc, vacc);
  if XNN_UNLIKELY(batch & (2 * sizeof(uint16_t))) {
    const float16x4_t vt = vreinterpret_f16_u32(vld1_dup_u32((const void*) i)); i += 2;
    vacc = vadd_f16(vacc, vt);
  }
  vacc = vpadd_f16(vacc, vacc);
  if XNN_UNLIKELY(batch & (1 * sizeof(uint16_t))) {
    const float16x4_t vt = vreinterpret_f16_u16(vld1_dup_u16(i));
    vacc = vadd_f16(vacc, vt);
  }
  vacc = vmul_f16(vacc, vscale);
  vst1_lane_u16(o, vreinterpret_u16_f16(vacc), 0);
}