// Copyright 2020 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 CHANNEL_TILE % 8 == 0
$assert CHANNEL_TILE >= 8
$assert ROW_TILE >= 1
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
#include <assert.h>

#include <arm_neon.h>

#include <xnnpack/math.h>
#include <xnnpack/prelu.h>


void xnn_f16_prelu_ukernel__neonfp16arith_${ROW_TILE}x${CHANNEL_TILE}(
    size_t rows,
    size_t channels,
    const void* restrict input,
    size_t input_stride,
    const void* restrict weights,
    void* restrict output,
    size_t output_stride) XNN_DISABLE_TSAN
{
  assert(rows != 0);
  assert(channels != 0);
  assert(channels % sizeof(__fp16) == 0);

  const __fp16* i0 = (const __fp16*) input;
  __fp16* o0 = (__fp16*) output;

  $for M in range(1, ROW_TILE):
    const __fp16* i${M} = (const __fp16*) ((uintptr_t) i${M-1} + input_stride);
    __fp16* o${M} = (__fp16*) ((uintptr_t) o${M-1} + output_stride);
    $if M % 2 == 0:
      if XNN_UNPREDICTABLE(rows <= ${M}) {
        i${M} = i${M-1};
        o${M} = o${M-1};
      }
    $else:
      if XNN_UNPREDICTABLE(rows < ${M+1}) {
        i${M} = i${M-1};
        o${M} = o${M-1};
      }

  const size_t input_increment = input_stride * ${ROW_TILE} - channels;
  const size_t output_increment = output_stride * ${ROW_TILE} - channels;

  do {
    const __fp16* w = (const __fp16*) weights;
    size_t c = channels;
    for (; c >= ${CHANNEL_TILE} * sizeof(__fp16); c -= ${CHANNEL_TILE} * sizeof(__fp16)) {
      $for C in range(0, CHANNEL_TILE, 8):
        const float16x8_t vw${ABC[C:C+8]} = vld1q_f16(w); w += 8;

      $for M in range(ROW_TILE):
        $for C in range(0, CHANNEL_TILE, 8):
          const float16x8_t vi${M}x0${ABC[C:C+8]} = vld1q_f16(i${M}); i${M} += 8;

      $for M in range(ROW_TILE):
        $for C in range(0, CHANNEL_TILE, 8):
          float16x8_t vacc${M}x0${ABC[C:C+8]} = vmulq_f16(vi${M}x0${ABC[C:C+8]}, vw${ABC[C:C+8]});
          const uint16x8_t vm${M}x0${ABC[C:C+8]} = vcltq_s16(vreinterpretq_s16_f16(vi${M}x0${ABC[C:C+8]}), vmovq_n_s16(0));

      $for M in range(ROW_TILE):
        $for C in range(0, CHANNEL_TILE, 8):
          vacc${M}x0${ABC[C:C+8]} = vbslq_f16(vm${M}x0${ABC[C:C+8]}, vacc${M}x0${ABC[C:C+8]}, vi${M}x0${ABC[C:C+8]});

      $for M in range(ROW_TILE):
        $for C in range(0, CHANNEL_TILE, 8):
          vst1q_f16(o${M}, vacc${M}x0${ABC[C:C+8]}); o${M} += 8;
    }
    $if CHANNEL_TILE != 8:
      for (; c >= 8 * sizeof(__fp16); c -= 8 * sizeof(__fp16)) {
        const float16x8_t vw01234567 = vld1q_f16(w); w += 8;

        $for M in range(ROW_TILE):
          const float16x8_t vi${M}x01234567 = vld1q_f16(i${M});
          i${M} += 8;

        $for M in range(ROW_TILE):
          float16x8_t vacc${M}x01234567 = vmulq_f16(vi${M}x01234567, vw01234567);
          const uint16x8_t vm${M}x01234567 = vcltq_s16(vreinterpretq_s16_f16(vi${M}x01234567), vmovq_n_s16(0));

        $for M in range(ROW_TILE):
          vacc${M}x01234567 = vbslq_f16(vm${M}x01234567, vacc${M}x01234567, vi${M}x01234567);

        $for M in range(ROW_TILE):
          vst1q_f16(o${M}, vacc${M}x01234567); o${M} += 8;
      }
    if XNN_UNLIKELY(c != 0) {
      const float16x8_t vw01234567 = vld1q_f16(w); w += 8;

      $for M in range(ROW_TILE):
        const float16x8_t vi${M}x01234567 = vld1q_f16(i${M});
        i${M} = (const __fp16*) ((uintptr_t) i${M} + c);

      $for M in range(ROW_TILE):
        float16x8_t vacc${M}x01234567 = vmulq_f16(vi${M}x01234567, vw01234567);
        const uint16x8_t vm${M}x01234567 = vcltq_s16(vreinterpretq_s16_f16(vi${M}x01234567), vmovq_n_s16(0));

      $for M in range(ROW_TILE):
        vacc${M}x01234567 = vbslq_f16(vm${M}x01234567, vacc${M}x01234567, vi${M}x01234567);

      $for M in range(ROW_TILE):
        float16x4_t vacc${M}x0123 = vget_low_f16(vacc${M}x01234567);
      if (c & (4 * sizeof(__fp16))) {
        $for M in range(ROW_TILE):
          vst1_f16(o${M}, vacc${M}x0123); o${M} += 4;

        $for M in range(ROW_TILE):
          vacc${M}x0123 = vget_high_f16(vacc${M}x01234567);
      }
      if (c & (2 * sizeof(__fp16))) {
        $for M in range(ROW_TILE):
          vst1_lane_u32(__builtin_assume_aligned(o${M}, 1), vreinterpret_u32_f16(vacc${M}x0123), 0); o${M} += 2;
          vacc${M}x0123 = vext_f16(vacc${M}x0123, vacc${M}x0123, 2);
      }
      if (c & (1 * sizeof(__fp16))) {
        $for M in range(ROW_TILE):
          vst1_lane_f16(o${M}, vacc${M}x0123, 0); o${M} += 1;
      }
    }
    $for M in range(ROW_TILE):
      i${M} = (const __fp16*) ((uintptr_t) i${M} + input_increment);
      o${M} = (__fp16*) ((uintptr_t) o${M} + output_increment);
      $if M % 2 == 1:
        if XNN_UNPREDICTABLE(rows < ${ROW_TILE+M+1}) {
          i${M} = i${M-1};
          o${M} = o${M-1};
        }
      $elif M != 0:
        if XNN_UNPREDICTABLE(rows <= ${ROW_TILE+M}) {
          i${M} = i${M-1};
          o${M} = o${M-1};
        }
    rows = doz(rows, ${ROW_TILE});
  } while (rows != 0);
}