// Copyright 2021 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
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
#include <assert.h>

#include <arm_neon.h>

#include "xnnpack/common.h"
#include "xnnpack/intrinsics-polyfill.h"
#include "xnnpack/vcvt.h"


$XINT8_T = {"QS8": "int8_t", "QU8": "uint8_t"}[DATATYPE]
$XINT8X8_T = {"QS8": "int8x8_t", "QU8": "uint8x8_t"}[DATATYPE]
$XINT8X16_T = {"QS8": "int8x16_t", "QU8": "uint8x16_t"}[DATATYPE]
$VLD1Q_DUP_X8 = {"QS8": "vld1q_dup_s8", "QU8": "vld1q_dup_u8"}[DATATYPE]
$VLD1_DUP_X8 = {"QS8": "vld1_dup_s8", "QU8": "vld1_dup_u8"}[DATATYPE]
$VST1Q_X8 = {"QS8": "vst1q_s8", "QU8": "vst1q_u8"}[DATATYPE]
$VST1_X8 = {"QS8": "vst1_s8", "QU8": "vst1_u8"}[DATATYPE]
$VST1_LANE_X8 = {"QS8": "vst1_lane_s8", "QU8": "vst1_lane_u8"}[DATATYPE]
$VQMOVXN_S16 = {"QS8": "vqmovn_s16", "QU8": "vqmovun_s16"}[DATATYPE]
$VEXT_X8 = {"QS8": "vext_s8", "QU8": "vext_u8"}[DATATYPE]
$VCOMBINE_X8 = {"QS8": "vcombine_s8", "QU8": "vcombine_u8"}[DATATYPE]
$VGET_LOW_X8 = {"QS8": "vget_low_s8", "QU8": "vget_low_u8"}[DATATYPE]
$VREINTERPRET_U16_X8 = {"QS8": "vreinterpret_u16_s8", "QU8": "vreinterpret_u16_u8"}[DATATYPE]
$VREINTERPRET_U32_X8 = {"QS8": "vreinterpret_u32_s8", "QU8": "vreinterpret_u32_u8"}[DATATYPE]
void xnn_f32_${DATATYPE.lower()}_vcvt_ukernel__neonv8_u${BATCH_TILE}(
    size_t batch,
    const float* input,
    ${XINT8_T}* output,
    const struct xnn_f32_${DATATYPE.lower()}_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(batch != 0);
  assert(batch % sizeof(float) == 0);
  assert(input != NULL);
  assert(output != NULL);

  const float32x4_t vscale = vld1q_dup_f32(&params->scalar.scale);
  const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->scalar.output_zero_point);
  $if BATCH_TILE > 8:
    for (; batch >= ${BATCH_TILE} * sizeof(float); batch -= ${BATCH_TILE} * sizeof(float)) {
      $for N in range(0, BATCH_TILE, 4):
        float32x4_t vx${ABC[N:N+4]} = vld1q_f32(input); input += 4;

      $for N in range(0, BATCH_TILE, 4):
        vx${ABC[N:N+4]} = vmulq_f32(vx${ABC[N:N+4]}, vscale);

      $for N in range(0, BATCH_TILE, 4):
        const int32x4_t vacc${ABC[N:N+4]} = vcvtnq_s32_f32(vx${ABC[N:N+4]});

      $for N in range(0, BATCH_TILE, 8):
        int16x8_t vacc${ABC[N:N+8]} = vcombine_s16(vqmovn_s32(vacc${ABC[N:N+4]}), vqmovn_s32(vacc${ABC[N+4:N+8]}));

      $for N in range(0, BATCH_TILE, 8):
        vacc${ABC[N:N+8]} = vqaddq_s16(vacc${ABC[N:N+8]}, voutput_zero_point);

      $for N in range(0, BATCH_TILE, 16):
        $if N + 8 < BATCH_TILE:
          ${XINT8X16_T} vy${ABC[N:N+16]} = ${VCOMBINE_X8}(${VQMOVXN_S16}(vacc${ABC[N:N+8]}), ${VQMOVXN_S16}(vacc${ABC[N+8:N+16]}));
        $else:
          ${XINT8X8_T} vy${ABC[N:N+8]} = ${VQMOVXN_S16}(vacc${ABC[N:N+8]});

      $for N in range(0, BATCH_TILE, 16):
        $if N + 8 < BATCH_TILE:
          ${VST1Q_X8}(output, vy${ABC[N:N+16]}); output += 16;
        $else:
          ${VST1_X8}(output, vy${ABC[N:N+8]}); output += 8;
    }
  for (; batch >= 8 * sizeof(float); batch -= 8 * sizeof(float)) {
    float32x4_t vx_lo = vld1q_f32(input); input += 4;
    float32x4_t vx_hi = vld1q_f32(input); input += 4;

    vx_lo = vmulq_f32(vx_lo, vscale);
    vx_hi = vmulq_f32(vx_hi, vscale);

    const int32x4_t vacc_lo = vcvtnq_s32_f32(vx_lo);
    const int32x4_t vacc_hi = vcvtnq_s32_f32(vx_hi);

    int16x8_t vacc = vcombine_s16(vqmovn_s32(vacc_lo), vqmovn_s32(vacc_hi));
    vacc = vqaddq_s16(vacc, voutput_zero_point);

    ${XINT8X8_T} vy = ${VQMOVXN_S16}(vacc);
    ${VST1_X8}(output, vy); output += 8;
  }
  if XNN_UNLIKELY(batch != 0) {
    assert(batch >= 1 * sizeof(float));
    assert(batch <= 7 * sizeof(float));
    float32x4_t vx_lo = vld1q_f32(input);
    const float* x_hi = (const float*) ((uintptr_t) input + (batch & (4 * sizeof(float))));
    float32x4_t vx_hi = vld1q_f32(x_hi);

    vx_lo = vmulq_f32(vx_lo, vscale);
    vx_hi = vmulq_f32(vx_hi, vscale);

    const int32x4_t vacc_lo = vcvtnq_s32_f32(vx_lo);
    const int32x4_t vacc_hi = vcvtnq_s32_f32(vx_hi);

    int16x8_t vacc = vcombine_s16(vqmovn_s32(vacc_lo), vqmovn_s32(vacc_hi));
    vacc = vqaddq_s16(vacc, voutput_zero_point);

    ${XINT8X8_T} vy = ${VQMOVXN_S16}(vacc);

    if (batch & (4 * sizeof(float))) {
      vst1_lane_u32((void*) output, ${VREINTERPRET_U32_X8}(vy), 0); output += 4;
      vy = ${VEXT_X8}(vy, vy, 4);
    }
    if (batch & (2 * sizeof(float))) {
      vst1_lane_u16((void*) output, ${VREINTERPRET_U16_X8}(vy), 0); output += 2;
      vy = ${VEXT_X8}(vy, vy, 2);
    }
    if (batch & (1 * sizeof(float))) {
      ${VST1_LANE_X8}(output, vy, 0);
    }
  }
}