// Copyright 2022 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 DATATYPE in ["QS8", "QU8"]
$assert BATCH_TILE % 4 == 0
$SIMD_TILE = BATCH_TILE // 4
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
#include <assert.h>

#include <arm_acle.h>

#include <xnnpack/intrinsics-polyfill.h>
#include <xnnpack/math.h>
#include <xnnpack/unaligned.h>
#include <xnnpack/vcvt.h>


$XINT8_T = {"QS8": "int8_t", "QU8": "uint8_t"}[DATATYPE]
$XINT8X4_T = {"QS8": "int8x4_t", "QU8": "uint8x4_t"}[DATATYPE]
$XINT16X2_T = {"QS8": "int16x2_t", "QU8": "uint16x2_t"}[DATATYPE]
$__XXTAB16 = {"QS8": "__sxtab16", "QU8": "__uxtab16"}[DATATYPE]
$__XSAT = {"QS8": "__ssat", "QU8": "__usat"}[DATATYPE]
void xnn_${DATATYPE.lower()}_vcvt_ukernel__armsimd32_x${BATCH_TILE}(
    size_t n,
    const ${XINT8_T}* x,
    ${XINT8_T}* y,
    const union xnn_${DATATYPE.lower()}_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  const ${XINT16X2_T} vminus_input_zero_point = (${XINT16X2_T}) params->armsimd32.minus_input_zero_point;
  const int32_t vbias = params->armsimd32.bias;
  const int32_t vmultiplier = params->armsimd32.multiplier;
  $if BATCH_TILE > 4:
    for (; n >= ${BATCH_TILE} * sizeof(${XINT8_T}); n -= ${BATCH_TILE} * sizeof(${XINT8_T})) {
      $for N in range(SIMD_TILE):
        const ${XINT8X4_T} vx${ABC[4*N:4*N+4]} = (${XINT8X4_T}) unaligned_indexed_load_u32(x, ${N});
      x += ${BATCH_TILE};

      $for N in range(0, BATCH_TILE, 4):
        const ${XINT16X2_T} vx${ABC[N]}${ABC[N+2]} = ${__XXTAB16}(vminus_input_zero_point, vx${ABC[N:N+4]});
        const ${XINT16X2_T} vx${ABC[N+1]}${ABC[N+3]} = ${__XXTAB16}(vminus_input_zero_point, __ror(vx${ABC[N:N+4]}, 8));

      $for N in range(0, BATCH_TILE, 4):
        int32_t vacc${ABC[N]} = __smlawb(vmultiplier, vx${ABC[N]}${ABC[N+2]}, vbias);
        int32_t vacc${ABC[N+1]} = __smlawb(vmultiplier, vx${ABC[N+1]}${ABC[N+3]}, vbias);
        int32_t vacc${ABC[N+2]} = __smlawt(vmultiplier, vx${ABC[N]}${ABC[N+2]}, vbias);
        int32_t vacc${ABC[N+3]} = __smlawt(vmultiplier, vx${ABC[N+1]}${ABC[N+3]}, vbias);

      $for N in range(BATCH_TILE):
        vacc${ABC[N]} = ${__XSAT}(math_asr_s32(vacc${ABC[N]}, 1), 8);

      $for N in range(BATCH_TILE):
        y[${N}] = (${XINT8_T}) vacc${ABC[N]};
      y += ${BATCH_TILE};
    }
  for (; n >= 4 * sizeof(${XINT8_T}); n -= 4 * sizeof(${XINT8_T})) {
    const ${XINT8X4_T} vx0123 = (${XINT8X4_T}) unaligned_load_u32(x);
    x += 4;

    const ${XINT16X2_T} vx02 = ${__XXTAB16}(vminus_input_zero_point, vx0123);
    const ${XINT16X2_T} vx13 = ${__XXTAB16}(vminus_input_zero_point, __ror(vx0123, 8));

    int32_t vacc0 = __smlawb(vmultiplier, vx02, vbias);
    int32_t vacc1 = __smlawb(vmultiplier, vx13, vbias);
    int32_t vacc2 = __smlawt(vmultiplier, vx02, vbias);
    int32_t vacc3 = __smlawt(vmultiplier, vx13, vbias);

    vacc0 = ${__XSAT}(math_asr_s32(vacc0, 1), 8);
    vacc1 = ${__XSAT}(math_asr_s32(vacc1, 1), 8);
    vacc2 = ${__XSAT}(math_asr_s32(vacc2, 1), 8);
    vacc3 = ${__XSAT}(math_asr_s32(vacc3, 1), 8);

    y[0] = (${XINT8_T}) vacc0;
    y[1] = (${XINT8_T}) vacc1;
    y[2] = (${XINT8_T}) vacc2;
    y[3] = (${XINT8_T}) vacc3;
    y += 4;
  }
  if XNN_UNLIKELY(n != 0) {
    const ${XINT8X4_T} vx0123 = (${XINT8X4_T}) unaligned_load_u32(x);

    const ${XINT16X2_T} vx02 = ${__XXTAB16}(vminus_input_zero_point, vx0123);
    const ${XINT16X2_T} vx13 = ${__XXTAB16}(vminus_input_zero_point, __ror(vx0123, 8));

    int32_t vacc0 = __smlawb(vmultiplier, vx02, vbias);
    int32_t vacc1 = __smlawb(vmultiplier, vx13, vbias);
    const int32_t vacc2 = __smlawt(vmultiplier, vx02, vbias);

    vacc0 = ${__XSAT}(math_asr_s32(vacc0, 1), 8);
    vacc1 = ${__XSAT}(math_asr_s32(vacc1, 1), 8);

    if (n & (2 * sizeof(${XINT8_T}))) {
      y[0] = (${XINT8_T}) vacc0;
      y[1] = (${XINT8_T}) vacc1;
      vacc0 = ${__XSAT}(math_asr_s32(vacc2, 1), 8);
      y += 2;
    }
    if (n & (1 * sizeof(${XINT8_T}))) {
      y[0] = (${XINT8_T}) vacc0;
    }
  }
}