// 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 REQUANTIZATION == "FP32"
$assert DATATYPE in ["QC8", "QS8", "QU8"]
$assert 1 <= MR <= 2
$assert 1 <= NR <= 2
#include <assert.h>

#include <arm_acle.h>

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


$PARAMS_STRUCT = REQUANTIZATION.lower() + "_armsimd32"
$PARAMS_UNION = "xnn_%s_conv_minmax_params" % DATATYPE.lower()
$__XXTB16 = "__uxtb16" if DATATYPE == "QU8" else "__sxtb16"
$__XSAT = "__usat" if DATATYPE == "QU8" else "__ssat"
$__XSUB8 = "__usub8" if DATATYPE == "QU8" else "__ssub8"
$XINT8_T = "uint8_t" if DATATYPE == "QU8" else "int8_t"
void xnn_${DATATYPE.lower()}_gemm_minmax_${REQUANTIZATION.lower()}_ukernel_${MR}x${NR}c4__armsimd32(
    size_t mr,
    size_t nc,
    size_t kc,
    const ${XINT8_T}* restrict a,
    size_t a_stride,
    const void* restrict w,
    ${XINT8_T}* restrict c,
    size_t cm_stride,
    size_t cn_stride,
    const union ${PARAMS_UNION} params[restrict XNN_MIN_ELEMENTS(1)])
{
  assert(mr != 0);
  assert(mr <= ${MR});
  assert(nc != 0);
  assert(kc != 0);

  kc = round_up_po2(kc, 4 * sizeof(int8_t));
  const ${XINT8_T}* a0 = a;
  ${XINT8_T}* c0 = c;
  $for M in range(1, MR):
    const ${XINT8_T}* a${M} = (const ${XINT8_T}*) ((uintptr_t) a${M-1} + a_stride);
    ${XINT8_T}* c${M} = (${XINT8_T}*) ((uintptr_t) c${M-1} + cm_stride);
    $if M % 2 == 0:
      if XNN_UNPREDICTABLE(mr <= ${M}) {
        a${M} = a${M-1};
        c${M} = c${M-1};
      }
    $elif M + 1 == MR:
      if XNN_UNPREDICTABLE(mr != ${M+1}) {
        a${M} = a${M-1};
        c${M} = c${M-1};
      }
    $else:
      if XNN_UNPREDICTABLE(mr < ${M+1}) {
        a${M} = a${M-1};
        c${M} = c${M-1};
      }

  $if DATATYPE == "QU8":
    const int16x2_t vb_minus_zero_point = (int16x2_t) params->${PARAMS_STRUCT}.minus_kernel_zero_point;
  $if REQUANTIZATION == "FP32":
    $if DATATYPE != "QC8":
      const float vscale = params->${PARAMS_STRUCT}.scale;
    const float vmagic_bias = params->${PARAMS_STRUCT}.magic_bias;
  do {
    $for N in range(NR):
      int32_t vacc0x${N} = ((const int32_t*) w)[${N}];
    $for M in range(1, MR):
      $for N in range(NR):
        int32_t vacc${M}x${N} = vacc0x${N};
    w = (const void*) ((const int32_t*) w + ${NR});

    size_t k = kc;
    do {
      $for M in range(MR):
        const int8x4_t va${M} = (int8x4_t) unaligned_load_s32(a${M}); a${M} += 4;

      $for M in range(MR):
        const int16x2_t va${M}c02 = ${__XXTB16}(va${M});
        const int16x2_t va${M}c13 = ${__XXTB16}(__ror(va${M}, 8));

      $for N in range(NR):
        const int8x4_t vb${N} = *((const int8x4_t*) w); w = (const int8_t*) w + 4;
        $if DATATYPE == "QU8":
          const int16x2_t vb${N}c02 = __uxtab16(vb_minus_zero_point, vb${N});
        $else:
          const int16x2_t vb${N}c02 = __sxtb16(vb${N});

        $for M in range(MR):
          vacc${M}x${N} = __smlad(va${M}c02, vb${N}c02, vacc${M}x${N});

        $if DATATYPE == "QU8":
          const int16x2_t vb${N}c13 = __uxtab16(vb_minus_zero_point, __ror(vb${N}, 8));
        $else:
          const int16x2_t vb${N}c13 = __sxtb16(__ror(vb${N}, 8));
        $for M in range(MR):
          vacc${M}x${N} = __smlad(va${M}c13, vb${N}c13, vacc${M}x${N});

      k -= 4 * sizeof(${XINT8_T});
    } while (k != 0);

    $for M in range(MR):
      $for N in range(NR):
        float vfpacc${M}x${N} = (float) vacc${M}x${N};

    $if DATATYPE == "QC8":
      $for N in range(NR):
        const float vscale${N} = ((const float*) w)[${N}];
        $for M in range(MR):
          vfpacc${M}x${N} *= vscale${N};
      w = (const void*) ((const float*) w + ${NR});
    $else:
      $for M in range(MR):
        $for N in range(NR):
          vfpacc${M}x${N} *= vscale;

    $for M in range(MR):
      $for N in range(NR):
        vfpacc${M}x${N} += vmagic_bias;

    $for M in range(MR):
      $for N in range(NR):
        int32_t vout${M}x${N} = (int32_t) float_as_uint32(vfpacc${M}x${N});

    const int32_t vmagic_bias_less_zero_point = params->${PARAMS_STRUCT}.magic_bias_less_zero_point;
    $for M in range(MR):
      $for N in range(NR):
        vout${M}x${N} = __qsub(vout${M}x${N}, vmagic_bias_less_zero_point);

    $for M in range(MR):
      $for N in range(NR):
        vout${M}x${N} = ${__XSAT}(vout${M}x${N}, 8);

    $for M in range(MR):
      $if NR == 1:
        const uint32_t vout${M} = (uint32_t) vout${M}x0;
      $else:
        const uint32_t vout${M} = (uint32_t) (uint8_t) vout${M}x0 | ((uint32_t) vout${M}x1 << 8);

    $if MR == 1:
      uint32_t vout = vout0;
    $else:
      uint32_t vout = (uint32_t) (uint16_t) vout0 | (vout1 << 16);

    const int8x4_t voutput_min = (int8x4_t) params->${PARAMS_STRUCT}.output_min;
    ${__XSUB8}((int8x4_t) vout, voutput_min);
    vout = (uint32_t) __sel((uint8x4_t) vout, (uint8x4_t) voutput_min);

    const int8x4_t voutput_max = (int8x4_t) params->${PARAMS_STRUCT}.output_max;
    ${__XSUB8}((int8x4_t) vout, voutput_max);
    vout = (uint32_t) __sel((uint8x4_t) voutput_max, (uint8x4_t) vout);

    $if NR == 2:
      if XNN_LIKELY(nc >= ${NR}) {
        $for M in range(MR):
          unaligned_store_u16(c${M}, (uint16_t) vout);
          $if M + 1 != MR:
            vout >>= 16;

        $for M in range(MR):
          a${M} = (const ${XINT8_T}*) ((uintptr_t) a${M} - kc);

        $for M in range(MR):
          c${M} = (${XINT8_T}*) ((uintptr_t) c${M} + cn_stride);

        nc -= ${NR};
      } else {
        $for M in range(MR):
          *c${M} = (${XINT8_T}) vout;
          $if M + 1 != MR:
            vout >>= 16;

        nc = 0;
      }
    $else:
      $for M in range(MR):
        *c${M} = (${XINT8_T}) vout;
        $if M + 1 != MR:
          vout >>= 16;

      $for M in range(MR):
        a${M} = (const ${XINT8_T}*) ((uintptr_t) a${M} - kc);

      $for M in range(MR):
        c${M} = (${XINT8_T}*) ((uintptr_t) c${M} + cn_stride);

      nc -= 1;
  } while (nc != 0);
}