src/qs8-gemm/scalar.c.in

// 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.

#include <assert.h>

#include <xnnpack/math.h>
#include <xnnpack/scalar-utils.h>
#include <xnnpack/gemm.h>


void xnn_qs8_gemm_minmax_gemmlowp_ukernel_${MR}x${NR}__scalar(
    size_t mr,
    size_t nc,
    size_t kc,
    const int8_t* restrict a,
    size_t a_stride,
    const void* restrict w,
    int8_t* restrict c,
    size_t cm_stride,
    size_t cn_stride,
    const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
{
  assert(mr != 0);
  assert(mr <= ${MR});
  assert(nc != 0);
  assert(kc != 0);

  const int8_t* a0 = a;
  int8_t* c0 = c;
  $for M in range(1, MR):
    const int8_t* a${M} = (const int8_t*) ((uintptr_t) a${M-1} + a_stride);
    int8_t* c${M} = (int8_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};
      }

  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*) ((uintptr_t) w + ${NR} * sizeof(int32_t));

    size_t k = kc;
    do {
      $for M in range(MR):
        const int32_t va${M} = (int32_t) *a${M}++;

      $for N in range(NR):
        const int32_t vb${N} = ((const int8_t*) w)[${N}];
      w = (const void*) ((uintptr_t) w + ${NR} * sizeof(int8_t));

      $for M in range(MR):
        $for N in range(NR):
          vacc${M}x${N} += va${M} * vb${N};

      k -= sizeof(int8_t);
    } while (k != 0);

    const int32_t vmultiplier = params->scalar.multiplier;
    $for M in range(MR):
      $for N in range(NR):
        const int64_t vproduct${M}x${N} = (int64_t) vacc${M}x${N} * (int64_t) vmultiplier;

    const int64_t vq31rounding = INT64_C(0x40000000);
    $for M in range(MR):
      $for N in range(NR):
        const int32_t vq31product${M}x${N} = (int32_t) (uint32_t) ((uint64_t) (vproduct${M}x${N} + vq31rounding) >> 31);

    const int32_t vremainder_mask = params->scalar.remainder_mask;
    $for M in range(MR):
      $for N in range(NR):
        const int32_t vremainder${M}x${N} = (vq31product${M}x${N} & vremainder_mask) - (int32_t) (vq31product${M}x${N} < 0);

    const uint32_t vshift = params->scalar.shift;
    const int32_t vremainder_threshold = params->scalar.remainder_threshold;
    $for M in range(MR):
      $for N in range(NR):
        int32_t vout${M}x${N} = asr_s32(vq31product${M}x${N}, vshift) + (int32_t) (vremainder${M}x${N} > vremainder_threshold);

    const int32_t vout_min = params->scalar.output_min_less_zero_point;
    $for M in range(MR):
      $for N in range(NR):
        vout${M}x${N} = math_max_s32(vout${M}x${N}, vout_min);

    const int32_t vout_max = params->scalar.output_max_less_zero_point;
    $for M in range(MR):
      $for N in range(NR):
        vout${M}x${N} = math_min_s32(vout${M}x${N}, vout_max);

    const int32_t voutput_zero_point = params->scalar.output_zero_point;
    $for M in range(MR):
      $for N in range(NR):
        vout${M}x${N} += voutput_zero_point;

    if XNN_LIKELY(nc >= ${NR}) {
      $for M in range(MR):
        $for N in range(NR):
          c${M}[${N}] = (int8_t) vout${M}x${N};

      $for M in range(MR):
        a${M} = (const int8_t*) ((uintptr_t) a${M} - kc);

      $for M in range(MR):
        c${M} = (int8_t*) ((uintptr_t) c${M} + cn_stride);

      nc -= ${NR};
    } else {
      $for LOG2N in reversed(range(NR.bit_length() - 1)):
        if (nc & ${1 << LOG2N}) {
          $for M in range(MR):
            $for N in range(1 << LOG2N):
              c${M}[${N}] = (int8_t) vout${M}x${N};
            $if LOG2N != 0:
              $for N in range(1 << (LOG2N - 1)):
                vout${M}x${N} = vout${M}x${N + (1 << LOG2N)};
              c${M} += ${1 << LOG2N};
        }

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