src/qu8-gemm/MRxNRc4-minmax-scalar.c.in

// 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.

#include <assert.h>

#include <xnnpack/gemm.h>

#include <xnnpack/scalar-utils.h>

// This kernel is a scalar model for a kernel using ARMv8.2 dot-product
// instructions.
//
// XNN_DISABLE_TSAN is used because this kernel reads up to 3 bytes past the
// bounds of the `a` matrix region, which may be a race condition with
// another thread. We deem this acceptable because the values that are
// read out of bounds do not affect the result, and the the compiler can't know
// about this undefined behavior.
void xnn_qu8_gemm_minmax_ukernel_${MR}x${NR}c4__scalar(
    size_t mr,
    size_t nc,
    size_t kc,
    const uint8_t* restrict a,
    size_t a_stride,
    const void* restrict w,
    uint8_t* restrict c,
    size_t cm_stride,
    size_t cn_stride,
    const union xnn_qu8_gemm_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN {
  assert(mr != 0);
  assert(mr <= ${MR});
  assert(nc != 0);
  assert(kc != 0);

  const uint8_t* a0 = a;
  uint8_t* c0 = c;
  $for M in range(1, MR):
    const uint8_t* a${M} = (const uint8_t*) ((uintptr_t) a${M-1} + a_stride);
    uint8_t* c${M} = (uint8_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};
      }

  const int32_t vb_zero_point = params->scalar.kernel_zero_point;

  // Loop over groups of ${NR} columns.
  do {
    // `vaccMN` is the accumulator at row `M`, column `N`.
    // Initialize accumulators with bias. ${NR} bias values are loaded from the
    // weight matrix, at the start of the group of ${NR} columns.
    $for N in range(NR):
      int32_t bias${N} = ((const int32_t*)w)[${N}];
      $for M in range(MR):
        int32_t vacc${M}${N} = bias${N};

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

    // Inner accumulation loop along the ${NR} columns.
    // Handle 4 rows at each iteration: this is key to modelling what an
    // actual kernel using ARMv8.2 dot-product instructions would look like.
    size_t k = 0;
    while (k < kc) {
      // Load a ${MR}x4 block of activations, and compute sums along rows.
      $for M in range(MR):
        int16_t vasum${M} = 0;
        $for K in range(4):
          int32_t va${M}${K} = *a${M}++;
          vasum${M} += (int16_t) va${M}${K};

      // Load a 4x${NR} block of weights.
      $for N in range(NR):
        $for K in range(4):
          int32_t vb${K}${N} = (int32_t) ((const uint8_t*)w)[${K}];

        w = (const void*) ((uintptr_t) w + 4 * sizeof(uint8_t));

      // Multiply-accumulate: ${MR}x4 * 4x${NR} --> ${MR}x${NR}. The inner size 4 here means
      // we're computing 4D dot-products, which makes this a model for
      // a ARMv8.2 dot-product kernel.
      $for M in range(MR):
        $for N in range(NR):
          $for K in range(4):
            vacc${M}${N} += va${M}${K} * vb${K}${N};
          vacc${M}${N} -= ((int32_t) vasum${M}) * vb_zero_point;

      k += 4 * sizeof(uint8_t);
    }
    // End of accumulation loop. The variable `k` contains the amount by which
    // we advanced the `va` pointers, so we rewind by this amount now.
    $for M in range(MR):
      a${M} = (const uint8_t*)((uintptr_t) a${M} - k);

    // Post-accumulation work

    const int32_t vmultiplier = params->scalar.multiplier;
    const int64_t vq31rounding = INT64_C(0x40000000);
    const int32_t vremainder_mask = params->scalar.remainder_mask;
    const uint32_t vshift = params->scalar.shift;
    const int32_t vremainder_threshold = params->scalar.remainder_threshold;
    const int32_t voutput_min = params->scalar.output_min_less_zero_point;
    const int32_t voutput_max = params->scalar.output_max_less_zero_point;
    const int32_t voutput_zero_point = params->scalar.output_zero_point;

    $for M in range(MR):
      $for N in range(NR):
        const int64_t vproduct${M}${N} = (int64_t)vacc${M}${N} * (int64_t)vmultiplier;

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

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

    $for M in range(MR):
      $for N in range(NR):
        int32_t vout${M}${N} = asr_s32(vq31product${M}${N}, vshift) + (int32_t)(vremainder${M}${N} > vremainder_threshold);

    $for M in range(MR):
      $for N in range(NR):
        vout${M}${N} = vout${M}${N} < voutput_min ? voutput_min : vout${M}${N};

    $for M in range(MR):
      $for N in range(NR):
        vout${M}${N} = vout${M}${N} > voutput_max ? voutput_max : vout${M}${N};

    $for M in range(MR):
      $for N in range(NR):
        vout${M}${N} += voutput_zero_point;

    if XNN_LIKELY (nc >= ${NR}) {
      // Main case where there the ${NR} columns fit in the destination.
      $for M in range(MR):
        $for N in range(NR):
          c${M}[${N}] = vout${M}${N};

      // Advance to the next ${NR} columns.
      $for M in range(MR):
        c${M} = (uint8_t*)((uintptr_t) c${M} + cn_stride);

      nc -= ${NR};
    } else {
      // Final case where not all of the ${NR} columns fit in the destination.
      $for N in range(NR):
        if (nc > ${N}) {
          $for M in range(MR):
            c${M}[${N}] = vout${M}${N};
        }

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