// Copyright 2019 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.

$ABC = "0123456789ABCDEFGHIJK"
#include <assert.h>

#include <xnnpack/math.h>
#include <xnnpack/spmm.h>


void xnn_f32_spmm_ukernel_${MR}x${NR}__scalar_pipelined(
    uint32_t m,
    uint32_t n,
    const float*restrict a,
    const float*restrict weights,
    const int32_t*restrict widx_dmap,
    const uint32_t*restrict nidx_nnzmap,
    float*restrict c,
    const union xnn_f32_output_params params[restrict static 1])
{
  assert(m != 0);

  const float vmin = params->scalar.min;
  const float vmax = params->scalar.max;
  size_t i = m;
  while XNN_LIKELY(i >= ${MR}) {
    const float*restrict w = weights;
    const int32_t* dmap = widx_dmap;
    const uint32_t* nnzmap = nidx_nnzmap;
    float vw = *w++;
    intptr_t diff = *dmap++;
    $for M in range(MR):
      float va${ABC[M]} = a[${M}];
    size_t j = n;
    do {
      uint32_t nnz = *nnzmap++;
      $for M in range(MR):
        float vacc${ABC[M]} = vw;
      vw = *w++;
      if XNN_LIKELY(nnz != 0) {
        do {
          $for M in range(MR):
            vacc${ABC[M]} += va${ABC[M]} * vw;
          a = (const float*restrict) ((uintptr_t) a + (uintptr_t) diff);

          diff = *dmap++;
          vw = *w++;
          $for M in range(MR):
            va${ABC[M]} = a[${M}];
        } while (--nnz != 0);
      }
      $for M in range(MR):
        float vout${ABC[M]} = math_min_f32(vacc${ABC[M]}, vmax);
      $for M in range(MR):
        vout${ABC[M]} = math_max_f32(vout${ABC[M]}, vmin);
      $for M in range(MR):
        c[${M}] = vout${ABC[M]};
      c += m;
    } while (--j != 0);
    c -= m * n;
    c += ${MR};
    a += ${MR};
    i -= ${MR};
  }
  if XNN_UNLIKELY(i != 0) {
    $for LOG2M in reversed(range((MR - 1).bit_length())):
      $SUBMR = 1 << LOG2M
      if (i & ${SUBMR}) {
        const float*restrict w = weights;
        const int32_t* dmap = widx_dmap;
        const uint32_t* nnzmap = nidx_nnzmap;
        float vw = *w++;
        intptr_t diff = *dmap++;
        $for M in range(SUBMR):
          float va${ABC[M]} = a[${M}];
        size_t j = n;
        do {
          uint32_t nnz = *nnzmap++;
          $for M in range(SUBMR):
            float vacc${ABC[M]} = vw;
          vw = *w++;
          if XNN_LIKELY(nnz != 0) {
            do {
              $for M in range(SUBMR):
                vacc${ABC[M]} += va${ABC[M]} * vw;
              a = (const float*restrict) ((uintptr_t) a + (uintptr_t) diff);

              diff = *dmap++;
              vw = *w++;
              $for M in range(SUBMR):
                va${ABC[M]} = a[${M}];
            } while (--nnz != 0);
          }
          $for M in range(SUBMR):
            float vout${ABC[M]} = math_min_f32(vacc${ABC[M]}, vmax);
          $for M in range(SUBMR):
            vout${ABC[M]} = math_max_f32(vout${ABC[M]}, vmin);
          $for M in range(SUBMR):
            c[${M}] = vout${ABC[M]};
          c += m;
        } while (--j != 0);
        c -= m * n;
        c += ${SUBMR};
        a += ${SUBMR};
      }
  }
}