// 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 = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
#include <assert.h>

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


void xnn_f32_spmm_minmax_ukernel_${MR}x${NR}__scalar_pipelined(
    size_t mc,
    size_t nc,
    const float*restrict input,
    const float*restrict weights,
    const int32_t*restrict widx_dmap,
    const uint32_t*restrict nidx_nnzmap,
    float*restrict output,
    size_t output_stride,
    const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
{
  assert(mc != 0);
  assert(mc % sizeof(float) == 0);
  assert(nc != 0);

  const float vmin = params->scalar.min;
  const float vmax = params->scalar.max;
  size_t output_decrement = output_stride * nc - ${MR} * sizeof(float);
  while XNN_LIKELY(mc >= ${MR} * sizeof(float)) {
    const float*restrict w = weights;
    const int32_t* dmap = widx_dmap;
    const uint32_t* nnzmap = nidx_nnzmap;
    float vw = *w++;
    intptr_t diff = *dmap++;
    $for M in range(MR):
      float vi${ABC[M]} = input[${M}];
    size_t n = nc;
    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]} += vi${ABC[M]} * vw;
          input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff);

          diff = *dmap++;
          vw = *w++;
          $for M in range(MR):
            vi${ABC[M]} = input[${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);
      $for M in range(MR):
        output[${M}] = vout${ABC[M]};
      output = (float*restrict) ((uintptr_t) output + output_stride);
    } while (--n != 0);
    output = (float*restrict) ((uintptr_t) output - output_decrement);
    input += ${MR};
    mc -= ${MR} * sizeof(float);
  }
  if XNN_UNLIKELY(mc != 0) {
    $for LOG2M in reversed(range((MR - 1).bit_length())):
      $SUBMR = 1 << LOG2M
      $if SUBMR * 2 >= MR:
        output_decrement += ${MR - SUBMR} * sizeof(float);
      $else:
        output_decrement += ${SUBMR} * sizeof(float);
      if (mc & (${SUBMR} * sizeof(float))) {
        const float*restrict w = weights;
        const int32_t* dmap = widx_dmap;
        const uint32_t* nnzmap = nidx_nnzmap;
        float vw = *w++;
        intptr_t diff = *dmap++;
        $for M in range(SUBMR):
          float vi${ABC[M]} = input[${M}];
        size_t n = nc;
        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]} += vi${ABC[M]} * vw;
              input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff);

              diff = *dmap++;
              vw = *w++;
              $for M in range(SUBMR):
                vi${ABC[M]} = input[${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);
          $for M in range(SUBMR):
            output[${M}] = vout${ABC[M]};
          output = (float*restrict) ((uintptr_t) output + output_stride);
        } while (--n != 0);
        output = (float*restrict) ((uintptr_t) output - output_decrement);
        input += ${SUBMR};
      }
  }
}