// 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.

$assert MR % 4 == 0
$assert NR in [1, 2, 4]
$ABC = "0123456789ABCDEFGHIJK"
#include <assert.h>

#include <arm_neon.h>

#include <xnnpack/spmm.h>


void xnn_f32_spmm_ukernel_${MR}x${NR}__${"neonfma" if FMA else "neon"}(
    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 float32x4_t vmin = vld1q_dup_f32(&params->scalar.min);
  const float32x4_t vmax = vld1q_dup_f32(&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;
    size_t j = n;
    while (j >= ${NR}) {
      uint32_t nnz = *nnzmap++;
      $for N in range(0, NR, 1):
        float32x4_t vacc${ABC[0:4]}c${N} = vld1q_dup_f32(w); w += 1;
        $for M in range(4, MR, 4):
          float32x4_t vacc${ABC[M:M+4]}c${N} = vacc${ABC[0:4]}c${N};
      if XNN_LIKELY(nnz != 0) {
        do {
          const intptr_t diff = *dmap++;
          const float32x4_t va${ABC[0:4]} = vld1q_f32(a);
          $for M in range(4, MR, 4):
            const float32x4_t va${ABC[M:M+4]} = vld1q_f32(a + ${M});
          $if MR == 16:
            __builtin_prefetch(a + 16);
          a = (const float*restrict) ((uintptr_t) a + (uintptr_t) diff);
          $if NR == 1:
            const float32x4_t vb = vld1q_dup_f32(w); w += 1;
          $elif NR == 2:
            const float32x2_t vb = vld1_f32(w); w += 2;
          $elif NR == 4:
            const float32x4_t vb = vld1q_f32(w); w += 4;

          $if NR == 1:
            $for M in range(0, MR, 4):
              vacc${ABC[M:M+4]}c0 = vfmaq_f32(vacc${ABC[M:M+4]}c0, va${ABC[M:M+4]}, vb);
          $else:
            $for N in range(NR):
              $for M in range(0, MR, 4):
                vacc${ABC[M:M+4]}c${N} = vfmaq_lane${"q" if NR == 4 else ""}_f32(vacc${ABC[M:M+4]}c${N}, va${ABC[M:M+4]}, vb, ${N});
        } while (--nnz != 0);
      }
      $for N in range(0, NR, 1):
        $for M in range(0, MR, 4):
          float32x4_t vout${ABC[M:M+4]}c${N} = vminq_f32(vacc${ABC[M:M+4]}c${N}, vmax);

      $for N in range(0, NR, 1):
        $for M in range(0, MR, 4):
          vout${ABC[M:M+4]}c${N} = vmaxq_f32(vout${ABC[M:M+4]}c${N}, vmin);

      $for N in range(0, NR, 1):
        $for M in range(0, MR, 4):
          vst1q_f32(c + ${N} * m + ${M}, vout${ABC[M:M+4]}c${N});
      c += ${NR} * m;
      j -= ${NR};
    }

    // clean up loop, fall back to nr=1
    if XNN_UNLIKELY(j != 0) {
      do {
        uint32_t nnz = *nnzmap++;
        float32x4_t vacc${ABC[0:4]} = vld1q_dup_f32(w); w += 1;
        $for M in range(4, MR, 4):
          float32x4_t vacc${ABC[M:M+4]} = vacc${ABC[0:4]};
        if XNN_LIKELY(nnz != 0) {
          do {
            const intptr_t diff = *dmap++;
            const float32x4_t va${ABC[0:4]} = vld1q_f32(a);
            $for M in range(4, MR, 4):
              const float32x4_t va${ABC[M:M+4]} = vld1q_f32(a + ${M});
            a = (const float*restrict) ((uintptr_t) a + (uintptr_t) diff);
            const float32x4_t vb = vld1q_dup_f32(w); w += 1;
            $for M in range(0, MR, 4):
              vacc${ABC[M:M+4]} = vfmaq_f32(vacc${ABC[M:M+4]}, va${ABC[M:M+4]}, vb);
          } while (--nnz != 0);
        }
        $for M in range(0, MR, 4):
          float32x4_t vout${ABC[M:M+4]} = vminq_f32(vacc${ABC[M:M+4]}, vmax);

        $for M in range(0, MR, 4):
          vout${ABC[M:M+4]} = vmaxq_f32(vout${ABC[M:M+4]}, vmin);

        $for M in range(0, MR, 4):
          vst1q_f32(c + ${M}, vout${ABC[M:M+4]});
        c += m;
        j -= 1;
      } 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;
        size_t j = n;
        while (j >= ${NR}) {
          uint32_t nnz = *nnzmap++;
          $for N in range(0, NR, 1):
            $if SUBMR < 4:
              float32x2_t vacc${ABC[0:SUBMR]}c${N} = vld1_dup_f32(w); w += 1;
            $else:
              float32x4_t vacc${ABC[0:4]}c${N} = vld1q_dup_f32(w); w += 1;
            $for M in range(4, SUBMR, 4):
              float32x4_t vacc${ABC[M:M+4]}c${N} = vacc${ABC[0:4]}c${N};
          if XNN_LIKELY(nnz != 0) {
            do {
              const intptr_t diff = *dmap++;
              $if SUBMR == 1:
                const float32x2_t va${ABC[0]} = vld1_dup_f32(a);
              $elif SUBMR == 2:
                const float32x2_t va${ABC[0:2]} = vld1_f32(a);
              $else:
                const float32x4_t va${ABC[0:4]} = vld1q_f32(a);
              $for M in range(4, SUBMR, 4):
                const float32x4_t va${ABC[M:M+4]} = vld1q_f32(a + ${M});
              a = (const float*restrict) ((uintptr_t) a + (uintptr_t) diff);
              $if NR == 1:
                $if SUBMR < 4:
                  const float32x2_t vb = vld1_dup_f32(w); w += 1;
                $else:
                  const float32x4_t vb = vld1q_dup_f32(w); w += 1;
              $elif NR == 2:
                const float32x2_t vb = vld1_f32(w); w += 2;
              $elif NR == 4:
                const float32x4_t vb = vld1q_f32(w); w += 4;

              $if NR == 1:
                $if SUBMR < 4:
                    vacc${ABC[0:SUBMR]}c0 = vfmaq_f32(vacc${ABC[0:SUBMR]}c0, va${ABC[0:SUBMR]}, vb);
                $else:
                  $for M in range(0, SUBMR, 4):
                    vacc${ABC[M:M+4]}c0 = vfmaq_f32(vacc${ABC[M:M+4]}c0, va${ABC[M:M+4]}, vb);
              $else:
                $for N in range(NR):
                  $if SUBMR < 4:
                    vacc${ABC[0:SUBMR]}c${N} = vfma_lane${"q" if NR == 4 else ""}_f32(vacc${ABC[0:SUBMR]}c${N}, va${ABC[0:SUBMR]}, vb, ${N});
                  $else:
                    $for M in range(0, SUBMR, 4):
                      vacc${ABC[M:M+4]}c${N} = vfmaq_lane${"q" if NR == 4 else ""}_f32(vacc${ABC[M:M+4]}c${N}, va${ABC[M:M+4]}, vb, ${N});
            } while (--nnz != 0);
          }
          $for N in range(0, NR, 1):
            $if SUBMR < 4:
              float32x2_t vout${ABC[0:SUBMR]}c${N} = vmin_f32(vacc${ABC[0:SUBMR]}c${N}, vget_low_f32(vmax));
            $else:
              $for M in range(0, SUBMR, 4):
                float32x4_t vout${ABC[M:M+4]}c${N} = vminq_f32(vacc${ABC[M:M+4]}c${N}, vmax);

          $for N in range(0, NR, 1):
            $if SUBMR < 4:
              vout${ABC[0:SUBMR]}c${N} = vmax_f32(vout${ABC[0:SUBMR]}c${N}, vget_low_f32(vmin));
            $else:
              $for M in range(0, SUBMR, 4):
                vout${ABC[M:M+4]}c${N} = vmaxq_f32(vout${ABC[M:M+4]}c${N}, vmin);

          $for N in range(0, NR, 1):
            $if SUBMR == 1:
              vst1_lane_f32(c + ${N} * m + ${M}, vout${ABC[0:SUBMR]}c${N}, 0);
            $elif SUBMR == 2:
              vst1_f32(c + ${N} * m + ${M}, vout${ABC[0:SUBMR]}c${N});
            $else:
              $for M in range(0, SUBMR, 4):
                vst1q_f32(c + ${N} * m + ${M}, vout${ABC[M:M+4]}c${N});
          c += ${NR} * m;
          j -= ${NR};
        }

        // clean up loop, fall back to nr=1
        if XNN_UNLIKELY(j != 0) {
          do {
            uint32_t nnz = *nnzmap++;
            $if SUBMR < 4:
              float32x2_t vacc${ABC[0:SUBMR]} = vld1_dup_f32(w); w += 1;
            $else:
              float32x4_t vacc${ABC[0:4]} = vld1q_dup_f32(w); w += 1;
            $for M in range(4, SUBMR, 4):
              float32x4_t vacc${ABC[M:M+4]} = vacc${ABC[0:4]};
            if XNN_LIKELY(nnz != 0) {
              do {
                const intptr_t diff = *dmap++;
                $if SUBMR == 1:
                  const float32x2_t va${ABC[0:1]} = vld1_dup_f32(a);
                $elif SUBMR == 2:
                  const float32x2_t va${ABC[0:2]} = vld1_f32(a);
                $else:
                  const float32x4_t va${ABC[0:4]} = vld1q_f32(a);
                $for M in range(4, SUBMR, 4):
                  const float32x4_t va${ABC[M:M+4]} = vld1q_f32(a + ${M});
                a = (const float*restrict) ((uintptr_t) a + (uintptr_t) diff);
                $if SUBMR < 4:
                  const float32x2_t vb = vld1_dup_f32(w); w += 1;
                  vacc${ABC[0:SUBMR]} = vfma_f32(vacc${ABC[0:SUBMR]}, va${ABC[0:SUBMR]}, vb);
                $else:
                  const float32x4_t vb = vld1q_dup_f32(w); w += 1;
                  $for M in range(0, SUBMR, 4):
                    vacc${ABC[M:M+4]} = vfmaq_f32(vacc${ABC[M:M+4]}, va${ABC[M:M+4]}, vb);
              } while (--nnz != 0);
            }
            $if SUBMR < 4:
              float32x2_t vout${ABC[0:SUBMR]} = vmin_f32(vacc${ABC[0:SUBMR]}, vget_low_f32(vmax));
              vout${ABC[0:SUBMR]} = vmax_f32(vout${ABC[0:SUBMR]}, vget_low_f32(vmin));
            $else:
              $for M in range(0, SUBMR, 4):
                float32x4_t vout${ABC[M:M+4]} = vminq_f32(vacc${ABC[M:M+4]}, vmax);

              $for M in range(0, SUBMR, 4):
                vout${ABC[M:M+4]} = vmaxq_f32(vout${ABC[M:M+4]}, vmin);

            $if SUBMR == 1:
              vst1_lane_f32(c, vout${ABC[0:1]}, 1);
            $elif SUBMR == 2:
              vst1_f32(c, vout${ABC[0:2]});
            $else:
              $for M in range(0, SUBMR, 4):
                vst1q_f32(c + ${M}, vout${ABC[M:M+4]});
            c += m;
            j -= 1;
          } while (j != 0);
        }
        c -= m * n;
        c += ${SUBMR};
        a += ${SUBMR};
      }
    }
}