// 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 NR % 4 == 0
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
$VMULADDQ_F32 = "vfmaq_f32" if FMA else "vmlaq_f32"
#include <assert.h>

#include <arm_neon.h>

#include <xnnpack/igemm.h>


void xnn_f32_igemm_minmax_ukernel_${MR}x${NR}s4__${"neonfma" if FMA else "neon"}(
    size_t mr,
    size_t nc,
    size_t kc,
    size_t ks,
    const float**restrict a,
    const float*restrict w,
    float*restrict c,
    size_t cm_stride,
    size_t cn_stride,
    size_t a_offset,
    const float* zero,
    const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(mr != 0);
  assert(mr <= ${MR});
  assert(nc != 0);
  assert(kc != 0);
  assert(kc % sizeof(float) == 0);
  assert(ks != 0);
  assert(ks % (${MR} * sizeof(void*)) == 0);
  assert(a_offset % sizeof(float) == 0);
  assert(a != NULL);
  assert(w != NULL);
  assert(c != NULL);

  float* c0 = c;
  $for M in range(1, MR):
    float* c${M} = (float*) ((uintptr_t) c${M-1} + cm_stride);
    $if M % 2 == 0:
      if XNN_UNPREDICTABLE(mr <= ${M}) {
        c${M} = c${M-1};
      }
    $elif M + 1 == MR:
      if XNN_UNPREDICTABLE(mr != ${M+1}) {
        c${M} = c${M-1};
      }
    $else:
      if XNN_UNPREDICTABLE(mr < ${M+1}) {
        c${M} = c${M-1};
      }

  do {
    $for N in range(0, NR, 4):
      float32x4_t vacc0x${ABC[N:N+4]} = vld1q_f32(w); w += 4;
    $for M in range(1, MR):
      $for N in range(0, NR, 4):
        float32x4_t vacc${M}x${ABC[N:N+4]} = vacc0x${ABC[N:N+4]};

    size_t p = ks;
    do {
      $for M in range(MR):
        const float* restrict a${M} = a[${M}];
        assert(a${M} != NULL);
        if XNN_UNPREDICTABLE(a${M} != zero) {
          a${M} = (const float*) ((uintptr_t) a${M} + a_offset);
        }
      a += ${MR};

      size_t k = kc;
      while (k >= 4 * sizeof(float)) {
        $for M in range(MR):
          float32x4_t va${M} = vld1q_f32(a${M}); a${M} += 4;

        $for L in range(4):

          $for N in range(0, NR, 4):
            const float32x4_t vb${ABC[N:N+4]}c${L} = vld1q_f32(w + ${L * NR + N});

          $for N in range(0, NR, 4):
            $for M in range(MR):
              vacc${M}x${ABC[N:N+4]} = ${VMULADDQ_F32}(vacc${M}x${ABC[N:N+4]}, va${M}, vb${ABC[N:N+4]}c${L});

          $if L + 1 != 4:
            $for M in range(MR):
              va${M} = vextq_f32(va${M}, va${M}, 1);

        w += ${4 * NR};
        k -= 4 * sizeof(float);
      }
      if XNN_UNLIKELY(k != 0) {
        $for M in range(MR):
          float32x4_t va${M} = vld1q_f32(a${M}); a${M} = (const float*) ((uintptr_t) a${M} + k);

        $for L in range(4):

          $for N in range(0, NR, 4):
            const float32x4_t vb${ABC[N:N+4]}c${L} = vld1q_f32(w); w += 4;

          $for N in range(0, NR, 4):
            $for M in range(MR):
              const float32x4_t vmska${M}x${ABC[N:N+4]}c${L} = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va${M}), vceqq_f32(vb${ABC[N:N+4]}c${L}, vmovq_n_f32(0.0f))));
              vacc${M}x${ABC[N:N+4]} = ${VMULADDQ_F32}(vacc${M}x${ABC[N:N+4]}, vmska${M}x${ABC[N:N+4]}c${L}, vb${ABC[N:N+4]}c${L});

          $if L + 1 != 4:
            $for M in range(MR):
              va${M} = vextq_f32(va${M}, va${M}, 1);
      }

      p -= ${MR} * sizeof(void*);
    } while (p != 0);

    const float32x4_t vmax = vld1q_dup_f32(&params->scalar.max);
    $for N in range(0, NR, 4):
      $for M in range(MR):
        vacc${M}x${ABC[N:N+4]} = vminq_f32(vacc${M}x${ABC[N:N+4]}, vmax);

    const float32x4_t vmin = vld1q_dup_f32(&params->scalar.min);
    $for N in range(0, NR, 4):
      $for M in range(MR):
        vacc${M}x${ABC[N:N+4]} = vmaxq_f32(vacc${M}x${ABC[N:N+4]}, vmin);

    if XNN_LIKELY(nc >= ${NR}) {
      $for M in reversed(range(MR)):
        vst1q_f32(c${M}, vacc${M}x${ABC[0:4]});
        $for N in range(4, NR, 4):
          vst1q_f32(c${M} + ${N}, vacc${M}x${ABC[N:N+4]});
        c${M} = (float*) ((uintptr_t) c${M} + cn_stride);

      a = (const float**restrict) ((uintptr_t) a - ks);
      nc -= ${NR};
    } else {
      $for LOG2N in reversed(range(NR.bit_length())):
        $if LOG2N == 1:
          $for M in reversed(range(MR)):
            float32x2_t vacc${M}x${ABC[0:2]} = vget_low_f32(vacc${M}x${ABC[0:4]});
        $if 1 << LOG2N != NR:
          if (nc & ${1 << LOG2N}) {
            $if LOG2N >= 2:
              $for N in range(0, 1 << LOG2N, 4):
                $for M in reversed(range(MR)):
                  vst1q_f32(c${M}, vacc${M}x${ABC[N:N+4]}); c${M} += 4;

              $for M in reversed(range(MR)):
                $for N in range(0, 1 << (LOG2N - 1), 4):
                  vacc${M}x${ABC[N:N+4]} = vacc${M}x${ABC[N + (1 << LOG2N):N + (1 << LOG2N)+4]};
            $elif LOG2N == 1:
              $for M in reversed(range(MR)):
                vst1_f32(c${M}, vacc${M}x${ABC[0:2]}); c${M} += 2;

              $for M in reversed(range(MR)):
                vacc${M}x${ABC[0:2]} = vget_high_f32(vacc${M}x${ABC[0:4]});
            $elif LOG2N == 0:
              $for M in reversed(range(MR)):
                vst1_lane_f32(c${M}, vacc${M}x${ABC[0:2]}, 0);
          }

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