// 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 == 2
#include <assert.h>

#include <arm_neon.h>

#include <xnnpack/common.h>
#include <xnnpack/igemm.h>


void xnn_f32_igemm_minmax_ukernel_${MR}x${NR}__${"neonfma" if FMA else "neon"}_lane_ld64(
    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)])
{
  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 {
    float32x2_t vacc0x01 = vld1_f32(w); w += 2;
    $for M in range(1, MR):
      float32x2_t vacc${M}x01 = vacc0x01;

    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;
      for (; k >= 2 * sizeof(float); k -= 2 * sizeof(float)) {
        $for M in range(MR):
          const float32x2_t va${M} = vld1_f32(a${M}); a${M} += 2;

        $for L in range(2):
          const float32x2_t vb01c${L} = vld1_f32(w); w += 2;

          $if FMA:
            #if XNN_ARCH_ARM64
              $for M in range(MR):
                vacc${M}x01 = vfma_lane_f32(vacc${M}x01, vb01c${L}, va${M}, ${L});
            #else
              $for M in range(MR):
                const float32x2_t va${M}c${L} = vdup_lane_f32(va${M}, ${L});
              $for M in range(MR):
                vacc${M}x01 = vfma_f32(vacc${M}x01, va${M}c${L}, vb01c${L});
            #endif
          $else:
            $for M in range(MR):
              vacc${M}x01 = vmla_lane_f32(vacc${M}x01, vb01c${L}, va${M}, ${L});
      }
      if XNN_UNLIKELY(k != 0) {
        $for M in range(MR):
          const float32x2_t va${M} = vld1_dup_f32(a${M});

        const float32x2_t vb01 = vld1_f32(w); w += 2;

        $for M in range(MR):
          $if FMA:
            vacc${M}x01 = vfma_f32(vacc${M}x01, va${M}, vb01);
          $else:
            vacc${M}x01 = vmla_f32(vacc${M}x01, va${M}, vb01);
      }
      p -= ${MR} * sizeof(void*);
    } while (p != 0);

    const float32x2_t vmax = vld1_dup_f32(&params->scalar.max);
    $for N in range(0, NR, 4):
      $for M in range(MR):
        vacc${M}x01 = vmin_f32(vacc${M}x01, vmax);

    const float32x2_t vmin = vld1_dup_f32(&params->scalar.min);
    $for N in range(0, NR, 4):
      $for M in range(MR):
        vacc${M}x01 = vmax_f32(vacc${M}x01, vmin);

    if XNN_LIKELY(nc >= ${NR}) {
      $for M in reversed(range(MR)):
        vst1_f32(c${M}, vacc${M}x01);
        c${M} = (float*) ((uintptr_t) c${M} + cn_stride);

      a = (const float**restrict) ((uintptr_t) a - ks);
      nc -= ${NR};
    } else {
      assert(nc == 1);
      $for M in reversed(range(MR)):
        vst1_lane_f32(c${M}, vacc${M}x01, 0);

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