// 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 % 8 == 0

#include <assert.h>

#include <arm_neon.h>

#include "xnnpack/igemm.h"
#include "xnnpack/intrinsics-polyfill.h"


void xnn_f16_igemm_minmax_ukernel_${MR}x${NR}__neonfp16arith_ld64(
    size_t mr,
    size_t nc,
    size_t kc,
    size_t ks,
    const xnn_float16** restrict a,
    const xnn_float16* restrict w,
    xnn_float16* restrict c,
    size_t cm_stride,
    size_t cn_stride,
    size_t a_offset,
    const xnn_float16* zero,
    const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
{
  assert(mr != 0);
  assert(mr <= ${MR});
  assert(nc != 0);
  assert(kc != 0);
  assert(kc % sizeof(uint16_t) == 0);
  assert(ks != 0);
  assert(ks % (${MR} * sizeof(void*)) == 0);
  assert(a_offset % sizeof(uint16_t) == 0);
  assert(a != NULL);
  assert(w != NULL);
  assert(c != NULL);

  uint16_t* c0 = (uint16_t*) c;
  $for M in range(1, MR):
    uint16_t* c${M} = (uint16_t*) ((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 {
    $for N in range(0, NR, 8):
      float16x8_t vacc0x${N//8} = vreinterpretq_f16_u16(vld1q_u16((const uint16_t*) w)); w = (const xnn_float16*) w + 8;
    $for M in range(1, MR):
      $for N in range(0, NR, 8):
        float16x8_t vacc${M}x${N//8} = vacc0x${N//8};

    size_t p = ks;
    do {
      $for M in range(MR):
        const uint16_t* restrict a${M} = (const uint16_t*) a[${M}];
        assert(a${M} != NULL);
        if XNN_UNPREDICTABLE(a${M} != (const uint16_t*) zero) {
          a${M} = (const uint16_t*) ((uintptr_t) a${M} + a_offset);
        }
      a += ${MR};

      size_t k = kc;
      for (; k >= 4 * sizeof(uint16_t); k -= 4 * sizeof(uint16_t)) {
        $for M in range(MR):
          const float16x4_t va${M} = vreinterpret_f16_u16(vld1_u16(a${M})); a${M} += 4;

        $for L in range(4):
          $for N in range(0, NR, 8):
            const float16x8_t vb${N//8}c${L} = vreinterpretq_f16_u16(vld1q_u16((const uint16_t*) w)); w = (const xnn_float16*) w + 8;

          #if XNN_ARCH_ARM64
            $for N in range(0, NR, 8):
              $for M in range(MR):
                vacc${M}x${N//8} = vfmaq_lane_f16(vacc${M}x${N//8}, vb${N//8}c${L}, va${M}, ${L});
          #else
            $for N in range(0, NR, 8):
              $for M in range(MR):
                vacc${M}x${N//8} = vmlaq_lane_f16(vacc${M}x${N//8}, vb${N//8}c${L}, va${M}, ${L});
          #endif
      }
      if XNN_UNLIKELY(k != 0) {
        do {
          $for M in range(MR):
            const float16x8_t va${M} = vreinterpretq_f16_u16(vld1q_dup_u16(a${M})); a${M} += 1;

          $for N in range(0, NR, 8):
            const float16x8_t vb${N//8} = vreinterpretq_f16_u16(vld1q_u16((const uint16_t*) w)); w = (const xnn_float16*) w + 8;

          $for N in range(0, NR, 8):
            $for M in range(MR):
              vacc${M}x${N//8} = vfmaq_f16(vacc${M}x${N//8}, va${M}, vb${N//8});

          k -= sizeof(uint16_t);
        } while (k != 0);
      }
      p -= ${MR} * sizeof(void*);
    } while (p != 0);


    const float16x8_t vmin = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.min));
    $for N in range(0, NR, 8):
      $for M in range(MR):
        vacc${M}x${N//8} = vmaxq_f16(vacc${M}x${N//8}, vmin);

    const float16x8_t vmax = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.max));
    $for N in range(0, NR, 8):
      $for M in range(MR):
        vacc${M}x${N//8} = vminq_f16(vacc${M}x${N//8}, vmax);

    if XNN_LIKELY(nc >= ${NR}) {
      $for M in reversed(range(MR)):
        vst1q_u16(c${M}, vreinterpretq_u16_f16(vacc${M}x0));
        $for N in range(8, NR, 8):
          vst1q_u16(c${M} + ${N}, vreinterpretq_u16_f16(vacc${M}x${N//8}));
        c${M} = (uint16_t*) ((uintptr_t) c${M} + cn_stride);

      a = (const xnn_float16**restrict) ((uintptr_t) a - ks);
      nc -= ${NR};
    } else {
      $for LOG2N in reversed(range(NR.bit_length())):
        $if NR != 1 << LOG2N:
          if (nc & ${1 << LOG2N}) {
            $if LOG2N >= 3:
              $for N in range(0, 1 << LOG2N, 8):
                $for M in reversed(range(MR)):
                  vst1q_u16(c${M}, vreinterpretq_u16_f16(vacc${M}x${N//8})); c${M} += 8;

              $for M in reversed(range(MR)):
                $for N in range(0, NR - (1 << LOG2N), 8):
                  vacc${M}x${N//8} = vacc${M}x${(N + (1 << LOG2N))//8};
            $elif LOG2N == 2:
              $for M in reversed(range(MR)):
                vst1_u16(c${M}, vreinterpret_u16_f16(vacc${M})); c${M} += 4;

              $for M in reversed(range(MR)):
                vacc${M} = vget_high_f16(vacc${M}x0);
            $elif LOG2N == 1:
              $for M in reversed(range(MR)):
                vst1_lane_u32((void*) c${M}, vreinterpret_u32_f16(vacc${M}), 0); c${M} += 2;

              $for M in reversed(range(MR)):
                vacc${M} = vext_f16(vacc${M}, vacc${M}, 2);
            $elif LOG2N == 0:
              $for M in reversed(range(MR)):
                vst1_lane_u16(c${M}, vreinterpret_u16_f16(vacc${M}), 0);
          }
        $if LOG2N == 3:
          $for M in reversed(range(MR)):
            float16x4_t vacc${M} = vget_low_f16(vacc${M}x0);

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