// Copyright 2020 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
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"

#include <assert.h>

#include <arm_neon.h>

#include <xnnpack/common.h>

#include <xnnpack/gemm.h>


void xnn_f16_gemm${"inc" if INC else ""}_minmax_ukernel_${MR}x${NR}__neonfp16arith_ld64(
    size_t mr,
    size_t nc,
    size_t kc,
    const void* restrict a,
    size_t a_stride,
    const void* restrict w,
    void* restrict c,
    size_t cm_stride,
    size_t cn_stride,
    $if INC:
      const void*restrict acc,
    const struct xnn_f16_scaleminmax_params params[restrict XNN_MIN_ELEMENTS(1)])
{
  assert(mr != 0);
  assert(mr <= ${MR});
  assert(nc != 0);
  assert(kc != 0);
  assert(kc % sizeof(__fp16) == 0);
  assert(a != NULL);
  assert(w != NULL);
  assert(c != NULL);
  $if INC:
    assert(acc != NULL);

  const __fp16* a0 = (const __fp16*) a;
  __fp16* c0 = (__fp16*) c;
  $for M in range(1, MR):
    const __fp16* a${M} = (const __fp16*) ((uintptr_t) a${M-1} + a_stride);
    __fp16* c${M} = (__fp16*) ((uintptr_t) c${M-1} + cm_stride);
    $if M % 2 == 0:
      if XNN_UNPREDICTABLE(mr <= ${M}) {
        a${M} = a${M-1};
        c${M} = c${M-1};
      }
    $elif M + 1 == MR:
      if XNN_UNPREDICTABLE(mr != ${M+1}) {
        a${M} = a${M-1};
        c${M} = c${M-1};
      }
    $else:
      if XNN_UNPREDICTABLE(mr < ${M+1}) {
        a${M} = a${M-1};
        c${M} = c${M-1};
      }

  do {
    $if INC:
      $for M in range(MR):
        $for N in range(0, NR, 8):
          float16x8_t vacc${M}x${ABC[N:N+8]} = vld1q_f16(acc); acc = (const void*) ((uintptr_t) acc + sizeof(float16x8_t));
    $else:
      $for N in range(0, NR, 8):
        float16x8_t vacc0x${ABC[N:N+8]} = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
      $for M in range(1, MR):
        $for N in range(0, NR, 8):
          float16x8_t vacc${M}x${ABC[N:N+8]} = vacc0x${ABC[N:N+8]};

    size_t k = kc;
    while (k >= 4 * sizeof(__fp16)) {
      $for M in range(MR):
        const float16x4_t va${M} = vld1_f16(a${M}); a${M} += 4;

      $for L in range(4):
        $for N in range(0, NR, 8):
          const float16x8_t vb${ABC[N:N+8]}c${L} = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));

        #if XNN_ARCH_ARM64
          $for N in range(0, NR, 8):
            $for M in range(MR):
              vacc${M}x${ABC[N:N+8]} = vfmaq_lane_f16(vacc${M}x${ABC[N:N+8]}, vb${ABC[N:N+8]}c${L}, va${M}, ${L});
        #else
          $for M in range(MR):
            const float16x8_t va${M}c${L} = vdupq_lane_f16(va${M}, ${L});

          $for N in range(0, NR, 8):
            $for M in range(MR):
              vacc${M}x${ABC[N:N+8]} = vfmaq_f16(vacc${M}x${ABC[N:N+8]}, va${M}c${L}, vb${ABC[N:N+8]}c${L});
        #endif

      k -= 4 * sizeof(__fp16);
    }
    if XNN_UNLIKELY(k != 0) {
      do {
        $for M in range(MR):
          const float16x8_t va${M} = vld1q_dup_f16(a${M}); a${M} += 1;

        $for N in range(0, NR, 8):
          const float16x8_t vb${ABC[N:N+8]} = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));

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

        k -= sizeof(__fp16);
      } while (k != 0);
    }

    const float16x8_t vscale = vld1q_dup_f16((const __fp16*) &params->scale);
    $for N in range(0, NR, 8):
      $for M in range(MR):
        vacc${M}x${ABC[N:N+8]} = vmulq_f16(vacc${M}x${ABC[N:N+8]}, vscale);

    const float16x8_t vmax = vld1q_dup_f16((const __fp16*) &params->max);
    $for N in range(0, NR, 8):
      $for M in range(MR):
        vacc${M}x${ABC[N:N+8]} = vminq_f16(vacc${M}x${ABC[N:N+8]}, vmax);

    const float16x8_t vmin = vld1q_dup_f16((const __fp16*) &params->min);
    $for N in range(0, NR, 8):
      $for M in range(MR):
        vacc${M}x${ABC[N:N+8]} = vmaxq_f16(vacc${M}x${ABC[N:N+8]}, vmin);

    if XNN_LIKELY(nc >= ${NR}) {
      $for M in range(MR):
        vst1q_f16(c${M}, vacc${M}x${ABC[0:8]});
        $for N in range(8, NR, 8):
          vst1q_f16(c${M} + ${N}, vacc${M}x${ABC[N:N+8]});
        c${M} = (__fp16*) ((uintptr_t) c${M} + cn_stride);

      $for M in range(MR):
        a${M} = (const __fp16*) ((uintptr_t) a${M} - kc);

      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):
                $for M in range(MR):
                  vst1q_f16(c${M}, vacc${M}x${ABC[N:N+8]}); c${M} += 8;

              $for M in range(MR):
                $for N in range(0, 1 << (LOG2N - 1), 8):
                  vacc${M}x${ABC[N:N+8]} = vacc${M}x${ABC[N + (1 << LOG2N):N + (1 << LOG2N)+8]};
            $elif LOG2N == 2:
              $for M in range(MR):
                vst1_f16(c${M}, vacc${M}x${ABC[0:4]}); c${M} += 4;

              $for M in range(MR):
                vacc${M}x${ABC[0:4]} = vget_high_f16(vacc${M}x${ABC[0:8]});
            $elif LOG2N == 1:
              $for M in range(MR):
                vst1_lane_u32(__builtin_assume_aligned(c${M}, 1), vreinterpret_u32_f16(vacc${M}x${ABC[0:4]}), 0); c${M} += 2;

              $for M in range(MR):
                vacc${M}x${ABC[0:4]} = vext_f16(vacc${M}x${ABC[0:4]}, vacc${M}x${ABC[0:4]}, 2);
            $elif LOG2N == 0:
              $for M in range(MR):
                vst1_lane_f16(c${M}, vacc${M}x${ABC[0:4]}, 0);
          }
        $if LOG2N == 3:
          $for M in range(MR):
            float16x4_t vacc${M}x${ABC[0:4]} = vget_low_f16(vacc${M}x${ABC[0:8]});

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