// Copyright 2024 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 DATATYPE in ["F16"]
$assert NR % 32 == 0
#include <assert.h>

#include <immintrin.h>

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


$RANGE_MR = list(range(MR))
$DATATYPE_SPEC = {"F16": "f16", "QC8": "f16_qc8w"}[DATATYPE]
void xnn_${DATATYPE_SPEC}_igemm_minmax_ukernel_${MR}x${NR}__avx512fp16_broadcast(
    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(xnn_float16) == 0);
  assert(a != NULL);
  assert(w != NULL);
  assert(c != NULL);

#if defined(__AVX512FP16__)
  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 {
    __m512h vacc0x0 = _mm512_load_ph(w);
    $for N in range(32, NR, 32):
      __m512h vacc0x${N//32} = _mm512_load_ph((const uint16_t*) w + ${N});
    $for M in range(1, MR):
      $for N in range(0, NR, 32):
        __m512h vacc${M}x${N//32} = vacc0x${N//32};
    w = (const xnn_float16*) w + ${NR};

    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;
      do {
        const __m512h vb0 = _mm512_load_ph(w);
        $for N in range(32, NR, 32):
          const __m512h vb${N//32} = _mm512_load_ph((const uint16_t*) w + ${N});
        w = (const xnn_float16*) w + ${NR};

        $for M in range(MR):
          const __m512h va${M} = _mm512_castsi512_ph(_mm512_set1_epi16(*a${M}));
          $for N in range(0, NR, 32):
            vacc${M}x${N//32} = _mm512_fmadd_ph(va${M}, vb${N//32}, vacc${M}x${N//32});
          a${M} += 1;

        k -= sizeof(uint16_t);
      } while (k != 0);

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

    const __m512h vmin = _mm512_castsi512_ph(_mm512_set1_epi16(*(const uint16_t*) &params->scalar.min));
    $for N in range(0, NR, 32):
      $for M in range(MR):
        vacc${M}x${N//32} = _mm512_max_ph(vmin, vacc${M}x${N//32});

    const __m512h vmax = _mm512_castsi512_ph(_mm512_set1_epi16(*(const uint16_t*) &params->scalar.max));
    $for N in range(0, NR, 32):
      $for M in range(MR):
        vacc${M}x${N//32} = _mm512_min_ph(vmax, vacc${M}x${N//32});

    if XNN_LIKELY(nc >= ${NR}) {
      $for M in reversed(range(MR)):
        _mm512_storeu_ph(c${M}, vacc${M}x0);
        $for N in range(32, NR, 32):
          _mm512_storeu_ph((uint16_t*) c${M} + ${N//32}, vacc${M}x${N//32});
        c${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(5, NR.bit_length())):
        $if NR != 1 << LOG2N:
          if (nc & ${1 << LOG2N}) {
            $if LOG2N >= 5:
              $for M in reversed(range(MR)):
                _mm512_storeu_ph(c${M}, vacc${M}x0);
                $for N in range(32, 1 << LOG2N, 32):
                  _mm512_storeu_ph((uint16_t*) c${M} + ${N//32}, vacc${M}x${N//32});

              $for M in reversed(range(MR)):
                $for N in range(0, NR - (1 << LOG2N), 32):
                  vacc${M}x${N//32} = vacc${M}x${(N+(1<<LOG2N))//32};
                c${M} += ${1 << LOG2N};
          }
        $if LOG2N == 5:
          $if NR > 32:
            if (nc & 31) {
              // Prepare mask for valid 16-bit elements (depends on nc).
              const __mmask32 vmask = _cvtu32_mask32((uint32_t) (UINT32_C(1) << (nc & 31)) - UINT32_C(1));
              $for M in reversed(range(MR)):
                _mm512_mask_storeu_epi16(c${M}, vmask, _mm512_castph_si512(vacc${M}x0));
            }
          $else:
            assert(nc != 0);
            assert(nc < ${NR});
            // Prepare mask for valid 16-bit elements (depends on nc).
            const __mmask32 vmask = _cvtu32_mask32((uint32_t) (UINT32_C(1) << nc) - UINT32_C(1));
            $for M in reversed(range(MR)):
              _mm512_mask_storeu_epi16(c${M}, vmask, _mm512_castph_si512(vacc${M}x0));
      nc = 0;
    }
  } while (nc != 0);
#endif  // defined(__AVX512FP16__)
}