// Copyright 2022 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
$assert ACCTYPE in ["F16", "F32"]
$ACC_SUFFIX = "_f32acc" if ACCTYPE == "F32" else ""
#include <assert.h>

#include <immintrin.h>

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


void xnn_f16${ACC_SUFFIX}_gemm_minmax_ukernel_${MR}x${NR}__avx2_broadcast(
    size_t mr,
    size_t nc,
    size_t kc,
    const xnn_float16* restrict a,
    size_t a_stride,
    const xnn_float16* restrict w,
    xnn_float16* restrict c,
    size_t cm_stride,
    size_t cn_stride,
    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(a != NULL);
  assert(w != NULL);
  assert(c != NULL);

  const uint16_t* a0 = (const uint16_t*) a;
  uint16_t* c0 = (uint16_t*) c;

  const __m256 vmin = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.min));
  const __m256 vmax = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.max));
  XNN_FORCE_REALIZATION(vmin);
  XNN_FORCE_REALIZATION(vmax);

  $for M in range(1, MR):
    const uint16_t* a${M} = (const uint16_t*) ((uintptr_t) a${M-1} + a_stride);
    uint16_t* c${M} = (uint16_t*) ((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 {
    __m256 vacc0x0 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) w));
    $for N in range(8, NR, 8):
      __m256 vacc0x${N//8} = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) ((const uint16_t*) w + ${N})));
    $for M in range(1, MR):
      $for N in range(0, NR, 8):
        __m256 vacc${M}x${N//8} = vacc0x${N//8};
    w = (const xnn_float16*) w + ${NR};

    size_t k = kc;
    do {
      $for M in range(MR):
        const __m256 va${M} = _mm256_cvtph_ps(_mm_set1_epi16((short) *a${M}));
        a${M} += 1;

      const __m256 vb0 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) w));
      $for N in range(8, NR, 8):
        const __m256 vb${N//8} = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) ((const uint16_t*) w + ${N})));
      w = (const xnn_float16*) w + ${NR};

      $for N in range(0, NR, 8):
        $for M in range(MR):
          $if ACCTYPE == "F32":
            vacc${M}x${N//8} = _mm256_fmadd_ps(va${M}, vb${N//8}, vacc${M}x${N//8});
          $else:
            vacc${M}x${N//8} = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(va${M}, vb${N//8}, vacc${M}x${N//8}), _MM_FROUND_TO_NEAREST_INT));

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

    $for N in range(0, NR, 8):
      $for M in range(MR):
        vacc${M}x${N//8} = _mm256_max_ps(vacc${M}x${N//8}, vmin);

    $for N in range(0, NR, 8):
      $for M in range(MR):
        vacc${M}x${N//8} = _mm256_min_ps(vacc${M}x${N//8}, vmax);

    if XNN_LIKELY(nc >= ${NR}) {
      $for M in range(MR):
        _mm_storeu_si128((__m128i*) c${M}, _mm256_cvtps_ph(vacc${M}x0, _MM_FROUND_TO_NEAREST_INT));
        $for N in range(8, NR, 8):
          _mm_storeu_si128((__m128i*) (c${M} + ${N}), _mm256_cvtps_ph(vacc${M}x${N//8}, _MM_FROUND_TO_NEAREST_INT));
        c${M} = (uint16_t*) ((uintptr_t) c${M} + cn_stride);

      $for M in range(MR):
        a${M} = (const uint16_t*) ((uintptr_t) a${M} - kc);

      nc -= ${NR};
    } else {
      $for LOG2N in reversed(range(NR.bit_length())):
        $if LOG2N == 3:
          $for M in range(MR):
            __m128i vh${M}x0 = _mm256_cvtps_ph(vacc${M}x0, _MM_FROUND_TO_NEAREST_INT);
        $if NR != 1 << LOG2N:
          if (nc & ${1 << LOG2N}) {
            $if LOG2N >= 4:
              $for M in range(MR):
                _mm_storeu_si128((__m128i*) c${M}, _mm256_cvtps_ph(vacc${M}x0, _MM_FROUND_TO_NEAREST_INT));
                $for N in range(8, 1 << LOG2N, 8):
                  _mm_storeu_si128((__m128i*) (c${M} + ${N}), _mm256_cvtps_ph(vacc${M}x${N//8}, _MM_FROUND_TO_NEAREST_INT));

              $for M in range(MR):
                $for N in range(0, NR - (1 << LOG2N), 8):
                  vacc${M}x${N//8} = vacc${M}x${(N + (1 << LOG2N))//8};

              $for M in range(MR):
                c${M} += ${1 << LOG2N};
            $elif LOG2N == 3:
              $for M in range(MR):
                _mm_storeu_si128((__m128i*) c${M}, vh${M}x0);

              $for M in range(MR):
                vh${M}x0 = _mm256_cvtps_ph(vacc${M}x1, _MM_FROUND_TO_NEAREST_INT);

              $for M in range(MR):
                c${M} += ${1 << LOG2N};
            $elif LOG2N == 2:
              $for M in range(MR):
                _mm_storel_epi64((__m128i*) c${M}, vh${M}x0);

              $for M in range(MR):
                vh${M}x0 = _mm_unpackhi_epi64(vh${M}x0, vh${M}x0);

              $for M in range(MR):
                c${M} += 4;
            $elif LOG2N == 1:
              $for M in range(MR):
                _mm_storeu_si32(c${M}, vh${M}x0);

              $for M in range(MR):
                vh${M}x0 = _mm_srli_epi64(vh${M}x0, 32);

              $for M in range(MR):
                c${M} += 2;
            $elif LOG2N == 0:
              $for M in range(MR):
                *c${M} = (uint16_t) _mm_extract_epi16(vh${M}x0, 0);
          }

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