// 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 DATATYPE in ["F32", "QC4", "QC8"]
$assert NR % 16 == 0
$assert 16 <= NR <= 64
#include <assert.h>

#include <immintrin.h>
$if DATATYPE in ["QC8", "QC4"]:
  #include <smmintrin.h>

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


$RANGE_MR = list(reversed(range(MR))) if INC else list(range(MR))
$ISA = {"F32": "avx512f", "QC8": "avx512skx", "QC4": "avx512skx"}[DATATYPE]
$DATATYPE_SPEC = {"F32": "f32", "QC8": "f32_qc8w", "QC4": "f32_qc4w"}[DATATYPE]
void xnn_${DATATYPE_SPEC}_gemm${"inc" if INC else ""}_minmax_ukernel_${MR}x${NR}__${ISA}_broadcast(
    size_t mr,
    size_t nc,
    size_t kc,
    const float* restrict a,
    size_t a_stride,
    $if DATATYPE == "F32":
      const float* restrict w,
    $else:
      const void* restrict w,
    float* restrict c,
    size_t cm_stride,
    size_t cn_stride,
    $if INC:
      const float* restrict acc,
    $if DATATYPE == "QC4":
      const struct xnn_f32_qc4w_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
    $else:
      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(a != NULL);
  assert(w != NULL);
  assert(c != NULL);
  $if INC:
    assert(acc != NULL);

  const float* a0 = a;
  float* c0 = c;
  $for M in range(1, MR):
    const float* a${M} = (const float*) ((uintptr_t) a${M-1} + a_stride);
    float* c${M} = (float*) ((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):
        $for N in range(0, NR, 16):
          __m512 vacc${M}x${N//16} = _mm512_load_ps(acc + ${M*NR+N});
      acc += ${MR*NR};
    $else:
      $if DATATYPE == "F32":
        __m512 vacc0x0 = _mm512_load_ps(w);
      $else:
        __m512 vacc0x0 = _mm512_loadu_ps(w);
      $for N in range(16, NR, 16):
        $if DATATYPE == "F32":
          __m512 vacc0x${N//16} = _mm512_load_ps(w + ${N});
        $else:
          __m512 vacc0x${N//16} = _mm512_loadu_ps((const float*) w + ${N});
      $for M in range(1, MR):
        $for N in range(0, NR, 16):
          __m512 vacc${M}x${N//16} = vacc0x${N//16};
      $if DATATYPE == "F32":
        w += ${NR};
      $else:
        w = (const float*) w + ${NR};

    size_t k = kc;
    do {
      $if DATATYPE == "F32":
        const __m512 vb0 = _mm512_load_ps(w);
        $for N in range(16, NR, 16):
          const __m512 vb${N//16} = _mm512_loadu_ps(w + ${N});
        w += ${NR};
      $else:
        const __m512i vbi0 = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) w));
        $for N in range(16, NR, 16):
          const __m512i vbi${N//16}  = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((const int8_t*) w + ${N})));
        $for N in range(0, NR, 16):
          const __m512 vb${N//16}  = _mm512_cvtepi32_ps(vbi${N//16});
        w = (const int8_t*) w + ${NR};

      $for M in range(MR):
        const __m512 va${M} = _mm512_set1_ps(*a${M});
        $for N in range(0, NR, 16):
          vacc${M}x${N//16} = _mm512_fmadd_ps(va${M}, vb${N//16}, vacc${M}x${N//16});

      $for M in range(MR):
        a${M} += 1;

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

    $if DATATYPE in ["QC8", "QC4"]:
      $for N in range(0, NR, 16):
        const __m512 vscale${N//16} = _mm512_loadu_ps((const float*) w + ${N});
        $for M in range(MR):
          vacc${M}x${N//16} = _mm512_mul_ps(vacc${M}x${N//16}, vscale${N//16});
      w = (const float*) w + ${NR};
    const __m512 vmin = _mm512_set1_ps(params->scalar.min);
    $for N in range(0, NR, 16):
      $for M in range(MR):
        vacc${M}x${N//16} = _mm512_max_ps(vmin, vacc${M}x${N//16});

    const __m512 vmax = _mm512_set1_ps(params->scalar.max);
    $for N in range(0, NR, 16):
      $for M in range(MR):
        vacc${M}x${N//16} = _mm512_min_ps(vmax, vacc${M}x${N//16});

    if XNN_LIKELY(nc >= ${NR}) {
      $for M in RANGE_MR:
        _mm512_storeu_ps(c${M}, vacc${M}x0);
        $for N in range(16, NR, 16):
          _mm512_storeu_ps(c${M} + ${N}, vacc${M}x${N//16});
        c${M} = (float*) ((uintptr_t) c${M} + cn_stride);

      $for M in RANGE_MR:
        a${M} = (const float*) ((uintptr_t) a${M} - kc);

      nc -= ${NR};
    } else {
      // NC remainder (1..${NR-1})
      assert(nc >= 1);
      assert(nc <= ${NR-1});
      // Prepare mask for valid 32-bit elements (depends on nc).
      $for N in range(0, NR, 16):
        const __mmask16 vmask${N//16} = _cvtu32_mask16((uint32_t) (((UINT64_C(1) << nc) - 1) >> ${N}));

      $for M in RANGE_MR:
        $for N in range(0, NR, 16):
          _mm512_mask_storeu_ps(c${M} + ${N}, vmask${N//16}, vacc${M}x${N//16});
      nc = 0;
    }
  } while (nc != 0);
}