// Copyright 2023 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 == "QC4"
$assert NR % 16 == 0
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
#include <assert.h>

#include <immintrin.h>

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


$ISA = "avx512skx"
$DATATYPE_SPEC = {"QC4": "f32_qc4w"}[DATATYPE]
void xnn_${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};
      }
  const __m512i vmagic_bias_c0 = _mm512_set1_epi32(0x4B0000F0);
  const __m512i vmagic_bias_c1 = _mm512_set1_epi32(0x4900000F);
  const __m512 vmagic_bias_plus_kernel_zero_point_c0 = _mm512_set1_ps(0x1.0001E0p+23f + (float) params->scalar.kernel_zero_point);
  const __m512 vmagic_bias_plus_kernel_zero_point_c1 = _mm512_set1_ps(0x1.00001Ep+19f + (float) params->scalar.kernel_zero_point);
  XNN_FORCE_REALIZATION(vmagic_bias_c0);
  XNN_FORCE_REALIZATION(vmagic_bias_c1);
  XNN_FORCE_REALIZATION(vmagic_bias_plus_kernel_zero_point_c0);
  XNN_FORCE_REALIZATION(vmagic_bias_plus_kernel_zero_point_c1);

  do {
    $for N in range(0, NR, 16):
      __m512 vacc0x${ABC[N: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${ABC[N:N+16]} = vacc0x${ABC[N:N+16]};
    w = (const float*) w + ${NR};

    size_t k = kc;
    for (; k >= 2 * sizeof(float); k -= sizeof(float) * 2) {
      $for L in range(2):
        $for M in range(MR):
          const __m512 va${M}c${L} = _mm512_set1_ps(*a${M});
          a${M} += 1;

      const __m512i vbi${ABC[0:16]}c01 = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) w));
      $for N in range(16, NR, 16):
        const __m512i vbi${ABC[N:N+16]}c01 = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) ((const int8_t*) w + ${N})));
      w = (const int8_t*) w + ${NR};

      $for L in range(2):
        $for N in range(0, NR, 16):
          const __m512 vbm${ABC[N:N+16]}c${L} =  _mm512_castsi512_ps(_mm512_or_si512(vbi${ABC[N:N+16]}c01, vmagic_bias_c${L}));

      $for L in range(2):
        $for N in range(0, NR, 16):
          const __m512 vb${ABC[N:N+16]}c${L} = _mm512_sub_ps(vbm${ABC[N:N+16]}c${L}, vmagic_bias_plus_kernel_zero_point_c${L});

      $for L in range(2):
        $for N in range(0, NR, 16):
          $for M in range(MR):
            vacc${M}x${ABC[N:N+16]} = _mm512_fmadd_ps(va${M}c${L}, vb${ABC[N:N+16]}c${L}, vacc${M}x${ABC[N:N+16]});
    }

    if XNN_UNLIKELY(k != 0) {
      $for M in range(MR):
        const __m512 va${M} = _mm512_set1_ps(*a${M});
        a${M} += 1;

      const __m512i vbi${ABC[0:16]} = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) w));
      $for N in range(16, NR, 16):
        const __m512i vbi${ABC[N:N+16]} = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) ((const int8_t*) w + ${N})));
      w = (const int8_t*) w + ${NR};

      $for N in range(0, NR, 16):
        const __m512 vbm${ABC[N:N+16]} =  _mm512_castsi512_ps(_mm512_or_si512(vbi${ABC[N:N+16]}, vmagic_bias_c0));

      $for N in range(0, NR, 16):
        const __m512 vb${ABC[N:N+16]} = _mm512_sub_ps(vbm${ABC[N:N+16]}, vmagic_bias_plus_kernel_zero_point_c0);

      $for N in range(0, NR, 16):
        $for M in range(MR):
          vacc${M}x${ABC[N:N+16]} = _mm512_fmadd_ps(va${M}, vb${ABC[N:N+16]}, vacc${M}x${ABC[N:N+16]});
    }

    $if DATATYPE in ["QC8", "QC4"]:
      $for N in range(0, NR, 16):
        const __m512 vscale${ABC[N:N+16]} = _mm512_loadu_ps((const float*) w + ${N});
        $for M in range(MR):
          vacc${M}x${ABC[N:N+16]} = _mm512_mul_ps(vacc${M}x${ABC[N:N+16]}, vscale${ABC[N: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${ABC[N:N+16]} = _mm512_max_ps(vmin, vacc${M}x${ABC[N: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${ABC[N:N+16]} = _mm512_min_ps(vmax, vacc${M}x${ABC[N:N+16]});

    if XNN_LIKELY(nc >= ${NR}) {
      $for M in reversed(range(MR)):
        _mm512_storeu_ps(c${M}, vacc${M}x${ABC[0:16]});
        $for N in range(16, NR, 16):
          _mm512_storeu_ps(c${M} + ${N}, vacc${M}x${ABC[N:N+16]});
        c${M} = (float*) ((uintptr_t) c${M} + cn_stride);

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

      nc -= ${NR};
    } else {
      $for LOG2N in reversed(range(4, NR.bit_length())):
        $if NR != 1 << LOG2N:
          if (nc & ${1 << LOG2N}) {
            $if LOG2N >= 4:
              $for M in reversed(range(MR)):
                _mm512_storeu_ps(c${M}, vacc${M}x${ABC[0:16]});
                $for N in range(16, 1 << LOG2N, 16):
                  _mm512_storeu_ps(c${M} + ${N}, vacc${M}x${ABC[N:N+16]});

              $for M in reversed(range(MR)):
                $for N in range(0, NR - (1 << LOG2N), 16):
                  vacc${M}x${ABC[N:N+16]} = vacc${M}x${ABC[N + (1 << LOG2N):N + (1 << LOG2N)+16]};

              $for M in reversed(range(MR)):
                c${M} += ${1 << LOG2N};
          }
        $if LOG2N == 4:
          $if NR > 16:
            if (nc & 15) {
              // Prepare mask for valid 32-bit elements (depends on nc).
              const __mmask16 vmask = _cvtu32_mask16((uint32_t) (UINT32_C(1) << (nc & 15)) - UINT32_C(1));
              $for M in reversed(range(MR)):
                _mm512_mask_storeu_ps(c${M}, vmask, vacc${M}x${ABC[0:16]});
            }
          $else:
            assert(nc != 0);
            assert(nc < ${NR});
            // Prepare mask for valid 32-bit elements (depends on nc).
            const __mmask16 vmask = _cvtu32_mask16((uint32_t) (UINT32_C(1) << nc) - UINT32_C(1));
            $for M in reversed(range(MR)):
              _mm512_mask_storeu_ps(c${M}, vmask, vacc${M}x${ABC[0:16]});
      nc = 0;
    }
  } while (nc != 0);
}