// 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.

$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
$assert REQUANTIZATION == "FP32" or not REQUANTIZATION
$assert DATATYPE in ["QD8", "QS8", "QC4"]
$UNROLL = locals().get("UNROLL", 1)
$assert ACCUMULATORS == UNROLL or ACCUMULATORS == 1
$AVX = 2
$GFNI = 1
$PREFETCH = 1
#include <assert.h>
#include <stdint.h>

#include <immintrin.h>

#include "xnnpack/common.h"
#include "xnnpack/gemm.h"
#include "xnnpack/intrinsics-polyfill.h"
#include "xnnpack/math.h"
#include "xnnpack/unaligned.h"
$if PREFETCH:
  #include "xnnpack/prefetch.h"


$DATATYPE_SPEC = {"QC8": "qs8_qc8w", "QD8": "qd8_f32_qc8w", "QS8": "qs8", "QU8": "qu8", "QC4": "qd8_f32_qc4w"}[DATATYPE]
$REQUANTIZATION_SPEC = "" if DATATYPE in ["QD8", "QC4"] else "_" + REQUANTIZATION.lower()
$PARAMS_STRUCT = REQUANTIZATION.lower() + "_scalar" if REQUANTIZATION else "scalar"
$PARAMS_TYPE = {"QC8": "union xnn_qs8_qc8w_conv_minmax_params", "QD8": "union xnn_f32_minmax_params", "QS8": "union xnn_qs8_conv_minmax_params", "QU8": "union xnn_qu8_conv_minmax_params", "QC4": "struct xnn_f32_qc4w_minmax_params"}[DATATYPE]
$XINT8_T = "uint8_t" if DATATYPE == "QU8" else "int8_t"
$OUT_T = "float" if DATATYPE in ["QD8", "QC4"] else XINT8_T
$def VACC(M,K=0):
$  return f"vacc{M}x{K}" if UNROLL > 1 else f"vacc{M}"
$ACC_POSTFIX=f"_acc{ACCUMULATORS}" if ACCUMULATORS > 1 else ""
$_MM256_DPBUSD_EPI32 = "_mm256_dpbusd_avx_epi32" if AVX == 2 else "_mm256_dpbusd_epi32"
$ISA = "avxvnni" if AVX == 2 else "avx512vnnigfni" if GFNI else "avx512vnni"
void xnn_${DATATYPE_SPEC}_gemm_minmax${REQUANTIZATION_SPEC}_ukernel_${MR}x8c4__${ISA}_u${UNROLL}${ACC_POSTFIX}(
    size_t mr,
    size_t nc,
    size_t kc,
    const ${XINT8_T}* restrict a,
    size_t a_stride,
    const void* restrict w,
    ${OUT_T}* restrict c,
    size_t cm_stride,
    size_t cn_stride,
    $if DATATYPE in ["QD8", "QC4"]:
      const ${PARAMS_TYPE} params[restrict XNN_MIN_ELEMENTS(1)],
      const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
    $else:
      const ${PARAMS_TYPE} params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(mr != 0);
  assert(mr <= ${MR});
  assert(nc != 0);
  assert(kc != 0);
  assert(kc % sizeof(${XINT8_T}) == 0);
  assert(a != NULL);
  assert(w != NULL);
  assert(c != NULL);

  kc = round_up_po2(kc, 4 * sizeof(${XINT8_T}));
  const ${XINT8_T}* a0 = a;
  ${OUT_T}* c0 = c;
  $for M in range(1, MR):
    const ${XINT8_T}* a${M} = (const ${XINT8_T}*) ((uintptr_t) a${M-1} + a_stride);
    ${OUT_T}* c${M} = (${OUT_T}*) ((uintptr_t) c${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};
      }

  $if DATATYPE in ["QD8", "QC4"]:
    $for M in range(MR):
      const __m256i vinput_zero_point${M} = _mm256_set1_epi32((int) quantization_params[${M}].zero_point);
    const __m256 voutput_min = _mm256_set1_ps(params->scalar.min);
    const __m256 voutput_max = _mm256_set1_ps(params->scalar.max);
    // XNN_FORCE_REALIZATION(voutput_min);
    // XNN_FORCE_REALIZATION(voutput_max);
    $if DATATYPE == "QC4":
      const __m256i vvalue_mask = _mm256_set1_epi8(0xF0);
      XNN_FORCE_REALIZATION(vvalue_mask);
      $if GFNI:
        const __m256i vshl4 = _mm256_set1_epi64x(0x01020408);
        XNN_FORCE_REALIZATION(vshl4);
  $else:
    const __m256i vsign_mask =_mm256_set1_epi8(0x80);
    XNN_FORCE_REALIZATION(vsign_mask);
    $if DATATYPE != "QC8":
      const __m256 vscale = _mm256_set1_ps(params->${PARAMS_STRUCT}.scale);
      // XNN_FORCE_REALIZATION(vscale);
    const __m256 voutput_max_less_zero_point = _mm256_set1_ps((int32_t) params->${PARAMS_STRUCT}.output_max - (int32_t) params->${PARAMS_STRUCT}.output_zero_point);
    const __m256i voutput_zero_point = _mm256_set1_epi32(params->${PARAMS_STRUCT}.output_zero_point);
    const __m128i voutput_min = _mm_set1_epi8(params->${PARAMS_STRUCT}.output_min);
    const __m128i vshuffle_control_mask = _mm_load_si128((const __m128i*) params->${PARAMS_STRUCT}.shuffle_control_mask);
    // XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
    // XNN_FORCE_REALIZATION(voutput_zero_point);
    // XNN_FORCE_REALIZATION(voutput_min);
  do {
    $if DATATYPE in ["QD8", "QC4"]:
      const __m256i vksum01234567 = _mm256_load_si256(w);
      $for M in range(MR):
        __m256i ${VACC(M)}x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point${M});
        $if ACCUMULATORS > 1:
          $for K in range(1, UNROLL):
            __m256i ${VACC(M,K)}x01234567 = _mm256_setzero_si256();
    $elif DATATYPE == "QS8":
      __m256i ${VACC(0)}x01234567 = _mm256_load_si256((const __m256i*) w);
      $for M in range(1, MR):
        __m256i ${VACC(M)}x01234567 = ${VACC(0)}x01234567;
    w = (const int32_t*) w + 8;

    size_t k = kc;
    $if UNROLL > 1:
      while (k >= ${UNROLL * 4} * sizeof(${XINT8_T})) {
        $for M in range(MR):
          $for K in range(UNROLL):
            $if DATATYPE in ["QD8", "QC4"]:
              __m256i va${M}x${K}x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a${M} + ${4 * K}));
            $else:
              va${M}x${K}x0123 = _mm256_xor_si256(va${M}x${K}x0123, vsign_mask);
          a${M} += ${4 * UNROLL};

        $if DATATYPE in ["QS8", "QD8"]:
          $for K in range(UNROLL):
            const __m256i vb${K}x01234567 = _mm256_load_si256((const __m256i*) ((const ${XINT8_T}*) w + ${32 * K}));
          w = (const ${XINT8_T}*) w + ${32 * UNROLL};
        $elif DATATYPE in ["QC4"]:
          $for K in range(UNROLL):
            __m128i vb4b${K}x01234567 = _mm128_loadu_si128(w);
            const __m128i vb4bodd${K} = _mm128_and_si128(vb4b${K}x01234567, vvalue_mask);
            const __m128i vb4beven${K} = _mm128_and_si128(_mm128_slli_epi16(vb4b${K}x01234567, 4), vvalue_mask);
            const __m128i vb${K}x012389AB = _mm128_unpacklo_epi8(vb4beven${K}, vb4bodd${K});
            const __m128i vb${K}x4567CDEF = _mm128_unpackhi_epi8(vb4beven${K}, vb4bodd${K});
            __m256i vb${K}x01234567 = _mm256_permutex2var_epi64(_mm256_castsi128_si256(vb${K}x012389AB), vpermute_mask, _mm256_castsi128_si256(vb${K}x4567CDEF));
            w = (const ${XINT8_T}*) w + 16;

        $if ACCUMULATORS > 1:
          $for M in range(MR):
            $for K in range(UNROLL):
              ${VACC(M,K)}x01234567 = ${_MM256_DPBUSD_EPI32}(${VACC(M,K)}x01234567, va${M}x${K}x0123, vb${K}x01234567);
        $else:
          $for K in range(UNROLL):
            $for M in range(MR):
              ${VACC(M)}x01234567 = ${_MM256_DPBUSD_EPI32}(${VACC(M)}x01234567, va${M}x${K}x0123, vb${K}x01234567);

        k -= ${UNROLL * 4} * sizeof(${XINT8_T});
      }
      $if ACCUMULATORS > 1:
        $PAIRS = [(i,) for i in range(UNROLL)]
        $while len(PAIRS) > 1:
          $TPLS=[PAIRS[i:i+2] for i in range(0, len(PAIRS), 2)]
          $PAIRS = [(P1[0],P2[0]) for P1, P2 in TPLS]
          $for K1, K2 in PAIRS:
            $for M in range(MR):
              ${VACC(M,K1)}x01234567 = _mm256_add_epi32(${VACC(M,K1)}x01234567, ${VACC(M,K2)}x01234567);
    while (k != 0) {
      $for M in range(MR):
        $if DATATYPE in ["QD8", "QC4"]:
          __m256i va${M}x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a${M}));
        $else:
          va${M}x0123 = _mm256_xor_si256(va${M}x0123, vsign_mask);
        a${M} += 4;

      $if DATATYPE in ["QS8", "QD8"]:
        const __m256i vb01234567 = _mm256_load_si256(w);
      $elif DATATYPE in ["QC4"]:
        __m128i vb4b01234567 = _mm128_loadu_si128(w);
        const __m128i vb4bodd = _mm128_and_si128(vb4b0123456789ABCDEF, vvalue_mask);
        const __m128i vb4beven = _mm128_and_si128(_mm128_slli_epi16(vb4b0123456789ABCDEF, 4), vvalue_mask);
        const __m128i vb012389AB = _mm128_unpacklo_epi8(vb4beven, vb4bodd);
        const __m128i vb4567CDEF = _mm128_unpackhi_epi8(vb4beven, vb4bodd);
        __m256i vb01234567 = _mm256_permutex2var_epi64(_mm256_castsi128_si256(vb012389AB), vpermute_mask, _mm256_castsi128_si256(vb4567CDEF));

      $for M in range(MR):
        ${VACC(M)}x01234567 = ${_MM256_DPBUSD_EPI32}(${VACC(M)}x01234567, va${M}x0123, vb01234567);

      $if DATATYPE in ["QS8", "QD8"]:
        w = (const ${XINT8_T}*) w + 32;
      $elif DATATYPE in ["QC4"]:
        w = (const ${XINT8_T}*) w + 16;
      k -= 4 * sizeof(${XINT8_T});
    }

    $if DATATYPE == "QC4":
      $for M in range(MR):
        ${VACC(M)}x01234567 = _mm256_srai_epi32(${VACC(M)}x01234567, 4);
    $for M in range(MR):
      __m256 vout${M}x01234567 = _mm256_cvtepi32_ps(${VACC(M)}x01234567);

    $if DATATYPE in ["QD8", "QC4"]:
      $for M in range(MR):
        vout${M}x01234567 = _mm256_mul_ps(vout${M}x01234567, _mm256_set1_ps(quantization_params[${M}].inv_scale));

      const __m256 vfilter_output_scale01234567 = _mm256_load_ps((const float*) w);
      const __m256 vbias01234567 = _mm256_load_ps((const float*) w + 8);
      w = (const float*) w + 16;

      $for M in range(MR):
        vout${M}x01234567 = _mm256_fmadd_ps(vout${M}x01234567, vfilter_output_scale01234567, vbias01234567);

      $for M in range(MR):
        vout${M}x01234567 = _mm256_max_ps(vout${M}x01234567, voutput_min);

      $for M in range(MR):
        vout${M}x01234567 = _mm256_min_ps(vout${M}x01234567, voutput_max);

      if XNN_LIKELY(nc >= 8) {
        $for M in range(MR):
          _mm256_storeu_ps(c${M}, vout${M}x01234567);
          a${M} = (const int8_t*) ((uintptr_t) a${M} - kc);
          c${M} = (float*) ((uintptr_t) c${M} + cn_stride);
        nc -= 8;
      } else {
        $if AVX == 2:
          $for M in range(MR):
            __m128 vout${M}x0123 = _mm256_castps256_ps128(vout${M}x01234567);
          if (nc & 4) {
            $for M in range(MR):
              _mm_storeu_ps(c${M}, vout${M}x0123);
              c${M} += 4;
            $for M in range(MR):
              vout${M}x0123 = _mm256_extractf128_ps(vout${M}x01234567, 1);
          }
          if (nc & 2) {
            $for M in range(MR):
              _mm_storel_pi((__m64*) c${M}, vout${M}x0123);
              c${M} += 2;
            $for M in range(MR):
              vout${M}x0123 = _mm_movehl_ps(vout${M}x0123, vout${M}x0123);
          }
          if (nc & 1) {
            $for M in range(MR):
              _mm_store_ss(c${M}, vout${M}x0123);
          }
        $else:
          // Prepare mask for valid 32-bit elements (depends on nc).
          const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - 1);
          $for M in range(MR):
            _mm256_mask_storeu_ps(c${M}, vmask, vout${M}x01234567);
        nc = 0;
      }
    $else:
      $if DATATYPE == "QC8":
        const __m256 vscale01234567 = _mm256_load_ps(w);
        w = (const float*) w + 8;
        $for M in range(MR):
          vout${M}x01234567 = _mm256_mul_ps(vout${M}x01234567, vscale01234567);
      $else:
        $for M in range(MR):
          vout${M}x01234567 = _mm256_mul_ps(vout${M}x01234567, vscale);

      $for M in range(MR):
        vout${M}x01234567 = _mm256_min_ps(vout${M}x01234567, voutput_max_less_zero_point);

      $for M in range(MR):
        vacc${M}x01234567 = _mm256_cvtps_epi32(vout${M}x01234567);

      $for M in range(MR):
        vacc${M}x01234567 = _mm256_add_epi32(vacc${M}x01234567, voutput_zero_point);

      $if AVX == 2:
        $for M in range(MR):
          vacc${M}x01234567 = _mm256_packs_epi32(vacc${M}x01234567, _mm256_castsi128_si256(_mm256_extracti128_si256(vacc${M}x01234567, 1)));
          __m128i voutb${M}x01234567 = _mm256_castsi256_si128(_mm256_packs_epi16(vacc${M}x01234567, vacc${M}x01234567));
      $else:
        $for M in range(MR):
          __m128i voutb${M}x01234567 = ${_MM256_CVTXEPI32_EPI8}(vacc${M}x01234567);

      $for M in range(MR):
        voutb${M}x01234567 = ${_MM_MAX_EPX8}(voutb${M}x01234567, voutput_min);

      if (nc >= 8) {
        $for M in range(MR):
          _mm_storel_epi64((__m128i*) c${M}, voutb${M}x01234567);
          c${M} = (${OUT_T}*) ((uintptr_t) c${M} + cn_stride);
          a${M} = (const ${XINT8_T}*) ((uintptr_t) a${M} - kc);

        nc -= 8;
      } else {
        $if AVX == 2:
          if (nc & 4) {
            $for M in range(MR):
              _mm_storeu_si32(c${M}, voutb${M}x01234567);
              c${M} += 4;
            $for M in range(MR):
              voutb${M}x01234567 = _mm_srli_epi64(voutb${M}x01234567, 32);
          }
          if (nc & 2) {
            $for M in range(MR):
              unaligned_store_u16(c${M}, (uint16_t) _mm_extract_epi16(voutb${M}x01234567, 0));
              c${M} += 2;
            $for M in range(MR):
              voutb${M}x01234567 = _mm_srli_epi32(voutb${M}x01234567, 16);
          }
          if (nc & 1) {
            $for M in range(MR):
              *c${M} = (${XINT8_T}) _mm_extract_epi8(voutb${M}x01234567, 0);
          }
        $else:
          // Prepare mask for valid 8-bit elements (depends on nc).
          const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - UINT32_C(1));
          $for M in range(MR):
            _mm_mask_storeu_epi8(c${M}, vmask, voutb${M}x01234567);
        nc = 0;
      }
  } while (nc != 0);
}