// Copyright 2020 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 SSE in [2, 3, 4]
$assert not AVX or SSE == 4
$assert REQUANTIZATION == "FP32" or not REQUANTIZATION
$assert DATATYPE in ["QC8", "QD8", "QS8", "QU8", "QC4", "QB4"]
$assert DATATYPE != "QC8" or REQUANTIZATION == "FP32"
$assert not DATATYPE in ["QD8", "QC4", "QB4"] or not REQUANTIZATION
$assert VARIANT in ["LD64", "LD128", "EXTENDED"]
$assert MR <= 4
#include <assert.h>

$if AVX:
  #ifdef _MSC_VER
    #include <intrin.h>
  #else
    #include <x86intrin.h>
  #endif
$else:
  $SSE_HEADER = {2: "emmintrin.h", 3: "tmmintrin.h", 4: "smmintrin.h"}[SSE]
  #include <${SSE_HEADER}>

#include "xnnpack/common.h"
#include "xnnpack/gemm.h"
#include "xnnpack/math.h"
$if DATATYPE != "QD8":
  #include "xnnpack/unaligned.h"


$#
$INDENT = 0
$_ = ""
$#
$def SET_INDENT(NEW_INDENT):
  $global INDENT
  $global _
  $INDENT=NEW_INDENT
  $_ = "  " * INDENT
$#
$def BEGIN():
  $SET_INDENT(INDENT + 1)
  $return '{'
$#
$def END():
  $SET_INDENT(INDENT - 1)
  $return _ + '}'
$#
$DATATYPE_SPEC = {"QC8": "qs8_qc8w", "QD8": "qd8_f32_qc8w", "QC4": "qd8_f32_qc4w", "QS8": "qs8", "QU8": "qu8", "QB4": "qd8_f32_qb4w"}[DATATYPE]
$LOAD_SUFFIX = {"LD128": "_ld128", "LD64": "_ld64", "EXTENDED": ""}[VARIANT]
$GEMM_SUFFIX = "_xw" if VARIANT == "EXTENDED" else ""
$REQUANTIZATION_SPEC = "_" + REQUANTIZATION.lower() if REQUANTIZATION else ""
$if DATATYPE in ["QD8", "QC4", "QB4"]:
  $PARAMS_STRUCT = "scalar"
$else:
  $PARAMS_STRUCT = REQUANTIZATION.lower() + "_scalar"
$PARAMS_TYPE = {"QC8": "union xnn_qs8_qc8w_conv_minmax_params", "QD8": "union xnn_f32_minmax_params", "QC4": "struct xnn_f32_qc4w_minmax_params", "QS8": "union xnn_qs8_conv_minmax_params", "QU8": "union xnn_qu8_conv_minmax_params", "QB4": "struct xnn_f32_qb4w_minmax_params"}[DATATYPE]
$XINT8_T = "uint8_t" if DATATYPE == "QU8" else "int8_t"
$OUT_T = "float" if DATATYPE in ["QD8", "QC4", "QB4"] else XINT8_T
$_MM_LOADX_SI128 = "_mm_loadu_si128" if DATATYPE == "QB4" else "_mm_load_si128"
$ISA = "avx" if AVX else {2: "sse2", 3: "ssse3", 4: "sse41"}[SSE]
void xnn_${DATATYPE_SPEC}_gemm${GEMM_SUFFIX}_minmax${REQUANTIZATION_SPEC}_ukernel_${MR}x4c8__${ISA}${LOAD_SUFFIX}(
    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", "QB4"]:
      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);

  $if DATATYPE == "QB4":
    size_t bl = params->scalar.blocksize;
    assert(bl <= round_up_po2(kc, 2));
    assert(bl != 0);
    assert(bl % 32 == 0);
  kc = round_up_po2(kc, 8 * sizeof(${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};
      }

  $if DATATYPE in ["QD8", "QC4", "QB4"]:
    const __m128 vmin = _mm_set1_ps(params->${PARAMS_STRUCT}.min);
    const __m128 vmax = _mm_set1_ps(params->${PARAMS_STRUCT}.max);
    XNN_FORCE_REALIZATION(vmin);
    XNN_FORCE_REALIZATION(vmax);
  $else:
    $if DATATYPE != "QC8":
      const __m128 vscale = _mm_set1_ps(params->${PARAMS_STRUCT}.scale);
      XNN_FORCE_REALIZATION(vscale);

    const __m128 voutput_max_less_zero_point = _mm_set1_ps((int32_t) params->${PARAMS_STRUCT}.output_max - (int32_t) params->${PARAMS_STRUCT}.output_zero_point);
    const __m128i voutput_zero_point = _mm_set1_epi16(params->${PARAMS_STRUCT}.output_zero_point);
    $if DATATYPE == "QU8":
      const __m128i voutput_min = _mm_set1_epi8(params->${PARAMS_STRUCT}.output_min);
    $else:
      $if SSE < 4:
        const __m128i voutput_min = _mm_set1_epi16(params->${PARAMS_STRUCT}.output_min);
      $if SSE == 4:
        const __m128i voutput_min = _mm_set1_epi8(params->${PARAMS_STRUCT}.output_min);
    XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
    XNN_FORCE_REALIZATION(voutput_zero_point);
    XNN_FORCE_REALIZATION(voutput_min);

  $if DATATYPE in ["QC4", "QB4"]:
    const __m128i vmask = _mm_set1_epi8(0xF0);
    XNN_FORCE_REALIZATION(vmask);

  $if DATATYPE == "QU8":
    const __m128i vb_zero_point = _mm_set1_epi16(params->${PARAMS_STRUCT}.kernel_zero_point);
    XNN_FORCE_REALIZATION(vb_zero_point);
  do {
    $if DATATYPE in ["QD8", "QC4"]:
      const __m128i vksum = _mm_load_si128((const __m128i*) w);
      $if AVX:
        $for M in range(MR):
          const __m128i vinput_zero_point${M} = _mm_castps_si128(_mm_broadcast_ss((const float*) &quantization_params[${M}].zero_point));
      $else:
        $for M in range(0, MR, 2):
          $if M + 1 == MR:
            __m128i vinput_zero_point${M} = _mm_cvtsi32_si128(*((const int*) &quantization_params[${M}].zero_point));
            vinput_zero_point${M} = _mm_shuffle_epi32(vinput_zero_point${M}, _MM_SHUFFLE(0, 0, 0, 0));
          $else:
            const __m128i vinput_zero_point${ABC[M:M+2]} = _mm_loadu_si128((const __m128i*) &quantization_params[${M}]);
            const __m128i vinput_zero_point${M} = _mm_shuffle_epi32(vinput_zero_point${ABC[M:M+2]}, _MM_SHUFFLE(0, 0, 0, 0));
            const __m128i vinput_zero_point${M+1} = _mm_shuffle_epi32(vinput_zero_point${ABC[M:M+2]}, _MM_SHUFFLE(2, 2, 2, 2));
      const __m128i vzero = _mm_setzero_si128();
      $if SSE == 4:
        $for M in range(0, MR):
          const __m128i vinit${M} = _mm_mullo_epi32(vksum, vinput_zero_point${M});
        $for M in range(0, MR):
          __m128i vacc${M}x0 = _mm_blend_epi16(vinit${M}, vzero, 0xFC);
          __m128i vacc${M}x1 = _mm_blend_epi16(vinit${M}, vzero, 0xF3);
          __m128i vacc${M}x2 = _mm_blend_epi16(vinit${M}, vzero, 0xCF);
          __m128i vacc${M}x3 = _mm_blend_epi16(vinit${M}, vzero, 0x3F);
      $else:
        const __m128i vksum_lo = _mm_srli_epi32(_mm_slli_epi32(vksum, 16), 16);
        const __m128i vksum_hi = _mm_srli_epi32(vksum, 16);

        $for M in range(0, MR):
          __m128i vzpprodksumhi${M} = _mm_mulhi_epu16(vinput_zero_point${M}, vksum_lo);
          const __m128i vzpprodksumlo${M} = _mm_mullo_epi16(vinput_zero_point${M}, vksum_lo);

        $for M in range(0, MR):
          vzpprodksumhi${M} = _mm_add_epi16(vzpprodksumhi${M}, _mm_mullo_epi16(vinput_zero_point${M}, vksum_hi));
          vzpprodksumhi${M} = _mm_sub_epi16(vzpprodksumhi${M}, _mm_and_si128(_mm_srai_epi16(vinput_zero_point${M}, 15), vksum_lo));

        $for M in range(0, MR):
          vzpprodksumhi${M} = _mm_slli_si128(vzpprodksumhi${M}, 2);

        $for M in range(0, MR):
          const __m128i vksumzp${M} = _mm_or_si128(vzpprodksumhi${M}, vzpprodksumlo${M});

        $for M in range(0, MR):
          const __m128i vksum01${M} = _mm_unpacklo_epi32(vksumzp${M}, vzero);
          const __m128i vksum23${M} = _mm_unpackhi_epi32(vksumzp${M}, vzero);

        $for M in range(0, MR):
          __m128i vacc${M}x0 = _mm_unpacklo_epi64(vksum01${M}, vzero);
          __m128i vacc${M}x1 = _mm_unpackhi_epi64(vksum01${M}, vzero);

        $for M in range(MR):
          __m128i vacc${M}x2 = _mm_unpacklo_epi64(vksum23${M}, vzero);
          __m128i vacc${M}x3 = _mm_unpackhi_epi64(vksum23${M}, vzero);
    $elif DATATYPE == "QB4":
      const __m128 vksum = _mm_load_ps((const float*) w);
      $if AVX:
        $for M in range(MR):
          const __m128i vinput_zero_point${M} = _mm_castps_si128(_mm_broadcast_ss((const float*) &quantization_params[${M}].zero_point));
      $else:
        $for M in range(0, MR, 2):
          $if M + 1 == MR:
            __m128i vinput_zero_point${M} = _mm_cvtsi32_si128(*((const int*) &quantization_params[${M}].zero_point));
            vinput_zero_point${M} = _mm_shuffle_epi32(vinput_zero_point${M}, _MM_SHUFFLE(0, 0, 0, 0));
          $else:
            const __m128i vinput_zero_point${ABC[M:M+2]} = _mm_loadu_si128((const __m128i*) &quantization_params[${M}]);
            const __m128i vinput_zero_point${M} = _mm_shuffle_epi32(vinput_zero_point${ABC[M:M+2]}, _MM_SHUFFLE(0, 0, 0, 0));
            const __m128i vinput_zero_point${M+1} = _mm_shuffle_epi32(vinput_zero_point${ABC[M:M+2]}, _MM_SHUFFLE(2, 2, 2, 2));

      $for M in range(0, MR):
        __m128 vinput_zero_point${M}_float = _mm_cvtepi32_ps(vinput_zero_point${M});
        __m128 vout${M}x0123 = _mm_mul_ps(vksum, vinput_zero_point${M}_float);
    $else:
      $for N in range(4):
        __m128i vacc0x${N} = _mm_cvtsi32_si128(((const int*) w)[${N}]);
      $for M in range(1, MR):
        $for N in range(4):
          __m128i vacc${M}x${N} = vacc0x${N};
    w = (const int32_t*) w + 4;

    $if DATATYPE == "QU8":
      $if SSE < 4 or VARIANT == "LD128":
        const __m128i vzero = _mm_setzero_si128();
    $if DATATYPE == "QB4":
      for (size_t kb=0; kb < kc; kb += bl) {
        $for M in range(MR):
          __m128i vacc${M}x0 = _mm_setzero_si128();
          __m128i vacc${M}x1 = _mm_setzero_si128();
          __m128i vacc${M}x2 = _mm_setzero_si128();
          __m128i vacc${M}x3 = _mm_setzero_si128();
        size_t k = bl;
        $SET_INDENT(1)
    $else:
      size_t k = kc;

    $if DATATYPE in ["QC4", "QB4"]:
      ${_}while (k >= 16 * sizeof(${XINT8_T})) {
        $for M in range(MR):
          ${_}const __m128i va${M}c0 = _mm_loadl_epi64((const __m128i*) a${M});
          $# SSE 4, AVX
          $if SSE == 4:
            ${_}const __m128i vxa${M}c0 = _mm_cvtepi8_epi16(va${M}c0);
          $# SSE 2 only
          $else:
            ${_}const __m128i vxa${M}c0 = _mm_srai_epi16(_mm_unpacklo_epi8(va${M}c0, va${M}c0), 8);
          ${_}a${M} += 8;

        $if VARIANT == "LD128":
          $for N in range(0, 4, 2):
            $if N == 0:
              ${_}const __m128i vb${N}${N+1}c01 = ${_MM_LOADX_SI128}((const __m128i*) w);
            $else:
              ${_}const __m128i vb${N}${N+1}c01 = ${_MM_LOADX_SI128}((const __m128i*) ((const ${XINT8_T}*) w + ${N * 8}));
            ${_}const __m128i vbs${N}${N+1}c0 = _mm_slli_epi32(vb${N}${N+1}c01, 4);
            ${_}const __m128i vb${N}${N+1}c0 = _mm_and_si128(vbs${N}${N+1}c0, vmask);
            ${_}const __m128i vsb${N}${N+1}c0 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb${N}${N+1}c0);
            ${_}const __m128i vxb${N}c0 = _mm_unpacklo_epi8(vb${N}${N+1}c0, vsb${N}${N+1}c0);
            ${_}const __m128i vxb${N+1}c0 = _mm_unpackhi_epi8(vb${N}${N+1}c0, vsb${N}${N+1}c0);

            $for M in range(MR):
              ${_}vacc${M}x${N} = _mm_add_epi32(vacc${M}x${N}, _mm_madd_epi16(vxa${M}c0, vxb${N}c0));
              ${_}vacc${M}x${N+1} = _mm_add_epi32(vacc${M}x${N+1}, _mm_madd_epi16(vxa${M}c0, vxb${N+1}c0));
        $else:
          $for N in range(4):
            $if VARIANT == "LD64":
              $if N == 0:
                ${_}const __m128i vb${N}c01 = _mm_loadl_epi64((const __m128i*) w);
              $else:
                ${_}const __m128i vb${N}c01 = _mm_loadl_epi64((const __m128i*) ((const ${XINT8_T}*) w + ${N * 8}));
            ${_}const __m128i vbs${N}c0 = _mm_slli_epi32(vb${N}c01, 4);
            ${_}const __m128i vb${N}c0 = _mm_and_si128(vbs${N}c0, vmask);
            $# SSE 4, AVX
            $if SSE == 4:
              ${_}const __m128i vxb${N}c0 = _mm_cvtepi8_epi16(vb${N}c0);
            $# SSE 2 only
            $else:
              ${_}const __m128i vxb${N}c0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb${N}c0, vb${N}c0), 8);

            $for M in range(MR):
              ${_}vacc${M}x${N} = _mm_add_epi32(vacc${M}x${N}, _mm_madd_epi16(vxa${M}c0, vxb${N}c0));

        $for M in range(MR):
          ${_}const __m128i va${M}c1 = _mm_loadl_epi64((const __m128i*) a${M});
          $# SSE 4, AVX
          $if SSE == 4:
            ${_}const __m128i vxa${M}c1 = _mm_cvtepi8_epi16(va${M}c1);
          $# SSE 2 only
          $else:
            ${_}const __m128i vxa${M}c1 = _mm_srai_epi16(_mm_unpacklo_epi8(va${M}c1, va${M}c1), 8);
          ${_}a${M} += 8;

        $if VARIANT == "LD128":
          $for N in range(0, 4, 2):
            ${_}const __m128i vb${N}${N+1}c1 = _mm_and_si128(vb${N}${N+1}c01, vmask);
            ${_}const __m128i vsb${N}${N+1}c1 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb${N}${N+1}c1);
            ${_}const __m128i vxb${N}c1 = _mm_unpacklo_epi8(vb${N}${N+1}c1, vsb${N}${N+1}c1);
            ${_}const __m128i vxb${N+1}c1 = _mm_unpackhi_epi8(vb${N}${N+1}c1, vsb${N}${N+1}c1);

            $for M in range(MR):
              ${_}vacc${M}x${N} = _mm_add_epi32(vacc${M}x${N}, _mm_madd_epi16(vxa${M}c1, vxb${N}c1));
              ${_}vacc${M}x${N+1} = _mm_add_epi32(vacc${M}x${N+1}, _mm_madd_epi16(vxa${M}c1, vxb${N+1}c1));
        $else:
          $for N in range(4):
            $if VARIANT == "LD64":
              ${_}const __m128i vb${N}c1 = _mm_and_si128(vb${N}c01, vmask);
              $# SSE 4, AVX
              $if SSE == 4:
                ${_}const __m128i vxb${N}c1 = _mm_cvtepi8_epi16(vb${N}c1);
              $# SSE 2 only
              $else:
                ${_}const __m128i vxb${N}c1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb${N}c1, vb${N}c1), 8);

            $for M in range(MR):
              ${_}vacc${M}x${N} = _mm_add_epi32(vacc${M}x${N}, _mm_madd_epi16(vxa${M}c1, vxb${N}c1));


        ${_}w = (const ${XINT8_T}*) w + 32;
        ${_}k -= 16 * sizeof(${XINT8_T});
      ${_}}

    ${_}while (k >= 8 * sizeof(${XINT8_T})) {
      $for M in range(MR):
        ${_}const __m128i va${M} = _mm_loadl_epi64((const __m128i*) a${M});
        $if DATATYPE == "QU8":
          $# SSE 4, AVX
          $if SSE == 4:
            ${_}const __m128i vxa${M} = _mm_cvtepu8_epi16(va${M});
          $# SSE 2 only
          $else:
            ${_}const __m128i vxa${M} = _mm_unpacklo_epi8(va${M}, vzero);
        $else:
          $# SSE 4, AVX
          $if SSE == 4:
            ${_}const __m128i vxa${M} = _mm_cvtepi8_epi16(va${M});
          $# SSE 2 only
          $else:
            ${_}const __m128i vxa${M} = _mm_srai_epi16(_mm_unpacklo_epi8(va${M}, va${M}), 8);
        ${_}a${M} += 8;

      $if VARIANT == "LD128":
        $for N in range(0, 4, 2):
          $if DATATYPE in ["QC4", "QB4"]:
            $if N == 0:
              ${_}__m128i vb${N}${N+1} = ${_MM_LOADX_SI128}((const __m128i*) w);
            $else:
              ${_}__m128i vb${N}${N+1} = ${_MM_LOADX_SI128}((const __m128i*) ((const ${XINT8_T}*) w + ${N * 8}));
            ${_}vb${N}${N+1} = _mm_slli_epi32(vb${N}${N+1}, 4);
            ${_}vb${N}${N+1} = _mm_and_si128(vb${N}${N+1}, vmask);
          $else:
            $if N == 0:
              ${_}const __m128i vb${N}${N+1} = _mm_load_si128((const __m128i*) w);
            $else:
              ${_}const __m128i vb${N}${N+1} = _mm_load_si128((const __m128i*) ((const ${XINT8_T}*) w + ${N * 8}));

          $if DATATYPE == "QU8":
            ${_}const __m128i vxb${N} = _mm_sub_epi16(_mm_unpacklo_epi8(vb${N}${N+1}, vzero), vb_zero_point);
            ${_}const __m128i vxb${N+1} = _mm_sub_epi16(_mm_unpackhi_epi8(vb${N}${N+1}, vzero), vb_zero_point);
          $# SSE 4, AVX
          $elif SSE == 4:
            ${_}const __m128i vxbm${N+1} = _mm_unpackhi_epi8(vb${N}${N+1}, vb${N}${N+1});
            ${_}const __m128i vxb${N} = _mm_cvtepi8_epi16(vb${N}${N+1});
            ${_}const __m128i vxb${N+1} = _mm_srai_epi16(vxbm${N+1}, 8);
          $# SSE 2 only
          $else:
            ${_}const __m128i vsb${N}${N+1} = _mm_cmpgt_epi8(_mm_setzero_si128(), vb${N}${N+1});
            ${_}const __m128i vxb${N} = _mm_unpacklo_epi8(vb${N}${N+1}, vsb${N}${N+1});
            ${_}const __m128i vxb${N+1} = _mm_unpackhi_epi8(vb${N}${N+1}, vsb${N}${N+1});

          $for M in range(MR):
            ${_}vacc${M}x${N} = _mm_add_epi32(vacc${M}x${N}, _mm_madd_epi16(vxa${M}, vxb${N}));
            ${_}vacc${M}x${N+1} = _mm_add_epi32(vacc${M}x${N+1}, _mm_madd_epi16(vxa${M}, vxb${N+1}));
      $else:
        $for N in range(4):
          $if VARIANT == "LD64":
            $if DATATYPE in ["QC4", "QB4"]:
              $if N == 0:
                ${_}__m128i vb${N} = _mm_loadl_epi64((const __m128i*) w);
              $else:
                ${_}__m128i vb${N} = _mm_loadl_epi64((const __m128i*) ((const ${XINT8_T}*) w + ${N * 8}));
              ${_}vb${N} = _mm_slli_epi32(vb${N}, 4);
              ${_}vb${N} = _mm_and_si128(vb${N}, vmask);
            $else:
              $if N == 0:
                ${_}const __m128i vb${N} = _mm_loadl_epi64((const __m128i*) w);
              $else:
                ${_}const __m128i vb${N} = _mm_loadl_epi64((const __m128i*) ((const ${XINT8_T}*) w + ${N * 8}));

            $if DATATYPE == "QU8":
              $# SSE 4, AVX
              $if SSE == 4:
                ${_}const __m128i vxb${N} = _mm_sub_epi16(_mm_cvtepu8_epi16(vb${N}), vb_zero_point);
              $# SSE 2 only
              $else:
                ${_}const __m128i vxb${N} = _mm_sub_epi16(_mm_unpacklo_epi8(vb${N}, vzero), vb_zero_point);
            $else:
              $# SSE 4, AVX
              $if SSE == 4:
                ${_}const __m128i vxb${N} = _mm_cvtepi8_epi16(vb${N});
              $# SSE 2 only
              $else:
                ${_}const __m128i vxb${N} = _mm_srai_epi16(_mm_unpacklo_epi8(vb${N}, vb${N}), 8);
          $elif VARIANT == "EXTENDED":
            $if N == 0:
              ${_}const __m128i vxb${N} = _mm_load_si128((const __m128i*) w);
            $else:
              ${_}const __m128i vxb${N} = _mm_load_si128((const __m128i*) ((const int16_t*) w + ${N * 8}));

          $for M in range(MR):
            ${_}vacc${M}x${N} = _mm_add_epi32(vacc${M}x${N}, _mm_madd_epi16(vxa${M}, vxb${N}));

      $if VARIANT == "EXTENDED":
        ${_}w = (const int16_t*) w + 32;
      $else:
        ${_}w = (const ${XINT8_T}*) w + 32;
      ${_}k -= 8 * sizeof(${XINT8_T});
    ${_}}
    $if DATATYPE == "QB4":
      ${_}// accumulate float
      $if SSE == 4:
        ${_}const __m128 vfilter_output_scale0123 = _mm_castsi128_ps(_mm_slli_epi32(_mm_cvtepu16_epi32(_mm_loadl_epi64((const __m128i*) w)), 16));
      $else:
        ${_}const __m128 vfilter_output_scale0123 = _mm_castsi128_ps(_mm_unpacklo_epi16(_mm_setzero_si128(), _mm_loadl_epi64((const __m128i*) w)));
      ${_}w = (const uint16_t*) w + 4;

    $if SSE >= 3:
      $for M in range(MR):
        ${_}const __m128i vacc${M}x01 = _mm_hadd_epi32(vacc${M}x0, vacc${M}x1);
        ${_}const __m128i vacc${M}x23 = _mm_hadd_epi32(vacc${M}x2, vacc${M}x3);

      $for M in range(MR):
        ${_}${_}__m128i vacc${M}x0123 = _mm_hadd_epi32(vacc${M}x01, vacc${M}x23);
    $else:
      $for M in range(MR):
        ${_}const __m128i vacc${M}x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc${M}x0, vacc${M}x2), _mm_unpackhi_epi32(vacc${M}x0, vacc${M}x2));
        ${_}const __m128i vacc${M}x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc${M}x1, vacc${M}x3), _mm_unpackhi_epi32(vacc${M}x1, vacc${M}x3));

      $for M in range(MR):
        ${_}__m128i vacc${M}x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc${M}x02, vacc${M}x13), _mm_unpackhi_epi32(vacc${M}x02, vacc${M}x13));

    $if DATATYPE == "QB4":
      $for M in range(MR):
        ${_}vout${M}x0123 = _mm_add_ps(vout${M}x0123, _mm_mul_ps(_mm_cvtepi32_ps(vacc${M}x0123), vfilter_output_scale0123));
      }
    $if DATATYPE in ["QD8", "QC4", "QB4"]:
      $if DATATYPE == "QC4":
        $for M in range(MR):
          vacc${M}x0123 = _mm_srai_epi32(vacc${M}x0123, 4);
      $if DATATYPE != "QB4":
        $for M in range(MR):
          __m128 vout${M}x0123 = _mm_cvtepi32_ps(vacc${M}x0123);

      $if AVX:
        $for M in range(MR):
          const __m128 vinput_scale${M} = _mm_broadcast_ss(&quantization_params[${M}].inv_scale);
      $else:
        $for M in range(0, MR, 2):
          $if M + 1 == MR:
            const __m128 vinput_scale${M} = _mm_load1_ps(&quantization_params[${M}].inv_scale);
          $else:
            const __m128i vinput_scale${ABC[M:M+2]} = _mm_loadu_si128((const __m128i*) &quantization_params[${M}]);
            const __m128 vinput_scale${M} = _mm_castsi128_ps(_mm_shuffle_epi32(vinput_scale${ABC[M:M+2]}, _MM_SHUFFLE(1, 1, 1, 1)));
            const __m128 vinput_scale${M+1} = _mm_castsi128_ps(_mm_shuffle_epi32(vinput_scale${ABC[M:M+2]}, _MM_SHUFFLE(3, 3, 3, 3)));

      $for M in range(MR):
        vout${M}x0123 = _mm_mul_ps(vout${M}x0123, vinput_scale${M});

      $if DATATYPE in ["QD8", "QC4"]:
        const __m128 vfilter_output_scale0123 = _mm_load_ps((const float*) w);
        $for M in range(MR):
          vout${M}x0123 = _mm_mul_ps(vout${M}x0123, vfilter_output_scale0123);

      $if DATATYPE == "QB4":
        const __m128 vbias0123 = _mm_loadu_ps((const float*) w);
        w = (const float*) w + 4;
      $else:
        const __m128 vbias0123 = _mm_load_ps((const float*) w + 4);
        w = (const float*) w + 8;
      $for M in range(MR):
        vout${M}x0123 = _mm_add_ps(vout${M}x0123, vbias0123);

      $for M in range(MR):
        vout${M}x0123 = _mm_max_ps(vout${M}x0123, vmin);

      $for M in range(MR):
        vout${M}x0123 = _mm_min_ps(vout${M}x0123, vmax);

      if XNN_LIKELY(nc >= 4) {
        $for M in range(MR):
          _mm_storeu_ps(c${M}, vout${M}x0123);

        $for M in range(MR):
          a${M} = (const int8_t*) ((uintptr_t) a${M} - kc);

        $for M in range(MR):
          c${M} = (float*) ((uintptr_t) c${M} + cn_stride);

        nc -= 4;
      } else {
        if (nc & 2) {
          $for M in range(MR):
            _mm_storel_pi((__m64*) c${M}, vout${M}x0123);
            vout${M}x0123 = _mm_unpackhi_ps(vout${M}x0123, vout${M}x0123);
            c${M} += 2;
        }
        if (nc & 1) {
          $for M in range(MR):
            _mm_store_ss(c${M}, vout${M}x0123);
        }
        nc = 0;
      }
    $else:
      $for M in range(MR):
        __m128 vscaled${M}x0123 = _mm_cvtepi32_ps(vacc${M}x0123);

      $if DATATYPE == "QC8":
        const __m128 vscale0123 = _mm_load_ps((const float*) w);
        w = (const float*) w + 4;
        $for M in range(MR):
          vscaled${M}x0123 = _mm_mul_ps(vscaled${M}x0123, vscale0123);
      $else:
        $for M in range(MR):
          vscaled${M}x0123 = _mm_mul_ps(vscaled${M}x0123, vscale);

      $for M in range(MR):
        vscaled${M}x0123 = _mm_min_ps(vscaled${M}x0123, voutput_max_less_zero_point);

      $for M in range(MR):
        vacc${M}x0123 = _mm_cvtps_epi32(vscaled${M}x0123);

      $for M in range(0, MR, 2):
        __m128i vacc${M}${min(M+1, MR-1)}x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc${M}x0123, vacc${min(M+1, MR-1)}x0123), voutput_zero_point);

      $if DATATYPE == "QU8":
        $if MR > 2:
          __m128i vout = _mm_packus_epi16(vacc0${min(1, MR-1)}x0123, vacc${min(2, MR-1)}${min(3, MR-1)}x0123);
        $else:
          __m128i vout = _mm_packus_epi16(vacc0${min(1, MR-1)}x0123, vacc0${min(1, MR-1)}x0123);

        vout = _mm_max_epu8(vout, voutput_min);
      $else:
        $if SSE < 4:
          $for M in range(0, MR, 2):
            vacc${M}${min(M+1, MR-1)}x0123 = _mm_max_epi16(vacc${M}${min(M+1, MR-1)}x0123, voutput_min);

        $if MR > 2:
          __m128i vout = _mm_packs_epi16(vacc0${min(1, MR-1)}x0123, vacc${min(2, MR-1)}${min(3, MR-1)}x0123);
        $else:
          __m128i vout = _mm_packs_epi16(vacc0${min(1, MR-1)}x0123, vacc0${min(1, MR-1)}x0123);

        $# SSE 4, AVX
        $if SSE == 4:
          vout = _mm_max_epi8(vout, voutput_min);

      if (nc >= 4) {
        unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
        $for M in range(1, MR):
          $# SSE 4, AVX
          $if SSE == 4:
            unaligned_store_u32(c${M}, (uint32_t) _mm_extract_epi32(vout, ${M}));
          $# SSE 2 only
          $else:
            vout = _mm_srli_si128(vout, 4);
            unaligned_store_u32(c${M}, (uint32_t) _mm_cvtsi128_si32(vout));

        $for M in range(MR):
          c${M} = (${OUT_T}*) ((uintptr_t) c${M} + cn_stride);

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

        nc -= 4;
      } else {
        if (nc & 2) {
          $for M in range(MR):
            unaligned_store_u16(c${M}, (uint16_t) _mm_extract_epi16(vout, ${M * 2}));
            c${M} += 2;
          vout = _mm_srli_epi32(vout, 16);
        }
        if (nc & 1) {
          $# SSE 4, AVX
          $if SSE == 4:
            $for M in range(MR):
              *c${M} = (${OUT_T}) _mm_extract_epi8(vout, ${M * 4});
          $# SSE 2 only
          $else:
            *c0 = (${OUT_T}) _mm_cvtsi128_si32(vout);
            $for M in range(1, MR):
              *c${M} = (${OUT_T}) _mm_extract_epi16(vout, ${M * 2});
        }

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