src/qs8-igemm/MRx16c8-avx512skx.c.in - platform/external/XNNPACK - Gitiles

 // 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 VARIANT in ["LD256", "EXTENDED"]
 $assert MR <= 4
 #include <assert.h>

 #include <immintrin.h>

 #include <xnnpack/igemm.h>
 #include <xnnpack/intrinsics-polyfill.h>


 $GEMM_SUFFIX = "_xw" if VARIANT == "EXTENDED" else ""
 void xnn_qs8_igemm${GEMM_SUFFIX}_minmax_ukernel_${MR}x16c8__avx512skx(
     size_t mr,
     size_t nc,
     size_t kc,
     size_t ks,
     const int8_t** restrict a,
     const void* restrict w,
     int8_t* restrict c,
     size_t cm_stride,
     size_t cn_stride,
     size_t a_offset,
     const int8_t* zero,
     const union xnn_qs8_gemm${GEMM_SUFFIX}_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
 {
   assert(mr != 0);
   assert(mr <= ${MR});
   assert(nc != 0);
   assert(kc != 0);
   assert(kc % sizeof(int8_t) == 0);
   assert(a != NULL);
   assert(w != NULL);
   assert(c != NULL);

   int8_t* c0 = c;
   $for M in range(1, MR):
     int8_t* c${M} = (int8_t*) ((uintptr_t) c${M-1} + cm_stride);
     $if M % 2 == 0:
       if XNN_UNPREDICTABLE(mr <= ${M}) {
         c${M} = c${M-1};
       }
     $elif M + 1 == MR:
       if XNN_UNPREDICTABLE(mr != ${M+1}) {
         c${M} = c${M-1};
       }
     $else:
       if XNN_UNPREDICTABLE(mr < ${M+1}) {
         c${M} = c${M-1};
       }

   const __mmask16 vbias_mask = _cvtu32_mask16(0x1111);
   const __mmask16 vblend_mask = _cvtu32_mask16(0xAAAA);
   const __m512i vmultiplier = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.multiplier));
   const __m512i vrounding = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.rounding));
   const __m512i vremainder_mask = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.remainder_mask));
   const __m512i vremainder_threshold = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.remainder_threshold));
   const __m128i vshift = _mm_load_si128((const __m128i*) params->sse2.shift);
   $if MR > 1:
     const __m512i voutput_zero_point = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.output_zero_point));
     const __m512i voutput_min = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.output_min));
     const __m512i voutput_max = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.output_max));
   $else:
     const __m256i voutput_zero_point = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.output_zero_point));
     const __m256i voutput_min = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.output_min));
     const __m256i voutput_max = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.output_max));
   do {
     __m512i vacc0x0123 = _mm512_maskz_expandloadu_epi32(vbias_mask, w);
     $for N in range(4, 16, 4):
       __m512i vacc0x${ABC[N:N+4]} = _mm512_maskz_expandloadu_epi32(vbias_mask, (const void*) ((uintptr_t) w + ${N} * sizeof(int32_t)));
     $for M in range(1, MR):
       $for N in range(0, 16, 4):
         __m512i vacc${M}x${ABC[N:N+4]} = vacc0x${ABC[N:N+4]};
     w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t));

     size_t p = ks;
     do {
       $for M in range(MR):
         const int8_t* restrict a${M} = a[${M}];
         if XNN_UNPREDICTABLE(a${M} != zero) {
           a${M} = (const int8_t*) ((uintptr_t) a${M} + a_offset);
         }
       a += ${MR};

       size_t k = 0;
       while (k < kc) {
         $for M in range(MR):
           const __m512i va${M} = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a${M})));
           a${M} += 8;

         $for N in range(0, 16, 4):
           $if VARIANT == "EXTENDED":
             $if N == 0:
               const __m512i vb${ABC[N:N+4]} = _mm512_load_si512((const __m512i*) w);
             $else:
               const __m512i vb${ABC[N:N+4]} = _mm512_load_si512((const __m512i*) ((uintptr_t) w + ${N * 8} * sizeof(int16_t)));
           $else:
             $if N == 0:
               const __m512i vb${ABC[N:N+4]} = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) w));
             $else:
               const __m512i vb${ABC[N:N+4]} = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) ((uintptr_t) w + ${N * 8} * sizeof(int8_t))));

           $for M in range(MR):
             vacc${M}x${ABC[N:N+4]} = _mm512_add_epi32(vacc${M}x${ABC[N:N+4]}, _mm512_madd_epi16(va${M}, vb${ABC[N:N+4]}));

         $if VARIANT == "EXTENDED":
           w = (const void*) ((uintptr_t) w + 128 * sizeof(int16_t));
         $else:
           w = (const void*) ((uintptr_t) w + 128 * sizeof(int8_t));
         k += 8 * sizeof(int8_t);
       }
       p -= ${MR} * sizeof(void*);
     } while (p != 0);

     $for M in range(MR):
       const __m512i vacc${M}x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc${M}x0123, vacc${M}x4567), _mm512_unpackhi_epi32(vacc${M}x0123, vacc${M}x4567));
       const __m512i vacc${M}x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc${M}x89AB, vacc${M}xCDEF), _mm512_unpackhi_epi32(vacc${M}x89AB, vacc${M}xCDEF));

     $for M in range(MR):
       __m512i vacc${M}x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc${M}x04152637, vacc${M}x8C9DAEBF), _mm512_unpackhi_epi32(vacc${M}x04152637, vacc${M}x8C9DAEBF));

     $for M in range(MR):
       const __m512i vacc${M}x88CC99DDAAEEBBFF = _mm512_shuffle_epi32(vacc${M}x084C195D2A6E3B7F, _MM_SHUFFLE(3, 3, 1, 1));

     $for M in range(MR):
       const __m512i vprod${M}x04152637 = _mm512_add_epi64(_mm512_mul_epi32(vacc${M}x084C195D2A6E3B7F, vmultiplier), vrounding);

     $for M in range(MR):
       const __m512i vprod${M}x8C9DAEBF = _mm512_add_epi64(_mm512_mul_epi32(vacc${M}x88CC99DDAAEEBBFF, vmultiplier), vrounding);

     $for M in range(MR):
       const __m512i vq31prod${M}x04152637 = _mm512_srli_epi64(vprod${M}x04152637, 31);
       const __m512i vq31prod${M}x8C9DAEBF = _mm512_add_epi64(vprod${M}x8C9DAEBF, vprod${M}x8C9DAEBF);

     $for M in range(MR):
       const __m512i vq31prod${M}x084C195D2A6E3B7F = _mm512_mask_blend_epi32(vblend_mask, vq31prod${M}x04152637, vq31prod${M}x8C9DAEBF);

     $for M in range(MR):
       const __m512i vrem${M}x084C195D2A6E3B7F =
         _mm512_add_epi32(_mm512_and_si512(vq31prod${M}x084C195D2A6E3B7F, vremainder_mask), _mm512_srai_epi32(vq31prod${M}x084C195D2A6E3B7F, 31));

     $for M in range(MR):
       vacc${M}x084C195D2A6E3B7F = _mm512_sra_epi32(vq31prod${M}x084C195D2A6E3B7F, vshift);

     const __m512i vminus_one = _mm512_set1_epi32(-1);
     $for M in range(MR):
       vacc${M}x084C195D2A6E3B7F =
         _mm512_mask_sub_epi32(vacc${M}x084C195D2A6E3B7F, _mm512_cmpgt_epi32_mask(vrem${M}x084C195D2A6E3B7F, vremainder_threshold), vacc${M}x084C195D2A6E3B7F, vminus_one);

     $if MR == 1:
       __m256i vacc0x084C2A6E195D3B7F = _mm256_adds_epi16(_mm256_packs_epi32(_mm512_castsi512_si256(vacc0x084C195D2A6E3B7F), _mm512_extracti32x8_epi32(vacc0x084C195D2A6E3B7F, 1)), voutput_zero_point);
       vacc0x084C2A6E195D3B7F = _mm256_min_epi16(_mm256_max_epi16(vacc0x084C2A6E195D3B7F, voutput_min), voutput_max);
     $else:
       $for M in range(0, MR, 2):
         __m512i vacc${M}${min(M+1, MR-1)}x084Cx195Dx2A6Ex3B7F = _mm512_adds_epi16(_mm512_packs_epi32(vacc${M}x084C195D2A6E3B7F, vacc${min(M+1, MR-1)}x084C195D2A6E3B7F), voutput_zero_point);

       $for M in range(0, MR, 2):
         vacc${M}${min(M+1, MR-1)}x084Cx195Dx2A6Ex3B7F = _mm512_min_epi16(_mm512_max_epi16(vacc${M}${min(M+1, MR-1)}x084Cx195Dx2A6Ex3B7F, voutput_min), voutput_max);

     $if MR > 2:
       __m512i vout012${min(3, MR-1)}x084Cx195Dx2A6Ex3B7F = _mm512_packs_epi16(vacc01x084Cx195Dx2A6Ex3B7F, vacc2${min(3, MR-1)}x084Cx195Dx2A6Ex3B7F);
       vout012${min(M+3, MR-1)}x084Cx195Dx2A6Ex3B7F = _mm512_permutexvar_epi32(_mm512_set_epi32(15, 11, 7, 3, 14, 10, 6, 2, 13, 9, 5, 1, 12, 8, 4, 0), vout012${min(3, MR-1)}x084Cx195Dx2A6Ex3B7F);
       const __m512i vout012${min(3, MR-1)}x0123456789ABCDEF = _mm512_shuffle_epi8(vout012${min(3, MR-1)}x084Cx195Dx2A6Ex3B7F, _mm512_set_epi8(15, 11, 7, 3, 13, 9, 5, 1, 14, 10, 6, 2, 12, 8, 4, 0, 15, 11, 7, 3, 13, 9, 5, 1, 14, 10, 6, 2, 12, 8, 4, 0, 15, 11, 7, 3, 13, 9, 5, 1, 14, 10, 6, 2, 12, 8, 4, 0, 15, 11, 7, 3, 13, 9, 5, 1, 14, 10, 6, 2, 12, 8, 4, 0));
     $elif MR == 2:
       const __m256i vout01x084Cx2A6Ex195Dx3B7F = _mm256_packs_epi16(_mm512_castsi512_si256(vacc01x084Cx195Dx2A6Ex3B7F), _mm512_extracti32x8_epi32(vacc01x084Cx195Dx2A6Ex3B7F, 1));
       const __m256i vout01x084C2A6E195D3B7F = _mm256_permutevar8x32_epi32(vout01x084Cx2A6Ex195Dx3B7F, _mm256_set_epi32(7, 5, 3, 1, 6, 4, 2, 0));
       const __m256i vout01x0123456789ABCDEF = _mm256_shuffle_epi8(vout01x084C2A6E195D3B7F, _mm256_set_epi8(15, 7, 11, 3, 13, 5, 9, 1, 14, 6, 10, 2, 12, 4, 8, 0, 15, 7, 11, 3, 13, 5, 9, 1, 14, 6, 10, 2, 12, 4, 8, 0));
     $elif MR == 1:
       const __m128i vout0x084C2A6E195D3B7F = _mm_packs_epi16(_mm256_castsi256_si128(vacc0x084C2A6E195D3B7F), _mm256_extracti128_si256(vacc0x084C2A6E195D3B7F, 1));
       const __m128i vout0x0123456789ABCDEF = _mm_shuffle_epi8(vout0x084C2A6E195D3B7F, _mm_set_epi8(15, 7, 11, 3, 13, 5, 9, 1, 14, 6, 10, 2, 12, 4, 8, 0));

     $if MR > 2:
       if (nc >= 16) {
         $for M in reversed(range(1, MR)):
           _mm_storeu_si128((__m128i*) c${M}, _mm512_extracti32x4_epi32(vout012${min(M+3, MR-1)}x0123456789ABCDEF, ${M}));
         _mm_storeu_si128((__m128i*) c0, _mm512_castsi512_si128(vout012${min(M+3, MR-1)}x0123456789ABCDEF));

         $for M in reversed(range(MR)):
           c${M} = (int8_t*) ((uintptr_t) c${M} + cn_stride);

         a = (const int8_t**restrict) ((uintptr_t) a - ks);

         nc -= 16;
       } else {
         // Prepare mask for valid 8-bit elements (depends on nc).
         __mmask64 vmask = _cvtu64_mask64((uint64_t) ((UINT64_C(1) << (nc + ${16 * (MR - 1)})) - (UINT64_C(1) << ${16 * (MR - 1)})));

         $for M in reversed(range(1, MR)):
           _mm512_mask_storeu_epi8(c${M} - ${M * 16}, vmask, vout012${min(3, MR-1)}x0123456789ABCDEF);
           vmask = _kshiftri_mask64(vmask, 16);
         _mm512_mask_storeu_epi8(c0, vmask, vout012${min(M+3, MR-1)}x0123456789ABCDEF);

         nc = 0;
       }
     $elif MR == 2:
       if (nc >= 16) {
         _mm_storeu_si128((__m128i*) c1, _mm256_extracti128_si256(vout01x0123456789ABCDEF, 1));
         _mm_storeu_si128((__m128i*) c0, _mm256_castsi256_si128(vout01x0123456789ABCDEF));

         $for M in reversed(range(MR)):
           c${M} = (int8_t*) ((uintptr_t) c${M} + cn_stride);

         a = (const int8_t**restrict) ((uintptr_t) a - ks);

         nc -= 16;
       } else {
         // Prepare mask for valid 8-bit elements (depends on nc).
         __mmask64 vmask = _cvtu64_mask64((uint64_t) ((UINT32_C(1) << (nc + 16)) - (UINT32_C(1) << 16)));

         _mm256_mask_storeu_epi8(c1 - 16, vmask, vout01x0123456789ABCDEF);
         vmask = _kshiftri_mask64(vmask, 16);
         _mm256_mask_storeu_epi8(c0, vmask, vout01x0123456789ABCDEF);

         nc = 0;
       }
     $elif MR == 1:
       if (nc >= 16) {
         _mm_storeu_si128((__m128i*) c0, vout0x0123456789ABCDEF);

         $for M in range(MR):
           c${M} = (int8_t*) ((uintptr_t) c${M} + cn_stride);

         a = (const int8_t**restrict) ((uintptr_t) a - ks);

         nc -= 16;
       } else {
         // Prepare mask for valid 8-bit elements (depends on nc).
         const __mmask64 vmask = _cvtu64_mask64((uint64_t) ((UINT32_C(1) << nc) - UINT32_C(1)));

         _mm_mask_storeu_epi8(c0, vmask, vout0x0123456789ABCDEF);

         nc = 0;
       }
   } while (nc != 0);
 }
	// 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 VARIANT in ["LD256", "EXTENDED"]
	$assert MR <= 4
	#include <assert.h>

	#include <immintrin.h>

	#include <xnnpack/igemm.h>
	#include <xnnpack/intrinsics-polyfill.h>


	$GEMM_SUFFIX = "_xw" if VARIANT == "EXTENDED" else ""
	void xnn_qs8_igemm${GEMM_SUFFIX}_minmax_ukernel_${MR}x16c8__avx512skx(
	size_t mr,
	size_t nc,
	size_t kc,
	size_t ks,
	const int8_t** restrict a,
	const void* restrict w,
	int8_t* restrict c,
	size_t cm_stride,
	size_t cn_stride,
	size_t a_offset,
	const int8_t* zero,
	const union xnn_qs8_gemm${GEMM_SUFFIX}_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
	{
	assert(mr != 0);
	assert(mr <= ${MR});
	assert(nc != 0);
	assert(kc != 0);
	assert(kc % sizeof(int8_t) == 0);
	assert(a != NULL);
	assert(w != NULL);
	assert(c != NULL);

	int8_t* c0 = c;
	$for M in range(1, MR):
	int8_t* c${M} = (int8_t*) ((uintptr_t) c${M-1} + cm_stride);
	$if M % 2 == 0:
	if XNN_UNPREDICTABLE(mr <= ${M}) {
	c${M} = c${M-1};
	}
	$elif M + 1 == MR:
	if XNN_UNPREDICTABLE(mr != ${M+1}) {
	c${M} = c${M-1};
	}
	$else:
	if XNN_UNPREDICTABLE(mr < ${M+1}) {
	c${M} = c${M-1};
	}

	const __mmask16 vbias_mask = _cvtu32_mask16(0x1111);
	const __mmask16 vblend_mask = _cvtu32_mask16(0xAAAA);
	const __m512i vmultiplier = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.multiplier));
	const __m512i vrounding = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.rounding));
	const __m512i vremainder_mask = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.remainder_mask));
	const __m512i vremainder_threshold = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.remainder_threshold));
	const __m128i vshift = _mm_load_si128((const __m128i*) params->sse2.shift);
	$if MR > 1:
	const __m512i voutput_zero_point = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.output_zero_point));
	const __m512i voutput_min = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.output_min));
	const __m512i voutput_max = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.output_max));
	$else:
	const __m256i voutput_zero_point = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.output_zero_point));
	const __m256i voutput_min = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.output_min));
	const __m256i voutput_max = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.output_max));
	do {
	__m512i vacc0x0123 = _mm512_maskz_expandloadu_epi32(vbias_mask, w);
	$for N in range(4, 16, 4):
	__m512i vacc0x${ABC[N:N+4]} = _mm512_maskz_expandloadu_epi32(vbias_mask, (const void) ((uintptr_t) w + ${N} sizeof(int32_t)));
	$for M in range(1, MR):
	$for N in range(0, 16, 4):
	__m512i vacc${M}x${ABC[N:N+4]} = vacc0x${ABC[N:N+4]};
	w = (const void) ((uintptr_t) w + 16 sizeof(int32_t));

	size_t p = ks;
	do {
	$for M in range(MR):
	const int8_t* restrict a${M} = a[${M}];
	if XNN_UNPREDICTABLE(a${M} != zero) {
	a${M} = (const int8_t*) ((uintptr_t) a${M} + a_offset);
	}
	a += ${MR};

	size_t k = 0;
	while (k < kc) {
	$for M in range(MR):
	const __m512i va${M} = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a${M})));
	a${M} += 8;

	$for N in range(0, 16, 4):
	$if VARIANT == "EXTENDED":
	$if N == 0:
	const __m512i vb${ABC[N:N+4]} = _mm512_load_si512((const __m512i*) w);
	$else:
	const __m512i vb${ABC[N:N+4]} = _mm512_load_si512((const __m512i) ((uintptr_t) w + ${N 8} * sizeof(int16_t)));
	$else:
	$if N == 0:
	const __m512i vb${ABC[N:N+4]} = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) w));
	$else:
	const __m512i vb${ABC[N:N+4]} = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i) ((uintptr_t) w + ${N 8} * sizeof(int8_t))));

	$for M in range(MR):
	vacc${M}x${ABC[N:N+4]} = _mm512_add_epi32(vacc${M}x${ABC[N:N+4]}, _mm512_madd_epi16(va${M}, vb${ABC[N:N+4]}));

	$if VARIANT == "EXTENDED":
	w = (const void) ((uintptr_t) w + 128 sizeof(int16_t));
	$else:
	w = (const void) ((uintptr_t) w + 128 sizeof(int8_t));
	k += 8 * sizeof(int8_t);
	}
	p -= ${MR} * sizeof(void*);
	} while (p != 0);

	$for M in range(MR):
	const __m512i vacc${M}x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc${M}x0123, vacc${M}x4567), _mm512_unpackhi_epi32(vacc${M}x0123, vacc${M}x4567));
	const __m512i vacc${M}x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc${M}x89AB, vacc${M}xCDEF), _mm512_unpackhi_epi32(vacc${M}x89AB, vacc${M}xCDEF));

	$for M in range(MR):
	__m512i vacc${M}x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc${M}x04152637, vacc${M}x8C9DAEBF), _mm512_unpackhi_epi32(vacc${M}x04152637, vacc${M}x8C9DAEBF));

	$for M in range(MR):
	const __m512i vacc${M}x88CC99DDAAEEBBFF = _mm512_shuffle_epi32(vacc${M}x084C195D2A6E3B7F, _MM_SHUFFLE(3, 3, 1, 1));

	$for M in range(MR):
	const __m512i vprod${M}x04152637 = _mm512_add_epi64(_mm512_mul_epi32(vacc${M}x084C195D2A6E3B7F, vmultiplier), vrounding);

	$for M in range(MR):
	const __m512i vprod${M}x8C9DAEBF = _mm512_add_epi64(_mm512_mul_epi32(vacc${M}x88CC99DDAAEEBBFF, vmultiplier), vrounding);

	$for M in range(MR):
	const __m512i vq31prod${M}x04152637 = _mm512_srli_epi64(vprod${M}x04152637, 31);
	const __m512i vq31prod${M}x8C9DAEBF = _mm512_add_epi64(vprod${M}x8C9DAEBF, vprod${M}x8C9DAEBF);

	$for M in range(MR):
	const __m512i vq31prod${M}x084C195D2A6E3B7F = _mm512_mask_blend_epi32(vblend_mask, vq31prod${M}x04152637, vq31prod${M}x8C9DAEBF);

	$for M in range(MR):
	const __m512i vrem${M}x084C195D2A6E3B7F =
	_mm512_add_epi32(_mm512_and_si512(vq31prod${M}x084C195D2A6E3B7F, vremainder_mask), _mm512_srai_epi32(vq31prod${M}x084C195D2A6E3B7F, 31));

	$for M in range(MR):
	vacc${M}x084C195D2A6E3B7F = _mm512_sra_epi32(vq31prod${M}x084C195D2A6E3B7F, vshift);

	const __m512i vminus_one = _mm512_set1_epi32(-1);
	$for M in range(MR):
	vacc${M}x084C195D2A6E3B7F =
	_mm512_mask_sub_epi32(vacc${M}x084C195D2A6E3B7F, _mm512_cmpgt_epi32_mask(vrem${M}x084C195D2A6E3B7F, vremainder_threshold), vacc${M}x084C195D2A6E3B7F, vminus_one);

	$if MR == 1:
	__m256i vacc0x084C2A6E195D3B7F = _mm256_adds_epi16(_mm256_packs_epi32(_mm512_castsi512_si256(vacc0x084C195D2A6E3B7F), _mm512_extracti32x8_epi32(vacc0x084C195D2A6E3B7F, 1)), voutput_zero_point);
	vacc0x084C2A6E195D3B7F = _mm256_min_epi16(_mm256_max_epi16(vacc0x084C2A6E195D3B7F, voutput_min), voutput_max);
	$else:
	$for M in range(0, MR, 2):
	__m512i vacc${M}${min(M+1, MR-1)}x084Cx195Dx2A6Ex3B7F = _mm512_adds_epi16(_mm512_packs_epi32(vacc${M}x084C195D2A6E3B7F, vacc${min(M+1, MR-1)}x084C195D2A6E3B7F), voutput_zero_point);

	$for M in range(0, MR, 2):
	vacc${M}${min(M+1, MR-1)}x084Cx195Dx2A6Ex3B7F = _mm512_min_epi16(_mm512_max_epi16(vacc${M}${min(M+1, MR-1)}x084Cx195Dx2A6Ex3B7F, voutput_min), voutput_max);

	$if MR > 2:
	__m512i vout012${min(3, MR-1)}x084Cx195Dx2A6Ex3B7F = _mm512_packs_epi16(vacc01x084Cx195Dx2A6Ex3B7F, vacc2${min(3, MR-1)}x084Cx195Dx2A6Ex3B7F);
	vout012${min(M+3, MR-1)}x084Cx195Dx2A6Ex3B7F = _mm512_permutexvar_epi32(_mm512_set_epi32(15, 11, 7, 3, 14, 10, 6, 2, 13, 9, 5, 1, 12, 8, 4, 0), vout012${min(3, MR-1)}x084Cx195Dx2A6Ex3B7F);
	const __m512i vout012${min(3, MR-1)}x0123456789ABCDEF = _mm512_shuffle_epi8(vout012${min(3, MR-1)}x084Cx195Dx2A6Ex3B7F, _mm512_set_epi8(15, 11, 7, 3, 13, 9, 5, 1, 14, 10, 6, 2, 12, 8, 4, 0, 15, 11, 7, 3, 13, 9, 5, 1, 14, 10, 6, 2, 12, 8, 4, 0, 15, 11, 7, 3, 13, 9, 5, 1, 14, 10, 6, 2, 12, 8, 4, 0, 15, 11, 7, 3, 13, 9, 5, 1, 14, 10, 6, 2, 12, 8, 4, 0));
	$elif MR == 2:
	const __m256i vout01x084Cx2A6Ex195Dx3B7F = _mm256_packs_epi16(_mm512_castsi512_si256(vacc01x084Cx195Dx2A6Ex3B7F), _mm512_extracti32x8_epi32(vacc01x084Cx195Dx2A6Ex3B7F, 1));
	const __m256i vout01x084C2A6E195D3B7F = _mm256_permutevar8x32_epi32(vout01x084Cx2A6Ex195Dx3B7F, _mm256_set_epi32(7, 5, 3, 1, 6, 4, 2, 0));
	const __m256i vout01x0123456789ABCDEF = _mm256_shuffle_epi8(vout01x084C2A6E195D3B7F, _mm256_set_epi8(15, 7, 11, 3, 13, 5, 9, 1, 14, 6, 10, 2, 12, 4, 8, 0, 15, 7, 11, 3, 13, 5, 9, 1, 14, 6, 10, 2, 12, 4, 8, 0));
	$elif MR == 1:
	const __m128i vout0x084C2A6E195D3B7F = _mm_packs_epi16(_mm256_castsi256_si128(vacc0x084C2A6E195D3B7F), _mm256_extracti128_si256(vacc0x084C2A6E195D3B7F, 1));
	const __m128i vout0x0123456789ABCDEF = _mm_shuffle_epi8(vout0x084C2A6E195D3B7F, _mm_set_epi8(15, 7, 11, 3, 13, 5, 9, 1, 14, 6, 10, 2, 12, 4, 8, 0));

	$if MR > 2:
	if (nc >= 16) {
	$for M in reversed(range(1, MR)):
	_mm_storeu_si128((__m128i*) c${M}, _mm512_extracti32x4_epi32(vout012${min(M+3, MR-1)}x0123456789ABCDEF, ${M}));
	_mm_storeu_si128((__m128i*) c0, _mm512_castsi512_si128(vout012${min(M+3, MR-1)}x0123456789ABCDEF));

	$for M in reversed(range(MR)):
	c${M} = (int8_t*) ((uintptr_t) c${M} + cn_stride);

	a = (const int8_t**restrict) ((uintptr_t) a - ks);

	nc -= 16;
	} else {
	// Prepare mask for valid 8-bit elements (depends on nc).
	__mmask64 vmask = _cvtu64_mask64((uint64_t) ((UINT64_C(1) << (nc + ${16 * (MR - 1)})) - (UINT64_C(1) << ${16 * (MR - 1)})));

	$for M in reversed(range(1, MR)):
	_mm512_mask_storeu_epi8(c${M} - ${M * 16}, vmask, vout012${min(3, MR-1)}x0123456789ABCDEF);
	vmask = _kshiftri_mask64(vmask, 16);
	_mm512_mask_storeu_epi8(c0, vmask, vout012${min(M+3, MR-1)}x0123456789ABCDEF);

	nc = 0;
	}
	$elif MR == 2:
	if (nc >= 16) {
	_mm_storeu_si128((__m128i*) c1, _mm256_extracti128_si256(vout01x0123456789ABCDEF, 1));
	_mm_storeu_si128((__m128i*) c0, _mm256_castsi256_si128(vout01x0123456789ABCDEF));

	$for M in reversed(range(MR)):
	c${M} = (int8_t*) ((uintptr_t) c${M} + cn_stride);

	a = (const int8_t**restrict) ((uintptr_t) a - ks);

	nc -= 16;
	} else {
	// Prepare mask for valid 8-bit elements (depends on nc).
	__mmask64 vmask = _cvtu64_mask64((uint64_t) ((UINT32_C(1) << (nc + 16)) - (UINT32_C(1) << 16)));

	_mm256_mask_storeu_epi8(c1 - 16, vmask, vout01x0123456789ABCDEF);
	vmask = _kshiftri_mask64(vmask, 16);
	_mm256_mask_storeu_epi8(c0, vmask, vout01x0123456789ABCDEF);

	nc = 0;
	}
	$elif MR == 1:
	if (nc >= 16) {
	_mm_storeu_si128((__m128i*) c0, vout0x0123456789ABCDEF);

	$for M in range(MR):
	c${M} = (int8_t*) ((uintptr_t) c${M} + cn_stride);

	a = (const int8_t**restrict) ((uintptr_t) a - ks);

	nc -= 16;
	} else {
	// Prepare mask for valid 8-bit elements (depends on nc).
	const __mmask64 vmask = _cvtu64_mask64((uint64_t) ((UINT32_C(1) << nc) - UINT32_C(1)));

	_mm_mask_storeu_epi8(c0, vmask, vout0x0123456789ABCDEF);

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