src/qs8-dwconv/unipass-avx2-mul16-vpmovsx.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 REQUANTIZATION in ["GEMMLOWP", "FP32"]
 $assert DATATYPE in ["QC8", "QS8"]
 $assert DATATYPE != "QC8" or REQUANTIZATION == "FP32"
 $assert CHANNEL_TILE % 16 == 0
 $assert CHANNEL_TILE >= 16
 $assert KERNEL_TILE >= 2
 #include <assert.h>

 #include <immintrin.h>

 #include <xnnpack/dwconv.h>


 $PARAMS_STRUCT = "avx2" if DATATYPE == "QC8" else REQUANTIZATION.lower() + "_avx2"
 $PARAMS_UNION = "xnn_qs8_minmax_params" if DATATYPE == "QC8" else "xnn_qs8_conv_minmax_params"
 void xnn_${DATATYPE.lower()}_dwconv_minmax_${REQUANTIZATION.lower()}_ukernel_up${CHANNEL_TILE}x${KERNEL_TILE}__avx2_mul16_vpmovsx(
     size_t channels,
     size_t output_width,
     const int8_t** input,
     const void* weights,
     int8_t* output,
     size_t input_stride,
     size_t output_increment,
     size_t input_offset,
     const int8_t* zero,
     const union ${PARAMS_UNION} params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
 {
   assert(channels != 0);
   assert(output_width != 0);

   do {
     $for K in range(KERNEL_TILE):
       const int8_t* i${K} = input[${K}];
       assert(i${K} != NULL);
       if XNN_UNPREDICTABLE(i${K} != zero) {
         i${K} = (const int8_t*) ((uintptr_t) i${K} + input_offset);
       }
     input = (const int8_t**) ((uintptr_t) input + input_stride);

     size_t c = channels;
     const void* w = weights;
     for (; c >= ${CHANNEL_TILE}; c -= ${CHANNEL_TILE}) {
       __m256i vacc${ABC[0:8]} = _mm256_loadu_si256((const __m256i*) w);
       $for C in range(8, CHANNEL_TILE, 8):
         __m256i vacc${ABC[C:C+8]} = _mm256_loadu_si256((const __m256i*) ((uintptr_t) w + ${C} * sizeof(int32_t)));

       $for K in range(KERNEL_TILE):

         $for C in range(0, CHANNEL_TILE, 16):
           $if C == 0:
             const __m256i vi${K}x${ABC[0:16]} = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i${K}));
           $else:
             const __m256i vi${K}x${ABC[C:C+16]} = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i${K} + ${C})));
           const __m256i vk${K}x${ABC[C:C+16]} = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + ${CHANNEL_TILE} * sizeof(int32_t) + ${K * CHANNEL_TILE + C} * sizeof(int8_t))));
         i${K} += ${CHANNEL_TILE};

         $for C in range(0, CHANNEL_TILE, 16):
           const __m256i vprod${K}x${ABC[C:C+16]} =  _mm256_mullo_epi16(vi${K}x${ABC[C:C+16]}, vk${K}x${ABC[C:C+16]});
           const __m128i vprod${K}x${ABC[C+8:C+16]} = _mm256_extracti128_si256(vprod${K}x${ABC[C:C+16]}, 1);
           vacc${ABC[C:C+8]} = _mm256_add_epi32(vacc${ABC[C:C+8]}, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod${K}x${ABC[C:C+16]})));
           vacc${ABC[C+8:C+16]} = _mm256_add_epi32(vacc${ABC[C+8:C+16]}, _mm256_cvtepi16_epi32(vprod${K}x${ABC[C+8:C+16]}));

       w = (const void*) ((uintptr_t) w + ${CHANNEL_TILE} * sizeof(int32_t) + ${KERNEL_TILE * CHANNEL_TILE} * sizeof(int8_t));
       $if REQUANTIZATION == "GEMMLOWP":
         const __m256i vmultiplier = _mm256_load_si256((const __m256i*) params->gemmlowp_avx2.multiplier);
         const __m256i vrounding = _mm256_load_si256((const __m256i*) params->gemmlowp_avx2.rounding);

         $for C in range(0, CHANNEL_TILE, 8):
           const __m256i vacc${ABC[C+1:C+8:2]} = _mm256_shuffle_epi32(vacc${ABC[C:C+8]}, _MM_SHUFFLE(3, 3, 1, 1));

         $for C in range(0, CHANNEL_TILE, 8):
           const __m256i vprod${ABC[C:C+8:2]} = _mm256_add_epi64(_mm256_mul_epi32(vacc${ABC[C:C+8]}, vmultiplier), vrounding);
           const __m256i vprod${ABC[C+1:C+8:2]} = _mm256_add_epi64(_mm256_mul_epi32(vacc${ABC[C+1:C+8:2]}, vmultiplier), vrounding);

         $for C in range(0, CHANNEL_TILE, 8):
           const __m256i vq31prod${ABC[C:C+8:2]} = _mm256_srli_epi64(vprod${ABC[C:C+8:2]}, 31);
           const __m256i vq31prod${ABC[C+1:C+8:2]} = _mm256_add_epi64(vprod${ABC[C+1:C+8:2]}, vprod${ABC[C+1:C+8:2]});

         $for C in range(0, CHANNEL_TILE, 8):
           const __m256i vq31prod${ABC[C:C+8]} = _mm256_blend_epi16(vq31prod${ABC[C:C+8:2]}, vq31prod${ABC[C+1:C+8:2]}, 0xCC);

         const __m256i vremainder_mask = _mm256_load_si256((const __m256i*) params->gemmlowp_avx2.remainder_mask);
         $for C in range(0, CHANNEL_TILE, 8):
           const __m256i vrem${ABC[C:C+8]} =
             _mm256_add_epi32(_mm256_and_si256(vq31prod${ABC[C:C+8]}, vremainder_mask), _mm256_cmpgt_epi32(_mm256_setzero_si256(), vq31prod${ABC[C:C+8]}));

         const __m256i vremainder_threshold = _mm256_load_si256((const __m256i*) params->gemmlowp_avx2.remainder_threshold);
         $if CHANNEL_TILE > 8:
           const __m128i vshift = _mm_loadl_epi64((const __m128i*) params->gemmlowp_avx2.shift);
         $else:
           const __m128i vshift = _mm_load_si128((const __m128i*) params->gemmlowp_avx2.shift);
         $for C in range(0, CHANNEL_TILE, 8):
           vacc${ABC[C:C+8]} =
             _mm256_sub_epi32(_mm256_sra_epi32(vq31prod${ABC[C:C+8]}, vshift), _mm256_cmpgt_epi32(vrem${ABC[C:C+8]}, vremainder_threshold));
       $elif REQUANTIZATION == "FP32":
         $for C in range(0, CHANNEL_TILE, 8):
           __m256 vfpacc${ABC[C:C+8]} = _mm256_cvtepi32_ps(vacc${ABC[C:C+8]});

         $if DATATYPE == "QC8":
           const __m256 vscale${ABC[0:8]} = _mm256_loadu_ps((const float*) w);
           $for C in range(8, CHANNEL_TILE, 8):
             const __m256 vscale${ABC[C:C+8]} = _mm256_loadu_ps((const float*) ((uintptr_t) w + ${C} * sizeof(float)));
           w = (const void*) ((uintptr_t) w + ${CHANNEL_TILE} * sizeof(float));
           $for C in range(0, CHANNEL_TILE, 8):
             vfpacc${ABC[C:C+8]} = _mm256_mul_ps(vfpacc${ABC[C:C+8]}, vscale${ABC[C:C+8]});
         $else:
           const __m256 vscale = _mm256_load_ps(params->fp32_avx2.scale);
           $for C in range(0, CHANNEL_TILE, 8):
             vfpacc${ABC[C:C+8]} = _mm256_mul_ps(vfpacc${ABC[C:C+8]}, vscale);

         $for C in range(0, CHANNEL_TILE, 8):
           vacc${ABC[C:C+8]} = _mm256_cvtps_epi32(vfpacc${ABC[C:C+8]});

       const __m256i voutput_zero_point = _mm256_load_si256((const __m256i*) params->${PARAMS_STRUCT}.output_zero_point);
       $for C in range(0, CHANNEL_TILE, 16):
         const __m256i vout${ABC[C:C+4]}${ABC[C+8:C+12]}${ABC[C+4:C+8]}${ABC[C+12:C+16]} = _mm256_adds_epi16(_mm256_packs_epi32(vacc${ABC[C:C+8]}, vacc${ABC[C+8:C+16]}), voutput_zero_point);

       $for C in range(0, CHANNEL_TILE, 16):
         __m128i vout${ABC[C:C+16]} = _mm_shuffle_epi32(_mm_packs_epi16(_mm256_castsi256_si128(vout${ABC[C:C+4]}${ABC[C+8:C+12]}${ABC[C+4:C+8]}${ABC[C+12:C+16]}), _mm256_extracti128_si256(vout${ABC[C:C+4]}${ABC[C+8:C+12]}${ABC[C+4:C+8]}${ABC[C+12:C+16]}, 1)), _MM_SHUFFLE(3, 1, 2, 0));

       const __m128i voutput_min = _mm_load_si128((const __m128i*) params->${PARAMS_STRUCT}.output_min);
       const __m128i voutput_max = _mm_load_si128((const __m128i*) params->${PARAMS_STRUCT}.output_max);
       $for C in range(0, CHANNEL_TILE, 16):
         vout${ABC[C:C+16]} = _mm_max_epi8(vout${ABC[C:C+16]}, voutput_min);
         vout${ABC[C:C+16]} = _mm_min_epi8(vout${ABC[C:C+16]}, voutput_max);

       _mm_storeu_si128((__m128i*) output, vout${ABC[0:16]});
       $for C in range(16, CHANNEL_TILE, 16):
         _mm_storeu_si128((__m128i*) (output + ${C}), vout${ABC[C:C+16]});
       output += ${CHANNEL_TILE};
     }
     if XNN_UNLIKELY(c != 0) {
       $if CHANNEL_TILE > 16:
         const int8_t* k = (const int8_t*) ((uintptr_t) w + ${CHANNEL_TILE} * sizeof(int32_t));
       ${"do " if CHANNEL_TILE > 16 else ""}{
         __m256i vacc${ABC[0:8]} = _mm256_loadu_si256((const __m256i*) w);
         __m256i vacc${ABC[8:16]} = _mm256_loadu_si256((const __m256i*) ((uintptr_t) w + 8 * sizeof(int32_t)));

         $for K in range(KERNEL_TILE):

           const __m256i vi${K}x${ABC[0:16]} = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i${K}));
           $if CHANNEL_TILE > 16:
             $if K == 0:
               const __m256i vk${K}x${ABC[0:16]} = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) k));
             $else:
               const __m256i vk${K}x${ABC[0:16]} = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + ${K * CHANNEL_TILE})));
           $else:
             const __m256i vk${K}x${ABC[0:16]} = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + ${CHANNEL_TILE} * sizeof(int32_t) + ${K * CHANNEL_TILE} * sizeof(int8_t))));
           $if CHANNEL_TILE > 16:
             i${K} += 16;

           const __m256i vprod${K}x${ABC[0:16]} = _mm256_mullo_epi16(vi${K}x${ABC[0:16]}, vk${K}x${ABC[0:16]});
           const __m128i vprod${K}x${ABC[8:16]} = _mm256_extracti128_si256(vprod${K}x${ABC[0:16]}, 1);
           vacc${ABC[0:8]} = _mm256_add_epi32(vacc${ABC[0:8]}, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod${K}x${ABC[0:16]})));
           vacc${ABC[8:16]} = _mm256_add_epi32(vacc${ABC[8:16]}, _mm256_cvtepi16_epi32(vprod${K}x${ABC[8:16]}));

         $if CHANNEL_TILE > 16:
           k += 16;

         $if REQUANTIZATION == "GEMMLOWP":
           const __m256i vmultiplier = _mm256_load_si256((const __m256i*) params->gemmlowp_avx2.multiplier);
           const __m256i vrounding = _mm256_load_si256((const __m256i*) params->gemmlowp_avx2.rounding);

           const __m256i vacc${ABC[1:8:2]} = _mm256_shuffle_epi32(vacc${ABC[0:8]}, _MM_SHUFFLE(3, 3, 1, 1));
           const __m256i vacc${ABC[9:16:2]} = _mm256_shuffle_epi32(vacc${ABC[8:16]}, _MM_SHUFFLE(3, 3, 1, 1));

           const __m256i vprod${ABC[0:8:2]} = _mm256_add_epi64(_mm256_mul_epi32(vacc${ABC[0:8]}, vmultiplier), vrounding);
           const __m256i vprod${ABC[1:8:2]} = _mm256_add_epi64(_mm256_mul_epi32(vacc${ABC[1:8:2]}, vmultiplier), vrounding);
           const __m256i vprod${ABC[8:16:2]} = _mm256_add_epi64(_mm256_mul_epi32(vacc${ABC[8:16]}, vmultiplier), vrounding);
           const __m256i vprod${ABC[9:16:2]} = _mm256_add_epi64(_mm256_mul_epi32(vacc${ABC[9:16:2]}, vmultiplier), vrounding);

           const __m256i vq31prod${ABC[0:8:2]} = _mm256_srli_epi64(vprod${ABC[0:8:2]}, 31);
           const __m256i vq31prod${ABC[1:8:2]} = _mm256_add_epi64(vprod${ABC[1:8:2]}, vprod${ABC[1:8:2]});
           const __m256i vq31prod${ABC[8:16:2]} = _mm256_srli_epi64(vprod${ABC[8:16:2]}, 31);
           const __m256i vq31prod${ABC[9:16:2]} = _mm256_add_epi64(vprod${ABC[9:16:2]}, vprod${ABC[9:16:2]});

           const __m256i vq31prod${ABC[0:8]} = _mm256_blend_epi16(vq31prod${ABC[0:8:2]}, vq31prod${ABC[1:8:2]}, 0xCC);
           const __m256i vq31prod${ABC[8:16]} = _mm256_blend_epi16(vq31prod${ABC[8:16:2]}, vq31prod${ABC[9:16:2]}, 0xCC);

           const __m256i vremainder_mask = _mm256_load_si256((const __m256i*) params->gemmlowp_avx2.remainder_mask);
           const __m256i vrem${ABC[0:8]} =
             _mm256_add_epi32(_mm256_and_si256(vq31prod${ABC[0:8]}, vremainder_mask), _mm256_cmpgt_epi32(_mm256_setzero_si256(), vq31prod${ABC[0:8]}));
           const __m256i vrem${ABC[8:16]} =
             _mm256_add_epi32(_mm256_and_si256(vq31prod${ABC[8:16]}, vremainder_mask), _mm256_cmpgt_epi32(_mm256_setzero_si256(), vq31prod${ABC[8:16]}));

           const __m256i vremainder_threshold = _mm256_load_si256((const __m256i*) params->gemmlowp_avx2.remainder_threshold);
           const __m128i vshift = _mm_loadl_epi64((const __m128i*) params->gemmlowp_avx2.shift);
           vacc${ABC[0:8]} =
             _mm256_sub_epi32(_mm256_sra_epi32(vq31prod${ABC[0:8]}, vshift), _mm256_cmpgt_epi32(vrem${ABC[0:8]}, vremainder_threshold));
           vacc${ABC[8:16]} =
             _mm256_sub_epi32(_mm256_sra_epi32(vq31prod${ABC[8:16]}, vshift), _mm256_cmpgt_epi32(vrem${ABC[8:16]}, vremainder_threshold));
         $elif REQUANTIZATION == "FP32":
           __m256 vfpacc${ABC[0:8]} = _mm256_cvtepi32_ps(vacc${ABC[0:8]});
           __m256 vfpacc${ABC[8:16]} = _mm256_cvtepi32_ps(vacc${ABC[8:16]});

           $if DATATYPE == "QC8":
             const __m256 vscale${ABC[0:8]} = _mm256_loadu_ps((const float*) ((uintptr_t) w + ${CHANNEL_TILE} * sizeof(int32_t) + ${CHANNEL_TILE * KERNEL_TILE} * sizeof(int8_t)));
             const __m256 vscale${ABC[8:16]} = _mm256_loadu_ps((const float*) ((uintptr_t) w + ${CHANNEL_TILE} * sizeof(int32_t) + ${CHANNEL_TILE * KERNEL_TILE} * sizeof(int8_t) + 8 * sizeof(float)));
             vfpacc${ABC[0:8]} = _mm256_mul_ps(vfpacc${ABC[0:8]}, vscale${ABC[0:8]});
             vfpacc${ABC[8:16]} = _mm256_mul_ps(vfpacc${ABC[8:16]}, vscale${ABC[8:16]});
           $else:
             const __m256 vscale = _mm256_load_ps(params->fp32_avx2.scale);
             vfpacc${ABC[0:8]} = _mm256_mul_ps(vfpacc${ABC[0:8]}, vscale);
             vfpacc${ABC[8:16]} = _mm256_mul_ps(vfpacc${ABC[8:16]}, vscale);

           vacc${ABC[0:8]} = _mm256_cvtps_epi32(vfpacc${ABC[0:8]});
           vacc${ABC[8:16]} = _mm256_cvtps_epi32(vfpacc${ABC[8:16]});

         $if CHANNEL_TILE > 16:
           w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t));

         const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->${PARAMS_STRUCT}.output_zero_point);
         __m128i vout${ABC[0:8]} = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc${ABC[0:8]}), _mm256_extracti128_si256(vacc${ABC[0:8]}, 1)), voutput_zero_point);
         __m128i vout${ABC[8:16]} = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc${ABC[8:16]}), _mm256_extracti128_si256(vacc${ABC[8:16]}, 1)), voutput_zero_point);

         const __m128i voutput_min = _mm_load_si128((const __m128i*) params->${PARAMS_STRUCT}.output_min);
         const __m128i voutput_max = _mm_load_si128((const __m128i*) params->${PARAMS_STRUCT}.output_max);

         __m128i vout${ABC[0:16]} = _mm_packs_epi16(vout${ABC[0:8]}, vout${ABC[8:16]});
         vout${ABC[0:16]} = _mm_min_epi8(_mm_max_epi8(vout${ABC[0:16]}, voutput_min), voutput_max);

         $if CHANNEL_TILE > 16:
           if XNN_LIKELY(c >= 16) {
             _mm_storeu_si128((__m128i*) output, vout${ABC[0:16]});
             output += 16;
             c -= 16;
           } else {
             if (c & 8) {
               _mm_storel_epi64((__m128i*) output, vout${ABC[0:16]});
               vout${ABC[0:16]} = _mm_unpackhi_epi64(vout${ABC[0:16]}, vout${ABC[0:16]});
               output += 8;
             }
             if (c & 4) {
               *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout${ABC[0:16]});
               vout${ABC[0:16]} = _mm_srli_epi64(vout${ABC[0:16]}, 32);
               output += 4;
             }
             if (c & 2) {
               *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout${ABC[0:16]}, 0);
               vout${ABC[0:16]} = _mm_srli_epi32(vout${ABC[0:16]}, 16);
               output += 2;
             }
             if (c & 1) {
               *output = (int8_t) _mm_extract_epi8(vout${ABC[0:16]}, 0);
               output += 1;
             }
             c = 0;
           }
         $else:
           if (c & 8) {
             _mm_storel_epi64((__m128i*) output, vout${ABC[0:16]});
             vout${ABC[0:16]} = _mm_unpackhi_epi64(vout${ABC[0:16]}, vout${ABC[0:16]});
             output += 8;
           }
           if (c & 4) {
             *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout${ABC[0:16]});
             vout${ABC[0:16]} = _mm_srli_epi64(vout${ABC[0:16]}, 32);
             output += 4;
           }
           if (c & 2) {
             *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout${ABC[0:16]}, 0);
             vout${ABC[0:16]} = _mm_srli_epi32(vout${ABC[0:16]}, 16);
             output += 2;
           }
           if (c & 1) {
             *output = (int8_t) _mm_extract_epi8(vout${ABC[0:16]}, 0);
             output += 1;
           }
       }${" while (c != 0);" if CHANNEL_TILE > 16 else ""}
     }

     output = (int8_t*) ((uintptr_t) output + output_increment);
   } while (--output_width != 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 REQUANTIZATION in ["GEMMLOWP", "FP32"]
	$assert DATATYPE in ["QC8", "QS8"]
	$assert DATATYPE != "QC8" or REQUANTIZATION == "FP32"
	$assert CHANNEL_TILE % 16 == 0
	$assert CHANNEL_TILE >= 16
	$assert KERNEL_TILE >= 2
	#include <assert.h>

	#include <immintrin.h>

	#include <xnnpack/dwconv.h>


	$PARAMS_STRUCT = "avx2" if DATATYPE == "QC8" else REQUANTIZATION.lower() + "_avx2"
	$PARAMS_UNION = "xnn_qs8_minmax_params" if DATATYPE == "QC8" else "xnn_qs8_conv_minmax_params"
	void xnn_${DATATYPE.lower()}_dwconv_minmax_${REQUANTIZATION.lower()}_ukernel_up${CHANNEL_TILE}x${KERNEL_TILE}__avx2_mul16_vpmovsx(
	size_t channels,
	size_t output_width,
	const int8_t** input,
	const void* weights,
	int8_t* output,
	size_t input_stride,
	size_t output_increment,
	size_t input_offset,
	const int8_t* zero,
	const union ${PARAMS_UNION} params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
	{
	assert(channels != 0);
	assert(output_width != 0);

	do {
	$for K in range(KERNEL_TILE):
	const int8_t* i${K} = input[${K}];
	assert(i${K} != NULL);
	if XNN_UNPREDICTABLE(i${K} != zero) {
	i${K} = (const int8_t*) ((uintptr_t) i${K} + input_offset);
	}
	input = (const int8_t**) ((uintptr_t) input + input_stride);

	size_t c = channels;
	const void* w = weights;
	for (; c >= ${CHANNEL_TILE}; c -= ${CHANNEL_TILE}) {
	__m256i vacc${ABC[0:8]} = _mm256_loadu_si256((const __m256i*) w);
	$for C in range(8, CHANNEL_TILE, 8):
	__m256i vacc${ABC[C:C+8]} = _mm256_loadu_si256((const __m256i) ((uintptr_t) w + ${C} sizeof(int32_t)));

	$for K in range(KERNEL_TILE):

	$for C in range(0, CHANNEL_TILE, 16):
	$if C == 0:
	const __m256i vi${K}x${ABC[0:16]} = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i${K}));
	$else:
	const __m256i vi${K}x${ABC[C:C+16]} = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i${K} + ${C})));
	const __m256i vk${K}x${ABC[C:C+16]} = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i) ((uintptr_t) w + ${CHANNEL_TILE} sizeof(int32_t) + ${K * CHANNEL_TILE + C} * sizeof(int8_t))));
	i${K} += ${CHANNEL_TILE};

	$for C in range(0, CHANNEL_TILE, 16):
	const __m256i vprod${K}x${ABC[C:C+16]} = _mm256_mullo_epi16(vi${K}x${ABC[C:C+16]}, vk${K}x${ABC[C:C+16]});
	const __m128i vprod${K}x${ABC[C+8:C+16]} = _mm256_extracti128_si256(vprod${K}x${ABC[C:C+16]}, 1);
	vacc${ABC[C:C+8]} = _mm256_add_epi32(vacc${ABC[C:C+8]}, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod${K}x${ABC[C:C+16]})));
	vacc${ABC[C+8:C+16]} = _mm256_add_epi32(vacc${ABC[C+8:C+16]}, _mm256_cvtepi16_epi32(vprod${K}x${ABC[C+8:C+16]}));

	w = (const void) ((uintptr_t) w + ${CHANNEL_TILE} sizeof(int32_t) + ${KERNEL_TILE * CHANNEL_TILE} * sizeof(int8_t));
	$if REQUANTIZATION == "GEMMLOWP":
	const __m256i vmultiplier = _mm256_load_si256((const __m256i*) params->gemmlowp_avx2.multiplier);
	const __m256i vrounding = _mm256_load_si256((const __m256i*) params->gemmlowp_avx2.rounding);

	$for C in range(0, CHANNEL_TILE, 8):
	const __m256i vacc${ABC[C+1:C+8:2]} = _mm256_shuffle_epi32(vacc${ABC[C:C+8]}, _MM_SHUFFLE(3, 3, 1, 1));

	$for C in range(0, CHANNEL_TILE, 8):
	const __m256i vprod${ABC[C:C+8:2]} = _mm256_add_epi64(_mm256_mul_epi32(vacc${ABC[C:C+8]}, vmultiplier), vrounding);
	const __m256i vprod${ABC[C+1:C+8:2]} = _mm256_add_epi64(_mm256_mul_epi32(vacc${ABC[C+1:C+8:2]}, vmultiplier), vrounding);

	$for C in range(0, CHANNEL_TILE, 8):
	const __m256i vq31prod${ABC[C:C+8:2]} = _mm256_srli_epi64(vprod${ABC[C:C+8:2]}, 31);
	const __m256i vq31prod${ABC[C+1:C+8:2]} = _mm256_add_epi64(vprod${ABC[C+1:C+8:2]}, vprod${ABC[C+1:C+8:2]});

	$for C in range(0, CHANNEL_TILE, 8):
	const __m256i vq31prod${ABC[C:C+8]} = _mm256_blend_epi16(vq31prod${ABC[C:C+8:2]}, vq31prod${ABC[C+1:C+8:2]}, 0xCC);

	const __m256i vremainder_mask = _mm256_load_si256((const __m256i*) params->gemmlowp_avx2.remainder_mask);
	$for C in range(0, CHANNEL_TILE, 8):
	const __m256i vrem${ABC[C:C+8]} =
	_mm256_add_epi32(_mm256_and_si256(vq31prod${ABC[C:C+8]}, vremainder_mask), _mm256_cmpgt_epi32(_mm256_setzero_si256(), vq31prod${ABC[C:C+8]}));

	const __m256i vremainder_threshold = _mm256_load_si256((const __m256i*) params->gemmlowp_avx2.remainder_threshold);
	$if CHANNEL_TILE > 8:
	const __m128i vshift = _mm_loadl_epi64((const __m128i*) params->gemmlowp_avx2.shift);
	$else:
	const __m128i vshift = _mm_load_si128((const __m128i*) params->gemmlowp_avx2.shift);
	$for C in range(0, CHANNEL_TILE, 8):
	vacc${ABC[C:C+8]} =
	_mm256_sub_epi32(_mm256_sra_epi32(vq31prod${ABC[C:C+8]}, vshift), _mm256_cmpgt_epi32(vrem${ABC[C:C+8]}, vremainder_threshold));
	$elif REQUANTIZATION == "FP32":
	$for C in range(0, CHANNEL_TILE, 8):
	__m256 vfpacc${ABC[C:C+8]} = _mm256_cvtepi32_ps(vacc${ABC[C:C+8]});

	$if DATATYPE == "QC8":
	const __m256 vscale${ABC[0:8]} = _mm256_loadu_ps((const float*) w);
	$for C in range(8, CHANNEL_TILE, 8):
	const __m256 vscale${ABC[C:C+8]} = _mm256_loadu_ps((const float) ((uintptr_t) w + ${C} sizeof(float)));
	w = (const void) ((uintptr_t) w + ${CHANNEL_TILE} sizeof(float));
	$for C in range(0, CHANNEL_TILE, 8):
	vfpacc${ABC[C:C+8]} = _mm256_mul_ps(vfpacc${ABC[C:C+8]}, vscale${ABC[C:C+8]});
	$else:
	const __m256 vscale = _mm256_load_ps(params->fp32_avx2.scale);
	$for C in range(0, CHANNEL_TILE, 8):
	vfpacc${ABC[C:C+8]} = _mm256_mul_ps(vfpacc${ABC[C:C+8]}, vscale);

	$for C in range(0, CHANNEL_TILE, 8):
	vacc${ABC[C:C+8]} = _mm256_cvtps_epi32(vfpacc${ABC[C:C+8]});

	const __m256i voutput_zero_point = _mm256_load_si256((const __m256i*) params->${PARAMS_STRUCT}.output_zero_point);
	$for C in range(0, CHANNEL_TILE, 16):
	const __m256i vout${ABC[C:C+4]}${ABC[C+8:C+12]}${ABC[C+4:C+8]}${ABC[C+12:C+16]} = _mm256_adds_epi16(_mm256_packs_epi32(vacc${ABC[C:C+8]}, vacc${ABC[C+8:C+16]}), voutput_zero_point);

	$for C in range(0, CHANNEL_TILE, 16):
	__m128i vout${ABC[C:C+16]} = _mm_shuffle_epi32(_mm_packs_epi16(_mm256_castsi256_si128(vout${ABC[C:C+4]}${ABC[C+8:C+12]}${ABC[C+4:C+8]}${ABC[C+12:C+16]}), _mm256_extracti128_si256(vout${ABC[C:C+4]}${ABC[C+8:C+12]}${ABC[C+4:C+8]}${ABC[C+12:C+16]}, 1)), _MM_SHUFFLE(3, 1, 2, 0));

	const __m128i voutput_min = _mm_load_si128((const __m128i*) params->${PARAMS_STRUCT}.output_min);
	const __m128i voutput_max = _mm_load_si128((const __m128i*) params->${PARAMS_STRUCT}.output_max);
	$for C in range(0, CHANNEL_TILE, 16):
	vout${ABC[C:C+16]} = _mm_max_epi8(vout${ABC[C:C+16]}, voutput_min);
	vout${ABC[C:C+16]} = _mm_min_epi8(vout${ABC[C:C+16]}, voutput_max);

	_mm_storeu_si128((__m128i*) output, vout${ABC[0:16]});
	$for C in range(16, CHANNEL_TILE, 16):
	_mm_storeu_si128((__m128i*) (output + ${C}), vout${ABC[C:C+16]});
	output += ${CHANNEL_TILE};
	}
	if XNN_UNLIKELY(c != 0) {
	$if CHANNEL_TILE > 16:
	const int8_t* k = (const int8_t) ((uintptr_t) w + ${CHANNEL_TILE} sizeof(int32_t));
	${"do " if CHANNEL_TILE > 16 else ""}{
	__m256i vacc${ABC[0:8]} = _mm256_loadu_si256((const __m256i*) w);
	__m256i vacc${ABC[8:16]} = _mm256_loadu_si256((const __m256i) ((uintptr_t) w + 8 sizeof(int32_t)));

	$for K in range(KERNEL_TILE):

	const __m256i vi${K}x${ABC[0:16]} = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i${K}));
	$if CHANNEL_TILE > 16:
	$if K == 0:
	const __m256i vk${K}x${ABC[0:16]} = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) k));
	$else:
	const __m256i vk${K}x${ABC[0:16]} = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i) (k + ${K CHANNEL_TILE})));
	$else:
	const __m256i vk${K}x${ABC[0:16]} = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i) ((uintptr_t) w + ${CHANNEL_TILE} sizeof(int32_t) + ${K * CHANNEL_TILE} * sizeof(int8_t))));
	$if CHANNEL_TILE > 16:
	i${K} += 16;

	const __m256i vprod${K}x${ABC[0:16]} = _mm256_mullo_epi16(vi${K}x${ABC[0:16]}, vk${K}x${ABC[0:16]});
	const __m128i vprod${K}x${ABC[8:16]} = _mm256_extracti128_si256(vprod${K}x${ABC[0:16]}, 1);
	vacc${ABC[0:8]} = _mm256_add_epi32(vacc${ABC[0:8]}, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod${K}x${ABC[0:16]})));
	vacc${ABC[8:16]} = _mm256_add_epi32(vacc${ABC[8:16]}, _mm256_cvtepi16_epi32(vprod${K}x${ABC[8:16]}));

	$if CHANNEL_TILE > 16:
	k += 16;

	$if REQUANTIZATION == "GEMMLOWP":
	const __m256i vmultiplier = _mm256_load_si256((const __m256i*) params->gemmlowp_avx2.multiplier);
	const __m256i vrounding = _mm256_load_si256((const __m256i*) params->gemmlowp_avx2.rounding);

	const __m256i vacc${ABC[1:8:2]} = _mm256_shuffle_epi32(vacc${ABC[0:8]}, _MM_SHUFFLE(3, 3, 1, 1));
	const __m256i vacc${ABC[9:16:2]} = _mm256_shuffle_epi32(vacc${ABC[8:16]}, _MM_SHUFFLE(3, 3, 1, 1));

	const __m256i vprod${ABC[0:8:2]} = _mm256_add_epi64(_mm256_mul_epi32(vacc${ABC[0:8]}, vmultiplier), vrounding);
	const __m256i vprod${ABC[1:8:2]} = _mm256_add_epi64(_mm256_mul_epi32(vacc${ABC[1:8:2]}, vmultiplier), vrounding);
	const __m256i vprod${ABC[8:16:2]} = _mm256_add_epi64(_mm256_mul_epi32(vacc${ABC[8:16]}, vmultiplier), vrounding);
	const __m256i vprod${ABC[9:16:2]} = _mm256_add_epi64(_mm256_mul_epi32(vacc${ABC[9:16:2]}, vmultiplier), vrounding);

	const __m256i vq31prod${ABC[0:8:2]} = _mm256_srli_epi64(vprod${ABC[0:8:2]}, 31);
	const __m256i vq31prod${ABC[1:8:2]} = _mm256_add_epi64(vprod${ABC[1:8:2]}, vprod${ABC[1:8:2]});
	const __m256i vq31prod${ABC[8:16:2]} = _mm256_srli_epi64(vprod${ABC[8:16:2]}, 31);
	const __m256i vq31prod${ABC[9:16:2]} = _mm256_add_epi64(vprod${ABC[9:16:2]}, vprod${ABC[9:16:2]});

	const __m256i vq31prod${ABC[0:8]} = _mm256_blend_epi16(vq31prod${ABC[0:8:2]}, vq31prod${ABC[1:8:2]}, 0xCC);
	const __m256i vq31prod${ABC[8:16]} = _mm256_blend_epi16(vq31prod${ABC[8:16:2]}, vq31prod${ABC[9:16:2]}, 0xCC);

	const __m256i vremainder_mask = _mm256_load_si256((const __m256i*) params->gemmlowp_avx2.remainder_mask);
	const __m256i vrem${ABC[0:8]} =
	_mm256_add_epi32(_mm256_and_si256(vq31prod${ABC[0:8]}, vremainder_mask), _mm256_cmpgt_epi32(_mm256_setzero_si256(), vq31prod${ABC[0:8]}));
	const __m256i vrem${ABC[8:16]} =
	_mm256_add_epi32(_mm256_and_si256(vq31prod${ABC[8:16]}, vremainder_mask), _mm256_cmpgt_epi32(_mm256_setzero_si256(), vq31prod${ABC[8:16]}));

	const __m256i vremainder_threshold = _mm256_load_si256((const __m256i*) params->gemmlowp_avx2.remainder_threshold);
	const __m128i vshift = _mm_loadl_epi64((const __m128i*) params->gemmlowp_avx2.shift);
	vacc${ABC[0:8]} =
	_mm256_sub_epi32(_mm256_sra_epi32(vq31prod${ABC[0:8]}, vshift), _mm256_cmpgt_epi32(vrem${ABC[0:8]}, vremainder_threshold));
	vacc${ABC[8:16]} =
	_mm256_sub_epi32(_mm256_sra_epi32(vq31prod${ABC[8:16]}, vshift), _mm256_cmpgt_epi32(vrem${ABC[8:16]}, vremainder_threshold));
	$elif REQUANTIZATION == "FP32":
	__m256 vfpacc${ABC[0:8]} = _mm256_cvtepi32_ps(vacc${ABC[0:8]});
	__m256 vfpacc${ABC[8:16]} = _mm256_cvtepi32_ps(vacc${ABC[8:16]});

	$if DATATYPE == "QC8":
	const __m256 vscale${ABC[0:8]} = _mm256_loadu_ps((const float) ((uintptr_t) w + ${CHANNEL_TILE} sizeof(int32_t) + ${CHANNEL_TILE * KERNEL_TILE} * sizeof(int8_t)));
	const __m256 vscale${ABC[8:16]} = _mm256_loadu_ps((const float) ((uintptr_t) w + ${CHANNEL_TILE} sizeof(int32_t) + ${CHANNEL_TILE * KERNEL_TILE} * sizeof(int8_t) + 8 * sizeof(float)));
	vfpacc${ABC[0:8]} = _mm256_mul_ps(vfpacc${ABC[0:8]}, vscale${ABC[0:8]});
	vfpacc${ABC[8:16]} = _mm256_mul_ps(vfpacc${ABC[8:16]}, vscale${ABC[8:16]});
	$else:
	const __m256 vscale = _mm256_load_ps(params->fp32_avx2.scale);
	vfpacc${ABC[0:8]} = _mm256_mul_ps(vfpacc${ABC[0:8]}, vscale);
	vfpacc${ABC[8:16]} = _mm256_mul_ps(vfpacc${ABC[8:16]}, vscale);

	vacc${ABC[0:8]} = _mm256_cvtps_epi32(vfpacc${ABC[0:8]});
	vacc${ABC[8:16]} = _mm256_cvtps_epi32(vfpacc${ABC[8:16]});

	$if CHANNEL_TILE > 16:
	w = (const void) ((uintptr_t) w + 16 sizeof(int32_t));

	const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->${PARAMS_STRUCT}.output_zero_point);
	__m128i vout${ABC[0:8]} = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc${ABC[0:8]}), _mm256_extracti128_si256(vacc${ABC[0:8]}, 1)), voutput_zero_point);
	__m128i vout${ABC[8:16]} = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc${ABC[8:16]}), _mm256_extracti128_si256(vacc${ABC[8:16]}, 1)), voutput_zero_point);

	const __m128i voutput_min = _mm_load_si128((const __m128i*) params->${PARAMS_STRUCT}.output_min);
	const __m128i voutput_max = _mm_load_si128((const __m128i*) params->${PARAMS_STRUCT}.output_max);

	__m128i vout${ABC[0:16]} = _mm_packs_epi16(vout${ABC[0:8]}, vout${ABC[8:16]});
	vout${ABC[0:16]} = _mm_min_epi8(_mm_max_epi8(vout${ABC[0:16]}, voutput_min), voutput_max);

	$if CHANNEL_TILE > 16:
	if XNN_LIKELY(c >= 16) {
	_mm_storeu_si128((__m128i*) output, vout${ABC[0:16]});
	output += 16;
	c -= 16;
	} else {
	if (c & 8) {
	_mm_storel_epi64((__m128i*) output, vout${ABC[0:16]});
	vout${ABC[0:16]} = _mm_unpackhi_epi64(vout${ABC[0:16]}, vout${ABC[0:16]});
	output += 8;
	}
	if (c & 4) {
	((uint32_t) output) = (uint32_t) _mm_cvtsi128_si32(vout${ABC[0:16]});
	vout${ABC[0:16]} = _mm_srli_epi64(vout${ABC[0:16]}, 32);
	output += 4;
	}
	if (c & 2) {
	((uint16_t) output) = (uint16_t) _mm_extract_epi16(vout${ABC[0:16]}, 0);
	vout${ABC[0:16]} = _mm_srli_epi32(vout${ABC[0:16]}, 16);
	output += 2;
	}
	if (c & 1) {
	*output = (int8_t) _mm_extract_epi8(vout${ABC[0:16]}, 0);
	output += 1;
	}
	c = 0;
	}
	$else:
	if (c & 8) {
	_mm_storel_epi64((__m128i*) output, vout${ABC[0:16]});
	vout${ABC[0:16]} = _mm_unpackhi_epi64(vout${ABC[0:16]}, vout${ABC[0:16]});
	output += 8;
	}
	if (c & 4) {
	((uint32_t) output) = (uint32_t) _mm_cvtsi128_si32(vout${ABC[0:16]});
	vout${ABC[0:16]} = _mm_srli_epi64(vout${ABC[0:16]}, 32);
	output += 4;
	}
	if (c & 2) {
	((uint16_t) output) = (uint16_t) _mm_extract_epi16(vout${ABC[0:16]}, 0);
	vout${ABC[0:16]} = _mm_srli_epi32(vout${ABC[0:16]}, 16);
	output += 2;
	}
	if (c & 1) {
	*output = (int8_t) _mm_extract_epi8(vout${ABC[0:16]}, 0);
	output += 1;
	}
	}${" while (c != 0);" if CHANNEL_TILE > 16 else ""}
	}

	output = (int8_t*) ((uintptr_t) output + output_increment);
	} while (--output_width != 0);
	}