src/qs8-dwconv/unipass-neon-mul16.c.in - platform/external/XNNPACK - Git at Google

 // 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.

 $assert REQUANTIZATION in ["FP32", "RNDNU"]
 $assert DATATYPE in ["QC8", "QS8", "QU8"]
 $assert DATATYPE != "QC8" or REQUANTIZATION == "FP32"
 $ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
 $assert CHANNEL_TILE % 8 == 0
 $assert CHANNEL_TILE >= 8
 $assert KERNEL_TILE >= 2
 #include <assert.h>

 #include <arm_neon.h>

 #include <xnnpack/dwconv.h>
 $if REQUANTIZATION == "FP32" and ARMV8:
   #include <xnnpack/intrinsics-polyfill.h>


 $PARAMS_STRUCT = REQUANTIZATION.lower() + "_" + ("neonv8" if ARMV8 else "neon")
 $PARAMS_UNION = "xnn_%s_conv_minmax_params" % DATATYPE.lower()
 $XINT8_T = "uint8_t" if DATATYPE == "QU8" else "int8_t"
 $XINT8X8_T = "uint8x8_t" if DATATYPE == "QU8" else "int8x8_t"
 $XINT8X16_T = "uint8x16_t" if DATATYPE == "QU8" else "int8x16_t"
 $VGET_LOW_X8 = "vget_low_u8" if DATATYPE == "QU8" else "vget_low_s8"
 $VCOMBINE_X8 = "vcombine_u8" if DATATYPE == "QU8" else "vcombine_s8"
 $VREINTERPRET_U32_X8 = "vreinterpret_u32_u8" if DATATYPE == "QU8" else "vreinterpret_u32_s8"
 $VREINTERPRET_U16_X8 = "vreinterpret_u16_u8" if DATATYPE == "QU8" else "vreinterpret_u16_s8"
 $VLD1_X8 = "vld1_u8" if DATATYPE == "QU8" else "vld1_s8"
 $VLD1_DUP_X8 = "vld1_dup_u8" if DATATYPE == "QU8" else "vld1_dup_s8"
 $VLD1Q_DUP_X8 = "vld1q_dup_u8" if DATATYPE == "QU8" else "vld1q_dup_s8"
 $VST1_X8 = "vst1_u8" if DATATYPE == "QU8" else "vst1_s8"
 $VST1Q_X8 = "vst1q_u8" if DATATYPE == "QU8" else "vst1q_s8"
 $VST1_LANE_X8 = "vst1_lane_u8" if DATATYPE == "QU8" else "vst1_lane_s8"
 $VST1Q_LANE_X8 = "vst1q_lane_u8" if DATATYPE == "QU8" else "vst1q_lane_s8"
 $VMIN_X8 = "vmin_u8" if DATATYPE == "QU8" else "vmin_s8"
 $VMAX_X8 = "vmax_u8" if DATATYPE == "QU8" else "vmax_s8"
 $VMINQ_X8 = "vminq_u8" if DATATYPE == "QU8" else "vminq_s8"
 $VMAXQ_X8 = "vmaxq_u8" if DATATYPE == "QU8" else "vmaxq_s8"
 $VEXT_X8 = "vext_u8" if DATATYPE == "QU8" else "vext_s8"
 $VQMOVXN_S16 = "vqmovun_s16" if DATATYPE == "QU8" else "vqmovn_s16"
 $VQMOVXN_HIGH_S16 = "vqmovun_high_s16" if DATATYPE == "QU8" else "vqmovn_high_s16"
 $ISA = "neonv8" if ARMV8 else "neon"
 void xnn_${DATATYPE.lower()}_dwconv_minmax_${REQUANTIZATION.lower()}_ukernel_up${CHANNEL_TILE}x${KERNEL_TILE}__${ISA}_mul16(
     size_t channels,
     size_t output_width,
     const ${XINT8_T}** input,
     const void* weights,
     ${XINT8_T}* output,
     size_t input_stride,
     size_t output_increment,
     size_t input_offset,
     const ${XINT8_T}* zero,
     const union ${PARAMS_UNION} params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
 {
   assert(channels != 0);
   assert(output_width != 0);

   $if DATATYPE == "QU8":
     const uint8x8_t vkernel_zero_point = vld1_dup_u8(params->${PARAMS_STRUCT}.kernel_zero_point);
   $if REQUANTIZATION == "RNDNU":
     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->${PARAMS_STRUCT}.right_pre_shift);
     const int32x4_t vmultiplier = vld1q_dup_s32(&params->${PARAMS_STRUCT}.multiplier);
     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->${PARAMS_STRUCT}.right_post_shift);
   $elif REQUANTIZATION == "FP32":
     $if DATATYPE != "QC8":
       const float32x4_t vscale = vld1q_dup_f32(&params->${PARAMS_STRUCT}.scale);
     $if not ARMV8:
       const float32x4_t vmagic_bias = vld1q_dup_f32(&params->${PARAMS_STRUCT}.magic_bias);
       const int32x4_t vmagic_bias_less_output_zero_point = vld1q_dup_s32(&params->${PARAMS_STRUCT}.magic_bias_less_output_zero_point);
   $if REQUANTIZATION != "FP32" or ARMV8:
     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->${PARAMS_STRUCT}.output_zero_point);
   $if CHANNEL_TILE == 8:
     const ${XINT8X8_T} voutput_min = ${VLD1_DUP_X8}(&params->${PARAMS_STRUCT}.output_min);
     const ${XINT8X8_T} voutput_max = ${VLD1_DUP_X8}(&params->${PARAMS_STRUCT}.output_max);
   $else:
     const ${XINT8X16_T} voutput_min = ${VLD1Q_DUP_X8}(&params->${PARAMS_STRUCT}.output_min);
     const ${XINT8X16_T} voutput_max = ${VLD1Q_DUP_X8}(&params->${PARAMS_STRUCT}.output_max);
   do {
     $for K in range(KERNEL_TILE):
       const ${XINT8_T}* i${K} = input[${K}];
       assert(i${K} != NULL);
       if XNN_UNPREDICTABLE(i${K} != zero) {
         i${K} = (const ${XINT8_T}*) ((uintptr_t) i${K} + input_offset);
       }
     input = (const ${XINT8_T}**) ((uintptr_t) input + input_stride);

     size_t c = channels;
     const void* w = weights;
     for (; c >= ${CHANNEL_TILE}; c -= ${CHANNEL_TILE}) {
       $for C in range(0, CHANNEL_TILE, 4):
         int32x4_t vacc${ABC[C:C+4]} = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);

       $for K in range(KERNEL_TILE):

         $for C in range(0, CHANNEL_TILE, 8):
           $if DATATYPE == "QU8":
             const int16x8_t vi${K}x${ABC[C:C+8]} = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i${K}))); i${K} += 8;
             const int16x8_t vk${K}x${ABC[C:C+8]} = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const ${XINT8_T}*) w + 8);
           $else:
             const int16x8_t vi${K}x${ABC[C:C+8]} = vmovl_s8(vld1_s8(i${K})); i${K} += 8;
             const int16x8_t vk${K}x${ABC[C:C+8]} = vmovl_s8(vld1_s8(w)); w = (const void*) ((const ${XINT8_T}*) w + 8);

         $for C in range(0, CHANNEL_TILE, 8):
           vacc${ABC[C:C+4]} = vmlal_s16(vacc${ABC[C:C+4]}, vget_low_s16(vi${K}x${ABC[C:C+8]}), vget_low_s16(vk${K}x${ABC[C:C+8]}));
           vacc${ABC[C+4:C+8]} = vmlal_s16(vacc${ABC[C+4:C+8]}, vget_high_s16(vi${K}x${ABC[C:C+8]}), vget_high_s16(vk${K}x${ABC[C:C+8]}));

       $if REQUANTIZATION == "RNDNU":
         $for C in range(0, CHANNEL_TILE, 4):
           vacc${ABC[C:C+4]} = vqshlq_s32(vacc${ABC[C:C+4]}, vright_pre_shift);

         $for C in range(0, CHANNEL_TILE, 4):
           vacc${ABC[C:C+4]} = vqdmulhq_s32(vacc${ABC[C:C+4]}, vmultiplier);

         $for C in range(0, CHANNEL_TILE, 4):
           vacc${ABC[C:C+4]} = vrshlq_s32(vacc${ABC[C:C+4]}, vright_post_shift);
       $elif REQUANTIZATION == "FP32":
         $for C in range(0, CHANNEL_TILE, 4):
           float32x4_t vfpacc${ABC[C:C+4]} = vcvtq_f32_s32(vacc${ABC[C:C+4]});

         $if DATATYPE == "QC8":
           $for C in range(0, CHANNEL_TILE, 4):
             const float32x4_t vscale${ABC[C:C+4]} = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);

           $for C in range(0, CHANNEL_TILE, 4):
             vfpacc${ABC[C:C+4]} = vmulq_f32(vfpacc${ABC[C:C+4]}, vscale${ABC[C:C+4]});
         $else:
           $for C in range(0, CHANNEL_TILE, 4):
             vfpacc${ABC[C:C+4]} = vmulq_f32(vfpacc${ABC[C:C+4]}, vscale);

         $if ARMV8:
           $for C in range(0, CHANNEL_TILE, 4):
             vacc${ABC[C:C+4]} = vcvtnq_s32_f32(vfpacc${ABC[C:C+4]});
         $else:
           $for C in range(0, CHANNEL_TILE, 4):
             vacc${ABC[C:C+4]} = vreinterpretq_s32_f32(vaddq_f32(vfpacc${ABC[C:C+4]}, vmagic_bias));

           $for C in range(0, CHANNEL_TILE, 4):
             vacc${ABC[C:C+4]} = vqsubq_s32(vacc${ABC[C:C+4]}, vmagic_bias_less_output_zero_point);

 #if XNN_ARCH_ARM64
       $for C in range(0, CHANNEL_TILE, 8):
         int16x8_t vacc${ABC[C:C+8]} = vqmovn_high_s32(vqmovn_s32(vacc${ABC[C:C+4]}), vacc${ABC[C+4:C+8]});

       $if REQUANTIZATION != "FP32" or ARMV8:
         $for C in range(0, CHANNEL_TILE, 8):
           vacc${ABC[C:C+8]} = vqaddq_s16(vacc${ABC[C:C+8]}, voutput_zero_point);

       $for C in range(0, CHANNEL_TILE, 16):
         $if C + 8 < CHANNEL_TILE:
           ${XINT8X16_T} vout${ABC[C:C+16]} = ${VQMOVXN_HIGH_S16}(${VQMOVXN_S16}(vacc${ABC[C:C+8]}), vacc${ABC[C+8:C+16]});
         $else:
           ${XINT8X8_T} vout${ABC[C:C+8]} = ${VQMOVXN_S16}(vacc${ABC[C:C+8]});
 #else  // !XNN_ARCH_ARM64
       $for C in range(0, CHANNEL_TILE, 8):
         int16x8_t vacc${ABC[C:C+8]} = vcombine_s16(vqmovn_s32(vacc${ABC[C:C+4]}), vqmovn_s32(vacc${ABC[C+4:C+8]}));

       $if REQUANTIZATION != "FP32" or ARMV8:
         $for C in range(0, CHANNEL_TILE, 8):
           vacc${ABC[C:C+8]} = vqaddq_s16(vacc${ABC[C:C+8]}, voutput_zero_point);

       $for C in range(0, CHANNEL_TILE, 16):
         $if C + 8 < CHANNEL_TILE:
           ${XINT8X16_T} vout${ABC[C:C+16]} = ${VCOMBINE_X8}(${VQMOVXN_S16}(vacc${ABC[C:C+8]}), ${VQMOVXN_S16}(vacc${ABC[C+8:C+16]}));
         $else:
           ${XINT8X8_T} vout${ABC[C:C+8]} = ${VQMOVXN_S16}(vacc${ABC[C:C+8]});
 #endif  // !XNN_ARCH_ARM64

       $for C in range(0, CHANNEL_TILE, 16):
         $if C + 8 < CHANNEL_TILE:
           vout${ABC[C:C+16]} = ${VMAXQ_X8}(vout${ABC[C:C+16]}, voutput_min);
         $else:
           $if CHANNEL_TILE == 8:
             vout${ABC[C:C+8]} = ${VMAX_X8}(vout${ABC[C:C+8]}, voutput_min);
           $else:
             vout${ABC[C:C+8]} = ${VMAX_X8}(vout${ABC[C:C+8]}, ${VGET_LOW_X8}(voutput_min));

       $for C in range(0, CHANNEL_TILE, 16):
         $if C + 8 < CHANNEL_TILE:
           vout${ABC[C:C+16]} = ${VMINQ_X8}(vout${ABC[C:C+16]}, voutput_max);
         $else:
           $if CHANNEL_TILE == 8:
             vout${ABC[C:C+8]} = ${VMIN_X8}(vout${ABC[C:C+8]}, voutput_max);
           $else:
             vout${ABC[C:C+8]} = ${VMIN_X8}(vout${ABC[C:C+8]}, ${VGET_LOW_X8}(voutput_max));

       $for C in range(0, CHANNEL_TILE, 16):
         $if C + 8 < CHANNEL_TILE:
           ${VST1Q_X8}(output, vout${ABC[C:C+16]}); output += 16;
         $else:
           ${VST1_X8}(output, vout${ABC[C:C+8]}); output += 8;
     }
     if XNN_UNLIKELY(c != 0) {
       $if CHANNEL_TILE > 8:
         const ${XINT8_T}* k = (const ${XINT8_T}*) ((const int32_t*) w + ${CHANNEL_TILE});
       ${"do " if CHANNEL_TILE > 8 else ""}{
         int32x4_t vacc${ABC[0:4]} = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
         int32x4_t vacc${ABC[4:8]} = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);

         $for K in range(KERNEL_TILE):
           $if CHANNEL_TILE > 8:
             $if DATATYPE == "QU8":
               const int16x8_t vi${K}x${ABC[0:8]} = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i${K}))); i${K} += 8;
             $else:
               const int16x8_t vi${K}x${ABC[0:8]} = vmovl_s8(vld1_s8(i${K})); i${K} += 8;
           $else:
             $if DATATYPE == "QU8":
               const int16x8_t vi${K}x${ABC[0:8]} = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i${K})));
             $else:
               const int16x8_t vi${K}x${ABC[0:8]} = vmovl_s8(vld1_s8(i${K}));
           $if CHANNEL_TILE > 8:
             $if K == 0:
               $if DATATYPE == "QU8":
                 const int16x8_t vk${K}x${ABC[0:8]} = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(k), vkernel_zero_point)); k += 8;
               $else:
                 const int16x8_t vk${K}x${ABC[0:8]} = vmovl_s8(vld1_s8(k)); k += 8;
             $else:
               $if DATATYPE == "QU8":
                 const int16x8_t vk${K}x${ABC[0:8]} = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + ${K * CHANNEL_TILE - 8})), vkernel_zero_point));
               $else:
                 const int16x8_t vk${K}x${ABC[0:8]} = vmovl_s8(vld1_s8((const void*) (k + ${K * CHANNEL_TILE - 8})));
           $else:
             $if K == 0:
               $if DATATYPE == "QU8":
                 const int16x8_t vk${K}x${ABC[0:8]} = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point));
               $else:
                 const int16x8_t vk${K}x${ABC[0:8]} = vmovl_s8(vld1_s8(w));
             $else:
               $if DATATYPE == "QU8":
                 const int16x8_t vk${K}x${ABC[0:8]} = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) ((const ${XINT8_T}*) w + ${K * CHANNEL_TILE})), vkernel_zero_point));
               $else:
                 const int16x8_t vk${K}x${ABC[0:8]} = vmovl_s8(vld1_s8((const void*) ((const ${XINT8_T}*) w + ${K * CHANNEL_TILE})));

           vacc${ABC[0:4]} = vmlal_s16(vacc${ABC[0:4]}, vget_low_s16(vi${K}x${ABC[0:8]}), vget_low_s16(vk${K}x${ABC[0:8]}));
           vacc${ABC[4:8]} = vmlal_s16(vacc${ABC[4:8]}, vget_high_s16(vi${K}x${ABC[0:8]}), vget_high_s16(vk${K}x${ABC[0:8]}));

         $if REQUANTIZATION == "RNDNU":
           vacc${ABC[0:4]} = vqshlq_s32(vacc${ABC[0:4]}, vright_pre_shift);
           vacc${ABC[4:8]} = vqshlq_s32(vacc${ABC[4:8]}, vright_pre_shift);

           vacc${ABC[0:4]} = vqdmulhq_s32(vacc${ABC[0:4]}, vmultiplier);
           vacc${ABC[4:8]} = vqdmulhq_s32(vacc${ABC[4:8]}, vmultiplier);

           vacc${ABC[0:4]} = vrshlq_s32(vacc${ABC[0:4]}, vright_post_shift);
           vacc${ABC[4:8]} = vrshlq_s32(vacc${ABC[4:8]}, vright_post_shift);
         $elif REQUANTIZATION == "FP32":
           float32x4_t vfpacc${ABC[0:4]} = vcvtq_f32_s32(vacc${ABC[0:4]});
           float32x4_t vfpacc${ABC[4:8]} = vcvtq_f32_s32(vacc${ABC[4:8]});

           $if DATATYPE == "QC8":
             const float32x4_t vscale${ABC[0:4]} = vld1q_f32((const float*) ((uintptr_t) w + ${CHANNEL_TILE - 8} * sizeof(int32_t) + ${CHANNEL_TILE * KERNEL_TILE} * sizeof(${XINT8_T})));
             const float32x4_t vscale${ABC[4:8]} = vld1q_f32((const float*) ((uintptr_t) w + ${CHANNEL_TILE - 8} * sizeof(int32_t) + ${CHANNEL_TILE * KERNEL_TILE} * sizeof(${XINT8_T}) + 4 * sizeof(float)));
             vfpacc${ABC[0:4]} = vmulq_f32(vfpacc${ABC[0:4]}, vscale${ABC[0:4]});
             vfpacc${ABC[4:8]} = vmulq_f32(vfpacc${ABC[4:8]}, vscale${ABC[4:8]});
           $else:
             vfpacc${ABC[0:4]} = vmulq_f32(vfpacc${ABC[0:4]}, vscale);
             vfpacc${ABC[4:8]} = vmulq_f32(vfpacc${ABC[4:8]}, vscale);

           $if ARMV8:
             vacc${ABC[0:4]} = vcvtnq_s32_f32(vfpacc${ABC[0:4]});
             vacc${ABC[4:8]} = vcvtnq_s32_f32(vfpacc${ABC[4:8]});
           $else:
             vacc${ABC[0:4]} = vreinterpretq_s32_f32(vaddq_f32(vfpacc${ABC[0:4]}, vmagic_bias));
             vacc${ABC[4:8]} = vreinterpretq_s32_f32(vaddq_f32(vfpacc${ABC[4:8]}, vmagic_bias));

             vacc${ABC[0:4]} = vqsubq_s32(vacc${ABC[0:4]}, vmagic_bias_less_output_zero_point);
             vacc${ABC[4:8]} = vqsubq_s32(vacc${ABC[4:8]}, vmagic_bias_less_output_zero_point);

 #if XNN_ARCH_ARM64
         int16x8_t vacc${ABC[0:8]} = vqmovn_high_s32(vqmovn_s32(vacc${ABC[0:4]}), vacc${ABC[4:8]});
 #else
         int16x8_t vacc${ABC[0:8]} = vcombine_s16(vqmovn_s32(vacc${ABC[0:4]}), vqmovn_s32(vacc${ABC[4:8]}));
 #endif
         $if REQUANTIZATION != "FP32" or ARMV8:
           vacc${ABC[0:8]} = vqaddq_s16(vacc${ABC[0:8]}, voutput_zero_point);

         ${XINT8X8_T} vout${ABC[0:8]} = ${VQMOVXN_S16}(vacc${ABC[0:8]});
         $if CHANNEL_TILE == 8:
           vout${ABC[0:8]} = ${VMAX_X8}(vout${ABC[0:8]}, voutput_min);
           vout${ABC[0:8]} = ${VMIN_X8}(vout${ABC[0:8]}, voutput_max);
         $else:
           vout${ABC[0:8]} = ${VMAX_X8}(vout${ABC[0:8]}, ${VGET_LOW_X8}(voutput_min));
           vout${ABC[0:8]} = ${VMIN_X8}(vout${ABC[0:8]}, ${VGET_LOW_X8}(voutput_max));

         $if CHANNEL_TILE > 8:
           if XNN_LIKELY(c >= 8) {
             ${VST1_X8}(output, vout${ABC[0:8]}); output += 8;
             c -= 8;
           } else {
             if (c & 4) {
               vst1_lane_u32((void*) output, ${VREINTERPRET_U32_X8}(vout${ABC[0:8]}), 0); output += 4;
               vout${ABC[0:8]} = ${VEXT_X8}(vout${ABC[0:8]}, vout${ABC[0:8]}, 4);
             }
             if (c & 2) {
               vst1_lane_u16((void*) output, ${VREINTERPRET_U16_X8}(vout${ABC[0:8]}), 0); output += 2;
               vout${ABC[0:8]} = ${VEXT_X8}(vout${ABC[0:8]}, vout${ABC[0:8]}, 2);
             }
             if (c & 1) {
               ${VST1_LANE_X8}(output, vout${ABC[0:8]}, 0); output += 1;
             }
             c = 0;
           }
         $else:
           if (c & 4) {
             vst1_lane_u32((void*) output, ${VREINTERPRET_U32_X8}(vout${ABC[0:8]}), 0); output += 4;
             vout${ABC[0:8]} = ${VEXT_X8}(vout${ABC[0:8]}, vout${ABC[0:8]}, 4);
           }
           if (c & 2) {
             vst1_lane_u16((void*) output, ${VREINTERPRET_U16_X8}(vout${ABC[0:8]}), 0); output += 2;
             vout${ABC[0:8]} = ${VEXT_X8}(vout${ABC[0:8]}, vout${ABC[0:8]}, 2);
           }
           if (c & 1) {
             ${VST1_LANE_X8}(output, vout${ABC[0:8]}, 0); output += 1;
           }
       }${" while (c != 0);" if CHANNEL_TILE > 8 else ""}
     }

     output = (${XINT8_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.

	$assert REQUANTIZATION in ["FP32", "RNDNU"]
	$assert DATATYPE in ["QC8", "QS8", "QU8"]
	$assert DATATYPE != "QC8" or REQUANTIZATION == "FP32"
	$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
	$assert CHANNEL_TILE % 8 == 0
	$assert CHANNEL_TILE >= 8
	$assert KERNEL_TILE >= 2
	#include <assert.h>

	#include <arm_neon.h>

	#include <xnnpack/dwconv.h>
	$if REQUANTIZATION == "FP32" and ARMV8:
	#include <xnnpack/intrinsics-polyfill.h>


	$PARAMS_STRUCT = REQUANTIZATION.lower() + "_" + ("neonv8" if ARMV8 else "neon")
	$PARAMS_UNION = "xnn_%s_conv_minmax_params" % DATATYPE.lower()
	$XINT8_T = "uint8_t" if DATATYPE == "QU8" else "int8_t"
	$XINT8X8_T = "uint8x8_t" if DATATYPE == "QU8" else "int8x8_t"
	$XINT8X16_T = "uint8x16_t" if DATATYPE == "QU8" else "int8x16_t"
	$VGET_LOW_X8 = "vget_low_u8" if DATATYPE == "QU8" else "vget_low_s8"
	$VCOMBINE_X8 = "vcombine_u8" if DATATYPE == "QU8" else "vcombine_s8"
	$VREINTERPRET_U32_X8 = "vreinterpret_u32_u8" if DATATYPE == "QU8" else "vreinterpret_u32_s8"
	$VREINTERPRET_U16_X8 = "vreinterpret_u16_u8" if DATATYPE == "QU8" else "vreinterpret_u16_s8"
	$VLD1_X8 = "vld1_u8" if DATATYPE == "QU8" else "vld1_s8"
	$VLD1_DUP_X8 = "vld1_dup_u8" if DATATYPE == "QU8" else "vld1_dup_s8"
	$VLD1Q_DUP_X8 = "vld1q_dup_u8" if DATATYPE == "QU8" else "vld1q_dup_s8"
	$VST1_X8 = "vst1_u8" if DATATYPE == "QU8" else "vst1_s8"
	$VST1Q_X8 = "vst1q_u8" if DATATYPE == "QU8" else "vst1q_s8"
	$VST1_LANE_X8 = "vst1_lane_u8" if DATATYPE == "QU8" else "vst1_lane_s8"
	$VST1Q_LANE_X8 = "vst1q_lane_u8" if DATATYPE == "QU8" else "vst1q_lane_s8"
	$VMIN_X8 = "vmin_u8" if DATATYPE == "QU8" else "vmin_s8"
	$VMAX_X8 = "vmax_u8" if DATATYPE == "QU8" else "vmax_s8"
	$VMINQ_X8 = "vminq_u8" if DATATYPE == "QU8" else "vminq_s8"
	$VMAXQ_X8 = "vmaxq_u8" if DATATYPE == "QU8" else "vmaxq_s8"
	$VEXT_X8 = "vext_u8" if DATATYPE == "QU8" else "vext_s8"
	$VQMOVXN_S16 = "vqmovun_s16" if DATATYPE == "QU8" else "vqmovn_s16"
	$VQMOVXN_HIGH_S16 = "vqmovun_high_s16" if DATATYPE == "QU8" else "vqmovn_high_s16"
	$ISA = "neonv8" if ARMV8 else "neon"
	void xnn_${DATATYPE.lower()}_dwconv_minmax_${REQUANTIZATION.lower()}_ukernel_up${CHANNEL_TILE}x${KERNEL_TILE}__${ISA}_mul16(
	size_t channels,
	size_t output_width,
	const ${XINT8_T}** input,
	const void* weights,
	${XINT8_T}* output,
	size_t input_stride,
	size_t output_increment,
	size_t input_offset,
	const ${XINT8_T}* zero,
	const union ${PARAMS_UNION} params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
	{
	assert(channels != 0);
	assert(output_width != 0);

	$if DATATYPE == "QU8":
	const uint8x8_t vkernel_zero_point = vld1_dup_u8(params->${PARAMS_STRUCT}.kernel_zero_point);
	$if REQUANTIZATION == "RNDNU":
	const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->${PARAMS_STRUCT}.right_pre_shift);
	const int32x4_t vmultiplier = vld1q_dup_s32(&params->${PARAMS_STRUCT}.multiplier);
	const int32x4_t vright_post_shift = vld1q_dup_s32(&params->${PARAMS_STRUCT}.right_post_shift);
	$elif REQUANTIZATION == "FP32":
	$if DATATYPE != "QC8":
	const float32x4_t vscale = vld1q_dup_f32(&params->${PARAMS_STRUCT}.scale);
	$if not ARMV8:
	const float32x4_t vmagic_bias = vld1q_dup_f32(&params->${PARAMS_STRUCT}.magic_bias);
	const int32x4_t vmagic_bias_less_output_zero_point = vld1q_dup_s32(&params->${PARAMS_STRUCT}.magic_bias_less_output_zero_point);
	$if REQUANTIZATION != "FP32" or ARMV8:
	const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->${PARAMS_STRUCT}.output_zero_point);
	$if CHANNEL_TILE == 8:
	const ${XINT8X8_T} voutput_min = ${VLD1_DUP_X8}(&params->${PARAMS_STRUCT}.output_min);
	const ${XINT8X8_T} voutput_max = ${VLD1_DUP_X8}(&params->${PARAMS_STRUCT}.output_max);
	$else:
	const ${XINT8X16_T} voutput_min = ${VLD1Q_DUP_X8}(&params->${PARAMS_STRUCT}.output_min);
	const ${XINT8X16_T} voutput_max = ${VLD1Q_DUP_X8}(&params->${PARAMS_STRUCT}.output_max);
	do {
	$for K in range(KERNEL_TILE):
	const ${XINT8_T}* i${K} = input[${K}];
	assert(i${K} != NULL);
	if XNN_UNPREDICTABLE(i${K} != zero) {
	i${K} = (const ${XINT8_T}*) ((uintptr_t) i${K} + input_offset);
	}
	input = (const ${XINT8_T}**) ((uintptr_t) input + input_stride);

	size_t c = channels;
	const void* w = weights;
	for (; c >= ${CHANNEL_TILE}; c -= ${CHANNEL_TILE}) {
	$for C in range(0, CHANNEL_TILE, 4):
	int32x4_t vacc${ABC[C:C+4]} = vld1q_s32(w); w = (const void) ((const int32_t) w + 4);

	$for K in range(KERNEL_TILE):

	$for C in range(0, CHANNEL_TILE, 8):
	$if DATATYPE == "QU8":
	const int16x8_t vi${K}x${ABC[C:C+8]} = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i${K}))); i${K} += 8;
	const int16x8_t vk${K}x${ABC[C:C+8]} = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void) ((const ${XINT8_T}) w + 8);
	$else:
	const int16x8_t vi${K}x${ABC[C:C+8]} = vmovl_s8(vld1_s8(i${K})); i${K} += 8;
	const int16x8_t vk${K}x${ABC[C:C+8]} = vmovl_s8(vld1_s8(w)); w = (const void) ((const ${XINT8_T}) w + 8);

	$for C in range(0, CHANNEL_TILE, 8):
	vacc${ABC[C:C+4]} = vmlal_s16(vacc${ABC[C:C+4]}, vget_low_s16(vi${K}x${ABC[C:C+8]}), vget_low_s16(vk${K}x${ABC[C:C+8]}));
	vacc${ABC[C+4:C+8]} = vmlal_s16(vacc${ABC[C+4:C+8]}, vget_high_s16(vi${K}x${ABC[C:C+8]}), vget_high_s16(vk${K}x${ABC[C:C+8]}));

	$if REQUANTIZATION == "RNDNU":
	$for C in range(0, CHANNEL_TILE, 4):
	vacc${ABC[C:C+4]} = vqshlq_s32(vacc${ABC[C:C+4]}, vright_pre_shift);

	$for C in range(0, CHANNEL_TILE, 4):
	vacc${ABC[C:C+4]} = vqdmulhq_s32(vacc${ABC[C:C+4]}, vmultiplier);

	$for C in range(0, CHANNEL_TILE, 4):
	vacc${ABC[C:C+4]} = vrshlq_s32(vacc${ABC[C:C+4]}, vright_post_shift);
	$elif REQUANTIZATION == "FP32":
	$for C in range(0, CHANNEL_TILE, 4):
	float32x4_t vfpacc${ABC[C:C+4]} = vcvtq_f32_s32(vacc${ABC[C:C+4]});

	$if DATATYPE == "QC8":
	$for C in range(0, CHANNEL_TILE, 4):
	const float32x4_t vscale${ABC[C:C+4]} = vld1q_f32((const float) w); w = (const void) ((const float*) w + 4);

	$for C in range(0, CHANNEL_TILE, 4):
	vfpacc${ABC[C:C+4]} = vmulq_f32(vfpacc${ABC[C:C+4]}, vscale${ABC[C:C+4]});
	$else:
	$for C in range(0, CHANNEL_TILE, 4):
	vfpacc${ABC[C:C+4]} = vmulq_f32(vfpacc${ABC[C:C+4]}, vscale);

	$if ARMV8:
	$for C in range(0, CHANNEL_TILE, 4):
	vacc${ABC[C:C+4]} = vcvtnq_s32_f32(vfpacc${ABC[C:C+4]});
	$else:
	$for C in range(0, CHANNEL_TILE, 4):
	vacc${ABC[C:C+4]} = vreinterpretq_s32_f32(vaddq_f32(vfpacc${ABC[C:C+4]}, vmagic_bias));

	$for C in range(0, CHANNEL_TILE, 4):
	vacc${ABC[C:C+4]} = vqsubq_s32(vacc${ABC[C:C+4]}, vmagic_bias_less_output_zero_point);

	#if XNN_ARCH_ARM64
	$for C in range(0, CHANNEL_TILE, 8):
	int16x8_t vacc${ABC[C:C+8]} = vqmovn_high_s32(vqmovn_s32(vacc${ABC[C:C+4]}), vacc${ABC[C+4:C+8]});

	$if REQUANTIZATION != "FP32" or ARMV8:
	$for C in range(0, CHANNEL_TILE, 8):
	vacc${ABC[C:C+8]} = vqaddq_s16(vacc${ABC[C:C+8]}, voutput_zero_point);

	$for C in range(0, CHANNEL_TILE, 16):
	$if C + 8 < CHANNEL_TILE:
	${XINT8X16_T} vout${ABC[C:C+16]} = ${VQMOVXN_HIGH_S16}(${VQMOVXN_S16}(vacc${ABC[C:C+8]}), vacc${ABC[C+8:C+16]});
	$else:
	${XINT8X8_T} vout${ABC[C:C+8]} = ${VQMOVXN_S16}(vacc${ABC[C:C+8]});
	#else // !XNN_ARCH_ARM64
	$for C in range(0, CHANNEL_TILE, 8):
	int16x8_t vacc${ABC[C:C+8]} = vcombine_s16(vqmovn_s32(vacc${ABC[C:C+4]}), vqmovn_s32(vacc${ABC[C+4:C+8]}));

	$if REQUANTIZATION != "FP32" or ARMV8:
	$for C in range(0, CHANNEL_TILE, 8):
	vacc${ABC[C:C+8]} = vqaddq_s16(vacc${ABC[C:C+8]}, voutput_zero_point);

	$for C in range(0, CHANNEL_TILE, 16):
	$if C + 8 < CHANNEL_TILE:
	${XINT8X16_T} vout${ABC[C:C+16]} = ${VCOMBINE_X8}(${VQMOVXN_S16}(vacc${ABC[C:C+8]}), ${VQMOVXN_S16}(vacc${ABC[C+8:C+16]}));
	$else:
	${XINT8X8_T} vout${ABC[C:C+8]} = ${VQMOVXN_S16}(vacc${ABC[C:C+8]});
	#endif // !XNN_ARCH_ARM64

	$for C in range(0, CHANNEL_TILE, 16):
	$if C + 8 < CHANNEL_TILE:
	vout${ABC[C:C+16]} = ${VMAXQ_X8}(vout${ABC[C:C+16]}, voutput_min);
	$else:
	$if CHANNEL_TILE == 8:
	vout${ABC[C:C+8]} = ${VMAX_X8}(vout${ABC[C:C+8]}, voutput_min);
	$else:
	vout${ABC[C:C+8]} = ${VMAX_X8}(vout${ABC[C:C+8]}, ${VGET_LOW_X8}(voutput_min));

	$for C in range(0, CHANNEL_TILE, 16):
	$if C + 8 < CHANNEL_TILE:
	vout${ABC[C:C+16]} = ${VMINQ_X8}(vout${ABC[C:C+16]}, voutput_max);
	$else:
	$if CHANNEL_TILE == 8:
	vout${ABC[C:C+8]} = ${VMIN_X8}(vout${ABC[C:C+8]}, voutput_max);
	$else:
	vout${ABC[C:C+8]} = ${VMIN_X8}(vout${ABC[C:C+8]}, ${VGET_LOW_X8}(voutput_max));

	$for C in range(0, CHANNEL_TILE, 16):
	$if C + 8 < CHANNEL_TILE:
	${VST1Q_X8}(output, vout${ABC[C:C+16]}); output += 16;
	$else:
	${VST1_X8}(output, vout${ABC[C:C+8]}); output += 8;
	}
	if XNN_UNLIKELY(c != 0) {
	$if CHANNEL_TILE > 8:
	const ${XINT8_T}* k = (const ${XINT8_T}) ((const int32_t) w + ${CHANNEL_TILE});
	${"do " if CHANNEL_TILE > 8 else ""}{
	int32x4_t vacc${ABC[0:4]} = vld1q_s32(w); w = (const void) ((const int32_t) w + 4);
	int32x4_t vacc${ABC[4:8]} = vld1q_s32(w); w = (const void) ((const int32_t) w + 4);

	$for K in range(KERNEL_TILE):
	$if CHANNEL_TILE > 8:
	$if DATATYPE == "QU8":
	const int16x8_t vi${K}x${ABC[0:8]} = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i${K}))); i${K} += 8;
	$else:
	const int16x8_t vi${K}x${ABC[0:8]} = vmovl_s8(vld1_s8(i${K})); i${K} += 8;
	$else:
	$if DATATYPE == "QU8":
	const int16x8_t vi${K}x${ABC[0:8]} = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i${K})));
	$else:
	const int16x8_t vi${K}x${ABC[0:8]} = vmovl_s8(vld1_s8(i${K}));
	$if CHANNEL_TILE > 8:
	$if K == 0:
	$if DATATYPE == "QU8":
	const int16x8_t vk${K}x${ABC[0:8]} = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(k), vkernel_zero_point)); k += 8;
	$else:
	const int16x8_t vk${K}x${ABC[0:8]} = vmovl_s8(vld1_s8(k)); k += 8;
	$else:
	$if DATATYPE == "QU8":
	const int16x8_t vk${K}x${ABC[0:8]} = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void) (k + ${K CHANNEL_TILE - 8})), vkernel_zero_point));
	$else:
	const int16x8_t vk${K}x${ABC[0:8]} = vmovl_s8(vld1_s8((const void) (k + ${K CHANNEL_TILE - 8})));
	$else:
	$if K == 0:
	$if DATATYPE == "QU8":
	const int16x8_t vk${K}x${ABC[0:8]} = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point));
	$else:
	const int16x8_t vk${K}x${ABC[0:8]} = vmovl_s8(vld1_s8(w));
	$else:
	$if DATATYPE == "QU8":
	const int16x8_t vk${K}x${ABC[0:8]} = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void) ((const ${XINT8_T}) w + ${K * CHANNEL_TILE})), vkernel_zero_point));
	$else:
	const int16x8_t vk${K}x${ABC[0:8]} = vmovl_s8(vld1_s8((const void) ((const ${XINT8_T}) w + ${K * CHANNEL_TILE})));

	vacc${ABC[0:4]} = vmlal_s16(vacc${ABC[0:4]}, vget_low_s16(vi${K}x${ABC[0:8]}), vget_low_s16(vk${K}x${ABC[0:8]}));
	vacc${ABC[4:8]} = vmlal_s16(vacc${ABC[4:8]}, vget_high_s16(vi${K}x${ABC[0:8]}), vget_high_s16(vk${K}x${ABC[0:8]}));

	$if REQUANTIZATION == "RNDNU":
	vacc${ABC[0:4]} = vqshlq_s32(vacc${ABC[0:4]}, vright_pre_shift);
	vacc${ABC[4:8]} = vqshlq_s32(vacc${ABC[4:8]}, vright_pre_shift);

	vacc${ABC[0:4]} = vqdmulhq_s32(vacc${ABC[0:4]}, vmultiplier);
	vacc${ABC[4:8]} = vqdmulhq_s32(vacc${ABC[4:8]}, vmultiplier);

	vacc${ABC[0:4]} = vrshlq_s32(vacc${ABC[0:4]}, vright_post_shift);
	vacc${ABC[4:8]} = vrshlq_s32(vacc${ABC[4:8]}, vright_post_shift);
	$elif REQUANTIZATION == "FP32":
	float32x4_t vfpacc${ABC[0:4]} = vcvtq_f32_s32(vacc${ABC[0:4]});
	float32x4_t vfpacc${ABC[4:8]} = vcvtq_f32_s32(vacc${ABC[4:8]});

	$if DATATYPE == "QC8":
	const float32x4_t vscale${ABC[0:4]} = vld1q_f32((const float) ((uintptr_t) w + ${CHANNEL_TILE - 8} sizeof(int32_t) + ${CHANNEL_TILE * KERNEL_TILE} * sizeof(${XINT8_T})));
	const float32x4_t vscale${ABC[4:8]} = vld1q_f32((const float) ((uintptr_t) w + ${CHANNEL_TILE - 8} sizeof(int32_t) + ${CHANNEL_TILE * KERNEL_TILE} * sizeof(${XINT8_T}) + 4 * sizeof(float)));
	vfpacc${ABC[0:4]} = vmulq_f32(vfpacc${ABC[0:4]}, vscale${ABC[0:4]});
	vfpacc${ABC[4:8]} = vmulq_f32(vfpacc${ABC[4:8]}, vscale${ABC[4:8]});
	$else:
	vfpacc${ABC[0:4]} = vmulq_f32(vfpacc${ABC[0:4]}, vscale);
	vfpacc${ABC[4:8]} = vmulq_f32(vfpacc${ABC[4:8]}, vscale);

	$if ARMV8:
	vacc${ABC[0:4]} = vcvtnq_s32_f32(vfpacc${ABC[0:4]});
	vacc${ABC[4:8]} = vcvtnq_s32_f32(vfpacc${ABC[4:8]});
	$else:
	vacc${ABC[0:4]} = vreinterpretq_s32_f32(vaddq_f32(vfpacc${ABC[0:4]}, vmagic_bias));
	vacc${ABC[4:8]} = vreinterpretq_s32_f32(vaddq_f32(vfpacc${ABC[4:8]}, vmagic_bias));

	vacc${ABC[0:4]} = vqsubq_s32(vacc${ABC[0:4]}, vmagic_bias_less_output_zero_point);
	vacc${ABC[4:8]} = vqsubq_s32(vacc${ABC[4:8]}, vmagic_bias_less_output_zero_point);

	#if XNN_ARCH_ARM64
	int16x8_t vacc${ABC[0:8]} = vqmovn_high_s32(vqmovn_s32(vacc${ABC[0:4]}), vacc${ABC[4:8]});
	#else
	int16x8_t vacc${ABC[0:8]} = vcombine_s16(vqmovn_s32(vacc${ABC[0:4]}), vqmovn_s32(vacc${ABC[4:8]}));
	#endif
	$if REQUANTIZATION != "FP32" or ARMV8:
	vacc${ABC[0:8]} = vqaddq_s16(vacc${ABC[0:8]}, voutput_zero_point);

	${XINT8X8_T} vout${ABC[0:8]} = ${VQMOVXN_S16}(vacc${ABC[0:8]});
	$if CHANNEL_TILE == 8:
	vout${ABC[0:8]} = ${VMAX_X8}(vout${ABC[0:8]}, voutput_min);
	vout${ABC[0:8]} = ${VMIN_X8}(vout${ABC[0:8]}, voutput_max);
	$else:
	vout${ABC[0:8]} = ${VMAX_X8}(vout${ABC[0:8]}, ${VGET_LOW_X8}(voutput_min));
	vout${ABC[0:8]} = ${VMIN_X8}(vout${ABC[0:8]}, ${VGET_LOW_X8}(voutput_max));

	$if CHANNEL_TILE > 8:
	if XNN_LIKELY(c >= 8) {
	${VST1_X8}(output, vout${ABC[0:8]}); output += 8;
	c -= 8;
	} else {
	if (c & 4) {
	vst1_lane_u32((void*) output, ${VREINTERPRET_U32_X8}(vout${ABC[0:8]}), 0); output += 4;
	vout${ABC[0:8]} = ${VEXT_X8}(vout${ABC[0:8]}, vout${ABC[0:8]}, 4);
	}
	if (c & 2) {
	vst1_lane_u16((void*) output, ${VREINTERPRET_U16_X8}(vout${ABC[0:8]}), 0); output += 2;
	vout${ABC[0:8]} = ${VEXT_X8}(vout${ABC[0:8]}, vout${ABC[0:8]}, 2);
	}
	if (c & 1) {
	${VST1_LANE_X8}(output, vout${ABC[0:8]}, 0); output += 1;
	}
	c = 0;
	}
	$else:
	if (c & 4) {
	vst1_lane_u32((void*) output, ${VREINTERPRET_U32_X8}(vout${ABC[0:8]}), 0); output += 4;
	vout${ABC[0:8]} = ${VEXT_X8}(vout${ABC[0:8]}, vout${ABC[0:8]}, 4);
	}
	if (c & 2) {
	vst1_lane_u16((void*) output, ${VREINTERPRET_U16_X8}(vout${ABC[0:8]}), 0); output += 2;
	vout${ABC[0:8]} = ${VEXT_X8}(vout${ABC[0:8]}, vout${ABC[0:8]}, 2);
	}
	if (c & 1) {
	${VST1_LANE_X8}(output, vout${ABC[0:8]}, 0); output += 1;
	}
	}${" while (c != 0);" if CHANNEL_TILE > 8 else ""}
	}

	output = (${XINT8_T}*) ((uintptr_t) output + output_increment);
	} while (--output_width != 0);
	}