src/qs8-dwconv/unipass-wasmsimd-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.

 $ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
 $assert REQUANTIZATION == "FP32"
 $assert DATATYPE in ["QC8", "QS8", "QU8"]
 $assert not ADD16 or DATATYPE != "QU8"
 $assert CHANNEL_TILE % 8 == 0
 $assert CHANNEL_TILE >= 8
 $assert KERNEL_TILE >= 2
 #include <assert.h>

 #include <wasm_simd128.h>

 #include <xnnpack/dwconv.h>


 $PARAMS_STRUCT = REQUANTIZATION.lower() + "_wasmsimd"
 $PARAMS_UNION = "xnn_%s_conv_minmax_params" % DATATYPE.lower()
 $XINT8_T = "uint8_t" if DATATYPE == "QU8" else "int8_t"
 $WASM_X16X8_LOAD8X8 = "wasm_u16x8_load8x8" if DATATYPE == "QU8" else "wasm_i16x8_load8x8"
 $WASM_X32X4_EXTEND_LOW_X16X8 = "wasm_u32x4_extend_low_u16x8" if DATATYPE == "QU8" else "wasm_i32x4_extend_low_i16x8"
 $WASM_X32X4_EXTEND_HIGH_X16X8 = "wasm_u32x4_extend_high_u16x8" if DATATYPE == "QU8" else "wasm_i32x4_extend_high_i16x8"
 $WASM_X8X16_NARROW_I16X8 = "wasm_u8x16_narrow_i16x8" if DATATYPE == "QU8" else "wasm_i8x16_narrow_i16x8"
 $WASM_X8X16_MIN = "wasm_u8x16_min" if DATATYPE == "QU8" else "wasm_i8x16_min"
 void xnn_${DATATYPE.lower()}_dwconv_minmax_${REQUANTIZATION.lower()}_ukernel_up${CHANNEL_TILE}x${KERNEL_TILE}__wasmsimd_mul16${"_add16" if ADD16 else ""}(
     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);

   $if DATATYPE == "QU8":
     const v128_t vkernel_zero_point = wasm_u32x4_load16x4(params->${PARAMS_STRUCT}.kernel_zero_point);
   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 -= ${CHANNEL_TILE}) {
       v128_t vacc${ABC[0:4]} = wasm_v128_load(w);
       $for C in range(4, CHANNEL_TILE, 4):
         v128_t vacc${ABC[C:C+4]} = wasm_v128_load((const void*) ((uintptr_t) w + ${C} * sizeof(int32_t)));

       $for K in range(KERNEL_TILE):

         $for C in range(0, CHANNEL_TILE, 8):
           $if C == 0:
             const v128_t vi${K}x${ABC[0:8]} = ${WASM_X16X8_LOAD8X8}(i${K});
           $else:
             const v128_t vi${K}x${ABC[C:C+8]} = ${WASM_X16X8_LOAD8X8}(i${K} + ${C});
           const v128_t vk${K}x${ABC[C:C+8]} = ${WASM_X16X8_LOAD8X8}((const void*) ((uintptr_t) w + ${CHANNEL_TILE} * sizeof(int32_t) + ${K * CHANNEL_TILE + C} * sizeof(${XINT8_T})));
         $if DATATYPE == "QU8":
           $for C in range(0, CHANNEL_TILE, 8):
             $if K == 1:
               v128_t vsumx${ABC[C:C+8]} = wasm_i16x8_add(vi0x${ABC[C:C+8]}, vi1x${ABC[C:C+8]});
             $elif K > 1:
               vsumx${ABC[C:C+8]} = wasm_i16x8_add(vsumx${ABC[C:C+8]}, vi${K}x${ABC[C:C+8]});
         i${K} += ${CHANNEL_TILE};

         $for C in range(0, CHANNEL_TILE, 8):
           $if K == 0:
             v128_t vprod${ABC[C:C+8]} = wasm_i16x8_mul(vi${K}x${ABC[C:C+8]}, vk${K}x${ABC[C:C+8]});
           $elif K % 2 == 0 or K + 1 == KERNEL_TILE or not ADD16:
             vprod${ABC[C:C+8]} = wasm_i16x8_mul(vi${K}x${ABC[C:C+8]}, vk${K}x${ABC[C:C+8]});
           $else:
             vprod${ABC[C:C+8]} = wasm_i16x8_add(vprod${ABC[C:C+8]}, wasm_i16x8_mul(vi${K}x${ABC[C:C+8]}, vk${K}x${ABC[C:C+8]}));

         $if not ADD16 or K % 2 == 1 or K + 1 == KERNEL_TILE:
           $for C in range(0, CHANNEL_TILE, 8):
             vacc${ABC[C:C+4]} = wasm_i32x4_add(vacc${ABC[C:C+4]}, ${WASM_X32X4_EXTEND_LOW_X16X8}(vprod${ABC[C:C+8]}));
             vacc${ABC[C+4:C+8]} = wasm_i32x4_add(vacc${ABC[C+4:C+8]}, ${WASM_X32X4_EXTEND_HIGH_X16X8}(vprod${ABC[C:C+8]}));

       $if DATATYPE == "QU8":
         $for C in range(0, CHANNEL_TILE, 8):
           vacc${ABC[C:C+4]} = wasm_i32x4_sub(vacc${ABC[C:C+4]}, wasm_i32x4_mul(wasm_u32x4_extend_low_u16x8(vsumx${ABC[C:C+8]}), vkernel_zero_point));
           vacc${ABC[C+4:C+8]} = wasm_i32x4_sub(vacc${ABC[C+4:C+8]}, wasm_i32x4_mul(wasm_u32x4_extend_high_u16x8(vsumx${ABC[C:C+8]}), vkernel_zero_point));

       w = (const void*) ((uintptr_t) w + ${CHANNEL_TILE} * sizeof(int32_t) + ${KERNEL_TILE * CHANNEL_TILE} * sizeof(${XINT8_T}));

       $for C in range(0, CHANNEL_TILE, 4):
         vacc${ABC[C:C+4]} = wasm_f32x4_convert_i32x4(vacc${ABC[C:C+4]});

       $if DATATYPE == "QC8":
         const v128_t vscale${ABC[0:4]} = wasm_v128_load(w);
         $for C in range(4, CHANNEL_TILE, 4):
           const v128_t vscale${ABC[C:C+4]} = wasm_v128_load((const float*) w + ${C});
         w = (const void*) ((const float*) w + ${CHANNEL_TILE});

         $for C in range(0, CHANNEL_TILE, 4):
           vacc${ABC[C:C+4]} = wasm_f32x4_mul(vacc${ABC[C:C+4]}, vscale${ABC[C:C+4]});
       $else:
         const v128_t vscale = wasm_v128_load64_splat(params->${PARAMS_STRUCT}.scale);
         $for C in range(0, CHANNEL_TILE, 4):
           vacc${ABC[C:C+4]} = wasm_f32x4_mul(vacc${ABC[C:C+4]}, vscale);

       const v128_t vmagic_bias = wasm_v128_load64_splat(params->${PARAMS_STRUCT}.magic_bias);
       $for C in range(0, CHANNEL_TILE, 4):
         vacc${ABC[C:C+4]} = wasm_f32x4_add(vacc${ABC[C:C+4]}, vmagic_bias);

       const v128_t vmagic_min = wasm_v128_load64_splat(params->${PARAMS_STRUCT}.magic_min);
       $for C in range(0, CHANNEL_TILE, 4):
         vacc${ABC[C:C+4]} = wasm_i32x4_max(vacc${ABC[C:C+4]}, vmagic_min);

       const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->${PARAMS_STRUCT}.magic_bias_less_output_zero_point);
       $for C in range(0, CHANNEL_TILE, 4):
         vacc${ABC[C:C+4]} = wasm_i32x4_sub(vacc${ABC[C:C+4]}, vmagic_bias_less_output_zero_point);

       $for C in range(0, CHANNEL_TILE, 8):
         v128_t vout${ABC[C:C+8]} = wasm_i16x8_narrow_i32x4(vacc${ABC[C:C+4]}, vacc${ABC[C+4:C+8]});

       $for C in range(0, CHANNEL_TILE, 16):
         $if C + 8 < CHANNEL_TILE:
           v128_t vout${ABC[C:C+16]} = ${WASM_X8X16_NARROW_I16X8}(vout${ABC[C:C+8]}, vout${ABC[C+8:C+16]});
         $else:
           v128_t vout${ABC[C:C+8]}${ABC[C:C+8]} = ${WASM_X8X16_NARROW_I16X8}(vout${ABC[C:C+8]}, vout${ABC[C:C+8]});

       const v128_t voutput_max = wasm_v128_load64_splat(params->${PARAMS_STRUCT}.output_max);
       $for C in range(0, CHANNEL_TILE, 16):
         $if C + 8 < CHANNEL_TILE:
           vout${ABC[C:C+16]} = ${WASM_X8X16_MIN}(vout${ABC[C:C+16]}, voutput_max);
         $else:
           vout${ABC[C:C+8]}${ABC[C:C+8]} = ${WASM_X8X16_MIN}(vout${ABC[C:C+8]}${ABC[C:C+8]}, voutput_max);

       $if CHANNEL_TILE > 8:
         wasm_v128_store(output, vout${ABC[0:16]});
       $else:
         *((double*) output) = wasm_f64x2_extract_lane(vout${ABC[0:8]}${ABC[0:8]}, 0);
       $for C in range(16, CHANNEL_TILE, 16):
         $if C + 8 < CHANNEL_TILE:
           wasm_v128_store(output + ${C}, vout${ABC[C:C+16]});
         $else:
           *((double*) (output + ${C})) = wasm_f64x2_extract_lane(vout${ABC[C:C+8]}${ABC[C:C+8]}, 0);
       output += ${CHANNEL_TILE};
     }
     if XNN_UNLIKELY(c != 0) {
       $if CHANNEL_TILE > 8:
         const ${XINT8_T}* k = (const ${XINT8_T}*) ((uintptr_t) w + ${CHANNEL_TILE} * sizeof(int32_t));
       ${"do " if CHANNEL_TILE > 8 else ""}{
         v128_t vacc${ABC[0:4]} = wasm_v128_load(w);
         v128_t vacc${ABC[4:8]} = wasm_v128_load((const void*) ((uintptr_t) w + 4 * sizeof(int32_t)));

         $for K in range(KERNEL_TILE):

           const v128_t vi${K}x${ABC[0:8]} = ${WASM_X16X8_LOAD8X8}(i${K});
           $if CHANNEL_TILE > 8:
             $if K == 0:
               const v128_t vk${K}x${ABC[0:8]} = ${WASM_X16X8_LOAD8X8}(k);
             $else:
               const v128_t vk${K}x${ABC[0:8]} = ${WASM_X16X8_LOAD8X8}((const void*) (k + ${K * CHANNEL_TILE}));
           $else:
             const v128_t vk${K}x${ABC[0:8]} = ${WASM_X16X8_LOAD8X8}((const void*) ((uintptr_t) w + ${CHANNEL_TILE} * sizeof(int32_t) + ${K * CHANNEL_TILE} * sizeof(${XINT8_T})));
           $if DATATYPE == "QU8":
             $if K == 1:
               v128_t vsumx${ABC[0:8]} = wasm_i16x8_add(vi0x${ABC[0:8]}, vi1x${ABC[0:8]});
             $elif K > 1:
               vsumx${ABC[0:8]} = wasm_i16x8_add(vsumx${ABC[0:8]}, vi${K}x${ABC[0:8]});
           $if CHANNEL_TILE > 8:
             i${K} += 8;

           $if K == 0:
             v128_t vprod${ABC[0:8]} = wasm_i16x8_mul(vi${K}x${ABC[0:8]}, vk${K}x${ABC[0:8]});
           $elif K % 2 == 0 or K + 1 == KERNEL_TILE or not ADD16:
             vprod${ABC[0:8]} = wasm_i16x8_mul(vi${K}x${ABC[0:8]}, vk${K}x${ABC[0:8]});
           $else:
             vprod${ABC[0:8]} = wasm_i16x8_add(vprod${ABC[0:8]}, wasm_i16x8_mul(vi${K}x${ABC[0:8]}, vk${K}x${ABC[0:8]}));

           $if not ADD16 or K % 2 == 1 or K + 1 == KERNEL_TILE:
             vacc${ABC[0:4]} = wasm_i32x4_add(vacc${ABC[0:4]}, ${WASM_X32X4_EXTEND_LOW_X16X8}(vprod${ABC[0:8]}));
             vacc${ABC[4:8]} = wasm_i32x4_add(vacc${ABC[4:8]}, ${WASM_X32X4_EXTEND_HIGH_X16X8}(vprod${ABC[0:8]}));

         $if CHANNEL_TILE > 8:
           k += 8;

       $if DATATYPE == "QU8":
         vacc${ABC[0:4]} = wasm_i32x4_sub(vacc${ABC[0:4]}, wasm_i32x4_mul(wasm_u32x4_extend_low_u16x8(vsumx${ABC[0:8]}), vkernel_zero_point));
         vacc${ABC[4:8]} = wasm_i32x4_sub(vacc${ABC[4:8]}, wasm_i32x4_mul(wasm_u32x4_extend_high_u16x8(vsumx${ABC[0:8]}), vkernel_zero_point));

       vacc${ABC[0:4]} = wasm_f32x4_convert_i32x4(vacc${ABC[0:4]});
       vacc${ABC[4:8]} = wasm_f32x4_convert_i32x4(vacc${ABC[4:8]});

       $if DATATYPE == "QC8":
         const v128_t vscale${ABC[0:4]} = wasm_v128_load((const float*) ((uintptr_t) w + ${CHANNEL_TILE} * sizeof(int32_t) + ${CHANNEL_TILE * KERNEL_TILE} * sizeof(${XINT8_T})));
         const v128_t vscale${ABC[4:8]} = wasm_v128_load((const float*) ((uintptr_t) w + ${CHANNEL_TILE} * sizeof(int32_t) + ${CHANNEL_TILE * KERNEL_TILE} * sizeof(${XINT8_T}) + 4 * sizeof(float)));

         vacc${ABC[0:4]} = wasm_f32x4_mul(vacc${ABC[0:4]}, vscale${ABC[0:4]});
         vacc${ABC[4:8]} = wasm_f32x4_mul(vacc${ABC[4:8]}, vscale${ABC[4:8]});
       $else:
         const v128_t vscale = wasm_v128_load64_splat(params->${PARAMS_STRUCT}.scale);
         vacc${ABC[0:4]} = wasm_f32x4_mul(vacc${ABC[0:4]}, vscale);
         vacc${ABC[4:8]} = wasm_f32x4_mul(vacc${ABC[4:8]}, vscale);

       const v128_t vmagic_bias = wasm_v128_load64_splat(params->${PARAMS_STRUCT}.magic_bias);
       vacc${ABC[0:4]} = wasm_f32x4_add(vacc${ABC[0:4]}, vmagic_bias);
       vacc${ABC[4:8]} = wasm_f32x4_add(vacc${ABC[4:8]}, vmagic_bias);

       const v128_t vmagic_min = wasm_v128_load64_splat(params->${PARAMS_STRUCT}.magic_min);
       vacc${ABC[0:4]} = wasm_i32x4_max(vacc${ABC[0:4]}, vmagic_min);
       vacc${ABC[4:8]} = wasm_i32x4_max(vacc${ABC[4:8]}, vmagic_min);

       const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->${PARAMS_STRUCT}.magic_bias_less_output_zero_point);
       vacc${ABC[0:4]} = wasm_i32x4_sub(vacc${ABC[0:4]}, vmagic_bias_less_output_zero_point);
       vacc${ABC[4:8]} = wasm_i32x4_sub(vacc${ABC[4:8]}, vmagic_bias_less_output_zero_point);

       v128_t vout${ABC[0:8]} = wasm_i16x8_narrow_i32x4(vacc${ABC[0:4]}, vacc${ABC[4:8]});
       v128_t vout${ABC[0:8]}${ABC[0:8]} = ${WASM_X8X16_NARROW_I16X8}(vout${ABC[0:8]}, vout${ABC[0:8]});

       const v128_t voutput_max = wasm_v128_load64_splat(params->${PARAMS_STRUCT}.output_max);
       vout${ABC[0:8]}${ABC[0:8]} = ${WASM_X8X16_MIN}(vout${ABC[0:8]}${ABC[0:8]}, voutput_max);

       $if CHANNEL_TILE > 8:
         w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t));

       $if CHANNEL_TILE > 8:
         if XNN_LIKELY(c >= 8) {
           *((double*) output) = wasm_f64x2_extract_lane(vout${ABC[0:8]}${ABC[0:8]}, 0);
           output += 8;
           c -= 8;
         } else {
           if (c & 4) {
             *((float*) output) = wasm_f32x4_extract_lane(vout${ABC[0:8]}${ABC[0:8]}, 0);
             vout${ABC[0:8]}${ABC[0:8]} = wasm_u64x2_shr(vout${ABC[0:8]}${ABC[0:8]}, 32);
             output += 4;
           }
           uint32_t vout${ABC[0:4]} = wasm_i32x4_extract_lane(vout${ABC[0:8]}${ABC[0:8]}, 0);
           if (c & 2) {
             *((uint16_t*) output) = (uint16_t) vout${ABC[0:4]};
             vout${ABC[0:4]} >>= 16;
             output += 2;
           }
           if (c & 1) {
             *output = (${XINT8_T}) vout${ABC[0:4]};
             output += 1;
           }
           c = 0;
         }
       $else:
         if (c & 4) {
           *((float*) output) = wasm_f32x4_extract_lane(vout${ABC[0:8]}${ABC[0:8]}, 0);
           vout${ABC[0:8]}${ABC[0:8]} = wasm_u64x2_shr(vout${ABC[0:8]}${ABC[0:8]}, 32);
           output += 4;
         }
         uint32_t vout${ABC[0:4]} = wasm_i32x4_extract_lane(vout${ABC[0:8]}${ABC[0:8]}, 0);
         if (c & 2) {
           *((uint16_t*) output) = (uint16_t) vout${ABC[0:4]};
           vout${ABC[0:4]} >>= 16;
           output += 2;
         }
         if (c & 1) {
           *output = (${XINT8_T}) vout${ABC[0:4]};
           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.

	$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
	$assert REQUANTIZATION == "FP32"
	$assert DATATYPE in ["QC8", "QS8", "QU8"]
	$assert not ADD16 or DATATYPE != "QU8"
	$assert CHANNEL_TILE % 8 == 0
	$assert CHANNEL_TILE >= 8
	$assert KERNEL_TILE >= 2
	#include <assert.h>

	#include <wasm_simd128.h>

	#include <xnnpack/dwconv.h>


	$PARAMS_STRUCT = REQUANTIZATION.lower() + "_wasmsimd"
	$PARAMS_UNION = "xnn_%s_conv_minmax_params" % DATATYPE.lower()
	$XINT8_T = "uint8_t" if DATATYPE == "QU8" else "int8_t"
	$WASM_X16X8_LOAD8X8 = "wasm_u16x8_load8x8" if DATATYPE == "QU8" else "wasm_i16x8_load8x8"
	$WASM_X32X4_EXTEND_LOW_X16X8 = "wasm_u32x4_extend_low_u16x8" if DATATYPE == "QU8" else "wasm_i32x4_extend_low_i16x8"
	$WASM_X32X4_EXTEND_HIGH_X16X8 = "wasm_u32x4_extend_high_u16x8" if DATATYPE == "QU8" else "wasm_i32x4_extend_high_i16x8"
	$WASM_X8X16_NARROW_I16X8 = "wasm_u8x16_narrow_i16x8" if DATATYPE == "QU8" else "wasm_i8x16_narrow_i16x8"
	$WASM_X8X16_MIN = "wasm_u8x16_min" if DATATYPE == "QU8" else "wasm_i8x16_min"
	void xnn_${DATATYPE.lower()}_dwconv_minmax_${REQUANTIZATION.lower()}_ukernel_up${CHANNEL_TILE}x${KERNEL_TILE}__wasmsimd_mul16${"_add16" if ADD16 else ""}(
	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);

	$if DATATYPE == "QU8":
	const v128_t vkernel_zero_point = wasm_u32x4_load16x4(params->${PARAMS_STRUCT}.kernel_zero_point);
	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 -= ${CHANNEL_TILE}) {
	v128_t vacc${ABC[0:4]} = wasm_v128_load(w);
	$for C in range(4, CHANNEL_TILE, 4):
	v128_t vacc${ABC[C:C+4]} = wasm_v128_load((const void) ((uintptr_t) w + ${C} sizeof(int32_t)));

	$for K in range(KERNEL_TILE):

	$for C in range(0, CHANNEL_TILE, 8):
	$if C == 0:
	const v128_t vi${K}x${ABC[0:8]} = ${WASM_X16X8_LOAD8X8}(i${K});
	$else:
	const v128_t vi${K}x${ABC[C:C+8]} = ${WASM_X16X8_LOAD8X8}(i${K} + ${C});
	const v128_t vk${K}x${ABC[C:C+8]} = ${WASM_X16X8_LOAD8X8}((const void) ((uintptr_t) w + ${CHANNEL_TILE} sizeof(int32_t) + ${K * CHANNEL_TILE + C} * sizeof(${XINT8_T})));
	$if DATATYPE == "QU8":
	$for C in range(0, CHANNEL_TILE, 8):
	$if K == 1:
	v128_t vsumx${ABC[C:C+8]} = wasm_i16x8_add(vi0x${ABC[C:C+8]}, vi1x${ABC[C:C+8]});
	$elif K > 1:
	vsumx${ABC[C:C+8]} = wasm_i16x8_add(vsumx${ABC[C:C+8]}, vi${K}x${ABC[C:C+8]});
	i${K} += ${CHANNEL_TILE};

	$for C in range(0, CHANNEL_TILE, 8):
	$if K == 0:
	v128_t vprod${ABC[C:C+8]} = wasm_i16x8_mul(vi${K}x${ABC[C:C+8]}, vk${K}x${ABC[C:C+8]});
	$elif K % 2 == 0 or K + 1 == KERNEL_TILE or not ADD16:
	vprod${ABC[C:C+8]} = wasm_i16x8_mul(vi${K}x${ABC[C:C+8]}, vk${K}x${ABC[C:C+8]});
	$else:
	vprod${ABC[C:C+8]} = wasm_i16x8_add(vprod${ABC[C:C+8]}, wasm_i16x8_mul(vi${K}x${ABC[C:C+8]}, vk${K}x${ABC[C:C+8]}));

	$if not ADD16 or K % 2 == 1 or K + 1 == KERNEL_TILE:
	$for C in range(0, CHANNEL_TILE, 8):
	vacc${ABC[C:C+4]} = wasm_i32x4_add(vacc${ABC[C:C+4]}, ${WASM_X32X4_EXTEND_LOW_X16X8}(vprod${ABC[C:C+8]}));
	vacc${ABC[C+4:C+8]} = wasm_i32x4_add(vacc${ABC[C+4:C+8]}, ${WASM_X32X4_EXTEND_HIGH_X16X8}(vprod${ABC[C:C+8]}));

	$if DATATYPE == "QU8":
	$for C in range(0, CHANNEL_TILE, 8):
	vacc${ABC[C:C+4]} = wasm_i32x4_sub(vacc${ABC[C:C+4]}, wasm_i32x4_mul(wasm_u32x4_extend_low_u16x8(vsumx${ABC[C:C+8]}), vkernel_zero_point));
	vacc${ABC[C+4:C+8]} = wasm_i32x4_sub(vacc${ABC[C+4:C+8]}, wasm_i32x4_mul(wasm_u32x4_extend_high_u16x8(vsumx${ABC[C:C+8]}), vkernel_zero_point));

	w = (const void) ((uintptr_t) w + ${CHANNEL_TILE} sizeof(int32_t) + ${KERNEL_TILE * CHANNEL_TILE} * sizeof(${XINT8_T}));

	$for C in range(0, CHANNEL_TILE, 4):
	vacc${ABC[C:C+4]} = wasm_f32x4_convert_i32x4(vacc${ABC[C:C+4]});

	$if DATATYPE == "QC8":
	const v128_t vscale${ABC[0:4]} = wasm_v128_load(w);
	$for C in range(4, CHANNEL_TILE, 4):
	const v128_t vscale${ABC[C:C+4]} = wasm_v128_load((const float*) w + ${C});
	w = (const void) ((const float) w + ${CHANNEL_TILE});

	$for C in range(0, CHANNEL_TILE, 4):
	vacc${ABC[C:C+4]} = wasm_f32x4_mul(vacc${ABC[C:C+4]}, vscale${ABC[C:C+4]});
	$else:
	const v128_t vscale = wasm_v128_load64_splat(params->${PARAMS_STRUCT}.scale);
	$for C in range(0, CHANNEL_TILE, 4):
	vacc${ABC[C:C+4]} = wasm_f32x4_mul(vacc${ABC[C:C+4]}, vscale);

	const v128_t vmagic_bias = wasm_v128_load64_splat(params->${PARAMS_STRUCT}.magic_bias);
	$for C in range(0, CHANNEL_TILE, 4):
	vacc${ABC[C:C+4]} = wasm_f32x4_add(vacc${ABC[C:C+4]}, vmagic_bias);

	const v128_t vmagic_min = wasm_v128_load64_splat(params->${PARAMS_STRUCT}.magic_min);
	$for C in range(0, CHANNEL_TILE, 4):
	vacc${ABC[C:C+4]} = wasm_i32x4_max(vacc${ABC[C:C+4]}, vmagic_min);

	const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->${PARAMS_STRUCT}.magic_bias_less_output_zero_point);
	$for C in range(0, CHANNEL_TILE, 4):
	vacc${ABC[C:C+4]} = wasm_i32x4_sub(vacc${ABC[C:C+4]}, vmagic_bias_less_output_zero_point);

	$for C in range(0, CHANNEL_TILE, 8):
	v128_t vout${ABC[C:C+8]} = wasm_i16x8_narrow_i32x4(vacc${ABC[C:C+4]}, vacc${ABC[C+4:C+8]});

	$for C in range(0, CHANNEL_TILE, 16):
	$if C + 8 < CHANNEL_TILE:
	v128_t vout${ABC[C:C+16]} = ${WASM_X8X16_NARROW_I16X8}(vout${ABC[C:C+8]}, vout${ABC[C+8:C+16]});
	$else:
	v128_t vout${ABC[C:C+8]}${ABC[C:C+8]} = ${WASM_X8X16_NARROW_I16X8}(vout${ABC[C:C+8]}, vout${ABC[C:C+8]});

	const v128_t voutput_max = wasm_v128_load64_splat(params->${PARAMS_STRUCT}.output_max);
	$for C in range(0, CHANNEL_TILE, 16):
	$if C + 8 < CHANNEL_TILE:
	vout${ABC[C:C+16]} = ${WASM_X8X16_MIN}(vout${ABC[C:C+16]}, voutput_max);
	$else:
	vout${ABC[C:C+8]}${ABC[C:C+8]} = ${WASM_X8X16_MIN}(vout${ABC[C:C+8]}${ABC[C:C+8]}, voutput_max);

	$if CHANNEL_TILE > 8:
	wasm_v128_store(output, vout${ABC[0:16]});
	$else:
	((double) output) = wasm_f64x2_extract_lane(vout${ABC[0:8]}${ABC[0:8]}, 0);
	$for C in range(16, CHANNEL_TILE, 16):
	$if C + 8 < CHANNEL_TILE:
	wasm_v128_store(output + ${C}, vout${ABC[C:C+16]});
	$else:
	((double) (output + ${C})) = wasm_f64x2_extract_lane(vout${ABC[C:C+8]}${ABC[C:C+8]}, 0);
	output += ${CHANNEL_TILE};
	}
	if XNN_UNLIKELY(c != 0) {
	$if CHANNEL_TILE > 8:
	const ${XINT8_T}* k = (const ${XINT8_T}) ((uintptr_t) w + ${CHANNEL_TILE} sizeof(int32_t));
	${"do " if CHANNEL_TILE > 8 else ""}{
	v128_t vacc${ABC[0:4]} = wasm_v128_load(w);
	v128_t vacc${ABC[4:8]} = wasm_v128_load((const void) ((uintptr_t) w + 4 sizeof(int32_t)));

	$for K in range(KERNEL_TILE):

	const v128_t vi${K}x${ABC[0:8]} = ${WASM_X16X8_LOAD8X8}(i${K});
	$if CHANNEL_TILE > 8:
	$if K == 0:
	const v128_t vk${K}x${ABC[0:8]} = ${WASM_X16X8_LOAD8X8}(k);
	$else:
	const v128_t vk${K}x${ABC[0:8]} = ${WASM_X16X8_LOAD8X8}((const void) (k + ${K CHANNEL_TILE}));
	$else:
	const v128_t vk${K}x${ABC[0:8]} = ${WASM_X16X8_LOAD8X8}((const void) ((uintptr_t) w + ${CHANNEL_TILE} sizeof(int32_t) + ${K * CHANNEL_TILE} * sizeof(${XINT8_T})));
	$if DATATYPE == "QU8":
	$if K == 1:
	v128_t vsumx${ABC[0:8]} = wasm_i16x8_add(vi0x${ABC[0:8]}, vi1x${ABC[0:8]});
	$elif K > 1:
	vsumx${ABC[0:8]} = wasm_i16x8_add(vsumx${ABC[0:8]}, vi${K}x${ABC[0:8]});
	$if CHANNEL_TILE > 8:
	i${K} += 8;

	$if K == 0:
	v128_t vprod${ABC[0:8]} = wasm_i16x8_mul(vi${K}x${ABC[0:8]}, vk${K}x${ABC[0:8]});
	$elif K % 2 == 0 or K + 1 == KERNEL_TILE or not ADD16:
	vprod${ABC[0:8]} = wasm_i16x8_mul(vi${K}x${ABC[0:8]}, vk${K}x${ABC[0:8]});
	$else:
	vprod${ABC[0:8]} = wasm_i16x8_add(vprod${ABC[0:8]}, wasm_i16x8_mul(vi${K}x${ABC[0:8]}, vk${K}x${ABC[0:8]}));

	$if not ADD16 or K % 2 == 1 or K + 1 == KERNEL_TILE:
	vacc${ABC[0:4]} = wasm_i32x4_add(vacc${ABC[0:4]}, ${WASM_X32X4_EXTEND_LOW_X16X8}(vprod${ABC[0:8]}));
	vacc${ABC[4:8]} = wasm_i32x4_add(vacc${ABC[4:8]}, ${WASM_X32X4_EXTEND_HIGH_X16X8}(vprod${ABC[0:8]}));

	$if CHANNEL_TILE > 8:
	k += 8;

	$if DATATYPE == "QU8":
	vacc${ABC[0:4]} = wasm_i32x4_sub(vacc${ABC[0:4]}, wasm_i32x4_mul(wasm_u32x4_extend_low_u16x8(vsumx${ABC[0:8]}), vkernel_zero_point));
	vacc${ABC[4:8]} = wasm_i32x4_sub(vacc${ABC[4:8]}, wasm_i32x4_mul(wasm_u32x4_extend_high_u16x8(vsumx${ABC[0:8]}), vkernel_zero_point));

	vacc${ABC[0:4]} = wasm_f32x4_convert_i32x4(vacc${ABC[0:4]});
	vacc${ABC[4:8]} = wasm_f32x4_convert_i32x4(vacc${ABC[4:8]});

	$if DATATYPE == "QC8":
	const v128_t vscale${ABC[0:4]} = wasm_v128_load((const float) ((uintptr_t) w + ${CHANNEL_TILE} sizeof(int32_t) + ${CHANNEL_TILE * KERNEL_TILE} * sizeof(${XINT8_T})));
	const v128_t vscale${ABC[4:8]} = wasm_v128_load((const float) ((uintptr_t) w + ${CHANNEL_TILE} sizeof(int32_t) + ${CHANNEL_TILE * KERNEL_TILE} * sizeof(${XINT8_T}) + 4 * sizeof(float)));

	vacc${ABC[0:4]} = wasm_f32x4_mul(vacc${ABC[0:4]}, vscale${ABC[0:4]});
	vacc${ABC[4:8]} = wasm_f32x4_mul(vacc${ABC[4:8]}, vscale${ABC[4:8]});
	$else:
	const v128_t vscale = wasm_v128_load64_splat(params->${PARAMS_STRUCT}.scale);
	vacc${ABC[0:4]} = wasm_f32x4_mul(vacc${ABC[0:4]}, vscale);
	vacc${ABC[4:8]} = wasm_f32x4_mul(vacc${ABC[4:8]}, vscale);

	const v128_t vmagic_bias = wasm_v128_load64_splat(params->${PARAMS_STRUCT}.magic_bias);
	vacc${ABC[0:4]} = wasm_f32x4_add(vacc${ABC[0:4]}, vmagic_bias);
	vacc${ABC[4:8]} = wasm_f32x4_add(vacc${ABC[4:8]}, vmagic_bias);

	const v128_t vmagic_min = wasm_v128_load64_splat(params->${PARAMS_STRUCT}.magic_min);
	vacc${ABC[0:4]} = wasm_i32x4_max(vacc${ABC[0:4]}, vmagic_min);
	vacc${ABC[4:8]} = wasm_i32x4_max(vacc${ABC[4:8]}, vmagic_min);

	const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->${PARAMS_STRUCT}.magic_bias_less_output_zero_point);
	vacc${ABC[0:4]} = wasm_i32x4_sub(vacc${ABC[0:4]}, vmagic_bias_less_output_zero_point);
	vacc${ABC[4:8]} = wasm_i32x4_sub(vacc${ABC[4:8]}, vmagic_bias_less_output_zero_point);

	v128_t vout${ABC[0:8]} = wasm_i16x8_narrow_i32x4(vacc${ABC[0:4]}, vacc${ABC[4:8]});
	v128_t vout${ABC[0:8]}${ABC[0:8]} = ${WASM_X8X16_NARROW_I16X8}(vout${ABC[0:8]}, vout${ABC[0:8]});

	const v128_t voutput_max = wasm_v128_load64_splat(params->${PARAMS_STRUCT}.output_max);
	vout${ABC[0:8]}${ABC[0:8]} = ${WASM_X8X16_MIN}(vout${ABC[0:8]}${ABC[0:8]}, voutput_max);

	$if CHANNEL_TILE > 8:
	w = (const void) ((uintptr_t) w + 8 sizeof(int32_t));

	$if CHANNEL_TILE > 8:
	if XNN_LIKELY(c >= 8) {
	((double) output) = wasm_f64x2_extract_lane(vout${ABC[0:8]}${ABC[0:8]}, 0);
	output += 8;
	c -= 8;
	} else {
	if (c & 4) {
	((float) output) = wasm_f32x4_extract_lane(vout${ABC[0:8]}${ABC[0:8]}, 0);
	vout${ABC[0:8]}${ABC[0:8]} = wasm_u64x2_shr(vout${ABC[0:8]}${ABC[0:8]}, 32);
	output += 4;
	}
	uint32_t vout${ABC[0:4]} = wasm_i32x4_extract_lane(vout${ABC[0:8]}${ABC[0:8]}, 0);
	if (c & 2) {
	((uint16_t) output) = (uint16_t) vout${ABC[0:4]};
	vout${ABC[0:4]} >>= 16;
	output += 2;
	}
	if (c & 1) {
	*output = (${XINT8_T}) vout${ABC[0:4]};
	output += 1;
	}
	c = 0;
	}
	$else:
	if (c & 4) {
	((float) output) = wasm_f32x4_extract_lane(vout${ABC[0:8]}${ABC[0:8]}, 0);
	vout${ABC[0:8]}${ABC[0:8]} = wasm_u64x2_shr(vout${ABC[0:8]}${ABC[0:8]}, 32);
	output += 4;
	}
	uint32_t vout${ABC[0:4]} = wasm_i32x4_extract_lane(vout${ABC[0:8]}${ABC[0:8]}, 0);
	if (c & 2) {
	((uint16_t) output) = (uint16_t) vout${ABC[0:4]};
	vout${ABC[0:4]} >>= 16;
	output += 2;
	}
	if (c & 1) {
	*output = (${XINT8_T}) vout${ABC[0:4]};
	output += 1;
	}
	}${" while (c != 0);" if CHANNEL_TILE > 8 else ""}
	}

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