// Copyright 2023 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
$CHANNEL_SUBTILE = 8
$assert CHANNEL_TILE % CHANNEL_SUBTILE == 0
$CHANNEL_ROUND = 8
$assert MIDDLE_PASS_TILE <= LAST_PASS_TILE
$assert FIRST_PASS_TILE >= 1
$assert MIDDLE_PASS_TILE >= 1
$assert LAST_PASS_TILE >= 1
#include <assert.h>

#include <wasm_simd128.h>

#include "xnnpack/dwconv.h"
#include "xnnpack/math.h"


$DATATYPE_SPEC = {"QS8": "qs8", "QC8": "qs8_qc8w", "QU8": "qu8"}[DATATYPE]
$PARAMS_STRUCT = REQUANTIZATION.lower() + "_scalar"
$PARAMS_TYPE = "union xnn_qs8_qc8w_conv_minmax_params" if DATATYPE == "QC8" else "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_SPEC}_dwconv_minmax_${REQUANTIZATION.lower()}_ukernel_${FIRST_PASS_TILE}f${MIDDLE_PASS_TILE}m${LAST_PASS_TILE}l${CHANNEL_TILE}c${CHANNEL_SUBTILE}s${CHANNEL_ROUND}r__wasmsimd_mul16${"_add16" if ADD16 else ""}(
    size_t channels,
    size_t output_width,
    const ${XINT8_T}** input,
    const void* weights,
    ${XINT8_T}* output,
    intptr_t input_stride,
    size_t output_increment,
    size_t input_offset,
    const ${XINT8_T}* zero,
    size_t kernel_size,
    int32_t* buffer,
    const ${PARAMS_TYPE} params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(channels != 0);
  assert(output_width != 0);
  assert(kernel_size > ${FIRST_PASS_TILE});

  $if DATATYPE == "QU8":
    const v128_t vkernel_zero_point = wasm_i32x4_splat(params->${PARAMS_STRUCT}.kernel_zero_point);
    XNN_FORCE_REALIZATION(vkernel_zero_point);

  $if DATATYPE != "QC8":
    const v128_t vscale = wasm_v128_load32_splat(&params->${PARAMS_STRUCT}.scale);
    XNN_FORCE_REALIZATION(vscale);

  const v128_t vmagic_bias = wasm_f32x4_const_splat(12582912.0f);
  const int32_t output_min_less_zero_point = (int32_t) params->${PARAMS_STRUCT}.output_min - (int32_t) params->${PARAMS_STRUCT}.output_zero_point;
  const v128_t vmagic_min = wasm_i32x4_splat((int32_t) float_as_uint32(12582912.0f + output_min_less_zero_point));
  const v128_t vmagic_bias_less_output_zero_point = wasm_i32x4_splat(INT32_C(0x4B400000) - (int32_t) params->${PARAMS_STRUCT}.output_zero_point);
  const v128_t voutput_max = wasm_i8x16_splat(params->${PARAMS_STRUCT}.output_max);
  XNN_FORCE_REALIZATION(vmagic_bias);
  XNN_FORCE_REALIZATION(vmagic_min);
  XNN_FORCE_REALIZATION(vmagic_bias_less_output_zero_point);
  XNN_FORCE_REALIZATION(voutput_max);

  do {
    const void* w = weights;

    // First pass to process ${FIRST_PASS_TILE} inputs.
    {
      int32_t* b = buffer;
      $for K in range(FIRST_PASS_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 += ${FIRST_PASS_TILE};

      size_t c = round_up_po2(channels, ${CHANNEL_ROUND});

      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(FIRST_PASS_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} + ${C});
            v128_t vk${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]});
            $elif K % 2 == 0 or not ADD16:
              vprod${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 == FIRST_PASS_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) + ${FIRST_PASS_TILE * CHANNEL_TILE} * sizeof(${XINT8_T}));

        wasm_v128_store(b, vacc${ABC[0:4]});
        $for C in range(4, CHANNEL_TILE, 4):
          wasm_v128_store(b + ${C}, vacc${ABC[C:C+4]});
        b += ${CHANNEL_TILE};
      }

      $if CHANNEL_TILE == 8:
          assert(c == 0);
      $else:
        if XNN_UNLIKELY(c != 0) {
          ${"do " if CHANNEL_TILE > 8 else ""}{
            v128_t vacc${ABC[0:4]} = wasm_v128_load(w);
            v128_t vacc${ABC[4:8]} = wasm_v128_load((const int32_t*) w + 4);
            $for K in range(FIRST_PASS_TILE):

              const v128_t vi${K}x${ABC[0:8]} = ${WASM_X16X8_LOAD8X8}(i${K});
              $if K == 0:
                v128_t vk${K}x${ABC[0:8]} = ${WASM_X16X8_LOAD8X8}((const void*) ((uintptr_t) w + ${CHANNEL_SUBTILE} * sizeof(int32_t)));
              $else:
                v128_t vk${K}x${ABC[0:8]} = ${WASM_X16X8_LOAD8X8}((const void*) ((uintptr_t) w + ${CHANNEL_SUBTILE} * sizeof(int32_t) + ${K * CHANNEL_SUBTILE} * sizeof(${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]});
              $elif K % 2 == 0 or not ADD16:
                vprod${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]}));

              $if not ADD16 or K % 2 == 1 or K + 1 == FIRST_PASS_TILE:
                vacc${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]}));

            $if DATATYPE == "QU8":
              vacc${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));

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

            wasm_v128_store(b, vacc${ABC[0:4]});
            wasm_v128_store(b + 4, vacc${ABC[4:8]});
            b += 8;
            c -= 8;
          }${" while (c != 0);" if CHANNEL_TILE > 8 else ""}
        }
    }


    // Middle pass to process ${MIDDLE_PASS_TILE} inputs in each iteration.
    for (size_t ks = kernel_size - ${FIRST_PASS_TILE}; ks > ${LAST_PASS_TILE}; ks -= ${MIDDLE_PASS_TILE}) {
      int32_t* b = buffer;
      $for K in range(MIDDLE_PASS_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 += ${MIDDLE_PASS_TILE};

      size_t c = round_up_po2(channels, ${CHANNEL_ROUND});

      for (; c >= ${CHANNEL_TILE}; c -= ${CHANNEL_TILE}) {
        v128_t vacc${ABC[0:4]} = wasm_v128_load(b);
        $for C in range(4, CHANNEL_TILE, 4):
          v128_t vacc${ABC[C:C+4]} = wasm_v128_load(b + ${C});
        $for K in range(MIDDLE_PASS_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} + ${C});
            v128_t vk${K}x${ABC[C:C+8]} = ${WASM_X16X8_LOAD8X8}((const void*) ((uintptr_t) w + ${K * CHANNEL_TILE + C} * sizeof(${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]});
            $elif K % 2 == 0 or not ADD16:
              vprod${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 == MIDDLE_PASS_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 + ${MIDDLE_PASS_TILE * CHANNEL_TILE} * sizeof(${XINT8_T}));

        wasm_v128_store(b, vacc${ABC[0:4]});
        $for C in range(4, CHANNEL_TILE, 4):
          wasm_v128_store(b + ${C}, vacc${ABC[C:C+4]});
        b += ${CHANNEL_TILE};
      }

      $if CHANNEL_TILE == 8:
          assert(c == 0);
      $else:
        if XNN_UNLIKELY(c != 0) {
          ${"do " if CHANNEL_TILE > 8 else ""}{
            v128_t vacc${ABC[0:4]} = wasm_v128_load(b);
            v128_t vacc${ABC[4:8]} = wasm_v128_load(b + 4);
            $for K in range(MIDDLE_PASS_TILE):

              const v128_t vi${K}x${ABC[0:8]} = ${WASM_X16X8_LOAD8X8}(i${K});
              $if K == 0:
                v128_t vk${K}x${ABC[0:8]} = ${WASM_X16X8_LOAD8X8}(w);
              $else:
                v128_t vk${K}x${ABC[0:8]} = ${WASM_X16X8_LOAD8X8}((const void*) ((uintptr_t) w + ${K * CHANNEL_SUBTILE} * sizeof(${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]});
              $elif K % 2 == 0 or not ADD16:
                vprod${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]}));

              $if not ADD16 or K % 2 == 1 or K + 1 == MIDDLE_PASS_TILE:
                vacc${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]}));

            $if DATATYPE == "QU8":
              vacc${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));

            w = (const void*) ((uintptr_t) w + ${MIDDLE_PASS_TILE * CHANNEL_SUBTILE} * sizeof(${XINT8_T}));

            wasm_v128_store(b, vacc${ABC[0:4]});
            wasm_v128_store(b + 4, vacc${ABC[4:8]});
            b += 8;
            c -= 8;
          }${" while (c != 0);" if CHANNEL_TILE > 8 else ""}
        }
    }

    // Last pass to process up to ${LAST_PASS_TILE} inputs.
    {
      const int32_t* b = buffer;
      $for K in range(LAST_PASS_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);
        }

      size_t c = channels;

      for (; c >= ${CHANNEL_TILE}; c -= ${CHANNEL_TILE}) {
        v128_t vacc${ABC[0:4]} = wasm_v128_load(b);
        $for C in range(4, CHANNEL_TILE, 4):
          v128_t vacc${ABC[C:C+4]} = wasm_v128_load(b + ${C});
        b += ${CHANNEL_TILE};
        $for K in range(LAST_PASS_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} + ${C});
            v128_t vk${K}x${ABC[C:C+8]} = ${WASM_X16X8_LOAD8X8}((const void*) ((uintptr_t) w + ${K * CHANNEL_TILE + C} * sizeof(${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]});
            $elif K % 2 == 0 or not ADD16:
              vprod${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 == LAST_PASS_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 + ${LAST_PASS_TILE * CHANNEL_TILE} * sizeof(${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:
          $for C in range(0, CHANNEL_TILE, 4):
            vacc${ABC[C:C+4]} = wasm_f32x4_mul(vacc${ABC[C:C+4]}, vscale);

        $for C in range(0, CHANNEL_TILE, 4):
          vacc${ABC[C:C+4]} = wasm_f32x4_add(vacc${ABC[C:C+4]}, vmagic_bias);

        $for C in range(0, CHANNEL_TILE, 4):
          vacc${ABC[C:C+4]} = wasm_i32x4_max(vacc${ABC[C:C+4]}, vmagic_min);

        $for C in range(0, CHANNEL_TILE, 4):
          vacc${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]});

        $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]});
        $else:
          wasm_v128_store64_lane(output, vout${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:
            wasm_v128_store64_lane(output + ${C}, vout${ABC[C:C+8]}${ABC[C:C+8]}, 0);
        output += ${CHANNEL_TILE};
      }

      if XNN_UNLIKELY(c != 0) {
        ${"do " if CHANNEL_TILE > 8 else ""}{
          v128_t vacc${ABC[0:4]} = wasm_v128_load(b);
          v128_t vacc${ABC[4:8]} = wasm_v128_load(b + 4);
          $if CHANNEL_TILE > 8:
            b += 8;
          $for K in range(LAST_PASS_TILE):

            const v128_t vi${K}x${ABC[0:8]} = ${WASM_X16X8_LOAD8X8}(i${K});
            $if K == 0:
              v128_t vk${K}x${ABC[0:8]} = ${WASM_X16X8_LOAD8X8}((const void*) ((uintptr_t) w));
            $else:
              v128_t vk${K}x${ABC[0:8]} = ${WASM_X16X8_LOAD8X8}((const void*) ((uintptr_t) w + ${K * CHANNEL_SUBTILE} * sizeof(${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]});
            $elif K % 2 == 0 or not ADD16:
              vprod${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]}));

            $if not ADD16 or K % 2 == 1 or K + 1 == LAST_PASS_TILE:
              vacc${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]}));

          $if DATATYPE == "QU8":
            vacc${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 + ${CHANNEL_SUBTILE * LAST_PASS_TILE} * sizeof(${XINT8_T})));
            const v128_t vscale${ABC[4:8]} = wasm_v128_load((const float*) ((uintptr_t) w + ${CHANNEL_SUBTILE * LAST_PASS_TILE} * sizeof(${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]});
          $else:
            vacc${ABC[0:4]} = wasm_f32x4_mul(vacc${ABC[0:4]}, vscale);
            vacc${ABC[4:8]} = wasm_f32x4_mul(vacc${ABC[4:8]}, vscale);

          $if CHANNEL_TILE > 8:
            $if DATATYPE == "QC8":
              w = (void*) ((uintptr_t) w + ${LAST_PASS_TILE * CHANNEL_SUBTILE} + 8 * sizeof(float));
            $else:
              w = (void*) ((uintptr_t) w + ${LAST_PASS_TILE * CHANNEL_SUBTILE});

          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);

          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);

          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]});

          vout${ABC[0:8]}${ABC[0:8]} = ${WASM_X8X16_MIN}(vout${ABC[0:8]}${ABC[0:8]}, voutput_max);

          $if CHANNEL_TILE > 8:
            if XNN_LIKELY(c >= 8) {
              wasm_v128_store64_lane(output, vout${ABC[0:8]}${ABC[0:8]}, 0);
              output += 8;
              c -= 8;
            } else {
              if (c & 4) {
                wasm_v128_store32_lane(output, vout${ABC[0:8]}${ABC[0:8]}, 0);
                vout${ABC[0:8]}${ABC[0:8]} = wasm_u64x2_shr(vout${ABC[0:8]}${ABC[0:8]}, 32);
                output += 4;
              }
              if (c & 2) {
                wasm_v128_store16_lane(output, vout${ABC[0:8]}${ABC[0:8]}, 0);
                vout${ABC[0:8]}${ABC[0:8]} = wasm_u32x4_shr(vout${ABC[0:8]}${ABC[0:8]}, 16);
                output += 2;
              }
              if (c & 1) {
                wasm_v128_store8_lane(output, vout${ABC[0:8]}${ABC[0:8]}, 0);
                output += 1;
              }
              c = 0;
            }
          $else:
            if (c & 4) {
              wasm_v128_store32_lane(output, vout${ABC[0:8]}${ABC[0:8]}, 0);
              vout${ABC[0:8]}${ABC[0:8]} = wasm_u64x2_shr(vout${ABC[0:8]}${ABC[0:8]}, 32);
              output += 4;
            }
            if (c & 2) {
              wasm_v128_store16_lane(output, vout${ABC[0:8]}${ABC[0:8]}, 0);
              vout${ABC[0:8]}${ABC[0:8]} = wasm_u32x4_shr(vout${ABC[0:8]}${ABC[0:8]}, 16);
              output += 2;
            }
            if (c & 1) {
              wasm_v128_store8_lane(output, vout${ABC[0:8]}${ABC[0:8]}, 0);
              output += 1;
            }
        }${" while (c != 0);" if CHANNEL_TILE > 8 else ""}
      }
    }

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