// Auto-generated file. Do not edit!
//   Template: src/qs8-dwconv/unipass-wasmsimd-mul16.c.in
//   Generator: tools/xngen
//
// 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.

#include <assert.h>

#include <wasm_simd128.h>

#include <xnnpack/dwconv.h>


void xnn_qu8_dwconv_minmax_fp32_ukernel_up8x9__wasmsimd_mul16(
    size_t channels,
    size_t output_width,
    const uint8_t** input,
    const void* weights,
    uint8_t* output,
    size_t input_stride,
    size_t output_increment,
    size_t input_offset,
    const uint8_t* zero,
    const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(channels != 0);
  assert(output_width != 0);

  const v128_t vkernel_zero_point = wasm_u32x4_load16x4(params->fp32_wasmsimd.kernel_zero_point);
  do {
    const uint8_t* i0 = input[0];
    assert(i0 != NULL);
    if XNN_UNPREDICTABLE(i0 != zero) {
      i0 = (const uint8_t*) ((uintptr_t) i0 + input_offset);
    }
    const uint8_t* i1 = input[1];
    assert(i1 != NULL);
    if XNN_UNPREDICTABLE(i1 != zero) {
      i1 = (const uint8_t*) ((uintptr_t) i1 + input_offset);
    }
    const uint8_t* i2 = input[2];
    assert(i2 != NULL);
    if XNN_UNPREDICTABLE(i2 != zero) {
      i2 = (const uint8_t*) ((uintptr_t) i2 + input_offset);
    }
    const uint8_t* i3 = input[3];
    assert(i3 != NULL);
    if XNN_UNPREDICTABLE(i3 != zero) {
      i3 = (const uint8_t*) ((uintptr_t) i3 + input_offset);
    }
    const uint8_t* i4 = input[4];
    assert(i4 != NULL);
    if XNN_UNPREDICTABLE(i4 != zero) {
      i4 = (const uint8_t*) ((uintptr_t) i4 + input_offset);
    }
    const uint8_t* i5 = input[5];
    assert(i5 != NULL);
    if XNN_UNPREDICTABLE(i5 != zero) {
      i5 = (const uint8_t*) ((uintptr_t) i5 + input_offset);
    }
    const uint8_t* i6 = input[6];
    assert(i6 != NULL);
    if XNN_UNPREDICTABLE(i6 != zero) {
      i6 = (const uint8_t*) ((uintptr_t) i6 + input_offset);
    }
    const uint8_t* i7 = input[7];
    assert(i7 != NULL);
    if XNN_UNPREDICTABLE(i7 != zero) {
      i7 = (const uint8_t*) ((uintptr_t) i7 + input_offset);
    }
    const uint8_t* i8 = input[8];
    assert(i8 != NULL);
    if XNN_UNPREDICTABLE(i8 != zero) {
      i8 = (const uint8_t*) ((uintptr_t) i8 + input_offset);
    }
    input = (const uint8_t**) ((uintptr_t) input + input_stride);

    size_t c = channels;
    const void* w = weights;
    for (; c >= 8; c -= 8) {
      v128_t vacc0123 = wasm_v128_load(w);
      v128_t vacc4567 = wasm_v128_load((const void*) ((uintptr_t) w + 4 * sizeof(int32_t)));


      const v128_t vi0x01234567 = wasm_u16x8_load8x8(i0);
      const v128_t vk0x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(uint8_t)));
      i0 += 8;

      v128_t vprod01234567 = wasm_i16x8_mul(vi0x01234567, vk0x01234567);

      vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
      vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));

      const v128_t vi1x01234567 = wasm_u16x8_load8x8(i1);
      const v128_t vk1x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(uint8_t)));
      v128_t vsumx01234567 = wasm_i16x8_add(vi0x01234567, vi1x01234567);
      i1 += 8;

      vprod01234567 = wasm_i16x8_mul(vi1x01234567, vk1x01234567);

      vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
      vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));

      const v128_t vi2x01234567 = wasm_u16x8_load8x8(i2);
      const v128_t vk2x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(uint8_t)));
      vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi2x01234567);
      i2 += 8;

      vprod01234567 = wasm_i16x8_mul(vi2x01234567, vk2x01234567);

      vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
      vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));

      const v128_t vi3x01234567 = wasm_u16x8_load8x8(i3);
      const v128_t vk3x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(uint8_t)));
      vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi3x01234567);
      i3 += 8;

      vprod01234567 = wasm_i16x8_mul(vi3x01234567, vk3x01234567);

      vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
      vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));

      const v128_t vi4x01234567 = wasm_u16x8_load8x8(i4);
      const v128_t vk4x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(uint8_t)));
      vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi4x01234567);
      i4 += 8;

      vprod01234567 = wasm_i16x8_mul(vi4x01234567, vk4x01234567);

      vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
      vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));

      const v128_t vi5x01234567 = wasm_u16x8_load8x8(i5);
      const v128_t vk5x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(uint8_t)));
      vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi5x01234567);
      i5 += 8;

      vprod01234567 = wasm_i16x8_mul(vi5x01234567, vk5x01234567);

      vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
      vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));

      const v128_t vi6x01234567 = wasm_u16x8_load8x8(i6);
      const v128_t vk6x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(uint8_t)));
      vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi6x01234567);
      i6 += 8;

      vprod01234567 = wasm_i16x8_mul(vi6x01234567, vk6x01234567);

      vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
      vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));

      const v128_t vi7x01234567 = wasm_u16x8_load8x8(i7);
      const v128_t vk7x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(uint8_t)));
      vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi7x01234567);
      i7 += 8;

      vprod01234567 = wasm_i16x8_mul(vi7x01234567, vk7x01234567);

      vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
      vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));

      const v128_t vi8x01234567 = wasm_u16x8_load8x8(i8);
      const v128_t vk8x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(uint8_t)));
      vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi8x01234567);
      i8 += 8;

      vprod01234567 = wasm_i16x8_mul(vi8x01234567, vk8x01234567);

      vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
      vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));

      vacc0123 = wasm_i32x4_sub(vacc0123, wasm_i32x4_mul(wasm_u32x4_extend_low_u16x8(vsumx01234567), vkernel_zero_point));
      vacc4567 = wasm_i32x4_sub(vacc4567, wasm_i32x4_mul(wasm_u32x4_extend_high_u16x8(vsumx01234567), vkernel_zero_point));

      w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(uint8_t));

      vacc0123 = wasm_f32x4_convert_i32x4(vacc0123);
      vacc4567 = wasm_f32x4_convert_i32x4(vacc4567);

      const v128_t vscale = wasm_v128_load64_splat(params->fp32_wasmsimd.scale);
      vacc0123 = wasm_f32x4_mul(vacc0123, vscale);
      vacc4567 = wasm_f32x4_mul(vacc4567, vscale);

      const v128_t vmagic_bias = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias);
      vacc0123 = wasm_f32x4_add(vacc0123, vmagic_bias);
      vacc4567 = wasm_f32x4_add(vacc4567, vmagic_bias);

      const v128_t vmagic_min = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_min);
      vacc0123 = wasm_i32x4_max(vacc0123, vmagic_min);
      vacc4567 = wasm_i32x4_max(vacc4567, vmagic_min);

      const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias_less_output_zero_point);
      vacc0123 = wasm_i32x4_sub(vacc0123, vmagic_bias_less_output_zero_point);
      vacc4567 = wasm_i32x4_sub(vacc4567, vmagic_bias_less_output_zero_point);

      v128_t vout01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);

      v128_t vout0123456701234567 = wasm_u8x16_narrow_i16x8(vout01234567, vout01234567);

      const v128_t voutput_max = wasm_v128_load64_splat(params->fp32_wasmsimd.output_max);
      vout0123456701234567 = wasm_u8x16_min(vout0123456701234567, voutput_max);

      *((double*) output) = wasm_f64x2_extract_lane(vout0123456701234567, 0);
      output += 8;
    }
    if XNN_UNLIKELY(c != 0) {
      {
        v128_t vacc0123 = wasm_v128_load(w);
        v128_t vacc4567 = wasm_v128_load((const void*) ((uintptr_t) w + 4 * sizeof(int32_t)));


        const v128_t vi0x01234567 = wasm_u16x8_load8x8(i0);
        const v128_t vk0x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(uint8_t)));

        v128_t vprod01234567 = wasm_i16x8_mul(vi0x01234567, vk0x01234567);

        vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
        vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));

        const v128_t vi1x01234567 = wasm_u16x8_load8x8(i1);
        const v128_t vk1x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(uint8_t)));
        v128_t vsumx01234567 = wasm_i16x8_add(vi0x01234567, vi1x01234567);

        vprod01234567 = wasm_i16x8_mul(vi1x01234567, vk1x01234567);

        vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
        vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));

        const v128_t vi2x01234567 = wasm_u16x8_load8x8(i2);
        const v128_t vk2x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(uint8_t)));
        vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi2x01234567);

        vprod01234567 = wasm_i16x8_mul(vi2x01234567, vk2x01234567);

        vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
        vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));

        const v128_t vi3x01234567 = wasm_u16x8_load8x8(i3);
        const v128_t vk3x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(uint8_t)));
        vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi3x01234567);

        vprod01234567 = wasm_i16x8_mul(vi3x01234567, vk3x01234567);

        vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
        vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));

        const v128_t vi4x01234567 = wasm_u16x8_load8x8(i4);
        const v128_t vk4x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(uint8_t)));
        vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi4x01234567);

        vprod01234567 = wasm_i16x8_mul(vi4x01234567, vk4x01234567);

        vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
        vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));

        const v128_t vi5x01234567 = wasm_u16x8_load8x8(i5);
        const v128_t vk5x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(uint8_t)));
        vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi5x01234567);

        vprod01234567 = wasm_i16x8_mul(vi5x01234567, vk5x01234567);

        vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
        vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));

        const v128_t vi6x01234567 = wasm_u16x8_load8x8(i6);
        const v128_t vk6x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(uint8_t)));
        vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi6x01234567);

        vprod01234567 = wasm_i16x8_mul(vi6x01234567, vk6x01234567);

        vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
        vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));

        const v128_t vi7x01234567 = wasm_u16x8_load8x8(i7);
        const v128_t vk7x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(uint8_t)));
        vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi7x01234567);

        vprod01234567 = wasm_i16x8_mul(vi7x01234567, vk7x01234567);

        vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
        vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));

        const v128_t vi8x01234567 = wasm_u16x8_load8x8(i8);
        const v128_t vk8x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(uint8_t)));
        vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi8x01234567);

        vprod01234567 = wasm_i16x8_mul(vi8x01234567, vk8x01234567);

        vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
        vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));


      vacc0123 = wasm_i32x4_sub(vacc0123, wasm_i32x4_mul(wasm_u32x4_extend_low_u16x8(vsumx01234567), vkernel_zero_point));
      vacc4567 = wasm_i32x4_sub(vacc4567, wasm_i32x4_mul(wasm_u32x4_extend_high_u16x8(vsumx01234567), vkernel_zero_point));

      vacc0123 = wasm_f32x4_convert_i32x4(vacc0123);
      vacc4567 = wasm_f32x4_convert_i32x4(vacc4567);

      const v128_t vscale = wasm_v128_load64_splat(params->fp32_wasmsimd.scale);
      vacc0123 = wasm_f32x4_mul(vacc0123, vscale);
      vacc4567 = wasm_f32x4_mul(vacc4567, vscale);

      const v128_t vmagic_bias = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias);
      vacc0123 = wasm_f32x4_add(vacc0123, vmagic_bias);
      vacc4567 = wasm_f32x4_add(vacc4567, vmagic_bias);

      const v128_t vmagic_min = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_min);
      vacc0123 = wasm_i32x4_max(vacc0123, vmagic_min);
      vacc4567 = wasm_i32x4_max(vacc4567, vmagic_min);

      const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias_less_output_zero_point);
      vacc0123 = wasm_i32x4_sub(vacc0123, vmagic_bias_less_output_zero_point);
      vacc4567 = wasm_i32x4_sub(vacc4567, vmagic_bias_less_output_zero_point);

      v128_t vout01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
      v128_t vout0123456701234567 = wasm_u8x16_narrow_i16x8(vout01234567, vout01234567);

      const v128_t voutput_max = wasm_v128_load64_splat(params->fp32_wasmsimd.output_max);
      vout0123456701234567 = wasm_u8x16_min(vout0123456701234567, voutput_max);


      if (c & 4) {
        *((float*) output) = wasm_f32x4_extract_lane(vout0123456701234567, 0);
        vout0123456701234567 = wasm_u64x2_shr(vout0123456701234567, 32);
        output += 4;
      }
      uint32_t vout0123 = wasm_i32x4_extract_lane(vout0123456701234567, 0);
      if (c & 2) {
        *((uint16_t*) output) = (uint16_t) vout0123;
        vout0123 >>= 16;
        output += 2;
      }
      if (c & 1) {
        *output = (uint8_t) vout0123;
        output += 1;
      }
      }
    }

    output = (uint8_t*) ((uintptr_t) output + output_increment);
  } while (--output_width != 0);
}