xref: /aosp_15_r20/external/XNNPACK/src/qc8-dwconv/gen/up16x3-minmax-fp32-wasmsimd-mul16-add16.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)

// 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_qc8_dwconv_minmax_fp32_ukernel_up16x3__wasmsimd_mul16_add16(
    size_t channels,
    size_t output_width,
    const int8_t** input,
    const void* weights,
    int8_t* output,
    size_t input_stride,
    size_t output_increment,
    size_t input_offset,
    const int8_t* zero,
    const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(channels != 0);
  assert(output_width != 0);

  do {
    const int8_t* i0 = input[0];
    assert(i0 != NULL);
    if XNN_UNPREDICTABLE(i0 != zero) {
      i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
    }
    const int8_t* i1 = input[1];
    assert(i1 != NULL);
    if XNN_UNPREDICTABLE(i1 != zero) {
      i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
    }
    const int8_t* i2 = input[2];
    assert(i2 != NULL);
    if XNN_UNPREDICTABLE(i2 != zero) {
      i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
    }
    input = (const int8_t**) ((uintptr_t) input + input_stride);

    size_t c = channels;
    const void* w = weights;
    for (; c >= 16; c -= 16) {
      v128_t vacc0123 = wasm_v128_load(w);
      v128_t vacc4567 = wasm_v128_load((const void*) ((uintptr_t) w + 4 * sizeof(int32_t)));
      v128_t vacc89AB = wasm_v128_load((const void*) ((uintptr_t) w + 8 * sizeof(int32_t)));
      v128_t vaccCDEF = wasm_v128_load((const void*) ((uintptr_t) w + 12 * sizeof(int32_t)));


      const v128_t vi0x01234567 = wasm_i16x8_load8x8(i0);
      const v128_t vk0x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(int8_t)));
      const v128_t vi0x89ABCDEF = wasm_i16x8_load8x8(i0 + 8);
      const v128_t vk0x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 8 * sizeof(int8_t)));
      i0 += 16;

      v128_t vprod01234567 = wasm_i16x8_mul(vi0x01234567, vk0x01234567);
      v128_t vprod89ABCDEF = wasm_i16x8_mul(vi0x89ABCDEF, vk0x89ABCDEF);


      const v128_t vi1x01234567 = wasm_i16x8_load8x8(i1);
      const v128_t vk1x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(int8_t)));
      const v128_t vi1x89ABCDEF = wasm_i16x8_load8x8(i1 + 8);
      const v128_t vk1x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 24 * sizeof(int8_t)));
      i1 += 16;

      vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi1x01234567, vk1x01234567));
      vprod89ABCDEF = wasm_i16x8_add(vprod89ABCDEF, wasm_i16x8_mul(vi1x89ABCDEF, vk1x89ABCDEF));

      vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
      vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
      vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
      vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));

      const v128_t vi2x01234567 = wasm_i16x8_load8x8(i2);
      const v128_t vk2x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(int8_t)));
      const v128_t vi2x89ABCDEF = wasm_i16x8_load8x8(i2 + 8);
      const v128_t vk2x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 40 * sizeof(int8_t)));
      i2 += 16;

      vprod01234567 = wasm_i16x8_mul(vi2x01234567, vk2x01234567);
      vprod89ABCDEF = wasm_i16x8_mul(vi2x89ABCDEF, vk2x89ABCDEF);

      vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
      vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
      vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
      vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));


      w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t));

      vacc0123 = wasm_f32x4_convert_i32x4(vacc0123);
      vacc4567 = wasm_f32x4_convert_i32x4(vacc4567);
      vacc89AB = wasm_f32x4_convert_i32x4(vacc89AB);
      vaccCDEF = wasm_f32x4_convert_i32x4(vaccCDEF);

      const v128_t vscale0123 = wasm_v128_load(w);
      const v128_t vscale4567 = wasm_v128_load((const float*) w + 4);
      const v128_t vscale89AB = wasm_v128_load((const float*) w + 8);
      const v128_t vscaleCDEF = wasm_v128_load((const float*) w + 12);
      w = (const void*) ((const float*) w + 16);

      vacc0123 = wasm_f32x4_mul(vacc0123, vscale0123);
      vacc4567 = wasm_f32x4_mul(vacc4567, vscale4567);
      vacc89AB = wasm_f32x4_mul(vacc89AB, vscale89AB);
      vaccCDEF = wasm_f32x4_mul(vaccCDEF, vscaleCDEF);

      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);
      vacc89AB = wasm_f32x4_add(vacc89AB, vmagic_bias);
      vaccCDEF = wasm_f32x4_add(vaccCDEF, 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);
      vacc89AB = wasm_i32x4_max(vacc89AB, vmagic_min);
      vaccCDEF = wasm_i32x4_max(vaccCDEF, 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);
      vacc89AB = wasm_i32x4_sub(vacc89AB, vmagic_bias_less_output_zero_point);
      vaccCDEF = wasm_i32x4_sub(vaccCDEF, vmagic_bias_less_output_zero_point);

      v128_t vout01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
      v128_t vout89ABCDEF = wasm_i16x8_narrow_i32x4(vacc89AB, vaccCDEF);

      v128_t vout0123456789ABCDEF = wasm_i8x16_narrow_i16x8(vout01234567, vout89ABCDEF);

      const v128_t voutput_max = wasm_v128_load64_splat(params->fp32_wasmsimd.output_max);
      vout0123456789ABCDEF = wasm_i8x16_min(vout0123456789ABCDEF, voutput_max);

      wasm_v128_store(output, vout0123456789ABCDEF);
      output += 16;
    }
    if XNN_UNLIKELY(c != 0) {
      const int8_t* k = (const int8_t*) ((uintptr_t) w + 16 * sizeof(int32_t));
      do {
        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_i16x8_load8x8(i0);
        const v128_t vk0x01234567 = wasm_i16x8_load8x8(k);
        i0 += 8;

        v128_t vprod01234567 = wasm_i16x8_mul(vi0x01234567, vk0x01234567);


        const v128_t vi1x01234567 = wasm_i16x8_load8x8(i1);
        const v128_t vk1x01234567 = wasm_i16x8_load8x8((const void*) (k + 16));
        i1 += 8;

        vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi1x01234567, vk1x01234567));

        vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
        vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));

        const v128_t vi2x01234567 = wasm_i16x8_load8x8(i2);
        const v128_t vk2x01234567 = wasm_i16x8_load8x8((const void*) (k + 32));
        i2 += 8;

        vprod01234567 = wasm_i16x8_mul(vi2x01234567, vk2x01234567);

        vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
        vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));

        k += 8;


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

      const v128_t vscale0123 = wasm_v128_load((const float*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t)));
      const v128_t vscale4567 = wasm_v128_load((const float*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t) + 4 * sizeof(float)));

      vacc0123 = wasm_f32x4_mul(vacc0123, vscale0123);
      vacc4567 = wasm_f32x4_mul(vacc4567, vscale4567);

      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_i8x16_narrow_i16x8(vout01234567, vout01234567);

      const v128_t voutput_max = wasm_v128_load64_splat(params->fp32_wasmsimd.output_max);
      vout0123456701234567 = wasm_i8x16_min(vout0123456701234567, voutput_max);

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

      if XNN_LIKELY(c >= 8) {
        *((double*) output) = wasm_f64x2_extract_lane(vout0123456701234567, 0);
        output += 8;
        c -= 8;
      } else {
        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 = (int8_t) vout0123;
          output += 1;
        }
        c = 0;
      }
      } while (c != 0);
    }

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