xref: /aosp_15_r20/external/XNNPACK/src/qc8-dwconv/gen/up1x3-minmax-fp32-scalar-fmagic.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)

// Auto-generated file. Do not edit!
//   Template: src/qs8-dwconv/unipass-scalar.c.in
//   Generator: tools/xngen
//
// Copyright 2021 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 <xnnpack/dwconv.h>
#include <xnnpack/math.h>
#include <xnnpack/unaligned.h>


void xnn_qc8_dwconv_minmax_fp32_ukernel_up1x3__scalar_fmagic(
    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)])
{
  assert(channels != 0);
  assert(output_width != 0);

  const float voutput_min_less_zero_point = params->fp32_scalar_fmagic.output_min_less_zero_point;
  const float voutput_max_less_zero_point = params->fp32_scalar_fmagic.output_max_less_zero_point;
  const float vmagic_bias = params->fp32_scalar_fmagic.magic_bias;
  const int32_t vmagic_bias_less_output_zero_point = params->fp32_scalar_fmagic.magic_bias_less_output_zero_point;
  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;
    do {
      int32_t vacc = unaligned_load_s32(w);

      const int32_t vi0 = (int32_t) *i0++;
      const int32_t vk0 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[0];
      vacc += vi0 * vk0;
      const int32_t vi1 = (int32_t) *i1++;
      const int32_t vk1 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[1];
      vacc += vi1 * vk1;
      const int32_t vi2 = (int32_t) *i2++;
      const int32_t vk2 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[2];
      vacc += vi2 * vk2;

      w = (const void*) ((uintptr_t) w + sizeof(int32_t) + 3 * sizeof(int8_t));

      const float vscale = unaligned_load_f32(w);
      w = (const void*) ((const float*) w + 1);
      float vfpacc = (float) vacc * vscale;

      vfpacc = math_max_f32(vfpacc, voutput_min_less_zero_point);
      vfpacc = math_min_f32(vfpacc, voutput_max_less_zero_point);
      vfpacc += vmagic_bias;
      int32_t vout = (int32_t) float_as_uint32(vfpacc) - vmagic_bias_less_output_zero_point;

      *output++ = (int8_t) vout;
    } while (--c != 0);

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