src/f16-avgpool/f16-avgpool-9x-minmax-f16c-c8.c

// Copyright 2022 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 <immintrin.h>

#include <xnnpack/avgpool.h>
#include <xnnpack/intrinsics-polyfill.h>


void xnn_f16_avgpool_minmax_ukernel_9x__f16c_c8(
    size_t output_pixels,
    size_t kernel_elements,
    size_t channels,
    const void** input,
    size_t input_offset,
    const void* zero,
    void* output,
    size_t input_increment,
    size_t output_increment,
    const union xnn_f16_scaleminmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(output_pixels != 0);
  assert(kernel_elements != 0);
  assert(kernel_elements <= 9);
  assert(channels != 0);

  const __m256 vscale = _mm256_load_ps(params->avx.scale);
  const __m256 vmin = _mm256_load_ps(params->avx.min);
  const __m256 vmax = _mm256_load_ps(params->avx.max);

  uint16_t* o = (uint16_t*) output;
  do {
    const uint16_t* i0 = (const uint16_t*) input[0];
    assert(i0 != NULL);
    const uint16_t* i1 = (const uint16_t*) input[1];
    const uint16_t* i2 = (const uint16_t*) input[2];
    const uint16_t* i3 = (const uint16_t*) input[3];
    const uint16_t* i4 = (const uint16_t*) input[4];
    const uint16_t* i5 = (const uint16_t*) input[5];
    const uint16_t* i6 = (const uint16_t*) input[6];
    const uint16_t* i7 = (const uint16_t*) input[7];
    const uint16_t* i8 = (const uint16_t*) input[8];
    input = (const void**) ((uintptr_t) input + input_increment);
    if (kernel_elements < 2) {
      i1 = (const uint16_t*) zero;
    }
    assert(i1 != NULL);
    if (kernel_elements <= 2) {
      i2 = (const uint16_t*) zero;
    }
    assert(i2 != NULL);
    if (kernel_elements < 4) {
      i3 = (const uint16_t*) zero;
    }
    assert(i3 != NULL);
    if (kernel_elements <= 4) {
      i4 = (const uint16_t*) zero;
    }
    assert(i4 != NULL);
    if (kernel_elements < 6) {
      i5 = (const uint16_t*) zero;
    }
    assert(i5 != NULL);
    if (kernel_elements <= 6) {
      i6 = (const uint16_t*) zero;
    }
    assert(i6 != NULL);
    if (kernel_elements < 8) {
      i7 = (const uint16_t*) zero;
    }
    assert(i7 != NULL);
    if (kernel_elements <= 8) {
      i8 = (const uint16_t*) zero;
    }
    assert(i8 != NULL);
    if XNN_UNPREDICTABLE(i0 != zero) {
      i0 = (const uint16_t*) ((uintptr_t) i0 + input_offset);
    }
    if XNN_UNPREDICTABLE(i1 != zero) {
      i1 = (const uint16_t*) ((uintptr_t) i1 + input_offset);
    }
    if XNN_UNPREDICTABLE(i2 != zero) {
      i2 = (const uint16_t*) ((uintptr_t) i2 + input_offset);
    }
    if XNN_UNPREDICTABLE(i3 != zero) {
      i3 = (const uint16_t*) ((uintptr_t) i3 + input_offset);
    }
    if XNN_UNPREDICTABLE(i4 != zero) {
      i4 = (const uint16_t*) ((uintptr_t) i4 + input_offset);
    }
    if XNN_UNPREDICTABLE(i5 != zero) {
      i5 = (const uint16_t*) ((uintptr_t) i5 + input_offset);
    }
    if XNN_UNPREDICTABLE(i6 != zero) {
      i6 = (const uint16_t*) ((uintptr_t) i6 + input_offset);
    }
    if XNN_UNPREDICTABLE(i7 != zero) {
      i7 = (const uint16_t*) ((uintptr_t) i7 + input_offset);
    }
    if XNN_UNPREDICTABLE(i8 != zero) {
      i8 = (const uint16_t*) ((uintptr_t) i8 + input_offset);
    }

    size_t c = channels;
    while (c >= 8) {
      const __m256 vi0 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0));
      i0 += 8;
      const __m256 vi1 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1));
      i1 += 8;
      const __m256 vi2 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2));
      i2 += 8;
      const __m256 vi3 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i3));
      i3 += 8;
      const __m256 vi4 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i4));
      i4 += 8;
      const __m256 vi5 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i5));
      i5 += 8;
      const __m256 vi6 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i6));
      i6 += 8;
      const __m256 vi7 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i7));
      i7 += 8;
      const __m256 vi8 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i8));
      i8 += 8;

      const __m256 vsum01 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vi0, vi1), _MM_FROUND_TO_NEAREST_INT));
      const __m256 vsum23 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vi2, vi3), _MM_FROUND_TO_NEAREST_INT));
      const __m256 vsum45 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vi4, vi5), _MM_FROUND_TO_NEAREST_INT));
      const __m256 vsum67 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vi6, vi7), _MM_FROUND_TO_NEAREST_INT));
      const __m256 vsum018 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vsum01, vi8), _MM_FROUND_TO_NEAREST_INT));
      const __m256 vsum2345 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vsum23, vsum45), _MM_FROUND_TO_NEAREST_INT));
      const __m256 vsum01678 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vsum018, vsum67), _MM_FROUND_TO_NEAREST_INT));
      const __m256 vsum = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vsum2345, vsum01678), _MM_FROUND_TO_NEAREST_INT));

      __m256 vout = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vsum, vscale), _MM_FROUND_TO_NEAREST_INT));
      vout = _mm256_max_ps(vout, vmin);
      vout = _mm256_min_ps(vout, vmax);

      _mm_storeu_si128((__m128i*) o, _mm256_cvtps_ph(vout, _MM_FROUND_TO_NEAREST_INT));
      o += 8;

      c -= 8;
    }
    if (c != 0) {
      const __m256 vi0 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0));
      const __m256 vi1 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1));
      const __m256 vi2 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2));
      const __m256 vi3 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i3));
      const __m256 vi4 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i4));
      const __m256 vi5 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i5));
      const __m256 vi6 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i6));
      const __m256 vi7 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i7));
      const __m256 vi8 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i8));

      const __m256 vsum01 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vi0, vi1), _MM_FROUND_TO_NEAREST_INT));
      const __m256 vsum23 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vi2, vi3), _MM_FROUND_TO_NEAREST_INT));
      const __m256 vsum45 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vi4, vi5), _MM_FROUND_TO_NEAREST_INT));
      const __m256 vsum67 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vi6, vi7), _MM_FROUND_TO_NEAREST_INT));
      const __m256 vsum018 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vsum01, vi8), _MM_FROUND_TO_NEAREST_INT));
      const __m256 vsum2345 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vsum23, vsum45), _MM_FROUND_TO_NEAREST_INT));
      const __m256 vsum01678 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vsum018, vsum67), _MM_FROUND_TO_NEAREST_INT));
      const __m256 vsum = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vsum2345, vsum01678), _MM_FROUND_TO_NEAREST_INT));

      __m256 vout = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vsum, vscale), _MM_FROUND_TO_NEAREST_INT));
      vout = _mm256_max_ps(vout, vmin);
      vout = _mm256_min_ps(vout, vmax);

      __m128i vh = _mm256_cvtps_ph(vout, _MM_FROUND_TO_NEAREST_INT);
      if (c & 4) {
        _mm_storel_epi64((__m128i*) o, vh);
        vh = _mm_unpackhi_epi64(vh, vh);
        o += 4;
      }
      if (c & 2) {
        _mm_storeu_si32(o, vh);
        vh = _mm_srli_epi64(vh, 32);
        o += 2;
      }
      if (c & 1) {
        *o = (uint16_t) _mm_extract_epi16(vh, 0);
        o += 1;
      }
    }
    o = (uint16_t*) ((uintptr_t) o + output_increment);
  } while (--output_pixels != 0);
}