src/f16-ibilinear/fma3.c.in

// 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.

$assert CHANNEL_TILE % 8 == 0
$assert CHANNEL_TILE >= 8
$assert PIXEL_TILE == 1
$ABC = "456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
#include <assert.h>

#include <immintrin.h>

#include <xnnpack/common.h>
#include <xnnpack/ibilinear.h>
#include <xnnpack/intrinsics-polyfill.h>


void xnn_f16_ibilinear_ukernel__fma3_c${CHANNEL_TILE}${"" if PIXEL_TILE == 1 else "x%d" % PIXEL_TILE}(
    size_t output_pixels,
    size_t channels,
    const void** restrict input,
    size_t input_offset,
    const void* restrict weights,
    void* restrict output,
    size_t output_increment) XNN_OOB_READS
{
  assert(output_pixels != 0);
  assert(channels != 0);
  assert(channels % sizeof(uint16_t) == 0);

  uint16_t* o = (uint16_t*) output;
  do {
    const uint16_t* i0 = (const uint16_t*) ((uintptr_t) input[0] + input_offset);
    const uint16_t* i1 = (const uint16_t*) ((uintptr_t) input[1] + input_offset);
    const uint16_t* i2 = (const uint16_t*) ((uintptr_t) input[2] + input_offset);
    const uint16_t* i3 = (const uint16_t*) ((uintptr_t) input[3] + input_offset);
    input += 4;

    const __m256 valphahv = _mm256_cvtph_ps(_mm_castps_si128(_mm_broadcast_ss(weights)));
    const __m256 valphah = _mm256_permute_ps(valphahv, _MM_SHUFFLE(2, 0, 2, 0));
    const __m256 valphav = _mm256_permute_ps(valphahv, _MM_SHUFFLE(3, 1, 3, 1));
    weights = (const uint16_t*) weights + 2;

    size_t c = channels;
    $if CHANNEL_TILE > 8:
      for (; c >= ${CHANNEL_TILE} * sizeof(uint16_t); c -= ${CHANNEL_TILE} * sizeof(uint16_t)) {
        const __m256 vtl${ABC[0:8]} = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0));
        const __m256 vtr${ABC[0:8]} = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1));
        const __m256 vbl${ABC[0:8]} = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2));
        const __m256 vbr${ABC[0:8]} = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i3));
        $for C in range(8, CHANNEL_TILE, 8):
          const __m256 vtl${ABC[C:C+8]} = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i0 + ${C})));
          const __m256 vtr${ABC[C:C+8]} = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i1 + ${C})));
          const __m256 vbl${ABC[C:C+8]} = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i2 + ${C})));
          const __m256 vbr${ABC[C:C+8]} = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i3 + ${C})));
        i0 += ${CHANNEL_TILE};
        i1 += ${CHANNEL_TILE};
        i2 += ${CHANNEL_TILE};
        i3 += ${CHANNEL_TILE};

        $for C in range(0, CHANNEL_TILE, 8):
          const __m256 vtd${ABC[C:C+8]} = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_sub_ps(vtr${ABC[C:C+8]}, vtl${ABC[C:C+8]}), _MM_FROUND_TO_NEAREST_INT));
          const __m256 vbd${ABC[C:C+8]} = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_sub_ps(vbr${ABC[C:C+8]}, vbl${ABC[C:C+8]}), _MM_FROUND_TO_NEAREST_INT));

        $for C in range(0, CHANNEL_TILE, 8):
          const __m256 vt${ABC[C:C+8]} = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vtd${ABC[C:C+8]}, valphah, vtl${ABC[C:C+8]}), _MM_FROUND_TO_NEAREST_INT));
          const __m256 vb${ABC[C:C+8]} = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vbd${ABC[C:C+8]}, valphah, vbl${ABC[C:C+8]}), _MM_FROUND_TO_NEAREST_INT));

        $for C in range(0, CHANNEL_TILE, 8):
          const __m256 vd${ABC[C:C+8]} = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_sub_ps(vb${ABC[C:C+8]}, vt${ABC[C:C+8]}), _MM_FROUND_TO_NEAREST_INT));

        $for C in range(0, CHANNEL_TILE, 8):
          const __m128i vo${ABC[C:C+8]} = _mm256_cvtps_ph(_mm256_fmadd_ps(vd${ABC[C:C+8]}, valphav, vt${ABC[C:C+8]}), _MM_FROUND_TO_NEAREST_INT);

        _mm_storeu_si128((__m128i*) o, vo${ABC[0:8]});
        $for C in range(8, CHANNEL_TILE, 8):
          _mm_storeu_si128((__m128i*) (o + ${C}), vo${ABC[C:C+8]});
        o += ${CHANNEL_TILE};
      }
    for (; c >= 8 * sizeof(uint16_t); c -= 8 * sizeof(uint16_t)) {
      const __m256 vtl = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0));
      i0 += 8;
      const __m256 vtr = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1));
      i1 += 8;
      const __m256 vbl = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2));
      i2 += 8;
      const __m256 vbr = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i3));
      i3 += 8;

      const __m256 vtd = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_sub_ps(vtr, vtl), _MM_FROUND_TO_NEAREST_INT));
      const __m256 vbd = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_sub_ps(vbr, vbl), _MM_FROUND_TO_NEAREST_INT));

      const __m256 vt = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vtd, valphah, vtl), _MM_FROUND_TO_NEAREST_INT));
      const __m256 vb = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vbd, valphah, vbl), _MM_FROUND_TO_NEAREST_INT));

      const __m256 vd = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_sub_ps(vb, vt), _MM_FROUND_TO_NEAREST_INT));

      const __m128i vo = _mm256_cvtps_ph(_mm256_fmadd_ps(vd, valphav, vt), _MM_FROUND_TO_NEAREST_INT);

      _mm_storeu_si128((__m128i*) o, vo);
      o += 8;
    }
    if XNN_UNLIKELY(c != 0) {
      const __m256 vtl = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0));
      i0 += 8;
      const __m256 vtr = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1));
      i1 += 8;
      const __m256 vbl = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2));
      i2 += 8;
      const __m256 vbr = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i3));
      i3 += 8;

      const __m256 vtd = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_sub_ps(vtr, vtl), _MM_FROUND_TO_NEAREST_INT));
      const __m256 vbd = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_sub_ps(vbr, vbl), _MM_FROUND_TO_NEAREST_INT));

      const __m256 vt = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vtd, valphah, vtl), _MM_FROUND_TO_NEAREST_INT));
      const __m256 vb = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vbd, valphah, vbl), _MM_FROUND_TO_NEAREST_INT));

      const __m256 vd = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_sub_ps(vb, vt), _MM_FROUND_TO_NEAREST_INT));

      __m128i vo = _mm256_cvtps_ph(_mm256_fmadd_ps(vd, valphav, vt), _MM_FROUND_TO_NEAREST_INT);
      if (c & (4 * sizeof(uint16_t))) {
        _mm_storel_epi64((__m128i*) o, vo);
        vo = _mm_unpackhi_epi64(vo, vo);
        o += 4;
      }
      if (c & (2 * sizeof(uint16_t))) {
        _mm_storeu_si32(o, vo);
        vo = _mm_srli_epi64(vo, 32);
        o += 2;
      }
      if (c & (1 * sizeof(uint16_t))) {
        *o = (uint16_t) _mm_extract_epi16(vo, 0);
        o += 1;
      }
    }

    o = (uint16_t*) ((uintptr_t) o + output_increment);
  } while (--output_pixels != 0);
}