src/f16-vrnd/f16c.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 BATCH_TILE % 8 == 0
$assert BATCH_TILE >= 8
$SIMD_TILE = BATCH_TILE // 8
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
$assert OP in ["RNDNE", "RNDZ", "RNDU", "RNDD"]
#include <assert.h>

#include <immintrin.h>

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


$_MM_FROUND_TO_FLAG = {
$  "RNDNE": "_MM_FROUND_TO_NEAREST_INT",
$  "RNDZ": "_MM_FROUND_TO_ZERO",
$  "RNDU": "_MM_FROUND_TO_POS_INF",
$  "RNDD": "_MM_FROUND_TO_NEG_INF",
$}[OP]
void xnn_f16_v${OP.lower()}_ukernel__f16c_x${BATCH_TILE}(
    size_t batch,
    const void* input,
    void* output,
    const union xnn_f16_rnd_params params[restrict XNN_MIN_ELEMENTS(1)])
{
  assert(batch != 0);
  assert(batch % sizeof(uint16_t) == 0);
  assert(input != NULL);
  assert(output != NULL);

  const uint16_t* i = (const uint16_t*) input;
  uint16_t* o = (uint16_t*) output;
  $if BATCH_TILE > 8:
    for (; batch >= ${BATCH_TILE} * sizeof(uint16_t); batch -= ${BATCH_TILE} * sizeof(uint16_t)) {
      __m256 vacc${ABC[0]} = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i));
      $for N in range(1, SIMD_TILE):
        __m256 vacc${ABC[N]} = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i + ${N * 8})));
      i += ${BATCH_TILE};

      $for N in range(SIMD_TILE):
        vacc${ABC[N]} = _mm256_round_ps(vacc${ABC[N]}, ${_MM_FROUND_TO_FLAG} | _MM_FROUND_TO_NEAREST_INT);

      _mm_storeu_si128((__m128i*) o, _mm256_cvtps_ph(vacc${ABC[0]}, _MM_FROUND_TO_NEAREST_INT));
      $for N in range(1, SIMD_TILE):
        _mm_storeu_si128((__m128i*) (o + ${N * 8}), _mm256_cvtps_ph(vacc${ABC[N]}, _MM_FROUND_TO_NEAREST_INT));
      o += ${BATCH_TILE};
    }
  for (; batch >= 8 * sizeof(uint16_t); batch -= 8 * sizeof(uint16_t)) {
    __m256 vacc = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i));
    i += 8;

    vacc = _mm256_round_ps(vacc, ${_MM_FROUND_TO_FLAG} | _MM_FROUND_TO_NEAREST_INT);

    _mm_storeu_si128((__m128i*) o, _mm256_cvtps_ph(vacc, _MM_FROUND_TO_NEAREST_INT));
    o += 8;
  }
  if XNN_UNLIKELY(batch != 0) {
    assert(batch >= 1 * sizeof(uint16_t));
    assert(batch <= 7 * sizeof(uint16_t));
    __m256 vacc = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i));
    vacc = _mm256_round_ps(vacc, ${_MM_FROUND_TO_FLAG} | _MM_FROUND_TO_NEAREST_INT);
    __m128i vh = _mm256_cvtps_ph(vacc, _MM_FROUND_TO_NEAREST_INT);
    if (batch & (4 * sizeof(uint16_t))) {
      _mm_storel_epi64((__m128i*) o, vh);
      vh = _mm_unpackhi_epi64(vh, vh);
      o += 4;
    }
    if (batch & (2 * sizeof(uint16_t))) {
      _mm_storeu_si32(o, vh);
      vh = _mm_srli_epi64(vh, 32);
      o += 2;
    }
    if (batch & (1 * sizeof(uint16_t))) {
      *o = (uint16_t) _mm_extract_epi16(vh, 0);
    }
  }
}