src/f32-clamp/avx.c

// Copyright 2019 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/clamp.h>


static const int32_t mask_table[14] = {-1, -1, -1, -1, -1, -1, -1, 0, 0, 0, 0, 0, 0, 0};

void xnn_f32_clamp_ukernel__avx(
    size_t n,
    const float* x,
    float* y,
    const union xnn_f32_output_params params[restrict static 1])
{
  assert(n != 0);
  assert(n % sizeof(float) == 0);

  const __m256 voutput_max = _mm256_broadcast_ps((const __m128*) params->sse.max);
  const __m256 voutput_min = _mm256_broadcast_ps((const __m128*) params->sse.min);

  for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
    const __m256 vx = _mm256_loadu_ps(x);
    x += 8;

    const __m256 vy = _mm256_min_ps(_mm256_max_ps(vx, voutput_min), voutput_max);

    _mm256_storeu_ps(y, vy);
    y += 8;
  }
  if (n != 0) {
    assert(n >= 1 * sizeof(float));
    assert(n <= 7 * sizeof(float));
    __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - n));

    const __m256 vx = _mm256_maskload_ps(x, vmask);

    const __m256 vy = _mm256_min_ps(_mm256_max_ps(vx, voutput_min), voutput_max);

    // _mm256_maskstore_ps(y, vmask, vy) could be used here, but triggers msan failures (probably an msan bug).
    __m128 vy_lo = _mm256_castps256_ps128(vy);
    if (n & (4 * sizeof(float))) {
      _mm_storeu_ps(y, vy_lo);
      vy_lo = _mm256_extractf128_ps(vy, 1);
      y += 4;
    }
    if (n & (2 * sizeof(float))) {
      _mm_storel_pi((__m64*) y, vy_lo);
      vy_lo = _mm_movehl_ps(vy_lo, vy_lo);
      y += 2;
    }
    if (n & (1 * sizeof(float))) {
      _mm_store_ss(y, vy_lo);
    }
  }
}