vpx/vpx_dsp/x86/sum_squares_sse2.c

/*
 *  Copyright (c) 2016 The WebM project authors. All Rights Reserved.
 *
 *  Use of this source code is governed by a BSD-style license
 *  that can be found in the LICENSE file in the root of the source
 *  tree. An additional intellectual property rights grant can be found
 *  in the file PATENTS.  All contributing project authors may
 *  be found in the AUTHORS file in the root of the source tree.
 */

#include <assert.h>
#include <emmintrin.h>
#include <stdio.h>

#include "./vpx_dsp_rtcd.h"

static uint64_t vpx_sum_squares_2d_i16_4x4_sse2(const int16_t *src,
                                                int stride) {
  const __m128i v_val_0_w =
      _mm_loadl_epi64((const __m128i *)(src + 0 * stride));
  const __m128i v_val_1_w =
      _mm_loadl_epi64((const __m128i *)(src + 1 * stride));
  const __m128i v_val_2_w =
      _mm_loadl_epi64((const __m128i *)(src + 2 * stride));
  const __m128i v_val_3_w =
      _mm_loadl_epi64((const __m128i *)(src + 3 * stride));

  const __m128i v_sq_0_d = _mm_madd_epi16(v_val_0_w, v_val_0_w);
  const __m128i v_sq_1_d = _mm_madd_epi16(v_val_1_w, v_val_1_w);
  const __m128i v_sq_2_d = _mm_madd_epi16(v_val_2_w, v_val_2_w);
  const __m128i v_sq_3_d = _mm_madd_epi16(v_val_3_w, v_val_3_w);

  const __m128i v_sum_01_d = _mm_add_epi32(v_sq_0_d, v_sq_1_d);
  const __m128i v_sum_23_d = _mm_add_epi32(v_sq_2_d, v_sq_3_d);
  const __m128i v_sum_0123_d = _mm_add_epi32(v_sum_01_d, v_sum_23_d);

  const __m128i v_sum_d =
      _mm_add_epi32(v_sum_0123_d, _mm_srli_epi64(v_sum_0123_d, 32));

  return (uint64_t)_mm_cvtsi128_si32(v_sum_d);
}

// TODO(jingning): Evaluate the performance impact here.
#ifdef __GNUC__
// This prevents GCC/Clang from inlining this function into
// vpx_sum_squares_2d_i16_sse2, which in turn saves some stack
// maintenance instructions in the common case of 4x4.
__attribute__((noinline))
#endif
static uint64_t
vpx_sum_squares_2d_i16_nxn_sse2(const int16_t *src, int stride, int size) {
  int r, c;
  const __m128i v_zext_mask_q = _mm_set_epi32(0, 0xffffffff, 0, 0xffffffff);
  __m128i v_acc_q = _mm_setzero_si128();

  for (r = 0; r < size; r += 8) {
    __m128i v_acc_d = _mm_setzero_si128();

    for (c = 0; c < size; c += 8) {
      const int16_t *b = src + c;
      const __m128i v_val_0_w =
          _mm_load_si128((const __m128i *)(b + 0 * stride));
      const __m128i v_val_1_w =
          _mm_load_si128((const __m128i *)(b + 1 * stride));
      const __m128i v_val_2_w =
          _mm_load_si128((const __m128i *)(b + 2 * stride));
      const __m128i v_val_3_w =
          _mm_load_si128((const __m128i *)(b + 3 * stride));
      const __m128i v_val_4_w =
          _mm_load_si128((const __m128i *)(b + 4 * stride));
      const __m128i v_val_5_w =
          _mm_load_si128((const __m128i *)(b + 5 * stride));
      const __m128i v_val_6_w =
          _mm_load_si128((const __m128i *)(b + 6 * stride));
      const __m128i v_val_7_w =
          _mm_load_si128((const __m128i *)(b + 7 * stride));

      const __m128i v_sq_0_d = _mm_madd_epi16(v_val_0_w, v_val_0_w);
      const __m128i v_sq_1_d = _mm_madd_epi16(v_val_1_w, v_val_1_w);
      const __m128i v_sq_2_d = _mm_madd_epi16(v_val_2_w, v_val_2_w);
      const __m128i v_sq_3_d = _mm_madd_epi16(v_val_3_w, v_val_3_w);
      const __m128i v_sq_4_d = _mm_madd_epi16(v_val_4_w, v_val_4_w);
      const __m128i v_sq_5_d = _mm_madd_epi16(v_val_5_w, v_val_5_w);
      const __m128i v_sq_6_d = _mm_madd_epi16(v_val_6_w, v_val_6_w);
      const __m128i v_sq_7_d = _mm_madd_epi16(v_val_7_w, v_val_7_w);

      const __m128i v_sum_01_d = _mm_add_epi32(v_sq_0_d, v_sq_1_d);
      const __m128i v_sum_23_d = _mm_add_epi32(v_sq_2_d, v_sq_3_d);
      const __m128i v_sum_45_d = _mm_add_epi32(v_sq_4_d, v_sq_5_d);
      const __m128i v_sum_67_d = _mm_add_epi32(v_sq_6_d, v_sq_7_d);

      const __m128i v_sum_0123_d = _mm_add_epi32(v_sum_01_d, v_sum_23_d);
      const __m128i v_sum_4567_d = _mm_add_epi32(v_sum_45_d, v_sum_67_d);

      v_acc_d = _mm_add_epi32(v_acc_d, v_sum_0123_d);
      v_acc_d = _mm_add_epi32(v_acc_d, v_sum_4567_d);
    }

    v_acc_q = _mm_add_epi64(v_acc_q, _mm_and_si128(v_acc_d, v_zext_mask_q));
    v_acc_q = _mm_add_epi64(v_acc_q, _mm_srli_epi64(v_acc_d, 32));

    src += 8 * stride;
  }

  v_acc_q = _mm_add_epi64(v_acc_q, _mm_srli_si128(v_acc_q, 8));

#if ARCH_X86_64
  return (uint64_t)_mm_cvtsi128_si64(v_acc_q);
#else
  {
    uint64_t tmp;
    _mm_storel_epi64((__m128i *)&tmp, v_acc_q);
    return tmp;
  }
#endif
}

uint64_t vpx_sum_squares_2d_i16_sse2(const int16_t *src, int stride, int size) {
  // 4 elements per row only requires half an XMM register, so this
  // must be a special case, but also note that over 75% of all calls
  // are with size == 4, so it is also the common case.
  if (size == 4) {
    return vpx_sum_squares_2d_i16_4x4_sse2(src, stride);
  } else {
    // Generic case
    return vpx_sum_squares_2d_i16_nxn_sse2(src, stride, size);
  }
}