genericopenlibs/liboil/src/sad8x8_sse.c
author Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
Tue, 31 Aug 2010 16:54:36 +0300
branchRCL_3
changeset 56 acd3cd4aaceb
permissions -rw-r--r--
Revision: 201021 Kit: 201035
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
56
acd3cd4aaceb Revision: 201021
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff changeset
     1
/*
acd3cd4aaceb Revision: 201021
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff changeset
     2
 * Copyright (c) 2005
acd3cd4aaceb Revision: 201021
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff changeset
     3
 *	Eric Anholt.  All rights reserved.
acd3cd4aaceb Revision: 201021
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff changeset
     4
 *
acd3cd4aaceb Revision: 201021
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff changeset
     5
 * Redistribution and use in source and binary forms, with or without
acd3cd4aaceb Revision: 201021
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff changeset
     6
 * modification, are permitted provided that the following conditions
acd3cd4aaceb Revision: 201021
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff changeset
     7
 * are met:
acd3cd4aaceb Revision: 201021
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff changeset
     8
 * 1. Redistributions of source code must retain the above copyright
acd3cd4aaceb Revision: 201021
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff changeset
     9
 *    notice, this list of conditions and the following disclaimer.
acd3cd4aaceb Revision: 201021
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff changeset
    10
 * 2. Redistributions in binary form must reproduce the above copyright
acd3cd4aaceb Revision: 201021
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff changeset
    11
 *    notice, this list of conditions and the following disclaimer in the
acd3cd4aaceb Revision: 201021
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff changeset
    12
 *    documentation and/or other materials provided with the distribution.
acd3cd4aaceb Revision: 201021
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff changeset
    13
 *
acd3cd4aaceb Revision: 201021
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff changeset
    14
 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND
acd3cd4aaceb Revision: 201021
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff changeset
    15
 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
acd3cd4aaceb Revision: 201021
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff changeset
    16
 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
acd3cd4aaceb Revision: 201021
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff changeset
    17
 * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR BE LIABLE
acd3cd4aaceb Revision: 201021
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff changeset
    18
 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
acd3cd4aaceb Revision: 201021
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff changeset
    19
 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
acd3cd4aaceb Revision: 201021
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff changeset
    20
 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
acd3cd4aaceb Revision: 201021
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff changeset
    21
 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
acd3cd4aaceb Revision: 201021
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff changeset
    22
 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
acd3cd4aaceb Revision: 201021
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff changeset
    23
 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
acd3cd4aaceb Revision: 201021
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff changeset
    24
 * SUCH DAMAGE.
acd3cd4aaceb Revision: 201021
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff changeset
    25
 */
acd3cd4aaceb Revision: 201021
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff changeset
    26
acd3cd4aaceb Revision: 201021
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff changeset
    27
#ifdef HAVE_CONFIG_H
acd3cd4aaceb Revision: 201021
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff changeset
    28
#include "config.h"
acd3cd4aaceb Revision: 201021
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff changeset
    29
#endif
acd3cd4aaceb Revision: 201021
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff changeset
    30
#include <liboil/liboilclasses.h>
acd3cd4aaceb Revision: 201021
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff changeset
    31
#include <liboil/liboilfunction.h>
acd3cd4aaceb Revision: 201021
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff changeset
    32
#include <emmintrin.h>
acd3cd4aaceb Revision: 201021
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff changeset
    33
acd3cd4aaceb Revision: 201021
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff changeset
    34
#define SSE_FUNCTION __attribute__((force_align_arg_pointer))
acd3cd4aaceb Revision: 201021
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff changeset
    35
acd3cd4aaceb Revision: 201021
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff changeset
    36
#ifdef ENABLE_BROKEN_IMPLS
acd3cd4aaceb Revision: 201021
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff changeset
    37
union m128_int {
acd3cd4aaceb Revision: 201021
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff changeset
    38
  __m128i m128;
acd3cd4aaceb Revision: 201021
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff changeset
    39
  uint32_t i[4];
acd3cd4aaceb Revision: 201021
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff changeset
    40
  uint16_t s[8];
acd3cd4aaceb Revision: 201021
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff changeset
    41
};
acd3cd4aaceb Revision: 201021
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff changeset
    42
acd3cd4aaceb Revision: 201021
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff changeset
    43
SSE_FUNCTION static void
acd3cd4aaceb Revision: 201021
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff changeset
    44
sad8x8_u8_sse (uint32_t *dest, uint8_t *src1, int sstr1, uint8_t *src2,
acd3cd4aaceb Revision: 201021
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff changeset
    45
    int sstr2)
acd3cd4aaceb Revision: 201021
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff changeset
    46
{
acd3cd4aaceb Revision: 201021
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff changeset
    47
  int i;
acd3cd4aaceb Revision: 201021
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff changeset
    48
  __m128i sum = _mm_setzero_si128();
acd3cd4aaceb Revision: 201021
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff changeset
    49
  union m128_int sumi;
acd3cd4aaceb Revision: 201021
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff changeset
    50
acd3cd4aaceb Revision: 201021
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff changeset
    51
  for (i = 0; i < 4; i++) {
acd3cd4aaceb Revision: 201021
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff changeset
    52
    __m128i xmm0, xmm1, xmm2, xmm3;
acd3cd4aaceb Revision: 201021
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff changeset
    53
    xmm0 = _mm_loadl_epi64((__m128i *)src1);
acd3cd4aaceb Revision: 201021
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff changeset
    54
    xmm1 = _mm_loadl_epi64((__m128i *)(src1 + sstr1));
acd3cd4aaceb Revision: 201021
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff changeset
    55
    xmm2 = _mm_loadl_epi64((__m128i *)src2);
acd3cd4aaceb Revision: 201021
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff changeset
    56
    xmm3 = _mm_loadl_epi64((__m128i *)(src2 + sstr2));
acd3cd4aaceb Revision: 201021
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff changeset
    57
    xmm0 = _mm_unpacklo_epi8(xmm0, xmm1);
acd3cd4aaceb Revision: 201021
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff changeset
    58
    xmm2 = _mm_unpacklo_epi8(xmm2, xmm3);
acd3cd4aaceb Revision: 201021
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff changeset
    59
    sum = _mm_add_epi64(sum, _mm_sad_epu8(xmm0, xmm2));
acd3cd4aaceb Revision: 201021
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff changeset
    60
    src1 += 2 * sstr1;
acd3cd4aaceb Revision: 201021
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff changeset
    61
    src2 += 2 * sstr2;
acd3cd4aaceb Revision: 201021
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff changeset
    62
  }
acd3cd4aaceb Revision: 201021
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff changeset
    63
  sumi.m128 = sum;
acd3cd4aaceb Revision: 201021
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff changeset
    64
  *dest = sumi.i[0] + sumi.i[2];
acd3cd4aaceb Revision: 201021
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff changeset
    65
}
acd3cd4aaceb Revision: 201021
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff changeset
    66
OIL_DEFINE_IMPL_FULL_WRAPPER(sad8x8_u8_sse, sad8x8_u8, OIL_IMPL_FLAG_SSE2);
acd3cd4aaceb Revision: 201021
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff changeset
    67
#endif
acd3cd4aaceb Revision: 201021
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff changeset
    68