media/libyuv/source/compare_neon.cc

changeset 0
6474c204b198
     1.1 --- /dev/null	Thu Jan 01 00:00:00 1970 +0000
     1.2 +++ b/media/libyuv/source/compare_neon.cc	Wed Dec 31 06:09:35 2014 +0100
     1.3 @@ -0,0 +1,61 @@
     1.4 +/*
     1.5 + *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
     1.6 + *
     1.7 + *  Use of this source code is governed by a BSD-style license
     1.8 + *  that can be found in the LICENSE file in the root of the source
     1.9 + *  tree. An additional intellectual property rights grant can be found
    1.10 + *  in the file PATENTS. All contributing project authors may
    1.11 + *  be found in the AUTHORS file in the root of the source tree.
    1.12 + */
    1.13 +
    1.14 +#include "libyuv/basic_types.h"
    1.15 +
    1.16 +#ifdef __cplusplus
    1.17 +namespace libyuv {
    1.18 +extern "C" {
    1.19 +#endif
    1.20 +
    1.21 +#if !defined(LIBYUV_DISABLE_NEON) && defined(__ARM_NEON__)
    1.22 +
    1.23 +uint32 SumSquareError_NEON(const uint8* src_a, const uint8* src_b, int count) {
    1.24 +  volatile uint32 sse;
    1.25 +  asm volatile (
    1.26 +    "vmov.u8    q8, #0                         \n"
    1.27 +    "vmov.u8    q10, #0                        \n"
    1.28 +    "vmov.u8    q9, #0                         \n"
    1.29 +    "vmov.u8    q11, #0                        \n"
    1.30 +
    1.31 +    ".p2align  2                               \n"
    1.32 +  "1:                                          \n"
    1.33 +    "vld1.8     {q0}, [%0]!                    \n"
    1.34 +    "vld1.8     {q1}, [%1]!                    \n"
    1.35 +    "subs       %2, %2, #16                    \n"
    1.36 +    "vsubl.u8   q2, d0, d2                     \n"
    1.37 +    "vsubl.u8   q3, d1, d3                     \n"
    1.38 +    "vmlal.s16  q8, d4, d4                     \n"
    1.39 +    "vmlal.s16  q9, d6, d6                     \n"
    1.40 +    "vmlal.s16  q10, d5, d5                    \n"
    1.41 +    "vmlal.s16  q11, d7, d7                    \n"
    1.42 +    "bgt        1b                             \n"
    1.43 +
    1.44 +    "vadd.u32   q8, q8, q9                     \n"
    1.45 +    "vadd.u32   q10, q10, q11                  \n"
    1.46 +    "vadd.u32   q11, q8, q10                   \n"
    1.47 +    "vpaddl.u32 q1, q11                        \n"
    1.48 +    "vadd.u64   d0, d2, d3                     \n"
    1.49 +    "vmov.32    %3, d0[0]                      \n"
    1.50 +    : "+r"(src_a),
    1.51 +      "+r"(src_b),
    1.52 +      "+r"(count),
    1.53 +      "=r"(sse)
    1.54 +    :
    1.55 +    : "memory", "cc", "q0", "q1", "q2", "q3", "q8", "q9", "q10", "q11");
    1.56 +  return sse;
    1.57 +}
    1.58 +
    1.59 +#endif  // __ARM_NEON__
    1.60 +
    1.61 +#ifdef __cplusplus
    1.62 +}  // extern "C"
    1.63 +}  // namespace libyuv
    1.64 +#endif

mercurial