michael@0: /*
michael@0:  *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
michael@0:  *
michael@0:  *  Use of this source code is governed by a BSD-style license
michael@0:  *  that can be found in the LICENSE file in the root of the source
michael@0:  *  tree. An additional intellectual property rights grant can be found
michael@0:  *  in the file PATENTS. All contributing project authors may
michael@0:  *  be found in the AUTHORS file in the root of the source tree.
michael@0:  */
michael@0: 
michael@0: #include "libyuv/basic_types.h"
michael@0: 
michael@0: #ifdef __cplusplus
michael@0: namespace libyuv {
michael@0: extern "C" {
michael@0: #endif
michael@0: 
michael@0: #if !defined(LIBYUV_DISABLE_NEON) && defined(__ARM_NEON__)
michael@0: 
michael@0: uint32 SumSquareError_NEON(const uint8* src_a, const uint8* src_b, int count) {
michael@0:   volatile uint32 sse;
michael@0:   asm volatile (
michael@0:     "vmov.u8    q8, #0                         \n"
michael@0:     "vmov.u8    q10, #0                        \n"
michael@0:     "vmov.u8    q9, #0                         \n"
michael@0:     "vmov.u8    q11, #0                        \n"
michael@0: 
michael@0:     ".p2align  2                               \n"
michael@0:   "1:                                          \n"
michael@0:     "vld1.8     {q0}, [%0]!                    \n"
michael@0:     "vld1.8     {q1}, [%1]!                    \n"
michael@0:     "subs       %2, %2, #16                    \n"
michael@0:     "vsubl.u8   q2, d0, d2                     \n"
michael@0:     "vsubl.u8   q3, d1, d3                     \n"
michael@0:     "vmlal.s16  q8, d4, d4                     \n"
michael@0:     "vmlal.s16  q9, d6, d6                     \n"
michael@0:     "vmlal.s16  q10, d5, d5                    \n"
michael@0:     "vmlal.s16  q11, d7, d7                    \n"
michael@0:     "bgt        1b                             \n"
michael@0: 
michael@0:     "vadd.u32   q8, q8, q9                     \n"
michael@0:     "vadd.u32   q10, q10, q11                  \n"
michael@0:     "vadd.u32   q11, q8, q10                   \n"
michael@0:     "vpaddl.u32 q1, q11                        \n"
michael@0:     "vadd.u64   d0, d2, d3                     \n"
michael@0:     "vmov.32    %3, d0[0]                      \n"
michael@0:     : "+r"(src_a),
michael@0:       "+r"(src_b),
michael@0:       "+r"(count),
michael@0:       "=r"(sse)
michael@0:     :
michael@0:     : "memory", "cc", "q0", "q1", "q2", "q3", "q8", "q9", "q10", "q11");
michael@0:   return sse;
michael@0: }
michael@0: 
michael@0: #endif  // __ARM_NEON__
michael@0: 
michael@0: #ifdef __cplusplus
michael@0: }  // extern "C"
michael@0: }  // namespace libyuv
michael@0: #endif