vp8/encoder/arm/neon/vp8_memcpy_neon.asm

   1 ;
   2 ;  Copyright (c) 2010 The VP8 project authors. All Rights Reserved.
   3 ;
   4 ;  Use of this source code is governed by a BSD-style license and patent
   5 ;  grant that can be found in the LICENSE file in the root of the source
   6 ;  tree. All contributing project authors may be found in the AUTHORS
   7 ;  file in the root of the source tree.
   8 ;
   9
  10
  11     EXPORT |vp8_memcpy_neon|
  12
  13     ARM
  14     REQUIRE8
  15     PRESERVE8
  16
  17     AREA ||.text||, CODE, READONLY, ALIGN=2
  18 ;=========================================
  19 ;void vp8_memcpy_neon(unsigned char *dst_ptr, unsigned char *src_ptr, int sz);
  20 |vp8_memcpy_neon| PROC
  21     ;pld                [r1]                        ;preload pred data
  22     ;pld                [r1, #128]
  23     ;pld                [r1, #256]
  24     ;pld                [r1, #384]
  25
  26     mov             r12, r2, lsr #8                 ;copy 256 bytes data at one time
  27
  28 memcpy_neon_loop
  29     vld1.8          {q0, q1}, [r1]!                 ;load src data
  30     subs            r12, r12, #1
  31     vld1.8          {q2, q3}, [r1]!
  32     vst1.8          {q0, q1}, [r0]!                 ;copy to dst_ptr
  33     vld1.8          {q4, q5}, [r1]!
  34     vst1.8          {q2, q3}, [r0]!
  35     vld1.8          {q6, q7}, [r1]!
  36     vst1.8          {q4, q5}, [r0]!
  37     vld1.8          {q8, q9}, [r1]!
  38     vst1.8          {q6, q7}, [r0]!
  39     vld1.8          {q10, q11}, [r1]!
  40     vst1.8          {q8, q9}, [r0]!
  41     vld1.8          {q12, q13}, [r1]!
  42     vst1.8          {q10, q11}, [r0]!
  43     vld1.8          {q14, q15}, [r1]!
  44     vst1.8          {q12, q13}, [r0]!
  45     vst1.8          {q14, q15}, [r0]!
  46
  47     ;pld                [r1]                        ;preload pred data -- need to adjust for real device
  48     ;pld                [r1, #128]
  49     ;pld                [r1, #256]
  50     ;pld                [r1, #384]
  51
  52     bne             memcpy_neon_loop
  53
  54     ands            r3, r2, #0xff                   ;extra copy
  55     beq             done_copy_neon_loop
  56
  57 extra_copy_neon_loop
  58     vld1.8          {q0}, [r1]!                 ;load src data
  59     subs            r3, r3, #16
  60     vst1.8          {q0}, [r0]!
  61     bne             extra_copy_neon_loop
  62
  63 done_copy_neon_loop
  64     bx              lr
  65     ENDP
  66
  67     END