vp8/common/arm/neon/dc_only_idct_add_neon.asm

   1 ;
   2 ;  Copyright (c) 2010 The WebM 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_dc_only_idct_add_neon|
  12     ARM
  13     REQUIRE8
  14     PRESERVE8
  15
  16     AREA ||.text||, CODE, READONLY, ALIGN=2
  17 ;void vp8_dc_only_idct_add_neon(short input_dc, unsigned char *pred_ptr,
  18 ;                               unsigned char *dst_ptr, int pitch, int stride)
  19 ; r0  input_dc
  20 ; r1  pred_ptr
  21 ; r2  dst_ptr
  22 ; r3  pitch
  23 ; sp  stride
  24 |vp8_dc_only_idct_add_neon| PROC
  25     add             r0, r0, #4
  26     asr             r0, r0, #3
  27     ldr             r12, [sp]
  28     vdup.16         q0, r0
  29
  30     vld1.32         {d2[0]}, [r1], r3
  31     vld1.32         {d2[1]}, [r1], r3
  32     vld1.32         {d4[0]}, [r1], r3
  33     vld1.32         {d4[1]}, [r1]
  34
  35     vaddw.u8        q1, q0, d2
  36     vaddw.u8        q2, q0, d4
  37
  38     vqmovun.s16     d2, q1
  39     vqmovun.s16     d4, q2
  40
  41     vst1.32         {d2[0]}, [r2], r12
  42     vst1.32         {d2[1]}, [r2], r12
  43     vst1.32         {d4[0]}, [r2], r12
  44     vst1.32         {d4[1]}, [r2]
  45
  46     bx             lr
  47
  48     ENDP
  49     END