modules/arm_neon/yuyv_i422.S

   1  @*****************************************************************************
   2  @ yuyv_i422.S : ARM NEONv1 packed to planar YUV422 conversion
   3  @*****************************************************************************
   4  @ Copyright (C) 2011 Rémi Denis-Courmont
   5  @
   6  @ This program is free software; you can redistribute it and/or modify
   7  @ it under the terms of the GNU Lesser General Public License as published by
   8  @ the Free Software Foundation; either version 2.1 of the License, or
   9  @ (at your option) any later version.
  10  @
  11  @ This program is distributed in the hope that it will be useful,
  12  @ but WITHOUT ANY WARRANTY; without even the implied warranty of
  13  @ MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  14  @ GNU Lesser General Public License for more details.
  15  @
  16  @ You should have received a copy of the GNU Lesser General Public License
  17  @ along with this program; if not, write to the Free Software Foundation,
  18  @ Inc., 51 Franklin Street, Fifth Floor, Boston MA 02110-1301, USA.
  19  @****************************************************************************/
  20
  21 #include "asm.S"
  22
  23         .syntax unified
  24 #if HAVE_AS_FPU_DIRECTIVE
  25         .fpu    neon
  26 #endif
  27         .text
  28
  29 #define I       r0
  30 #define IPAD    r1
  31 #define WIDTH   r2
  32 #define HEIGHT  r3
  33 #define Y       r4
  34 #define U       r5
  35 #define V       r6
  36 #define COUNT   ip
  37 #define YPAD    lr
  38
  39         .align 2
  40 function yuyv_i422_neon
  41         push            {r4-r6,lr}
  42         ldmia           r0,     {Y, U, V, YPAD}
  43         ldmia           r1,     {I, IPAD}
  44         cmp             HEIGHT, #0
  45         sub             YPAD,   YPAD,   WIDTH
  46         sub             IPAD,   IPAD,   WIDTH,  lsl #1
  47 1:
  48         ite             gt
  49         movsgt          COUNT,  WIDTH
  50         pople           {r4-r6,pc}
  51 2:
  52         pld             [I, #64]
  53         subs            COUNT,  COUNT,  #16
  54         vld1.u8         {q0-q1},        [I,:128]!
  55         vuzp.u8         q0,     q1
  56         @ TODO: unroll (1 cycle stall)
  57         vuzp.u8         d2,     d3
  58         vst1.u8         {q0},           [Y,:128]!
  59         vst1.u8         {d2},           [U,:64]!
  60         vst1.u8         {d3},           [V,:64]!
  61         bgt             2b
  62
  63         subs            HEIGHT, #1
  64         add             I,      I,      IPAD
  65         add             Y,      Y,      YPAD
  66         add             U,      U,      YPAD,   lsr #1
  67         add             V,      V,      YPAD,   lsr #1
  68         b               1b
  69
  70 function uyvy_i422_neon
  71         push            {r4-r6,lr}
  72         ldmia           r0,     {Y, U, V, YPAD}
  73         ldmia           r1,     {I, IPAD}
  74         cmp             HEIGHT, #0
  75         sub             YPAD,   YPAD,   WIDTH
  76         sub             IPAD,   IPAD,   WIDTH,  lsl #1
  77 1:
  78         ite             gt
  79         movsgt          COUNT,  WIDTH
  80         pople           {r4-r6,pc}
  81 2:
  82         pld             [I, #64]
  83         subs            COUNT,  COUNT,  #16
  84         vld1.u8         {q0-q1},        [I,:128]!
  85         vuzp.u8         q0,     q1
  86         vuzp.u8         d0,     d1
  87         vst1.u8         {q1},           [Y,:128]!
  88         vst1.u8         {d0},           [U,:64]!
  89         vst1.u8         {d1},           [V,:64]!
  90         bgt             2b
  91
  92         subs            HEIGHT, #1
  93         add             I,      I,      IPAD
  94         add             Y,      Y,      YPAD
  95         add             U,      U,      YPAD,   lsr #1
  96         add             V,      V,      YPAD,   lsr #1
  97         b               1b