vp8/common/arm/armv6/iwalsh_v6.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
   5 ;  that can be found in the LICENSE file in the root of the source
   6 ;  tree. An additional intellectual property rights grant can be found
   7 ;  in the file PATENTS.  All contributing project authors may
   8 ;  be found in the AUTHORS file in the root of the source tree.
   9 ;
  10
  11     EXPORT |vp8_short_inv_walsh4x4_v6|
  12     EXPORT |vp8_short_inv_walsh4x4_1_v6|
  13
  14     ARM
  15     REQUIRE8
  16     PRESERVE8
  17
  18     AREA    |.text|, CODE, READONLY  ; name this block of code
  19
  20 ;short vp8_short_inv_walsh4x4_v6(short *input, short *output)
  21 |vp8_short_inv_walsh4x4_v6| PROC
  22
  23     stmdb       sp!, {r4 - r11, lr}
  24
  25     ldr         r2, [r0], #4         ; [1  |  0]
  26     ldr         r3, [r0], #4         ; [3  |  2]
  27     ldr         r4, [r0], #4         ; [5  |  4]
  28     ldr         r5, [r0], #4         ; [7  |  6]
  29     ldr         r6, [r0], #4         ; [9  |  8]
  30     ldr         r7, [r0], #4         ; [11 | 10]
  31     ldr         r8, [r0], #4         ; [13 | 12]
  32     ldr         r9, [r0]             ; [15 | 14]
  33
  34     qadd16      r10, r2, r8          ; a1 [1+13  |  0+12]
  35     qadd16      r11, r4, r6          ; b1 [5+9   |  4+8]
  36     qsub16      r12, r4, r6          ; c1 [5-9   |  4-8]
  37     qsub16      lr, r2, r8           ; d1 [1-13  |  0-12]
  38
  39     qadd16      r2, r10, r11         ; a1 + b1 [1  |  0]
  40     qadd16      r4, r12, lr          ; c1 + d1 [5  |  4]
  41     qsub16      r6, r10, r11         ; a1 - b1 [9  |  8]
  42     qsub16      r8, lr, r12          ; d1 - c1 [13 | 12]
  43
  44     qadd16      r10, r3, r9          ; a1 [3+15  |  2+14]
  45     qadd16      r11, r5, r7          ; b1 [7+11  |  6+10]
  46     qsub16      r12, r5, r7          ; c1 [7-11  |  6-10]
  47     qsub16      lr, r3, r9           ; d1 [3-15  |  2-14]
  48
  49     qadd16      r3, r10, r11         ; a1 + b1 [3  |  2]
  50     qadd16      r5, r12, lr          ; c1 + d1 [7  |  6]
  51     qsub16      r7, r10, r11         ; a1 - b1 [11 | 10]
  52     qsub16      r9, lr, r12          ; d1 - c1 [15 | 14]
  53
  54     ; first transform complete
  55
  56     qsubaddx    r10, r2, r3          ; [c1|a1] [1-2   |   0+3]
  57     qaddsubx    r11, r2, r3          ; [b1|d1] [1+2   |   0-3]
  58     qsubaddx    r12, r4, r5          ; [c1|a1] [5-6   |   4+7]
  59     qaddsubx    lr, r4, r5           ; [b1|d1] [5+6   |   4-7]
  60
  61     qaddsubx    r2, r10, r11         ; [b2|c2] [c1+d1 | a1-b1]
  62     qaddsubx    r3, r11, r10         ; [a2|d2] [b1+a1 | d1-c1]
  63     ldr         r10, c0x00030003
  64     qaddsubx    r4, r12, lr          ; [b2|c2] [c1+d1 | a1-b1]
  65     qaddsubx    r5, lr, r12          ; [a2|d2] [b1+a1 | d1-c1]
  66
  67     qadd16      r2, r2, r10          ; [b2+3|c2+3]
  68     qadd16      r3, r3, r10          ; [a2+3|d2+3]
  69     qadd16      r4, r4, r10          ; [b2+3|c2+3]
  70     qadd16      r5, r5, r10          ; [a2+3|d2+3]
  71
  72     asr         r12, r2, #3          ; [1  |  x]
  73     pkhtb       r12, r12, r3, asr #19; [1  |  0]
  74     lsl         lr, r3, #16          ; [~3 |  x]
  75     lsl         r2, r2, #16          ; [~2 |  x]
  76     asr         lr, lr, #3           ; [3  |  x]
  77     pkhtb       lr, lr, r2, asr #19  ; [3  |  2]
  78
  79     asr         r2, r4, #3           ; [5  |  x]
  80     pkhtb       r2, r2, r5, asr #19  ; [5  |  4]
  81     lsl         r3, r5, #16          ; [~7 |  x]
  82     lsl         r4, r4, #16          ; [~6 |  x]
  83     asr         r3, r3, #3           ; [7  |  x]
  84     pkhtb       r3, r3, r4, asr #19  ; [7  |  6]
  85
  86     str         r12, [r1], #4
  87     str         lr, [r1], #4
  88     str         r2, [r1], #4
  89     str         r3, [r1], #4
  90
  91     qsubaddx    r2, r6, r7           ; [c1|a1] [9-10  |  8+11]
  92     qaddsubx    r3, r6, r7           ; [b1|d1] [9+10  |  8-11]
  93     qsubaddx    r4, r8, r9           ; [c1|a1] [13-14 | 12+15]
  94     qaddsubx    r5, r8, r9           ; [b1|d1] [13+14 | 12-15]
  95
  96     qaddsubx    r6, r2, r3           ; [b2|c2] [c1+d1 | a1-b1]
  97     qaddsubx    r7, r3, r2           ; [a2|d2] [b1+a1 | d1-c1]
  98     qaddsubx    r8, r4, r5           ; [b2|c2] [c1+d1 | a1-b1]
  99     qaddsubx    r9, r5, r4           ; [a2|d2] [b1+a1 | d1-c1]
 100
 101     qadd16      r6, r6, r10          ; [b2+3|c2+3]
 102     qadd16      r7, r7, r10          ; [a2+3|d2+3]
 103     qadd16      r8, r8, r10          ; [b2+3|c2+3]
 104     qadd16      r9, r9, r10          ; [a2+3|d2+3]
 105
 106     asr         r2, r6, #3           ; [9  |  x]
 107     pkhtb       r2, r2, r7, asr #19  ; [9  |  8]
 108     lsl         r3, r7, #16          ; [~11|  x]
 109     lsl         r4, r6, #16          ; [~10|  x]
 110     asr         r3, r3, #3           ; [11 |  x]
 111     pkhtb       r3, r3, r4, asr #19  ; [11 | 10]
 112
 113     asr         r4, r8, #3           ; [13 |  x]
 114     pkhtb       r4, r4, r9, asr #19  ; [13 | 12]
 115     lsl         r5, r9, #16          ; [~15|  x]
 116     lsl         r6, r8, #16          ; [~14|  x]
 117     asr         r5, r5, #3           ; [15 |  x]
 118     pkhtb       r5, r5, r6, asr #19  ; [15 | 14]
 119
 120     str         r2, [r1], #4
 121     str         r3, [r1], #4
 122     str         r4, [r1], #4
 123     str         r5, [r1]
 124
 125     ldmia       sp!, {r4 - r11, pc}
 126     ENDP        ; |vp8_short_inv_walsh4x4_v6|
 127
 128
 129 ;short vp8_short_inv_walsh4x4_1_v6(short *input, short *output)
 130 |vp8_short_inv_walsh4x4_1_v6| PROC
 131
 132     ldrsh       r2, [r0]             ; [0]
 133     add         r2, r2, #3           ; [0] + 3
 134     asr         r2, r2, #3           ; a1 ([0]+3) >> 3
 135     lsl         r2, r2, #16          ; [a1 |  x]
 136     orr         r2, r2, r2, lsr #16  ; [a1 | a1]
 137
 138     str         r2, [r1], #4
 139     str         r2, [r1], #4
 140     str         r2, [r1], #4
 141     str         r2, [r1], #4
 142     str         r2, [r1], #4
 143     str         r2, [r1], #4
 144     str         r2, [r1], #4
 145     str         r2, [r1]
 146
 147     bx          lr
 148     ENDP        ; |vp8_short_inv_walsh4x4_1_v6|
 149
 150 ; Constant Pool
 151 c0x00030003 DCD 0x00030003
 152     END