50 |
# |
# |
51 |
#} |
#} |
52 |
|
|
53 |
|
.data |
54 |
|
.skip: |
55 |
|
.word 0,-1 |
56 |
|
.align 4 |
57 |
|
|
58 |
.text |
.text |
59 |
.global calc_cbp_ppc |
.global calc_cbp_ppc |
60 |
calc_cbp_ppc: |
calc_cbp_ppc: |
61 |
# r9 will contain coeffs addr |
# r9 will contain coeffs addr |
62 |
mr 9,3 |
mr %r9,%r3 |
63 |
# r8 is the loop counter (rows) |
# r8 is the loop counter (rows) |
64 |
li 8,5 |
li %r8,5 |
65 |
# r3 contains the result, therefore we set it to 0 |
# r3 contains the result, therefore we set it to 0 |
66 |
xor 3,3,3 |
li %r3,0 |
67 |
.loop: |
.loop: |
68 |
# CTR is the loop2 counter |
# CTR is the loop2 counter |
69 |
li 4,15 |
li %r4,16 |
70 |
mtctr 4 |
mtctr %r4 |
71 |
# r6 is coeff pointer for this line |
# r6 is coeff pointer for this line |
72 |
mr 6,9 |
mr %r6,%r9 |
73 |
|
lis %r7,.skip@ha |
74 |
|
addi %r7,%r7,.skip@l |
75 |
|
lwz %r7,0(%r7) |
76 |
.loop2: |
.loop2: |
77 |
# coeffs is a matrix of 16 bits cells |
# coeffs is a matrix of 16 bits cells |
78 |
lha 4,2(6) |
lwz %r4,0(%r6) |
79 |
lha 5,4(6) |
and %r4,%r4,%r7 |
80 |
or 4,5,4 |
li %r7,-1 |
81 |
lha 5,6(6) |
|
82 |
or 4,5,4 |
lwz %r5,4(%r6) |
|
lha 5,8(6) |
|
83 |
# or. updates CR0 |
# or. updates CR0 |
84 |
or. 4,5,4 |
or. %r4,%r5,%r4 |
85 |
# testing bit 2 (is zero) of CR0 |
# testing bit 2 (is zero) of CR0 |
86 |
bf 2,.cbp |
bf 2,.cbp |
87 |
addi 6,6,8 |
addi %r6,%r6,8 |
88 |
bdnz .loop2 |
bdnz .loop2 |
89 |
# latest line coeffs |
b .newline |
|
lha 4,2(6) |
|
|
lha 5,4(6) |
|
|
or 4,5,4 |
|
|
lha 5,6(6) |
|
|
# or. updates CR0 |
|
|
or. 4,5,4 |
|
|
# testing bit 2 (is zero) of CR0 |
|
|
bt 2,.newline |
|
90 |
.cbp: |
.cbp: |
91 |
li 4,1 |
li %r4,1 |
92 |
slw 4,4,8 |
slw %r4,%r4,%r8 |
93 |
or 3,3,4 |
or %r3,%r3,%r4 |
94 |
.newline: |
.newline: |
95 |
addi 9,9,128 |
addi %r9,%r9,128 |
96 |
# updates CR0, blabla |
# updates CR0, blabla |
97 |
subic. 8,8,1 |
subic. %r8,%r8,1 |
98 |
bf 0,.loop |
bf 0,.loop |
99 |
blr |
blr |