ref: 467a0b99abb8bbf9c64cc517eff78b96aa81818c
parent: c75f3993c00772b6ec5ee00a1e6cbde91243c016
author: Johann <johannkoenig@google.com>
date: Thu Aug 19 07:29:21 EDT 2010
fix armv6 simpleloop filter test cases were causing a crash because the count was being read incorrectly. after fixing that, noticed that the output was not matching. fixed that. Change-Id: Idb0edb887736bd566a3cf6d4aa1a03ea8d20eb27
--- a/vp8/common/arm/armv6/simpleloopfilter_v6.asm
+++ b/vp8/common/arm/armv6/simpleloopfilter_v6.asm
@@ -63,23 +63,22 @@
;-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-
stmdb sp!, {r4 - r11, lr}- sub src, src, pstep, lsl #1 ; move src pointer down by 2 lines
+ ldr r12, [r3] ; limit
+ ldr r3, [src, -pstep, lsl #1] ; p1
- ldr r12, [r3], #4 ; limit
- ldr r3, [src], pstep ; p1
+ ldr r9, [sp, #40] ; count for 8-in-parallel
+ ldr r4, [src, -pstep] ; p0
- ldr r9, [sp, #36] ; count for 8-in-parallel
- ldr r4, [src], pstep ; p0
-
- ldr r7, [r2], #4 ; flimit
- ldr r5, [src], pstep ; q0
+ ldr r7, [r2] ; flimit
+ ldr r5, [src] ; q0
ldr r2, c0x80808080
- ldr r6, [src] ; q1
+ ldr r6, [src, pstep] ; q1
uadd8 r7, r7, r7 ; flimit * 2
- mov r9, r9, lsl #1 ; 4-in-parallel
+ mov r9, r9, lsl #1 ; double the count. we're doing 4 at a time
uadd8 r12, r7, r12 ; flimit * 2 + limit
+ mov lr, #0
|simple_hnext8|
; vp8_simple_filter_mask() function
@@ -89,22 +88,19 @@
uqsub8 r10, r4, r5 ; p0 - q0
uqsub8 r11, r5, r4 ; q0 - p0
orr r8, r8, r7 ; abs(p1 - q1)
- ldr lr, c0x7F7F7F7F ; 01111111 mask
orr r10, r10, r11 ; abs(p0 - q0)
- and r8, lr, r8, lsr #1 ; abs(p1 - q1) / 2
+ uhadd8 r8, r8, lr ; abs(p1 - q2) >> 1
uqadd8 r10, r10, r10 ; abs(p0 - q0) * 2
- mvn lr, #0 ; r10 == -1
+ ; STALL waiting on r10
uqadd8 r10, r10, r8 ; abs(p0 - q0)*2 + abs(p1 - q1)/2
- ; STALL waiting on r10 :(
- uqsub8 r10, r10, r12 ; compare to flimit
- mov r8, #0
-
- usub8 r10, r8, r10 ; use usub8 instead of ssub8
- ; STALL (maybe?) when are flags set? :/
- sel r10, lr, r8 ; filter mask: lr
-
+ ; STALL waiting on r10
+ mvn r8, #0
+ uqsub8 r10, r10, r12 ; compare to flimit. need to do this twice because uqsub8 doesn't set GE flags
+ ; and usub8 doesn't saturate
+ usub8 r10, lr, r10 ; set GE flags for each byte
+ sel r10, r8, lr ; filter mask: F or 0
cmp r10, #0
- beq simple_hskip_filter ; skip filtering
+ beq simple_hskip_filter ; skip filtering if we're &ing with 0s. would just write out the same values
;vp8_simple_filter() function
@@ -113,56 +109,46 @@
eor r4, r4, r2 ; p0 offset to convert to a signed value
eor r5, r5, r2 ; q0 offset to convert to a signed value
- qsub8 r3, r3, r6 ; vp8_filter (r3) = vp8_signed_char_clamp(p1-q1)
- qsub8 r6, r5, r4 ; vp8_filter = vp8_signed_char_clamp(vp8_filter + 3 * ( q0 - p0))
+ qsub8 r3, r3, r6 ; vp8_signed_char_clamp(p1-q1)
+ qsub8 r6, r5, r4 ; vp8_signed_char_clamp(q0-p0)
+ qadd8 r3, r3, r6 ; += q0-p0
+ qadd8 r3, r3, r6 ; += q0-p0
+ qadd8 r3, r3, r6 ; p1-q1 + 3*(q0-p0))
+ and r3, r3, r10 ; &= mask
- qadd8 r3, r3, r6
- ldr r8, c0x03030303 ; r8 = 3
-
- qadd8 r3, r3, r6
ldr r7, c0x04040404
+ ldr r8, c0x03030303
- qadd8 r3, r3, r6
- and r3, r3, lr ; vp8_filter &= mask;
-
;save bottom 3 bits so that we round one side +4 and the other +3
+ qadd8 r7 , r3 , r7 ; Filter1 (r3) = vp8_signed_char_clamp(vp8_filter+4)
qadd8 r8 , r3 , r8 ; Filter2 (r8) = vp8_signed_char_clamp(vp8_filter+3)
- qadd8 r3 , r3 , r7 ; Filter1 (r3) = vp8_signed_char_clamp(vp8_filter+4)
- mov r7, #0
- shadd8 r8 , r8 , r7 ; Filter2 >>= 3
- shadd8 r3 , r3 , r7 ; Filter1 >>= 3
- shadd8 r8 , r8 , r7
- shadd8 r3 , r3 , r7
- shadd8 r8 , r8 , r7 ; r8: Filter2
- shadd8 r3 , r3 , r7 ; r7: filter1
+ mov r3, #0
+ shadd8 r7 , r7 , r3
+ shadd8 r8 , r8 , r3
+ shadd8 r7 , r7 , r3
+ shadd8 r8 , r8 , r3
+ shadd8 r7 , r7 , r3 ; Filter1 >>= 3
+ shadd8 r8 , r8 , r3 ; Filter2 >>= 3
- ;calculate output
- sub src, src, pstep, lsl #1
+ qsub8 r5 ,r5, r7 ; u = vp8_signed_char_clamp(q0 - Filter1)
qadd8 r4, r4, r8 ; u = vp8_signed_char_clamp(p0 + Filter2)
- qsub8 r5 ,r5, r3 ; u = vp8_signed_char_clamp(q0 - Filter1)
- eor r4, r4, r2 ; *op0 = u^0x80
- str r4, [src], pstep ; store op0 result
eor r5, r5, r2 ; *oq0 = u^0x80
- str r5, [src], pstep ; store oq0 result
+ str r5, [src] ; store oq0 result
+ eor r4, r4, r2 ; *op0 = u^0x80
+ str r4, [src, -pstep] ; store op0 result
|simple_hskip_filter|
- add src, src, #4
- sub src, src, pstep
- sub src, src, pstep, lsl #1
subs r9, r9, #1
+ addne src, src, #4 ; next row
- ;pld [src]
- ;pld [src, pstep]
- ;pld [src, pstep, lsl #1]
+ ldrne r3, [src, -pstep, lsl #1] ; p1
+ ldrne r4, [src, -pstep] ; p0
+ ldrne r5, [src] ; q0
+ ldrne r6, [src, pstep] ; q1
- ldrne r3, [src], pstep ; p1
- ldrne r4, [src], pstep ; p0
- ldrne r5, [src], pstep ; q0
- ldrne r6, [src] ; q1
-
bne simple_hnext8
ldmia sp!, {r4 - r11, pc}@@ -174,9 +160,9 @@
;-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-
stmdb sp!, {r4 - r11, lr}- ldr r12, [r2], #4 ; r12: flimit
+ ldr r12, [r2] ; r12: flimit
ldr r2, c0x80808080
- ldr r7, [r3], #4 ; limit
+ ldr r7, [r3] ; limit
; load soure data to r7, r8, r9, r10
ldrh r3, [src, #-2]
@@ -213,16 +199,15 @@
uqsub8 r10, r5, r4 ; q0 - p0
orr r7, r7, r8 ; abs(p1 - q1)
orr r9, r9, r10 ; abs(p0 - q0)
- ldr lr, c0x7F7F7F7F ; 0111 1111 mask
- uqadd8 r9, r9, r9 ; abs(p0 - q0) * 2
- and r7, lr, r7, lsr #1 ; abs(p1 - q1) / 2
mov r8, #0
+ uqadd8 r9, r9, r9 ; abs(p0 - q0) * 2
+ uhadd8 r7, r7, r8 ; abs(p1 - q1) / 2
uqadd8 r7, r7, r9 ; abs(p0 - q0)*2 + abs(p1 - q1)/2
mvn r10, #0 ; r10 == -1
uqsub8 r7, r7, r12 ; compare to flimit
usub8 r7, r8, r7
- sel r7, r10, r8 ; filter mask: lr
+ sel lr, r10, r8 ; filter mask
cmp lr, #0
beq simple_vskip_filter ; skip filtering
@@ -286,10 +271,6 @@
|simple_vskip_filter|
subs r11, r11, #1
- ;pld [src]
- ;pld [src, pstep]
- ;pld [src, pstep, lsl #1]
-
; load soure data to r7, r8, r9, r10
ldrneh r3, [src, #-2]
ldrneh r4, [src], pstep
@@ -316,7 +297,5 @@
c0x80808080 DCD 0x80808080
c0x03030303 DCD 0x03030303
c0x04040404 DCD 0x04040404
-c0x01010101 DCD 0x01010101
-c0x7F7F7F7F DCD 0x7F7F7F7F
END