generic-library/vpx
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

Optimize vp9_tm_predictor_8x8_neon function

Browse Source 
Change-Id: Ia12aae491202098ff66366145aa0c3da38dc97e5
This commit is contained in:
Frank Galligan 2014-01-24 11:07:14 -08:00
parent 8b0537f631
commit 56a8a0b54b
1 changed files with 31 additions and 32 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

63
vp9/common/arm/neon/vp9_reconintra_neon.asm
View File

@ -353,53 +353,52 @@ loop_h
vsubl.u8 q3, d2, d0 vsubl.u8 q3, d2, d0
; Load left row by row and compute left + (above - ytop_left) ; Load left row by row and compute left + (above - ytop_left)
vld1.u8 {d6}, [r3]
; 1st row and 2nd row ; 1st row and 2nd row
ldrb r12, [r3], #1 vdup.u8 d0, d6[0]
ldrb r2, [r3], #1 vdup.u8 d1, d6[1]
vdup.u16 q1, r12 vaddw.s16 q1, q3, d0
vdup.u16 q2, r2 vaddw.s16 q2, q3, d1
vadd.s16 q1, q1, q3
vadd.s16 q2, q2, q3
vqshrun.s16 d0, q1, #0
vqshrun.s16 d1, q2, #0
vst1.64 {d0}, [r0], r1
vst1.64 {d1}, [r0], r1
; 3rd row and 4th row ; 3rd row and 4th row
ldrb r12, [r3], #1 vdup.u8 d0, d6[2]
ldrb r2, [r3], #1 vdup.u8 d1, d6[3]
vdup.u16 q1, r12 vaddw.s16 q8, q3, d0
vdup.u16 q2, r2 vaddw.s16 q9, q3, d1
vadd.s16 q1, q1, q3
vadd.s16 q2, q2, q3
vqshrun.s16 d0, q1, #0 vqshrun.s16 d0, q1, #0
vqshrun.s16 d1, q2, #0 vqshrun.s16 d1, q2, #0
vqshrun.s16 d2, q8, #0
vqshrun.s16 d3, q9, #0
vst1.64 {d0}, [r0], r1 vst1.64 {d0}, [r0], r1
vst1.64 {d1}, [r0], r1 vst1.64 {d1}, [r0], r1
vst1.64 {d2}, [r0], r1
vst1.64 {d3}, [r0], r1
; 5th row and 6th row ; 5th row and 6th row
ldrb r12, [r3], #1 vdup.u8 d0, d6[4]
ldrb r2, [r3], #1 vdup.u8 d1, d6[5]
vdup.u16 q1, r12 vaddw.s16 q1, q3, d0
vdup.u16 q2, r2 vaddw.s16 q2, q3, d1
vadd.s16 q1, q1, q3
vadd.s16 q2, q2, q3
vqshrun.s16 d0, q1, #0
vqshrun.s16 d1, q2, #0
vst1.64 {d0}, [r0], r1
vst1.64 {d1}, [r0], r1
; 7rd row and 8th row ; 7rd row and 8th row
ldrb r12, [r3], #1 vdup.u8 d0, d6[6]
ldrb r2, [r3], #1 vdup.u8 d1, d6[7]
vdup.u16 q1, r12 vaddw.s16 q8, q3, d0
vdup.u16 q2, r2 vaddw.s16 q9, q3, d1
vadd.s16 q1, q1, q3
vadd.s16 q2, q2, q3
vqshrun.s16 d0, q1, #0 vqshrun.s16 d0, q1, #0
vqshrun.s16 d1, q2, #0 vqshrun.s16 d1, q2, #0
vqshrun.s16 d2, q8, #0
vqshrun.s16 d3, q9, #0
vst1.64 {d0}, [r0], r1 vst1.64 {d0}, [r0], r1
vst1.64 {d1}, [r0], r1 vst1.64 {d1}, [r0], r1
vst1.64 {d2}, [r0], r1
vst1.64 {d3}, [r0], r1
bx lr bx lr
ENDP ; |vp9_tm_predictor_8x8_neon| ENDP ; |vp9_tm_predictor_8x8_neon|