Skip to content
GitLab
Commit f4f60f60 authored by Frank Galligan's avatar Frank Galligan
Browse files

Neon: Update mbfilter if all vectors follow one branch. 

Change the mbfilter Neon code from executing both branches if all
vectors follow only one branch.

The code is about 5% faster when executing only one branch and about
1% slower when executing both branches.

-PS5: Remove local stack space from mbfilter.

Change-Id: I6a23f9b318a9f4568a2718b4c9348db988fe2182
parent 30bac896
v1.14.0-linphone 1.4.X experimental feature/update_to_v1.9.0-linphone feature/uwp_nuget forest frame_parallel highbitdepth indianrunnerduck javanwhistlingduck khakicampbell linphone linphone-android linphone-old longtailedduck m31-baseline m49-2623 m52-2743 m54-2840 m56-2924 m66-3359 m68-3440 mandarinduck mcw mcw2 nextgen nextgenv2 pcs-2013 playground sandbox/Jingning/experimental sandbox/Jingning/transcode sandbox/Jingning/vpx sandbox/aconverse@google.com/ansbench sandbox/debargha/playground sandbox/hkuang/frame_parallel sandbox/hkuang@google.com/decode sandbox/jimbankoski@google.com/proposed-aom sandbox/jingning@google.com/decoder_test_suite sandbox/jingning@google.com/experimental sandbox/jzern@google.com/test sandbox/wangch@google.com/vp9 sandbox/yaowu@google.com/mergeaom stable-vp9-decoder v1.12.0-linphone v1.6.1_linphone v1.7.0-linphone v1.9.0-linphone v1.9.0 v1.9.0-rc1 v1.8.2 v1.8.1 v1.8.0 v1.7.0 v1.6.1 v1.6.0 v1.5.0 v1.4.0 v1.3.0
No related merge requests found
Hide whitespace changes
Inline Side-by-side
Showing
with 83 additions and 12 deletions
vp9/common/arm/neon/vp9_loopfilter_neon.asm
+ 83
− 12
View file @ f4f60f60
......@@ -163,7 +163,7 @@ end_vp9_lf_v_edge
; necessary load, transpose (if necessary) and store. The function does not use
; registers d8-d15.
;
; r0-r3 PRESERVE
; r0-r3, r12 PRESERVE
; d0 blimit
; d1 limit
; d2 thresh
......@@ -275,14 +275,14 @@ end_vp9_lf_v_edge
; sp const uint8_t *thresh,
; sp+4 int count
|vp9_mbloop_filter_horizontal_edge_neon| PROC
push {lr}
push {r4-r5, lr}
ldr r12, [sp,#8] ; load count
ldr r12, [sp,#16] ; load count
cmp r12, #0
beq end_vp9_mblf_h_edge
vld1.8 {d0[]}, [r2] ; duplicate *blimit
ldr r2, [sp, #4] ; load thresh
ldr r2, [sp, #12] ; load thresh
vld1.8 {d1[]}, [r3] ; duplicate *limit
vld1.8 {d2[]}, [r2] ; duplicate *thresh
......@@ -317,7 +317,7 @@ count_mblf_h_loop
bne count_mblf_h_loop
end_vp9_mblf_h_edge
pop {pc}
pop {r4-r5, pc}
ENDP ; |vp9_mbloop_filter_horizontal_edge_neon|
......@@ -335,14 +335,14 @@ end_vp9_mblf_h_edge
; sp const uint8_t *thresh,
; sp+4 int count
|vp9_mbloop_filter_vertical_edge_neon| PROC
push {lr}
push {r4-r5, lr}
ldr r12, [sp,#8] ; load count
ldr r12, [sp,#16] ; load count
cmp r12, #0
beq end_vp9_mblf_v_edge
vld1.8 {d0[]}, [r2] ; duplicate *blimit
ldr r2, [sp, #4] ; load thresh
ldr r2, [sp, #12] ; load thresh
vld1.8 {d1[]}, [r3] ; duplicate *limit
vld1.8 {d2[]}, [r2] ; duplicate *thresh
......@@ -404,7 +404,7 @@ count_mblf_v_loop
bne count_mblf_v_loop
end_vp9_mblf_v_edge
pop {pc}
pop {r4-r5, pc}
ENDP ; |vp9_mbloop_filter_vertical_edge_neon|
; void vp9_mbloop_filter_neon();
......@@ -412,7 +412,7 @@ end_vp9_mblf_v_edge
; necessary load, transpose (if necessary) and store. The function does not use
; registers d8-d15.
;
; r0-r3 PRESERVE
; r0-r3, r12 PRESERVE
; d0 blimit
; d1 limit
; d2 thresh
......@@ -471,6 +471,19 @@ end_vp9_mblf_v_edge
vand d19, d19, d24 ; mask
vand d20, d20, d19 ; flat & mask
; This instruction will truncate the "flat & mask" masks down to 4 bits
; each to fit into one 32 bit arm register. The values are stored in
; q10.64[0].
vshrn.u16 d30, q10, #4
vmov.u32 r4, d30[0] ; flat & mask 4bits
adds r5, r4, #1 ; Check for all 1's
beq power_branch_only
cmp r4, #0 ; Check for 0, set flag for later
; hevmask
vcgt.u8 d21, d21, d2 ; (abs(p1 - p0) > thresh)*-1
vcgt.u8 d22, d22, d2 ; (abs(q1 - q0) > thresh)*-1
......@@ -511,8 +524,6 @@ end_vp9_mblf_v_edge
vshr.s8 d30, d30, #3 ; filter2 >>= 3
vshr.s8 d29, d29, #3 ; filter1 >>= 3
vand d20, d20, d19 ; flat & mask
vqadd.s8 d24, d24, d30 ; op0 = clamp(ps0 + filter2)
vqsub.s8 d23, d23, d29 ; oq0 = clamp(qs0 - filter1)
......@@ -523,6 +534,8 @@ end_vp9_mblf_v_edge
vqadd.s8 d25, d25, d29 ; op1 = clamp(ps1 + filter)
vqsub.s8 d26, d26, d29 ; oq1 = clamp(qs1 - filter)
beq filter_branch_only
veor d24, d24, d22 ; *f_op0 = u^0x80
veor d23, d23, d22 ; *f_oq0 = u^0x80
veor d25, d25, d22 ; *f_op1 = u^0x80
......@@ -588,6 +601,64 @@ end_vp9_mblf_v_edge
vbif d7, d17, d20 ; oq2 |= oq2 & ~(flat & mask)
bx lr
power_branch_only
vmov.u8 d27, #3
vmov.u8 d21, #2
vaddl.u8 q14, d6, d7 ; op2 = p0 + q0
vmlal.u8 q14, d3, d27 ; op2 += p3 * 3
vmlal.u8 q14, d4, d21 ; op2 += p2 * 2
vaddw.u8 q14, d5 ; op2 += p1
vqrshrn.u16 d2, q14, #3 ; op2
vsubw.u8 q14, d3 ; op1 = op2 - p3
vsubw.u8 q14, d4 ; op1 -= p2
vaddw.u8 q14, d5 ; op1 += p1
vaddw.u8 q14, d16 ; op1 += q1
vqrshrn.u16 d31, q14, #3 ; op1
vsubw.u8 q14, d3 ; op0 = op1 - p3
vsubw.u8 q14, d5 ; op0 -= p1
vaddw.u8 q14, d6 ; op0 += p0
vaddw.u8 q14, d17 ; op0 += q2
vqrshrn.u16 d21, q14, #3 ; op0
vsubw.u8 q14, d3 ; oq0 = op0 - p3
vsubw.u8 q14, d6 ; oq0 -= p0
vaddw.u8 q14, d7 ; oq0 += q0
vaddw.u8 q14, d18 ; oq0 += q3
vqrshrn.u16 d22, q14, #3 ; oq0
vsubw.u8 q14, d4 ; oq1 = oq0 - p2
vsubw.u8 q14, d7 ; oq1 -= q0
vaddw.u8 q14, d16 ; oq1 += q1
vaddw.u8 q14, d18 ; oq1 += q3
vqrshrn.u16 d6, q14, #3 ; oq1
vsubw.u8 q14, d5 ; oq2 = oq0 - p1
vsubw.u8 q14, d16 ; oq2 -= q1
vaddw.u8 q14, d17 ; oq2 += q2
vaddw.u8 q14, d18 ; oq2 += q3
vqrshrn.u16 d7, q14, #3 ; oq2
vswp d3, d31
vswp d4, d21
vswp d5, d22
bx lr
filter_branch_only
; TODO(fgalligan): See if we can rearange registers so we do not need to
; do the 2 vswp.
vswp d2, d4 ; op2
vswp d7, d17 ; oq2
veor d4, d24, d22 ; *op0 = u^0x80
veor d5, d23, d22 ; *oq0 = u^0x80
veor d3, d25, d22 ; *op1 = u^0x80
veor d6, d26, d22 ; *oq1 = u^0x80
bx lr
ENDP ; |vp9_mbloop_filter_neon|
END
Supports Markdown
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment