ref: 99c611fea60100be888363ea0e08481247de59b3
parent: 14b322e466fb7e20be2fa4b302440457704acae7
parent: b7dc9398f2b5b749a1369312ed4a4666e4e326cf
author: John Koleszar <jkoleszar@google.com>
date: Mon Sep 20 20:05:03 EDT 2010
Merge remote branch 'internal/upstream' into HEAD
--- a/vp8/common/x86/loopfilter_sse2.asm
+++ b/vp8/common/x86/loopfilter_sse2.asm
@@ -11,6 +11,8 @@
%include "vpx_ports/x86_abi_support.asm"
+; Use of pmaxub instead of psubusb to compute filter mask was seen
+; in ffvp8
%macro LFH_FILTER_MASK 1
%if %1
@@ -33,8 +35,6 @@
psubusb xmm2, xmm6 ; q3-=q2
por xmm1, xmm2 ; abs(q3-q2)
- psubusb xmm1, xmm7
-
%if %1
movdqa xmm4, [rsi+rax] ; q1
%else
@@ -49,10 +49,8 @@
psubusb xmm4, xmm6 ; q1-=q2
psubusb xmm6, xmm3 ; q2-=q1
por xmm4, xmm6 ; abs(q2-q1)
- psubusb xmm4, xmm7
+ pmaxub xmm1, xmm4
- por xmm1, xmm4
-
%if %1
movdqa xmm4, [rsi] ; q0
%else
@@ -67,10 +65,8 @@
psubusb xmm3, xmm0 ; q1-=q0
por xmm4, xmm3 ; abs(q0-q1)
movdqa t0, xmm4 ; save to t0
+ pmaxub xmm1, xmm4
- psubusb xmm4, xmm7
- por xmm1, xmm4
-
%if %1
neg rax ; negate pitch to deal with above border
@@ -95,10 +91,8 @@
psubusb xmm4, xmm2 ; p2-=p3
psubusb xmm2, xmm5 ; p3-=p2
por xmm4, xmm2 ; abs(p3 - p2)
+ pmaxub xmm1, xmm4
- psubusb xmm4, xmm7
- por xmm1, xmm4
-
%if %1
movdqa xmm4, [rsi+2*rax] ; p1
%else
@@ -113,9 +107,8 @@
psubusb xmm4, xmm5 ; p1-=p2
psubusb xmm5, xmm3 ; p2-=p1
por xmm4, xmm5 ; abs(p2 - p1)
- psubusb xmm4, xmm7
+ pmaxub xmm1, xmm4
- por xmm1, xmm4
movdqa xmm2, xmm3 ; p1
%if %1
@@ -133,8 +126,8 @@
por xmm4, xmm3 ; abs(p1 - p0)
movdqa t1, xmm4 ; save to t1
- psubusb xmm4, xmm7
- por xmm1, xmm4
+ pmaxub xmm1, xmm4
+ psubusb xmm1, xmm7
%if %1
movdqa xmm3, [rdi] ; q1
@@ -872,19 +865,18 @@
psubusb xmm0, xmm7 ; q2-q3
psubusb xmm7, xmm6 ; q3-q2
- por xmm7, xmm0 ; abs (q3-q2)
-
movdqa xmm4, xmm5 ; q1
- psubusb xmm4, xmm6 ; q1-q2
- psubusb xmm6, xmm5 ; q2-q1
- por xmm6, xmm4 ; abs (q2-q1)
+ por xmm7, xmm0 ; abs (q3-q2)
+ psubusb xmm4, xmm6 ; q1-q2
movdqa xmm0, xmm1
+ psubusb xmm6, xmm5 ; q2-q1
+ por xmm6, xmm4 ; abs (q2-q1)
psubusb xmm0, xmm2 ; p2 - p3;
- psubusb xmm2, xmm1 ; p3 - p2;
+ psubusb xmm2, xmm1 ; p3 - p2;
por xmm0, xmm2 ; abs(p2-p3)
%if %1
movdqa xmm2, [rdx] ; p1
@@ -892,31 +884,20 @@
movdqa xmm2, [rdx+32] ; p1
%endif
movdqa xmm5, xmm2 ; p1
+ pmaxub xmm0, xmm7
psubusb xmm5, xmm1 ; p1-p2
psubusb xmm1, xmm2 ; p2-p1
+ movdqa xmm7, xmm3 ; p0
+ psubusb xmm7, xmm2 ; p0-p1
+
por xmm1, xmm5 ; abs(p2-p1)
+ pmaxub xmm0, xmm6
- mov rdx, arg(3) ; limit
- movdqa xmm4, [rdx] ; limit
-
- psubusb xmm7, xmm4
-
- psubusb xmm0, xmm4 ; abs(p3-p2) > limit
- psubusb xmm1, xmm4 ; abs(p2-p1) > limit
-
- psubusb xmm6, xmm4 ; abs(q2-q1) > limit
- por xmm7, xmm6 ; or
-
- por xmm0, xmm1
- por xmm0, xmm7 ; abs(q3-q2) > limit || abs(p3-p2) > limit ||abs(p2-p1) > limit || abs(q2-q1) > limit
-
+ pmaxub xmm0, xmm1
movdqa xmm1, xmm2 ; p1
- movdqa xmm7, xmm3 ; p0
- psubusb xmm7, xmm2 ; p0-p1
-
psubusb xmm2, xmm3 ; p1-p0
por xmm2, xmm7 ; abs(p1-p0)
@@ -923,8 +904,8 @@
movdqa t0, xmm2 ; save abs(p1-p0)
lea rdx, srct
- psubusb xmm2, xmm4 ; abs(p1-p0)>limit
- por xmm0, xmm2 ; mask
+ pmaxub xmm0, xmm2
+
%if %1
movdqa xmm5, [rdx+32] ; q0
movdqa xmm7, [rdx+48] ; q1
@@ -940,9 +921,12 @@
por xmm7, xmm5 ; abs(q1-q0)
movdqa t1, xmm7 ; save abs(q1-q0)
- psubusb xmm7, xmm4 ; abs(q1-q0)> limit
- por xmm0, xmm7 ; mask
+ mov rdx, arg(3) ; limit
+ movdqa xmm4, [rdx] ; limit
+
+ pmaxub xmm0, xmm7
+ psubusb xmm0, xmm4
movdqa xmm5, xmm2 ; q1
psubusb xmm5, xmm1 ; q1-=p1
--- a/vp8/common/x86/subpixel_ssse3.asm
+++ b/vp8/common/x86/subpixel_ssse3.asm
@@ -70,28 +70,36 @@
sub rdi, rdx
;xmm3 free
filter_block1d8_h6_rowloop_ssse3:
- movdqu xmm0, XMMWORD PTR [rsi - 2]
+ movq xmm0, MMWORD PTR [rsi - 2] ; -2 -1 0 1 2 3 4 5
- movdqa xmm1, xmm0
- pshufb xmm0, [shuf1b GLOBAL]
+ movq xmm2, MMWORD PTR [rsi + 3] ; 3 4 5 6 7 8 9 10
- movdqa xmm2, xmm1
- pshufb xmm1, [shuf2b GLOBAL]
- pmaddubsw xmm0, xmm4
- pmaddubsw xmm1, xmm5
+ punpcklbw xmm0, xmm2 ; -2 3 -1 4 0 5 1 6 2 7 3 8 4 9 5 10
- pshufb xmm2, [shuf3b GLOBAL]
- add rdi, rdx
- pmaddubsw xmm2, xmm6
+ movdqa xmm1, xmm0
+ pmaddubsw xmm0, xmm4
+ movdqa xmm2, xmm1
+ pshufb xmm1, [shuf2bfrom1 GLOBAL]
+
+ pshufb xmm2, [shuf3bfrom1 GLOBAL]
+ pmaddubsw xmm1, xmm5
+
+ lea rdi, [rdi + rdx]
+ pmaddubsw xmm2, xmm6
+
lea rsi, [rsi + rax]
dec rcx
- paddsw xmm0, xmm1
- paddsw xmm0, xmm7
- paddsw xmm0, xmm2
- psraw xmm0, 7
- packuswb xmm0, xmm0
+ paddsw xmm0, xmm1
+ paddsw xmm2, xmm7
+
+ paddsw xmm0, xmm2
+
+ psraw xmm0, 7
+
+ packuswb xmm0, xmm0
+
movq MMWORD Ptr [rdi], xmm0
jnz filter_block1d8_h6_rowloop_ssse3
@@ -107,8 +115,8 @@
movdqa xmm5, XMMWORD PTR [rax+256] ;k2_k4
movdqa xmm6, XMMWORD PTR [rax+128] ;k1_k3
- movdqa xmm3, XMMWORD PTR [shuf2b GLOBAL]
- movdqa xmm4, XMMWORD PTR [shuf3b GLOBAL]
+ movdqa xmm3, XMMWORD PTR [shuf2bfrom1 GLOBAL]
+ movdqa xmm4, XMMWORD PTR [shuf3bfrom1 GLOBAL]
mov rsi, arg(0) ;src_ptr
@@ -118,25 +126,34 @@
movsxd rdx, dword ptr arg(3) ;output_pitch
sub rdi, rdx
-;xmm3 free
+
filter_block1d8_h4_rowloop_ssse3:
- movdqu xmm0, XMMWORD PTR [rsi - 2]
+ movq xmm0, MMWORD PTR [rsi - 2] ; -2 -1 0 1 2 3 4 5
- movdqa xmm2, xmm0
- pshufb xmm0, xmm3 ;[shuf2b GLOBAL]
- pshufb xmm2, xmm4 ;[shuf3b GLOBAL]
+ movq xmm1, MMWORD PTR [rsi + 3] ; 3 4 5 6 7 8 9 10
- pmaddubsw xmm0, xmm5
- add rdi, rdx
- pmaddubsw xmm2, xmm6
+ punpcklbw xmm0, xmm1 ; -2 3 -1 4 0 5 1 6 2 7 3 8 4 9 5 10
+ movdqa xmm2, xmm0
+ pshufb xmm0, xmm3
+
+ pshufb xmm2, xmm4
+ pmaddubsw xmm0, xmm5
+
+ lea rdi, [rdi + rdx]
+ pmaddubsw xmm2, xmm6
+
lea rsi, [rsi + rax]
dec rcx
- paddsw xmm0, xmm7
- paddsw xmm0, xmm2
- psraw xmm0, 7
- packuswb xmm0, xmm0
+ paddsw xmm0, xmm7
+
+ paddsw xmm0, xmm2
+
+ psraw xmm0, 7
+
+ packuswb xmm0, xmm0
+
movq MMWORD Ptr [rdi], xmm0
jnz filter_block1d8_h4_rowloop_ssse3
@@ -168,7 +185,7 @@
push rdi
; end prolog
- movsxd rdx, DWORD PTR arg(5) ;table index
+ movsxd rdx, DWORD PTR arg(5) ;table index
xor rsi, rsi
shl rdx, 4 ;
@@ -175,67 +192,81 @@
lea rax, [k0_k5 GLOBAL]
add rax, rdx
- mov rdi, arg(2) ;output_ptr
- movdqa xmm7, [rd GLOBAL]
+ mov rdi, arg(2) ;output_ptr
;;
;; cmp esi, DWORD PTR [rax]
;; je vp8_filter_block1d16_h4_ssse3
- mov rsi, arg(0) ;src_ptr
+ mov rsi, arg(0) ;src_ptr
movdqa xmm4, XMMWORD PTR [rax] ;k0_k5
movdqa xmm5, XMMWORD PTR [rax+256] ;k2_k4
movdqa xmm6, XMMWORD PTR [rax+128] ;k1_k3
- movsxd rax, dword ptr arg(1) ;src_pixels_per_line
- movsxd rcx, dword ptr arg(4) ;output_height
- movsxd rdx, dword ptr arg(3) ;output_pitch
+ movsxd rax, dword ptr arg(1) ;src_pixels_per_line
+ movsxd rcx, dword ptr arg(4) ;output_height
+ movsxd rdx, dword ptr arg(3) ;output_pitch
filter_block1d16_h6_rowloop_ssse3:
- movdqu xmm0, XMMWORD PTR [rsi - 2]
+ movq xmm0, MMWORD PTR [rsi - 2] ; -2 -1 0 1 2 3 4 5
- movdqa xmm1, xmm0
- pshufb xmm0, [shuf1b GLOBAL]
- movdqa xmm2, xmm1
- pmaddubsw xmm0, xmm4
- pshufb xmm1, [shuf2b GLOBAL]
- pshufb xmm2, [shuf3b GLOBAL]
- pmaddubsw xmm1, xmm5
+ movq xmm3, MMWORD PTR [rsi + 3] ; 3 4 5 6 7 8 9 10
- movdqu xmm3, XMMWORD PTR [rsi + 6]
+ punpcklbw xmm0, xmm3 ; -2 3 -1 4 0 5 1 6 2 7 3 8 4 9 5 10
- pmaddubsw xmm2, xmm6
- paddsw xmm0, xmm1
- movdqa xmm1, xmm3
- pshufb xmm3, [shuf1b GLOBAL]
- paddsw xmm0, xmm7
- pmaddubsw xmm3, xmm4
- paddsw xmm0, xmm2
- movdqa xmm2, xmm1
- pshufb xmm1, [shuf2b GLOBAL]
- pshufb xmm2, [shuf3b GLOBAL]
- pmaddubsw xmm1, xmm5
- pmaddubsw xmm2, xmm6
+ movdqa xmm1, xmm0
+ pmaddubsw xmm0, xmm4
- psraw xmm0, 7
- packuswb xmm0, xmm0
+ movdqa xmm2, xmm1
+ pshufb xmm1, [shuf2bfrom1 GLOBAL]
+
+ pshufb xmm2, [shuf3bfrom1 GLOBAL]
+ movq xmm3, MMWORD PTR [rsi + 6]
+
+ pmaddubsw xmm1, xmm5
+ movq xmm7, MMWORD PTR [rsi + 11]
+
+ pmaddubsw xmm2, xmm6
+ punpcklbw xmm3, xmm7
+
+ paddsw xmm0, xmm1
+ movdqa xmm1, xmm3
+
+ pmaddubsw xmm3, xmm4
+ paddsw xmm0, xmm2
+
+ movdqa xmm2, xmm1
+ paddsw xmm0, [rd GLOBAL]
+
+ pshufb xmm1, [shuf2bfrom1 GLOBAL]
+ pshufb xmm2, [shuf3bfrom1 GLOBAL]
+
+ psraw xmm0, 7
+ pmaddubsw xmm1, xmm5
+
+ pmaddubsw xmm2, xmm6
+ packuswb xmm0, xmm0
+
lea rsi, [rsi + rax]
- paddsw xmm3, xmm1
- paddsw xmm3, xmm7
- paddsw xmm3, xmm2
- psraw xmm3, 7
- packuswb xmm3, xmm3
+ paddsw xmm3, xmm1
- punpcklqdq xmm0, xmm3
+ paddsw xmm3, xmm2
+ paddsw xmm3, [rd GLOBAL]
+
+ psraw xmm3, 7
+
+ packuswb xmm3, xmm3
+
+ punpcklqdq xmm0, xmm3
+
movdqa XMMWORD Ptr [rdi], xmm0
- add rdi, rdx
+ lea rdi, [rdi + rdx]
dec rcx
jnz filter_block1d16_h6_rowloop_ssse3
-
; begin epilog
pop rdi
pop rsi
@@ -268,7 +299,7 @@
pshufb xmm3, [shuf3b GLOBAL]
pshufb xmm0, [shuf2b GLOBAL]
- paddsw xmm1, xmm7
+ paddsw xmm1, [rd GLOBAL]
paddsw xmm1, xmm2
pmaddubsw xmm0, xmm5
@@ -278,7 +309,7 @@
packuswb xmm1, xmm1
lea rsi, [rsi + rax]
paddsw xmm3, xmm0
- paddsw xmm3, xmm7
+ paddsw xmm3, [rd GLOBAL]
psraw xmm3, 7
packuswb xmm3, xmm3
@@ -939,17 +970,19 @@
%if ABI_IS_32BIT=0
movsxd r8, dword ptr arg(5) ; dst_pitch
%endif
- movdqu xmm3, [rsi] ; 00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15
+ movq xmm3, [rsi] ; 00 01 02 03 04 05 06 07
+ movq xmm5, [rsi+1] ; 01 02 03 04 05 06 07 08
- movdqa xmm4, xmm3
+ punpcklbw xmm3, xmm5 ; 00 01 01 02 02 03 03 04 04 05 05 06 06 07 07 08
+ movq xmm4, [rsi+8] ; 08 09 10 11 12 13 14 15
- movdqu xmm5, [rsi+1] ; 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16
+ movq xmm5, [rsi+9] ; 09 10 11 12 13 14 15 16
+
lea rsi, [rsi + rdx] ; next line
- punpcklbw xmm3, xmm5 ; 00 01 01 02 02 03 03 04 04 05 05 06 06 07 07 08
pmaddubsw xmm3, xmm1 ; 00 02 04 06 08 10 12 14
- punpckhbw xmm4, xmm5 ; 08 09 09 10 10 11 11 12 12 13 13 14 14 15 15 16
+ punpcklbw xmm4, xmm5 ; 08 09 09 10 10 11 11 12 12 13 13 14 14 15 15 16
pmaddubsw xmm4, xmm1 ; 01 03 05 07 09 11 13 15
paddw xmm3, [rd GLOBAL] ; xmm3 += round value
@@ -962,17 +995,18 @@
packuswb xmm7, xmm4 ; 00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15
.next_row:
- movdqu xmm6, [rsi] ; 00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15
+ movq xmm6, [rsi] ; 00 01 02 03 04 05 06 07
+ movq xmm5, [rsi+1] ; 01 02 03 04 05 06 07 08
- movdqa xmm4, xmm6
+ punpcklbw xmm6, xmm5
+ movq xmm4, [rsi+8] ; 08 09 10 11 12 13 14 15
- movdqu xmm5, [rsi+1] ; 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16
+ movq xmm5, [rsi+9] ; 09 10 11 12 13 14 15 16
lea rsi, [rsi + rdx] ; next line
- punpcklbw xmm6, xmm5
pmaddubsw xmm6, xmm1
- punpckhbw xmm4, xmm5
+ punpcklbw xmm4, xmm5
pmaddubsw xmm4, xmm1
paddw xmm6, [rd GLOBAL] ; xmm6 += round value
@@ -1027,49 +1061,51 @@
movsxd rax, dword ptr arg(1) ; src_pixels_per_line
; get the first horizontal line done
- movdqu xmm2, [rsi] ; load row 0
+ movq xmm4, [rsi] ; load row 0
+ movq xmm2, [rsi + 8] ; load row 0
lea rsi, [rsi + rax] ; next line
.next_row:
- movdqu xmm3, [rsi] ; load row + 1
+ movq xmm3, [rsi] ; load row + 1
+ movq xmm5, [rsi + 8] ; load row + 1
- movdqu xmm4, xmm2
punpcklbw xmm4, xmm3
+ punpcklbw xmm2, xmm5
pmaddubsw xmm4, xmm1
- movdqu xmm7, [rsi + rax] ; load row + 2
+ movq xmm7, [rsi + rax] ; load row + 2
- punpckhbw xmm2, xmm3
- movdqu xmm6, xmm3
-
pmaddubsw xmm2, xmm1
- punpcklbw xmm6, xmm7
+ movq xmm6, [rsi + rax + 8] ; load row + 2
+ punpcklbw xmm3, xmm7
+ punpcklbw xmm5, xmm6
+
+ pmaddubsw xmm3, xmm1
paddw xmm4, [rd GLOBAL]
- pmaddubsw xmm6, xmm1
- psraw xmm4, VP8_FILTER_SHIFT
- punpckhbw xmm3, xmm7
-
+ pmaddubsw xmm5, xmm1
paddw xmm2, [rd GLOBAL]
- pmaddubsw xmm3, xmm1
+ psraw xmm4, VP8_FILTER_SHIFT
psraw xmm2, VP8_FILTER_SHIFT
- paddw xmm6, [rd GLOBAL]
packuswb xmm4, xmm2
- psraw xmm6, VP8_FILTER_SHIFT
+ paddw xmm3, [rd GLOBAL]
movdqa [rdi], xmm4 ; store row 0
- paddw xmm3, [rd GLOBAL]
+ paddw xmm5, [rd GLOBAL]
psraw xmm3, VP8_FILTER_SHIFT
- lea rsi, [rsi + 2*rax]
+ psraw xmm5, VP8_FILTER_SHIFT
- packuswb xmm6, xmm3
- movdqa xmm2, xmm7
+ packuswb xmm3, xmm5
+ movdqa xmm4, xmm7
- movdqa [rdi + rdx],xmm6 ; store row 1
+ movdqa [rdi + rdx],xmm3 ; store row 1
+ lea rsi, [rsi + 2*rax]
+
+ movdqa xmm2, xmm6
lea rdi, [rdi + 2*rdx]
cmp rdi, rcx
@@ -1083,32 +1119,35 @@
movsxd rax, dword ptr arg(1) ; src_pixels_per_line
.next_row:
- movdqu xmm2, [rsi] ; row 0
- movdqa xmm3, xmm2
+ movq xmm2, [rsi] ; 00 01 02 03 04 05 06 07
+ movq xmm4, [rsi+1] ; 01 02 03 04 05 06 07 08
- movdqu xmm4, [rsi + 1] ; row 0 + 1
- lea rsi, [rsi + rax] ; next line
-
punpcklbw xmm2, xmm4
- movdqu xmm5, [rsi] ; row 1
+ movq xmm3, [rsi+8] ; 08 09 10 11 12 13 14 15
pmaddubsw xmm2, xmm1
- movdqa xmm6, xmm5
+ movq xmm4, [rsi+9] ; 09 10 11 12 13 14 15 16
- punpckhbw xmm3, xmm4
- movdqu xmm7, [rsi + 1] ; row 1 + 1
+ lea rsi, [rsi + rax] ; next line
+ punpcklbw xmm3, xmm4
pmaddubsw xmm3, xmm1
+ movq xmm5, [rsi]
+
paddw xmm2, [rd GLOBAL]
+ movq xmm7, [rsi+1]
+ movq xmm6, [rsi+8]
psraw xmm2, VP8_FILTER_SHIFT
+
punpcklbw xmm5, xmm7
+ movq xmm7, [rsi+9]
paddw xmm3, [rd GLOBAL]
pmaddubsw xmm5, xmm1
psraw xmm3, VP8_FILTER_SHIFT
- punpckhbw xmm6, xmm7
+ punpcklbw xmm6, xmm7
packuswb xmm2, xmm3
pmaddubsw xmm6, xmm1
@@ -1462,6 +1501,13 @@
db 2, 4, 3, 5, 4, 6, 5, 7, 6, 8, 7, 9, 8, 10, 9, 11
shuf3b:
db 1, 3, 2, 4, 3, 5, 4, 6, 5, 7, 6, 8, 7, 9, 8, 10
+
+align 16
+shuf2bfrom1:
+ db 4, 8, 6, 1, 8, 3, 1, 5, 3, 7, 5, 9, 7,11, 9,13
+align 16
+shuf3bfrom1:
+ db 2, 6, 4, 8, 6, 1, 8, 3, 1, 5, 3, 7, 5, 9, 7,11
align 16
rd:
--- a/vp8/decoder/arm/neon/dequant_dc_idct_neon.asm
+++ /dev/null
@@ -1,136 +1,0 @@
-;
-; Copyright (c) 2010 The WebM project authors. All Rights Reserved.
-;
-; Use of this source code is governed by a BSD-style license
-; that can be found in the LICENSE file in the root of the source
-; tree. An additional intellectual property rights grant can be found
-; in the file PATENTS. All contributing project authors may
-; be found in the AUTHORS file in the root of the source tree.
-;
-
-
- EXPORT |vp8_dequant_dc_idct_add_neon|
- ARM
- REQUIRE8
- PRESERVE8
-
- AREA ||.text||, CODE, READONLY, ALIGN=2
-;void vp8_dequant_dc_idct_add_neon(short *input, short *dq, unsigned char *pred,
-; unsigned char *dest, int pitch, int stride,
-; int Dc);
-; r0 short *input,
-; r1 short *dq,
-; r2 unsigned char *pred
-; r3 unsigned char *dest
-; sp int pitch
-; sp+4 int stride
-; sp+8 int Dc
-|vp8_dequant_dc_idct_add_neon| PROC
- vld1.16 {q3, q4}, [r0]- vld1.16 {q5, q6}, [r1]-
- ldr r1, [sp, #8] ;load Dc from stack
-
- ldr r12, _CONSTANTS_
-
- vmul.i16 q1, q3, q5 ;input for short_idct4x4llm_neon
- vmul.i16 q2, q4, q6
-
- vmov.16 d2[0], r1
-
- ldr r1, [sp] ; pitch
- vld1.32 {d14[0]}, [r2], r1- vld1.32 {d14[1]}, [r2], r1- vld1.32 {d15[0]}, [r2], r1- vld1.32 {d15[1]}, [r2]-
- ldr r1, [sp, #4] ; stride
-
-;|short_idct4x4llm_neon| PROC
- vld1.16 {d0}, [r12]- vswp d3, d4 ;q2(vp[4] vp[12])
-
- vqdmulh.s16 q3, q2, d0[2]
- vqdmulh.s16 q4, q2, d0[0]
-
- vqadd.s16 d12, d2, d3 ;a1
- vqsub.s16 d13, d2, d3 ;b1
-
- vshr.s16 q3, q3, #1
- vshr.s16 q4, q4, #1
-
- vqadd.s16 q3, q3, q2
- vqadd.s16 q4, q4, q2
-
- vqsub.s16 d10, d6, d9 ;c1
- vqadd.s16 d11, d7, d8 ;d1
-
- vqadd.s16 d2, d12, d11
- vqadd.s16 d3, d13, d10
- vqsub.s16 d4, d13, d10
- vqsub.s16 d5, d12, d11
-
- vtrn.32 d2, d4
- vtrn.32 d3, d5
- vtrn.16 d2, d3
- vtrn.16 d4, d5
-
-; memset(input, 0, 32) -- 32bytes
- vmov.i16 q14, #0
-
- vswp d3, d4
- vqdmulh.s16 q3, q2, d0[2]
- vqdmulh.s16 q4, q2, d0[0]
-
- vqadd.s16 d12, d2, d3 ;a1
- vqsub.s16 d13, d2, d3 ;b1
-
- vmov q15, q14
-
- vshr.s16 q3, q3, #1
- vshr.s16 q4, q4, #1
-
- vqadd.s16 q3, q3, q2
- vqadd.s16 q4, q4, q2
-
- vqsub.s16 d10, d6, d9 ;c1
- vqadd.s16 d11, d7, d8 ;d1
-
- vqadd.s16 d2, d12, d11
- vqadd.s16 d3, d13, d10
- vqsub.s16 d4, d13, d10
- vqsub.s16 d5, d12, d11
-
- vst1.16 {q14, q15}, [r0]-
- vrshr.s16 d2, d2, #3
- vrshr.s16 d3, d3, #3
- vrshr.s16 d4, d4, #3
- vrshr.s16 d5, d5, #3
-
- vtrn.32 d2, d4
- vtrn.32 d3, d5
- vtrn.16 d2, d3
- vtrn.16 d4, d5
-
- vaddw.u8 q1, q1, d14
- vaddw.u8 q2, q2, d15
-
- vqmovun.s16 d0, q1
- vqmovun.s16 d1, q2
-
- vst1.32 {d0[0]}, [r3], r1- vst1.32 {d0[1]}, [r3], r1- vst1.32 {d1[0]}, [r3], r1- vst1.32 {d1[1]}, [r3]-
- bx lr
-
- ENDP ; |vp8_dequant_dc_idct_add_neon|
-
-; Constant Pool
-_CONSTANTS_ DCD cospi8sqrt2minus1
-cospi8sqrt2minus1 DCD 0x4e7b4e7b
-sinpi8sqrt2 DCD 0x8a8c8a8c
-
- END
--- a/vp8/decoder/arm/neon/idct_blk_neon.c
+++ b/vp8/decoder/arm/neon/idct_blk_neon.c
@@ -12,6 +12,21 @@
#include "idct.h"
#include "dequantize.h"
+/* place these declarations here because we don't want to maintain them
+ * outside of this scope
+ */
+void idct_dequant_dc_full_2x_neon
+ (short *input, short *dq, unsigned char *pre, unsigned char *dst,
+ int stride, short *dc);
+void idct_dequant_dc_0_2x_neon
+ (short *dc, unsigned char *pre, unsigned char *dst, int stride);
+void idct_dequant_full_2x_neon
+ (short *q, short *dq, unsigned char *pre, unsigned char *dst,
+ int pitch, int stride);
+void idct_dequant_0_2x_neon
+ (short *q, short dq, unsigned char *pre, int pitch,
+ unsigned char *dst, int stride);
+
void vp8_dequant_dc_idct_add_y_block_neon
(short *q, short *dq, unsigned char *pre,
unsigned char *dst, int stride, char *eobs, short *dc)
@@ -20,26 +35,16 @@
for (i = 0; i < 4; i++)
{- if (eobs[0] > 1)
- vp8_dequant_dc_idct_add_neon (q, dq, pre, dst, 16, stride, dc[0]);
+ if (((short *)eobs)[0] & 0xfefe)
+ idct_dequant_dc_full_2x_neon (q, dq, pre, dst, stride, dc);
else
- vp8_dc_only_idct_add_neon (dc[0], pre, dst, 16, stride);
+ idct_dequant_dc_0_2x_neon(dc, pre, dst, stride);
- if (eobs[1] > 1)
- vp8_dequant_dc_idct_add_neon (q+16, dq, pre+4, dst+4, 16, stride, dc[1]);
+ if (((short *)eobs)[1] & 0xfefe)
+ idct_dequant_dc_full_2x_neon (q+32, dq, pre+8, dst+8, stride, dc+2);
else
- vp8_dc_only_idct_add_neon (dc[1], pre+4, dst+4, 16, stride);
+ idct_dequant_dc_0_2x_neon(dc+2, pre+8, dst+8, stride);
- if (eobs[2] > 1)
- vp8_dequant_dc_idct_add_neon (q+32, dq, pre+8, dst+8, 16, stride, dc[2]);
- else
- vp8_dc_only_idct_add_neon (dc[2], pre+8, dst+8, 16, stride);
-
- if (eobs[3] > 1)
- vp8_dequant_dc_idct_add_neon (q+48, dq, pre+12, dst+12, 16, stride, dc[3]);
- else
- vp8_dc_only_idct_add_neon (dc[3], pre+12, dst+12, 16, stride);
-
q += 64;
dc += 4;
pre += 64;
@@ -56,38 +61,16 @@
for (i = 0; i < 4; i++)
{- if (eobs[0] > 1)
- vp8_dequant_idct_add_neon (q, dq, pre, dst, 16, stride);
+ if (((short *)eobs)[0] & 0xfefe)
+ idct_dequant_full_2x_neon (q, dq, pre, dst, 16, stride);
else
- {- vp8_dc_only_idct_add_neon (q[0]*dq[0], pre, dst, 16, stride);
- ((int *)q)[0] = 0;
- }
+ idct_dequant_0_2x_neon (q, dq[0], pre, 16, dst, stride);
- if (eobs[1] > 1)
- vp8_dequant_idct_add_neon (q+16, dq, pre+4, dst+4, 16, stride);
+ if (((short *)eobs)[1] & 0xfefe)
+ idct_dequant_full_2x_neon (q+32, dq, pre+8, dst+8, 16, stride);
else
- {- vp8_dc_only_idct_add_neon (q[16]*dq[0], pre+4, dst+4, 16, stride);
- ((int *)(q+16))[0] = 0;
- }
+ idct_dequant_0_2x_neon (q+32, dq[0], pre+8, 16, dst+8, stride);
- if (eobs[2] > 1)
- vp8_dequant_idct_add_neon (q+32, dq, pre+8, dst+8, 16, stride);
- else
- {- vp8_dc_only_idct_add_neon (q[32]*dq[0], pre+8, dst+8, 16, stride);
- ((int *)(q+32))[0] = 0;
- }
-
- if (eobs[3] > 1)
- vp8_dequant_idct_add_neon (q+48, dq, pre+12, dst+12, 16, stride);
- else
- {- vp8_dc_only_idct_add_neon (q[48]*dq[0], pre+12, dst+12, 16, stride);
- ((int *)(q+48))[0] = 0;
- }
-
q += 64;
pre += 64;
dst += 4*stride;
@@ -99,53 +82,34 @@
(short *q, short *dq, unsigned char *pre,
unsigned char *dstu, unsigned char *dstv, int stride, char *eobs)
{- int i;
+ if (((short *)eobs)[0] & 0xfefe)
+ idct_dequant_full_2x_neon (q, dq, pre, dstu, 8, stride);
+ else
+ idct_dequant_0_2x_neon (q, dq[0], pre, 8, dstu, stride);
- for (i = 0; i < 2; i++)
- {- if (eobs[0] > 1)
- vp8_dequant_idct_add_neon (q, dq, pre, dstu, 8, stride);
- else
- {- vp8_dc_only_idct_add_neon (q[0]*dq[0], pre, dstu, 8, stride);
- ((int *)q)[0] = 0;
- }
+ q += 32;
+ pre += 32;
+ dstu += 4*stride;
- if (eobs[1] > 1)
- vp8_dequant_idct_add_neon (q+16, dq, pre+4, dstu+4, 8, stride);
- else
- {- vp8_dc_only_idct_add_neon (q[16]*dq[0], pre+4, dstu+4, 8, stride);
- ((int *)(q+16))[0] = 0;
- }
+ if (((short *)eobs)[1] & 0xfefe)
+ idct_dequant_full_2x_neon (q, dq, pre, dstu, 8, stride);
+ else
+ idct_dequant_0_2x_neon (q, dq[0], pre, 8, dstu, stride);
- q += 32;
- pre += 32;
- dstu += 4*stride;
- eobs += 2;
- }
+ q += 32;
+ pre += 32;
- for (i = 0; i < 2; i++)
- {- if (eobs[0] > 1)
- vp8_dequant_idct_add_neon (q, dq, pre, dstv, 8, stride);
- else
- {- vp8_dc_only_idct_add_neon (q[0]*dq[0], pre, dstv, 8, stride);
- ((int *)q)[0] = 0;
- }
+ if (((short *)eobs)[2] & 0xfefe)
+ idct_dequant_full_2x_neon (q, dq, pre, dstv, 8, stride);
+ else
+ idct_dequant_0_2x_neon (q, dq[0], pre, 8, dstv, stride);
- if (eobs[1] > 1)
- vp8_dequant_idct_add_neon (q+16, dq, pre+4, dstv+4, 8, stride);
- else
- {- vp8_dc_only_idct_add_neon (q[16]*dq[0], pre+4, dstv+4, 8, stride);
- ((int *)(q+16))[0] = 0;
- }
+ q += 32;
+ pre += 32;
+ dstv += 4*stride;
- q += 32;
- pre += 32;
- dstv += 4*stride;
- eobs += 2;
- }
+ if (((short *)eobs)[3] & 0xfefe)
+ idct_dequant_full_2x_neon (q, dq, pre, dstv, 8, stride);
+ else
+ idct_dequant_0_2x_neon (q, dq[0], pre, 8, dstv, stride);
}
--- /dev/null
+++ b/vp8/decoder/arm/neon/idct_dequant_0_2x_neon.asm
@@ -1,0 +1,79 @@
+;
+; Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+; Use of this source code is governed by a BSD-style license and patent
+; grant that can be found in the LICENSE file in the root of the source
+; tree. All contributing project authors may be found in the AUTHORS
+; file in the root of the source tree.
+;
+
+
+ EXPORT |idct_dequant_0_2x_neon|
+ ARM
+ REQUIRE8
+ PRESERVE8
+
+ AREA ||.text||, CODE, READONLY, ALIGN=2
+;void idct_dequant_0_2x_neon(short *q, short dq, unsigned char *pre,
+; int pitch, unsigned char *dst, int stride);
+; r0 *q
+; r1 dq
+; r2 *pre
+; r3 pitch
+; sp *dst
+; sp+4 stride
+|idct_dequant_0_2x_neon| PROC
+ add r12, r2, #4
+ vld1.32 {d2[0]}, [r2], r3+ vld1.32 {d2[1]}, [r2], r3+ vld1.32 {d4[0]}, [r2], r3+ vld1.32 {d4[1]}, [r2]+ vld1.32 {d8[0]}, [r12], r3+ vld1.32 {d8[1]}, [r12], r3+ vld1.32 {d10[0]}, [r12], r3+ vld1.32 {d10[1]}, [r12]+
+ ldrh r12, [r0] ; lo q
+ ldrh r2, [r0, #32] ; hi q
+ mov r3, #0
+ strh r3, [r0]
+ strh r3, [r0, #32]
+
+ sxth r12, r12 ; lo
+ mul r0, r12, r1
+ add r0, r0, #4
+ asr r0, r0, #3
+ vdup.16 q0, r0
+ sxth r2, r2 ; hi
+ mul r0, r2, r1
+ add r0, r0, #4
+ asr r0, r0, #3
+ vdup.16 q3, r0
+
+ vaddw.u8 q1, q0, d2 ; lo
+ vaddw.u8 q2, q0, d4
+ vaddw.u8 q4, q3, d8 ; hi
+ vaddw.u8 q5, q3, d10
+
+ ldr r2, [sp] ; dst
+ ldr r3, [sp, #4] ; stride
+
+ vqmovun.s16 d2, q1 ; lo
+ vqmovun.s16 d4, q2
+ vqmovun.s16 d8, q4 ; hi
+ vqmovun.s16 d10, q5
+
+ add r0, r2, #4
+ vst1.32 {d2[0]}, [r2], r3 ; lo+ vst1.32 {d2[1]}, [r2], r3+ vst1.32 {d4[0]}, [r2], r3+ vst1.32 {d4[1]}, [r2]+ vst1.32 {d8[0]}, [r0], r3 ; hi+ vst1.32 {d8[1]}, [r0], r3+ vst1.32 {d10[0]}, [r0], r3+ vst1.32 {d10[1]}, [r0]+
+ bx lr
+
+ ENDP ; |idct_dequant_0_2x_neon|
+ END
--- /dev/null
+++ b/vp8/decoder/arm/neon/idct_dequant_dc_0_2x_neon.asm
@@ -1,0 +1,69 @@
+;
+; Copyright (c) 2010 The Webm project authors. All Rights Reserved.
+;
+; Use of this source code is governed by a BSD-style license and patent
+; grant that can be found in the LICENSE file in the root of the source
+; tree. All contributing project authors may be found in the AUTHORS
+; file in the root of the source tree.
+;
+
+
+ EXPORT |idct_dequant_dc_0_2x_neon|
+ ARM
+ REQUIRE8
+ PRESERVE8
+
+ AREA ||.text||, CODE, READONLY, ALIGN=2
+;void idct_dequant_dc_0_2x_neon(short *dc, unsigned char *pre,
+; unsigned char *dst, int stride);
+; r0 *dc
+; r1 *pre
+; r2 *dst
+; r3 stride
+|idct_dequant_dc_0_2x_neon| PROC
+ ldr r0, [r0] ; *dc
+ mov r12, #16
+
+ vld1.32 {d2[0]}, [r1], r12 ; lo+ vld1.32 {d2[1]}, [r1], r12+ vld1.32 {d4[0]}, [r1], r12+ vld1.32 {d4[1]}, [r1]+ sub r1, r1, #44
+ vld1.32 {d8[0]}, [r1], r12 ; hi+ vld1.32 {d8[1]}, [r1], r12+ vld1.32 {d10[0]}, [r1], r12+ vld1.32 {d10[1]}, [r1]+
+ sxth r1, r0 ; lo *dc
+ add r1, r1, #4
+ asr r1, r1, #3
+ vdup.16 q0, r1
+ sxth r0, r0, ror #16 ; hi *dc
+ add r0, r0, #4
+ asr r0, r0, #3
+ vdup.16 q3, r0
+
+ vaddw.u8 q1, q0, d2 ; lo
+ vaddw.u8 q2, q0, d4
+ vaddw.u8 q4, q3, d8 ; hi
+ vaddw.u8 q5, q3, d10
+
+ vqmovun.s16 d2, q1 ; lo
+ vqmovun.s16 d4, q2
+ vqmovun.s16 d8, q4 ; hi
+ vqmovun.s16 d10, q5
+
+ add r0, r2, #4
+ vst1.32 {d2[0]}, [r2], r3 ; lo+ vst1.32 {d2[1]}, [r2], r3+ vst1.32 {d4[0]}, [r2], r3+ vst1.32 {d4[1]}, [r2]+ vst1.32 {d8[0]}, [r0], r3 ; hi+ vst1.32 {d8[1]}, [r0], r3+ vst1.32 {d10[0]}, [r0], r3+ vst1.32 {d10[1]}, [r0]+
+ bx lr
+
+ ENDP ;|idct_dequant_dc_0_2x_neon|
+ END
--- /dev/null
+++ b/vp8/decoder/arm/neon/idct_dequant_dc_full_2x_neon.asm
@@ -1,0 +1,206 @@
+;
+; Copyright (c) 2010 The Webm project authors. All Rights Reserved.
+;
+; Use of this source code is governed by a BSD-style license
+; that can be found in the LICENSE file in the root of the source
+; tree. An additional intellectual property rights grant can be found
+; in the file PATENTS. All contributing project authors may
+; be found in the AUTHORS file in the root of the source tree.
+;
+
+
+ EXPORT |idct_dequant_dc_full_2x_neon|
+ ARM
+ REQUIRE8
+ PRESERVE8
+
+ AREA ||.text||, CODE, READONLY, ALIGN=2
+;void idct_dequant_dc_full_2x_neon(short *q, short *dq, unsigned char *pre,
+; unsigned char *dst, int stride, short *dc);
+; r0 *q,
+; r1 *dq,
+; r2 *pre
+; r3 *dst
+; sp stride
+; sp+4 *dc
+|idct_dequant_dc_full_2x_neon| PROC
+ vld1.16 {q0, q1}, [r1] ; dq (same l/r)+ vld1.16 {q2, q3}, [r0] ; l q+ mov r1, #16 ; pitch
+ add r0, r0, #32
+ vld1.16 {q4, q5}, [r0] ; r q+ add r12, r2, #4
+ ; interleave the predictors
+ vld1.32 {d28[0]}, [r2], r1 ; l pre+ vld1.32 {d28[1]}, [r12], r1 ; r pre+ vld1.32 {d29[0]}, [r2], r1+ vld1.32 {d29[1]}, [r12], r1+ vld1.32 {d30[0]}, [r2], r1+ vld1.32 {d30[1]}, [r12], r1+ vld1.32 {d31[0]}, [r2]+ ldr r1, [sp, #4]
+ vld1.32 {d31[1]}, [r12]+
+ ldr r2, _CONSTANTS_
+
+ ldrh r12, [r1], #2 ; lo *dc
+ ldrh r1, [r1] ; hi *dc
+
+ ; dequant: q[i] = q[i] * dq[i]
+ vmul.i16 q2, q2, q0
+ vmul.i16 q3, q3, q1
+ vmul.i16 q4, q4, q0
+ vmul.i16 q5, q5, q1
+
+ ; move dc up to neon and overwrite first element
+ vmov.16 d4[0], r12
+ vmov.16 d8[0], r1
+
+ vld1.16 {d0}, [r2]+
+ ; q2: l0r0 q3: l8r8
+ ; q4: l4r4 q5: l12r12
+ vswp d5, d8
+ vswp d7, d10
+
+ ; _CONSTANTS_ * 4,12 >> 16
+ ; q6: 4 * sinpi : c1/temp1
+ ; q7: 12 * sinpi : d1/temp2
+ ; q8: 4 * cospi
+ ; q9: 12 * cospi
+ vqdmulh.s16 q6, q4, d0[2] ; sinpi8sqrt2
+ vqdmulh.s16 q7, q5, d0[2]
+ vqdmulh.s16 q8, q4, d0[0] ; cospi8sqrt2minus1
+ vqdmulh.s16 q9, q5, d0[0]
+
+ vqadd.s16 q10, q2, q3 ; a1 = 0 + 8
+ vqsub.s16 q11, q2, q3 ; b1 = 0 - 8
+
+ ; vqdmulh only accepts signed values. this was a problem because
+ ; our constant had the high bit set, and was treated as a negative value.
+ ; vqdmulh also doubles the value before it shifts by 16. we need to
+ ; compensate for this. in the case of sinpi8sqrt2, the lowest bit is 0,
+ ; so we can shift the constant without losing precision. this avoids
+ ; shift again afterward, but also avoids the sign issue. win win!
+ ; for cospi8sqrt2minus1 the lowest bit is 1, so we lose precision if we
+ ; pre-shift it
+ vshr.s16 q8, q8, #1
+ vshr.s16 q9, q9, #1
+
+ ; q4: 4 + 4 * cospi : d1/temp1
+ ; q5: 12 + 12 * cospi : c1/temp2
+ vqadd.s16 q4, q4, q8
+ vqadd.s16 q5, q5, q9
+
+ ; c1 = temp1 - temp2
+ ; d1 = temp1 + temp2
+ vqsub.s16 q2, q6, q5
+ vqadd.s16 q3, q4, q7
+
+ ; [0]: a1+d1
+ ; [1]: b1+c1
+ ; [2]: b1-c1
+ ; [3]: a1-d1
+ vqadd.s16 q4, q10, q3
+ vqadd.s16 q5, q11, q2
+ vqsub.s16 q6, q11, q2
+ vqsub.s16 q7, q10, q3
+
+ ; rotate
+ vtrn.32 q4, q6
+ vtrn.32 q5, q7
+ vtrn.16 q4, q5
+ vtrn.16 q6, q7
+ ; idct loop 2
+ ; q4: l 0, 4, 8,12 r 0, 4, 8,12
+ ; q5: l 1, 5, 9,13 r 1, 5, 9,13
+ ; q6: l 2, 6,10,14 r 2, 6,10,14
+ ; q7: l 3, 7,11,15 r 3, 7,11,15
+
+ ; q8: 1 * sinpi : c1/temp1
+ ; q9: 3 * sinpi : d1/temp2
+ ; q10: 1 * cospi
+ ; q11: 3 * cospi
+ vqdmulh.s16 q8, q5, d0[2] ; sinpi8sqrt2
+ vqdmulh.s16 q9, q7, d0[2]
+ vqdmulh.s16 q10, q5, d0[0] ; cospi8sqrt2minus1
+ vqdmulh.s16 q11, q7, d0[0]
+
+ vqadd.s16 q2, q4, q6 ; a1 = 0 + 2
+ vqsub.s16 q3, q4, q6 ; b1 = 0 - 2
+
+ ; see note on shifting above
+ vshr.s16 q10, q10, #1
+ vshr.s16 q11, q11, #1
+
+ ; q10: 1 + 1 * cospi : d1/temp1
+ ; q11: 3 + 3 * cospi : c1/temp2
+ vqadd.s16 q10, q5, q10
+ vqadd.s16 q11, q7, q11
+
+ ; q8: c1 = temp1 - temp2
+ ; q9: d1 = temp1 + temp2
+ vqsub.s16 q8, q8, q11
+ vqadd.s16 q9, q10, q9
+
+ ; a1+d1
+ ; b1+c1
+ ; b1-c1
+ ; a1-d1
+ vqadd.s16 q4, q2, q9
+ vqadd.s16 q5, q3, q8
+ vqsub.s16 q6, q3, q8
+ vqsub.s16 q7, q2, q9
+
+ ; +4 >> 3 (rounding)
+ vrshr.s16 q4, q4, #3 ; lo
+ vrshr.s16 q5, q5, #3
+ vrshr.s16 q6, q6, #3 ; hi
+ vrshr.s16 q7, q7, #3
+
+ vtrn.32 q4, q6
+ vtrn.32 q5, q7
+ vtrn.16 q4, q5
+ vtrn.16 q6, q7
+
+ ; adding pre
+ ; input is still packed. pre was read interleaved
+ vaddw.u8 q4, q4, d28
+ vaddw.u8 q5, q5, d29
+ vaddw.u8 q6, q6, d30
+ vaddw.u8 q7, q7, d31
+
+ vmov.i16 q14, #0
+ vmov q15, q14
+ vst1.16 {q14, q15}, [r0] ; write over high input+ sub r0, r0, #32
+ vst1.16 {q14, q15}, [r0] ; write over low input+
+ ;saturate and narrow
+ vqmovun.s16 d0, q4 ; lo
+ vqmovun.s16 d1, q5
+ vqmovun.s16 d2, q6 ; hi
+ vqmovun.s16 d3, q7
+
+ ldr r1, [sp] ; stride
+ add r2, r3, #4 ; hi
+ vst1.32 {d0[0]}, [r3], r1 ; lo+ vst1.32 {d0[1]}, [r2], r1 ; hi+ vst1.32 {d1[0]}, [r3], r1+ vst1.32 {d1[1]}, [r2], r1+ vst1.32 {d2[0]}, [r3], r1+ vst1.32 {d2[1]}, [r2], r1+ vst1.32 {d3[0]}, [r3]+ vst1.32 {d3[1]}, [r2]+
+ bx lr
+
+ ENDP ; |idct_dequant_dc_full_2x_neon|
+
+; Constant Pool
+_CONSTANTS_ DCD cospi8sqrt2minus1
+cospi8sqrt2minus1 DCD 0x4e7b
+; because the lowest bit in 0x8a8c is 0, we can pre-shift this
+sinpi8sqrt2 DCD 0x4546
+
+ END
--- /dev/null
+++ b/vp8/decoder/arm/neon/idct_dequant_full_2x_neon.asm
@@ -1,0 +1,198 @@
+;
+; Copyright (c) 2010 The Webm project authors. All Rights Reserved.
+;
+; Use of this source code is governed by a BSD-style license
+; that can be found in the LICENSE file in the root of the source
+; tree. An additional intellectual property rights grant can be found
+; in the file PATENTS. All contributing project authors may
+; be found in the AUTHORS file in the root of the source tree.
+;
+
+
+ EXPORT |idct_dequant_full_2x_neon|
+ ARM
+ REQUIRE8
+ PRESERVE8
+
+ AREA ||.text||, CODE, READONLY, ALIGN=2
+;void idct_dequant_full_2x_neon(short *q, short *dq, unsigned char *pre,
+; unsigned char *dst, int pitch, int stride);
+; r0 *q,
+; r1 *dq,
+; r2 *pre
+; r3 *dst
+; sp pitch
+; sp+4 stride
+|idct_dequant_full_2x_neon| PROC
+ vld1.16 {q0, q1}, [r1] ; dq (same l/r)+ vld1.16 {q2, q3}, [r0] ; l q+ ldr r1, [sp] ; pitch
+ add r0, r0, #32
+ vld1.16 {q4, q5}, [r0] ; r q+ add r12, r2, #4
+ ; interleave the predictors
+ vld1.32 {d28[0]}, [r2], r1 ; l pre+ vld1.32 {d28[1]}, [r12], r1 ; r pre+ vld1.32 {d29[0]}, [r2], r1+ vld1.32 {d29[1]}, [r12], r1+ vld1.32 {d30[0]}, [r2], r1+ vld1.32 {d30[1]}, [r12], r1+ vld1.32 {d31[0]}, [r2]+ vld1.32 {d31[1]}, [r12]+
+ ldr r2, _CONSTANTS_
+
+ ; dequant: q[i] = q[i] * dq[i]
+ vmul.i16 q2, q2, q0
+ vmul.i16 q3, q3, q1
+ vmul.i16 q4, q4, q0
+ vmul.i16 q5, q5, q1
+
+ vld1.16 {d0}, [r2]+
+ ; q2: l0r0 q3: l8r8
+ ; q4: l4r4 q5: l12r12
+ vswp d5, d8
+ vswp d7, d10
+
+ ; _CONSTANTS_ * 4,12 >> 16
+ ; q6: 4 * sinpi : c1/temp1
+ ; q7: 12 * sinpi : d1/temp2
+ ; q8: 4 * cospi
+ ; q9: 12 * cospi
+ vqdmulh.s16 q6, q4, d0[2] ; sinpi8sqrt2
+ vqdmulh.s16 q7, q5, d0[2]
+ vqdmulh.s16 q8, q4, d0[0] ; cospi8sqrt2minus1
+ vqdmulh.s16 q9, q5, d0[0]
+
+ vqadd.s16 q10, q2, q3 ; a1 = 0 + 8
+ vqsub.s16 q11, q2, q3 ; b1 = 0 - 8
+
+ ; vqdmulh only accepts signed values. this was a problem because
+ ; our constant had the high bit set, and was treated as a negative value.
+ ; vqdmulh also doubles the value before it shifts by 16. we need to
+ ; compensate for this. in the case of sinpi8sqrt2, the lowest bit is 0,
+ ; so we can shift the constant without losing precision. this avoids
+ ; shift again afterward, but also avoids the sign issue. win win!
+ ; for cospi8sqrt2minus1 the lowest bit is 1, so we lose precision if we
+ ; pre-shift it
+ vshr.s16 q8, q8, #1
+ vshr.s16 q9, q9, #1
+
+ ; q4: 4 + 4 * cospi : d1/temp1
+ ; q5: 12 + 12 * cospi : c1/temp2
+ vqadd.s16 q4, q4, q8
+ vqadd.s16 q5, q5, q9
+
+ ; c1 = temp1 - temp2
+ ; d1 = temp1 + temp2
+ vqsub.s16 q2, q6, q5
+ vqadd.s16 q3, q4, q7
+
+ ; [0]: a1+d1
+ ; [1]: b1+c1
+ ; [2]: b1-c1
+ ; [3]: a1-d1
+ vqadd.s16 q4, q10, q3
+ vqadd.s16 q5, q11, q2
+ vqsub.s16 q6, q11, q2
+ vqsub.s16 q7, q10, q3
+
+ ; rotate
+ vtrn.32 q4, q6
+ vtrn.32 q5, q7
+ vtrn.16 q4, q5
+ vtrn.16 q6, q7
+ ; idct loop 2
+ ; q4: l 0, 4, 8,12 r 0, 4, 8,12
+ ; q5: l 1, 5, 9,13 r 1, 5, 9,13
+ ; q6: l 2, 6,10,14 r 2, 6,10,14
+ ; q7: l 3, 7,11,15 r 3, 7,11,15
+
+ ; q8: 1 * sinpi : c1/temp1
+ ; q9: 3 * sinpi : d1/temp2
+ ; q10: 1 * cospi
+ ; q11: 3 * cospi
+ vqdmulh.s16 q8, q5, d0[2] ; sinpi8sqrt2
+ vqdmulh.s16 q9, q7, d0[2]
+ vqdmulh.s16 q10, q5, d0[0] ; cospi8sqrt2minus1
+ vqdmulh.s16 q11, q7, d0[0]
+
+ vqadd.s16 q2, q4, q6 ; a1 = 0 + 2
+ vqsub.s16 q3, q4, q6 ; b1 = 0 - 2
+
+ ; see note on shifting above
+ vshr.s16 q10, q10, #1
+ vshr.s16 q11, q11, #1
+
+ ; q10: 1 + 1 * cospi : d1/temp1
+ ; q11: 3 + 3 * cospi : c1/temp2
+ vqadd.s16 q10, q5, q10
+ vqadd.s16 q11, q7, q11
+
+ ; q8: c1 = temp1 - temp2
+ ; q9: d1 = temp1 + temp2
+ vqsub.s16 q8, q8, q11
+ vqadd.s16 q9, q10, q9
+
+ ; a1+d1
+ ; b1+c1
+ ; b1-c1
+ ; a1-d1
+ vqadd.s16 q4, q2, q9
+ vqadd.s16 q5, q3, q8
+ vqsub.s16 q6, q3, q8
+ vqsub.s16 q7, q2, q9
+
+ ; +4 >> 3 (rounding)
+ vrshr.s16 q4, q4, #3 ; lo
+ vrshr.s16 q5, q5, #3
+ vrshr.s16 q6, q6, #3 ; hi
+ vrshr.s16 q7, q7, #3
+
+ vtrn.32 q4, q6
+ vtrn.32 q5, q7
+ vtrn.16 q4, q5
+ vtrn.16 q6, q7
+
+ ; adding pre
+ ; input is still packed. pre was read interleaved
+ vaddw.u8 q4, q4, d28
+ vaddw.u8 q5, q5, d29
+ vaddw.u8 q6, q6, d30
+ vaddw.u8 q7, q7, d31
+
+ vmov.i16 q14, #0
+ vmov q15, q14
+ vst1.16 {q14, q15}, [r0] ; write over high input+ sub r0, r0, #32
+ vst1.16 {q14, q15}, [r0] ; write over low input+
+ ;saturate and narrow
+ vqmovun.s16 d0, q4 ; lo
+ vqmovun.s16 d1, q5
+ vqmovun.s16 d2, q6 ; hi
+ vqmovun.s16 d3, q7
+
+ ldr r1, [sp, #4] ; stride
+ add r2, r3, #4 ; hi
+ vst1.32 {d0[0]}, [r3], r1 ; lo+ vst1.32 {d0[1]}, [r2], r1 ; hi+ vst1.32 {d1[0]}, [r3], r1+ vst1.32 {d1[1]}, [r2], r1+ vst1.32 {d2[0]}, [r3], r1+ vst1.32 {d2[1]}, [r2], r1+ vst1.32 {d3[0]}, [r3]+ vst1.32 {d3[1]}, [r2]+
+ bx lr
+
+ ENDP ; |idct_dequant_full_2x_neon|
+
+; Constant Pool
+_CONSTANTS_ DCD cospi8sqrt2minus1
+cospi8sqrt2minus1 DCD 0x4e7b
+; because the lowest bit in 0x8a8c is 0, we can pre-shift this
+sinpi8sqrt2 DCD 0x4546
+
+ END
--- a/vp8/vp8dx_arm.mk
+++ b/vp8/vp8dx_arm.mk
@@ -25,7 +25,10 @@
VP8_DX_SRCS-$(HAVE_ARMV6) += decoder/arm/armv6/idct_blk_v6.c
#File list for neon
-VP8_DX_SRCS-$(HAVE_ARMV7) += decoder/arm/neon/dequant_dc_idct_neon$(ASM)
+VP8_DX_SRCS-$(HAVE_ARMV7) += decoder/arm/neon/idct_dequant_dc_full_2x_neon$(ASM)
+VP8_DX_SRCS-$(HAVE_ARMV7) += decoder/arm/neon/idct_dequant_dc_0_2x_neon$(ASM)
VP8_DX_SRCS-$(HAVE_ARMV7) += decoder/arm/neon/dequant_idct_neon$(ASM)
+VP8_DX_SRCS-$(HAVE_ARMV7) += decoder/arm/neon/idct_dequant_full_2x_neon$(ASM)
+VP8_DX_SRCS-$(HAVE_ARMV7) += decoder/arm/neon/idct_dequant_0_2x_neon$(ASM)
VP8_DX_SRCS-$(HAVE_ARMV7) += decoder/arm/neon/dequantizeb_neon$(ASM)
VP8_DX_SRCS-$(HAVE_ARMV7) += decoder/arm/neon/idct_blk_neon.c
--
⑨