generic-library/vpx
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity
c7cfde42a9
vpx/vp8/encoder/x86/variance_impl_sse2.asm

1476 lines
49 KiB
NASM
Raw Normal View History

Initial WebM release
2010-05-18 17:58:33 +02:00
;
Use WebM in copyright notice for consistency Changes 'The VP8 project' to 'The WebM project', for consistency with other webmproject.org repositories. Fixes issue #97. Change-Id: I37c13ed5fbdb9d334ceef71c6350e9febed9bbba
2010-09-09 14:16:39 +02:00
; Copyright (c) 2010 The WebM project authors. All Rights Reserved.
Initial WebM release
2010-05-18 17:58:33 +02:00
;
cosmetics: trim trailing whitespace When the license headers were updated, they accidentally contained trailing whitespace, so unfortunately we have to touch all the files again. Change-Id: I236c05fade06589e417179c0444cb39b09e4200d
2010-06-18 18:39:21 +02:00
; Use of this source code is governed by a BSD-style license
LICENSE: update with latest text Change-Id: Ieebea089095d9073b3a94932791099f614ce120c
2010-06-04 22:19:40 +02:00
; that can be found in the LICENSE file in the root of the source
; tree. An additional intellectual property rights grant can be found
cosmetics: trim trailing whitespace When the license headers were updated, they accidentally contained trailing whitespace, so unfortunately we have to touch all the files again. Change-Id: I236c05fade06589e417179c0444cb39b09e4200d
2010-06-18 18:39:21 +02:00
; in the file PATENTS. All contributing project authors may
LICENSE: update with latest text Change-Id: Ieebea089095d9073b3a94932791099f614ce120c
2010-06-04 22:19:40 +02:00
; be found in the AUTHORS file in the root of the source tree.
Initial WebM release
2010-05-18 17:58:33 +02:00
;
%include "vpx_ports/x86_abi_support.asm"
%define xmm_filter_shift 7
;unsigned int vp8_get_mb_ss_sse2
;(
; short *src_ptr
;)
global sym(vp8_get_mb_ss_sse2)
sym(vp8_get_mb_ss_sse2):
push rbp
mov rbp, rsp
SHADOW_ARGS_TO_STACK 1
GET_GOT rbx
push rsi
push rdi
sub rsp, 16
; end prolog
mov rax, arg(0) ;[src_ptr]
mov rcx, 8
pxor xmm4, xmm4
NEXTROW:
movdqa xmm0, [rax]
movdqa xmm1, [rax+16]
movdqa xmm2, [rax+32]
movdqa xmm3, [rax+48]
pmaddwd xmm0, xmm0
pmaddwd xmm1, xmm1
pmaddwd xmm2, xmm2
pmaddwd xmm3, xmm3
paddd xmm0, xmm1
paddd xmm2, xmm3
paddd xmm4, xmm0
paddd xmm4, xmm2
add rax, 0x40
dec rcx
ja NEXTROW
movdqa xmm3,xmm4
psrldq xmm4,8
paddd xmm4,xmm3
movdqa xmm3,xmm4
psrldq xmm4,4
paddd xmm4,xmm3
nasm: match instruction length (movd/movq) to parameters nasm requires the instruction length (movd/movq) to match to its parameters. I find it more clear to really use 64bit instructions when we use 64bit registers in the assembly. Provide nasm compatibility. No binary change by this patch with yasm on {x86_64,i686}-fedora13-linux-gnu. Few longer opcodes with nasm on {x86_64,i686}-fedora13-linux-gnu have been checked as safe. Change-Id: Id9b1a5cdfb1bc05697e523c317a296df43d42a91
2010-10-04 23:19:33 +02:00
movq rax,xmm4
Initial WebM release
2010-05-18 17:58:33 +02:00
; begin epilog
add rsp, 16
pop rdi
pop rsi
RESTORE_GOT
UNSHADOW_ARGS
pop rbp
ret
;unsigned int vp8_get16x16var_sse2
;(
; unsigned char * src_ptr,
; int source_stride,
; unsigned char * ref_ptr,
; int recon_stride,
; unsigned int * SSE,
; int * Sum
;)
global sym(vp8_get16x16var_sse2)
sym(vp8_get16x16var_sse2):
push rbp
mov rbp, rsp
SHADOW_ARGS_TO_STACK 6
Add save/restore xmm registers in x86 assembly code Went through the code and fixed it. Verified on Windows. Where possible, remove dependencies on xmm[67] Current code relies on pushing rbp to the stack to get 16 byte alignment. This broke when rbp wasn't pushed (vp8/encoder/x86/sad_sse3.asm). Work around this by using unaligned memory accesses. Revisit this and the offsets in vp8/encoder/x86/sad_sse3.asm in another change to SAVE_XMM. Change-Id: I5f940994d3ebfd977c3d68446cef20fd78b07877
2011-04-07 19:17:22 +02:00
SAVE_XMM
Add prefetch before variance calculation This improved encoding performance by 0.5% (good, speed 1) to 1.5% (good, speed 5). Change-Id: I843d72a0d68a90b5f694adf770943e4a4618f50e
2011-02-23 00:01:08 +01:00
push rbx
Initial WebM release
2010-05-18 17:58:33 +02:00
push rsi
push rdi
; end prolog
mov rsi, arg(0) ;[src_ptr]
mov rdi, arg(2) ;[ref_ptr]
movsxd rax, DWORD PTR arg(1) ;[source_stride]
movsxd rdx, DWORD PTR arg(3) ;[recon_stride]
Add prefetch before variance calculation This improved encoding performance by 0.5% (good, speed 1) to 1.5% (good, speed 5). Change-Id: I843d72a0d68a90b5f694adf770943e4a4618f50e
2011-02-23 00:01:08 +01:00
; Prefetch data
lea rcx, [rax+rax*2]
prefetcht0 [rsi]
prefetcht0 [rsi+rax]
prefetcht0 [rsi+rax*2]
prefetcht0 [rsi+rcx]
lea rbx, [rsi+rax*4]
prefetcht0 [rbx]
prefetcht0 [rbx+rax]
prefetcht0 [rbx+rax*2]
prefetcht0 [rbx+rcx]
lea rcx, [rdx+rdx*2]
prefetcht0 [rdi]
prefetcht0 [rdi+rdx]
prefetcht0 [rdi+rdx*2]
prefetcht0 [rdi+rcx]
lea rbx, [rdi+rdx*4]
prefetcht0 [rbx]
prefetcht0 [rbx+rdx]
prefetcht0 [rbx+rdx*2]
prefetcht0 [rbx+rcx]
Initial WebM release
2010-05-18 17:58:33 +02:00
pxor xmm0, xmm0 ; clear xmm0 for unpack
pxor xmm7, xmm7 ; clear xmm7 for accumulating diffs
pxor xmm6, xmm6 ; clear xmm6 for accumulating sse
mov rcx, 16
var16loop:
movdqu xmm1, XMMWORD PTR [rsi]
movdqu xmm2, XMMWORD PTR [rdi]
Add prefetch before variance calculation This improved encoding performance by 0.5% (good, speed 1) to 1.5% (good, speed 5). Change-Id: I843d72a0d68a90b5f694adf770943e4a4618f50e
2011-02-23 00:01:08 +01:00
prefetcht0 [rsi+rax*8]
prefetcht0 [rdi+rdx*8]
Initial WebM release
2010-05-18 17:58:33 +02:00
movdqa xmm3, xmm1
movdqa xmm4, xmm2
punpcklbw xmm1, xmm0
punpckhbw xmm3, xmm0
punpcklbw xmm2, xmm0
punpckhbw xmm4, xmm0
psubw xmm1, xmm2
psubw xmm3, xmm4
paddw xmm7, xmm1
pmaddwd xmm1, xmm1
paddw xmm7, xmm3
pmaddwd xmm3, xmm3
paddd xmm6, xmm1
paddd xmm6, xmm3
add rsi, rax
add rdi, rdx
sub rcx, 1
jnz var16loop
movdqa xmm1, xmm6
pxor xmm6, xmm6
pxor xmm5, xmm5
punpcklwd xmm6, xmm7
punpckhwd xmm5, xmm7
psrad xmm5, 16
psrad xmm6, 16
paddd xmm6, xmm5
movdqa xmm2, xmm1
punpckldq xmm1, xmm0
punpckhdq xmm2, xmm0
movdqa xmm7, xmm6
paddd xmm1, xmm2
punpckldq xmm6, xmm0
punpckhdq xmm7, xmm0
paddd xmm6, xmm7
movdqa xmm2, xmm1
movdqa xmm7, xmm6
psrldq xmm1, 8
psrldq xmm6, 8
paddd xmm7, xmm6
paddd xmm1, xmm2
mov rax, arg(5) ;[Sum]
mov rdi, arg(4) ;[SSE]
movd DWORD PTR [rax], xmm7
movd DWORD PTR [rdi], xmm1
; begin epilog
pop rdi
pop rsi
Add prefetch before variance calculation This improved encoding performance by 0.5% (good, speed 1) to 1.5% (good, speed 5). Change-Id: I843d72a0d68a90b5f694adf770943e4a4618f50e
2011-02-23 00:01:08 +01:00
pop rbx
Add save/restore xmm registers in x86 assembly code Went through the code and fixed it. Verified on Windows. Where possible, remove dependencies on xmm[67] Current code relies on pushing rbp to the stack to get 16 byte alignment. This broke when rbp wasn't pushed (vp8/encoder/x86/sad_sse3.asm). Work around this by using unaligned memory accesses. Revisit this and the offsets in vp8/encoder/x86/sad_sse3.asm in another change to SAVE_XMM. Change-Id: I5f940994d3ebfd977c3d68446cef20fd78b07877
2011-04-07 19:17:22 +02:00
RESTORE_XMM
Initial WebM release
2010-05-18 17:58:33 +02:00
UNSHADOW_ARGS
pop rbp
ret
;unsigned int vp8_get16x16pred_error_sse2
;(
; unsigned char *src_ptr,
; int src_stride,
; unsigned char *ref_ptr,
; int ref_stride
;)
global sym(vp8_get16x16pred_error_sse2)
sym(vp8_get16x16pred_error_sse2):
push rbp
mov rbp, rsp
SHADOW_ARGS_TO_STACK 4
Add save/restore xmm registers in x86 assembly code Went through the code and fixed it. Verified on Windows. Where possible, remove dependencies on xmm[67] Current code relies on pushing rbp to the stack to get 16 byte alignment. This broke when rbp wasn't pushed (vp8/encoder/x86/sad_sse3.asm). Work around this by using unaligned memory accesses. Revisit this and the offsets in vp8/encoder/x86/sad_sse3.asm in another change to SAVE_XMM. Change-Id: I5f940994d3ebfd977c3d68446cef20fd78b07877
2011-04-07 19:17:22 +02:00
SAVE_XMM
Initial WebM release
2010-05-18 17:58:33 +02:00
GET_GOT rbx
push rsi
push rdi
sub rsp, 16
; end prolog
mov rsi, arg(0) ;[src_ptr]
mov rdi, arg(2) ;[ref_ptr]
movsxd rax, DWORD PTR arg(1) ;[src_stride]
movsxd rdx, DWORD PTR arg(3) ;[ref_stride]
pxor xmm0, xmm0 ; clear xmm0 for unpack
pxor xmm7, xmm7 ; clear xmm7 for accumulating diffs
pxor xmm6, xmm6 ; clear xmm6 for accumulating sse
mov rcx, 16
var16peloop:
movdqu xmm1, XMMWORD PTR [rsi]
movdqu xmm2, XMMWORD PTR [rdi]
movdqa xmm3, xmm1
movdqa xmm4, xmm2
punpcklbw xmm1, xmm0
punpckhbw xmm3, xmm0
punpcklbw xmm2, xmm0
punpckhbw xmm4, xmm0
psubw xmm1, xmm2
psubw xmm3, xmm4
paddw xmm7, xmm1
pmaddwd xmm1, xmm1
paddw xmm7, xmm3
pmaddwd xmm3, xmm3
paddd xmm6, xmm1
paddd xmm6, xmm3
add rsi, rax
add rdi, rdx
sub rcx, 1
jnz var16peloop
movdqa xmm1, xmm6
pxor xmm6, xmm6
pxor xmm5, xmm5
punpcklwd xmm6, xmm7
punpckhwd xmm5, xmm7
psrad xmm5, 16
psrad xmm6, 16
paddd xmm6, xmm5
movdqa xmm2, xmm1
punpckldq xmm1, xmm0
punpckhdq xmm2, xmm0
movdqa xmm7, xmm6
paddd xmm1, xmm2
punpckldq xmm6, xmm0
punpckhdq xmm7, xmm0
paddd xmm6, xmm7
movdqa xmm2, xmm1
movdqa xmm7, xmm6
psrldq xmm1, 8
psrldq xmm6, 8
paddd xmm7, xmm6
paddd xmm1, xmm2
movd DWORD PTR [rsp], xmm7 ;Sum
movd DWORD PTR [rsp+4], xmm1 ;SSE
; return (SSE-((Sum*Sum)>>8));
movsxd rdx, dword ptr [rsp]
imul rdx, rdx
sar rdx, 8
movsxd rax, dword ptr [rsp + 4]
sub rax, rdx
; begin epilog
add rsp, 16
pop rdi
pop rsi
RESTORE_GOT
Add save/restore xmm registers in x86 assembly code Went through the code and fixed it. Verified on Windows. Where possible, remove dependencies on xmm[67] Current code relies on pushing rbp to the stack to get 16 byte alignment. This broke when rbp wasn't pushed (vp8/encoder/x86/sad_sse3.asm). Work around this by using unaligned memory accesses. Revisit this and the offsets in vp8/encoder/x86/sad_sse3.asm in another change to SAVE_XMM. Change-Id: I5f940994d3ebfd977c3d68446cef20fd78b07877
2011-04-07 19:17:22 +02:00
RESTORE_XMM
Initial WebM release
2010-05-18 17:58:33 +02:00
UNSHADOW_ARGS
pop rbp
ret
;unsigned int vp8_get8x8var_sse2
;(
; unsigned char * src_ptr,
; int source_stride,
; unsigned char * ref_ptr,
; int recon_stride,
; unsigned int * SSE,
; int * Sum
;)
global sym(vp8_get8x8var_sse2)
sym(vp8_get8x8var_sse2):
push rbp
mov rbp, rsp
SHADOW_ARGS_TO_STACK 6
Add save/restore xmm registers in x86 assembly code Went through the code and fixed it. Verified on Windows. Where possible, remove dependencies on xmm[67] Current code relies on pushing rbp to the stack to get 16 byte alignment. This broke when rbp wasn't pushed (vp8/encoder/x86/sad_sse3.asm). Work around this by using unaligned memory accesses. Revisit this and the offsets in vp8/encoder/x86/sad_sse3.asm in another change to SAVE_XMM. Change-Id: I5f940994d3ebfd977c3d68446cef20fd78b07877
2011-04-07 19:17:22 +02:00
SAVE_XMM
Initial WebM release
2010-05-18 17:58:33 +02:00
GET_GOT rbx
push rsi
push rdi
sub rsp, 16
; end prolog
mov rsi, arg(0) ;[src_ptr]
mov rdi, arg(2) ;[ref_ptr]
movsxd rax, DWORD PTR arg(1) ;[source_stride]
movsxd rdx, DWORD PTR arg(3) ;[recon_stride]
pxor xmm0, xmm0 ; clear xmm0 for unpack
pxor xmm7, xmm7 ; clear xmm7 for accumulating diffs
movq xmm1, QWORD PTR [rsi]
movq xmm2, QWORD PTR [rdi]
punpcklbw xmm1, xmm0
punpcklbw xmm2, xmm0
psubsw xmm1, xmm2
paddw xmm7, xmm1
pmaddwd xmm1, xmm1
movq xmm2, QWORD PTR[rsi + rax]
movq xmm3, QWORD PTR[rdi + rdx]
punpcklbw xmm2, xmm0
punpcklbw xmm3, xmm0
psubsw xmm2, xmm3
paddw xmm7, xmm2
pmaddwd xmm2, xmm2
paddd xmm1, xmm2
movq xmm2, QWORD PTR[rsi + rax * 2]
movq xmm3, QWORD PTR[rdi + rdx * 2]
punpcklbw xmm2, xmm0
punpcklbw xmm3, xmm0
psubsw xmm2, xmm3
paddw xmm7, xmm2
pmaddwd xmm2, xmm2
paddd xmm1, xmm2
lea rsi, [rsi + rax * 2]
lea rdi, [rdi + rdx * 2]
movq xmm2, QWORD PTR[rsi + rax]
movq xmm3, QWORD PTR[rdi + rdx]
punpcklbw xmm2, xmm0
punpcklbw xmm3, xmm0
psubsw xmm2, xmm3
paddw xmm7, xmm2
pmaddwd xmm2, xmm2
paddd xmm1, xmm2
movq xmm2, QWORD PTR[rsi + rax *2]
movq xmm3, QWORD PTR[rdi + rdx *2]
punpcklbw xmm2, xmm0
punpcklbw xmm3, xmm0
psubsw xmm2, xmm3
paddw xmm7, xmm2
pmaddwd xmm2, xmm2
paddd xmm1, xmm2
lea rsi, [rsi + rax * 2]
lea rdi, [rdi + rdx * 2]
movq xmm2, QWORD PTR[rsi + rax]
movq xmm3, QWORD PTR[rdi + rdx]
punpcklbw xmm2, xmm0
punpcklbw xmm3, xmm0
psubsw xmm2, xmm3
paddw xmm7, xmm2
pmaddwd xmm2, xmm2
paddd xmm1, xmm2
movq xmm2, QWORD PTR[rsi + rax *2]
movq xmm3, QWORD PTR[rdi + rdx *2]
punpcklbw xmm2, xmm0
punpcklbw xmm3, xmm0
psubsw xmm2, xmm3
paddw xmm7, xmm2
pmaddwd xmm2, xmm2
paddd xmm1, xmm2
lea rsi, [rsi + rax * 2]
lea rdi, [rdi + rdx * 2]
movq xmm2, QWORD PTR[rsi + rax]
movq xmm3, QWORD PTR[rdi + rdx]
punpcklbw xmm2, xmm0
punpcklbw xmm3, xmm0
psubsw xmm2, xmm3
paddw xmm7, xmm2
pmaddwd xmm2, xmm2
paddd xmm1, xmm2
movdqa xmm6, xmm7
punpcklwd xmm6, xmm0
punpckhwd xmm7, xmm0
movdqa xmm2, xmm1
paddw xmm6, xmm7
punpckldq xmm1, xmm0
punpckhdq xmm2, xmm0
movdqa xmm7, xmm6
paddd xmm1, xmm2
punpckldq xmm6, xmm0
punpckhdq xmm7, xmm0
paddw xmm6, xmm7
movdqa xmm2, xmm1
movdqa xmm7, xmm6
psrldq xmm1, 8
psrldq xmm6, 8
paddw xmm7, xmm6
paddd xmm1, xmm2
mov rax, arg(5) ;[Sum]
mov rdi, arg(4) ;[SSE]
nasm: match instruction length (movd/movq) to parameters nasm requires the instruction length (movd/movq) to match to its parameters. I find it more clear to really use 64bit instructions when we use 64bit registers in the assembly. Provide nasm compatibility. No binary change by this patch with yasm on {x86_64,i686}-fedora13-linux-gnu. Few longer opcodes with nasm on {x86_64,i686}-fedora13-linux-gnu have been checked as safe. Change-Id: Id9b1a5cdfb1bc05697e523c317a296df43d42a91
2010-10-04 23:19:33 +02:00
movq rdx, xmm7
Initial WebM release
2010-05-18 17:58:33 +02:00
movsx rcx, dx
mov dword ptr [rax], ecx
movd DWORD PTR [rdi], xmm1
; begin epilog
add rsp, 16
pop rdi
pop rsi
RESTORE_GOT
Add save/restore xmm registers in x86 assembly code Went through the code and fixed it. Verified on Windows. Where possible, remove dependencies on xmm[67] Current code relies on pushing rbp to the stack to get 16 byte alignment. This broke when rbp wasn't pushed (vp8/encoder/x86/sad_sse3.asm). Work around this by using unaligned memory accesses. Revisit this and the offsets in vp8/encoder/x86/sad_sse3.asm in another change to SAVE_XMM. Change-Id: I5f940994d3ebfd977c3d68446cef20fd78b07877
2011-04-07 19:17:22 +02:00
RESTORE_XMM
Initial WebM release
2010-05-18 17:58:33 +02:00
UNSHADOW_ARGS
pop rbp
ret
;void vp8_filter_block2d_bil_var_sse2
;(
; unsigned char *ref_ptr,
; int ref_pixels_per_line,
; unsigned char *src_ptr,
; int src_pixels_per_line,
; unsigned int Height,
Modify sub-pixel filters to eliminate unnecessary calculations In sub-pixel calculation, xoffset and yoffset mostly take some specific values. Modified sub-pixel filter functions according to these possible values to improve performance. Change-Id: I83083570af8b00ff65093467914fbb97a4e9ea21
2011-01-20 19:01:30 +01:00
; int xoffset,
; int yoffset,
Initial WebM release
2010-05-18 17:58:33 +02:00
; int *sum,
; unsigned int *sumsquared;;
;
;)
global sym(vp8_filter_block2d_bil_var_sse2)
sym(vp8_filter_block2d_bil_var_sse2):
push rbp
mov rbp, rsp
SHADOW_ARGS_TO_STACK 9
Modify sub-pixel filters to eliminate unnecessary calculations In sub-pixel calculation, xoffset and yoffset mostly take some specific values. Modified sub-pixel filter functions according to these possible values to improve performance. Change-Id: I83083570af8b00ff65093467914fbb97a4e9ea21
2011-01-20 19:01:30 +01:00
SAVE_XMM
Initial WebM release
2010-05-18 17:58:33 +02:00
GET_GOT rbx
push rsi
push rdi
Modify sub-pixel filters to eliminate unnecessary calculations In sub-pixel calculation, xoffset and yoffset mostly take some specific values. Modified sub-pixel filter functions according to these possible values to improve performance. Change-Id: I83083570af8b00ff65093467914fbb97a4e9ea21
2011-01-20 19:01:30 +01:00
push rbx
Initial WebM release
2010-05-18 17:58:33 +02:00
; end prolog
pxor xmm6, xmm6 ;
pxor xmm7, xmm7 ;
Modify sub-pixel filters to eliminate unnecessary calculations In sub-pixel calculation, xoffset and yoffset mostly take some specific values. Modified sub-pixel filter functions according to these possible values to improve performance. Change-Id: I83083570af8b00ff65093467914fbb97a4e9ea21
2011-01-20 19:01:30 +01:00
lea rsi, [GLOBAL(xmm_bi_rd)] ; rounding
movdqa xmm4, XMMWORD PTR [rsi]
Initial WebM release
2010-05-18 17:58:33 +02:00
Modify sub-pixel filters to eliminate unnecessary calculations In sub-pixel calculation, xoffset and yoffset mostly take some specific values. Modified sub-pixel filter functions according to these possible values to improve performance. Change-Id: I83083570af8b00ff65093467914fbb97a4e9ea21
2011-01-20 19:01:30 +01:00
lea rcx, [GLOBAL(vp8_bilinear_filters_sse2)]
movsxd rax, dword ptr arg(5) ; xoffset
cmp rax, 0 ; skip first_pass filter if xoffset=0
je filter_block2d_bil_var_sse2_sp_only
shl rax, 5 ; point to filter coeff with xoffset
lea rax, [rax + rcx] ; HFilter
movsxd rdx, dword ptr arg(6) ; yoffset
cmp rdx, 0 ; skip second_pass filter if yoffset=0
je filter_block2d_bil_var_sse2_fp_only
shl rdx, 5
lea rdx, [rdx + rcx] ; VFilter
mov rsi, arg(0) ;ref_ptr
mov rdi, arg(2) ;src_ptr
movsxd rcx, dword ptr arg(4) ;Height
Initial WebM release
2010-05-18 17:58:33 +02:00
pxor xmm0, xmm0 ;
Modify sub-pixel filters to eliminate unnecessary calculations In sub-pixel calculation, xoffset and yoffset mostly take some specific values. Modified sub-pixel filter functions according to these possible values to improve performance. Change-Id: I83083570af8b00ff65093467914fbb97a4e9ea21
2011-01-20 19:01:30 +01:00
movq xmm1, QWORD PTR [rsi] ;
movq xmm3, QWORD PTR [rsi+1] ;
Initial WebM release
2010-05-18 17:58:33 +02:00
punpcklbw xmm1, xmm0 ;
Modify sub-pixel filters to eliminate unnecessary calculations In sub-pixel calculation, xoffset and yoffset mostly take some specific values. Modified sub-pixel filter functions according to these possible values to improve performance. Change-Id: I83083570af8b00ff65093467914fbb97a4e9ea21
2011-01-20 19:01:30 +01:00
pmullw xmm1, [rax] ;
Initial WebM release
2010-05-18 17:58:33 +02:00
punpcklbw xmm3, xmm0
pmullw xmm3, [rax+16] ;
Modify sub-pixel filters to eliminate unnecessary calculations In sub-pixel calculation, xoffset and yoffset mostly take some specific values. Modified sub-pixel filter functions according to these possible values to improve performance. Change-Id: I83083570af8b00ff65093467914fbb97a4e9ea21
2011-01-20 19:01:30 +01:00
paddw xmm1, xmm3 ;
paddw xmm1, xmm4 ;
psraw xmm1, xmm_filter_shift ;
Initial WebM release
2010-05-18 17:58:33 +02:00
movdqa xmm5, xmm1
Modify sub-pixel filters to eliminate unnecessary calculations In sub-pixel calculation, xoffset and yoffset mostly take some specific values. Modified sub-pixel filter functions according to these possible values to improve performance. Change-Id: I83083570af8b00ff65093467914fbb97a4e9ea21
2011-01-20 19:01:30 +01:00
movsxd rbx, dword ptr arg(1) ;ref_pixels_per_line
lea rsi, [rsi + rbx]
%if ABI_IS_32BIT=0
movsxd r9, dword ptr arg(3) ;src_pixels_per_line
Initial WebM release
2010-05-18 17:58:33 +02:00
%endif
Modify sub-pixel filters to eliminate unnecessary calculations In sub-pixel calculation, xoffset and yoffset mostly take some specific values. Modified sub-pixel filter functions according to these possible values to improve performance. Change-Id: I83083570af8b00ff65093467914fbb97a4e9ea21
2011-01-20 19:01:30 +01:00
filter_block2d_bil_var_sse2_loop:
Initial WebM release
2010-05-18 17:58:33 +02:00
movq xmm1, QWORD PTR [rsi] ;
movq xmm3, QWORD PTR [rsi+1] ;
punpcklbw xmm1, xmm0 ;
pmullw xmm1, [rax] ;
punpcklbw xmm3, xmm0 ;
pmullw xmm3, [rax+16] ;
paddw xmm1, xmm3 ;
Modify sub-pixel filters to eliminate unnecessary calculations In sub-pixel calculation, xoffset and yoffset mostly take some specific values. Modified sub-pixel filter functions according to these possible values to improve performance. Change-Id: I83083570af8b00ff65093467914fbb97a4e9ea21
2011-01-20 19:01:30 +01:00
paddw xmm1, xmm4 ;
Initial WebM release
2010-05-18 17:58:33 +02:00
psraw xmm1, xmm_filter_shift ;
Modify sub-pixel filters to eliminate unnecessary calculations In sub-pixel calculation, xoffset and yoffset mostly take some specific values. Modified sub-pixel filter functions according to these possible values to improve performance. Change-Id: I83083570af8b00ff65093467914fbb97a4e9ea21
2011-01-20 19:01:30 +01:00
movdqa xmm3, xmm5 ;
Initial WebM release
2010-05-18 17:58:33 +02:00
movdqa xmm5, xmm1 ;
Modify sub-pixel filters to eliminate unnecessary calculations In sub-pixel calculation, xoffset and yoffset mostly take some specific values. Modified sub-pixel filter functions according to these possible values to improve performance. Change-Id: I83083570af8b00ff65093467914fbb97a4e9ea21
2011-01-20 19:01:30 +01:00
pmullw xmm3, [rdx] ;
Initial WebM release
2010-05-18 17:58:33 +02:00
pmullw xmm1, [rdx+16] ;
paddw xmm1, xmm3 ;
Modify sub-pixel filters to eliminate unnecessary calculations In sub-pixel calculation, xoffset and yoffset mostly take some specific values. Modified sub-pixel filter functions according to these possible values to improve performance. Change-Id: I83083570af8b00ff65093467914fbb97a4e9ea21
2011-01-20 19:01:30 +01:00
paddw xmm1, xmm4 ;
Initial WebM release
2010-05-18 17:58:33 +02:00
psraw xmm1, xmm_filter_shift ;
movq xmm3, QWORD PTR [rdi] ;
punpcklbw xmm3, xmm0 ;
psubw xmm1, xmm3 ;
paddw xmm6, xmm1 ;
pmaddwd xmm1, xmm1 ;
paddd xmm7, xmm1 ;
Modify sub-pixel filters to eliminate unnecessary calculations In sub-pixel calculation, xoffset and yoffset mostly take some specific values. Modified sub-pixel filter functions according to these possible values to improve performance. Change-Id: I83083570af8b00ff65093467914fbb97a4e9ea21
2011-01-20 19:01:30 +01:00
lea rsi, [rsi + rbx] ;ref_pixels_per_line
Initial WebM release
2010-05-18 17:58:33 +02:00
%if ABI_IS_32BIT
Modify sub-pixel filters to eliminate unnecessary calculations In sub-pixel calculation, xoffset and yoffset mostly take some specific values. Modified sub-pixel filter functions according to these possible values to improve performance. Change-Id: I83083570af8b00ff65093467914fbb97a4e9ea21
2011-01-20 19:01:30 +01:00
add rdi, dword ptr arg(3) ;src_pixels_per_line
Initial WebM release
2010-05-18 17:58:33 +02:00
%else
Modify sub-pixel filters to eliminate unnecessary calculations In sub-pixel calculation, xoffset and yoffset mostly take some specific values. Modified sub-pixel filter functions according to these possible values to improve performance. Change-Id: I83083570af8b00ff65093467914fbb97a4e9ea21
2011-01-20 19:01:30 +01:00
lea rdi, [rdi + r9]
Initial WebM release
2010-05-18 17:58:33 +02:00
%endif
sub rcx, 1 ;
jnz filter_block2d_bil_var_sse2_loop ;
Modify sub-pixel filters to eliminate unnecessary calculations In sub-pixel calculation, xoffset and yoffset mostly take some specific values. Modified sub-pixel filter functions according to these possible values to improve performance. Change-Id: I83083570af8b00ff65093467914fbb97a4e9ea21
2011-01-20 19:01:30 +01:00
jmp filter_block2d_bil_variance
filter_block2d_bil_var_sse2_sp_only:
movsxd rdx, dword ptr arg(6) ; yoffset
Add zero offset checking in SSE2 sub-pixel filter function Skip filter at zero offset. Change-Id: I95fc7e211869bc0ab5bcfb7ab2e3259d1c0ccf38
2011-03-08 21:22:07 +01:00
cmp rdx, 0 ; skip all if both xoffset=0 and yoffset=0
je filter_block2d_bil_var_sse2_full_pixel
Modify sub-pixel filters to eliminate unnecessary calculations In sub-pixel calculation, xoffset and yoffset mostly take some specific values. Modified sub-pixel filter functions according to these possible values to improve performance. Change-Id: I83083570af8b00ff65093467914fbb97a4e9ea21
2011-01-20 19:01:30 +01:00
shl rdx, 5
lea rdx, [rdx + rcx] ; VFilter
mov rsi, arg(0) ;ref_ptr
mov rdi, arg(2) ;src_ptr
movsxd rcx, dword ptr arg(4) ;Height
movsxd rax, dword ptr arg(1) ;ref_pixels_per_line
pxor xmm0, xmm0 ;
movq xmm1, QWORD PTR [rsi] ;
punpcklbw xmm1, xmm0 ;
movsxd rbx, dword ptr arg(3) ;src_pixels_per_line
lea rsi, [rsi + rax]
filter_block2d_bil_sp_only_loop:
movq xmm3, QWORD PTR [rsi] ;
punpcklbw xmm3, xmm0 ;
movdqa xmm5, xmm3
pmullw xmm1, [rdx] ;
pmullw xmm3, [rdx+16] ;
paddw xmm1, xmm3 ;
paddw xmm1, xmm4 ;
psraw xmm1, xmm_filter_shift ;
movq xmm3, QWORD PTR [rdi] ;
punpcklbw xmm3, xmm0 ;
psubw xmm1, xmm3 ;
paddw xmm6, xmm1 ;
pmaddwd xmm1, xmm1 ;
paddd xmm7, xmm1 ;
movdqa xmm1, xmm5 ;
lea rsi, [rsi + rax] ;ref_pixels_per_line
lea rdi, [rdi + rbx] ;src_pixels_per_line
sub rcx, 1 ;
jnz filter_block2d_bil_sp_only_loop ;
jmp filter_block2d_bil_variance
Add zero offset checking in SSE2 sub-pixel filter function Skip filter at zero offset. Change-Id: I95fc7e211869bc0ab5bcfb7ab2e3259d1c0ccf38
2011-03-08 21:22:07 +01:00
filter_block2d_bil_var_sse2_full_pixel:
mov rsi, arg(0) ;ref_ptr
mov rdi, arg(2) ;src_ptr
movsxd rcx, dword ptr arg(4) ;Height
movsxd rax, dword ptr arg(1) ;ref_pixels_per_line
movsxd rbx, dword ptr arg(3) ;src_pixels_per_line
pxor xmm0, xmm0 ;
filter_block2d_bil_full_pixel_loop:
movq xmm1, QWORD PTR [rsi] ;
punpcklbw xmm1, xmm0 ;
movq xmm2, QWORD PTR [rdi] ;
punpcklbw xmm2, xmm0 ;
psubw xmm1, xmm2 ;
paddw xmm6, xmm1 ;
pmaddwd xmm1, xmm1 ;
paddd xmm7, xmm1 ;
lea rsi, [rsi + rax] ;ref_pixels_per_line
lea rdi, [rdi + rbx] ;src_pixels_per_line
sub rcx, 1 ;
jnz filter_block2d_bil_full_pixel_loop ;
jmp filter_block2d_bil_variance
Modify sub-pixel filters to eliminate unnecessary calculations In sub-pixel calculation, xoffset and yoffset mostly take some specific values. Modified sub-pixel filter functions according to these possible values to improve performance. Change-Id: I83083570af8b00ff65093467914fbb97a4e9ea21
2011-01-20 19:01:30 +01:00
filter_block2d_bil_var_sse2_fp_only:
mov rsi, arg(0) ;ref_ptr
mov rdi, arg(2) ;src_ptr
movsxd rcx, dword ptr arg(4) ;Height
movsxd rdx, dword ptr arg(1) ;ref_pixels_per_line
pxor xmm0, xmm0 ;
movsxd rbx, dword ptr arg(3) ;src_pixels_per_line
filter_block2d_bil_fp_only_loop:
movq xmm1, QWORD PTR [rsi] ;
movq xmm3, QWORD PTR [rsi+1] ;
punpcklbw xmm1, xmm0 ;
pmullw xmm1, [rax] ;
punpcklbw xmm3, xmm0 ;
pmullw xmm3, [rax+16] ;
paddw xmm1, xmm3 ;
paddw xmm1, xmm4 ;
psraw xmm1, xmm_filter_shift ;
movq xmm3, QWORD PTR [rdi] ;
punpcklbw xmm3, xmm0 ;
psubw xmm1, xmm3 ;
paddw xmm6, xmm1 ;
pmaddwd xmm1, xmm1 ;
paddd xmm7, xmm1 ;
lea rsi, [rsi + rdx]
lea rdi, [rdi + rbx] ;src_pixels_per_line
sub rcx, 1 ;
jnz filter_block2d_bil_fp_only_loop ;
jmp filter_block2d_bil_variance
Initial WebM release
2010-05-18 17:58:33 +02:00
Modify sub-pixel filters to eliminate unnecessary calculations In sub-pixel calculation, xoffset and yoffset mostly take some specific values. Modified sub-pixel filter functions according to these possible values to improve performance. Change-Id: I83083570af8b00ff65093467914fbb97a4e9ea21
2011-01-20 19:01:30 +01:00
filter_block2d_bil_variance:
Initial WebM release
2010-05-18 17:58:33 +02:00
movdq2q mm6, xmm6 ;
movdq2q mm7, xmm7 ;
psrldq xmm6, 8
psrldq xmm7, 8
movdq2q mm2, xmm6
movdq2q mm3, xmm7
paddw mm6, mm2
paddd mm7, mm3
pxor mm3, mm3 ;
pxor mm2, mm2 ;
punpcklwd mm2, mm6 ;
punpckhwd mm3, mm6 ;
paddd mm2, mm3 ;
movq mm6, mm2 ;
psrlq mm6, 32 ;
paddd mm2, mm6 ;
psrad mm2, 16 ;
movq mm4, mm7 ;
psrlq mm4, 32 ;
paddd mm4, mm7 ;
mov rsi, arg(7) ; sum
mov rdi, arg(8) ; sumsquared
movd [rsi], mm2 ; xsum
movd [rdi], mm4 ; xxsum
; begin epilog
Modify sub-pixel filters to eliminate unnecessary calculations In sub-pixel calculation, xoffset and yoffset mostly take some specific values. Modified sub-pixel filter functions according to these possible values to improve performance. Change-Id: I83083570af8b00ff65093467914fbb97a4e9ea21
2011-01-20 19:01:30 +01:00
pop rbx
Initial WebM release
2010-05-18 17:58:33 +02:00
pop rdi
pop rsi
RESTORE_GOT
Modify sub-pixel filters to eliminate unnecessary calculations In sub-pixel calculation, xoffset and yoffset mostly take some specific values. Modified sub-pixel filter functions according to these possible values to improve performance. Change-Id: I83083570af8b00ff65093467914fbb97a4e9ea21
2011-01-20 19:01:30 +01:00
RESTORE_XMM
Initial WebM release
2010-05-18 17:58:33 +02:00
UNSHADOW_ARGS
pop rbp
ret
Improve SSE2 half-pixel filter funtions Rewrote these functions to process 16 pixels once instead of 8. Change-Id: Ic67e80124467a446a3df4cfecfb76a4248602adb
2011-03-08 22:25:06 +01:00
;void vp8_half_horiz_vert_variance8x_h_sse2
Initial WebM release
2010-05-18 17:58:33 +02:00
;(
; unsigned char *ref_ptr,
; int ref_pixels_per_line,
; unsigned char *src_ptr,
; int src_pixels_per_line,
; unsigned int Height,
; int *sum,
; unsigned int *sumsquared
;)
Improve SSE2 half-pixel filter funtions Rewrote these functions to process 16 pixels once instead of 8. Change-Id: Ic67e80124467a446a3df4cfecfb76a4248602adb
2011-03-08 22:25:06 +01:00
global sym(vp8_half_horiz_vert_variance8x_h_sse2)
sym(vp8_half_horiz_vert_variance8x_h_sse2):
Initial WebM release
2010-05-18 17:58:33 +02:00
push rbp
mov rbp, rsp
SHADOW_ARGS_TO_STACK 7
Add save/restore xmm registers in x86 assembly code Went through the code and fixed it. Verified on Windows. Where possible, remove dependencies on xmm[67] Current code relies on pushing rbp to the stack to get 16 byte alignment. This broke when rbp wasn't pushed (vp8/encoder/x86/sad_sse3.asm). Work around this by using unaligned memory accesses. Revisit this and the offsets in vp8/encoder/x86/sad_sse3.asm in another change to SAVE_XMM. Change-Id: I5f940994d3ebfd977c3d68446cef20fd78b07877
2011-04-07 19:17:22 +02:00
SAVE_XMM
Initial WebM release
2010-05-18 17:58:33 +02:00
GET_GOT rbx
push rsi
push rdi
; end prolog
%if ABI_IS_32BIT=0
movsxd r8, dword ptr arg(1) ;ref_pixels_per_line
movsxd r9, dword ptr arg(3) ;src_pixels_per_line
%endif
pxor xmm6, xmm6 ; error accumulator
pxor xmm7, xmm7 ; sse eaccumulator
mov rsi, arg(0) ;ref_ptr ;
mov rdi, arg(2) ;src_ptr ;
movsxd rcx, dword ptr arg(4) ;Height ;
movsxd rax, dword ptr arg(1) ;ref_pixels_per_line
pxor xmm0, xmm0 ;
movq xmm5, QWORD PTR [rsi] ; xmm5 = s0,s1,s2..s8
movq xmm3, QWORD PTR [rsi+1] ; xmm3 = s1,s2,s3..s9
pavgb xmm5, xmm3 ; xmm5 = avg(xmm1,xmm3) horizontal line 1
%if ABI_IS_32BIT
add rsi, dword ptr arg(1) ;ref_pixels_per_line ; next source
%else
add rsi, r8
%endif
Improve SSE2 half-pixel filter funtions Rewrote these functions to process 16 pixels once instead of 8. Change-Id: Ic67e80124467a446a3df4cfecfb76a4248602adb
2011-03-08 22:25:06 +01:00
vp8_half_horiz_vert_variance8x_h_1:
Initial WebM release
2010-05-18 17:58:33 +02:00
movq xmm1, QWORD PTR [rsi] ;
movq xmm2, QWORD PTR [rsi+1] ;
pavgb xmm1, xmm2 ; xmm1 = avg(xmm1,xmm3) horizontal line i+1
pavgb xmm5, xmm1 ; xmm = vertical average of the above
punpcklbw xmm5, xmm0 ; xmm5 = words of above
movq xmm3, QWORD PTR [rdi] ; xmm3 = d0,d1,d2..d8
punpcklbw xmm3, xmm0 ; xmm3 = words of above
psubw xmm5, xmm3 ; xmm5 -= xmm3
paddw xmm6, xmm5 ; xmm6 += accumulated column differences
pmaddwd xmm5, xmm5 ; xmm5 *= xmm5
paddd xmm7, xmm5 ; xmm7 += accumulated square column differences
movdqa xmm5, xmm1 ; save xmm1 for use on the next row
%if ABI_IS_32BIT
add esi, dword ptr arg(1) ;ref_pixels_per_line ; next source
add edi, dword ptr arg(3) ;src_pixels_per_line ; next destination
%else
add rsi, r8
add rdi, r9
%endif
sub rcx, 1 ;
Improve SSE2 half-pixel filter funtions Rewrote these functions to process 16 pixels once instead of 8. Change-Id: Ic67e80124467a446a3df4cfecfb76a4248602adb
2011-03-08 22:25:06 +01:00
jnz vp8_half_horiz_vert_variance8x_h_1 ;
Initial WebM release
2010-05-18 17:58:33 +02:00
movdq2q mm6, xmm6 ;
movdq2q mm7, xmm7 ;
psrldq xmm6, 8
psrldq xmm7, 8
movdq2q mm2, xmm6
movdq2q mm3, xmm7
paddw mm6, mm2
paddd mm7, mm3
pxor mm3, mm3 ;
pxor mm2, mm2 ;
punpcklwd mm2, mm6 ;
punpckhwd mm3, mm6 ;
paddd mm2, mm3 ;
movq mm6, mm2 ;
psrlq mm6, 32 ;
paddd mm2, mm6 ;
psrad mm2, 16 ;
movq mm4, mm7 ;
psrlq mm4, 32 ;
paddd mm4, mm7 ;
mov rsi, arg(5) ; sum
mov rdi, arg(6) ; sumsquared
movd [rsi], mm2 ;
movd [rdi], mm4 ;
; begin epilog
pop rdi
pop rsi
RESTORE_GOT
Add save/restore xmm registers in x86 assembly code Went through the code and fixed it. Verified on Windows. Where possible, remove dependencies on xmm[67] Current code relies on pushing rbp to the stack to get 16 byte alignment. This broke when rbp wasn't pushed (vp8/encoder/x86/sad_sse3.asm). Work around this by using unaligned memory accesses. Revisit this and the offsets in vp8/encoder/x86/sad_sse3.asm in another change to SAVE_XMM. Change-Id: I5f940994d3ebfd977c3d68446cef20fd78b07877
2011-04-07 19:17:22 +02:00
RESTORE_XMM
Initial WebM release
2010-05-18 17:58:33 +02:00
UNSHADOW_ARGS
pop rbp
ret
Improve SSE2 half-pixel filter funtions Rewrote these functions to process 16 pixels once instead of 8. Change-Id: Ic67e80124467a446a3df4cfecfb76a4248602adb
2011-03-08 22:25:06 +01:00
;void vp8_half_horiz_vert_variance16x_h_sse2
;(
; unsigned char *ref_ptr,
; int ref_pixels_per_line,
; unsigned char *src_ptr,
; int src_pixels_per_line,
; unsigned int Height,
; int *sum,
; unsigned int *sumsquared
;)
global sym(vp8_half_horiz_vert_variance16x_h_sse2)
sym(vp8_half_horiz_vert_variance16x_h_sse2):
push rbp
mov rbp, rsp
SHADOW_ARGS_TO_STACK 7
SAVE_XMM
GET_GOT rbx
push rsi
push rdi
; end prolog
Initial WebM release
2010-05-18 17:58:33 +02:00
Improve SSE2 half-pixel filter funtions Rewrote these functions to process 16 pixels once instead of 8. Change-Id: Ic67e80124467a446a3df4cfecfb76a4248602adb
2011-03-08 22:25:06 +01:00
pxor xmm6, xmm6 ; error accumulator
pxor xmm7, xmm7 ; sse eaccumulator
mov rsi, arg(0) ;ref_ptr ;
mov rdi, arg(2) ;src_ptr ;
movsxd rcx, dword ptr arg(4) ;Height ;
movsxd rax, dword ptr arg(1) ;ref_pixels_per_line
movsxd rdx, dword ptr arg(3) ;src_pixels_per_line
pxor xmm0, xmm0 ;
movdqu xmm5, XMMWORD PTR [rsi]
movdqu xmm3, XMMWORD PTR [rsi+1]
pavgb xmm5, xmm3 ; xmm5 = avg(xmm1,xmm3) horizontal line 1
lea rsi, [rsi + rax]
vp8_half_horiz_vert_variance16x_h_1:
movdqu xmm1, XMMWORD PTR [rsi] ;
movdqu xmm2, XMMWORD PTR [rsi+1] ;
pavgb xmm1, xmm2 ; xmm1 = avg(xmm1,xmm3) horizontal line i+1
pavgb xmm5, xmm1 ; xmm = vertical average of the above
movdqa xmm4, xmm5
punpcklbw xmm5, xmm0 ; xmm5 = words of above
punpckhbw xmm4, xmm0
movq xmm3, QWORD PTR [rdi] ; xmm3 = d0,d1,d2..d7
punpcklbw xmm3, xmm0 ; xmm3 = words of above
psubw xmm5, xmm3 ; xmm5 -= xmm3
movq xmm3, QWORD PTR [rdi+8]
punpcklbw xmm3, xmm0
psubw xmm4, xmm3
paddw xmm6, xmm5 ; xmm6 += accumulated column differences
paddw xmm6, xmm4
pmaddwd xmm5, xmm5 ; xmm5 *= xmm5
pmaddwd xmm4, xmm4
paddd xmm7, xmm5 ; xmm7 += accumulated square column differences
paddd xmm7, xmm4
movdqa xmm5, xmm1 ; save xmm1 for use on the next row
lea rsi, [rsi + rax]
lea rdi, [rdi + rdx]
sub rcx, 1 ;
jnz vp8_half_horiz_vert_variance16x_h_1 ;
pxor xmm1, xmm1
pxor xmm5, xmm5
punpcklwd xmm0, xmm6
punpckhwd xmm1, xmm6
psrad xmm0, 16
psrad xmm1, 16
paddd xmm0, xmm1
movdqa xmm1, xmm0
movdqa xmm6, xmm7
punpckldq xmm6, xmm5
punpckhdq xmm7, xmm5
paddd xmm6, xmm7
punpckldq xmm0, xmm5
punpckhdq xmm1, xmm5
paddd xmm0, xmm1
movdqa xmm7, xmm6
movdqa xmm1, xmm0
psrldq xmm7, 8
psrldq xmm1, 8
paddd xmm6, xmm7
paddd xmm0, xmm1
mov rsi, arg(5) ;[Sum]
mov rdi, arg(6) ;[SSE]
movd [rsi], xmm0
movd [rdi], xmm6
; begin epilog
pop rdi
pop rsi
RESTORE_GOT
RESTORE_XMM
UNSHADOW_ARGS
pop rbp
ret
;void vp8_half_vert_variance8x_h_sse2
Initial WebM release
2010-05-18 17:58:33 +02:00
;(
; unsigned char *ref_ptr,
; int ref_pixels_per_line,
; unsigned char *src_ptr,
; int src_pixels_per_line,
; unsigned int Height,
; int *sum,
; unsigned int *sumsquared
;)
Improve SSE2 half-pixel filter funtions Rewrote these functions to process 16 pixels once instead of 8. Change-Id: Ic67e80124467a446a3df4cfecfb76a4248602adb
2011-03-08 22:25:06 +01:00
global sym(vp8_half_vert_variance8x_h_sse2)
sym(vp8_half_vert_variance8x_h_sse2):
Initial WebM release
2010-05-18 17:58:33 +02:00
push rbp
mov rbp, rsp
SHADOW_ARGS_TO_STACK 7
Add save/restore xmm registers in x86 assembly code Went through the code and fixed it. Verified on Windows. Where possible, remove dependencies on xmm[67] Current code relies on pushing rbp to the stack to get 16 byte alignment. This broke when rbp wasn't pushed (vp8/encoder/x86/sad_sse3.asm). Work around this by using unaligned memory accesses. Revisit this and the offsets in vp8/encoder/x86/sad_sse3.asm in another change to SAVE_XMM. Change-Id: I5f940994d3ebfd977c3d68446cef20fd78b07877
2011-04-07 19:17:22 +02:00
SAVE_XMM
Initial WebM release
2010-05-18 17:58:33 +02:00
GET_GOT rbx
push rsi
push rdi
; end prolog
%if ABI_IS_32BIT=0
movsxd r8, dword ptr arg(1) ;ref_pixels_per_line
movsxd r9, dword ptr arg(3) ;src_pixels_per_line
%endif
pxor xmm6, xmm6 ; error accumulator
pxor xmm7, xmm7 ; sse eaccumulator
mov rsi, arg(0) ;ref_ptr ;
mov rdi, arg(2) ;src_ptr ;
movsxd rcx, dword ptr arg(4) ;Height ;
movsxd rax, dword ptr arg(1) ;ref_pixels_per_line
pxor xmm0, xmm0 ;
Improve SSE2 half-pixel filter funtions Rewrote these functions to process 16 pixels once instead of 8. Change-Id: Ic67e80124467a446a3df4cfecfb76a4248602adb
2011-03-08 22:25:06 +01:00
vp8_half_vert_variance8x_h_1:
Initial WebM release
2010-05-18 17:58:33 +02:00
movq xmm5, QWORD PTR [rsi] ; xmm5 = s0,s1,s2..s8
movq xmm3, QWORD PTR [rsi+rax] ; xmm3 = s1,s2,s3..s9
pavgb xmm5, xmm3 ; xmm5 = avg(xmm1,xmm3)
punpcklbw xmm5, xmm0 ; xmm5 = words of above
movq xmm3, QWORD PTR [rdi] ; xmm3 = d0,d1,d2..d8
punpcklbw xmm3, xmm0 ; xmm3 = words of above
psubw xmm5, xmm3 ; xmm5 -= xmm3
paddw xmm6, xmm5 ; xmm6 += accumulated column differences
pmaddwd xmm5, xmm5 ; xmm5 *= xmm5
paddd xmm7, xmm5 ; xmm7 += accumulated square column differences
%if ABI_IS_32BIT
add esi, dword ptr arg(1) ;ref_pixels_per_line ; next source
add edi, dword ptr arg(3) ;src_pixels_per_line ; next destination
%else
add rsi, r8
add rdi, r9
%endif
sub rcx, 1 ;
Improve SSE2 half-pixel filter funtions Rewrote these functions to process 16 pixels once instead of 8. Change-Id: Ic67e80124467a446a3df4cfecfb76a4248602adb
2011-03-08 22:25:06 +01:00
jnz vp8_half_vert_variance8x_h_1 ;
Initial WebM release
2010-05-18 17:58:33 +02:00
movdq2q mm6, xmm6 ;
movdq2q mm7, xmm7 ;
psrldq xmm6, 8
psrldq xmm7, 8
movdq2q mm2, xmm6
movdq2q mm3, xmm7
paddw mm6, mm2
paddd mm7, mm3
pxor mm3, mm3 ;
pxor mm2, mm2 ;
punpcklwd mm2, mm6 ;
punpckhwd mm3, mm6 ;
paddd mm2, mm3 ;
movq mm6, mm2 ;
psrlq mm6, 32 ;
paddd mm2, mm6 ;
psrad mm2, 16 ;
movq mm4, mm7 ;
psrlq mm4, 32 ;
paddd mm4, mm7 ;
mov rsi, arg(5) ; sum
mov rdi, arg(6) ; sumsquared
movd [rsi], mm2 ;
movd [rdi], mm4 ;
; begin epilog
pop rdi
pop rsi
RESTORE_GOT
Add save/restore xmm registers in x86 assembly code Went through the code and fixed it. Verified on Windows. Where possible, remove dependencies on xmm[67] Current code relies on pushing rbp to the stack to get 16 byte alignment. This broke when rbp wasn't pushed (vp8/encoder/x86/sad_sse3.asm). Work around this by using unaligned memory accesses. Revisit this and the offsets in vp8/encoder/x86/sad_sse3.asm in another change to SAVE_XMM. Change-Id: I5f940994d3ebfd977c3d68446cef20fd78b07877
2011-04-07 19:17:22 +02:00
RESTORE_XMM
Initial WebM release
2010-05-18 17:58:33 +02:00
UNSHADOW_ARGS
pop rbp
ret
Improve SSE2 half-pixel filter funtions Rewrote these functions to process 16 pixels once instead of 8. Change-Id: Ic67e80124467a446a3df4cfecfb76a4248602adb
2011-03-08 22:25:06 +01:00
;void vp8_half_vert_variance16x_h_sse2
;(
; unsigned char *ref_ptr,
; int ref_pixels_per_line,
; unsigned char *src_ptr,
; int src_pixels_per_line,
; unsigned int Height,
; int *sum,
; unsigned int *sumsquared
;)
global sym(vp8_half_vert_variance16x_h_sse2)
sym(vp8_half_vert_variance16x_h_sse2):
push rbp
mov rbp, rsp
SHADOW_ARGS_TO_STACK 7
SAVE_XMM
GET_GOT rbx
push rsi
push rdi
; end prolog
Initial WebM release
2010-05-18 17:58:33 +02:00
Improve SSE2 half-pixel filter funtions Rewrote these functions to process 16 pixels once instead of 8. Change-Id: Ic67e80124467a446a3df4cfecfb76a4248602adb
2011-03-08 22:25:06 +01:00
pxor xmm6, xmm6 ; error accumulator
pxor xmm7, xmm7 ; sse eaccumulator
mov rsi, arg(0) ;ref_ptr
mov rdi, arg(2) ;src_ptr
movsxd rcx, dword ptr arg(4) ;Height
movsxd rax, dword ptr arg(1) ;ref_pixels_per_line
movsxd rdx, dword ptr arg(3) ;src_pixels_per_line
movdqu xmm5, XMMWORD PTR [rsi]
lea rsi, [rsi + rax ]
pxor xmm0, xmm0
vp8_half_vert_variance16x_h_1:
movdqu xmm3, XMMWORD PTR [rsi]
pavgb xmm5, xmm3 ; xmm5 = avg(xmm1,xmm3)
movdqa xmm4, xmm5
punpcklbw xmm5, xmm0
punpckhbw xmm4, xmm0
movq xmm2, QWORD PTR [rdi]
punpcklbw xmm2, xmm0
psubw xmm5, xmm2
movq xmm2, QWORD PTR [rdi+8]
punpcklbw xmm2, xmm0
psubw xmm4, xmm2
paddw xmm6, xmm5 ; xmm6 += accumulated column differences
paddw xmm6, xmm4
pmaddwd xmm5, xmm5 ; xmm5 *= xmm5
pmaddwd xmm4, xmm4
paddd xmm7, xmm5 ; xmm7 += accumulated square column differences
paddd xmm7, xmm4
movdqa xmm5, xmm3
lea rsi, [rsi + rax]
lea rdi, [rdi + rdx]
sub rcx, 1
jnz vp8_half_vert_variance16x_h_1
pxor xmm1, xmm1
pxor xmm5, xmm5
punpcklwd xmm0, xmm6
punpckhwd xmm1, xmm6
psrad xmm0, 16
psrad xmm1, 16
paddd xmm0, xmm1
movdqa xmm1, xmm0
movdqa xmm6, xmm7
punpckldq xmm6, xmm5
punpckhdq xmm7, xmm5
paddd xmm6, xmm7
punpckldq xmm0, xmm5
punpckhdq xmm1, xmm5
paddd xmm0, xmm1
movdqa xmm7, xmm6
movdqa xmm1, xmm0
psrldq xmm7, 8
psrldq xmm1, 8
paddd xmm6, xmm7
paddd xmm0, xmm1
mov rsi, arg(5) ;[Sum]
mov rdi, arg(6) ;[SSE]
movd [rsi], xmm0
movd [rdi], xmm6
; begin epilog
pop rdi
pop rsi
RESTORE_GOT
RESTORE_XMM
UNSHADOW_ARGS
pop rbp
ret
;void vp8_half_horiz_variance8x_h_sse2
Initial WebM release
2010-05-18 17:58:33 +02:00
;(
; unsigned char *ref_ptr,
; int ref_pixels_per_line,
; unsigned char *src_ptr,
; int src_pixels_per_line,
; unsigned int Height,
; int *sum,
; unsigned int *sumsquared
;)
Improve SSE2 half-pixel filter funtions Rewrote these functions to process 16 pixels once instead of 8. Change-Id: Ic67e80124467a446a3df4cfecfb76a4248602adb
2011-03-08 22:25:06 +01:00
global sym(vp8_half_horiz_variance8x_h_sse2)
sym(vp8_half_horiz_variance8x_h_sse2):
Initial WebM release
2010-05-18 17:58:33 +02:00
push rbp
mov rbp, rsp
SHADOW_ARGS_TO_STACK 7
Add save/restore xmm registers in x86 assembly code Went through the code and fixed it. Verified on Windows. Where possible, remove dependencies on xmm[67] Current code relies on pushing rbp to the stack to get 16 byte alignment. This broke when rbp wasn't pushed (vp8/encoder/x86/sad_sse3.asm). Work around this by using unaligned memory accesses. Revisit this and the offsets in vp8/encoder/x86/sad_sse3.asm in another change to SAVE_XMM. Change-Id: I5f940994d3ebfd977c3d68446cef20fd78b07877
2011-04-07 19:17:22 +02:00
SAVE_XMM
Initial WebM release
2010-05-18 17:58:33 +02:00
GET_GOT rbx
push rsi
push rdi
; end prolog
%if ABI_IS_32BIT=0
movsxd r8, dword ptr arg(1) ;ref_pixels_per_line
movsxd r9, dword ptr arg(3) ;src_pixels_per_line
%endif
pxor xmm6, xmm6 ; error accumulator
pxor xmm7, xmm7 ; sse eaccumulator
mov rsi, arg(0) ;ref_ptr ;
mov rdi, arg(2) ;src_ptr ;
movsxd rcx, dword ptr arg(4) ;Height ;
pxor xmm0, xmm0 ;
Improve SSE2 half-pixel filter funtions Rewrote these functions to process 16 pixels once instead of 8. Change-Id: Ic67e80124467a446a3df4cfecfb76a4248602adb
2011-03-08 22:25:06 +01:00
vp8_half_horiz_variance8x_h_1:
Initial WebM release
2010-05-18 17:58:33 +02:00
movq xmm5, QWORD PTR [rsi] ; xmm5 = s0,s1,s2..s8
movq xmm3, QWORD PTR [rsi+1] ; xmm3 = s1,s2,s3..s9
pavgb xmm5, xmm3 ; xmm5 = avg(xmm1,xmm3)
punpcklbw xmm5, xmm0 ; xmm5 = words of above
movq xmm3, QWORD PTR [rdi] ; xmm3 = d0,d1,d2..d8
punpcklbw xmm3, xmm0 ; xmm3 = words of above
psubw xmm5, xmm3 ; xmm5 -= xmm3
paddw xmm6, xmm5 ; xmm6 += accumulated column differences
pmaddwd xmm5, xmm5 ; xmm5 *= xmm5
paddd xmm7, xmm5 ; xmm7 += accumulated square column differences
%if ABI_IS_32BIT
add esi, dword ptr arg(1) ;ref_pixels_per_line ; next source
add edi, dword ptr arg(3) ;src_pixels_per_line ; next destination
%else
add rsi, r8
add rdi, r9
%endif
sub rcx, 1 ;
Improve SSE2 half-pixel filter funtions Rewrote these functions to process 16 pixels once instead of 8. Change-Id: Ic67e80124467a446a3df4cfecfb76a4248602adb
2011-03-08 22:25:06 +01:00
jnz vp8_half_horiz_variance8x_h_1 ;
Initial WebM release
2010-05-18 17:58:33 +02:00
movdq2q mm6, xmm6 ;
movdq2q mm7, xmm7 ;
psrldq xmm6, 8
psrldq xmm7, 8
movdq2q mm2, xmm6
movdq2q mm3, xmm7
paddw mm6, mm2
paddd mm7, mm3
pxor mm3, mm3 ;
pxor mm2, mm2 ;
punpcklwd mm2, mm6 ;
punpckhwd mm3, mm6 ;
paddd mm2, mm3 ;
movq mm6, mm2 ;
psrlq mm6, 32 ;
paddd mm2, mm6 ;
psrad mm2, 16 ;
movq mm4, mm7 ;
psrlq mm4, 32 ;
paddd mm4, mm7 ;
mov rsi, arg(5) ; sum
mov rdi, arg(6) ; sumsquared
movd [rsi], mm2 ;
movd [rdi], mm4 ;
; begin epilog
pop rdi
pop rsi
RESTORE_GOT
Add save/restore xmm registers in x86 assembly code Went through the code and fixed it. Verified on Windows. Where possible, remove dependencies on xmm[67] Current code relies on pushing rbp to the stack to get 16 byte alignment. This broke when rbp wasn't pushed (vp8/encoder/x86/sad_sse3.asm). Work around this by using unaligned memory accesses. Revisit this and the offsets in vp8/encoder/x86/sad_sse3.asm in another change to SAVE_XMM. Change-Id: I5f940994d3ebfd977c3d68446cef20fd78b07877
2011-04-07 19:17:22 +02:00
RESTORE_XMM
Initial WebM release
2010-05-18 17:58:33 +02:00
UNSHADOW_ARGS
pop rbp
ret
Improve SSE2 half-pixel filter funtions Rewrote these functions to process 16 pixels once instead of 8. Change-Id: Ic67e80124467a446a3df4cfecfb76a4248602adb
2011-03-08 22:25:06 +01:00
;void vp8_half_horiz_variance16x_h_sse2
;(
; unsigned char *ref_ptr,
; int ref_pixels_per_line,
; unsigned char *src_ptr,
; int src_pixels_per_line,
; unsigned int Height,
; int *sum,
; unsigned int *sumsquared
;)
global sym(vp8_half_horiz_variance16x_h_sse2)
sym(vp8_half_horiz_variance16x_h_sse2):
push rbp
mov rbp, rsp
SHADOW_ARGS_TO_STACK 7
SAVE_XMM
GET_GOT rbx
push rsi
push rdi
; end prolog
pxor xmm6, xmm6 ; error accumulator
pxor xmm7, xmm7 ; sse eaccumulator
mov rsi, arg(0) ;ref_ptr ;
mov rdi, arg(2) ;src_ptr ;
movsxd rcx, dword ptr arg(4) ;Height ;
movsxd rax, dword ptr arg(1) ;ref_pixels_per_line
movsxd rdx, dword ptr arg(3) ;src_pixels_per_line
pxor xmm0, xmm0 ;
vp8_half_horiz_variance16x_h_1:
movdqu xmm5, XMMWORD PTR [rsi] ; xmm5 = s0,s1,s2..s15
movdqu xmm3, XMMWORD PTR [rsi+1] ; xmm3 = s1,s2,s3..s16
pavgb xmm5, xmm3 ; xmm5 = avg(xmm1,xmm3)
movdqa xmm1, xmm5
punpcklbw xmm5, xmm0 ; xmm5 = words of above
punpckhbw xmm1, xmm0
movq xmm3, QWORD PTR [rdi] ; xmm3 = d0,d1,d2..d7
punpcklbw xmm3, xmm0 ; xmm3 = words of above
movq xmm2, QWORD PTR [rdi+8]
punpcklbw xmm2, xmm0
psubw xmm5, xmm3 ; xmm5 -= xmm3
psubw xmm1, xmm2
paddw xmm6, xmm5 ; xmm6 += accumulated column differences
paddw xmm6, xmm1
pmaddwd xmm5, xmm5 ; xmm5 *= xmm5
pmaddwd xmm1, xmm1
paddd xmm7, xmm5 ; xmm7 += accumulated square column differences
paddd xmm7, xmm1
lea rsi, [rsi + rax]
lea rdi, [rdi + rdx]
sub rcx, 1 ;
jnz vp8_half_horiz_variance16x_h_1 ;
pxor xmm1, xmm1
pxor xmm5, xmm5
punpcklwd xmm0, xmm6
punpckhwd xmm1, xmm6
psrad xmm0, 16
psrad xmm1, 16
paddd xmm0, xmm1
movdqa xmm1, xmm0
movdqa xmm6, xmm7
punpckldq xmm6, xmm5
punpckhdq xmm7, xmm5
paddd xmm6, xmm7
punpckldq xmm0, xmm5
punpckhdq xmm1, xmm5
paddd xmm0, xmm1
movdqa xmm7, xmm6
movdqa xmm1, xmm0
psrldq xmm7, 8
psrldq xmm1, 8
paddd xmm6, xmm7
paddd xmm0, xmm1
mov rsi, arg(5) ;[Sum]
mov rdi, arg(6) ;[SSE]
movd [rsi], xmm0
movd [rdi], xmm6
; begin epilog
pop rdi
pop rsi
RESTORE_GOT
RESTORE_XMM
UNSHADOW_ARGS
pop rbp
ret
Initial WebM release
2010-05-18 17:58:33 +02:00
SECTION_RODATA
; short xmm_bi_rd[8] = { 64, 64, 64, 64,64, 64, 64, 64};
align 16
xmm_bi_rd:
times 8 dw 64
Modify sub-pixel filters to eliminate unnecessary calculations In sub-pixel calculation, xoffset and yoffset mostly take some specific values. Modified sub-pixel filter functions according to these possible values to improve performance. Change-Id: I83083570af8b00ff65093467914fbb97a4e9ea21
2011-01-20 19:01:30 +01:00
align 16
vp8_bilinear_filters_sse2:
dw 128, 128, 128, 128, 128, 128, 128, 128, 0, 0, 0, 0, 0, 0, 0, 0
dw 112, 112, 112, 112, 112, 112, 112, 112, 16, 16, 16, 16, 16, 16, 16, 16
dw 96, 96, 96, 96, 96, 96, 96, 96, 32, 32, 32, 32, 32, 32, 32, 32
dw 80, 80, 80, 80, 80, 80, 80, 80, 48, 48, 48, 48, 48, 48, 48, 48
dw 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64
dw 48, 48, 48, 48, 48, 48, 48, 48, 80, 80, 80, 80, 80, 80, 80, 80
dw 32, 32, 32, 32, 32, 32, 32, 32, 96, 96, 96, 96, 96, 96, 96, 96
dw 16, 16, 16, 16, 16, 16, 16, 16, 112, 112, 112, 112, 112, 112, 112, 112
Reference in New Issue Copy Permalink