generic-library/vpx
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity
b72373de79
vpx/vp8/common/x86/postproc_sse2.asm

722 lines
21 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"
post-proc: deblock filter optimization 1. Algorithm modification: Instead of having same filter threshold for a whole frame, now we allow the thresholds to be adjusted for each macroblock. In current implementation, to avoid excessive blur on background as reported in issue480(http://code.google.com/p/webm/issues/detail?id=480), we reduce the thresholds for skipped macroblocks. 2. SSE2 optimization: As started in issue479(http://code.google.com/p/webm/issues/detail?id=479), the filter calculation was adjusted for better performance. The c code was also modified accordingly. This made the deblock filter 2x faster, and the decoder was 1.2x faster overall. Next, the demacroblock filter will be modified similarly. Change-Id: I05e54c3f580ccd427487d085096b3174f2ab7e86
2012-09-28 19:13:07 +02:00
;macro in deblock functions
%macro FIRST_2_ROWS 0
movdqa xmm4, xmm0
movdqa xmm6, xmm0
movdqa xmm5, xmm1
pavgb xmm5, xmm3
;calculate absolute value
psubusb xmm4, xmm1
psubusb xmm1, xmm0
psubusb xmm6, xmm3
psubusb xmm3, xmm0
paddusb xmm4, xmm1
paddusb xmm6, xmm3
;get threshold
movdqa xmm2, flimit
pxor xmm1, xmm1
movdqa xmm7, xmm2
;get mask
psubusb xmm2, xmm4
psubusb xmm7, xmm6
pcmpeqb xmm2, xmm1
pcmpeqb xmm7, xmm1
por xmm7, xmm2
%endmacro
%macro SECOND_2_ROWS 0
movdqa xmm6, xmm0
movdqa xmm4, xmm0
movdqa xmm2, xmm1
pavgb xmm1, xmm3
;calculate absolute value
psubusb xmm6, xmm2
psubusb xmm2, xmm0
psubusb xmm4, xmm3
psubusb xmm3, xmm0
paddusb xmm6, xmm2
paddusb xmm4, xmm3
pavgb xmm5, xmm1
;get threshold
movdqa xmm2, flimit
pxor xmm1, xmm1
movdqa xmm3, xmm2
;get mask
psubusb xmm2, xmm6
psubusb xmm3, xmm4
pcmpeqb xmm2, xmm1
pcmpeqb xmm3, xmm1
por xmm7, xmm2
por xmm7, xmm3
pavgb xmm5, xmm0
;decide if or not to use filtered value
pand xmm0, xmm7
pandn xmm7, xmm5
paddusb xmm0, xmm7
%endmacro
%macro UPDATE_FLIMIT 0
movdqa xmm2, XMMWORD PTR [rbx]
movdqa [rsp], xmm2
add rbx, 16
%endmacro
;void vp8_post_proc_down_and_across_mb_row_sse2
Initial WebM release
2010-05-18 17:58:33 +02:00
;(
; unsigned char *src_ptr,
; unsigned char *dst_ptr,
; int src_pixels_per_line,
; int dst_pixels_per_line,
; int cols,
post-proc: deblock filter optimization 1. Algorithm modification: Instead of having same filter threshold for a whole frame, now we allow the thresholds to be adjusted for each macroblock. In current implementation, to avoid excessive blur on background as reported in issue480(http://code.google.com/p/webm/issues/detail?id=480), we reduce the thresholds for skipped macroblocks. 2. SSE2 optimization: As started in issue479(http://code.google.com/p/webm/issues/detail?id=479), the filter calculation was adjusted for better performance. The c code was also modified accordingly. This made the deblock filter 2x faster, and the decoder was 1.2x faster overall. Next, the demacroblock filter will be modified similarly. Change-Id: I05e54c3f580ccd427487d085096b3174f2ab7e86
2012-09-28 19:13:07 +02:00
; int *flimits,
; int size
Initial WebM release
2010-05-18 17:58:33 +02:00
;)
post-proc: deblock filter optimization 1. Algorithm modification: Instead of having same filter threshold for a whole frame, now we allow the thresholds to be adjusted for each macroblock. In current implementation, to avoid excessive blur on background as reported in issue480(http://code.google.com/p/webm/issues/detail?id=480), we reduce the thresholds for skipped macroblocks. 2. SSE2 optimization: As started in issue479(http://code.google.com/p/webm/issues/detail?id=479), the filter calculation was adjusted for better performance. The c code was also modified accordingly. This made the deblock filter 2x faster, and the decoder was 1.2x faster overall. Next, the demacroblock filter will be modified similarly. Change-Id: I05e54c3f580ccd427487d085096b3174f2ab7e86
2012-09-28 19:13:07 +02:00
global sym(vp8_post_proc_down_and_across_mb_row_sse2) PRIVATE
sym(vp8_post_proc_down_and_across_mb_row_sse2):
Initial WebM release
2010-05-18 17:58:33 +02:00
push rbp
mov rbp, rsp
SHADOW_ARGS_TO_STACK 7
modify SAVE_XMM for potential 64bit use the win64 abi requires saving and restoring xmm6:xmm15. currently SAVE_XMM and RESTORE XMM only allow for saving xmm6:xmm7. allow specifying the highest register used and if the stack is unaligned. Change-Id: Ica5699622ffe3346d3a486f48eef0206c51cf867
2011-04-15 16:05:20 +02:00
SAVE_XMM 7
post-proc: deblock filter optimization 1. Algorithm modification: Instead of having same filter threshold for a whole frame, now we allow the thresholds to be adjusted for each macroblock. In current implementation, to avoid excessive blur on background as reported in issue480(http://code.google.com/p/webm/issues/detail?id=480), we reduce the thresholds for skipped macroblocks. 2. SSE2 optimization: As started in issue479(http://code.google.com/p/webm/issues/detail?id=479), the filter calculation was adjusted for better performance. The c code was also modified accordingly. This made the deblock filter 2x faster, and the decoder was 1.2x faster overall. Next, the demacroblock filter will be modified similarly. Change-Id: I05e54c3f580ccd427487d085096b3174f2ab7e86
2012-09-28 19:13:07 +02:00
push rbx
Initial WebM release
2010-05-18 17:58:33 +02:00
push rsi
push rdi
; end prolog
ALIGN_STACK 16, rax
sub rsp, 16
post-proc: deblock filter optimization 1. Algorithm modification: Instead of having same filter threshold for a whole frame, now we allow the thresholds to be adjusted for each macroblock. In current implementation, to avoid excessive blur on background as reported in issue480(http://code.google.com/p/webm/issues/detail?id=480), we reduce the thresholds for skipped macroblocks. 2. SSE2 optimization: As started in issue479(http://code.google.com/p/webm/issues/detail?id=479), the filter calculation was adjusted for better performance. The c code was also modified accordingly. This made the deblock filter 2x faster, and the decoder was 1.2x faster overall. Next, the demacroblock filter will be modified similarly. Change-Id: I05e54c3f580ccd427487d085096b3174f2ab7e86
2012-09-28 19:13:07 +02:00
; put flimit on stack
mov rbx, arg(5) ;flimits ptr
UPDATE_FLIMIT
Initial WebM release
2010-05-18 17:58:33 +02:00
post-proc: deblock filter optimization 1. Algorithm modification: Instead of having same filter threshold for a whole frame, now we allow the thresholds to be adjusted for each macroblock. In current implementation, to avoid excessive blur on background as reported in issue480(http://code.google.com/p/webm/issues/detail?id=480), we reduce the thresholds for skipped macroblocks. 2. SSE2 optimization: As started in issue479(http://code.google.com/p/webm/issues/detail?id=479), the filter calculation was adjusted for better performance. The c code was also modified accordingly. This made the deblock filter 2x faster, and the decoder was 1.2x faster overall. Next, the demacroblock filter will be modified similarly. Change-Id: I05e54c3f580ccd427487d085096b3174f2ab7e86
2012-09-28 19:13:07 +02:00
%define flimit [rsp]
Initial WebM release
2010-05-18 17:58:33 +02:00
post-proc: deblock filter optimization 1. Algorithm modification: Instead of having same filter threshold for a whole frame, now we allow the thresholds to be adjusted for each macroblock. In current implementation, to avoid excessive blur on background as reported in issue480(http://code.google.com/p/webm/issues/detail?id=480), we reduce the thresholds for skipped macroblocks. 2. SSE2 optimization: As started in issue479(http://code.google.com/p/webm/issues/detail?id=479), the filter calculation was adjusted for better performance. The c code was also modified accordingly. This made the deblock filter 2x faster, and the decoder was 1.2x faster overall. Next, the demacroblock filter will be modified similarly. Change-Id: I05e54c3f580ccd427487d085096b3174f2ab7e86
2012-09-28 19:13:07 +02:00
mov rsi, arg(0) ;src_ptr
mov rdi, arg(1) ;dst_ptr
Initial WebM release
2010-05-18 17:58:33 +02:00
post-proc: deblock filter optimization 1. Algorithm modification: Instead of having same filter threshold for a whole frame, now we allow the thresholds to be adjusted for each macroblock. In current implementation, to avoid excessive blur on background as reported in issue480(http://code.google.com/p/webm/issues/detail?id=480), we reduce the thresholds for skipped macroblocks. 2. SSE2 optimization: As started in issue479(http://code.google.com/p/webm/issues/detail?id=479), the filter calculation was adjusted for better performance. The c code was also modified accordingly. This made the deblock filter 2x faster, and the decoder was 1.2x faster overall. Next, the demacroblock filter will be modified similarly. Change-Id: I05e54c3f580ccd427487d085096b3174f2ab7e86
2012-09-28 19:13:07 +02:00
movsxd rax, DWORD PTR arg(2) ;src_pixels_per_line
movsxd rcx, DWORD PTR arg(6) ;rows in a macroblock
Use local labels for jumps/loops in x86 assembly. Prepend . to local labels in assembly code. This allows non unique labels within a file. Also makes profiling information more informative by keeping the function name with the loop name. Change-Id: I7a983cb3a5ba2413d5dafd0a37936b268fb9e37f
2011-08-23 00:29:41 +02:00
.nextrow:
post-proc: deblock filter optimization 1. Algorithm modification: Instead of having same filter threshold for a whole frame, now we allow the thresholds to be adjusted for each macroblock. In current implementation, to avoid excessive blur on background as reported in issue480(http://code.google.com/p/webm/issues/detail?id=480), we reduce the thresholds for skipped macroblocks. 2. SSE2 optimization: As started in issue479(http://code.google.com/p/webm/issues/detail?id=479), the filter calculation was adjusted for better performance. The c code was also modified accordingly. This made the deblock filter 2x faster, and the decoder was 1.2x faster overall. Next, the demacroblock filter will be modified similarly. Change-Id: I05e54c3f580ccd427487d085096b3174f2ab7e86
2012-09-28 19:13:07 +02:00
xor rdx, rdx ;col
Use local labels for jumps/loops in x86 assembly. Prepend . to local labels in assembly code. This allows non unique labels within a file. Also makes profiling information more informative by keeping the function name with the loop name. Change-Id: I7a983cb3a5ba2413d5dafd0a37936b268fb9e37f
2011-08-23 00:29:41 +02:00
.nextcol:
post-proc: deblock filter optimization 1. Algorithm modification: Instead of having same filter threshold for a whole frame, now we allow the thresholds to be adjusted for each macroblock. In current implementation, to avoid excessive blur on background as reported in issue480(http://code.google.com/p/webm/issues/detail?id=480), we reduce the thresholds for skipped macroblocks. 2. SSE2 optimization: As started in issue479(http://code.google.com/p/webm/issues/detail?id=479), the filter calculation was adjusted for better performance. The c code was also modified accordingly. This made the deblock filter 2x faster, and the decoder was 1.2x faster overall. Next, the demacroblock filter will be modified similarly. Change-Id: I05e54c3f580ccd427487d085096b3174f2ab7e86
2012-09-28 19:13:07 +02:00
;load current and next 2 rows
movdqu xmm0, XMMWORD PTR [rsi]
movdqu xmm1, XMMWORD PTR [rsi + rax]
movdqu xmm3, XMMWORD PTR [rsi + 2*rax]
Initial WebM release
2010-05-18 17:58:33 +02:00
post-proc: deblock filter optimization 1. Algorithm modification: Instead of having same filter threshold for a whole frame, now we allow the thresholds to be adjusted for each macroblock. In current implementation, to avoid excessive blur on background as reported in issue480(http://code.google.com/p/webm/issues/detail?id=480), we reduce the thresholds for skipped macroblocks. 2. SSE2 optimization: As started in issue479(http://code.google.com/p/webm/issues/detail?id=479), the filter calculation was adjusted for better performance. The c code was also modified accordingly. This made the deblock filter 2x faster, and the decoder was 1.2x faster overall. Next, the demacroblock filter will be modified similarly. Change-Id: I05e54c3f580ccd427487d085096b3174f2ab7e86
2012-09-28 19:13:07 +02:00
FIRST_2_ROWS
Initial WebM release
2010-05-18 17:58:33 +02:00
post-proc: deblock filter optimization 1. Algorithm modification: Instead of having same filter threshold for a whole frame, now we allow the thresholds to be adjusted for each macroblock. In current implementation, to avoid excessive blur on background as reported in issue480(http://code.google.com/p/webm/issues/detail?id=480), we reduce the thresholds for skipped macroblocks. 2. SSE2 optimization: As started in issue479(http://code.google.com/p/webm/issues/detail?id=479), the filter calculation was adjusted for better performance. The c code was also modified accordingly. This made the deblock filter 2x faster, and the decoder was 1.2x faster overall. Next, the demacroblock filter will be modified similarly. Change-Id: I05e54c3f580ccd427487d085096b3174f2ab7e86
2012-09-28 19:13:07 +02:00
;load above 2 rows
Initial WebM release
2010-05-18 17:58:33 +02:00
neg rax
post-proc: deblock filter optimization 1. Algorithm modification: Instead of having same filter threshold for a whole frame, now we allow the thresholds to be adjusted for each macroblock. In current implementation, to avoid excessive blur on background as reported in issue480(http://code.google.com/p/webm/issues/detail?id=480), we reduce the thresholds for skipped macroblocks. 2. SSE2 optimization: As started in issue479(http://code.google.com/p/webm/issues/detail?id=479), the filter calculation was adjusted for better performance. The c code was also modified accordingly. This made the deblock filter 2x faster, and the decoder was 1.2x faster overall. Next, the demacroblock filter will be modified similarly. Change-Id: I05e54c3f580ccd427487d085096b3174f2ab7e86
2012-09-28 19:13:07 +02:00
movdqu xmm1, XMMWORD PTR [rsi + 2*rax]
movdqu xmm3, XMMWORD PTR [rsi + rax]
Initial WebM release
2010-05-18 17:58:33 +02:00
post-proc: deblock filter optimization 1. Algorithm modification: Instead of having same filter threshold for a whole frame, now we allow the thresholds to be adjusted for each macroblock. In current implementation, to avoid excessive blur on background as reported in issue480(http://code.google.com/p/webm/issues/detail?id=480), we reduce the thresholds for skipped macroblocks. 2. SSE2 optimization: As started in issue479(http://code.google.com/p/webm/issues/detail?id=479), the filter calculation was adjusted for better performance. The c code was also modified accordingly. This made the deblock filter 2x faster, and the decoder was 1.2x faster overall. Next, the demacroblock filter will be modified similarly. Change-Id: I05e54c3f580ccd427487d085096b3174f2ab7e86
2012-09-28 19:13:07 +02:00
SECOND_2_ROWS
Initial WebM release
2010-05-18 17:58:33 +02:00
post-proc: deblock filter optimization 1. Algorithm modification: Instead of having same filter threshold for a whole frame, now we allow the thresholds to be adjusted for each macroblock. In current implementation, to avoid excessive blur on background as reported in issue480(http://code.google.com/p/webm/issues/detail?id=480), we reduce the thresholds for skipped macroblocks. 2. SSE2 optimization: As started in issue479(http://code.google.com/p/webm/issues/detail?id=479), the filter calculation was adjusted for better performance. The c code was also modified accordingly. This made the deblock filter 2x faster, and the decoder was 1.2x faster overall. Next, the demacroblock filter will be modified similarly. Change-Id: I05e54c3f580ccd427487d085096b3174f2ab7e86
2012-09-28 19:13:07 +02:00
movdqu XMMWORD PTR [rdi], xmm0
Initial WebM release
2010-05-18 17:58:33 +02:00
post-proc: deblock filter optimization 1. Algorithm modification: Instead of having same filter threshold for a whole frame, now we allow the thresholds to be adjusted for each macroblock. In current implementation, to avoid excessive blur on background as reported in issue480(http://code.google.com/p/webm/issues/detail?id=480), we reduce the thresholds for skipped macroblocks. 2. SSE2 optimization: As started in issue479(http://code.google.com/p/webm/issues/detail?id=479), the filter calculation was adjusted for better performance. The c code was also modified accordingly. This made the deblock filter 2x faster, and the decoder was 1.2x faster overall. Next, the demacroblock filter will be modified similarly. Change-Id: I05e54c3f580ccd427487d085096b3174f2ab7e86
2012-09-28 19:13:07 +02:00
neg rax ; positive stride
add rsi, 16
add rdi, 16
Initial WebM release
2010-05-18 17:58:33 +02:00
post-proc: deblock filter optimization 1. Algorithm modification: Instead of having same filter threshold for a whole frame, now we allow the thresholds to be adjusted for each macroblock. In current implementation, to avoid excessive blur on background as reported in issue480(http://code.google.com/p/webm/issues/detail?id=480), we reduce the thresholds for skipped macroblocks. 2. SSE2 optimization: As started in issue479(http://code.google.com/p/webm/issues/detail?id=479), the filter calculation was adjusted for better performance. The c code was also modified accordingly. This made the deblock filter 2x faster, and the decoder was 1.2x faster overall. Next, the demacroblock filter will be modified similarly. Change-Id: I05e54c3f580ccd427487d085096b3174f2ab7e86
2012-09-28 19:13:07 +02:00
add rdx, 16
cmp edx, dword arg(4) ;cols
postproc_sse2: avoid reading off the end of the limits array Rather than unconditionally reading in the next MB's limits, test the loop exit condition first. Change-Id: I105d1e92698fb5561aa87160816787604aed03a2
2012-10-23 19:59:05 +02:00
jge .downdone
UPDATE_FLIMIT
jmp .nextcol
Initial WebM release
2010-05-18 17:58:33 +02:00
postproc_sse2: avoid reading off the end of the limits array Rather than unconditionally reading in the next MB's limits, test the loop exit condition first. Change-Id: I105d1e92698fb5561aa87160816787604aed03a2
2012-10-23 19:59:05 +02:00
.downdone:
Initial WebM release
2010-05-18 17:58:33 +02:00
; done with the all cols, start the across filtering in place
sub rsi, rdx
sub rdi, rdx
post-proc: deblock filter optimization 1. Algorithm modification: Instead of having same filter threshold for a whole frame, now we allow the thresholds to be adjusted for each macroblock. In current implementation, to avoid excessive blur on background as reported in issue480(http://code.google.com/p/webm/issues/detail?id=480), we reduce the thresholds for skipped macroblocks. 2. SSE2 optimization: As started in issue479(http://code.google.com/p/webm/issues/detail?id=479), the filter calculation was adjusted for better performance. The c code was also modified accordingly. This made the deblock filter 2x faster, and the decoder was 1.2x faster overall. Next, the demacroblock filter will be modified similarly. Change-Id: I05e54c3f580ccd427487d085096b3174f2ab7e86
2012-09-28 19:13:07 +02:00
mov rbx, arg(5) ; flimits
UPDATE_FLIMIT
vp8d - valgrind warnings in mb post processor Solved by extending the border in the postproc buffer as necessary Change-Id: Ic3f61397fe5bc8e4db6fc78050b0b160bd0aee86
2012-01-18 02:27:39 +01:00
; dup the first byte into the left border 8 times
movq mm1, [rdi]
punpcklbw mm1, mm1
punpcklwd mm1, mm1
punpckldq mm1, mm1
mov rdx, -8
movq [rdi+rdx], mm1
; dup the last byte into the right border
post-proc: deblock filter optimization 1. Algorithm modification: Instead of having same filter threshold for a whole frame, now we allow the thresholds to be adjusted for each macroblock. In current implementation, to avoid excessive blur on background as reported in issue480(http://code.google.com/p/webm/issues/detail?id=480), we reduce the thresholds for skipped macroblocks. 2. SSE2 optimization: As started in issue479(http://code.google.com/p/webm/issues/detail?id=479), the filter calculation was adjusted for better performance. The c code was also modified accordingly. This made the deblock filter 2x faster, and the decoder was 1.2x faster overall. Next, the demacroblock filter will be modified similarly. Change-Id: I05e54c3f580ccd427487d085096b3174f2ab7e86
2012-09-28 19:13:07 +02:00
movsxd rdx, dword arg(4)
vp8d - valgrind warnings in mb post processor Solved by extending the border in the postproc buffer as necessary Change-Id: Ic3f61397fe5bc8e4db6fc78050b0b160bd0aee86
2012-01-18 02:27:39 +01:00
movq mm1, [rdi + rdx + -1]
punpcklbw mm1, mm1
punpcklwd mm1, mm1
punpckldq mm1, mm1
movq [rdi+rdx], mm1
Initial WebM release
2010-05-18 17:58:33 +02:00
xor rdx, rdx
post-proc: deblock filter optimization 1. Algorithm modification: Instead of having same filter threshold for a whole frame, now we allow the thresholds to be adjusted for each macroblock. In current implementation, to avoid excessive blur on background as reported in issue480(http://code.google.com/p/webm/issues/detail?id=480), we reduce the thresholds for skipped macroblocks. 2. SSE2 optimization: As started in issue479(http://code.google.com/p/webm/issues/detail?id=479), the filter calculation was adjusted for better performance. The c code was also modified accordingly. This made the deblock filter 2x faster, and the decoder was 1.2x faster overall. Next, the demacroblock filter will be modified similarly. Change-Id: I05e54c3f580ccd427487d085096b3174f2ab7e86
2012-09-28 19:13:07 +02:00
movq mm0, QWORD PTR [rdi-16];
movq mm1, QWORD PTR [rdi-8];
Initial WebM release
2010-05-18 17:58:33 +02:00
Use local labels for jumps/loops in x86 assembly. Prepend . to local labels in assembly code. This allows non unique labels within a file. Also makes profiling information more informative by keeping the function name with the loop name. Change-Id: I7a983cb3a5ba2413d5dafd0a37936b268fb9e37f
2011-08-23 00:29:41 +02:00
.acrossnextcol:
post-proc: deblock filter optimization 1. Algorithm modification: Instead of having same filter threshold for a whole frame, now we allow the thresholds to be adjusted for each macroblock. In current implementation, to avoid excessive blur on background as reported in issue480(http://code.google.com/p/webm/issues/detail?id=480), we reduce the thresholds for skipped macroblocks. 2. SSE2 optimization: As started in issue479(http://code.google.com/p/webm/issues/detail?id=479), the filter calculation was adjusted for better performance. The c code was also modified accordingly. This made the deblock filter 2x faster, and the decoder was 1.2x faster overall. Next, the demacroblock filter will be modified similarly. Change-Id: I05e54c3f580ccd427487d085096b3174f2ab7e86
2012-09-28 19:13:07 +02:00
movdqu xmm0, XMMWORD PTR [rdi + rdx]
movdqu xmm1, XMMWORD PTR [rdi + rdx -2]
movdqu xmm3, XMMWORD PTR [rdi + rdx -1]
FIRST_2_ROWS
movdqu xmm1, XMMWORD PTR [rdi + rdx +1]
movdqu xmm3, XMMWORD PTR [rdi + rdx +2]
SECOND_2_ROWS
movq QWORD PTR [rdi+rdx-16], mm0 ; store previous 8 bytes
movq QWORD PTR [rdi+rdx-8], mm1 ; store previous 8 bytes
movdq2q mm0, xmm0
psrldq xmm0, 8
movdq2q mm1, xmm0
add rdx, 16
cmp edx, dword arg(4) ;cols
postproc_sse2: avoid reading off the end of the limits array Rather than unconditionally reading in the next MB's limits, test the loop exit condition first. Change-Id: I105d1e92698fb5561aa87160816787604aed03a2
2012-10-23 19:59:05 +02:00
jge .acrossdone
UPDATE_FLIMIT
jmp .acrossnextcol
Initial WebM release
2010-05-18 17:58:33 +02:00
postproc_sse2: avoid reading off the end of the limits array Rather than unconditionally reading in the next MB's limits, test the loop exit condition first. Change-Id: I105d1e92698fb5561aa87160816787604aed03a2
2012-10-23 19:59:05 +02:00
.acrossdone
post-proc: deblock filter optimization 1. Algorithm modification: Instead of having same filter threshold for a whole frame, now we allow the thresholds to be adjusted for each macroblock. In current implementation, to avoid excessive blur on background as reported in issue480(http://code.google.com/p/webm/issues/detail?id=480), we reduce the thresholds for skipped macroblocks. 2. SSE2 optimization: As started in issue479(http://code.google.com/p/webm/issues/detail?id=479), the filter calculation was adjusted for better performance. The c code was also modified accordingly. This made the deblock filter 2x faster, and the decoder was 1.2x faster overall. Next, the demacroblock filter will be modified similarly. Change-Id: I05e54c3f580ccd427487d085096b3174f2ab7e86
2012-09-28 19:13:07 +02:00
; last 16 pixels
movq QWORD PTR [rdi+rdx-16], mm0
Initial WebM release
2010-05-18 17:58:33 +02:00
post-proc: deblock filter optimization 1. Algorithm modification: Instead of having same filter threshold for a whole frame, now we allow the thresholds to be adjusted for each macroblock. In current implementation, to avoid excessive blur on background as reported in issue480(http://code.google.com/p/webm/issues/detail?id=480), we reduce the thresholds for skipped macroblocks. 2. SSE2 optimization: As started in issue479(http://code.google.com/p/webm/issues/detail?id=479), the filter calculation was adjusted for better performance. The c code was also modified accordingly. This made the deblock filter 2x faster, and the decoder was 1.2x faster overall. Next, the demacroblock filter will be modified similarly. Change-Id: I05e54c3f580ccd427487d085096b3174f2ab7e86
2012-09-28 19:13:07 +02:00
cmp edx, dword arg(4)
jne .throw_last_8
movq QWORD PTR [rdi+rdx-8], mm1
.throw_last_8:
Initial WebM release
2010-05-18 17:58:33 +02:00
; done with this rwo
post-proc: deblock filter optimization 1. Algorithm modification: Instead of having same filter threshold for a whole frame, now we allow the thresholds to be adjusted for each macroblock. In current implementation, to avoid excessive blur on background as reported in issue480(http://code.google.com/p/webm/issues/detail?id=480), we reduce the thresholds for skipped macroblocks. 2. SSE2 optimization: As started in issue479(http://code.google.com/p/webm/issues/detail?id=479), the filter calculation was adjusted for better performance. The c code was also modified accordingly. This made the deblock filter 2x faster, and the decoder was 1.2x faster overall. Next, the demacroblock filter will be modified similarly. Change-Id: I05e54c3f580ccd427487d085096b3174f2ab7e86
2012-09-28 19:13:07 +02:00
add rsi,rax ;next src line
mov eax, dword arg(3) ;dst_pixels_per_line
add rdi,rax ;next destination
mov eax, dword arg(2) ;src_pixels_per_line
Initial WebM release
2010-05-18 17:58:33 +02:00
post-proc: deblock filter optimization 1. Algorithm modification: Instead of having same filter threshold for a whole frame, now we allow the thresholds to be adjusted for each macroblock. In current implementation, to avoid excessive blur on background as reported in issue480(http://code.google.com/p/webm/issues/detail?id=480), we reduce the thresholds for skipped macroblocks. 2. SSE2 optimization: As started in issue479(http://code.google.com/p/webm/issues/detail?id=479), the filter calculation was adjusted for better performance. The c code was also modified accordingly. This made the deblock filter 2x faster, and the decoder was 1.2x faster overall. Next, the demacroblock filter will be modified similarly. Change-Id: I05e54c3f580ccd427487d085096b3174f2ab7e86
2012-09-28 19:13:07 +02:00
mov rbx, arg(5) ;flimits
UPDATE_FLIMIT
Initial WebM release
2010-05-18 17:58:33 +02:00
post-proc: deblock filter optimization 1. Algorithm modification: Instead of having same filter threshold for a whole frame, now we allow the thresholds to be adjusted for each macroblock. In current implementation, to avoid excessive blur on background as reported in issue480(http://code.google.com/p/webm/issues/detail?id=480), we reduce the thresholds for skipped macroblocks. 2. SSE2 optimization: As started in issue479(http://code.google.com/p/webm/issues/detail?id=479), the filter calculation was adjusted for better performance. The c code was also modified accordingly. This made the deblock filter 2x faster, and the decoder was 1.2x faster overall. Next, the demacroblock filter will be modified similarly. Change-Id: I05e54c3f580ccd427487d085096b3174f2ab7e86
2012-09-28 19:13:07 +02:00
dec rcx ;decrement count
jnz .nextrow ;next row
add rsp, 16
Initial WebM release
2010-05-18 17:58:33 +02:00
pop rsp
; begin epilog
pop rdi
pop rsi
post-proc: deblock filter optimization 1. Algorithm modification: Instead of having same filter threshold for a whole frame, now we allow the thresholds to be adjusted for each macroblock. In current implementation, to avoid excessive blur on background as reported in issue480(http://code.google.com/p/webm/issues/detail?id=480), we reduce the thresholds for skipped macroblocks. 2. SSE2 optimization: As started in issue479(http://code.google.com/p/webm/issues/detail?id=479), the filter calculation was adjusted for better performance. The c code was also modified accordingly. This made the deblock filter 2x faster, and the decoder was 1.2x faster overall. Next, the demacroblock filter will be modified similarly. Change-Id: I05e54c3f580ccd427487d085096b3174f2ab7e86
2012-09-28 19:13:07 +02:00
pop rbx
some XMM registers are non-volatile on windows x64 ABI XMM6 to XMM15 are non-volatile on Windows x64 ABI. We have to save these registers. Change-Id: I4676309f1350af25c8a35f0c81b1f0499ab99076
2010-06-11 11:32:28 +02:00
RESTORE_XMM
Initial WebM release
2010-05-18 17:58:33 +02:00
UNSHADOW_ARGS
pop rbp
ret
post-proc: deblock filter optimization 1. Algorithm modification: Instead of having same filter threshold for a whole frame, now we allow the thresholds to be adjusted for each macroblock. In current implementation, to avoid excessive blur on background as reported in issue480(http://code.google.com/p/webm/issues/detail?id=480), we reduce the thresholds for skipped macroblocks. 2. SSE2 optimization: As started in issue479(http://code.google.com/p/webm/issues/detail?id=479), the filter calculation was adjusted for better performance. The c code was also modified accordingly. This made the deblock filter 2x faster, and the decoder was 1.2x faster overall. Next, the demacroblock filter will be modified similarly. Change-Id: I05e54c3f580ccd427487d085096b3174f2ab7e86
2012-09-28 19:13:07 +02:00
%undef flimit
Initial WebM release
2010-05-18 17:58:33 +02:00
;void vp8_mbpost_proc_down_xmm(unsigned char *dst,
; int pitch, int rows, int cols,int flimit)
extern sym(vp8_rv)
Make libvpx Chromium build friendly Add PRIVATE macro for adding private_extern directive for yasm to hide global symbols. This is only enabled if -DCHROMIUM is used with YASM. Also fixed a small problem with rtcd_defs.sh to guard TEMPORAL_DENOISING. Change-Id: I9027fce3ebddcf20078293e4b86b396f21da7857
2012-05-24 00:52:34 +02:00
global sym(vp8_mbpost_proc_down_xmm) PRIVATE
Initial WebM release
2010-05-18 17:58:33 +02:00
sym(vp8_mbpost_proc_down_xmm):
push rbp
mov rbp, rsp
SHADOW_ARGS_TO_STACK 5
modify SAVE_XMM for potential 64bit use the win64 abi requires saving and restoring xmm6:xmm15. currently SAVE_XMM and RESTORE XMM only allow for saving xmm6:xmm7. allow specifying the highest register used and if the stack is unaligned. Change-Id: Ica5699622ffe3346d3a486f48eef0206c51cf867
2011-04-15 16:05:20 +02:00
SAVE_XMM 7
Initial WebM release
2010-05-18 17:58:33 +02:00
GET_GOT rbx
push rsi
push rdi
; end prolog
ALIGN_STACK 16, rax
sub rsp, 128+16
; unsigned char d[16][8] at [rsp]
; create flimit2 at [rsp+128]
mov eax, dword ptr arg(4) ;flimit
mov [rsp+128], eax
mov [rsp+128+4], eax
mov [rsp+128+8], eax
mov [rsp+128+12], eax
%define flimit4 [rsp+128]
%if ABI_IS_32BIT=0
nasm: address labels 'rel label' vice 'wrt rip' nasm does not support `label wrt rip', it requires `rel label'. It is still fully compatible with yasm. 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: I488773a4e930a56e43b0cc72d867ee5291215f50
2010-10-04 23:18:58 +02:00
lea r8, [GLOBAL(sym(vp8_rv))]
Initial WebM release
2010-05-18 17:58:33 +02:00
%endif
;rows +=8;
add dword arg(2), 8
;for(c=0; c<cols; c+=8)
Use local labels for jumps/loops in x86 assembly. Prepend . to local labels in assembly code. This allows non unique labels within a file. Also makes profiling information more informative by keeping the function name with the loop name. Change-Id: I7a983cb3a5ba2413d5dafd0a37936b268fb9e37f
2011-08-23 00:29:41 +02:00
.loop_col:
Initial WebM release
2010-05-18 17:58:33 +02:00
mov rsi, arg(0) ; s
pxor xmm0, xmm0 ;
movsxd rax, dword ptr arg(1) ;pitch ;
vp8d - valgrind warnings in mb post processor Solved by extending the border in the postproc buffer as necessary Change-Id: Ic3f61397fe5bc8e4db6fc78050b0b160bd0aee86
2012-01-18 02:27:39 +01:00
; this copies the last row down into the border 8 rows
mov rdi, rsi
mov rdx, arg(2)
sub rdx, 9
imul rdx, rax
lea rdi, [rdi+rdx]
movq xmm1, QWORD ptr[rdi] ; first row
mov rcx, 8
.init_borderd ; initialize borders
lea rdi, [rdi + rax]
movq [rdi], xmm1
dec rcx
jne .init_borderd
Initial WebM release
2010-05-18 17:58:33 +02:00
neg rax ; rax = -pitch
vp8d - valgrind warnings in mb post processor Solved by extending the border in the postproc buffer as necessary Change-Id: Ic3f61397fe5bc8e4db6fc78050b0b160bd0aee86
2012-01-18 02:27:39 +01:00
; this copies the first row up into the border 8 rows
mov rdi, rsi
movq xmm1, QWORD ptr[rdi] ; first row
mov rcx, 8
.init_border ; initialize borders
lea rdi, [rdi + rax]
movq [rdi], xmm1
dec rcx
jne .init_border
Initial WebM release
2010-05-18 17:58:33 +02:00
lea rsi, [rsi + rax*8]; ; rdi = s[-pitch*8]
neg rax
pxor xmm5, xmm5
pxor xmm6, xmm6 ;
pxor xmm7, xmm7 ;
mov rdi, rsi
mov rcx, 15 ;
Use local labels for jumps/loops in x86 assembly. Prepend . to local labels in assembly code. This allows non unique labels within a file. Also makes profiling information more informative by keeping the function name with the loop name. Change-Id: I7a983cb3a5ba2413d5dafd0a37936b268fb9e37f
2011-08-23 00:29:41 +02:00
.loop_initvar:
Initial WebM release
2010-05-18 17:58:33 +02:00
movq xmm1, QWORD PTR [rdi];
punpcklbw xmm1, xmm0 ;
paddw xmm5, xmm1 ;
pmullw xmm1, xmm1 ;
movdqa xmm2, xmm1 ;
punpcklwd xmm1, xmm0 ;
punpckhwd xmm2, xmm0 ;
paddd xmm6, xmm1 ;
paddd xmm7, xmm2 ;
lea rdi, [rdi+rax] ;
dec rcx
Use local labels for jumps/loops in x86 assembly. Prepend . to local labels in assembly code. This allows non unique labels within a file. Also makes profiling information more informative by keeping the function name with the loop name. Change-Id: I7a983cb3a5ba2413d5dafd0a37936b268fb9e37f
2011-08-23 00:29:41 +02:00
jne .loop_initvar
Initial WebM release
2010-05-18 17:58:33 +02:00
;save the var and sum
xor rdx, rdx
Use local labels for jumps/loops in x86 assembly. Prepend . to local labels in assembly code. This allows non unique labels within a file. Also makes profiling information more informative by keeping the function name with the loop name. Change-Id: I7a983cb3a5ba2413d5dafd0a37936b268fb9e37f
2011-08-23 00:29:41 +02:00
.loop_row:
Initial WebM release
2010-05-18 17:58:33 +02:00
movq xmm1, QWORD PTR [rsi] ; [s-pitch*8]
movq xmm2, QWORD PTR [rdi] ; [s+pitch*7]
punpcklbw xmm1, xmm0
punpcklbw xmm2, xmm0
paddw xmm5, xmm2
psubw xmm5, xmm1
pmullw xmm2, xmm2
movdqa xmm4, xmm2
punpcklwd xmm2, xmm0
punpckhwd xmm4, xmm0
paddd xmm6, xmm2
paddd xmm7, xmm4
pmullw xmm1, xmm1
movdqa xmm2, xmm1
punpcklwd xmm1, xmm0
psubd xmm6, xmm1
punpckhwd xmm2, xmm0
psubd xmm7, xmm2
movdqa xmm3, xmm6
pslld xmm3, 4
psubd xmm3, xmm6
movdqa xmm1, xmm5
movdqa xmm4, xmm5
pmullw xmm1, xmm1
pmulhw xmm4, xmm4
movdqa xmm2, xmm1
punpcklwd xmm1, xmm4
punpckhwd xmm2, xmm4
movdqa xmm4, xmm7
pslld xmm4, 4
psubd xmm4, xmm7
psubd xmm3, xmm1
psubd xmm4, xmm2
psubd xmm3, flimit4
psubd xmm4, flimit4
psrad xmm3, 31
psrad xmm4, 31
packssdw xmm3, xmm4
packsswb xmm3, xmm0
movq xmm1, QWORD PTR [rsi+rax*8]
movq xmm2, xmm1
punpcklbw xmm1, xmm0
paddw xmm1, xmm5
mov rcx, rdx
and rcx, 127
%if ABI_IS_32BIT=1 && CONFIG_PIC=1
push rax
nasm: address labels 'rel label' vice 'wrt rip' nasm does not support `label wrt rip', it requires `rel label'. It is still fully compatible with yasm. 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: I488773a4e930a56e43b0cc72d867ee5291215f50
2010-10-04 23:18:58 +02:00
lea rax, [GLOBAL(sym(vp8_rv))]
Initial WebM release
2010-05-18 17:58:33 +02:00
movdqu xmm4, [rax + rcx*2] ;vp8_rv[rcx*2]
pop rax
%elif ABI_IS_32BIT=0
movdqu xmm4, [r8 + rcx*2] ;vp8_rv[rcx*2]
%else
movdqu xmm4, [sym(vp8_rv) + rcx*2]
%endif
paddw xmm1, xmm4
;paddw xmm1, eight8s
psraw xmm1, 4
packuswb xmm1, xmm0
pand xmm1, xmm3
pandn xmm3, xmm2
por xmm1, xmm3
and rcx, 15
movq QWORD PTR [rsp + rcx*8], xmm1 ;d[rcx*8]
mov rcx, rdx
sub rcx, 8
and rcx, 15
movq mm0, [rsp + rcx*8] ;d[rcx*8]
movq [rsi], mm0
lea rsi, [rsi+rax]
lea rdi, [rdi+rax]
add rdx, 1
cmp edx, dword arg(2) ;rows
Use local labels for jumps/loops in x86 assembly. Prepend . to local labels in assembly code. This allows non unique labels within a file. Also makes profiling information more informative by keeping the function name with the loop name. Change-Id: I7a983cb3a5ba2413d5dafd0a37936b268fb9e37f
2011-08-23 00:29:41 +02:00
jl .loop_row
Initial WebM release
2010-05-18 17:58:33 +02:00
add dword arg(0), 8 ; s += 8
sub dword arg(3), 8 ; cols -= 8
cmp dword arg(3), 0
Use local labels for jumps/loops in x86 assembly. Prepend . to local labels in assembly code. This allows non unique labels within a file. Also makes profiling information more informative by keeping the function name with the loop name. Change-Id: I7a983cb3a5ba2413d5dafd0a37936b268fb9e37f
2011-08-23 00:29:41 +02:00
jg .loop_col
Initial WebM release
2010-05-18 17:58:33 +02:00
add rsp, 128+16
pop rsp
; begin epilog
pop rdi
pop rsi
RESTORE_GOT
some XMM registers are non-volatile on windows x64 ABI XMM6 to XMM15 are non-volatile on Windows x64 ABI. We have to save these registers. Change-Id: I4676309f1350af25c8a35f0c81b1f0499ab99076
2010-06-11 11:32:28 +02:00
RESTORE_XMM
Initial WebM release
2010-05-18 17:58:33 +02:00
UNSHADOW_ARGS
pop rbp
ret
%undef flimit4
;void vp8_mbpost_proc_across_ip_xmm(unsigned char *src,
; int pitch, int rows, int cols,int flimit)
Make libvpx Chromium build friendly Add PRIVATE macro for adding private_extern directive for yasm to hide global symbols. This is only enabled if -DCHROMIUM is used with YASM. Also fixed a small problem with rtcd_defs.sh to guard TEMPORAL_DENOISING. Change-Id: I9027fce3ebddcf20078293e4b86b396f21da7857
2012-05-24 00:52:34 +02:00
global sym(vp8_mbpost_proc_across_ip_xmm) PRIVATE
Initial WebM release
2010-05-18 17:58:33 +02:00
sym(vp8_mbpost_proc_across_ip_xmm):
push rbp
mov rbp, rsp
SHADOW_ARGS_TO_STACK 5
modify SAVE_XMM for potential 64bit use the win64 abi requires saving and restoring xmm6:xmm15. currently SAVE_XMM and RESTORE XMM only allow for saving xmm6:xmm7. allow specifying the highest register used and if the stack is unaligned. Change-Id: Ica5699622ffe3346d3a486f48eef0206c51cf867
2011-04-15 16:05:20 +02:00
SAVE_XMM 7
Initial WebM release
2010-05-18 17:58:33 +02:00
GET_GOT rbx
push rsi
push rdi
; end prolog
ALIGN_STACK 16, rax
sub rsp, 16
; create flimit4 at [rsp]
mov eax, dword ptr arg(4) ;flimit
mov [rsp], eax
mov [rsp+4], eax
mov [rsp+8], eax
mov [rsp+12], eax
%define flimit4 [rsp]
;for(r=0;r<rows;r++)
Use local labels for jumps/loops in x86 assembly. Prepend . to local labels in assembly code. This allows non unique labels within a file. Also makes profiling information more informative by keeping the function name with the loop name. Change-Id: I7a983cb3a5ba2413d5dafd0a37936b268fb9e37f
2011-08-23 00:29:41 +02:00
.ip_row_loop:
Initial WebM release
2010-05-18 17:58:33 +02:00
xor rdx, rdx ;sumsq=0;
xor rcx, rcx ;sum=0;
mov rsi, arg(0); s
vp8d - valgrind warnings in mb post processor Solved by extending the border in the postproc buffer as necessary Change-Id: Ic3f61397fe5bc8e4db6fc78050b0b160bd0aee86
2012-01-18 02:27:39 +01:00
; dup the first byte into the left border 8 times
movq mm1, [rsi]
punpcklbw mm1, mm1
punpcklwd mm1, mm1
punpckldq mm1, mm1
Initial WebM release
2010-05-18 17:58:33 +02:00
mov rdi, -8
vp8d - valgrind warnings in mb post processor Solved by extending the border in the postproc buffer as necessary Change-Id: Ic3f61397fe5bc8e4db6fc78050b0b160bd0aee86
2012-01-18 02:27:39 +01:00
movq [rsi+rdi], mm1
; dup the last byte into the right border
movsxd rdx, dword arg(3)
movq mm1, [rsi + rdx + -1]
punpcklbw mm1, mm1
punpcklwd mm1, mm1
punpckldq mm1, mm1
movq [rsi+rdx], mm1
Use local labels for jumps/loops in x86 assembly. Prepend . to local labels in assembly code. This allows non unique labels within a file. Also makes profiling information more informative by keeping the function name with the loop name. Change-Id: I7a983cb3a5ba2413d5dafd0a37936b268fb9e37f
2011-08-23 00:29:41 +02:00
.ip_var_loop:
Initial WebM release
2010-05-18 17:58:33 +02:00
;for(i=-8;i<=6;i++)
;{
; sumsq += s[i]*s[i];
; sum += s[i];
;}
movzx eax, byte [rsi+rdi]
add ecx, eax
mul al
add edx, eax
add rdi, 1
cmp rdi, 6
Use local labels for jumps/loops in x86 assembly. Prepend . to local labels in assembly code. This allows non unique labels within a file. Also makes profiling information more informative by keeping the function name with the loop name. Change-Id: I7a983cb3a5ba2413d5dafd0a37936b268fb9e37f
2011-08-23 00:29:41 +02:00
jle .ip_var_loop
Initial WebM release
2010-05-18 17:58:33 +02:00
;mov rax, sumsq
;movd xmm7, rax
movd xmm7, edx
;mov rax, sum
;movd xmm6, rax
movd xmm6, ecx
mov rsi, arg(0) ;s
xor rcx, rcx
movsxd rdx, dword arg(3) ;cols
add rdx, 8
pxor mm0, mm0
pxor mm1, mm1
pxor xmm0, xmm0
Use local labels for jumps/loops in x86 assembly. Prepend . to local labels in assembly code. This allows non unique labels within a file. Also makes profiling information more informative by keeping the function name with the loop name. Change-Id: I7a983cb3a5ba2413d5dafd0a37936b268fb9e37f
2011-08-23 00:29:41 +02:00
.nextcol4:
Initial WebM release
2010-05-18 17:58:33 +02:00
movd xmm1, DWORD PTR [rsi+rcx-8] ; -8 -7 -6 -5
movd xmm2, DWORD PTR [rsi+rcx+7] ; +7 +8 +9 +10
punpcklbw xmm1, xmm0 ; expanding
punpcklbw xmm2, xmm0 ; expanding
punpcklwd xmm1, xmm0 ; expanding to dwords
punpcklwd xmm2, xmm0 ; expanding to dwords
psubd xmm2, xmm1 ; 7--8 8--7 9--6 10--5
paddd xmm1, xmm1 ; -8*2 -7*2 -6*2 -5*2
paddd xmm1, xmm2 ; 7+-8 8+-7 9+-6 10+-5
pmaddwd xmm1, xmm2 ; squared of 7+-8 8+-7 9+-6 10+-5
paddd xmm6, xmm2
paddd xmm7, xmm1
pshufd xmm6, xmm6, 0 ; duplicate the last ones
pshufd xmm7, xmm7, 0 ; duplicate the last ones
psrldq xmm1, 4 ; 8--7 9--6 10--5 0000
psrldq xmm2, 4 ; 8--7 9--6 10--5 0000
pshufd xmm3, xmm1, 3 ; 0000 8--7 8--7 8--7 squared
pshufd xmm4, xmm2, 3 ; 0000 8--7 8--7 8--7 squared
paddd xmm6, xmm4
paddd xmm7, xmm3
pshufd xmm3, xmm1, 01011111b ; 0000 0000 9--6 9--6 squared
pshufd xmm4, xmm2, 01011111b ; 0000 0000 9--6 9--6 squared
paddd xmm7, xmm3
paddd xmm6, xmm4
pshufd xmm3, xmm1, 10111111b ; 0000 0000 8--7 8--7 squared
pshufd xmm4, xmm2, 10111111b ; 0000 0000 8--7 8--7 squared
paddd xmm7, xmm3
paddd xmm6, xmm4
movdqa xmm3, xmm6
pmaddwd xmm3, xmm3
movdqa xmm5, xmm7
pslld xmm5, 4
psubd xmm5, xmm7
psubd xmm5, xmm3
psubd xmm5, flimit4
psrad xmm5, 31
packssdw xmm5, xmm0
packsswb xmm5, xmm0
movd xmm1, DWORD PTR [rsi+rcx]
movq xmm2, xmm1
punpcklbw xmm1, xmm0
punpcklwd xmm1, xmm0
paddd xmm1, xmm6
nasm: address labels 'rel label' vice 'wrt rip' nasm does not support `label wrt rip', it requires `rel label'. It is still fully compatible with yasm. 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: I488773a4e930a56e43b0cc72d867ee5291215f50
2010-10-04 23:18:58 +02:00
paddd xmm1, [GLOBAL(four8s)]
Initial WebM release
2010-05-18 17:58:33 +02:00
psrad xmm1, 4
packssdw xmm1, xmm0
packuswb xmm1, xmm0
pand xmm1, xmm5
pandn xmm5, xmm2
por xmm5, xmm1
movd [rsi+rcx-8], mm0
movq mm0, mm1
movdq2q mm1, xmm5
psrldq xmm7, 12
psrldq xmm6, 12
add rcx, 4
cmp rcx, rdx
Use local labels for jumps/loops in x86 assembly. Prepend . to local labels in assembly code. This allows non unique labels within a file. Also makes profiling information more informative by keeping the function name with the loop name. Change-Id: I7a983cb3a5ba2413d5dafd0a37936b268fb9e37f
2011-08-23 00:29:41 +02:00
jl .nextcol4
Initial WebM release
2010-05-18 17:58:33 +02:00
;s+=pitch;
movsxd rax, dword arg(1)
add arg(0), rax
sub dword arg(2), 1 ;rows-=1
cmp dword arg(2), 0
Use local labels for jumps/loops in x86 assembly. Prepend . to local labels in assembly code. This allows non unique labels within a file. Also makes profiling information more informative by keeping the function name with the loop name. Change-Id: I7a983cb3a5ba2413d5dafd0a37936b268fb9e37f
2011-08-23 00:29:41 +02:00
jg .ip_row_loop
Initial WebM release
2010-05-18 17:58:33 +02:00
add rsp, 16
pop rsp
; begin epilog
pop rdi
pop rsi
RESTORE_GOT
some XMM registers are non-volatile on windows x64 ABI XMM6 to XMM15 are non-volatile on Windows x64 ABI. We have to save these registers. Change-Id: I4676309f1350af25c8a35f0c81b1f0499ab99076
2010-06-11 11:32:28 +02:00
RESTORE_XMM
Initial WebM release
2010-05-18 17:58:33 +02:00
UNSHADOW_ARGS
pop rbp
ret
%undef flimit4
;void vp8_plane_add_noise_wmt (unsigned char *Start, unsigned char *noise,
; unsigned char blackclamp[16],
; unsigned char whiteclamp[16],
; unsigned char bothclamp[16],
; unsigned int Width, unsigned int Height, int Pitch)
extern sym(rand)
Make libvpx Chromium build friendly Add PRIVATE macro for adding private_extern directive for yasm to hide global symbols. This is only enabled if -DCHROMIUM is used with YASM. Also fixed a small problem with rtcd_defs.sh to guard TEMPORAL_DENOISING. Change-Id: I9027fce3ebddcf20078293e4b86b396f21da7857
2012-05-24 00:52:34 +02:00
global sym(vp8_plane_add_noise_wmt) PRIVATE
Initial WebM release
2010-05-18 17:58:33 +02:00
sym(vp8_plane_add_noise_wmt):
push rbp
mov rbp, rsp
SHADOW_ARGS_TO_STACK 8
GET_GOT rbx
push rsi
push rdi
; end prolog
Use local labels for jumps/loops in x86 assembly. Prepend . to local labels in assembly code. This allows non unique labels within a file. Also makes profiling information more informative by keeping the function name with the loop name. Change-Id: I7a983cb3a5ba2413d5dafd0a37936b268fb9e37f
2011-08-23 00:29:41 +02:00
.addnoise_loop:
Initial WebM release
2010-05-18 17:58:33 +02:00
call sym(rand) WRT_PLT
mov rcx, arg(1) ;noise
and rax, 0xff
add rcx, rax
; we rely on the fact that the clamping vectors are stored contiguously
; in black/white/both order. Note that we have to reload this here because
; rdx could be trashed by rand()
mov rdx, arg(2) ; blackclamp
mov rdi, rcx
movsxd rcx, dword arg(5) ;[Width]
mov rsi, arg(0) ;Pos
xor rax,rax
Use local labels for jumps/loops in x86 assembly. Prepend . to local labels in assembly code. This allows non unique labels within a file. Also makes profiling information more informative by keeping the function name with the loop name. Change-Id: I7a983cb3a5ba2413d5dafd0a37936b268fb9e37f
2011-08-23 00:29:41 +02:00
.addnoise_nextset:
Initial WebM release
2010-05-18 17:58:33 +02:00
movdqu xmm1,[rsi+rax] ; get the source
psubusb xmm1, [rdx] ;blackclamp ; clamp both sides so we don't outrange adding noise
paddusb xmm1, [rdx+32] ;bothclamp
psubusb xmm1, [rdx+16] ;whiteclamp
movdqu xmm2,[rdi+rax] ; get the noise for this line
paddb xmm1,xmm2 ; add it in
movdqu [rsi+rax],xmm1 ; store the result
add rax,16 ; move to the next line
cmp rax, rcx
Use local labels for jumps/loops in x86 assembly. Prepend . to local labels in assembly code. This allows non unique labels within a file. Also makes profiling information more informative by keeping the function name with the loop name. Change-Id: I7a983cb3a5ba2413d5dafd0a37936b268fb9e37f
2011-08-23 00:29:41 +02:00
jl .addnoise_nextset
Initial WebM release
2010-05-18 17:58:33 +02:00
movsxd rax, dword arg(7) ; Pitch
add arg(0), rax ; Start += Pitch
sub dword arg(6), 1 ; Height -= 1
Use local labels for jumps/loops in x86 assembly. Prepend . to local labels in assembly code. This allows non unique labels within a file. Also makes profiling information more informative by keeping the function name with the loop name. Change-Id: I7a983cb3a5ba2413d5dafd0a37936b268fb9e37f
2011-08-23 00:29:41 +02:00
jg .addnoise_loop
Initial WebM release
2010-05-18 17:58:33 +02:00
; begin epilog
pop rdi
pop rsi
RESTORE_GOT
UNSHADOW_ARGS
pop rbp
ret
SECTION_RODATA
align 16
four8s:
times 4 dd 8
Reference in New Issue Copy Permalink