generic-library/vpx
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity
0fe79aab19
vpx/vp8/common/x86/subpixel_sse2.asm

1373 lines
46 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"
Move shared data to shared location Storing vp8_bilinear_filters_mmx in an mmx file and using it in an sse2 file is bad Moving towards allowing --disable-mmx Change-Id: I20493b35bdedcdcfc0915e6f05fdbe6c81a4a742
2011-11-18 20:50:13 +01:00
extern sym(vp8_bilinear_filters_x86_8)
Initial WebM release
2010-05-18 17:58:33 +02:00
%define BLOCK_HEIGHT_WIDTH 4
%define VP8_FILTER_WEIGHT 128
%define VP8_FILTER_SHIFT 7
;/************************************************************************************
; Notes: filter_block1d_h6 applies a 6 tap filter horizontally to the input pixels. The
; input pixel array has output_height rows. This routine assumes that output_height is an
; even number. This function handles 8 pixels in horizontal direction, calculating ONE
; rows each iteration to take advantage of the 128 bits operations.
;*************************************************************************************/
;void vp8_filter_block1d8_h6_sse2
;(
; unsigned char *src_ptr,
; unsigned short *output_ptr,
; unsigned int src_pixels_per_line,
; unsigned int pixel_step,
; unsigned int output_height,
; unsigned int output_width,
; short *vp8_filter
;)
global sym(vp8_filter_block1d8_h6_sse2)
sym(vp8_filter_block1d8_h6_sse2):
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
Initial WebM release
2010-05-18 17:58:33 +02:00
GET_GOT rbx
push rsi
push rdi
; end prolog
mov rdx, arg(6) ;vp8_filter
mov rsi, arg(0) ;src_ptr
mov rdi, arg(1) ;output_ptr
movsxd rcx, dword ptr arg(4) ;output_height
movsxd rax, dword ptr arg(2) ;src_pixels_per_line ; Pitch for Source
%if ABI_IS_32BIT=0
movsxd r8, dword ptr arg(5) ;output_width
%endif
pxor xmm0, xmm0 ; clear xmm0 for unpack
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
.filter_block1d8_h6_rowloop:
Initial WebM release
2010-05-18 17:58:33 +02:00
movq xmm3, MMWORD PTR [rsi - 2]
movq xmm1, MMWORD PTR [rsi + 6]
prefetcht2 [rsi+rax-2]
pslldq xmm1, 8
por xmm1, xmm3
movdqa xmm4, xmm1
movdqa xmm5, xmm1
movdqa xmm6, xmm1
movdqa xmm7, xmm1
punpcklbw xmm3, xmm0 ; xx05 xx04 xx03 xx02 xx01 xx01 xx-1 xx-2
psrldq xmm4, 1 ; xx 0d 0c 0b 0a 09 08 07 06 05 04 03 02 01 00 -1
pmullw xmm3, XMMWORD PTR [rdx] ; x[-2] * H[-2]; Tap 1
punpcklbw xmm4, xmm0 ; xx06 xx05 xx04 xx03 xx02 xx01 xx00 xx-1
psrldq xmm5, 2 ; xx xx 0d 0c 0b 0a 09 08 07 06 05 04 03 02 01 00
pmullw xmm4, XMMWORD PTR [rdx+16] ; x[-1] * H[-1]; Tap 2
punpcklbw xmm5, xmm0 ; xx07 xx06 xx05 xx04 xx03 xx02 xx01 xx00
psrldq xmm6, 3 ; xx xx xx 0d 0c 0b 0a 09 08 07 06 05 04 03 02 01
pmullw xmm5, [rdx+32] ; x[ 0] * H[ 0]; Tap 3
punpcklbw xmm6, xmm0 ; xx08 xx07 xx06 xx05 xx04 xx03 xx02 xx01
psrldq xmm7, 4 ; xx xx xx xx 0d 0c 0b 0a 09 08 07 06 05 04 03 02
pmullw xmm6, [rdx+48] ; x[ 1] * h[ 1] ; Tap 4
punpcklbw xmm7, xmm0 ; xx09 xx08 xx07 xx06 xx05 xx04 xx03 xx02
psrldq xmm1, 5 ; xx xx xx xx xx 0d 0c 0b 0a 09 08 07 06 05 04 03
pmullw xmm7, [rdx+64] ; x[ 2] * h[ 2] ; Tap 5
punpcklbw xmm1, xmm0 ; xx0a xx09 xx08 xx07 xx06 xx05 xx04 xx03
pmullw xmm1, [rdx+80] ; x[ 3] * h[ 3] ; Tap 6
paddsw xmm4, xmm7
paddsw xmm4, xmm5
paddsw xmm4, xmm3
paddsw xmm4, xmm6
paddsw xmm4, xmm1
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
paddsw xmm4, [GLOBAL(rd)]
Initial WebM release
2010-05-18 17:58:33 +02:00
psraw xmm4, 7
packuswb xmm4, xmm0
punpcklbw xmm4, xmm0
movdqa XMMWORD Ptr [rdi], xmm4
lea rsi, [rsi + rax]
%if ABI_IS_32BIT
add rdi, DWORD Ptr arg(5) ;[output_width]
%else
add rdi, r8
%endif
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
jnz .filter_block1d8_h6_rowloop ; next row
Initial WebM release
2010-05-18 17:58:33 +02:00
; 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
;void vp8_filter_block1d16_h6_sse2
;(
; unsigned char *src_ptr,
; unsigned short *output_ptr,
; unsigned int src_pixels_per_line,
; unsigned int pixel_step,
; unsigned int output_height,
; unsigned int output_width,
; short *vp8_filter
;)
;/************************************************************************************
; Notes: filter_block1d_h6 applies a 6 tap filter horizontally to the input pixels. The
; input pixel array has output_height rows. This routine assumes that output_height is an
; even number. This function handles 8 pixels in horizontal direction, calculating ONE
; rows each iteration to take advantage of the 128 bits operations.
;*************************************************************************************/
global sym(vp8_filter_block1d16_h6_sse2)
sym(vp8_filter_block1d16_h6_sse2):
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
Initial WebM release
2010-05-18 17:58:33 +02:00
GET_GOT rbx
push rsi
push rdi
; end prolog
mov rdx, arg(6) ;vp8_filter
mov rsi, arg(0) ;src_ptr
mov rdi, arg(1) ;output_ptr
movsxd rcx, dword ptr arg(4) ;output_height
movsxd rax, dword ptr arg(2) ;src_pixels_per_line ; Pitch for Source
%if ABI_IS_32BIT=0
movsxd r8, dword ptr arg(5) ;output_width
%endif
pxor xmm0, xmm0 ; clear xmm0 for unpack
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
.filter_block1d16_h6_sse2_rowloop:
Initial WebM release
2010-05-18 17:58:33 +02:00
movq xmm3, MMWORD PTR [rsi - 2]
movq xmm1, MMWORD PTR [rsi + 6]
movq xmm2, MMWORD PTR [rsi +14]
pslldq xmm2, 8
por xmm2, xmm1
prefetcht2 [rsi+rax-2]
pslldq xmm1, 8
por xmm1, xmm3
movdqa xmm4, xmm1
movdqa xmm5, xmm1
movdqa xmm6, xmm1
movdqa xmm7, xmm1
punpcklbw xmm3, xmm0 ; xx05 xx04 xx03 xx02 xx01 xx01 xx-1 xx-2
psrldq xmm4, 1 ; xx 0d 0c 0b 0a 09 08 07 06 05 04 03 02 01 00 -1
pmullw xmm3, XMMWORD PTR [rdx] ; x[-2] * H[-2]; Tap 1
punpcklbw xmm4, xmm0 ; xx06 xx05 xx04 xx03 xx02 xx01 xx00 xx-1
psrldq xmm5, 2 ; xx xx 0d 0c 0b 0a 09 08 07 06 05 04 03 02 01 00
pmullw xmm4, XMMWORD PTR [rdx+16] ; x[-1] * H[-1]; Tap 2
punpcklbw xmm5, xmm0 ; xx07 xx06 xx05 xx04 xx03 xx02 xx01 xx00
psrldq xmm6, 3 ; xx xx xx 0d 0c 0b 0a 09 08 07 06 05 04 03 02 01
pmullw xmm5, [rdx+32] ; x[ 0] * H[ 0]; Tap 3
punpcklbw xmm6, xmm0 ; xx08 xx07 xx06 xx05 xx04 xx03 xx02 xx01
psrldq xmm7, 4 ; xx xx xx xx 0d 0c 0b 0a 09 08 07 06 05 04 03 02
pmullw xmm6, [rdx+48] ; x[ 1] * h[ 1] ; Tap 4
punpcklbw xmm7, xmm0 ; xx09 xx08 xx07 xx06 xx05 xx04 xx03 xx02
psrldq xmm1, 5 ; xx xx xx xx xx 0d 0c 0b 0a 09 08 07 06 05 04 03
pmullw xmm7, [rdx+64] ; x[ 2] * h[ 2] ; Tap 5
punpcklbw xmm1, xmm0 ; xx0a xx09 xx08 xx07 xx06 xx05 xx04 xx03
pmullw xmm1, [rdx+80] ; x[ 3] * h[ 3] ; Tap 6
paddsw xmm4, xmm7
paddsw xmm4, xmm5
paddsw xmm4, xmm3
paddsw xmm4, xmm6
paddsw xmm4, xmm1
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
paddsw xmm4, [GLOBAL(rd)]
Initial WebM release
2010-05-18 17:58:33 +02:00
psraw xmm4, 7
packuswb xmm4, xmm0
punpcklbw xmm4, xmm0
movdqa XMMWORD Ptr [rdi], xmm4
movdqa xmm3, xmm2
movdqa xmm4, xmm2
movdqa xmm5, xmm2
movdqa xmm6, xmm2
movdqa xmm7, xmm2
punpcklbw xmm3, xmm0 ; xx05 xx04 xx03 xx02 xx01 xx01 xx-1 xx-2
psrldq xmm4, 1 ; xx 0d 0c 0b 0a 09 08 07 06 05 04 03 02 01 00 -1
pmullw xmm3, XMMWORD PTR [rdx] ; x[-2] * H[-2]; Tap 1
punpcklbw xmm4, xmm0 ; xx06 xx05 xx04 xx03 xx02 xx01 xx00 xx-1
psrldq xmm5, 2 ; xx xx 0d 0c 0b 0a 09 08 07 06 05 04 03 02 01 00
pmullw xmm4, XMMWORD PTR [rdx+16] ; x[-1] * H[-1]; Tap 2
punpcklbw xmm5, xmm0 ; xx07 xx06 xx05 xx04 xx03 xx02 xx01 xx00
psrldq xmm6, 3 ; xx xx xx 0d 0c 0b 0a 09 08 07 06 05 04 03 02 01
pmullw xmm5, [rdx+32] ; x[ 0] * H[ 0]; Tap 3
punpcklbw xmm6, xmm0 ; xx08 xx07 xx06 xx05 xx04 xx03 xx02 xx01
psrldq xmm7, 4 ; xx xx xx xx 0d 0c 0b 0a 09 08 07 06 05 04 03 02
pmullw xmm6, [rdx+48] ; x[ 1] * h[ 1] ; Tap 4
punpcklbw xmm7, xmm0 ; xx09 xx08 xx07 xx06 xx05 xx04 xx03 xx02
psrldq xmm2, 5 ; xx xx xx xx xx 0d 0c 0b 0a 09 08 07 06 05 04 03
pmullw xmm7, [rdx+64] ; x[ 2] * h[ 2] ; Tap 5
punpcklbw xmm2, xmm0 ; xx0a xx09 xx08 xx07 xx06 xx05 xx04 xx03
pmullw xmm2, [rdx+80] ; x[ 3] * h[ 3] ; Tap 6
paddsw xmm4, xmm7
paddsw xmm4, xmm5
paddsw xmm4, xmm3
paddsw xmm4, xmm6
paddsw xmm4, xmm2
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
paddsw xmm4, [GLOBAL(rd)]
Initial WebM release
2010-05-18 17:58:33 +02:00
psraw xmm4, 7
packuswb xmm4, xmm0
punpcklbw xmm4, xmm0
movdqa XMMWORD Ptr [rdi+16], xmm4
lea rsi, [rsi + rax]
%if ABI_IS_32BIT
add rdi, DWORD Ptr arg(5) ;[output_width]
%else
add rdi, r8
%endif
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
jnz .filter_block1d16_h6_sse2_rowloop ; next row
Initial WebM release
2010-05-18 17:58:33 +02:00
; 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
;void vp8_filter_block1d8_v6_sse2
;(
; short *src_ptr,
; unsigned char *output_ptr,
; int dst_ptich,
; unsigned int pixels_per_line,
; unsigned int pixel_step,
; unsigned int output_height,
; unsigned int output_width,
; short * vp8_filter
;)
;/************************************************************************************
; Notes: filter_block1d8_v6 applies a 6 tap filter vertically to the input pixels. The
; input pixel array has output_height rows.
;*************************************************************************************/
global sym(vp8_filter_block1d8_v6_sse2)
sym(vp8_filter_block1d8_v6_sse2):
push rbp
mov rbp, rsp
SHADOW_ARGS_TO_STACK 8
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
mov rax, arg(7) ;vp8_filter
movsxd rdx, dword ptr arg(3) ;pixels_per_line
mov rdi, arg(1) ;output_ptr
mov rsi, arg(0) ;src_ptr
sub rsi, rdx
sub rsi, rdx
movsxd rcx, DWORD PTR arg(5) ;[output_height]
pxor xmm0, xmm0 ; clear xmm0
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
movdqa xmm7, XMMWORD PTR [GLOBAL(rd)]
Initial WebM release
2010-05-18 17:58:33 +02:00
%if ABI_IS_32BIT=0
movsxd r8, dword ptr arg(2) ; dst_ptich
%endif
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
.vp8_filter_block1d8_v6_sse2_loop:
Initial WebM release
2010-05-18 17:58:33 +02:00
movdqa xmm1, XMMWORD PTR [rsi]
pmullw xmm1, [rax]
movdqa xmm2, XMMWORD PTR [rsi + rdx]
pmullw xmm2, [rax + 16]
movdqa xmm3, XMMWORD PTR [rsi + rdx * 2]
pmullw xmm3, [rax + 32]
movdqa xmm5, XMMWORD PTR [rsi + rdx * 4]
pmullw xmm5, [rax + 64]
add rsi, rdx
movdqa xmm4, XMMWORD PTR [rsi + rdx * 2]
pmullw xmm4, [rax + 48]
movdqa xmm6, XMMWORD PTR [rsi + rdx * 4]
pmullw xmm6, [rax + 80]
paddsw xmm2, xmm5
paddsw xmm2, xmm3
paddsw xmm2, xmm1
paddsw xmm2, xmm4
paddsw xmm2, xmm6
paddsw xmm2, xmm7
psraw xmm2, 7
packuswb xmm2, xmm0 ; pack and saturate
movq QWORD PTR [rdi], xmm2 ; store the results in the destination
%if ABI_IS_32BIT
add rdi, DWORD PTR arg(2) ;[dst_ptich]
%else
add rdi, r8
%endif
dec rcx ; decrement count
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
jnz .vp8_filter_block1d8_v6_sse2_loop ; next row
Initial WebM release
2010-05-18 17:58:33 +02:00
; 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
Improve vp8_sixtap_predict functions Restructure vp8_sixtap_predict functions to eliminate extra 5-line calculation while doing first-pass only. Also, combline functions to eliminate usage of intermediate buffer. This gives decoder a 3% performance gain on my test clips. Change-Id: I13de49638884d1a57d0855c63aea719316d08c1b
2010-06-10 17:48:48 +02:00
;void vp8_filter_block1d16_v6_sse2
;(
; unsigned short *src_ptr,
; unsigned char *output_ptr,
; int dst_ptich,
; unsigned int pixels_per_line,
; unsigned int pixel_step,
; unsigned int output_height,
; unsigned int output_width,
; const short *vp8_filter
;)
;/************************************************************************************
; Notes: filter_block1d16_v6 applies a 6 tap filter vertically to the input pixels. The
; input pixel array has output_height rows.
;*************************************************************************************/
global sym(vp8_filter_block1d16_v6_sse2)
sym(vp8_filter_block1d16_v6_sse2):
push rbp
mov rbp, rsp
SHADOW_ARGS_TO_STACK 8
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
Improve vp8_sixtap_predict functions Restructure vp8_sixtap_predict functions to eliminate extra 5-line calculation while doing first-pass only. Also, combline functions to eliminate usage of intermediate buffer. This gives decoder a 3% performance gain on my test clips. Change-Id: I13de49638884d1a57d0855c63aea719316d08c1b
2010-06-10 17:48:48 +02:00
GET_GOT rbx
push rsi
push rdi
; end prolog
mov rax, arg(7) ;vp8_filter
movsxd rdx, dword ptr arg(3) ;pixels_per_line
mov rdi, arg(1) ;output_ptr
mov rsi, arg(0) ;src_ptr
sub rsi, rdx
sub rsi, rdx
movsxd rcx, DWORD PTR arg(5) ;[output_height]
%if ABI_IS_32BIT=0
movsxd r8, dword ptr arg(2) ; dst_ptich
%endif
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
.vp8_filter_block1d16_v6_sse2_loop:
Improve vp8_sixtap_predict functions Restructure vp8_sixtap_predict functions to eliminate extra 5-line calculation while doing first-pass only. Also, combline functions to eliminate usage of intermediate buffer. This gives decoder a 3% performance gain on my test clips. Change-Id: I13de49638884d1a57d0855c63aea719316d08c1b
2010-06-10 17:48:48 +02:00
; The order for adding 6-tap is 2 5 3 1 4 6. Read in data in that order.
movdqa xmm1, XMMWORD PTR [rsi + rdx] ; line 2
movdqa xmm2, XMMWORD PTR [rsi + rdx + 16]
pmullw xmm1, [rax + 16]
pmullw xmm2, [rax + 16]
movdqa xmm3, XMMWORD PTR [rsi + rdx * 4] ; line 5
movdqa xmm4, XMMWORD PTR [rsi + rdx * 4 + 16]
pmullw xmm3, [rax + 64]
pmullw xmm4, [rax + 64]
movdqa xmm5, XMMWORD PTR [rsi + rdx * 2] ; line 3
movdqa xmm6, XMMWORD PTR [rsi + rdx * 2 + 16]
pmullw xmm5, [rax + 32]
pmullw xmm6, [rax + 32]
movdqa xmm7, XMMWORD PTR [rsi] ; line 1
movdqa xmm0, XMMWORD PTR [rsi + 16]
pmullw xmm7, [rax]
pmullw xmm0, [rax]
paddsw xmm1, xmm3
paddsw xmm2, xmm4
paddsw xmm1, xmm5
paddsw xmm2, xmm6
paddsw xmm1, xmm7
paddsw xmm2, xmm0
add rsi, rdx
movdqa xmm3, XMMWORD PTR [rsi + rdx * 2] ; line 4
movdqa xmm4, XMMWORD PTR [rsi + rdx * 2 + 16]
pmullw xmm3, [rax + 48]
pmullw xmm4, [rax + 48]
movdqa xmm5, XMMWORD PTR [rsi + rdx * 4] ; line 6
movdqa xmm6, XMMWORD PTR [rsi + rdx * 4 + 16]
pmullw xmm5, [rax + 80]
pmullw xmm6, [rax + 80]
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
movdqa xmm7, XMMWORD PTR [GLOBAL(rd)]
Improve vp8_sixtap_predict functions Restructure vp8_sixtap_predict functions to eliminate extra 5-line calculation while doing first-pass only. Also, combline functions to eliminate usage of intermediate buffer. This gives decoder a 3% performance gain on my test clips. Change-Id: I13de49638884d1a57d0855c63aea719316d08c1b
2010-06-10 17:48:48 +02:00
pxor xmm0, xmm0 ; clear xmm0
paddsw xmm1, xmm3
paddsw xmm2, xmm4
paddsw xmm1, xmm5
paddsw xmm2, xmm6
paddsw xmm1, xmm7
paddsw xmm2, xmm7
psraw xmm1, 7
psraw xmm2, 7
packuswb xmm1, xmm2 ; pack and saturate
movdqa XMMWORD PTR [rdi], xmm1 ; store the results in the destination
%if ABI_IS_32BIT
add rdi, DWORD PTR arg(2) ;[dst_ptich]
%else
add rdi, r8
%endif
dec rcx ; decrement count
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
jnz .vp8_filter_block1d16_v6_sse2_loop ; next row
Improve vp8_sixtap_predict functions Restructure vp8_sixtap_predict functions to eliminate extra 5-line calculation while doing first-pass only. Also, combline functions to eliminate usage of intermediate buffer. This gives decoder a 3% performance gain on my test clips. Change-Id: I13de49638884d1a57d0855c63aea719316d08c1b
2010-06-10 17:48:48 +02:00
; 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
Improve vp8_sixtap_predict functions Restructure vp8_sixtap_predict functions to eliminate extra 5-line calculation while doing first-pass only. Also, combline functions to eliminate usage of intermediate buffer. This gives decoder a 3% performance gain on my test clips. Change-Id: I13de49638884d1a57d0855c63aea719316d08c1b
2010-06-10 17:48:48 +02:00
UNSHADOW_ARGS
pop rbp
ret
;void vp8_filter_block1d8_h6_only_sse2
Initial WebM release
2010-05-18 17:58:33 +02:00
;(
; unsigned char *src_ptr,
; unsigned int src_pixels_per_line,
Improve vp8_sixtap_predict functions Restructure vp8_sixtap_predict functions to eliminate extra 5-line calculation while doing first-pass only. Also, combline functions to eliminate usage of intermediate buffer. This gives decoder a 3% performance gain on my test clips. Change-Id: I13de49638884d1a57d0855c63aea719316d08c1b
2010-06-10 17:48:48 +02:00
; unsigned char *output_ptr,
; int dst_ptich,
Initial WebM release
2010-05-18 17:58:33 +02:00
; unsigned int output_height,
Improve vp8_sixtap_predict functions Restructure vp8_sixtap_predict functions to eliminate extra 5-line calculation while doing first-pass only. Also, combline functions to eliminate usage of intermediate buffer. This gives decoder a 3% performance gain on my test clips. Change-Id: I13de49638884d1a57d0855c63aea719316d08c1b
2010-06-10 17:48:48 +02:00
; const short *vp8_filter
Initial WebM release
2010-05-18 17:58:33 +02:00
;)
Improve vp8_sixtap_predict functions Restructure vp8_sixtap_predict functions to eliminate extra 5-line calculation while doing first-pass only. Also, combline functions to eliminate usage of intermediate buffer. This gives decoder a 3% performance gain on my test clips. Change-Id: I13de49638884d1a57d0855c63aea719316d08c1b
2010-06-10 17:48:48 +02:00
; First-pass filter only when yoffset==0
global sym(vp8_filter_block1d8_h6_only_sse2)
sym(vp8_filter_block1d8_h6_only_sse2):
Initial WebM release
2010-05-18 17:58:33 +02:00
push rbp
mov rbp, rsp
Improve vp8_sixtap_predict functions Restructure vp8_sixtap_predict functions to eliminate extra 5-line calculation while doing first-pass only. Also, combline functions to eliminate usage of intermediate buffer. This gives decoder a 3% performance gain on my test clips. Change-Id: I13de49638884d1a57d0855c63aea719316d08c1b
2010-06-10 17:48:48 +02:00
SHADOW_ARGS_TO_STACK 6
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
Improve vp8_sixtap_predict functions Restructure vp8_sixtap_predict functions to eliminate extra 5-line calculation while doing first-pass only. Also, combline functions to eliminate usage of intermediate buffer. This gives decoder a 3% performance gain on my test clips. Change-Id: I13de49638884d1a57d0855c63aea719316d08c1b
2010-06-10 17:48:48 +02:00
mov rdx, arg(5) ;vp8_filter
Initial WebM release
2010-05-18 17:58:33 +02:00
mov rsi, arg(0) ;src_ptr
Improve vp8_sixtap_predict functions Restructure vp8_sixtap_predict functions to eliminate extra 5-line calculation while doing first-pass only. Also, combline functions to eliminate usage of intermediate buffer. This gives decoder a 3% performance gain on my test clips. Change-Id: I13de49638884d1a57d0855c63aea719316d08c1b
2010-06-10 17:48:48 +02:00
mov rdi, arg(2) ;output_ptr
Initial WebM release
2010-05-18 17:58:33 +02:00
Improve vp8_sixtap_predict functions Restructure vp8_sixtap_predict functions to eliminate extra 5-line calculation while doing first-pass only. Also, combline functions to eliminate usage of intermediate buffer. This gives decoder a 3% performance gain on my test clips. Change-Id: I13de49638884d1a57d0855c63aea719316d08c1b
2010-06-10 17:48:48 +02:00
movsxd rcx, dword ptr arg(4) ;output_height
movsxd rax, dword ptr arg(1) ;src_pixels_per_line ; Pitch for Source
Initial WebM release
2010-05-18 17:58:33 +02:00
%if ABI_IS_32BIT=0
Improve vp8_sixtap_predict functions Restructure vp8_sixtap_predict functions to eliminate extra 5-line calculation while doing first-pass only. Also, combline functions to eliminate usage of intermediate buffer. This gives decoder a 3% performance gain on my test clips. Change-Id: I13de49638884d1a57d0855c63aea719316d08c1b
2010-06-10 17:48:48 +02:00
movsxd r8, dword ptr arg(3) ;dst_ptich
Initial WebM release
2010-05-18 17:58:33 +02:00
%endif
Improve vp8_sixtap_predict functions Restructure vp8_sixtap_predict functions to eliminate extra 5-line calculation while doing first-pass only. Also, combline functions to eliminate usage of intermediate buffer. This gives decoder a 3% performance gain on my test clips. Change-Id: I13de49638884d1a57d0855c63aea719316d08c1b
2010-06-10 17:48:48 +02:00
pxor xmm0, xmm0 ; clear xmm0 for unpack
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
.filter_block1d8_h6_only_rowloop:
Improve vp8_sixtap_predict functions Restructure vp8_sixtap_predict functions to eliminate extra 5-line calculation while doing first-pass only. Also, combline functions to eliminate usage of intermediate buffer. This gives decoder a 3% performance gain on my test clips. Change-Id: I13de49638884d1a57d0855c63aea719316d08c1b
2010-06-10 17:48:48 +02:00
movq xmm3, MMWORD PTR [rsi - 2]
movq xmm1, MMWORD PTR [rsi + 6]
prefetcht2 [rsi+rax-2]
pslldq xmm1, 8
por xmm1, xmm3
movdqa xmm4, xmm1
movdqa xmm5, xmm1
movdqa xmm6, xmm1
movdqa xmm7, xmm1
Initial WebM release
2010-05-18 17:58:33 +02:00
punpcklbw xmm3, xmm0 ; xx05 xx04 xx03 xx02 xx01 xx01 xx-1 xx-2
Improve vp8_sixtap_predict functions Restructure vp8_sixtap_predict functions to eliminate extra 5-line calculation while doing first-pass only. Also, combline functions to eliminate usage of intermediate buffer. This gives decoder a 3% performance gain on my test clips. Change-Id: I13de49638884d1a57d0855c63aea719316d08c1b
2010-06-10 17:48:48 +02:00
psrldq xmm4, 1 ; xx 0d 0c 0b 0a 09 08 07 06 05 04 03 02 01 00 -1
Initial WebM release
2010-05-18 17:58:33 +02:00
Improve vp8_sixtap_predict functions Restructure vp8_sixtap_predict functions to eliminate extra 5-line calculation while doing first-pass only. Also, combline functions to eliminate usage of intermediate buffer. This gives decoder a 3% performance gain on my test clips. Change-Id: I13de49638884d1a57d0855c63aea719316d08c1b
2010-06-10 17:48:48 +02:00
pmullw xmm3, XMMWORD PTR [rdx] ; x[-2] * H[-2]; Tap 1
punpcklbw xmm4, xmm0 ; xx06 xx05 xx04 xx03 xx02 xx01 xx00 xx-1
psrldq xmm5, 2 ; xx xx 0d 0c 0b 0a 09 08 07 06 05 04 03 02 01 00
pmullw xmm4, XMMWORD PTR [rdx+16] ; x[-1] * H[-1]; Tap 2
punpcklbw xmm5, xmm0 ; xx07 xx06 xx05 xx04 xx03 xx02 xx01 xx00
psrldq xmm6, 3 ; xx xx xx 0d 0c 0b 0a 09 08 07 06 05 04 03 02 01
pmullw xmm5, [rdx+32] ; x[ 0] * H[ 0]; Tap 3
punpcklbw xmm6, xmm0 ; xx08 xx07 xx06 xx05 xx04 xx03 xx02 xx01
psrldq xmm7, 4 ; xx xx xx xx 0d 0c 0b 0a 09 08 07 06 05 04 03 02
pmullw xmm6, [rdx+48] ; x[ 1] * h[ 1] ; Tap 4
punpcklbw xmm7, xmm0 ; xx09 xx08 xx07 xx06 xx05 xx04 xx03 xx02
psrldq xmm1, 5 ; xx xx xx xx xx 0d 0c 0b 0a 09 08 07 06 05 04 03
pmullw xmm7, [rdx+64] ; x[ 2] * h[ 2] ; Tap 5
punpcklbw xmm1, xmm0 ; xx0a xx09 xx08 xx07 xx06 xx05 xx04 xx03
pmullw xmm1, [rdx+80] ; x[ 3] * h[ 3] ; Tap 6
Initial WebM release
2010-05-18 17:58:33 +02:00
Improve vp8_sixtap_predict functions Restructure vp8_sixtap_predict functions to eliminate extra 5-line calculation while doing first-pass only. Also, combline functions to eliminate usage of intermediate buffer. This gives decoder a 3% performance gain on my test clips. Change-Id: I13de49638884d1a57d0855c63aea719316d08c1b
2010-06-10 17:48:48 +02:00
paddsw xmm4, xmm7
paddsw xmm4, xmm5
paddsw xmm4, xmm3
paddsw xmm4, xmm6
paddsw xmm4, xmm1
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
paddsw xmm4, [GLOBAL(rd)]
Improve vp8_sixtap_predict functions Restructure vp8_sixtap_predict functions to eliminate extra 5-line calculation while doing first-pass only. Also, combline functions to eliminate usage of intermediate buffer. This gives decoder a 3% performance gain on my test clips. Change-Id: I13de49638884d1a57d0855c63aea719316d08c1b
2010-06-10 17:48:48 +02:00
psraw xmm4, 7
packuswb xmm4, xmm0
movq QWORD PTR [rdi], xmm4 ; store the results in the destination
Initial WebM release
2010-05-18 17:58:33 +02:00
lea rsi, [rsi + rax]
Improve vp8_sixtap_predict functions Restructure vp8_sixtap_predict functions to eliminate extra 5-line calculation while doing first-pass only. Also, combline functions to eliminate usage of intermediate buffer. This gives decoder a 3% performance gain on my test clips. Change-Id: I13de49638884d1a57d0855c63aea719316d08c1b
2010-06-10 17:48:48 +02:00
Initial WebM release
2010-05-18 17:58:33 +02:00
%if ABI_IS_32BIT
Improve vp8_sixtap_predict functions Restructure vp8_sixtap_predict functions to eliminate extra 5-line calculation while doing first-pass only. Also, combline functions to eliminate usage of intermediate buffer. This gives decoder a 3% performance gain on my test clips. Change-Id: I13de49638884d1a57d0855c63aea719316d08c1b
2010-06-10 17:48:48 +02:00
add rdi, DWORD Ptr arg(3) ;dst_ptich
Initial WebM release
2010-05-18 17:58:33 +02:00
%else
add rdi, r8
%endif
dec rcx
Improve vp8_sixtap_predict functions Restructure vp8_sixtap_predict functions to eliminate extra 5-line calculation while doing first-pass only. Also, combline functions to eliminate usage of intermediate buffer. This gives decoder a 3% performance gain on my test clips. Change-Id: I13de49638884d1a57d0855c63aea719316d08c1b
2010-06-10 17:48:48 +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
jnz .filter_block1d8_h6_only_rowloop ; next row
Initial WebM release
2010-05-18 17:58:33 +02:00
; begin epilog
pop rdi
pop rsi
RESTORE_GOT
More on "some XMM registers are non-volatile on windows x64 ABI" Add same fix in subpixel_sse2.asm. Change-Id: Icfda6103cbf74ec43308e96961dd738aa823c14d
2010-06-15 15:11:26 +02:00
RESTORE_XMM
Initial WebM release
2010-05-18 17:58:33 +02:00
UNSHADOW_ARGS
pop rbp
ret
Improve vp8_sixtap_predict functions Restructure vp8_sixtap_predict functions to eliminate extra 5-line calculation while doing first-pass only. Also, combline functions to eliminate usage of intermediate buffer. This gives decoder a 3% performance gain on my test clips. Change-Id: I13de49638884d1a57d0855c63aea719316d08c1b
2010-06-10 17:48:48 +02:00
;void vp8_filter_block1d16_h6_only_sse2
Initial WebM release
2010-05-18 17:58:33 +02:00
;(
; unsigned char *src_ptr,
; unsigned int src_pixels_per_line,
Improve vp8_sixtap_predict functions Restructure vp8_sixtap_predict functions to eliminate extra 5-line calculation while doing first-pass only. Also, combline functions to eliminate usage of intermediate buffer. This gives decoder a 3% performance gain on my test clips. Change-Id: I13de49638884d1a57d0855c63aea719316d08c1b
2010-06-10 17:48:48 +02:00
; unsigned char *output_ptr,
; int dst_ptich,
Initial WebM release
2010-05-18 17:58:33 +02:00
; unsigned int output_height,
Improve vp8_sixtap_predict functions Restructure vp8_sixtap_predict functions to eliminate extra 5-line calculation while doing first-pass only. Also, combline functions to eliminate usage of intermediate buffer. This gives decoder a 3% performance gain on my test clips. Change-Id: I13de49638884d1a57d0855c63aea719316d08c1b
2010-06-10 17:48:48 +02:00
; const short *vp8_filter
Initial WebM release
2010-05-18 17:58:33 +02:00
;)
Improve vp8_sixtap_predict functions Restructure vp8_sixtap_predict functions to eliminate extra 5-line calculation while doing first-pass only. Also, combline functions to eliminate usage of intermediate buffer. This gives decoder a 3% performance gain on my test clips. Change-Id: I13de49638884d1a57d0855c63aea719316d08c1b
2010-06-10 17:48:48 +02:00
; First-pass filter only when yoffset==0
global sym(vp8_filter_block1d16_h6_only_sse2)
sym(vp8_filter_block1d16_h6_only_sse2):
Initial WebM release
2010-05-18 17:58:33 +02:00
push rbp
mov rbp, rsp
Improve vp8_sixtap_predict functions Restructure vp8_sixtap_predict functions to eliminate extra 5-line calculation while doing first-pass only. Also, combline functions to eliminate usage of intermediate buffer. This gives decoder a 3% performance gain on my test clips. Change-Id: I13de49638884d1a57d0855c63aea719316d08c1b
2010-06-10 17:48:48 +02:00
SHADOW_ARGS_TO_STACK 6
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
Improve vp8_sixtap_predict functions Restructure vp8_sixtap_predict functions to eliminate extra 5-line calculation while doing first-pass only. Also, combline functions to eliminate usage of intermediate buffer. This gives decoder a 3% performance gain on my test clips. Change-Id: I13de49638884d1a57d0855c63aea719316d08c1b
2010-06-10 17:48:48 +02:00
mov rdx, arg(5) ;vp8_filter
Initial WebM release
2010-05-18 17:58:33 +02:00
mov rsi, arg(0) ;src_ptr
Improve vp8_sixtap_predict functions Restructure vp8_sixtap_predict functions to eliminate extra 5-line calculation while doing first-pass only. Also, combline functions to eliminate usage of intermediate buffer. This gives decoder a 3% performance gain on my test clips. Change-Id: I13de49638884d1a57d0855c63aea719316d08c1b
2010-06-10 17:48:48 +02:00
mov rdi, arg(2) ;output_ptr
Initial WebM release
2010-05-18 17:58:33 +02:00
Improve vp8_sixtap_predict functions Restructure vp8_sixtap_predict functions to eliminate extra 5-line calculation while doing first-pass only. Also, combline functions to eliminate usage of intermediate buffer. This gives decoder a 3% performance gain on my test clips. Change-Id: I13de49638884d1a57d0855c63aea719316d08c1b
2010-06-10 17:48:48 +02:00
movsxd rcx, dword ptr arg(4) ;output_height
movsxd rax, dword ptr arg(1) ;src_pixels_per_line ; Pitch for Source
Initial WebM release
2010-05-18 17:58:33 +02:00
%if ABI_IS_32BIT=0
Improve vp8_sixtap_predict functions Restructure vp8_sixtap_predict functions to eliminate extra 5-line calculation while doing first-pass only. Also, combline functions to eliminate usage of intermediate buffer. This gives decoder a 3% performance gain on my test clips. Change-Id: I13de49638884d1a57d0855c63aea719316d08c1b
2010-06-10 17:48:48 +02:00
movsxd r8, dword ptr arg(3) ;dst_ptich
Initial WebM release
2010-05-18 17:58:33 +02:00
%endif
Improve vp8_sixtap_predict functions Restructure vp8_sixtap_predict functions to eliminate extra 5-line calculation while doing first-pass only. Also, combline functions to eliminate usage of intermediate buffer. This gives decoder a 3% performance gain on my test clips. Change-Id: I13de49638884d1a57d0855c63aea719316d08c1b
2010-06-10 17:48:48 +02:00
pxor xmm0, xmm0 ; clear xmm0 for unpack
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
.filter_block1d16_h6_only_sse2_rowloop:
Improve vp8_sixtap_predict functions Restructure vp8_sixtap_predict functions to eliminate extra 5-line calculation while doing first-pass only. Also, combline functions to eliminate usage of intermediate buffer. This gives decoder a 3% performance gain on my test clips. Change-Id: I13de49638884d1a57d0855c63aea719316d08c1b
2010-06-10 17:48:48 +02:00
movq xmm3, MMWORD PTR [rsi - 2]
movq xmm1, MMWORD PTR [rsi + 6]
movq xmm2, MMWORD PTR [rsi +14]
pslldq xmm2, 8
Initial WebM release
2010-05-18 17:58:33 +02:00
Improve vp8_sixtap_predict functions Restructure vp8_sixtap_predict functions to eliminate extra 5-line calculation while doing first-pass only. Also, combline functions to eliminate usage of intermediate buffer. This gives decoder a 3% performance gain on my test clips. Change-Id: I13de49638884d1a57d0855c63aea719316d08c1b
2010-06-10 17:48:48 +02:00
por xmm2, xmm1
prefetcht2 [rsi+rax-2]
pslldq xmm1, 8
por xmm1, xmm3
movdqa xmm4, xmm1
movdqa xmm5, xmm1
movdqa xmm6, xmm1
movdqa xmm7, xmm1
punpcklbw xmm3, xmm0 ; xx05 xx04 xx03 xx02 xx01 xx01 xx-1 xx-2
psrldq xmm4, 1 ; xx 0d 0c 0b 0a 09 08 07 06 05 04 03 02 01 00 -1
pmullw xmm3, XMMWORD PTR [rdx] ; x[-2] * H[-2]; Tap 1
punpcklbw xmm4, xmm0 ; xx06 xx05 xx04 xx03 xx02 xx01 xx00 xx-1
psrldq xmm5, 2 ; xx xx 0d 0c 0b 0a 09 08 07 06 05 04 03 02 01 00
pmullw xmm4, XMMWORD PTR [rdx+16] ; x[-1] * H[-1]; Tap 2
punpcklbw xmm5, xmm0 ; xx07 xx06 xx05 xx04 xx03 xx02 xx01 xx00
psrldq xmm6, 3 ; xx xx xx 0d 0c 0b 0a 09 08 07 06 05 04 03 02 01
pmullw xmm5, [rdx+32] ; x[ 0] * H[ 0]; Tap 3
punpcklbw xmm6, xmm0 ; xx08 xx07 xx06 xx05 xx04 xx03 xx02 xx01
psrldq xmm7, 4 ; xx xx xx xx 0d 0c 0b 0a 09 08 07 06 05 04 03 02
pmullw xmm6, [rdx+48] ; x[ 1] * h[ 1] ; Tap 4
punpcklbw xmm7, xmm0 ; xx09 xx08 xx07 xx06 xx05 xx04 xx03 xx02
psrldq xmm1, 5 ; xx xx xx xx xx 0d 0c 0b 0a 09 08 07 06 05 04 03
pmullw xmm7, [rdx+64] ; x[ 2] * h[ 2] ; Tap 5
punpcklbw xmm1, xmm0 ; xx0a xx09 xx08 xx07 xx06 xx05 xx04 xx03
pmullw xmm1, [rdx+80] ; x[ 3] * h[ 3] ; Tap 6
paddsw xmm4, xmm7
paddsw xmm4, xmm5
paddsw xmm4, xmm3
paddsw xmm4, xmm6
paddsw xmm4, xmm1
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
paddsw xmm4, [GLOBAL(rd)]
Improve vp8_sixtap_predict functions Restructure vp8_sixtap_predict functions to eliminate extra 5-line calculation while doing first-pass only. Also, combline functions to eliminate usage of intermediate buffer. This gives decoder a 3% performance gain on my test clips. Change-Id: I13de49638884d1a57d0855c63aea719316d08c1b
2010-06-10 17:48:48 +02:00
psraw xmm4, 7
packuswb xmm4, xmm0 ; lower 8 bytes
movq QWORD Ptr [rdi], xmm4 ; store the results in the destination
movdqa xmm3, xmm2
movdqa xmm4, xmm2
movdqa xmm5, xmm2
movdqa xmm6, xmm2
movdqa xmm7, xmm2
punpcklbw xmm3, xmm0 ; xx05 xx04 xx03 xx02 xx01 xx01 xx-1 xx-2
psrldq xmm4, 1 ; xx 0d 0c 0b 0a 09 08 07 06 05 04 03 02 01 00 -1
pmullw xmm3, XMMWORD PTR [rdx] ; x[-2] * H[-2]; Tap 1
punpcklbw xmm4, xmm0 ; xx06 xx05 xx04 xx03 xx02 xx01 xx00 xx-1
psrldq xmm5, 2 ; xx xx 0d 0c 0b 0a 09 08 07 06 05 04 03 02 01 00
pmullw xmm4, XMMWORD PTR [rdx+16] ; x[-1] * H[-1]; Tap 2
punpcklbw xmm5, xmm0 ; xx07 xx06 xx05 xx04 xx03 xx02 xx01 xx00
psrldq xmm6, 3 ; xx xx xx 0d 0c 0b 0a 09 08 07 06 05 04 03 02 01
pmullw xmm5, [rdx+32] ; x[ 0] * H[ 0]; Tap 3
punpcklbw xmm6, xmm0 ; xx08 xx07 xx06 xx05 xx04 xx03 xx02 xx01
psrldq xmm7, 4 ; xx xx xx xx 0d 0c 0b 0a 09 08 07 06 05 04 03 02
pmullw xmm6, [rdx+48] ; x[ 1] * h[ 1] ; Tap 4
punpcklbw xmm7, xmm0 ; xx09 xx08 xx07 xx06 xx05 xx04 xx03 xx02
psrldq xmm2, 5 ; xx xx xx xx xx 0d 0c 0b 0a 09 08 07 06 05 04 03
pmullw xmm7, [rdx+64] ; x[ 2] * h[ 2] ; Tap 5
punpcklbw xmm2, xmm0 ; xx0a xx09 xx08 xx07 xx06 xx05 xx04 xx03
pmullw xmm2, [rdx+80] ; x[ 3] * h[ 3] ; Tap 6
paddsw xmm4, xmm7
paddsw xmm4, xmm5
paddsw xmm4, xmm3
paddsw xmm4, xmm6
paddsw xmm4, xmm2
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
paddsw xmm4, [GLOBAL(rd)]
Improve vp8_sixtap_predict functions Restructure vp8_sixtap_predict functions to eliminate extra 5-line calculation while doing first-pass only. Also, combline functions to eliminate usage of intermediate buffer. This gives decoder a 3% performance gain on my test clips. Change-Id: I13de49638884d1a57d0855c63aea719316d08c1b
2010-06-10 17:48:48 +02:00
psraw xmm4, 7
packuswb xmm4, xmm0 ; higher 8 bytes
movq QWORD Ptr [rdi+8], xmm4 ; store the results in the destination
lea rsi, [rsi + rax]
Initial WebM release
2010-05-18 17:58:33 +02:00
%if ABI_IS_32BIT
Improve vp8_sixtap_predict functions Restructure vp8_sixtap_predict functions to eliminate extra 5-line calculation while doing first-pass only. Also, combline functions to eliminate usage of intermediate buffer. This gives decoder a 3% performance gain on my test clips. Change-Id: I13de49638884d1a57d0855c63aea719316d08c1b
2010-06-10 17:48:48 +02:00
add rdi, DWORD Ptr arg(3) ;dst_ptich
Initial WebM release
2010-05-18 17:58:33 +02:00
%else
add rdi, r8
%endif
Improve vp8_sixtap_predict functions Restructure vp8_sixtap_predict functions to eliminate extra 5-line calculation while doing first-pass only. Also, combline functions to eliminate usage of intermediate buffer. This gives decoder a 3% performance gain on my test clips. Change-Id: I13de49638884d1a57d0855c63aea719316d08c1b
2010-06-10 17:48:48 +02:00
Initial WebM release
2010-05-18 17:58:33 +02:00
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
jnz .filter_block1d16_h6_only_sse2_rowloop ; next row
Initial WebM release
2010-05-18 17:58:33 +02:00
; begin epilog
pop rdi
pop rsi
RESTORE_GOT
More on "some XMM registers are non-volatile on windows x64 ABI" Add same fix in subpixel_sse2.asm. Change-Id: Icfda6103cbf74ec43308e96961dd738aa823c14d
2010-06-15 15:11:26 +02:00
RESTORE_XMM
Initial WebM release
2010-05-18 17:58:33 +02:00
UNSHADOW_ARGS
pop rbp
ret
Improve vp8_sixtap_predict functions Restructure vp8_sixtap_predict functions to eliminate extra 5-line calculation while doing first-pass only. Also, combline functions to eliminate usage of intermediate buffer. This gives decoder a 3% performance gain on my test clips. Change-Id: I13de49638884d1a57d0855c63aea719316d08c1b
2010-06-10 17:48:48 +02:00
;void vp8_filter_block1d8_v6_only_sse2
Initial WebM release
2010-05-18 17:58:33 +02:00
;(
Improve vp8_sixtap_predict functions Restructure vp8_sixtap_predict functions to eliminate extra 5-line calculation while doing first-pass only. Also, combline functions to eliminate usage of intermediate buffer. This gives decoder a 3% performance gain on my test clips. Change-Id: I13de49638884d1a57d0855c63aea719316d08c1b
2010-06-10 17:48:48 +02:00
; unsigned char *src_ptr,
; unsigned int src_pixels_per_line,
Initial WebM release
2010-05-18 17:58:33 +02:00
; unsigned char *output_ptr,
; int dst_ptich,
; unsigned int output_height,
Improve vp8_sixtap_predict functions Restructure vp8_sixtap_predict functions to eliminate extra 5-line calculation while doing first-pass only. Also, combline functions to eliminate usage of intermediate buffer. This gives decoder a 3% performance gain on my test clips. Change-Id: I13de49638884d1a57d0855c63aea719316d08c1b
2010-06-10 17:48:48 +02:00
; const short *vp8_filter
Initial WebM release
2010-05-18 17:58:33 +02:00
;)
Improve vp8_sixtap_predict functions Restructure vp8_sixtap_predict functions to eliminate extra 5-line calculation while doing first-pass only. Also, combline functions to eliminate usage of intermediate buffer. This gives decoder a 3% performance gain on my test clips. Change-Id: I13de49638884d1a57d0855c63aea719316d08c1b
2010-06-10 17:48:48 +02:00
; Second-pass filter only when xoffset==0
global sym(vp8_filter_block1d8_v6_only_sse2)
sym(vp8_filter_block1d8_v6_only_sse2):
Initial WebM release
2010-05-18 17:58:33 +02:00
push rbp
mov rbp, rsp
SHADOW_ARGS_TO_STACK 6
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
mov rsi, arg(0) ;src_ptr
Improve vp8_sixtap_predict functions Restructure vp8_sixtap_predict functions to eliminate extra 5-line calculation while doing first-pass only. Also, combline functions to eliminate usage of intermediate buffer. This gives decoder a 3% performance gain on my test clips. Change-Id: I13de49638884d1a57d0855c63aea719316d08c1b
2010-06-10 17:48:48 +02:00
mov rdi, arg(2) ;output_ptr
movsxd rcx, dword ptr arg(4) ;output_height
movsxd rdx, dword ptr arg(1) ;src_pixels_per_line
mov rax, arg(5) ;vp8_filter
pxor xmm0, xmm0 ; clear xmm0
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
movdqa xmm7, XMMWORD PTR [GLOBAL(rd)]
Initial WebM release
2010-05-18 17:58:33 +02:00
%if ABI_IS_32BIT=0
Improve vp8_sixtap_predict functions Restructure vp8_sixtap_predict functions to eliminate extra 5-line calculation while doing first-pass only. Also, combline functions to eliminate usage of intermediate buffer. This gives decoder a 3% performance gain on my test clips. Change-Id: I13de49638884d1a57d0855c63aea719316d08c1b
2010-06-10 17:48:48 +02:00
movsxd r8, dword ptr arg(3) ; dst_ptich
Initial WebM release
2010-05-18 17:58:33 +02:00
%endif
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
.vp8_filter_block1d8_v6_only_sse2_loop:
Improve vp8_sixtap_predict functions Restructure vp8_sixtap_predict functions to eliminate extra 5-line calculation while doing first-pass only. Also, combline functions to eliminate usage of intermediate buffer. This gives decoder a 3% performance gain on my test clips. Change-Id: I13de49638884d1a57d0855c63aea719316d08c1b
2010-06-10 17:48:48 +02:00
movq xmm1, MMWORD PTR [rsi]
movq xmm2, MMWORD PTR [rsi + rdx]
movq xmm3, MMWORD PTR [rsi + rdx * 2]
movq xmm5, MMWORD PTR [rsi + rdx * 4]
add rsi, rdx
movq xmm4, MMWORD PTR [rsi + rdx * 2]
movq xmm6, MMWORD PTR [rsi + rdx * 4]
punpcklbw xmm1, xmm0
pmullw xmm1, [rax]
punpcklbw xmm2, xmm0
pmullw xmm2, [rax + 16]
punpcklbw xmm3, xmm0
pmullw xmm3, [rax + 32]
punpcklbw xmm5, xmm0
pmullw xmm5, [rax + 64]
punpcklbw xmm4, xmm0
pmullw xmm4, [rax + 48]
punpcklbw xmm6, xmm0
pmullw xmm6, [rax + 80]
paddsw xmm2, xmm5
paddsw xmm2, xmm3
paddsw xmm2, xmm1
paddsw xmm2, xmm4
paddsw xmm2, xmm6
paddsw xmm2, xmm7
Initial WebM release
2010-05-18 17:58:33 +02:00
Improve vp8_sixtap_predict functions Restructure vp8_sixtap_predict functions to eliminate extra 5-line calculation while doing first-pass only. Also, combline functions to eliminate usage of intermediate buffer. This gives decoder a 3% performance gain on my test clips. Change-Id: I13de49638884d1a57d0855c63aea719316d08c1b
2010-06-10 17:48:48 +02:00
psraw xmm2, 7
packuswb xmm2, xmm0 ; pack and saturate
Initial WebM release
2010-05-18 17:58:33 +02:00
Improve vp8_sixtap_predict functions Restructure vp8_sixtap_predict functions to eliminate extra 5-line calculation while doing first-pass only. Also, combline functions to eliminate usage of intermediate buffer. This gives decoder a 3% performance gain on my test clips. Change-Id: I13de49638884d1a57d0855c63aea719316d08c1b
2010-06-10 17:48:48 +02:00
movq QWORD PTR [rdi], xmm2 ; store the results in the destination
Initial WebM release
2010-05-18 17:58:33 +02:00
%if ABI_IS_32BIT
Improve vp8_sixtap_predict functions Restructure vp8_sixtap_predict functions to eliminate extra 5-line calculation while doing first-pass only. Also, combline functions to eliminate usage of intermediate buffer. This gives decoder a 3% performance gain on my test clips. Change-Id: I13de49638884d1a57d0855c63aea719316d08c1b
2010-06-10 17:48:48 +02:00
add rdi, DWORD PTR arg(3) ;[dst_ptich]
Initial WebM release
2010-05-18 17:58:33 +02:00
%else
add rdi, r8
%endif
dec rcx ; decrement count
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
jnz .vp8_filter_block1d8_v6_only_sse2_loop ; next row
Initial WebM release
2010-05-18 17:58:33 +02:00
; 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
Improve vp8_sixtap_predict functions Restructure vp8_sixtap_predict functions to eliminate extra 5-line calculation while doing first-pass only. Also, combline functions to eliminate usage of intermediate buffer. This gives decoder a 3% performance gain on my test clips. Change-Id: I13de49638884d1a57d0855c63aea719316d08c1b
2010-06-10 17:48:48 +02:00
;void vp8_unpack_block1d16_h6_sse2
Initial WebM release
2010-05-18 17:58:33 +02:00
;(
Improve vp8_sixtap_predict functions Restructure vp8_sixtap_predict functions to eliminate extra 5-line calculation while doing first-pass only. Also, combline functions to eliminate usage of intermediate buffer. This gives decoder a 3% performance gain on my test clips. Change-Id: I13de49638884d1a57d0855c63aea719316d08c1b
2010-06-10 17:48:48 +02:00
; unsigned char *src_ptr,
; unsigned short *output_ptr,
; unsigned int src_pixels_per_line,
; unsigned int output_height,
; unsigned int output_width
Initial WebM release
2010-05-18 17:58:33 +02:00
;)
Improve vp8_sixtap_predict functions Restructure vp8_sixtap_predict functions to eliminate extra 5-line calculation while doing first-pass only. Also, combline functions to eliminate usage of intermediate buffer. This gives decoder a 3% performance gain on my test clips. Change-Id: I13de49638884d1a57d0855c63aea719316d08c1b
2010-06-10 17:48:48 +02:00
global sym(vp8_unpack_block1d16_h6_sse2)
sym(vp8_unpack_block1d16_h6_sse2):
Initial WebM release
2010-05-18 17:58:33 +02:00
push rbp
mov rbp, rsp
Improve vp8_sixtap_predict functions Restructure vp8_sixtap_predict functions to eliminate extra 5-line calculation while doing first-pass only. Also, combline functions to eliminate usage of intermediate buffer. This gives decoder a 3% performance gain on my test clips. Change-Id: I13de49638884d1a57d0855c63aea719316d08c1b
2010-06-10 17:48:48 +02:00
SHADOW_ARGS_TO_STACK 5
Initial WebM release
2010-05-18 17:58:33 +02:00
GET_GOT rbx
push rsi
push rdi
; end prolog
Improve vp8_sixtap_predict functions Restructure vp8_sixtap_predict functions to eliminate extra 5-line calculation while doing first-pass only. Also, combline functions to eliminate usage of intermediate buffer. This gives decoder a 3% performance gain on my test clips. Change-Id: I13de49638884d1a57d0855c63aea719316d08c1b
2010-06-10 17:48:48 +02:00
mov rsi, arg(0) ;src_ptr
Initial WebM release
2010-05-18 17:58:33 +02:00
mov rdi, arg(1) ;output_ptr
Improve vp8_sixtap_predict functions Restructure vp8_sixtap_predict functions to eliminate extra 5-line calculation while doing first-pass only. Also, combline functions to eliminate usage of intermediate buffer. This gives decoder a 3% performance gain on my test clips. Change-Id: I13de49638884d1a57d0855c63aea719316d08c1b
2010-06-10 17:48:48 +02:00
movsxd rcx, dword ptr arg(3) ;output_height
movsxd rax, dword ptr arg(2) ;src_pixels_per_line ; Pitch for Source
pxor xmm0, xmm0 ; clear xmm0 for unpack
Initial WebM release
2010-05-18 17:58:33 +02:00
%if ABI_IS_32BIT=0
Improve vp8_sixtap_predict functions Restructure vp8_sixtap_predict functions to eliminate extra 5-line calculation while doing first-pass only. Also, combline functions to eliminate usage of intermediate buffer. This gives decoder a 3% performance gain on my test clips. Change-Id: I13de49638884d1a57d0855c63aea719316d08c1b
2010-06-10 17:48:48 +02:00
movsxd r8, dword ptr arg(4) ;output_width ; Pitch for Source
Initial WebM release
2010-05-18 17:58:33 +02:00
%endif
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
.unpack_block1d16_h6_sse2_rowloop:
Improve vp8_sixtap_predict functions Restructure vp8_sixtap_predict functions to eliminate extra 5-line calculation while doing first-pass only. Also, combline functions to eliminate usage of intermediate buffer. This gives decoder a 3% performance gain on my test clips. Change-Id: I13de49638884d1a57d0855c63aea719316d08c1b
2010-06-10 17:48:48 +02:00
movq xmm1, MMWORD PTR [rsi] ; 0d 0c 0b 0a 09 08 07 06 05 04 03 02 01 00 -1 -2
movq xmm3, MMWORD PTR [rsi+8] ; make copy of xmm1
punpcklbw xmm3, xmm0 ; xx05 xx04 xx03 xx02 xx01 xx01 xx-1 xx-2
punpcklbw xmm1, xmm0
Initial WebM release
2010-05-18 17:58:33 +02:00
Improve vp8_sixtap_predict functions Restructure vp8_sixtap_predict functions to eliminate extra 5-line calculation while doing first-pass only. Also, combline functions to eliminate usage of intermediate buffer. This gives decoder a 3% performance gain on my test clips. Change-Id: I13de49638884d1a57d0855c63aea719316d08c1b
2010-06-10 17:48:48 +02:00
movdqa XMMWORD Ptr [rdi], xmm1
movdqa XMMWORD Ptr [rdi + 16], xmm3
Initial WebM release
2010-05-18 17:58:33 +02:00
Improve vp8_sixtap_predict functions Restructure vp8_sixtap_predict functions to eliminate extra 5-line calculation while doing first-pass only. Also, combline functions to eliminate usage of intermediate buffer. This gives decoder a 3% performance gain on my test clips. Change-Id: I13de49638884d1a57d0855c63aea719316d08c1b
2010-06-10 17:48:48 +02:00
lea rsi, [rsi + rax]
Initial WebM release
2010-05-18 17:58:33 +02:00
%if ABI_IS_32BIT
Improve vp8_sixtap_predict functions Restructure vp8_sixtap_predict functions to eliminate extra 5-line calculation while doing first-pass only. Also, combline functions to eliminate usage of intermediate buffer. This gives decoder a 3% performance gain on my test clips. Change-Id: I13de49638884d1a57d0855c63aea719316d08c1b
2010-06-10 17:48:48 +02:00
add rdi, DWORD Ptr arg(4) ;[output_width]
Initial WebM release
2010-05-18 17:58:33 +02:00
%else
add rdi, r8
%endif
Improve vp8_sixtap_predict functions Restructure vp8_sixtap_predict functions to eliminate extra 5-line calculation while doing first-pass only. Also, combline functions to eliminate usage of intermediate buffer. This gives decoder a 3% performance gain on my test clips. Change-Id: I13de49638884d1a57d0855c63aea719316d08c1b
2010-06-10 17:48:48 +02:00
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
jnz .unpack_block1d16_h6_sse2_rowloop ; next row
Initial WebM release
2010-05-18 17:58:33 +02:00
; begin epilog
pop rdi
pop rsi
RESTORE_GOT
UNSHADOW_ARGS
pop rbp
ret
;void vp8_bilinear_predict16x16_sse2
;(
; unsigned char *src_ptr,
; int src_pixels_per_line,
; int xoffset,
; int yoffset,
; unsigned char *dst_ptr,
; int dst_pitch
;)
Move shared data to shared location Storing vp8_bilinear_filters_mmx in an mmx file and using it in an sse2 file is bad Moving towards allowing --disable-mmx Change-Id: I20493b35bdedcdcfc0915e6f05fdbe6c81a4a742
2011-11-18 20:50:13 +01:00
extern sym(vp8_bilinear_filters_x86_8)
Initial WebM release
2010-05-18 17:58:33 +02:00
global sym(vp8_bilinear_predict16x16_sse2)
sym(vp8_bilinear_predict16x16_sse2):
push rbp
mov rbp, rsp
SHADOW_ARGS_TO_STACK 6
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
Move shared data to shared location Storing vp8_bilinear_filters_mmx in an mmx file and using it in an sse2 file is bad Moving towards allowing --disable-mmx Change-Id: I20493b35bdedcdcfc0915e6f05fdbe6c81a4a742
2011-11-18 20:50:13 +01:00
;const short *HFilter = vp8_bilinear_filters_x86_8[xoffset]
;const short *VFilter = vp8_bilinear_filters_x86_8[yoffset]
Initial WebM release
2010-05-18 17:58:33 +02:00
Move shared data to shared location Storing vp8_bilinear_filters_mmx in an mmx file and using it in an sse2 file is bad Moving towards allowing --disable-mmx Change-Id: I20493b35bdedcdcfc0915e6f05fdbe6c81a4a742
2011-11-18 20:50:13 +01:00
lea rcx, [GLOBAL(sym(vp8_bilinear_filters_x86_8))]
Initial WebM release
2010-05-18 17:58:33 +02:00
movsxd rax, dword ptr arg(2) ;xoffset
cmp rax, 0 ;skip first_pass filter if xoffset=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
je .b16x16_sp_only
Initial WebM release
2010-05-18 17:58:33 +02:00
shl rax, 5
add rax, rcx ;HFilter
mov rdi, arg(4) ;dst_ptr
mov rsi, arg(0) ;src_ptr
movsxd rdx, dword ptr arg(5) ;dst_pitch
movdqa xmm1, [rax]
movdqa xmm2, [rax+16]
movsxd rax, dword ptr arg(3) ;yoffset
cmp rax, 0 ;skip second_pass filter if yoffset=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
je .b16x16_fp_only
Initial WebM release
2010-05-18 17:58:33 +02:00
shl rax, 5
add rax, rcx ;VFilter
lea rcx, [rdi+rdx*8]
lea rcx, [rcx+rdx*8]
movsxd rdx, dword ptr arg(1) ;src_pixels_per_line
pxor xmm0, xmm0
%if ABI_IS_32BIT=0
movsxd r8, dword ptr arg(5) ;dst_pitch
%endif
; get the first horizontal line done
movdqu xmm3, [rsi] ; xx 00 01 02 03 04 05 06 07 08 09 10 11 12 13 14
movdqa xmm4, xmm3 ; make a copy of current line
punpcklbw xmm3, xmm0 ; xx 00 01 02 03 04 05 06
punpckhbw xmm4, xmm0
pmullw xmm3, xmm1
pmullw xmm4, xmm1
movdqu xmm5, [rsi+1]
movdqa xmm6, xmm5
punpcklbw xmm5, xmm0
punpckhbw xmm6, xmm0
pmullw xmm5, xmm2
pmullw xmm6, xmm2
paddw xmm3, xmm5
paddw xmm4, 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
paddw xmm3, [GLOBAL(rd)] ; xmm3 += round value
Initial WebM release
2010-05-18 17:58:33 +02:00
psraw xmm3, VP8_FILTER_SHIFT ; xmm3 /= 128
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
paddw xmm4, [GLOBAL(rd)]
Initial WebM release
2010-05-18 17:58:33 +02:00
psraw xmm4, VP8_FILTER_SHIFT
movdqa xmm7, xmm3
packuswb xmm7, xmm4
add rsi, rdx ; next line
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
.next_row:
Initial WebM release
2010-05-18 17:58:33 +02:00
movdqu xmm3, [rsi] ; xx 00 01 02 03 04 05 06 07 08 09 10 11 12 13 14
movdqa xmm4, xmm3 ; make a copy of current line
punpcklbw xmm3, xmm0 ; xx 00 01 02 03 04 05 06
punpckhbw xmm4, xmm0
pmullw xmm3, xmm1
pmullw xmm4, xmm1
movdqu xmm5, [rsi+1]
movdqa xmm6, xmm5
punpcklbw xmm5, xmm0
punpckhbw xmm6, xmm0
pmullw xmm5, xmm2
pmullw xmm6, xmm2
paddw xmm3, xmm5
paddw xmm4, xmm6
movdqa xmm5, xmm7
movdqa xmm6, xmm7
punpcklbw xmm5, xmm0
punpckhbw xmm6, xmm0
pmullw xmm5, [rax]
pmullw xmm6, [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
paddw xmm3, [GLOBAL(rd)] ; xmm3 += round value
Initial WebM release
2010-05-18 17:58:33 +02:00
psraw xmm3, VP8_FILTER_SHIFT ; xmm3 /= 128
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
paddw xmm4, [GLOBAL(rd)]
Initial WebM release
2010-05-18 17:58:33 +02:00
psraw xmm4, VP8_FILTER_SHIFT
movdqa xmm7, xmm3
packuswb xmm7, xmm4
pmullw xmm3, [rax+16]
pmullw xmm4, [rax+16]
paddw xmm3, xmm5
paddw xmm4, 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
paddw xmm3, [GLOBAL(rd)] ; xmm3 += round value
Initial WebM release
2010-05-18 17:58:33 +02:00
psraw xmm3, VP8_FILTER_SHIFT ; xmm3 /= 128
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
paddw xmm4, [GLOBAL(rd)]
Initial WebM release
2010-05-18 17:58:33 +02:00
psraw xmm4, VP8_FILTER_SHIFT
packuswb xmm3, xmm4
movdqa [rdi], xmm3 ; store the results in the destination
add rsi, rdx ; next line
%if ABI_IS_32BIT
add rdi, DWORD PTR arg(5) ;dst_pitch
%else
add rdi, r8
%endif
cmp rdi, 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 .next_row
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
jmp .done
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
.b16x16_sp_only:
Initial WebM release
2010-05-18 17:58:33 +02:00
movsxd rax, dword ptr arg(3) ;yoffset
shl rax, 5
add rax, rcx ;VFilter
mov rdi, arg(4) ;dst_ptr
mov rsi, arg(0) ;src_ptr
movsxd rdx, dword ptr arg(5) ;dst_pitch
movdqa xmm1, [rax]
movdqa xmm2, [rax+16]
lea rcx, [rdi+rdx*8]
lea rcx, [rcx+rdx*8]
movsxd rax, dword ptr arg(1) ;src_pixels_per_line
pxor xmm0, xmm0
; get the first horizontal line done
movdqu xmm7, [rsi] ; xx 00 01 02 03 04 05 06 07 08 09 10 11 12 13 14
add rsi, rax ; next line
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
.next_row_spo:
Initial WebM release
2010-05-18 17:58:33 +02:00
movdqu xmm3, [rsi] ; xx 00 01 02 03 04 05 06 07 08 09 10 11 12 13 14
movdqa xmm5, xmm7
movdqa xmm6, xmm7
movdqa xmm4, xmm3 ; make a copy of current line
movdqa xmm7, xmm3
punpcklbw xmm5, xmm0
punpckhbw xmm6, xmm0
punpcklbw xmm3, xmm0 ; xx 00 01 02 03 04 05 06
punpckhbw xmm4, xmm0
pmullw xmm5, xmm1
pmullw xmm6, xmm1
pmullw xmm3, xmm2
pmullw xmm4, xmm2
paddw xmm3, xmm5
paddw xmm4, 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
paddw xmm3, [GLOBAL(rd)] ; xmm3 += round value
Initial WebM release
2010-05-18 17:58:33 +02:00
psraw xmm3, VP8_FILTER_SHIFT ; xmm3 /= 128
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
paddw xmm4, [GLOBAL(rd)]
Initial WebM release
2010-05-18 17:58:33 +02:00
psraw xmm4, VP8_FILTER_SHIFT
packuswb xmm3, xmm4
movdqa [rdi], xmm3 ; store the results in the destination
add rsi, rax ; next line
add rdi, rdx ;dst_pitch
cmp rdi, 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 .next_row_spo
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
jmp .done
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
.b16x16_fp_only:
Initial WebM release
2010-05-18 17:58:33 +02:00
lea rcx, [rdi+rdx*8]
lea rcx, [rcx+rdx*8]
movsxd rax, dword ptr arg(1) ;src_pixels_per_line
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
.next_row_fpo:
Initial WebM release
2010-05-18 17:58:33 +02:00
movdqu xmm3, [rsi] ; xx 00 01 02 03 04 05 06 07 08 09 10 11 12 13 14
movdqa xmm4, xmm3 ; make a copy of current line
punpcklbw xmm3, xmm0 ; xx 00 01 02 03 04 05 06
punpckhbw xmm4, xmm0
pmullw xmm3, xmm1
pmullw xmm4, xmm1
movdqu xmm5, [rsi+1]
movdqa xmm6, xmm5
punpcklbw xmm5, xmm0
punpckhbw xmm6, xmm0
pmullw xmm5, xmm2
pmullw xmm6, xmm2
paddw xmm3, xmm5
paddw xmm4, 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
paddw xmm3, [GLOBAL(rd)] ; xmm3 += round value
Initial WebM release
2010-05-18 17:58:33 +02:00
psraw xmm3, VP8_FILTER_SHIFT ; xmm3 /= 128
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
paddw xmm4, [GLOBAL(rd)]
Initial WebM release
2010-05-18 17:58:33 +02:00
psraw xmm4, VP8_FILTER_SHIFT
packuswb xmm3, xmm4
movdqa [rdi], xmm3 ; store the results in the destination
add rsi, rax ; next line
add rdi, rdx ; dst_pitch
cmp rdi, 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 .next_row_fpo
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
.done:
Initial WebM release
2010-05-18 17:58:33 +02:00
; begin epilog
pop rdi
pop rsi
RESTORE_GOT
More on "some XMM registers are non-volatile on windows x64 ABI" Add same fix in subpixel_sse2.asm. Change-Id: Icfda6103cbf74ec43308e96961dd738aa823c14d
2010-06-15 15:11:26 +02:00
RESTORE_XMM
Initial WebM release
2010-05-18 17:58:33 +02:00
UNSHADOW_ARGS
pop rbp
ret
;void vp8_bilinear_predict8x8_sse2
;(
; unsigned char *src_ptr,
; int src_pixels_per_line,
; int xoffset,
; int yoffset,
; unsigned char *dst_ptr,
; int dst_pitch
;)
global sym(vp8_bilinear_predict8x8_sse2)
sym(vp8_bilinear_predict8x8_sse2):
push rbp
mov rbp, rsp
SHADOW_ARGS_TO_STACK 6
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, 144 ; reserve 144 bytes
Move shared data to shared location Storing vp8_bilinear_filters_mmx in an mmx file and using it in an sse2 file is bad Moving towards allowing --disable-mmx Change-Id: I20493b35bdedcdcfc0915e6f05fdbe6c81a4a742
2011-11-18 20:50:13 +01:00
;const short *HFilter = vp8_bilinear_filters_x86_8[xoffset]
;const short *VFilter = vp8_bilinear_filters_x86_8[yoffset]
lea rcx, [GLOBAL(sym(vp8_bilinear_filters_x86_8))]
Initial WebM release
2010-05-18 17:58:33 +02:00
mov rsi, arg(0) ;src_ptr
movsxd rdx, dword ptr arg(1) ;src_pixels_per_line
;Read 9-line unaligned data in and put them on stack. This gives a big
;performance boost.
movdqu xmm0, [rsi]
lea rax, [rdx + rdx*2]
movdqu xmm1, [rsi+rdx]
movdqu xmm2, [rsi+rdx*2]
add rsi, rax
movdqu xmm3, [rsi]
movdqu xmm4, [rsi+rdx]
movdqu xmm5, [rsi+rdx*2]
add rsi, rax
movdqu xmm6, [rsi]
movdqu xmm7, [rsi+rdx]
movdqa XMMWORD PTR [rsp], xmm0
movdqu xmm0, [rsi+rdx*2]
movdqa XMMWORD PTR [rsp+16], xmm1
movdqa XMMWORD PTR [rsp+32], xmm2
movdqa XMMWORD PTR [rsp+48], xmm3
movdqa XMMWORD PTR [rsp+64], xmm4
movdqa XMMWORD PTR [rsp+80], xmm5
movdqa XMMWORD PTR [rsp+96], xmm6
movdqa XMMWORD PTR [rsp+112], xmm7
movdqa XMMWORD PTR [rsp+128], xmm0
movsxd rax, dword ptr arg(2) ;xoffset
shl rax, 5
add rax, rcx ;HFilter
mov rdi, arg(4) ;dst_ptr
movsxd rdx, dword ptr arg(5) ;dst_pitch
movdqa xmm1, [rax]
movdqa xmm2, [rax+16]
movsxd rax, dword ptr arg(3) ;yoffset
shl rax, 5
add rax, rcx ;VFilter
lea rcx, [rdi+rdx*8]
movdqa xmm5, [rax]
movdqa xmm6, [rax+16]
pxor xmm0, xmm0
; get the first horizontal line done
movdqa xmm3, XMMWORD PTR [rsp]
movdqa xmm4, xmm3 ; make a copy of current line
psrldq xmm4, 1
punpcklbw xmm3, xmm0 ; 00 01 02 03 04 05 06 07
punpcklbw xmm4, xmm0 ; 01 02 03 04 05 06 07 08
pmullw xmm3, xmm1
pmullw xmm4, xmm2
paddw xmm3, xmm4
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
paddw xmm3, [GLOBAL(rd)] ; xmm3 += round value
Initial WebM release
2010-05-18 17:58:33 +02:00
psraw xmm3, VP8_FILTER_SHIFT ; xmm3 /= 128
movdqa xmm7, xmm3
add rsp, 16 ; next line
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
.next_row8x8:
Initial WebM release
2010-05-18 17:58:33 +02:00
movdqa xmm3, XMMWORD PTR [rsp] ; 00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15
movdqa xmm4, xmm3 ; make a copy of current line
psrldq xmm4, 1
punpcklbw xmm3, xmm0 ; 00 01 02 03 04 05 06 07
punpcklbw xmm4, xmm0 ; 01 02 03 04 05 06 07 08
pmullw xmm3, xmm1
pmullw xmm4, xmm2
paddw xmm3, xmm4
pmullw xmm7, xmm5
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
paddw xmm3, [GLOBAL(rd)] ; xmm3 += round value
Initial WebM release
2010-05-18 17:58:33 +02:00
psraw xmm3, VP8_FILTER_SHIFT ; xmm3 /= 128
movdqa xmm4, xmm3
pmullw xmm3, xmm6
paddw xmm3, xmm7
movdqa xmm7, xmm4
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
paddw xmm3, [GLOBAL(rd)] ; xmm3 += round value
Initial WebM release
2010-05-18 17:58:33 +02:00
psraw xmm3, VP8_FILTER_SHIFT ; xmm3 /= 128
packuswb xmm3, xmm0
movq [rdi], xmm3 ; store the results in the destination
add rsp, 16 ; next line
add rdi, rdx
cmp rdi, 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 .next_row8x8
Initial WebM release
2010-05-18 17:58:33 +02:00
;add rsp, 144
pop rsp
; begin epilog
pop rdi
pop rsi
RESTORE_GOT
More on "some XMM registers are non-volatile on windows x64 ABI" Add same fix in subpixel_sse2.asm. Change-Id: Icfda6103cbf74ec43308e96961dd738aa823c14d
2010-06-15 15:11:26 +02:00
RESTORE_XMM
Initial WebM release
2010-05-18 17:58:33 +02:00
UNSHADOW_ARGS
pop rbp
ret
SECTION_RODATA
align 16
rd:
times 8 dw 0x40
Reference in New Issue Copy Permalink