Chiebot-Mirror
/
FFmpeg
mirror of https://github.com/FFmpeg/FFmpeg.git
8
0
Fork 0
Code Issues Projects Releases Wiki Activity
You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
43792 Commits
37 Branches
400 Tags
465 MiB
Tag: Branch: Tree: 5801f9ed24
Branches Tags
${ item.name }
Create tag ${ searchTerm }
Create branch ${ searchTerm }
from '5801f9ed24'
${ noResults }
FFmpeg/libavcodec/x86/h264_idct.asm

1105 lines
28 KiB
Raw Normal View History Unescape

Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
;*****************************************************************************
;* MMX/SSE2-optimized H.264 iDCT
;*****************************************************************************
;* Copyright (C) 2004-2005 Michael Niedermayer, Loren Merritt
;* Copyright (C) 2003-2008 x264 project
;*
;* Authors: Laurent Aimar <fenrir@via.ecp.fr>
;* Loren Merritt <lorenm@u.washington.edu>
;* Holger Lubitz <hal@duncan.ol.sub.de>
;* Min Chen <chenm001.163.com>
;*
Replace FFmpeg with Libav in licence headers Signed-off-by: Mans Rullgard <mans@mansr.com>
14 years ago
;* This file is part of Libav.
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
;*
Replace FFmpeg with Libav in licence headers Signed-off-by: Mans Rullgard <mans@mansr.com>
14 years ago
;* Libav is free software; you can redistribute it and/or
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
;* modify it under the terms of the GNU Lesser General Public
;* License as published by the Free Software Foundation; either
;* version 2.1 of the License, or (at your option) any later version.
;*
Replace FFmpeg with Libav in licence headers Signed-off-by: Mans Rullgard <mans@mansr.com>
14 years ago
;* Libav is distributed in the hope that it will be useful,
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
;* but WITHOUT ANY WARRANTY; without even the implied warranty of
;* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
;* Lesser General Public License for more details.
;*
;* You should have received a copy of the GNU Lesser General Public
Replace FFmpeg with Libav in licence headers Signed-off-by: Mans Rullgard <mans@mansr.com>
14 years ago
;* License along with Libav; if not, write to the Free Software
Fix FSF address copy paste error in some license headers.
14 years ago
;* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
;*****************************************************************************
x86: yasm: Use complete source path for macro helper %includes This is more consistent with the way we handle C #includes and it simplifies the build system.
13 years ago
%include "libavutil/x86/x86util.asm"
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
SECTION_RODATA
4:4:4 H.264 decoding support Note: this is 4:4:4 from the 2007 spec revision, not the previous (now deprecated) 4:4:4 mode in H.264.
14 years ago
scan8_mem: db 4+ 1*8, 5+ 1*8, 4+ 2*8, 5+ 2*8
db 6+ 1*8, 7+ 1*8, 6+ 2*8, 7+ 2*8
db 4+ 3*8, 5+ 3*8, 4+ 4*8, 5+ 4*8
db 6+ 3*8, 7+ 3*8, 6+ 4*8, 7+ 4*8
db 4+ 6*8, 5+ 6*8, 4+ 7*8, 5+ 7*8
db 6+ 6*8, 7+ 6*8, 6+ 7*8, 7+ 7*8
db 4+ 8*8, 5+ 8*8, 4+ 9*8, 5+ 9*8
db 6+ 8*8, 7+ 8*8, 6+ 9*8, 7+ 9*8
db 4+11*8, 5+11*8, 4+12*8, 5+12*8
db 6+11*8, 7+11*8, 6+12*8, 7+12*8
db 4+13*8, 5+13*8, 4+14*8, 5+14*8
db 6+13*8, 7+13*8, 6+14*8, 7+14*8
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
%ifdef PIC
x86inc improvements for 64-bit Add support for all x86-64 registers Prefer caller-saved register over callee-saved on WIN64 Support up to 15 function arguments Also (by Ronald S. Bultje) Fix up our asm to work with new x86inc.asm. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com> Signed-off-by: Justin Ruggles <justin.ruggles@gmail.com>
13 years ago
%define npicregs 1
%define scan8 picregq
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
%else
x86inc improvements for 64-bit Add support for all x86-64 registers Prefer caller-saved register over callee-saved on WIN64 Support up to 15 function arguments Also (by Ronald S. Bultje) Fix up our asm to work with new x86inc.asm. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com> Signed-off-by: Justin Ruggles <justin.ruggles@gmail.com>
13 years ago
%define npicregs 0
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
%define scan8 scan8_mem
%endif
cextern pw_32
H.264: split luma dc idct out and implement MMX/SSE2 versions About 2.5x the speed. NOTE: the way that the asm code handles large qmuls is a bit suboptimal. If x264-style dequant was used (separate shift and qmul values), it might be possible to get some extra speed. Originally committed as revision 26336 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
cextern pw_1
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
SECTION .text
; %1=uint8_t *dst, %2=int16_t *block, %3=int stride
%macro IDCT4_ADD 3
; Load dct coeffs
movq m0, [%2]
movq m1, [%2+8]
movq m2, [%2+16]
movq m3, [%2+24]
Modify x86util.asm to ease transitioning to 10-bit H.264 assembly. Arguments for variable size instructions are added to many macros, along with other various changes. The x86util.asm code was ported from x264. Signed-off-by: Diego Biurrun <diego@biurrun.de>
14 years ago
IDCT4_1D w, 0, 1, 2, 3, 4, 5
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
mova m6, [pw_32]
TRANSPOSE4x4W 0, 1, 2, 3, 4
paddw m0, m6
Modify x86util.asm to ease transitioning to 10-bit H.264 assembly. Arguments for variable size instructions are added to many macros, along with other various changes. The x86util.asm code was ported from x264. Signed-off-by: Diego Biurrun <diego@biurrun.de>
14 years ago
IDCT4_1D w, 0, 1, 2, 3, 4, 5
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
pxor m7, m7
h264: Integrate clear_blocks calls with IDCT The non-intra-pcm branch in hl_decode_mb (simple, 8bpp) goes from 700 to 672 cycles, and the complete loop of decode_mb_cabac and hl_decode_mb (in the decode_slice loop) goes from 1759 to 1733 cycles on the clip tested (cathedral), i.e. almost 30 cycles per mb faster. Signed-off-by: Martin Storsjö <martin@martin.st>
12 years ago
movq [%2+ 0], m7
movq [%2+ 8], m7
movq [%2+16], m7
movq [%2+24], m7
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
STORE_DIFFx2 m0, m1, m4, m5, m7, 6, %1, %3
lea %1, [%1+%3*2]
STORE_DIFFx2 m2, m3, m4, m5, m7, 6, %1, %3
%endmacro
x86: h264_idct: port to cpuflags
13 years ago
INIT_MMX mmx
x86: Make function prototype comments in assembly code consistent This helps grepping for functions, among other things.
11 years ago
; void ff_h264_idct_add_8_mmx(uint8_t *dst, int16_t *block, int stride)
x86: h264_idct: port to cpuflags
13 years ago
cglobal h264_idct_add_8, 3, 3, 0
x86: Add missing movsxd for the int stride parameter Signed-off-by: Martin Storsjö <martin@martin.st>
9 years ago
movsxdifnidn r2, r2d
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
IDCT4_ADD r0, r1, r2
RET
%macro IDCT8_1D 2
mova m0, m1
psraw m1, 1
H.264: tweak some other x86 asm for Atom
14 years ago
mova m4, m5
psraw m4, 1
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
paddw m4, m5
paddw m1, m0
paddw m4, m7
paddw m1, m5
psubw m4, m0
paddw m1, m3
psubw m0, m3
psubw m5, m3
H.264: tweak some other x86 asm for Atom
14 years ago
psraw m3, 1
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
paddw m0, m7
psubw m5, m7
psraw m7, 1
psubw m0, m3
psubw m5, m7
mova m7, m1
psraw m1, 2
H.264: tweak some other x86 asm for Atom
14 years ago
mova m3, m4
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
psraw m3, 2
paddw m3, m0
psraw m0, 2
paddw m1, m5
psraw m5, 2
psubw m0, m4
psubw m7, m5
mova m5, m6
psraw m6, 1
H.264: tweak some other x86 asm for Atom
14 years ago
mova m4, m2
psraw m4, 1
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
paddw m6, m2
H.264: tweak some other x86 asm for Atom
14 years ago
psubw m4, m5
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
mova m2, %1
mova m5, %2
Modify x86util.asm to ease transitioning to 10-bit H.264 assembly. Arguments for variable size instructions are added to many macros, along with other various changes. The x86util.asm code was ported from x264. Signed-off-by: Diego Biurrun <diego@biurrun.de>
14 years ago
SUMSUB_BA w, 5, 2
SUMSUB_BA w, 6, 5
SUMSUB_BA w, 4, 2
SUMSUB_BA w, 7, 6
SUMSUB_BA w, 0, 4
SUMSUB_BA w, 3, 2
SUMSUB_BA w, 1, 5
Update 8-bit H.264 IDCT function names to reflect bit-depth. Signed-off-by: Ronald S. Bultje <rbultje@google.com>
14 years ago
SWAP 7, 6, 4, 5, 2, 3, 1, 0 ; 70315246 -> 01234567
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
%endmacro
%macro IDCT8_1D_FULL 1
mova m7, [%1+112]
mova m6, [%1+ 96]
mova m5, [%1+ 80]
mova m3, [%1+ 48]
mova m2, [%1+ 32]
mova m1, [%1+ 16]
IDCT8_1D [%1], [%1+ 64]
%endmacro
; %1=int16_t *block, %2=int16_t *dstblock
%macro IDCT8_ADD_MMX_START 2
IDCT8_1D_FULL %1
mova [%1], m7
TRANSPOSE4x4W 0, 1, 2, 3, 7
mova m7, [%1]
mova [%2 ], m0
mova [%2+16], m1
mova [%2+32], m2
mova [%2+48], m3
TRANSPOSE4x4W 4, 5, 6, 7, 3
mova [%2+ 8], m4
mova [%2+24], m5
mova [%2+40], m6
mova [%2+56], m7
%endmacro
; %1=uint8_t *dst, %2=int16_t *block, %3=int stride
h264: Integrate clear_blocks calls with IDCT The non-intra-pcm branch in hl_decode_mb (simple, 8bpp) goes from 700 to 672 cycles, and the complete loop of decode_mb_cabac and hl_decode_mb (in the decode_slice loop) goes from 1759 to 1733 cycles on the clip tested (cathedral), i.e. almost 30 cycles per mb faster. Signed-off-by: Martin Storsjö <martin@martin.st>
12 years ago
%macro IDCT8_ADD_MMX_END 3-4
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
IDCT8_1D_FULL %2
mova [%2 ], m5
mova [%2+16], m6
mova [%2+32], m7
pxor m7, m7
h264: Integrate clear_blocks calls with IDCT The non-intra-pcm branch in hl_decode_mb (simple, 8bpp) goes from 700 to 672 cycles, and the complete loop of decode_mb_cabac and hl_decode_mb (in the decode_slice loop) goes from 1759 to 1733 cycles on the clip tested (cathedral), i.e. almost 30 cycles per mb faster. Signed-off-by: Martin Storsjö <martin@martin.st>
12 years ago
%if %0 == 4
movq [%4+ 0], m7
movq [%4+ 8], m7
movq [%4+ 16], m7
movq [%4+ 24], m7
movq [%4+ 32], m7
movq [%4+ 40], m7
movq [%4+ 48], m7
movq [%4+ 56], m7
movq [%4+ 64], m7
movq [%4+ 72], m7
movq [%4+ 80], m7
movq [%4+ 88], m7
movq [%4+ 96], m7
movq [%4+104], m7
movq [%4+112], m7
movq [%4+120], m7
%endif
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
STORE_DIFFx2 m0, m1, m5, m6, m7, 6, %1, %3
lea %1, [%1+%3*2]
STORE_DIFFx2 m2, m3, m5, m6, m7, 6, %1, %3
mova m0, [%2 ]
mova m1, [%2+16]
mova m2, [%2+32]
lea %1, [%1+%3*2]
STORE_DIFFx2 m4, m0, m5, m6, m7, 6, %1, %3
lea %1, [%1+%3*2]
STORE_DIFFx2 m1, m2, m5, m6, m7, 6, %1, %3
%endmacro
x86: h264_idct: port to cpuflags
13 years ago
INIT_MMX mmx
x86: Make function prototype comments in assembly code consistent This helps grepping for functions, among other things.
11 years ago
; void ff_h264_idct8_add_8_mmx(uint8_t *dst, int16_t *block, int stride)
x86: h264_idct: port to cpuflags
13 years ago
cglobal h264_idct8_add_8, 3, 4, 0
x86: Add missing movsxd for the int stride parameter Signed-off-by: Martin Storsjö <martin@martin.st>
9 years ago
movsxdifnidn r2, r2d
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
%assign pad 128+4-(stack_offset&7)
SUB rsp, pad
add word [r1], 32
IDCT8_ADD_MMX_START r1 , rsp
IDCT8_ADD_MMX_START r1+8, rsp+64
lea r3, [r0+4]
h264: Integrate clear_blocks calls with IDCT The non-intra-pcm branch in hl_decode_mb (simple, 8bpp) goes from 700 to 672 cycles, and the complete loop of decode_mb_cabac and hl_decode_mb (in the decode_slice loop) goes from 1759 to 1733 cycles on the clip tested (cathedral), i.e. almost 30 cycles per mb faster. Signed-off-by: Martin Storsjö <martin@martin.st>
12 years ago
IDCT8_ADD_MMX_END r0 , rsp, r2, r1
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
IDCT8_ADD_MMX_END r3 , rsp+8, r2
ADD rsp, pad
RET
; %1=uint8_t *dst, %2=int16_t *block, %3=int stride
%macro IDCT8_ADD_SSE 4
IDCT8_1D_FULL %2
config.asm: change %ifdef directives to %if directives. This allows combining multiple conditionals in a single statement.
13 years ago
%if ARCH_X86_64
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
TRANSPOSE8x8W 0, 1, 2, 3, 4, 5, 6, 7, 8
%else
TRANSPOSE8x8W 0, 1, 2, 3, 4, 5, 6, 7, [%2], [%2+16]
%endif
paddw m0, [pw_32]
config.asm: change %ifdef directives to %if directives. This allows combining multiple conditionals in a single statement.
13 years ago
%if ARCH_X86_64 == 0
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
mova [%2 ], m0
mova [%2+16], m4
IDCT8_1D [%2], [%2+ 16]
mova [%2 ], m6
mova [%2+16], m7
%else
SWAP 0, 8
SWAP 4, 9
IDCT8_1D m8, m9
SWAP 6, 8
SWAP 7, 9
%endif
pxor m7, m7
lea %4, [%3*3]
STORE_DIFF m0, m6, m7, [%1 ]
STORE_DIFF m1, m6, m7, [%1+%3 ]
STORE_DIFF m2, m6, m7, [%1+%3*2]
STORE_DIFF m3, m6, m7, [%1+%4 ]
config.asm: change %ifdef directives to %if directives. This allows combining multiple conditionals in a single statement.
13 years ago
%if ARCH_X86_64 == 0
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
mova m0, [%2 ]
mova m1, [%2+16]
%else
SWAP 0, 8
SWAP 1, 9
%endif
h264: Integrate clear_blocks calls with IDCT The non-intra-pcm branch in hl_decode_mb (simple, 8bpp) goes from 700 to 672 cycles, and the complete loop of decode_mb_cabac and hl_decode_mb (in the decode_slice loop) goes from 1759 to 1733 cycles on the clip tested (cathedral), i.e. almost 30 cycles per mb faster. Signed-off-by: Martin Storsjö <martin@martin.st>
12 years ago
mova [%2+ 0], m7
mova [%2+ 16], m7
mova [%2+ 32], m7
mova [%2+ 48], m7
mova [%2+ 64], m7
mova [%2+ 80], m7
mova [%2+ 96], m7
mova [%2+112], m7
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
lea %1, [%1+%3*4]
STORE_DIFF m4, m6, m7, [%1 ]
STORE_DIFF m5, m6, m7, [%1+%3 ]
STORE_DIFF m0, m6, m7, [%1+%3*2]
STORE_DIFF m1, m6, m7, [%1+%4 ]
%endmacro
x86: h264_idct: port to cpuflags
13 years ago
INIT_XMM sse2
x86: Make function prototype comments in assembly code consistent This helps grepping for functions, among other things.
11 years ago
; void ff_h264_idct8_add_8_sse2(uint8_t *dst, int16_t *block, int stride)
x86: h264_idct: port to cpuflags
13 years ago
cglobal h264_idct8_add_8, 3, 4, 10
x86: Add missing movsxd for the int stride parameter Signed-off-by: Martin Storsjö <martin@martin.st>
9 years ago
movsxdifnidn r2, r2d
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
IDCT8_ADD_SSE r0, r1, r2, r3
RET
h264: Integrate clear_blocks calls with IDCT The non-intra-pcm branch in hl_decode_mb (simple, 8bpp) goes from 700 to 672 cycles, and the complete loop of decode_mb_cabac and hl_decode_mb (in the decode_slice loop) goes from 1759 to 1733 cycles on the clip tested (cathedral), i.e. almost 30 cycles per mb faster. Signed-off-by: Martin Storsjö <martin@martin.st>
12 years ago
%macro DC_ADD_MMXEXT_INIT 2
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
add %1, 32
sar %1, 6
Add d suffix to movd target register to make it work with nasm. Originally committed as revision 25206 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
movd m0, %1d
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
lea %1, [%2*3]
pshufw m0, m0, 0
pxor m1, m1
psubw m1, m0
packuswb m0, m0
packuswb m1, m1
%endmacro
x86: MMX2 ---> MMXEXT in macro names
13 years ago
%macro DC_ADD_MMXEXT_OP 4
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
%1 m2, [%2 ]
%1 m3, [%2+%3 ]
%1 m4, [%2+%3*2]
%1 m5, [%2+%4 ]
paddusb m2, m0
paddusb m3, m0
paddusb m4, m0
paddusb m5, m0
psubusb m2, m1
psubusb m3, m1
psubusb m4, m1
psubusb m5, m1
%1 [%2 ], m2
%1 [%2+%3 ], m3
%1 [%2+%3*2], m4
%1 [%2+%4 ], m5
%endmacro
x86: h264_idct: port to cpuflags
13 years ago
INIT_MMX mmxext
x86: Make function prototype comments in assembly code consistent This helps grepping for functions, among other things.
11 years ago
; void ff_h264_idct_dc_add_8_mmxext(uint8_t *dst, int16_t *block, int stride)
h264: Integrate clear_blocks calls with IDCT The non-intra-pcm branch in hl_decode_mb (simple, 8bpp) goes from 700 to 672 cycles, and the complete loop of decode_mb_cabac and hl_decode_mb (in the decode_slice loop) goes from 1759 to 1733 cycles on the clip tested (cathedral), i.e. almost 30 cycles per mb faster. Signed-off-by: Martin Storsjö <martin@martin.st>
12 years ago
%if ARCH_X86_64
cglobal h264_idct_dc_add_8, 3, 4, 0
x86: Add missing movsxd for the int stride parameter Signed-off-by: Martin Storsjö <martin@martin.st>
9 years ago
movsxd r2, r2d
h264: Integrate clear_blocks calls with IDCT The non-intra-pcm branch in hl_decode_mb (simple, 8bpp) goes from 700 to 672 cycles, and the complete loop of decode_mb_cabac and hl_decode_mb (in the decode_slice loop) goes from 1759 to 1733 cycles on the clip tested (cathedral), i.e. almost 30 cycles per mb faster. Signed-off-by: Martin Storsjö <martin@martin.st>
12 years ago
movsx r3, word [r1]
mov dword [r1], 0
DC_ADD_MMXEXT_INIT r3, r2
DC_ADD_MMXEXT_OP movh, r0, r2, r3
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
RET
x86: Make function prototype comments in assembly code consistent This helps grepping for functions, among other things.
11 years ago
; void ff_h264_idct8_dc_add_8_mmxext(uint8_t *dst, int16_t *block, int stride)
h264: Integrate clear_blocks calls with IDCT The non-intra-pcm branch in hl_decode_mb (simple, 8bpp) goes from 700 to 672 cycles, and the complete loop of decode_mb_cabac and hl_decode_mb (in the decode_slice loop) goes from 1759 to 1733 cycles on the clip tested (cathedral), i.e. almost 30 cycles per mb faster. Signed-off-by: Martin Storsjö <martin@martin.st>
12 years ago
cglobal h264_idct8_dc_add_8, 3, 4, 0
x86: Add missing movsxd for the int stride parameter Signed-off-by: Martin Storsjö <martin@martin.st>
9 years ago
movsxd r2, r2d
h264: Integrate clear_blocks calls with IDCT The non-intra-pcm branch in hl_decode_mb (simple, 8bpp) goes from 700 to 672 cycles, and the complete loop of decode_mb_cabac and hl_decode_mb (in the decode_slice loop) goes from 1759 to 1733 cycles on the clip tested (cathedral), i.e. almost 30 cycles per mb faster. Signed-off-by: Martin Storsjö <martin@martin.st>
12 years ago
movsx r3, word [r1]
mov dword [r1], 0
DC_ADD_MMXEXT_INIT r3, r2
DC_ADD_MMXEXT_OP mova, r0, r2, r3
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
lea r0, [r0+r2*4]
h264: Integrate clear_blocks calls with IDCT The non-intra-pcm branch in hl_decode_mb (simple, 8bpp) goes from 700 to 672 cycles, and the complete loop of decode_mb_cabac and hl_decode_mb (in the decode_slice loop) goes from 1759 to 1733 cycles on the clip tested (cathedral), i.e. almost 30 cycles per mb faster. Signed-off-by: Martin Storsjö <martin@martin.st>
12 years ago
DC_ADD_MMXEXT_OP mova, r0, r2, r3
RET
%else
x86: Make function prototype comments in assembly code consistent This helps grepping for functions, among other things.
11 years ago
; void ff_h264_idct_dc_add_8_mmxext(uint8_t *dst, int16_t *block, int stride)
h264: Integrate clear_blocks calls with IDCT The non-intra-pcm branch in hl_decode_mb (simple, 8bpp) goes from 700 to 672 cycles, and the complete loop of decode_mb_cabac and hl_decode_mb (in the decode_slice loop) goes from 1759 to 1733 cycles on the clip tested (cathedral), i.e. almost 30 cycles per mb faster. Signed-off-by: Martin Storsjö <martin@martin.st>
12 years ago
cglobal h264_idct_dc_add_8, 2, 3, 0
movsx r2, word [r1]
mov dword [r1], 0
mov r1, r2m
DC_ADD_MMXEXT_INIT r2, r1
DC_ADD_MMXEXT_OP movh, r0, r1, r2
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
RET
x86: Make function prototype comments in assembly code consistent This helps grepping for functions, among other things.
11 years ago
; void ff_h264_idct8_dc_add_8_mmxext(uint8_t *dst, int16_t *block, int stride)
h264: Integrate clear_blocks calls with IDCT The non-intra-pcm branch in hl_decode_mb (simple, 8bpp) goes from 700 to 672 cycles, and the complete loop of decode_mb_cabac and hl_decode_mb (in the decode_slice loop) goes from 1759 to 1733 cycles on the clip tested (cathedral), i.e. almost 30 cycles per mb faster. Signed-off-by: Martin Storsjö <martin@martin.st>
12 years ago
cglobal h264_idct8_dc_add_8, 2, 3, 0
movsx r2, word [r1]
mov dword [r1], 0
mov r1, r2m
DC_ADD_MMXEXT_INIT r2, r1
DC_ADD_MMXEXT_OP mova, r0, r1, r2
lea r0, [r0+r1*4]
DC_ADD_MMXEXT_OP mova, r0, r1, r2
RET
%endif
x86: h264_idct: port to cpuflags
13 years ago
INIT_MMX mmx
x86: Make function prototype comments in assembly code consistent This helps grepping for functions, among other things.
11 years ago
; void ff_h264_idct_add16_8_mmx(uint8_t *dst, const int *block_offset,
; int16_t *block, int stride,
; const uint8_t nnzc[6 * 8])
x86: h264_idct: port to cpuflags
13 years ago
cglobal h264_idct_add16_8, 5, 7 + npicregs, 0, dst, block_offset, block, stride, nnzc, cntr, coeff, picreg
x86: Add missing movsxd for the int stride parameter Signed-off-by: Martin Storsjö <martin@martin.st>
9 years ago
movsxdifnidn r3, r3d
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
xor r5, r5
%ifdef PIC
x86inc improvements for 64-bit Add support for all x86-64 registers Prefer caller-saved register over callee-saved on WIN64 Support up to 15 function arguments Also (by Ronald S. Bultje) Fix up our asm to work with new x86inc.asm. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com> Signed-off-by: Justin Ruggles <justin.ruggles@gmail.com>
13 years ago
lea picregq, [scan8_mem]
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
%endif
x86: add colons after labels nasm prints a warning if the colon is missing. Signed-off-by: Mans Rullgard <mans@mansr.com>
12 years ago
.nextblock:
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
movzx r6, byte [scan8+r5]
movzx r6, byte [r4+r6]
test r6, r6
jz .skipblock
mov r6d, dword [r1+r5*4]
lea r6, [r0+r6]
IDCT4_ADD r6, r2, r3
x86: add colons after labels nasm prints a warning if the colon is missing. Signed-off-by: Mans Rullgard <mans@mansr.com>
12 years ago
.skipblock:
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
inc r5
add r2, 32
cmp r5, 16
jl .nextblock
REP_RET
x86: Make function prototype comments in assembly code consistent This helps grepping for functions, among other things.
11 years ago
; void ff_h264_idct8_add4_8_mmx(uint8_t *dst, const int *block_offset,
; int16_t *block, int stride,
; const uint8_t nnzc[6 * 8])
x86: h264_idct: port to cpuflags
13 years ago
cglobal h264_idct8_add4_8, 5, 7 + npicregs, 0, dst, block_offset, block, stride, nnzc, cntr, coeff, picreg
x86: Add missing movsxd for the int stride parameter Signed-off-by: Martin Storsjö <martin@martin.st>
9 years ago
movsxdifnidn r3, r3d
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
%assign pad 128+4-(stack_offset&7)
SUB rsp, pad
xor r5, r5
%ifdef PIC
x86inc improvements for 64-bit Add support for all x86-64 registers Prefer caller-saved register over callee-saved on WIN64 Support up to 15 function arguments Also (by Ronald S. Bultje) Fix up our asm to work with new x86inc.asm. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com> Signed-off-by: Justin Ruggles <justin.ruggles@gmail.com>
13 years ago
lea picregq, [scan8_mem]
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
%endif
x86: add colons after labels nasm prints a warning if the colon is missing. Signed-off-by: Mans Rullgard <mans@mansr.com>
12 years ago
.nextblock:
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
movzx r6, byte [scan8+r5]
movzx r6, byte [r4+r6]
test r6, r6
jz .skipblock
mov r6d, dword [r1+r5*4]
H.264: tweak some other x86 asm for Atom
14 years ago
add r6, r0
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
add word [r2], 32
IDCT8_ADD_MMX_START r2 , rsp
IDCT8_ADD_MMX_START r2+8, rsp+64
h264: Integrate clear_blocks calls with IDCT The non-intra-pcm branch in hl_decode_mb (simple, 8bpp) goes from 700 to 672 cycles, and the complete loop of decode_mb_cabac and hl_decode_mb (in the decode_slice loop) goes from 1759 to 1733 cycles on the clip tested (cathedral), i.e. almost 30 cycles per mb faster. Signed-off-by: Martin Storsjö <martin@martin.st>
12 years ago
IDCT8_ADD_MMX_END r6 , rsp, r3, r2
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
mov r6d, dword [r1+r5*4]
lea r6, [r0+r6+4]
IDCT8_ADD_MMX_END r6 , rsp+8, r3
x86: add colons after labels nasm prints a warning if the colon is missing. Signed-off-by: Mans Rullgard <mans@mansr.com>
12 years ago
.skipblock:
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
add r5, 4
add r2, 128
cmp r5, 16
jl .nextblock
ADD rsp, pad
RET
x86: h264_idct: port to cpuflags
13 years ago
INIT_MMX mmxext
x86: Make function prototype comments in assembly code consistent This helps grepping for functions, among other things.
11 years ago
; void ff_h264_idct_add16_8_mmxext(uint8_t *dst, const int *block_offset,
; int16_t *block, int stride,
; const uint8_t nnzc[6 * 8])
x86: h264_idct: port to cpuflags
13 years ago
cglobal h264_idct_add16_8, 5, 8 + npicregs, 0, dst1, block_offset, block, stride, nnzc, cntr, coeff, dst2, picreg
x86: Add missing movsxd for the int stride parameter Signed-off-by: Martin Storsjö <martin@martin.st>
9 years ago
movsxdifnidn r3, r3d
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
xor r5, r5
%ifdef PIC
x86inc improvements for 64-bit Add support for all x86-64 registers Prefer caller-saved register over callee-saved on WIN64 Support up to 15 function arguments Also (by Ronald S. Bultje) Fix up our asm to work with new x86inc.asm. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com> Signed-off-by: Justin Ruggles <justin.ruggles@gmail.com>
13 years ago
lea picregq, [scan8_mem]
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
%endif
x86: add colons after labels nasm prints a warning if the colon is missing. Signed-off-by: Mans Rullgard <mans@mansr.com>
12 years ago
.nextblock:
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
movzx r6, byte [scan8+r5]
movzx r6, byte [r4+r6]
test r6, r6
jz .skipblock
cmp r6, 1
jnz .no_dc
movsx r6, word [r2]
test r6, r6
jz .no_dc
h264: Integrate clear_blocks calls with IDCT The non-intra-pcm branch in hl_decode_mb (simple, 8bpp) goes from 700 to 672 cycles, and the complete loop of decode_mb_cabac and hl_decode_mb (in the decode_slice loop) goes from 1759 to 1733 cycles on the clip tested (cathedral), i.e. almost 30 cycles per mb faster. Signed-off-by: Martin Storsjö <martin@martin.st>
12 years ago
mov word [r2], 0
DC_ADD_MMXEXT_INIT r6, r3
x86inc improvements for 64-bit Add support for all x86-64 registers Prefer caller-saved register over callee-saved on WIN64 Support up to 15 function arguments Also (by Ronald S. Bultje) Fix up our asm to work with new x86inc.asm. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com> Signed-off-by: Justin Ruggles <justin.ruggles@gmail.com>
13 years ago
%if ARCH_X86_64 == 0
%define dst2q r1
%define dst2d r1d
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
%endif
x86inc improvements for 64-bit Add support for all x86-64 registers Prefer caller-saved register over callee-saved on WIN64 Support up to 15 function arguments Also (by Ronald S. Bultje) Fix up our asm to work with new x86inc.asm. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com> Signed-off-by: Justin Ruggles <justin.ruggles@gmail.com>
13 years ago
mov dst2d, dword [r1+r5*4]
lea dst2q, [r0+dst2q]
x86: MMX2 ---> MMXEXT in macro names
13 years ago
DC_ADD_MMXEXT_OP movh, dst2q, r3, r6
config.asm: change %ifdef directives to %if directives. This allows combining multiple conditionals in a single statement.
13 years ago
%if ARCH_X86_64 == 0
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
mov r1, r1m
%endif
inc r5
add r2, 32
cmp r5, 16
jl .nextblock
REP_RET
x86: add colons after labels nasm prints a warning if the colon is missing. Signed-off-by: Mans Rullgard <mans@mansr.com>
12 years ago
.no_dc:
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
mov r6d, dword [r1+r5*4]
H.264: tweak some other x86 asm for Atom
14 years ago
add r6, r0
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
IDCT4_ADD r6, r2, r3
x86: add colons after labels nasm prints a warning if the colon is missing. Signed-off-by: Mans Rullgard <mans@mansr.com>
12 years ago
.skipblock:
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
inc r5
add r2, 32
cmp r5, 16
jl .nextblock
REP_RET
x86: h264_idct: port to cpuflags
13 years ago
INIT_MMX mmx
x86: Make function prototype comments in assembly code consistent This helps grepping for functions, among other things.
11 years ago
; void ff_h264_idct_add16intra_8_mmx(uint8_t *dst, const int *block_offset,
; int16_t *block, int stride,
; const uint8_t nnzc[6 * 8])
x86: h264_idct: port to cpuflags
13 years ago
cglobal h264_idct_add16intra_8, 5, 7 + npicregs, 0, dst, block_offset, block, stride, nnzc, cntr, coeff, picreg
x86: Add missing movsxd for the int stride parameter Signed-off-by: Martin Storsjö <martin@martin.st>
9 years ago
movsxdifnidn r3, r3d
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
xor r5, r5
%ifdef PIC
x86inc improvements for 64-bit Add support for all x86-64 registers Prefer caller-saved register over callee-saved on WIN64 Support up to 15 function arguments Also (by Ronald S. Bultje) Fix up our asm to work with new x86inc.asm. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com> Signed-off-by: Justin Ruggles <justin.ruggles@gmail.com>
13 years ago
lea picregq, [scan8_mem]
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
%endif
x86: add colons after labels nasm prints a warning if the colon is missing. Signed-off-by: Mans Rullgard <mans@mansr.com>
12 years ago
.nextblock:
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
movzx r6, byte [scan8+r5]
movzx r6, byte [r4+r6]
or r6w, word [r2]
test r6, r6
jz .skipblock
mov r6d, dword [r1+r5*4]
H.264: tweak some other x86 asm for Atom
14 years ago
add r6, r0
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
IDCT4_ADD r6, r2, r3
x86: add colons after labels nasm prints a warning if the colon is missing. Signed-off-by: Mans Rullgard <mans@mansr.com>
12 years ago
.skipblock:
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
inc r5
add r2, 32
cmp r5, 16
jl .nextblock
REP_RET
x86: h264_idct: port to cpuflags
13 years ago
INIT_MMX mmxext
x86: Make function prototype comments in assembly code consistent This helps grepping for functions, among other things.
11 years ago
; void ff_h264_idct_add16intra_8_mmxext(uint8_t *dst, const int *block_offset,
; int16_t *block, int stride,
; const uint8_t nnzc[6 * 8])
x86: h264_idct: port to cpuflags
13 years ago
cglobal h264_idct_add16intra_8, 5, 8 + npicregs, 0, dst1, block_offset, block, stride, nnzc, cntr, coeff, dst2, picreg
x86: Add missing movsxd for the int stride parameter Signed-off-by: Martin Storsjö <martin@martin.st>
9 years ago
movsxdifnidn r3, r3d
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
xor r5, r5
%ifdef PIC
x86inc improvements for 64-bit Add support for all x86-64 registers Prefer caller-saved register over callee-saved on WIN64 Support up to 15 function arguments Also (by Ronald S. Bultje) Fix up our asm to work with new x86inc.asm. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com> Signed-off-by: Justin Ruggles <justin.ruggles@gmail.com>
13 years ago
lea picregq, [scan8_mem]
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
%endif
x86: add colons after labels nasm prints a warning if the colon is missing. Signed-off-by: Mans Rullgard <mans@mansr.com>
12 years ago
.nextblock:
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
movzx r6, byte [scan8+r5]
movzx r6, byte [r4+r6]
test r6, r6
jz .try_dc
mov r6d, dword [r1+r5*4]
lea r6, [r0+r6]
IDCT4_ADD r6, r2, r3
inc r5
add r2, 32
cmp r5, 16
jl .nextblock
REP_RET
x86: add colons after labels nasm prints a warning if the colon is missing. Signed-off-by: Mans Rullgard <mans@mansr.com>
12 years ago
.try_dc:
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
movsx r6, word [r2]
test r6, r6
jz .skipblock
h264: Integrate clear_blocks calls with IDCT The non-intra-pcm branch in hl_decode_mb (simple, 8bpp) goes from 700 to 672 cycles, and the complete loop of decode_mb_cabac and hl_decode_mb (in the decode_slice loop) goes from 1759 to 1733 cycles on the clip tested (cathedral), i.e. almost 30 cycles per mb faster. Signed-off-by: Martin Storsjö <martin@martin.st>
12 years ago
mov word [r2], 0
DC_ADD_MMXEXT_INIT r6, r3
x86inc improvements for 64-bit Add support for all x86-64 registers Prefer caller-saved register over callee-saved on WIN64 Support up to 15 function arguments Also (by Ronald S. Bultje) Fix up our asm to work with new x86inc.asm. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com> Signed-off-by: Justin Ruggles <justin.ruggles@gmail.com>
13 years ago
%if ARCH_X86_64 == 0
%define dst2q r1
%define dst2d r1d
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
%endif
x86inc improvements for 64-bit Add support for all x86-64 registers Prefer caller-saved register over callee-saved on WIN64 Support up to 15 function arguments Also (by Ronald S. Bultje) Fix up our asm to work with new x86inc.asm. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com> Signed-off-by: Justin Ruggles <justin.ruggles@gmail.com>
13 years ago
mov dst2d, dword [r1+r5*4]
add dst2q, r0
x86: MMX2 ---> MMXEXT in macro names
13 years ago
DC_ADD_MMXEXT_OP movh, dst2q, r3, r6
config.asm: change %ifdef directives to %if directives. This allows combining multiple conditionals in a single statement.
13 years ago
%if ARCH_X86_64 == 0
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
mov r1, r1m
%endif
x86: add colons after labels nasm prints a warning if the colon is missing. Signed-off-by: Mans Rullgard <mans@mansr.com>
12 years ago
.skipblock:
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
inc r5
add r2, 32
cmp r5, 16
jl .nextblock
REP_RET
x86: Make function prototype comments in assembly code consistent This helps grepping for functions, among other things.
11 years ago
; void ff_h264_idct8_add4_8_mmxext(uint8_t *dst, const int *block_offset,
; int16_t *block, int stride,
; const uint8_t nnzc[6 * 8])
x86: h264_idct: port to cpuflags
13 years ago
cglobal h264_idct8_add4_8, 5, 8 + npicregs, 0, dst1, block_offset, block, stride, nnzc, cntr, coeff, dst2, picreg
x86: Add missing movsxd for the int stride parameter Signed-off-by: Martin Storsjö <martin@martin.st>
9 years ago
movsxdifnidn r3, r3d
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
%assign pad 128+4-(stack_offset&7)
SUB rsp, pad
xor r5, r5
%ifdef PIC
x86inc improvements for 64-bit Add support for all x86-64 registers Prefer caller-saved register over callee-saved on WIN64 Support up to 15 function arguments Also (by Ronald S. Bultje) Fix up our asm to work with new x86inc.asm. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com> Signed-off-by: Justin Ruggles <justin.ruggles@gmail.com>
13 years ago
lea picregq, [scan8_mem]
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
%endif
x86: add colons after labels nasm prints a warning if the colon is missing. Signed-off-by: Mans Rullgard <mans@mansr.com>
12 years ago
.nextblock:
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
movzx r6, byte [scan8+r5]
movzx r6, byte [r4+r6]
test r6, r6
jz .skipblock
cmp r6, 1
jnz .no_dc
movsx r6, word [r2]
test r6, r6
jz .no_dc
h264: Integrate clear_blocks calls with IDCT The non-intra-pcm branch in hl_decode_mb (simple, 8bpp) goes from 700 to 672 cycles, and the complete loop of decode_mb_cabac and hl_decode_mb (in the decode_slice loop) goes from 1759 to 1733 cycles on the clip tested (cathedral), i.e. almost 30 cycles per mb faster. Signed-off-by: Martin Storsjö <martin@martin.st>
12 years ago
mov word [r2], 0
DC_ADD_MMXEXT_INIT r6, r3
x86inc improvements for 64-bit Add support for all x86-64 registers Prefer caller-saved register over callee-saved on WIN64 Support up to 15 function arguments Also (by Ronald S. Bultje) Fix up our asm to work with new x86inc.asm. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com> Signed-off-by: Justin Ruggles <justin.ruggles@gmail.com>
13 years ago
%if ARCH_X86_64 == 0
%define dst2q r1
%define dst2d r1d
%endif
mov dst2d, dword [r1+r5*4]
lea dst2q, [r0+dst2q]
x86: MMX2 ---> MMXEXT in macro names
13 years ago
DC_ADD_MMXEXT_OP mova, dst2q, r3, r6
x86inc improvements for 64-bit Add support for all x86-64 registers Prefer caller-saved register over callee-saved on WIN64 Support up to 15 function arguments Also (by Ronald S. Bultje) Fix up our asm to work with new x86inc.asm. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com> Signed-off-by: Justin Ruggles <justin.ruggles@gmail.com>
13 years ago
lea dst2q, [dst2q+r3*4]
x86: MMX2 ---> MMXEXT in macro names
13 years ago
DC_ADD_MMXEXT_OP mova, dst2q, r3, r6
config.asm: change %ifdef directives to %if directives. This allows combining multiple conditionals in a single statement.
13 years ago
%if ARCH_X86_64 == 0
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
mov r1, r1m
%endif
add r5, 4
add r2, 128
cmp r5, 16
jl .nextblock
ADD rsp, pad
RET
x86: add colons after labels nasm prints a warning if the colon is missing. Signed-off-by: Mans Rullgard <mans@mansr.com>
12 years ago
.no_dc:
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
mov r6d, dword [r1+r5*4]
H.264: tweak some other x86 asm for Atom
14 years ago
add r6, r0
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
add word [r2], 32
IDCT8_ADD_MMX_START r2 , rsp
IDCT8_ADD_MMX_START r2+8, rsp+64
h264: Integrate clear_blocks calls with IDCT The non-intra-pcm branch in hl_decode_mb (simple, 8bpp) goes from 700 to 672 cycles, and the complete loop of decode_mb_cabac and hl_decode_mb (in the decode_slice loop) goes from 1759 to 1733 cycles on the clip tested (cathedral), i.e. almost 30 cycles per mb faster. Signed-off-by: Martin Storsjö <martin@martin.st>
12 years ago
IDCT8_ADD_MMX_END r6 , rsp, r3, r2
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
mov r6d, dword [r1+r5*4]
lea r6, [r0+r6+4]
IDCT8_ADD_MMX_END r6 , rsp+8, r3
x86: add colons after labels nasm prints a warning if the colon is missing. Signed-off-by: Mans Rullgard <mans@mansr.com>
12 years ago
.skipblock:
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
add r5, 4
add r2, 128
cmp r5, 16
jl .nextblock
ADD rsp, pad
RET
x86: h264_idct: port to cpuflags
13 years ago
INIT_XMM sse2
x86: Make function prototype comments in assembly code consistent This helps grepping for functions, among other things.
11 years ago
; void ff_h264_idct8_add4_8_sse2(uint8_t *dst, const int *block_offset,
; int16_t *block, int stride,
; const uint8_t nnzc[6 * 8])
x86: h264_idct: port to cpuflags
13 years ago
cglobal h264_idct8_add4_8, 5, 8 + npicregs, 10, dst1, block_offset, block, stride, nnzc, cntr, coeff, dst2, picreg
x86: Add missing movsxd for the int stride parameter Signed-off-by: Martin Storsjö <martin@martin.st>
9 years ago
movsxdifnidn r3, r3d
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
xor r5, r5
%ifdef PIC
x86inc improvements for 64-bit Add support for all x86-64 registers Prefer caller-saved register over callee-saved on WIN64 Support up to 15 function arguments Also (by Ronald S. Bultje) Fix up our asm to work with new x86inc.asm. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com> Signed-off-by: Justin Ruggles <justin.ruggles@gmail.com>
13 years ago
lea picregq, [scan8_mem]
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
%endif
x86: add colons after labels nasm prints a warning if the colon is missing. Signed-off-by: Mans Rullgard <mans@mansr.com>
12 years ago
.nextblock:
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
movzx r6, byte [scan8+r5]
movzx r6, byte [r4+r6]
test r6, r6
jz .skipblock
cmp r6, 1
jnz .no_dc
movsx r6, word [r2]
test r6, r6
jz .no_dc
x86: h264_idct: port to cpuflags
13 years ago
INIT_MMX cpuname
h264: Integrate clear_blocks calls with IDCT The non-intra-pcm branch in hl_decode_mb (simple, 8bpp) goes from 700 to 672 cycles, and the complete loop of decode_mb_cabac and hl_decode_mb (in the decode_slice loop) goes from 1759 to 1733 cycles on the clip tested (cathedral), i.e. almost 30 cycles per mb faster. Signed-off-by: Martin Storsjö <martin@martin.st>
12 years ago
mov word [r2], 0
DC_ADD_MMXEXT_INIT r6, r3
x86inc improvements for 64-bit Add support for all x86-64 registers Prefer caller-saved register over callee-saved on WIN64 Support up to 15 function arguments Also (by Ronald S. Bultje) Fix up our asm to work with new x86inc.asm. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com> Signed-off-by: Justin Ruggles <justin.ruggles@gmail.com>
13 years ago
%if ARCH_X86_64 == 0
%define dst2q r1
%define dst2d r1d
%endif
mov dst2d, dword [r1+r5*4]
add dst2q, r0
x86: MMX2 ---> MMXEXT in macro names
13 years ago
DC_ADD_MMXEXT_OP mova, dst2q, r3, r6
x86inc improvements for 64-bit Add support for all x86-64 registers Prefer caller-saved register over callee-saved on WIN64 Support up to 15 function arguments Also (by Ronald S. Bultje) Fix up our asm to work with new x86inc.asm. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com> Signed-off-by: Justin Ruggles <justin.ruggles@gmail.com>
13 years ago
lea dst2q, [dst2q+r3*4]
x86: MMX2 ---> MMXEXT in macro names
13 years ago
DC_ADD_MMXEXT_OP mova, dst2q, r3, r6
config.asm: change %ifdef directives to %if directives. This allows combining multiple conditionals in a single statement.
13 years ago
%if ARCH_X86_64 == 0
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
mov r1, r1m
%endif
add r5, 4
add r2, 128
cmp r5, 16
jl .nextblock
REP_RET
x86: add colons after labels nasm prints a warning if the colon is missing. Signed-off-by: Mans Rullgard <mans@mansr.com>
12 years ago
.no_dc:
x86: h264_idct: port to cpuflags
13 years ago
INIT_XMM cpuname
x86inc improvements for 64-bit Add support for all x86-64 registers Prefer caller-saved register over callee-saved on WIN64 Support up to 15 function arguments Also (by Ronald S. Bultje) Fix up our asm to work with new x86inc.asm. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com> Signed-off-by: Justin Ruggles <justin.ruggles@gmail.com>
13 years ago
mov dst2d, dword [r1+r5*4]
add dst2q, r0
IDCT8_ADD_SSE dst2q, r2, r3, r6
config.asm: change %ifdef directives to %if directives. This allows combining multiple conditionals in a single statement.
13 years ago
%if ARCH_X86_64 == 0
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
mov r1, r1m
%endif
x86: add colons after labels nasm prints a warning if the colon is missing. Signed-off-by: Mans Rullgard <mans@mansr.com>
12 years ago
.skipblock:
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
add r5, 4
add r2, 128
cmp r5, 16
jl .nextblock
REP_RET
x86: h264_idct: port to cpuflags
13 years ago
INIT_MMX mmx
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
h264_idct_add8_mmx_plane:
x86: Add missing movsxd for the int stride parameter Signed-off-by: Martin Storsjö <martin@martin.st>
9 years ago
movsxdifnidn r3, r3d
x86: add colons after labels nasm prints a warning if the colon is missing. Signed-off-by: Mans Rullgard <mans@mansr.com>
12 years ago
.nextblock:
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
movzx r6, byte [scan8+r5]
movzx r6, byte [r4+r6]
or r6w, word [r2]
test r6, r6
jz .skipblock
config.asm: change %ifdef directives to %if directives. This allows combining multiple conditionals in a single statement.
13 years ago
%if ARCH_X86_64
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
mov r0d, dword [r1+r5*4]
x86inc improvements for 64-bit Add support for all x86-64 registers Prefer caller-saved register over callee-saved on WIN64 Support up to 15 function arguments Also (by Ronald S. Bultje) Fix up our asm to work with new x86inc.asm. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com> Signed-off-by: Justin Ruggles <justin.ruggles@gmail.com>
13 years ago
add r0, [dst2q]
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
%else
mov r0, r1m ; XXX r1m here is actually r0m of the calling func
mov r0, [r0]
add r0, dword [r1+r5*4]
%endif
IDCT4_ADD r0, r2, r3
x86: add colons after labels nasm prints a warning if the colon is missing. Signed-off-by: Mans Rullgard <mans@mansr.com>
12 years ago
.skipblock:
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
inc r5
add r2, 32
test r5, 3
jnz .nextblock
rep ret
x86: Make function prototype comments in assembly code consistent This helps grepping for functions, among other things.
11 years ago
; void ff_h264_idct_add8_8_mmx(uint8_t **dest, const int *block_offset,
; int16_t *block, int stride,
; const uint8_t nnzc[6 * 8])
x86: h264_idct: port to cpuflags
13 years ago
cglobal h264_idct_add8_8, 5, 8 + npicregs, 0, dst1, block_offset, block, stride, nnzc, cntr, coeff, dst2, picreg
x86: Add missing movsxd for the int stride parameter Signed-off-by: Martin Storsjö <martin@martin.st>
9 years ago
movsxdifnidn r3, r3d
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
mov r5, 16
add r2, 512
%ifdef PIC
x86inc improvements for 64-bit Add support for all x86-64 registers Prefer caller-saved register over callee-saved on WIN64 Support up to 15 function arguments Also (by Ronald S. Bultje) Fix up our asm to work with new x86inc.asm. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com> Signed-off-by: Justin Ruggles <justin.ruggles@gmail.com>
13 years ago
lea picregq, [scan8_mem]
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
%endif
config.asm: change %ifdef directives to %if directives. This allows combining multiple conditionals in a single statement.
13 years ago
%if ARCH_X86_64
x86inc improvements for 64-bit Add support for all x86-64 registers Prefer caller-saved register over callee-saved on WIN64 Support up to 15 function arguments Also (by Ronald S. Bultje) Fix up our asm to work with new x86inc.asm. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com> Signed-off-by: Justin Ruggles <justin.ruggles@gmail.com>
13 years ago
mov dst2q, r0
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
%endif
call h264_idct_add8_mmx_plane
4:4:4 H.264 decoding support Note: this is 4:4:4 from the 2007 spec revision, not the previous (now deprecated) 4:4:4 mode in H.264.
14 years ago
mov r5, 32
add r2, 384
config.asm: change %ifdef directives to %if directives. This allows combining multiple conditionals in a single statement.
13 years ago
%if ARCH_X86_64
x86inc improvements for 64-bit Add support for all x86-64 registers Prefer caller-saved register over callee-saved on WIN64 Support up to 15 function arguments Also (by Ronald S. Bultje) Fix up our asm to work with new x86inc.asm. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com> Signed-off-by: Justin Ruggles <justin.ruggles@gmail.com>
13 years ago
add dst2q, gprsize
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
%else
add r0mp, gprsize
%endif
call h264_idct_add8_mmx_plane
RET
x86: mmx2 ---> mmxext in asm constructs
13 years ago
h264_idct_add8_mmxext_plane:
x86: Add missing movsxd for the int stride parameter Signed-off-by: Martin Storsjö <martin@martin.st>
9 years ago
movsxdifnidn r3, r3d
x86: add colons after labels nasm prints a warning if the colon is missing. Signed-off-by: Mans Rullgard <mans@mansr.com>
12 years ago
.nextblock:
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
movzx r6, byte [scan8+r5]
movzx r6, byte [r4+r6]
test r6, r6
jz .try_dc
config.asm: change %ifdef directives to %if directives. This allows combining multiple conditionals in a single statement.
13 years ago
%if ARCH_X86_64
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
mov r0d, dword [r1+r5*4]
x86inc improvements for 64-bit Add support for all x86-64 registers Prefer caller-saved register over callee-saved on WIN64 Support up to 15 function arguments Also (by Ronald S. Bultje) Fix up our asm to work with new x86inc.asm. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com> Signed-off-by: Justin Ruggles <justin.ruggles@gmail.com>
13 years ago
add r0, [dst2q]
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
%else
mov r0, r1m ; XXX r1m here is actually r0m of the calling func
mov r0, [r0]
add r0, dword [r1+r5*4]
%endif
IDCT4_ADD r0, r2, r3
inc r5
add r2, 32
test r5, 3
jnz .nextblock
rep ret
x86: add colons after labels nasm prints a warning if the colon is missing. Signed-off-by: Mans Rullgard <mans@mansr.com>
12 years ago
.try_dc:
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
movsx r6, word [r2]
test r6, r6
jz .skipblock
h264: Integrate clear_blocks calls with IDCT The non-intra-pcm branch in hl_decode_mb (simple, 8bpp) goes from 700 to 672 cycles, and the complete loop of decode_mb_cabac and hl_decode_mb (in the decode_slice loop) goes from 1759 to 1733 cycles on the clip tested (cathedral), i.e. almost 30 cycles per mb faster. Signed-off-by: Martin Storsjö <martin@martin.st>
12 years ago
mov word [r2], 0
DC_ADD_MMXEXT_INIT r6, r3
config.asm: change %ifdef directives to %if directives. This allows combining multiple conditionals in a single statement.
13 years ago
%if ARCH_X86_64
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
mov r0d, dword [r1+r5*4]
x86inc improvements for 64-bit Add support for all x86-64 registers Prefer caller-saved register over callee-saved on WIN64 Support up to 15 function arguments Also (by Ronald S. Bultje) Fix up our asm to work with new x86inc.asm. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com> Signed-off-by: Justin Ruggles <justin.ruggles@gmail.com>
13 years ago
add r0, [dst2q]
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
%else
mov r0, r1m ; XXX r1m here is actually r0m of the calling func
mov r0, [r0]
add r0, dword [r1+r5*4]
%endif
x86: MMX2 ---> MMXEXT in macro names
13 years ago
DC_ADD_MMXEXT_OP movh, r0, r3, r6
x86: add colons after labels nasm prints a warning if the colon is missing. Signed-off-by: Mans Rullgard <mans@mansr.com>
12 years ago
.skipblock:
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
inc r5
add r2, 32
test r5, 3
jnz .nextblock
rep ret
x86: h264_idct: port to cpuflags
13 years ago
INIT_MMX mmxext
x86: Make function prototype comments in assembly code consistent This helps grepping for functions, among other things.
11 years ago
; void ff_h264_idct_add8_8_mmxext(uint8_t **dest, const int *block_offset,
; int16_t *block, int stride,
; const uint8_t nnzc[6 * 8])
x86: h264_idct: port to cpuflags
13 years ago
cglobal h264_idct_add8_8, 5, 8 + npicregs, 0, dst1, block_offset, block, stride, nnzc, cntr, coeff, dst2, picreg
x86: Add missing movsxd for the int stride parameter Signed-off-by: Martin Storsjö <martin@martin.st>
9 years ago
movsxdifnidn r3, r3d
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
mov r5, 16
add r2, 512
config.asm: change %ifdef directives to %if directives. This allows combining multiple conditionals in a single statement.
13 years ago
%if ARCH_X86_64
x86inc improvements for 64-bit Add support for all x86-64 registers Prefer caller-saved register over callee-saved on WIN64 Support up to 15 function arguments Also (by Ronald S. Bultje) Fix up our asm to work with new x86inc.asm. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com> Signed-off-by: Justin Ruggles <justin.ruggles@gmail.com>
13 years ago
mov dst2q, r0
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
%endif
%ifdef PIC
x86inc improvements for 64-bit Add support for all x86-64 registers Prefer caller-saved register over callee-saved on WIN64 Support up to 15 function arguments Also (by Ronald S. Bultje) Fix up our asm to work with new x86inc.asm. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com> Signed-off-by: Justin Ruggles <justin.ruggles@gmail.com>
13 years ago
lea picregq, [scan8_mem]
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
%endif
x86: mmx2 ---> mmxext in asm constructs
13 years ago
call h264_idct_add8_mmxext_plane
4:4:4 H.264 decoding support Note: this is 4:4:4 from the 2007 spec revision, not the previous (now deprecated) 4:4:4 mode in H.264.
14 years ago
mov r5, 32
add r2, 384
config.asm: change %ifdef directives to %if directives. This allows combining multiple conditionals in a single statement.
13 years ago
%if ARCH_X86_64
x86inc improvements for 64-bit Add support for all x86-64 registers Prefer caller-saved register over callee-saved on WIN64 Support up to 15 function arguments Also (by Ronald S. Bultje) Fix up our asm to work with new x86inc.asm. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com> Signed-off-by: Justin Ruggles <justin.ruggles@gmail.com>
13 years ago
add dst2q, gprsize
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
%else
add r0mp, gprsize
%endif
x86: mmx2 ---> mmxext in asm constructs
13 years ago
call h264_idct_add8_mmxext_plane
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
RET
; r0 = uint8_t *dst, r2 = int16_t *block, r3 = int stride, r6=clobbered
x86: mmx2 ---> mmxext in asm constructs
13 years ago
h264_idct_dc_add8_mmxext:
x86: Add missing movsxd for the int stride parameter Signed-off-by: Martin Storsjö <martin@martin.st>
9 years ago
movsxdifnidn r3, r3d
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
movd m0, [r2 ] ; 0 0 X D
h264: Integrate clear_blocks calls with IDCT The non-intra-pcm branch in hl_decode_mb (simple, 8bpp) goes from 700 to 672 cycles, and the complete loop of decode_mb_cabac and hl_decode_mb (in the decode_slice loop) goes from 1759 to 1733 cycles on the clip tested (cathedral), i.e. almost 30 cycles per mb faster. Signed-off-by: Martin Storsjö <martin@martin.st>
12 years ago
mov word [r2+ 0], 0
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
punpcklwd m0, [r2+32] ; x X d D
h264: Integrate clear_blocks calls with IDCT The non-intra-pcm branch in hl_decode_mb (simple, 8bpp) goes from 700 to 672 cycles, and the complete loop of decode_mb_cabac and hl_decode_mb (in the decode_slice loop) goes from 1759 to 1733 cycles on the clip tested (cathedral), i.e. almost 30 cycles per mb faster. Signed-off-by: Martin Storsjö <martin@martin.st>
12 years ago
mov word [r2+32], 0
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
paddsw m0, [pw_32]
psraw m0, 6
punpcklwd m0, m0 ; d d D D
pxor m1, m1 ; 0 0 0 0
psubw m1, m0 ; -d-d-D-D
packuswb m0, m1 ; -d-d-D-D d d D D
pshufw m1, m0, 0xFA ; -d-d-d-d-D-D-D-D
punpcklwd m0, m0 ; d d d d D D D D
lea r6, [r3*3]
x86: MMX2 ---> MMXEXT in macro names
13 years ago
DC_ADD_MMXEXT_OP movq, r0, r3, r6
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
ret
ALIGN 16
x86: h264_idct: port to cpuflags
13 years ago
INIT_XMM sse2
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
; r0 = uint8_t *dst (clobbered), r2 = int16_t *block, r3 = int stride
x86: h264_idct: Rename x264_add8x4_idct_sse2 --> h264_add8x4_idct_sse2
12 years ago
h264_add8x4_idct_sse2:
x86: Add missing movsxd for the int stride parameter Signed-off-by: Martin Storsjö <martin@martin.st>
9 years ago
movsxdifnidn r3, r3d
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
movq m0, [r2+ 0]
movq m1, [r2+ 8]
movq m2, [r2+16]
movq m3, [r2+24]
movhps m0, [r2+32]
movhps m1, [r2+40]
movhps m2, [r2+48]
movhps m3, [r2+56]
Modify x86util.asm to ease transitioning to 10-bit H.264 assembly. Arguments for variable size instructions are added to many macros, along with other various changes. The x86util.asm code was ported from x264. Signed-off-by: Diego Biurrun <diego@biurrun.de>
14 years ago
IDCT4_1D w,0,1,2,3,4,5
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
TRANSPOSE2x4x4W 0,1,2,3,4
paddw m0, [pw_32]
Modify x86util.asm to ease transitioning to 10-bit H.264 assembly. Arguments for variable size instructions are added to many macros, along with other various changes. The x86util.asm code was ported from x264. Signed-off-by: Diego Biurrun <diego@biurrun.de>
14 years ago
IDCT4_1D w,0,1,2,3,4,5
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
pxor m7, m7
h264: Integrate clear_blocks calls with IDCT The non-intra-pcm branch in hl_decode_mb (simple, 8bpp) goes from 700 to 672 cycles, and the complete loop of decode_mb_cabac and hl_decode_mb (in the decode_slice loop) goes from 1759 to 1733 cycles on the clip tested (cathedral), i.e. almost 30 cycles per mb faster. Signed-off-by: Martin Storsjö <martin@martin.st>
12 years ago
mova [r2+ 0], m7
mova [r2+16], m7
mova [r2+32], m7
mova [r2+48], m7
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
STORE_DIFFx2 m0, m1, m4, m5, m7, 6, r0, r3
lea r0, [r0+r3*2]
STORE_DIFFx2 m2, m3, m4, m5, m7, 6, r0, r3
ret
%macro add16_sse2_cycle 2
movzx r0, word [r4+%2]
test r0, r0
jz .cycle%1end
mov r0d, dword [r1+%1*8]
config.asm: change %ifdef directives to %if directives. This allows combining multiple conditionals in a single statement.
13 years ago
%if ARCH_X86_64
x86inc improvements for 64-bit Add support for all x86-64 registers Prefer caller-saved register over callee-saved on WIN64 Support up to 15 function arguments Also (by Ronald S. Bultje) Fix up our asm to work with new x86inc.asm. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com> Signed-off-by: Justin Ruggles <justin.ruggles@gmail.com>
13 years ago
add r0, r5
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
%else
add r0, r0m
%endif
x86: h264_idct: Rename x264_add8x4_idct_sse2 --> h264_add8x4_idct_sse2
12 years ago
call h264_add8x4_idct_sse2
x86: add colons after labels nasm prints a warning if the colon is missing. Signed-off-by: Mans Rullgard <mans@mansr.com>
12 years ago
.cycle%1end:
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
%if %1 < 7
add r2, 64
%endif
%endmacro
x86: Make function prototype comments in assembly code consistent This helps grepping for functions, among other things.
11 years ago
; void ff_h264_idct_add16_8_sse2(uint8_t *dst, const int *block_offset,
; int16_t *block, int stride,
; const uint8_t nnzc[6 * 8])
x86: h264_idct: port to cpuflags
13 years ago
cglobal h264_idct_add16_8, 5, 5 + ARCH_X86_64, 8
x86: Add missing movsxd for the int stride parameter Signed-off-by: Martin Storsjö <martin@martin.st>
9 years ago
movsxdifnidn r3, r3d
config.asm: change %ifdef directives to %if directives. This allows combining multiple conditionals in a single statement.
13 years ago
%if ARCH_X86_64
x86inc improvements for 64-bit Add support for all x86-64 registers Prefer caller-saved register over callee-saved on WIN64 Support up to 15 function arguments Also (by Ronald S. Bultje) Fix up our asm to work with new x86inc.asm. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com> Signed-off-by: Justin Ruggles <justin.ruggles@gmail.com>
13 years ago
mov r5, r0
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
%endif
; unrolling of the loop leads to an average performance gain of
; 20-25%
add16_sse2_cycle 0, 0xc
add16_sse2_cycle 1, 0x14
add16_sse2_cycle 2, 0xe
add16_sse2_cycle 3, 0x16
add16_sse2_cycle 4, 0x1c
add16_sse2_cycle 5, 0x24
add16_sse2_cycle 6, 0x1e
add16_sse2_cycle 7, 0x26
RET
Unroll loop in h264_idct_add16intra_sse2(). Basically identical to r25171, this inlines scan8[] and removes loop setup. 15% faster, 0.4% overall. See "[PATCH] unroll loop in h264_idct_add8_sse2()" thread on ML. Originally committed as revision 25172 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
%macro add16intra_sse2_cycle 2
movzx r0, word [r4+%2]
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
test r0, r0
Unroll loop in h264_idct_add16intra_sse2(). Basically identical to r25171, this inlines scan8[] and removes loop setup. 15% faster, 0.4% overall. See "[PATCH] unroll loop in h264_idct_add8_sse2()" thread on ML. Originally committed as revision 25172 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
jz .try%1dc
mov r0d, dword [r1+%1*8]
config.asm: change %ifdef directives to %if directives. This allows combining multiple conditionals in a single statement.
13 years ago
%if ARCH_X86_64
x86inc improvements for 64-bit Add support for all x86-64 registers Prefer caller-saved register over callee-saved on WIN64 Support up to 15 function arguments Also (by Ronald S. Bultje) Fix up our asm to work with new x86inc.asm. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com> Signed-off-by: Justin Ruggles <justin.ruggles@gmail.com>
13 years ago
add r0, r7
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
%else
add r0, r0m
%endif
x86: h264_idct: Rename x264_add8x4_idct_sse2 --> h264_add8x4_idct_sse2
12 years ago
call h264_add8x4_idct_sse2
Unroll loop in h264_idct_add16intra_sse2(). Basically identical to r25171, this inlines scan8[] and removes loop setup. 15% faster, 0.4% overall. See "[PATCH] unroll loop in h264_idct_add8_sse2()" thread on ML. Originally committed as revision 25172 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
jmp .cycle%1end
x86: add colons after labels nasm prints a warning if the colon is missing. Signed-off-by: Mans Rullgard <mans@mansr.com>
12 years ago
.try%1dc:
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
movsx r0, word [r2 ]
or r0w, word [r2+32]
Unroll loop in h264_idct_add16intra_sse2(). Basically identical to r25171, this inlines scan8[] and removes loop setup. 15% faster, 0.4% overall. See "[PATCH] unroll loop in h264_idct_add8_sse2()" thread on ML. Originally committed as revision 25172 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
jz .cycle%1end
mov r0d, dword [r1+%1*8]
config.asm: change %ifdef directives to %if directives. This allows combining multiple conditionals in a single statement.
13 years ago
%if ARCH_X86_64
x86inc improvements for 64-bit Add support for all x86-64 registers Prefer caller-saved register over callee-saved on WIN64 Support up to 15 function arguments Also (by Ronald S. Bultje) Fix up our asm to work with new x86inc.asm. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com> Signed-off-by: Justin Ruggles <justin.ruggles@gmail.com>
13 years ago
add r0, r7
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
%else
add r0, r0m
%endif
x86: mmx2 ---> mmxext in asm constructs
13 years ago
call h264_idct_dc_add8_mmxext
x86: add colons after labels nasm prints a warning if the colon is missing. Signed-off-by: Mans Rullgard <mans@mansr.com>
12 years ago
.cycle%1end:
Unroll loop in h264_idct_add16intra_sse2(). Basically identical to r25171, this inlines scan8[] and removes loop setup. 15% faster, 0.4% overall. See "[PATCH] unroll loop in h264_idct_add8_sse2()" thread on ML. Originally committed as revision 25172 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
%if %1 < 7
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
add r2, 64
Unroll loop in h264_idct_add16intra_sse2(). Basically identical to r25171, this inlines scan8[] and removes loop setup. 15% faster, 0.4% overall. See "[PATCH] unroll loop in h264_idct_add8_sse2()" thread on ML. Originally committed as revision 25172 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
%endif
%endmacro
x86: Make function prototype comments in assembly code consistent This helps grepping for functions, among other things.
11 years ago
; void ff_h264_idct_add16intra_8_sse2(uint8_t *dst, const int *block_offset,
; int16_t *block, int stride,
; const uint8_t nnzc[6 * 8])
x86: h264_idct: port to cpuflags
13 years ago
cglobal h264_idct_add16intra_8, 5, 7 + ARCH_X86_64, 8
x86: Add missing movsxd for the int stride parameter Signed-off-by: Martin Storsjö <martin@martin.st>
9 years ago
movsxdifnidn r3, r3d
config.asm: change %ifdef directives to %if directives. This allows combining multiple conditionals in a single statement.
13 years ago
%if ARCH_X86_64
x86inc improvements for 64-bit Add support for all x86-64 registers Prefer caller-saved register over callee-saved on WIN64 Support up to 15 function arguments Also (by Ronald S. Bultje) Fix up our asm to work with new x86inc.asm. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com> Signed-off-by: Justin Ruggles <justin.ruggles@gmail.com>
13 years ago
mov r7, r0
Unroll loop in h264_idct_add16intra_sse2(). Basically identical to r25171, this inlines scan8[] and removes loop setup. 15% faster, 0.4% overall. See "[PATCH] unroll loop in h264_idct_add8_sse2()" thread on ML. Originally committed as revision 25172 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
%endif
add16intra_sse2_cycle 0, 0xc
add16intra_sse2_cycle 1, 0x14
add16intra_sse2_cycle 2, 0xe
add16intra_sse2_cycle 3, 0x16
add16intra_sse2_cycle 4, 0x1c
add16intra_sse2_cycle 5, 0x24
add16intra_sse2_cycle 6, 0x1e
add16intra_sse2_cycle 7, 0x26
RET
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
Unroll loop in h264_idct_add8_sse2(). This means we can inline scan8[] in the code directly also and remove loop setup. 20% faster in function, 0.8% overall. See "[PATCH] unroll loop in h264_idct_add8_sse2()" thread on ML. Originally committed as revision 25171 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
%macro add8_sse2_cycle 2
movzx r0, word [r4+%2]
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
test r0, r0
Unroll loop in h264_idct_add8_sse2(). This means we can inline scan8[] in the code directly also and remove loop setup. 20% faster in function, 0.8% overall. See "[PATCH] unroll loop in h264_idct_add8_sse2()" thread on ML. Originally committed as revision 25171 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
jz .try%1dc
config.asm: change %ifdef directives to %if directives. This allows combining multiple conditionals in a single statement.
13 years ago
%if ARCH_X86_64
4:4:4 H.264 decoding support Note: this is 4:4:4 from the 2007 spec revision, not the previous (now deprecated) 4:4:4 mode in H.264.
14 years ago
mov r0d, dword [r1+(%1&1)*8+64*(1+(%1>>1))]
x86inc improvements for 64-bit Add support for all x86-64 registers Prefer caller-saved register over callee-saved on WIN64 Support up to 15 function arguments Also (by Ronald S. Bultje) Fix up our asm to work with new x86inc.asm. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com> Signed-off-by: Justin Ruggles <justin.ruggles@gmail.com>
13 years ago
add r0, [r7]
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
%else
Unroll loop in h264_idct_add8_sse2(). This means we can inline scan8[] in the code directly also and remove loop setup. 20% faster in function, 0.8% overall. See "[PATCH] unroll loop in h264_idct_add8_sse2()" thread on ML. Originally committed as revision 25171 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
mov r0, r0m
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
mov r0, [r0]
4:4:4 H.264 decoding support Note: this is 4:4:4 from the 2007 spec revision, not the previous (now deprecated) 4:4:4 mode in H.264.
14 years ago
add r0, dword [r1+(%1&1)*8+64*(1+(%1>>1))]
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
%endif
x86: h264_idct: Rename x264_add8x4_idct_sse2 --> h264_add8x4_idct_sse2
12 years ago
call h264_add8x4_idct_sse2
Unroll loop in h264_idct_add8_sse2(). This means we can inline scan8[] in the code directly also and remove loop setup. 20% faster in function, 0.8% overall. See "[PATCH] unroll loop in h264_idct_add8_sse2()" thread on ML. Originally committed as revision 25171 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
jmp .cycle%1end
x86: add colons after labels nasm prints a warning if the colon is missing. Signed-off-by: Mans Rullgard <mans@mansr.com>
12 years ago
.try%1dc:
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
movsx r0, word [r2 ]
or r0w, word [r2+32]
Unroll loop in h264_idct_add8_sse2(). This means we can inline scan8[] in the code directly also and remove loop setup. 20% faster in function, 0.8% overall. See "[PATCH] unroll loop in h264_idct_add8_sse2()" thread on ML. Originally committed as revision 25171 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
jz .cycle%1end
config.asm: change %ifdef directives to %if directives. This allows combining multiple conditionals in a single statement.
13 years ago
%if ARCH_X86_64
4:4:4 H.264 decoding support Note: this is 4:4:4 from the 2007 spec revision, not the previous (now deprecated) 4:4:4 mode in H.264.
14 years ago
mov r0d, dword [r1+(%1&1)*8+64*(1+(%1>>1))]
x86inc improvements for 64-bit Add support for all x86-64 registers Prefer caller-saved register over callee-saved on WIN64 Support up to 15 function arguments Also (by Ronald S. Bultje) Fix up our asm to work with new x86inc.asm. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com> Signed-off-by: Justin Ruggles <justin.ruggles@gmail.com>
13 years ago
add r0, [r7]
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
%else
Unroll loop in h264_idct_add8_sse2(). This means we can inline scan8[] in the code directly also and remove loop setup. 20% faster in function, 0.8% overall. See "[PATCH] unroll loop in h264_idct_add8_sse2()" thread on ML. Originally committed as revision 25171 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
mov r0, r0m
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
mov r0, [r0]
4:4:4 H.264 decoding support Note: this is 4:4:4 from the 2007 spec revision, not the previous (now deprecated) 4:4:4 mode in H.264.
14 years ago
add r0, dword [r1+(%1&1)*8+64*(1+(%1>>1))]
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
%endif
x86: mmx2 ---> mmxext in asm constructs
13 years ago
call h264_idct_dc_add8_mmxext
x86: add colons after labels nasm prints a warning if the colon is missing. Signed-off-by: Mans Rullgard <mans@mansr.com>
12 years ago
.cycle%1end:
4:4:4 H.264 decoding support Note: this is 4:4:4 from the 2007 spec revision, not the previous (now deprecated) 4:4:4 mode in H.264.
14 years ago
%if %1 == 1
add r2, 384+64
%elif %1 < 3
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
add r2, 64
Unroll loop in h264_idct_add8_sse2(). This means we can inline scan8[] in the code directly also and remove loop setup. 20% faster in function, 0.8% overall. See "[PATCH] unroll loop in h264_idct_add8_sse2()" thread on ML. Originally committed as revision 25171 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
%endif
%endmacro
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
x86: Make function prototype comments in assembly code consistent This helps grepping for functions, among other things.
11 years ago
; void ff_h264_idct_add8_8_sse2(uint8_t **dest, const int *block_offset,
; int16_t *block, int stride,
; const uint8_t nnzc[6 * 8])
x86: h264_idct: port to cpuflags
13 years ago
cglobal h264_idct_add8_8, 5, 7 + ARCH_X86_64, 8
x86: Add missing movsxd for the int stride parameter Signed-off-by: Martin Storsjö <martin@martin.st>
9 years ago
movsxdifnidn r3, r3d
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
add r2, 512
config.asm: change %ifdef directives to %if directives. This allows combining multiple conditionals in a single statement.
13 years ago
%if ARCH_X86_64
x86inc improvements for 64-bit Add support for all x86-64 registers Prefer caller-saved register over callee-saved on WIN64 Support up to 15 function arguments Also (by Ronald S. Bultje) Fix up our asm to work with new x86inc.asm. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com> Signed-off-by: Justin Ruggles <justin.ruggles@gmail.com>
13 years ago
mov r7, r0
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
%endif
4:4:4 H.264 decoding support Note: this is 4:4:4 from the 2007 spec revision, not the previous (now deprecated) 4:4:4 mode in H.264.
14 years ago
add8_sse2_cycle 0, 0x34
add8_sse2_cycle 1, 0x3c
config.asm: change %ifdef directives to %if directives. This allows combining multiple conditionals in a single statement.
13 years ago
%if ARCH_X86_64
x86inc improvements for 64-bit Add support for all x86-64 registers Prefer caller-saved register over callee-saved on WIN64 Support up to 15 function arguments Also (by Ronald S. Bultje) Fix up our asm to work with new x86inc.asm. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com> Signed-off-by: Justin Ruggles <justin.ruggles@gmail.com>
13 years ago
add r7, gprsize
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
%else
add r0mp, gprsize
%endif
4:4:4 H.264 decoding support Note: this is 4:4:4 from the 2007 spec revision, not the previous (now deprecated) 4:4:4 mode in H.264.
14 years ago
add8_sse2_cycle 2, 0x5c
add8_sse2_cycle 3, 0x64
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%. Originally committed as revision 25119 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
RET
H.264: split luma dc idct out and implement MMX/SSE2 versions About 2.5x the speed. NOTE: the way that the asm code handles large qmuls is a bit suboptimal. If x264-style dequant was used (separate shift and qmul values), it might be possible to get some extra speed. Originally committed as revision 26336 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
Drop DCTELEM typedef It does not help as an abstraction and adds dsputil dependencies. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
12 years ago
;void ff_h264_luma_dc_dequant_idct_mmx(int16_t *output, int16_t *input, int qmul)
H.264: split luma dc idct out and implement MMX/SSE2 versions About 2.5x the speed. NOTE: the way that the asm code handles large qmuls is a bit suboptimal. If x264-style dequant was used (separate shift and qmul values), it might be possible to get some extra speed. Originally committed as revision 26336 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
%macro WALSH4_1D 5
Modify x86util.asm to ease transitioning to 10-bit H.264 assembly. Arguments for variable size instructions are added to many macros, along with other various changes. The x86util.asm code was ported from x264. Signed-off-by: Diego Biurrun <diego@biurrun.de>
14 years ago
SUMSUB_BADC w, %4, %3, %2, %1, %5
SUMSUB_BADC w, %4, %2, %3, %1, %5
H.264: split luma dc idct out and implement MMX/SSE2 versions About 2.5x the speed. NOTE: the way that the asm code handles large qmuls is a bit suboptimal. If x264-style dequant was used (separate shift and qmul values), it might be possible to get some extra speed. Originally committed as revision 26336 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
SWAP %1, %4, %3
%endmacro
%macro DEQUANT_MMX 3
mova m7, [pw_1]
mova m4, %1
punpcklwd %1, m7
punpckhwd m4, m7
mova m5, %2
punpcklwd %2, m7
punpckhwd m5, m7
movd m7, t3d
punpckldq m7, m7
pmaddwd %1, m7
pmaddwd %2, m7
pmaddwd m4, m7
pmaddwd m5, m7
psrad %1, %3
psrad %2, %3
psrad m4, %3
psrad m5, %3
packssdw %1, m4
packssdw %2, m5
%endmacro
x86: h264_idct: port to cpuflags
13 years ago
%macro STORE_WORDS 5-9
%if cpuflag(sse)
H.264: split luma dc idct out and implement MMX/SSE2 versions About 2.5x the speed. NOTE: the way that the asm code handles large qmuls is a bit suboptimal. If x264-style dequant was used (separate shift and qmul values), it might be possible to get some extra speed. Originally committed as revision 26336 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
movd t0d, %1
psrldq %1, 4
movd t1d, %1
psrldq %1, 4
mov [t2+%2*32], t0w
mov [t2+%4*32], t1w
shr t0d, 16
shr t1d, 16
mov [t2+%3*32], t0w
mov [t2+%5*32], t1w
movd t0d, %1
psrldq %1, 4
movd t1d, %1
mov [t2+%6*32], t0w
mov [t2+%8*32], t1w
shr t0d, 16
shr t1d, 16
mov [t2+%7*32], t0w
mov [t2+%9*32], t1w
x86: h264_idct: port to cpuflags
13 years ago
%else
movd t0d, %1
psrlq %1, 32
movd t1d, %1
mov [t2+%2*32], t0w
mov [t2+%4*32], t1w
shr t0d, 16
shr t1d, 16
mov [t2+%3*32], t0w
mov [t2+%5*32], t1w
%endif
H.264: split luma dc idct out and implement MMX/SSE2 versions About 2.5x the speed. NOTE: the way that the asm code handles large qmuls is a bit suboptimal. If x264-style dequant was used (separate shift and qmul values), it might be possible to get some extra speed. Originally committed as revision 26336 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
%endmacro
x86: h264_idct: port to cpuflags
13 years ago
%macro DEQUANT_STORE 1
%if cpuflag(sse2)
H.264: split luma dc idct out and implement MMX/SSE2 versions About 2.5x the speed. NOTE: the way that the asm code handles large qmuls is a bit suboptimal. If x264-style dequant was used (separate shift and qmul values), it might be possible to get some extra speed. Originally committed as revision 26336 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
movd xmm4, t3d
movq xmm5, [pw_1]
pshufd xmm4, xmm4, 0
movq2dq xmm0, m0
movq2dq xmm1, m1
movq2dq xmm2, m2
movq2dq xmm3, m3
punpcklwd xmm0, xmm5
punpcklwd xmm1, xmm5
punpcklwd xmm2, xmm5
punpcklwd xmm3, xmm5
pmaddwd xmm0, xmm4
pmaddwd xmm1, xmm4
pmaddwd xmm2, xmm4
pmaddwd xmm3, xmm4
psrad xmm0, %1
psrad xmm1, %1
psrad xmm2, %1
psrad xmm3, %1
packssdw xmm0, xmm1
packssdw xmm2, xmm3
x86: h264_idct: port to cpuflags
13 years ago
STORE_WORDS xmm0, 0, 1, 4, 5, 2, 3, 6, 7
STORE_WORDS xmm2, 8, 9, 12, 13, 10, 11, 14, 15
%else
DEQUANT_MMX m0, m1, %1
STORE_WORDS m0, 0, 1, 4, 5
STORE_WORDS m1, 2, 3, 6, 7
DEQUANT_MMX m2, m3, %1
STORE_WORDS m2, 8, 9, 12, 13
STORE_WORDS m3, 10, 11, 14, 15
%endif
H.264: split luma dc idct out and implement MMX/SSE2 versions About 2.5x the speed. NOTE: the way that the asm code handles large qmuls is a bit suboptimal. If x264-style dequant was used (separate shift and qmul values), it might be possible to get some extra speed. Originally committed as revision 26336 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
%endmacro
x86: h264_idct: port to cpuflags
13 years ago
%macro IDCT_DC_DEQUANT 1
cglobal h264_luma_dc_dequant_idct, 3, 4, %1
h264: manually save/restore XMM registers for functions using INIT_MMX. On Win64, these registers are callee-save, so not saving/restoring them correctly is a violation of ABI and can lead to crashes or corrupt data.
13 years ago
; manually spill XMM registers for Win64 because
; the code here is initialized with INIT_MMX
x86: h264_idct: port to cpuflags
13 years ago
WIN64_SPILL_XMM %1
H.264: split luma dc idct out and implement MMX/SSE2 versions About 2.5x the speed. NOTE: the way that the asm code handles large qmuls is a bit suboptimal. If x264-style dequant was used (separate shift and qmul values), it might be possible to get some extra speed. Originally committed as revision 26336 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
movq m3, [r1+24]
movq m2, [r1+16]
movq m1, [r1+ 8]
movq m0, [r1+ 0]
WALSH4_1D 0,1,2,3,4
TRANSPOSE4x4W 0,1,2,3,4
WALSH4_1D 0,1,2,3,4
; shift, tmp, output, qmul
config.asm: change %ifdef directives to %if directives. This allows combining multiple conditionals in a single statement.
13 years ago
%if WIN64
H.264: split luma dc idct out and implement MMX/SSE2 versions About 2.5x the speed. NOTE: the way that the asm code handles large qmuls is a bit suboptimal. If x264-style dequant was used (separate shift and qmul values), it might be possible to get some extra speed. Originally committed as revision 26336 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
DECLARE_REG_TMP 0,3,1,2
; we can't avoid this, because r0 is the shift register (ecx) on win64
xchg r0, t2
config.asm: change %ifdef directives to %if directives. This allows combining multiple conditionals in a single statement.
13 years ago
%elif ARCH_X86_64
H.264: split luma dc idct out and implement MMX/SSE2 versions About 2.5x the speed. NOTE: the way that the asm code handles large qmuls is a bit suboptimal. If x264-style dequant was used (separate shift and qmul values), it might be possible to get some extra speed. Originally committed as revision 26336 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
DECLARE_REG_TMP 3,1,0,2
%else
DECLARE_REG_TMP 1,3,0,2
%endif
cmp t3d, 32767
jg .big_qmul
add t3d, 128 << 16
x86: h264_idct: port to cpuflags
13 years ago
DEQUANT_STORE 8
H.264: split luma dc idct out and implement MMX/SSE2 versions About 2.5x the speed. NOTE: the way that the asm code handles large qmuls is a bit suboptimal. If x264-style dequant was used (separate shift and qmul values), it might be possible to get some extra speed. Originally committed as revision 26336 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
RET
.big_qmul:
bsr t0d, t3d
add t3d, 128 << 16
mov t1d, 7
cmp t0d, t1d
cmovg t0d, t1d
inc t1d
shr t3d, t0b
sub t1d, t0d
x86: h264_idct: port to cpuflags
13 years ago
%if cpuflag(sse2)
H.264: split luma dc idct out and implement MMX/SSE2 versions About 2.5x the speed. NOTE: the way that the asm code handles large qmuls is a bit suboptimal. If x264-style dequant was used (separate shift and qmul values), it might be possible to get some extra speed. Originally committed as revision 26336 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
movd xmm6, t1d
x86: h264_idct: port to cpuflags
13 years ago
DEQUANT_STORE xmm6
%else
movd m6, t1d
DEQUANT_STORE m6
H.264: split luma dc idct out and implement MMX/SSE2 versions About 2.5x the speed. NOTE: the way that the asm code handles large qmuls is a bit suboptimal. If x264-style dequant was used (separate shift and qmul values), it might be possible to get some extra speed. Originally committed as revision 26336 to svn://svn.ffmpeg.org/ffmpeg/trunk
14 years ago
%endif
RET
%endmacro
x86: h264_idct: port to cpuflags
13 years ago
INIT_MMX mmx
IDCT_DC_DEQUANT 0
INIT_MMX sse2
IDCT_DC_DEQUANT 7