generic-library/ffmpeg
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity
0a216bd1dd
ffmpeg/libavcodec/x86/h264_idct.asm

857 lines
21 KiB
NASM
Raw Normal View History

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
2010-09-14 15:36:26 +02:00
;*****************************************************************************
;* 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>
;*
;* This file is part of FFmpeg.
;*
;* FFmpeg is free software; you can redistribute it and/or
;* 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.
;*
;* FFmpeg is distributed in the hope that it will be useful,
;* 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
;* License along with FFmpeg; if not, write to the Free Software
;* 51, Inc., Foundation Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
;*****************************************************************************
%include "x86inc.asm"
%include "x86util.asm"
SECTION_RODATA
; FIXME this table is a duplicate from h264data.h, and will be removed once the tables from, h264 have been split
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 1+1*8, 2+1*8
db 1+2*8, 2+2*8
db 1+4*8, 2+4*8
db 1+5*8, 2+5*8
%ifdef PIC
%define scan8 r11
%else
%define scan8 scan8_mem
%endif
cextern pw_32
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]
IDCT4_1D 0, 1, 2, 3, 4, 5
mova m6, [pw_32]
TRANSPOSE4x4W 0, 1, 2, 3, 4
paddw m0, m6
IDCT4_1D 0, 1, 2, 3, 4, 5
pxor m7, m7
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
INIT_MMX
; ff_h264_idct_add_mmx(uint8_t *dst, int16_t *block, int stride)
cglobal h264_idct_add_mmx, 3, 3, 0
IDCT4_ADD r0, r1, r2
RET
%macro IDCT8_1D 2
mova m4, m5
mova m0, m1
psraw m4, 1
psraw m1, 1
paddw m4, m5
paddw m1, m0
paddw m4, m7
paddw m1, m5
psubw m4, m0
paddw m1, m3
psubw m0, m3
psubw m5, m3
paddw m0, m7
psubw m5, m7
psraw m3, 1
psraw m7, 1
psubw m0, m3
psubw m5, m7
mova m3, m4
mova m7, m1
psraw m1, 2
psraw m3, 2
paddw m3, m0
psraw m0, 2
paddw m1, m5
psraw m5, 2
psubw m0, m4
psubw m7, m5
mova m4, m2
mova m5, m6
psraw m4, 1
psraw m6, 1
psubw m4, m5
paddw m6, m2
mova m2, %1
mova m5, %2
SUMSUB_BA m5, m2
SUMSUB_BA m6, m5
SUMSUB_BA m4, m2
SUMSUB_BA m7, m6
SUMSUB_BA m0, m4
SUMSUB_BA m3, m2
SUMSUB_BA m1, m5
SWAP 7, 6, 4, 5, 2, 3, 1, 0 ; 70315246 -> 01234567
%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
%macro IDCT8_ADD_MMX_END 3
IDCT8_1D_FULL %2
mova [%2 ], m5
mova [%2+16], m6
mova [%2+32], m7
pxor m7, m7
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
INIT_MMX
; ff_h264_idct8_add_mmx(uint8_t *dst, int16_t *block, int stride)
cglobal h264_idct8_add_mmx, 3, 4, 0
%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]
IDCT8_ADD_MMX_END r0 , rsp, r2
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
%ifdef ARCH_X86_64
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]
%ifndef ARCH_X86_64
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 ]
%ifndef ARCH_X86_64
mova m0, [%2 ]
mova m1, [%2+16]
%else
SWAP 0, 8
SWAP 1, 9
%endif
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
INIT_XMM
; ff_h264_idct8_add_sse2(uint8_t *dst, int16_t *block, int stride)
cglobal h264_idct8_add_sse2, 3, 4, 10
IDCT8_ADD_SSE r0, r1, r2, r3
RET
%macro DC_ADD_MMX2_INIT 2-3
%if %0 == 2
movsx %1, word [%1]
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
2010-09-26 11:15:18 +02:00
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
2010-09-14 15:36:26 +02:00
lea %1, [%2*3]
%else
add %3, 32
sar %3, 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
2010-09-26 11:15:18 +02:00
movd m0, %3d
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
2010-09-14 15:36:26 +02:00
lea %3, [%2*3]
%endif
pshufw m0, m0, 0
pxor m1, m1
psubw m1, m0
packuswb m0, m0
packuswb m1, m1
%endmacro
%macro DC_ADD_MMX2_OP 3-4
%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
INIT_MMX
; ff_h264_idct_dc_add_mmx2(uint8_t *dst, int16_t *block, int stride)
cglobal h264_idct_dc_add_mmx2, 3, 3, 0
DC_ADD_MMX2_INIT r1, r2
DC_ADD_MMX2_OP movh, r0, r2, r1
RET
; ff_h264_idct8_dc_add_mmx2(uint8_t *dst, int16_t *block, int stride)
cglobal h264_idct8_dc_add_mmx2, 3, 3, 0
DC_ADD_MMX2_INIT r1, r2
DC_ADD_MMX2_OP mova, r0, r2, r1
lea r0, [r0+r2*4]
DC_ADD_MMX2_OP mova, r0, r2, r1
RET
; ff_h264_idct_add16_mmx(uint8_t *dst, const int *block_offset,
; DCTELEM *block, int stride, const uint8_t nnzc[6*8])
cglobal h264_idct_add16_mmx, 5, 7, 0
xor r5, r5
%ifdef PIC
lea r11, [scan8_mem]
%endif
.nextblock
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
.skipblock
inc r5
add r2, 32
cmp r5, 16
jl .nextblock
REP_RET
; ff_h264_idct8_add4_mmx(uint8_t *dst, const int *block_offset,
; DCTELEM *block, int stride, const uint8_t nnzc[6*8])
cglobal h264_idct8_add4_mmx, 5, 7, 0
%assign pad 128+4-(stack_offset&7)
SUB rsp, pad
xor r5, r5
%ifdef PIC
lea r11, [scan8_mem]
%endif
.nextblock
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]
add word [r2], 32
IDCT8_ADD_MMX_START r2 , rsp
IDCT8_ADD_MMX_START r2+8, rsp+64
IDCT8_ADD_MMX_END r6 , rsp, r3
mov r6d, dword [r1+r5*4]
lea r6, [r0+r6+4]
IDCT8_ADD_MMX_END r6 , rsp+8, r3
.skipblock
add r5, 4
add r2, 128
cmp r5, 16
jl .nextblock
ADD rsp, pad
RET
; ff_h264_idct_add16_mmx2(uint8_t *dst, const int *block_offset,
; DCTELEM *block, int stride, const uint8_t nnzc[6*8])
cglobal h264_idct_add16_mmx2, 5, 7, 0
xor r5, r5
%ifdef PIC
lea r11, [scan8_mem]
%endif
.nextblock
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
DC_ADD_MMX2_INIT r2, r3, r6
%ifdef ARCH_X86_64
%define dst_reg r10
%define dst_regd r10d
%else
%define dst_reg r1
%define dst_regd r1d
%endif
mov dst_regd, dword [r1+r5*4]
lea dst_reg, [r0+dst_reg]
DC_ADD_MMX2_OP movh, dst_reg, r3, r6
%ifndef ARCH_X86_64
mov r1, r1m
%endif
inc r5
add r2, 32
cmp r5, 16
jl .nextblock
REP_RET
.no_dc
mov r6d, dword [r1+r5*4]
lea r6, [r0+r6]
IDCT4_ADD r6, r2, r3
.skipblock
inc r5
add r2, 32
cmp r5, 16
jl .nextblock
REP_RET
; ff_h264_idct_add16intra_mmx(uint8_t *dst, const int *block_offset,
; DCTELEM *block, int stride, const uint8_t nnzc[6*8])
cglobal h264_idct_add16intra_mmx, 5, 7, 0
xor r5, r5
%ifdef PIC
lea r11, [scan8_mem]
%endif
.nextblock
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]
lea r6, [r0+r6]
IDCT4_ADD r6, r2, r3
.skipblock
inc r5
add r2, 32
cmp r5, 16
jl .nextblock
REP_RET
; ff_h264_idct_add16intra_mmx2(uint8_t *dst, const int *block_offset,
; DCTELEM *block, int stride, const uint8_t nnzc[6*8])
cglobal h264_idct_add16intra_mmx2, 5, 7, 0
xor r5, r5
%ifdef PIC
lea r11, [scan8_mem]
%endif
.nextblock
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
.try_dc
movsx r6, word [r2]
test r6, r6
jz .skipblock
DC_ADD_MMX2_INIT r2, r3, r6
%ifdef ARCH_X86_64
%define dst_reg r10
%define dst_regd r10d
%else
%define dst_reg r1
%define dst_regd r1d
%endif
mov dst_regd, dword [r1+r5*4]
lea dst_reg, [r0+dst_reg]
DC_ADD_MMX2_OP movh, dst_reg, r3, r6
%ifndef ARCH_X86_64
mov r1, r1m
%endif
.skipblock
inc r5
add r2, 32
cmp r5, 16
jl .nextblock
REP_RET
; ff_h264_idct8_add4_mmx2(uint8_t *dst, const int *block_offset,
; DCTELEM *block, int stride, const uint8_t nnzc[6*8])
cglobal h264_idct8_add4_mmx2, 5, 7, 0
%assign pad 128+4-(stack_offset&7)
SUB rsp, pad
xor r5, r5
%ifdef PIC
lea r11, [scan8_mem]
%endif
.nextblock
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
DC_ADD_MMX2_INIT r2, r3, r6
%ifdef ARCH_X86_64
%define dst_reg r10
%define dst_regd r10d
%else
%define dst_reg r1
%define dst_regd r1d
%endif
mov dst_regd, dword [r1+r5*4]
lea dst_reg, [r0+dst_reg]
DC_ADD_MMX2_OP mova, dst_reg, r3, r6
lea dst_reg, [dst_reg+r3*4]
DC_ADD_MMX2_OP mova, dst_reg, r3, r6
%ifndef ARCH_X86_64
mov r1, r1m
%endif
add r5, 4
add r2, 128
cmp r5, 16
jl .nextblock
ADD rsp, pad
RET
.no_dc
mov r6d, dword [r1+r5*4]
lea r6, [r0+r6]
add word [r2], 32
IDCT8_ADD_MMX_START r2 , rsp
IDCT8_ADD_MMX_START r2+8, rsp+64
IDCT8_ADD_MMX_END r6 , rsp, r3
mov r6d, dword [r1+r5*4]
lea r6, [r0+r6+4]
IDCT8_ADD_MMX_END r6 , rsp+8, r3
.skipblock
add r5, 4
add r2, 128
cmp r5, 16
jl .nextblock
ADD rsp, pad
RET
INIT_XMM
; ff_h264_idct8_add4_sse2(uint8_t *dst, const int *block_offset,
; DCTELEM *block, int stride, const uint8_t nnzc[6*8])
cglobal h264_idct8_add4_sse2, 5, 7, 10
xor r5, r5
%ifdef PIC
lea r11, [scan8_mem]
%endif
.nextblock
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
INIT_MMX
DC_ADD_MMX2_INIT r2, r3, r6
%ifdef ARCH_X86_64
%define dst_reg r10
%define dst_regd r10d
%else
%define dst_reg r1
%define dst_regd r1d
%endif
mov dst_regd, dword [r1+r5*4]
lea dst_reg, [r0+dst_reg]
DC_ADD_MMX2_OP mova, dst_reg, r3, r6
lea dst_reg, [dst_reg+r3*4]
DC_ADD_MMX2_OP mova, dst_reg, r3, r6
%ifndef ARCH_X86_64
mov r1, r1m
%endif
add r5, 4
add r2, 128
cmp r5, 16
jl .nextblock
REP_RET
.no_dc
INIT_XMM
mov dst_regd, dword [r1+r5*4]
lea dst_reg, [r0+dst_reg]
IDCT8_ADD_SSE dst_reg, r2, r3, r6
%ifndef ARCH_X86_64
mov r1, r1m
%endif
.skipblock
add r5, 4
add r2, 128
cmp r5, 16
jl .nextblock
REP_RET
INIT_MMX
h264_idct_add8_mmx_plane:
.nextblock
movzx r6, byte [scan8+r5]
movzx r6, byte [r4+r6]
or r6w, word [r2]
test r6, r6
jz .skipblock
%ifdef ARCH_X86_64
mov r0d, dword [r1+r5*4]
add r0, [r10]
%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
.skipblock
inc r5
add r2, 32
test r5, 3
jnz .nextblock
rep ret
; ff_h264_idct_add8_mmx(uint8_t **dest, const int *block_offset,
; DCTELEM *block, int stride, const uint8_t nnzc[6*8])
cglobal h264_idct_add8_mmx, 5, 7, 0
mov r5, 16
add r2, 512
%ifdef PIC
lea r11, [scan8_mem]
%endif
%ifdef ARCH_X86_64
mov r10, r0
%endif
call h264_idct_add8_mmx_plane
%ifdef ARCH_X86_64
add r10, gprsize
%else
add r0mp, gprsize
%endif
call h264_idct_add8_mmx_plane
RET
h264_idct_add8_mmx2_plane
.nextblock
movzx r6, byte [scan8+r5]
movzx r6, byte [r4+r6]
test r6, r6
jz .try_dc
%ifdef ARCH_X86_64
mov r0d, dword [r1+r5*4]
add r0, [r10]
%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
.try_dc
movsx r6, word [r2]
test r6, r6
jz .skipblock
DC_ADD_MMX2_INIT r2, r3, r6
%ifdef ARCH_X86_64
mov r0d, dword [r1+r5*4]
add r0, [r10]
%else
mov r0, r1m ; XXX r1m here is actually r0m of the calling func
mov r0, [r0]
add r0, dword [r1+r5*4]
%endif
DC_ADD_MMX2_OP movh, r0, r3, r6
.skipblock
inc r5
add r2, 32
test r5, 3
jnz .nextblock
rep ret
; ff_h264_idct_add8_mmx2(uint8_t **dest, const int *block_offset,
; DCTELEM *block, int stride, const uint8_t nnzc[6*8])
cglobal h264_idct_add8_mmx2, 5, 7, 0
mov r5, 16
add r2, 512
%ifdef ARCH_X86_64
mov r10, r0
%endif
%ifdef PIC
lea r11, [scan8_mem]
%endif
call h264_idct_add8_mmx2_plane
%ifdef ARCH_X86_64
add r10, gprsize
%else
add r0mp, gprsize
%endif
call h264_idct_add8_mmx2_plane
RET
INIT_MMX
; r0 = uint8_t *dst, r2 = int16_t *block, r3 = int stride, r6=clobbered
h264_idct_dc_add8_mmx2:
movd m0, [r2 ] ; 0 0 X D
punpcklwd m0, [r2+32] ; x X d D
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]
DC_ADD_MMX2_OP movq, r0, r3, r6
ret
ALIGN 16
INIT_XMM
; r0 = uint8_t *dst (clobbered), r2 = int16_t *block, r3 = int stride
x264_add8x4_idct_sse2:
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]
IDCT4_1D 0,1,2,3,4,5
TRANSPOSE2x4x4W 0,1,2,3,4
paddw m0, [pw_32]
IDCT4_1D 0,1,2,3,4,5
pxor m7, m7
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]
%ifdef ARCH_X86_64
add r0, r10
%else
add r0, r0m
%endif
call x264_add8x4_idct_sse2
.cycle%1end
%if %1 < 7
add r2, 64
%endif
%endmacro
; ff_h264_idct_add16_sse2(uint8_t *dst, const int *block_offset,
; DCTELEM *block, int stride, const uint8_t nnzc[6*8])
cglobal h264_idct_add16_sse2, 5, 5, 8
%ifdef ARCH_X86_64
mov r10, r0
%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
2010-09-24 16:07:23 +02:00
%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
2010-09-14 15:36:26 +02:00
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
2010-09-24 16:07:23 +02:00
jz .try%1dc
mov r0d, dword [r1+%1*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
2010-09-14 15:36:26 +02:00
%ifdef ARCH_X86_64
add r0, r10
%else
add r0, r0m
%endif
call x264_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
2010-09-24 16:07:23 +02:00
jmp .cycle%1end
.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
2010-09-14 15:36:26 +02:00
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
2010-09-24 16:07:23 +02:00
jz .cycle%1end
mov r0d, dword [r1+%1*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
2010-09-14 15:36:26 +02:00
%ifdef ARCH_X86_64
add r0, r10
%else
add r0, r0m
%endif
call h264_idct_dc_add8_mmx2
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
2010-09-24 16:07:23 +02:00
.cycle%1end
%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
2010-09-14 15:36:26 +02:00
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
2010-09-24 16:07:23 +02:00
%endif
%endmacro
; ff_h264_idct_add16intra_sse2(uint8_t *dst, const int *block_offset,
; DCTELEM *block, int stride, const uint8_t nnzc[6*8])
cglobal h264_idct_add16intra_sse2, 5, 7, 8
%ifdef ARCH_X86_64
mov r10, r0
%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
2010-09-14 15:36:26 +02:00
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
2010-09-24 16:05:45 +02:00
%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
2010-09-14 15:36:26 +02:00
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
2010-09-24 16:05:45 +02:00
jz .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
2010-09-14 15:36:26 +02:00
%ifdef ARCH_X86_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
2010-09-24 16:05:45 +02:00
mov r0d, dword [r1+%1*8+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
2010-09-14 15:36:26 +02:00
add r0, [r10]
%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
2010-09-24 16:05:45 +02:00
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
2010-09-14 15:36:26 +02:00
mov 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
2010-09-24 16:05:45 +02:00
add r0, dword [r1+%1*8+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
2010-09-14 15:36:26 +02:00
%endif
call x264_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
2010-09-24 16:05:45 +02:00
jmp .cycle%1end
.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
2010-09-14 15:36:26 +02:00
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
2010-09-24 16:05:45 +02:00
jz .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
2010-09-14 15:36:26 +02:00
%ifdef ARCH_X86_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
2010-09-24 16:05:45 +02:00
mov r0d, dword [r1+%1*8+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
2010-09-14 15:36:26 +02:00
add r0, [r10]
%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
2010-09-24 16:05:45 +02:00
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
2010-09-14 15:36:26 +02:00
mov 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
2010-09-24 16:05:45 +02:00
add r0, dword [r1+%1*8+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
2010-09-14 15:36:26 +02:00
%endif
call h264_idct_dc_add8_mmx2
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
2010-09-24 16:05:45 +02:00
.cycle%1end
%if %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
2010-09-14 15:36:26 +02:00
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
2010-09-24 16:05:45 +02:00
%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
2010-09-14 15:36:26 +02:00
; ff_h264_idct_add8_sse2(uint8_t **dest, const int *block_offset,
; DCTELEM *block, int stride, const uint8_t nnzc[6*8])
cglobal h264_idct_add8_sse2, 5, 7, 8
add r2, 512
%ifdef ARCH_X86_64
mov r10, r0
%endif
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
2010-09-24 16:05:45 +02:00
add8_sse2_cycle 0, 0x09
add8_sse2_cycle 1, 0x11
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
2010-09-14 15:36:26 +02:00
%ifdef ARCH_X86_64
add r10, gprsize
%else
add r0mp, gprsize
%endif
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
2010-09-24 16:05:45 +02:00
add8_sse2_cycle 2, 0x21
add8_sse2_cycle 3, 0x29
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
2010-09-14 15:36:26 +02:00
RET
Reference in New Issue Copy Permalink