ffmpeg/libavcodec/x86/vc1dsp_mc.asm
Lynne bbe95f7353
x86: replace explicit REP_RETs with RETs
From x86inc:
> On AMD cpus <=K10, an ordinary ret is slow if it immediately follows either
> a branch or a branch target. So switch to a 2-byte form of ret in that case.
> We can automatically detect "follows a branch", but not a branch target.
> (SSSE3 is a sufficient condition to know that your cpu doesn't have this problem.)

x86inc can automatically determine whether to use REP_RET rather than
REP in most of these cases, so impact is minimal. Additionally, a few
REP_RETs were used unnecessary, despite the return being nowhere near a
branch.

The only CPUs affected were AMD K10s, made between 2007 and 2011, 16
years ago and 12 years ago, respectively.

In the future, everyone involved with x86inc should consider dropping
REP_RETs altogether.
2023-02-01 04:23:55 +01:00

293 lines
8.6 KiB
NASM

;******************************************************************************
;* VC1 motion compensation optimizations
;* Copyright (c) 2007 Christophe GISQUET <christophe.gisquet@free.fr>
;*
;* 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
;* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
;******************************************************************************
%include "libavutil/x86/x86util.asm"
cextern pw_9
cextern pw_128
SECTION .text
%if HAVE_MMX_INLINE
; XXX some of these macros are not used right now, but they will in the future
; when more functions are ported.
%macro OP_PUT 2 ; dst, src
%endmacro
%macro OP_AVG 2 ; dst, src
pavgb %1, %2
%endmacro
%macro NORMALIZE_MMX 1 ; shift
paddw m3, m7 ; +bias-r
paddw m4, m7 ; +bias-r
psraw m3, %1
psraw m4, %1
%endmacro
%macro TRANSFER_DO_PACK 2 ; op, dst
packuswb m3, m4
%1 m3, [%2]
mova [%2], m3
%endmacro
%macro TRANSFER_DONT_PACK 2 ; op, dst
%1 m3, [%2]
%1 m3, [%2 + mmsize]
mova [%2], m3
mova [mmsize + %2], m4
%endmacro
; see MSPEL_FILTER13_CORE for use as UNPACK macro
%macro DO_UNPACK 1 ; reg
punpcklbw %1, m0
%endmacro
%macro DONT_UNPACK 1 ; reg
%endmacro
; Compute the rounder 32-r or 8-r and unpacks it to m7
%macro LOAD_ROUNDER_MMX 1 ; round
movd m7, %1
punpcklwd m7, m7
punpckldq m7, m7
%endmacro
%macro SHIFT2_LINE 5 ; off, r0, r1, r2, r3
paddw m%3, m%4
movh m%2, [srcq + stride_neg2]
pmullw m%3, m6
punpcklbw m%2, m0
movh m%5, [srcq + strideq]
psubw m%3, m%2
punpcklbw m%5, m0
paddw m%3, m7
psubw m%3, m%5
psraw m%3, shift
movu [dstq + %1], m%3
add srcq, strideq
%endmacro
INIT_MMX mmx
; void ff_vc1_put_ver_16b_shift2_mmx(int16_t *dst, const uint8_t *src,
; x86_reg stride, int rnd, int64_t shift)
; Sacrificing m6 makes it possible to pipeline loads from src
%if ARCH_X86_32
cglobal vc1_put_ver_16b_shift2, 3,6,0, dst, src, stride
DECLARE_REG_TMP 3, 4, 5
%define rnd r3mp
%define shift qword r4m
%else ; X86_64
cglobal vc1_put_ver_16b_shift2, 4,7,0, dst, src, stride
DECLARE_REG_TMP 4, 5, 6
%define rnd r3d
; We need shift either in memory or in a mm reg as it's used in psraw
; On WIN64, the arg is already on the stack
; On UNIX64, m5 doesn't seem to be used
%if WIN64
%define shift r4mp
%else ; UNIX64
%define shift m5
mova shift, r4q
%endif ; WIN64
%endif ; X86_32
%define stride_neg2 t0q
%define stride_9minus4 t1q
%define i t2q
mov stride_neg2, strideq
neg stride_neg2
add stride_neg2, stride_neg2
lea stride_9minus4, [strideq * 9 - 4]
mov i, 3
LOAD_ROUNDER_MMX rnd
mova m6, [pw_9]
pxor m0, m0
.loop:
movh m2, [srcq]
add srcq, strideq
movh m3, [srcq]
punpcklbw m2, m0
punpcklbw m3, m0
SHIFT2_LINE 0, 1, 2, 3, 4
SHIFT2_LINE 24, 2, 3, 4, 1
SHIFT2_LINE 48, 3, 4, 1, 2
SHIFT2_LINE 72, 4, 1, 2, 3
SHIFT2_LINE 96, 1, 2, 3, 4
SHIFT2_LINE 120, 2, 3, 4, 1
SHIFT2_LINE 144, 3, 4, 1, 2
SHIFT2_LINE 168, 4, 1, 2, 3
sub srcq, stride_9minus4
add dstq, 8
dec i
jnz .loop
RET
%undef rnd
%undef shift
%undef stride_neg2
%undef stride_9minus4
%undef i
; void ff_vc1_*_hor_16b_shift2_mmx(uint8_t *dst, x86_reg stride,
; const int16_t *src, int rnd);
; Data is already unpacked, so some operations can directly be made from
; memory.
%macro HOR_16B_SHIFT2 2 ; op, opname
cglobal vc1_%2_hor_16b_shift2, 4, 5, 0, dst, stride, src, rnd, h
mov hq, 8
sub srcq, 2
sub rndd, (-1+9+9-1) * 1024 ; add -1024 bias
LOAD_ROUNDER_MMX rndd
mova m5, [pw_9]
mova m6, [pw_128]
pxor m0, m0
.loop:
mova m1, [srcq + 2 * 0]
mova m2, [srcq + 2 * 0 + mmsize]
mova m3, [srcq + 2 * 1]
mova m4, [srcq + 2 * 1 + mmsize]
paddw m3, [srcq + 2 * 2]
paddw m4, [srcq + 2 * 2 + mmsize]
paddw m1, [srcq + 2 * 3]
paddw m2, [srcq + 2 * 3 + mmsize]
pmullw m3, m5
pmullw m4, m5
psubw m3, m1
psubw m4, m2
NORMALIZE_MMX 7
; remove bias
paddw m3, m6
paddw m4, m6
TRANSFER_DO_PACK %1, dstq
add srcq, 24
add dstq, strideq
dec hq
jnz .loop
RET
%endmacro
INIT_MMX mmx
HOR_16B_SHIFT2 OP_PUT, put
INIT_MMX mmxext
HOR_16B_SHIFT2 OP_AVG, avg
%endif ; HAVE_MMX_INLINE
%macro INV_TRANS_INIT 0
movsxdifnidn linesizeq, linesized
movd m0, blockd
SPLATW m0, m0
pxor m1, m1
psubw m1, m0
packuswb m0, m0
packuswb m1, m1
DEFINE_ARGS dest, linesize, linesize3
lea linesize3q, [linesizeq*3]
%endmacro
%macro INV_TRANS_PROCESS 1
mov%1 m2, [destq+linesizeq*0]
mov%1 m3, [destq+linesizeq*1]
mov%1 m4, [destq+linesizeq*2]
mov%1 m5, [destq+linesize3q]
paddusb m2, m0
paddusb m3, m0
paddusb m4, m0
paddusb m5, m0
psubusb m2, m1
psubusb m3, m1
psubusb m4, m1
psubusb m5, m1
mov%1 [linesizeq*0+destq], m2
mov%1 [linesizeq*1+destq], m3
mov%1 [linesizeq*2+destq], m4
mov%1 [linesize3q +destq], m5
%endmacro
; ff_vc1_inv_trans_?x?_dc_mmxext(uint8_t *dest, ptrdiff_t linesize, int16_t *block)
INIT_MMX mmxext
cglobal vc1_inv_trans_4x4_dc, 3,4,0, dest, linesize, block
movsx r3d, WORD [blockq]
mov blockd, r3d ; dc
shl blockd, 4 ; 16 * dc
lea blockd, [blockq+r3+4] ; 17 * dc + 4
sar blockd, 3 ; >> 3
mov r3d, blockd ; dc
shl blockd, 4 ; 16 * dc
lea blockd, [blockq+r3+64] ; 17 * dc + 64
sar blockd, 7 ; >> 7
INV_TRANS_INIT
INV_TRANS_PROCESS h
RET
INIT_MMX mmxext
cglobal vc1_inv_trans_4x8_dc, 3,4,0, dest, linesize, block
movsx r3d, WORD [blockq]
mov blockd, r3d ; dc
shl blockd, 4 ; 16 * dc
lea blockd, [blockq+r3+4] ; 17 * dc + 4
sar blockd, 3 ; >> 3
shl blockd, 2 ; 4 * dc
lea blockd, [blockq*3+64] ; 12 * dc + 64
sar blockd, 7 ; >> 7
INV_TRANS_INIT
INV_TRANS_PROCESS h
lea destq, [destq+linesizeq*4]
INV_TRANS_PROCESS h
RET
INIT_MMX mmxext
cglobal vc1_inv_trans_8x4_dc, 3,4,0, dest, linesize, block
movsx blockd, WORD [blockq] ; dc
lea blockd, [blockq*3+1] ; 3 * dc + 1
sar blockd, 1 ; >> 1
mov r3d, blockd ; dc
shl blockd, 4 ; 16 * dc
lea blockd, [blockq+r3+64] ; 17 * dc + 64
sar blockd, 7 ; >> 7
INV_TRANS_INIT
INV_TRANS_PROCESS a
RET
INIT_MMX mmxext
cglobal vc1_inv_trans_8x8_dc, 3,3,0, dest, linesize, block
movsx blockd, WORD [blockq] ; dc
lea blockd, [blockq*3+1] ; 3 * dc + 1
sar blockd, 1 ; >> 1
lea blockd, [blockq*3+16] ; 3 * dc + 16
sar blockd, 5 ; >> 5
INV_TRANS_INIT
INV_TRANS_PROCESS a
lea destq, [destq+linesizeq*4]
INV_TRANS_PROCESS a
RET