From 2999bd7da2cf6343592d4eb6da3cd0748354775d Mon Sep 17 00:00:00 2001 From: Christophe Gisquet Date: Tue, 10 Mar 2015 23:11:51 +0000 Subject: x86: xvid_idct: port SSE2 iDCT to yasm The main difference consists in renaming properly labels, and letting yasm select the gprs for skipping 1D transforms. Previous-version-reviewed-by: James Almer Signed-off-by: Michael Niedermayer --- libavcodec/x86/Makefile | 4 +- libavcodec/x86/dct-test.c | 4 +- libavcodec/x86/xvididct.asm | 379 ++++++++++++++++++++++++++++++++++++++ libavcodec/x86/xvididct_init.c | 18 +- libavcodec/x86/xvididct_sse2.c | 406 ----------------------------------------- 5 files changed, 398 insertions(+), 413 deletions(-) create mode 100644 libavcodec/x86/xvididct.asm delete mode 100644 libavcodec/x86/xvididct_sse2.c diff --git a/libavcodec/x86/Makefile b/libavcodec/x86/Makefile index 6b9164ab12..f46c7d572f 100644 --- a/libavcodec/x86/Makefile +++ b/libavcodec/x86/Makefile @@ -73,8 +73,7 @@ MMX-OBJS-$(CONFIG_FDCTDSP) += x86/fdct.o MMX-OBJS-$(CONFIG_IDCTDSP) += x86/simple_idct.o # decoders/encoders -MMX-OBJS-$(CONFIG_MPEG4_DECODER) += x86/xvididct_mmx.o \ - x86/xvididct_sse2.o +MMX-OBJS-$(CONFIG_MPEG4_DECODER) += x86/xvididct_mmx.o MMX-OBJS-$(CONFIG_SNOW_DECODER) += x86/snowdsp.o MMX-OBJS-$(CONFIG_SNOW_ENCODER) += x86/snowdsp.o MMX-OBJS-$(CONFIG_VC1_DECODER) += x86/vc1dsp_mmx.o @@ -141,6 +140,7 @@ YASM-OBJS-$(CONFIG_HEVC_DECODER) += x86/hevc_mc.o \ x86/hevc_res_add.o \ x86/hevc_sao.o YASM-OBJS-$(CONFIG_MLP_DECODER) += x86/mlpdsp.o +YASM-OBJS-$(CONFIG_MPEG4_DECODER) += x86/xvididct.o YASM-OBJS-$(CONFIG_PNG_DECODER) += x86/pngdsp.o YASM-OBJS-$(CONFIG_PRORES_DECODER) += x86/proresdsp.o YASM-OBJS-$(CONFIG_PRORES_LGPL_DECODER) += x86/proresdsp.o diff --git a/libavcodec/x86/dct-test.c b/libavcodec/x86/dct-test.c index 3414cb0cb0..e14ce9aac3 100644 --- a/libavcodec/x86/dct-test.c +++ b/libavcodec/x86/dct-test.c @@ -67,9 +67,9 @@ static const struct algo idct_tab_arch[] = { #if HAVE_MMXEXT_INLINE { "XVID-MMXEXT", ff_xvid_idct_mmxext, FF_IDCT_PERM_NONE, AV_CPU_FLAG_MMXEXT, 1 }, #endif -#if HAVE_SSE2_INLINE +#if HAVE_SSE2_EXTERNAL { "XVID-SSE2", ff_xvid_idct_sse2, FF_IDCT_PERM_SSE2, AV_CPU_FLAG_SSE2, 1 }, -#if ARCH_X86_64 && HAVE_YASM +#if ARCH_X86_64 { "PR-SSE2", ff_prores_idct_put_10_sse2_wrap, FF_IDCT_PERM_TRANSPOSE, AV_CPU_FLAG_SSE2, 1 }, #endif #endif diff --git a/libavcodec/x86/xvididct.asm b/libavcodec/x86/xvididct.asm new file mode 100644 index 0000000000..d16db345a4 --- /dev/null +++ b/libavcodec/x86/xvididct.asm @@ -0,0 +1,379 @@ +; XVID MPEG-4 VIDEO CODEC +; - SSE2 inverse discrete cosine transform - +; +; Copyright(C) 2003 Pascal Massimino +; +; Conversion to gcc syntax with modifications +; by Alexander Strange +; +; Originally from dct/x86_asm/fdct_sse2_skal.asm in Xvid. +; +; This file is part of FFmpeg. +; +; Vertical pass is an implementation of the scheme: +; Loeffler C., Ligtenberg A., and Moschytz C.S.: +; Practical Fast 1D DCT Algorithm with Eleven Multiplications, +; Proc. ICASSP 1989, 988-991. +; +; Horizontal pass is a double 4x4 vector/matrix multiplication, +; (see also Intel's Application Note 922: +; http://developer.intel.com/vtune/cbts/strmsimd/922down.htm +; Copyright (C) 1999 Intel Corporation) +; +; More details at http://skal.planet-d.net/coding/dct.html +; +; 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" + +SECTION_RODATA +tan1: times 8 dw 13036 +tan2: times 8 dw 27146 +tan3: times 8 dw 43790 +sqrt2: times 8 dw 23170 + +iTab1: dw 0x4000, 0x539f, 0xc000, 0xac61, 0x4000, 0xdd5d, 0x4000, 0xdd5d + dw 0x4000, 0x22a3, 0x4000, 0x22a3, 0xc000, 0x539f, 0x4000, 0xac61 + dw 0x3249, 0x11a8, 0x4b42, 0xee58, 0x11a8, 0x4b42, 0x11a8, 0xcdb7 + dw 0x58c5, 0x4b42, 0xa73b, 0xcdb7, 0x3249, 0xa73b, 0x4b42, 0xa73b +iTab2: dw 0x58c5, 0x73fc, 0xa73b, 0x8c04, 0x58c5, 0xcff5, 0x58c5, 0xcff5 + dw 0x58c5, 0x300b, 0x58c5, 0x300b, 0xa73b, 0x73fc, 0x58c5, 0x8c04 + dw 0x45bf, 0x187e, 0x6862, 0xe782, 0x187e, 0x6862, 0x187e, 0xba41 + dw 0x7b21, 0x6862, 0x84df, 0xba41, 0x45bf, 0x84df, 0x6862, 0x84df +iTab3: dw 0x539f, 0x6d41, 0xac61, 0x92bf, 0x539f, 0xd2bf, 0x539f, 0xd2bf + dw 0x539f, 0x2d41, 0x539f, 0x2d41, 0xac61, 0x6d41, 0x539f, 0x92bf + dw 0x41b3, 0x1712, 0x6254, 0xe8ee, 0x1712, 0x6254, 0x1712, 0xbe4d + dw 0x73fc, 0x6254, 0x8c04, 0xbe4d, 0x41b3, 0x8c04, 0x6254, 0x8c04 +iTab4: dw 0x4b42, 0x6254, 0xb4be, 0x9dac, 0x4b42, 0xd746, 0x4b42, 0xd746 + dw 0x4b42, 0x28ba, 0x4b42, 0x28ba, 0xb4be, 0x6254, 0x4b42, 0x9dac + dw 0x3b21, 0x14c3, 0x587e, 0xeb3d, 0x14c3, 0x587e, 0x14c3, 0xc4df + dw 0x6862, 0x587e, 0x979e, 0xc4df, 0x3b21, 0x979e, 0x587e, 0x979e + +walkenIdctRounders: times 4 dd 65536 + times 4 dd 3597 + times 4 dd 2260 + times 4 dd 1203 + times 4 dd 120 + times 4 dd 512 + +pb_127: times 8 db 127 + +SECTION .text + +; Temporary storage before the column pass +%define ROW1 xmm6 +%define ROW3 xmm4 +%define ROW5 xmm5 +%define ROW7 xmm7 + +%macro CLEAR_ODD 1 + pxor %1, %1 +%endmacro +%macro PUT_ODD 1 + pshufhw %1, xmm2, 0x1B +%endmacro + +%macro MOV32 2 +%if ARCH_X86_32 + movdqa %2, %1 +%endif +%endmacro + +%macro CLEAR_EVEN 1 +%if ARCH_X86_64 + CLEAR_ODD %1 +%endif +%endmacro + +%macro PUT_EVEN 1 +%if ARCH_X86_64 + PUT_ODD %1 +%else + pshufhw xmm2, xmm2, 0x1B + movdqa %1, xmm2 +%endif +%endmacro + +%if ARCH_X86_64 +%define ROW0 xmm8 +%define REG0 ROW0 +%define ROW2 xmm9 +%define REG2 ROW2 +%define ROW4 xmm10 +%define REG4 ROW4 +%define ROW6 xmm11 +%define REG6 ROW6 +%define XMMS xmm12 +%define SREG2 REG2 +%define TAN3 xmm13 +%define TAN1 xmm14 +%else +%define ROW0 [r0 + 0*16] +%define REG0 xmm4 +%define ROW2 [r0 + 2*16] +%define REG2 xmm4 +%define ROW4 [r0 + 4*16] +%define REG4 xmm6 +%define ROW6 [r0 + 6*16] +%define REG6 xmm6 +%define XMMS xmm2 +%define SREG2 xmm7 +%define TAN3 xmm0 +%define TAN1 xmm2 +%endif + +%macro JZ 2 + test %1, %1 + jz .%2 +%endmacro + +%macro JNZ 2 + test %1, %1 + jnz .%2 +%endmacro + +%macro TEST_ONE_ROW 4 ; src, reg, clear, arg + %3 %4 + movq mm1, [%1] + por mm1, [%1 + 8] + paddusb mm1, mm0 + pmovmskb %2, mm1 +%endmacro + +;row1, row2, reg1, reg2, clear1, arg1, clear2, arg2 +%macro TEST_TWO_ROWS 8 + %5 %6 + %7 %8 + movq mm1, [%1 + 0] + por mm1, [%1 + 8] + movq mm2, [%2 + 0] + por mm2, [%2 + 8] + paddusb mm1, mm0 + paddusb mm2, mm0 + pmovmskb %3, mm1 + pmovmskb %4, mm2 +%endmacro + +; IDCT pass on rows. +%macro iMTX_MULT 4-5 ; src, table, put, arg, rounder + movdqa xmm3, [%1] + movdqa xmm0, xmm3 + pshufd xmm1, xmm3, 0x11 ; 4602 + punpcklqdq xmm0, xmm0 ; 0246 + pmaddwd xmm0, [%2] + pmaddwd xmm1, [%2+16] + pshufd xmm2, xmm3, 0xBB ; 5713 + punpckhqdq xmm3, xmm3 ; 1357 + pmaddwd xmm2, [%2+32] + pmaddwd xmm3, [%2+48] + paddd xmm0, xmm1 + paddd xmm2, xmm3 +%if %0 == 5 + paddd xmm0, [walkenIdctRounders+%5] +%endif + movdqa xmm3, xmm2 + paddd xmm2, xmm0 + psubd xmm0, xmm3 + psrad xmm2, 11 + psrad xmm0, 11 + packssdw xmm2, xmm0 + %3 %4 +%endmacro + +%macro iLLM_HEAD 0 + movdqa TAN3, [tan3] + movdqa TAN1, [tan1] +%endmacro + +; IDCT pass on columns. +%macro iLLM_PASS 1 ;dct + movdqa xmm1, TAN3 + movdqa xmm3, TAN1 + pmulhw TAN3, xmm4 + pmulhw xmm1, xmm5 + paddsw TAN3, xmm4 + paddsw xmm1, xmm5 + psubsw TAN3, xmm5 + paddsw xmm1, xmm4 + pmulhw xmm3, xmm7 + pmulhw TAN1, xmm6 + paddsw xmm3, xmm6 + psubsw TAN1, xmm7 + movdqa xmm7, xmm3 + movdqa xmm6, TAN1 + psubsw xmm3, xmm1 + psubsw TAN1, TAN3 + paddsw xmm1, xmm7 + paddsw TAN3, xmm6 + movdqa xmm6, xmm3 + psubsw xmm3, TAN3 + paddsw TAN3, xmm6 + movdqa xmm4, [sqrt2] + pmulhw xmm3, xmm4 + pmulhw TAN3, xmm4 + paddsw TAN3, TAN3 + paddsw xmm3, xmm3 + movdqa xmm7, [tan2] + MOV32 ROW2, REG2 + MOV32 ROW6, REG6 + movdqa xmm5, xmm7 + pmulhw xmm7, REG6 + pmulhw xmm5, REG2 + paddsw xmm7, REG2 + psubsw xmm5, REG6 + MOV32 ROW0, REG0 + MOV32 ROW4, REG4 + MOV32 TAN1, [r0] + movdqa XMMS, REG0 + psubsw REG0, REG4 + paddsw REG4, XMMS + movdqa XMMS, REG4 + psubsw REG4, xmm7 + paddsw xmm7, XMMS + movdqa XMMS, REG0 + psubsw REG0, xmm5 + paddsw xmm5, XMMS + movdqa XMMS, xmm5 + psubsw xmm5, TAN3 + paddsw TAN3, XMMS + movdqa XMMS, REG0 + psubsw REG0, xmm3 + paddsw xmm3, XMMS + MOV32 [r0], TAN1 + psraw xmm5, 6 + psraw REG0, 6 + psraw TAN3, 6 + psraw xmm3, 6 + movdqa [%1+1*16], TAN3 + movdqa [%1+2*16], xmm3 + movdqa [%1+5*16], REG0 + movdqa [%1+6*16], xmm5 + movdqa xmm0, xmm7 + movdqa xmm4, REG4 + psubsw xmm7, xmm1 + psubsw REG4, TAN1 + paddsw xmm1, xmm0 + paddsw TAN1, xmm4 + psraw xmm1, 6 + psraw xmm7, 6 + psraw TAN1, 6 + psraw REG4, 6 + movdqa [%1+0*16], xmm1 + movdqa [%1+3*16], TAN1 + movdqa [%1+4*16], REG4 + movdqa [%1+7*16], xmm7 +%endmacro + +; IDCT pass on columns, assuming rows 4-7 are zero +%macro iLLM_PASS_SPARSE 1 ;dct + pmulhw TAN3, xmm4 + paddsw TAN3, xmm4 + movdqa xmm3, xmm6 + pmulhw TAN1, xmm6 + movdqa xmm1, xmm4 + psubsw xmm3, xmm1 + paddsw xmm1, xmm6 + movdqa xmm6, TAN1 + psubsw TAN1, TAN3 + paddsw TAN3, xmm6 + movdqa xmm6, xmm3 + psubsw xmm3, TAN3 + paddsw TAN3, xmm6 + movdqa xmm4, [sqrt2] + pmulhw xmm3, xmm4 + pmulhw TAN3, xmm4 + paddsw TAN3, TAN3 + paddsw xmm3, xmm3 + movdqa xmm5, [tan2] + MOV32 ROW2, SREG2 + pmulhw xmm5, SREG2 + MOV32 ROW0, REG0 + movdqa xmm6, REG0 + psubsw xmm6, SREG2 + paddsw SREG2, REG0 + MOV32 TAN1, [r0] + movdqa XMMS, REG0 + psubsw REG0, xmm5 + paddsw xmm5, XMMS + movdqa XMMS, xmm5 + psubsw xmm5, TAN3 + paddsw TAN3, XMMS + movdqa XMMS, REG0 + psubsw REG0, xmm3 + paddsw xmm3, XMMS + MOV32 [r0], TAN1 + psraw xmm5, 6 + psraw REG0, 6 + psraw TAN3, 6 + psraw xmm3, 6 + movdqa [%1+1*16], TAN3 + movdqa [%1+2*16], xmm3 + movdqa [%1+5*16], REG0 + movdqa [%1+6*16], xmm5 + movdqa xmm0, SREG2 + movdqa xmm4, xmm6 + psubsw SREG2, xmm1 + psubsw xmm6, TAN1 + paddsw xmm1, xmm0 + paddsw TAN1, xmm4 + psraw xmm1, 6 + psraw SREG2, 6 + psraw TAN1, 6 + psraw xmm6, 6 + movdqa [%1+0*16], xmm1 + movdqa [%1+3*16], TAN1 + movdqa [%1+4*16], xmm6 + movdqa [%1+7*16], SREG2 +%endmacro + +INIT_XMM sse2 +cglobal xvid_idct, 1, 5, 8+7*ARCH_X86_64, block + movq mm0, [pb_127] + iMTX_MULT r0 + 0*16, iTab1, PUT_EVEN, ROW0, 0*16 + iMTX_MULT r0 + 1*16, iTab2, PUT_ODD, ROW1, 1*16 + iMTX_MULT r0 + 2*16, iTab3, PUT_EVEN, ROW2, 2*16 + + TEST_TWO_ROWS r0 + 3*16, r0 + 4*16, r1d, r2d, CLEAR_ODD, ROW3, CLEAR_EVEN, ROW4 ; a, c + JZ r1d, col1 + iMTX_MULT r0 + 3*16, iTab4, PUT_ODD, ROW3, 3*16 +.col1: + TEST_TWO_ROWS r0 + 5*16, r0 + 6*16, r1d, r3d, CLEAR_ODD, ROW5, CLEAR_EVEN, ROW6 ; a, d + TEST_ONE_ROW r0 + 7*16, r4d, CLEAR_ODD, ROW7 ; esi + + iLLM_HEAD + JNZ r2d, 2 + JNZ r1d, 3 + JNZ r3d, 4 + JNZ r4d, 5 + iLLM_PASS_SPARSE r0 + jmp .6 +.2: + iMTX_MULT r0 + 4*16, iTab1, PUT_EVEN, ROW4 +.3: + iMTX_MULT r0 + 5*16, iTab4, PUT_ODD, ROW5, 4*16 + JZ r3d, col2 +.4: + iMTX_MULT r0 + 6*16, iTab3, PUT_EVEN, ROW6, 5*16 +.col2: + JZ r4d, col3 +.5: + iMTX_MULT r0 + 7*16, iTab2, PUT_ODD, ROW7, 5*16 +.col3: +%if ARCH_X86_32 + iLLM_HEAD +%endif + iLLM_PASS r0 +.6: + RET diff --git a/libavcodec/x86/xvididct_init.c b/libavcodec/x86/xvididct_init.c index 8c227844b5..ae65667a31 100644 --- a/libavcodec/x86/xvididct_init.c +++ b/libavcodec/x86/xvididct_init.c @@ -26,6 +26,18 @@ #include "idctdsp.h" #include "xvididct.h" +static void xvid_idct_sse2_put(uint8_t *dest, int line_size, short *block) +{ + ff_xvid_idct_sse2(block); + ff_put_pixels_clamped(block, dest, line_size); +} + +static void xvid_idct_sse2_add(uint8_t *dest, int line_size, short *block) +{ + ff_xvid_idct_sse2(block); + ff_add_pixels_clamped(block, dest, line_size); +} + av_cold void ff_xvid_idct_init_x86(IDCTDSPContext *c, AVCodecContext *avctx, unsigned high_bit_depth) { @@ -50,9 +62,9 @@ av_cold void ff_xvid_idct_init_x86(IDCTDSPContext *c, AVCodecContext *avctx, c->perm_type = FF_IDCT_PERM_NONE; } - if (INLINE_SSE2(cpu_flags)) { - c->idct_put = ff_xvid_idct_sse2_put; - c->idct_add = ff_xvid_idct_sse2_add; + if (EXTERNAL_SSE2(cpu_flags)) { + c->idct_put = xvid_idct_sse2_put; + c->idct_add = xvid_idct_sse2_add; c->idct = ff_xvid_idct_sse2; c->perm_type = FF_IDCT_PERM_SSE2; } diff --git a/libavcodec/x86/xvididct_sse2.c b/libavcodec/x86/xvididct_sse2.c deleted file mode 100644 index 51a5d9d6b0..0000000000 --- a/libavcodec/x86/xvididct_sse2.c +++ /dev/null @@ -1,406 +0,0 @@ -/* - * XVID MPEG-4 VIDEO CODEC - * - SSE2 inverse discrete cosine transform - - * - * Copyright(C) 2003 Pascal Massimino - * - * Conversion to gcc syntax with modifications - * by Alexander Strange - * - * Originally from dct/x86_asm/fdct_sse2_skal.asm in Xvid. - * - * This file is part of FFmpeg. - * - * Vertical pass is an implementation of the scheme: - * Loeffler C., Ligtenberg A., and Moschytz C.S.: - * Practical Fast 1D DCT Algorithm with Eleven Multiplications, - * Proc. ICASSP 1989, 988-991. - * - * Horizontal pass is a double 4x4 vector/matrix multiplication, - * (see also Intel's Application Note 922: - * http://developer.intel.com/vtune/cbts/strmsimd/922down.htm - * Copyright (C) 1999 Intel Corporation) - * - * More details at http://skal.planet-d.net/coding/dct.html - * - * 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/mem.h" -#include "libavutil/x86/asm.h" - -#include "libavcodec/idctdsp.h" - -#include "idctdsp.h" -#include "xvididct.h" - -#if HAVE_SSE2_INLINE - -/** - * @file - * @brief SSE2 IDCT compatible with the Xvid IDCT - */ - -#define X8(x) x, x, x, x, x, x, x, x - -DECLARE_ASM_CONST(16, int16_t, tan1)[] = { X8(13036) }; // tan( pi/16) -DECLARE_ASM_CONST(16, int16_t, tan2)[] = { X8(27146) }; // tan(2pi/16) = sqrt(2)-1 -DECLARE_ASM_CONST(16, int16_t, tan3)[] = { X8(43790) }; // tan(3pi/16)-1 -DECLARE_ASM_CONST(16, int16_t, sqrt2)[] = { X8(23170) }; // 0.5/sqrt(2) -DECLARE_ASM_CONST(8, uint8_t, m127)[] = { X8(127) }; - -DECLARE_ASM_CONST(16, int16_t, iTab1)[] = { - 0x4000, 0x539f, 0xc000, 0xac61, 0x4000, 0xdd5d, 0x4000, 0xdd5d, - 0x4000, 0x22a3, 0x4000, 0x22a3, 0xc000, 0x539f, 0x4000, 0xac61, - 0x3249, 0x11a8, 0x4b42, 0xee58, 0x11a8, 0x4b42, 0x11a8, 0xcdb7, - 0x58c5, 0x4b42, 0xa73b, 0xcdb7, 0x3249, 0xa73b, 0x4b42, 0xa73b -}; - -DECLARE_ASM_CONST(16, int16_t, iTab2)[] = { - 0x58c5, 0x73fc, 0xa73b, 0x8c04, 0x58c5, 0xcff5, 0x58c5, 0xcff5, - 0x58c5, 0x300b, 0x58c5, 0x300b, 0xa73b, 0x73fc, 0x58c5, 0x8c04, - 0x45bf, 0x187e, 0x6862, 0xe782, 0x187e, 0x6862, 0x187e, 0xba41, - 0x7b21, 0x6862, 0x84df, 0xba41, 0x45bf, 0x84df, 0x6862, 0x84df -}; - -DECLARE_ASM_CONST(16, int16_t, iTab3)[] = { - 0x539f, 0x6d41, 0xac61, 0x92bf, 0x539f, 0xd2bf, 0x539f, 0xd2bf, - 0x539f, 0x2d41, 0x539f, 0x2d41, 0xac61, 0x6d41, 0x539f, 0x92bf, - 0x41b3, 0x1712, 0x6254, 0xe8ee, 0x1712, 0x6254, 0x1712, 0xbe4d, - 0x73fc, 0x6254, 0x8c04, 0xbe4d, 0x41b3, 0x8c04, 0x6254, 0x8c04 -}; - -DECLARE_ASM_CONST(16, int16_t, iTab4)[] = { - 0x4b42, 0x6254, 0xb4be, 0x9dac, 0x4b42, 0xd746, 0x4b42, 0xd746, - 0x4b42, 0x28ba, 0x4b42, 0x28ba, 0xb4be, 0x6254, 0x4b42, 0x9dac, - 0x3b21, 0x14c3, 0x587e, 0xeb3d, 0x14c3, 0x587e, 0x14c3, 0xc4df, - 0x6862, 0x587e, 0x979e, 0xc4df, 0x3b21, 0x979e, 0x587e, 0x979e -}; - -DECLARE_ASM_CONST(16, int32_t, walkenIdctRounders)[] = { - 65536, 65536, 65536, 65536, - 3597, 3597, 3597, 3597, - 2260, 2260, 2260, 2260, - 1203, 1203, 1203, 1203, - 120, 120, 120, 120, - 512, 512, 512, 512 -}; - -// Temporary storage before the column pass -#define ROW1 "%%xmm6" -#define ROW3 "%%xmm4" -#define ROW5 "%%xmm5" -#define ROW7 "%%xmm7" - -#define CLEAR_ODD(r) "pxor "r","r" \n\t" -#define PUT_ODD(dst) "pshufhw $0x1B, %%xmm2, "dst" \n\t" - -#if ARCH_X86_64 - -# define ROW0 "%%xmm8" -# define REG0 ROW0 -# define ROW2 "%%xmm9" -# define REG2 ROW2 -# define ROW4 "%%xmm10" -# define REG4 ROW4 -# define ROW6 "%%xmm11" -# define REG6 ROW6 -# define CLEAR_EVEN(r) CLEAR_ODD(r) -# define PUT_EVEN(dst) PUT_ODD(dst) -# define XMMS "%%xmm12" -# define MOV_32_ONLY "#" -# define SREG2 REG2 -# define TAN3 "%%xmm13" -# define TAN1 "%%xmm14" - -#else - -# define ROW0 "(%0)" -# define REG0 "%%xmm4" -# define ROW2 "2*16(%0)" -# define REG2 "%%xmm4" -# define ROW4 "4*16(%0)" -# define REG4 "%%xmm6" -# define ROW6 "6*16(%0)" -# define REG6 "%%xmm6" -# define CLEAR_EVEN(r) -# define PUT_EVEN(dst) \ - "pshufhw $0x1B, %%xmm2, %%xmm2 \n\t" \ - "movdqa %%xmm2, "dst" \n\t" -# define XMMS "%%xmm2" -# define MOV_32_ONLY "movdqa " -# define SREG2 "%%xmm7" -# define TAN3 "%%xmm0" -# define TAN1 "%%xmm2" - -#endif - -#define ROUND(x) "paddd "x - -#define JZ(reg, to) \ - "testl "reg","reg" \n\t" \ - "jz "to" \n\t" - -#define JNZ(reg, to) \ - "testl "reg","reg" \n\t" \ - "jnz "to" \n\t" - -#define TEST_ONE_ROW(src, reg, clear) \ - clear \ - "movq "src", %%mm1 \n\t" \ - "por 8+"src", %%mm1 \n\t" \ - "paddusb %%mm0, %%mm1 \n\t" \ - "pmovmskb %%mm1, "reg" \n\t" - -#define TEST_TWO_ROWS(row1, row2, reg1, reg2, clear1, clear2) \ - clear1 \ - clear2 \ - "movq "row1", %%mm1 \n\t" \ - "por 8+"row1", %%mm1 \n\t" \ - "movq "row2", %%mm2 \n\t" \ - "por 8+"row2", %%mm2 \n\t" \ - "paddusb %%mm0, %%mm1 \n\t" \ - "paddusb %%mm0, %%mm2 \n\t" \ - "pmovmskb %%mm1, "reg1" \n\t" \ - "pmovmskb %%mm2, "reg2" \n\t" - -/// IDCT pass on rows. -#define iMTX_MULT(src, table, rounder, put) \ - "movdqa "src", %%xmm3 \n\t" \ - "movdqa %%xmm3, %%xmm0 \n\t" \ - "pshufd $0x11, %%xmm3, %%xmm1 \n\t" /* 4602 */ \ - "punpcklqdq %%xmm0, %%xmm0 \n\t" /* 0246 */ \ - "pmaddwd "table", %%xmm0 \n\t" \ - "pmaddwd 16+"table", %%xmm1 \n\t" \ - "pshufd $0xBB, %%xmm3, %%xmm2 \n\t" /* 5713 */ \ - "punpckhqdq %%xmm3, %%xmm3 \n\t" /* 1357 */ \ - "pmaddwd 32+"table", %%xmm2 \n\t" \ - "pmaddwd 48+"table", %%xmm3 \n\t" \ - "paddd %%xmm1, %%xmm0 \n\t" \ - "paddd %%xmm3, %%xmm2 \n\t" \ - rounder", %%xmm0 \n\t" \ - "movdqa %%xmm2, %%xmm3 \n\t" \ - "paddd %%xmm0, %%xmm2 \n\t" \ - "psubd %%xmm3, %%xmm0 \n\t" \ - "psrad $11, %%xmm2 \n\t" \ - "psrad $11, %%xmm0 \n\t" \ - "packssdw %%xmm0, %%xmm2 \n\t" \ - put \ - "1: \n\t" - -#define iLLM_HEAD \ - "movdqa "MANGLE(tan3)", "TAN3" \n\t" \ - "movdqa "MANGLE(tan1)", "TAN1" \n\t" \ - -/// IDCT pass on columns. -#define iLLM_PASS(dct) \ - "movdqa "TAN3", %%xmm1 \n\t" \ - "movdqa "TAN1", %%xmm3 \n\t" \ - "pmulhw %%xmm4, "TAN3" \n\t" \ - "pmulhw %%xmm5, %%xmm1 \n\t" \ - "paddsw %%xmm4, "TAN3" \n\t" \ - "paddsw %%xmm5, %%xmm1 \n\t" \ - "psubsw %%xmm5, "TAN3" \n\t" \ - "paddsw %%xmm4, %%xmm1 \n\t" \ - "pmulhw %%xmm7, %%xmm3 \n\t" \ - "pmulhw %%xmm6, "TAN1" \n\t" \ - "paddsw %%xmm6, %%xmm3 \n\t" \ - "psubsw %%xmm7, "TAN1" \n\t" \ - "movdqa %%xmm3, %%xmm7 \n\t" \ - "movdqa "TAN1", %%xmm6 \n\t" \ - "psubsw %%xmm1, %%xmm3 \n\t" \ - "psubsw "TAN3", "TAN1" \n\t" \ - "paddsw %%xmm7, %%xmm1 \n\t" \ - "paddsw %%xmm6, "TAN3" \n\t" \ - "movdqa %%xmm3, %%xmm6 \n\t" \ - "psubsw "TAN3", %%xmm3 \n\t" \ - "paddsw %%xmm6, "TAN3" \n\t" \ - "movdqa "MANGLE(sqrt2)", %%xmm4 \n\t" \ - "pmulhw %%xmm4, %%xmm3 \n\t" \ - "pmulhw %%xmm4, "TAN3" \n\t" \ - "paddsw "TAN3", "TAN3" \n\t" \ - "paddsw %%xmm3, %%xmm3 \n\t" \ - "movdqa "MANGLE(tan2)", %%xmm7 \n\t" \ - MOV_32_ONLY ROW2", "REG2" \n\t" \ - MOV_32_ONLY ROW6", "REG6" \n\t" \ - "movdqa %%xmm7, %%xmm5 \n\t" \ - "pmulhw "REG6", %%xmm7 \n\t" \ - "pmulhw "REG2", %%xmm5 \n\t" \ - "paddsw "REG2", %%xmm7 \n\t" \ - "psubsw "REG6", %%xmm5 \n\t" \ - MOV_32_ONLY ROW0", "REG0" \n\t" \ - MOV_32_ONLY ROW4", "REG4" \n\t" \ - MOV_32_ONLY" "TAN1", (%0) \n\t" \ - "movdqa "REG0", "XMMS" \n\t" \ - "psubsw "REG4", "REG0" \n\t" \ - "paddsw "XMMS", "REG4" \n\t" \ - "movdqa "REG4", "XMMS" \n\t" \ - "psubsw %%xmm7, "REG4" \n\t" \ - "paddsw "XMMS", %%xmm7 \n\t" \ - "movdqa "REG0", "XMMS" \n\t" \ - "psubsw %%xmm5, "REG0" \n\t" \ - "paddsw "XMMS", %%xmm5 \n\t" \ - "movdqa %%xmm5, "XMMS" \n\t" \ - "psubsw "TAN3", %%xmm5 \n\t" \ - "paddsw "XMMS", "TAN3" \n\t" \ - "movdqa "REG0", "XMMS" \n\t" \ - "psubsw %%xmm3, "REG0" \n\t" \ - "paddsw "XMMS", %%xmm3 \n\t" \ - MOV_32_ONLY" (%0), "TAN1" \n\t" \ - "psraw $6, %%xmm5 \n\t" \ - "psraw $6, "REG0" \n\t" \ - "psraw $6, "TAN3" \n\t" \ - "psraw $6, %%xmm3 \n\t" \ - "movdqa "TAN3", 1*16("dct") \n\t" \ - "movdqa %%xmm3, 2*16("dct") \n\t" \ - "movdqa "REG0", 5*16("dct") \n\t" \ - "movdqa %%xmm5, 6*16("dct") \n\t" \ - "movdqa %%xmm7, %%xmm0 \n\t" \ - "movdqa "REG4", %%xmm4 \n\t" \ - "psubsw %%xmm1, %%xmm7 \n\t" \ - "psubsw "TAN1", "REG4" \n\t" \ - "paddsw %%xmm0, %%xmm1 \n\t" \ - "paddsw %%xmm4, "TAN1" \n\t" \ - "psraw $6, %%xmm1 \n\t" \ - "psraw $6, %%xmm7 \n\t" \ - "psraw $6, "TAN1" \n\t" \ - "psraw $6, "REG4" \n\t" \ - "movdqa %%xmm1, ("dct") \n\t" \ - "movdqa "TAN1", 3*16("dct") \n\t" \ - "movdqa "REG4", 4*16("dct") \n\t" \ - "movdqa %%xmm7, 7*16("dct") \n\t" - -/// IDCT pass on columns, assuming rows 4-7 are zero. -#define iLLM_PASS_SPARSE(dct) \ - "pmulhw %%xmm4, "TAN3" \n\t" \ - "paddsw %%xmm4, "TAN3" \n\t" \ - "movdqa %%xmm6, %%xmm3 \n\t" \ - "pmulhw %%xmm6, "TAN1" \n\t" \ - "movdqa %%xmm4, %%xmm1 \n\t" \ - "psubsw %%xmm1, %%xmm3 \n\t" \ - "paddsw %%xmm6, %%xmm1 \n\t" \ - "movdqa "TAN1", %%xmm6 \n\t" \ - "psubsw "TAN3", "TAN1" \n\t" \ - "paddsw %%xmm6, "TAN3" \n\t" \ - "movdqa %%xmm3, %%xmm6 \n\t" \ - "psubsw "TAN3", %%xmm3 \n\t" \ - "paddsw %%xmm6, "TAN3" \n\t" \ - "movdqa "MANGLE(sqrt2)", %%xmm4 \n\t" \ - "pmulhw %%xmm4, %%xmm3 \n\t" \ - "pmulhw %%xmm4, "TAN3" \n\t" \ - "paddsw "TAN3", "TAN3" \n\t" \ - "paddsw %%xmm3, %%xmm3 \n\t" \ - "movdqa "MANGLE(tan2)", %%xmm5 \n\t" \ - MOV_32_ONLY ROW2", "SREG2" \n\t" \ - "pmulhw "SREG2", %%xmm5 \n\t" \ - MOV_32_ONLY ROW0", "REG0" \n\t" \ - "movdqa "REG0", %%xmm6 \n\t" \ - "psubsw "SREG2", %%xmm6 \n\t" \ - "paddsw "REG0", "SREG2" \n\t" \ - MOV_32_ONLY" "TAN1", (%0) \n\t" \ - "movdqa "REG0", "XMMS" \n\t" \ - "psubsw %%xmm5, "REG0" \n\t" \ - "paddsw "XMMS", %%xmm5 \n\t" \ - "movdqa %%xmm5, "XMMS" \n\t" \ - "psubsw "TAN3", %%xmm5 \n\t" \ - "paddsw "XMMS", "TAN3" \n\t" \ - "movdqa "REG0", "XMMS" \n\t" \ - "psubsw %%xmm3, "REG0" \n\t" \ - "paddsw "XMMS", %%xmm3 \n\t" \ - MOV_32_ONLY" (%0), "TAN1" \n\t" \ - "psraw $6, %%xmm5 \n\t" \ - "psraw $6, "REG0" \n\t" \ - "psraw $6, "TAN3" \n\t" \ - "psraw $6, %%xmm3 \n\t" \ - "movdqa "TAN3", 1*16("dct") \n\t" \ - "movdqa %%xmm3, 2*16("dct") \n\t" \ - "movdqa "REG0", 5*16("dct") \n\t" \ - "movdqa %%xmm5, 6*16("dct") \n\t" \ - "movdqa "SREG2", %%xmm0 \n\t" \ - "movdqa %%xmm6, %%xmm4 \n\t" \ - "psubsw %%xmm1, "SREG2" \n\t" \ - "psubsw "TAN1", %%xmm6 \n\t" \ - "paddsw %%xmm0, %%xmm1 \n\t" \ - "paddsw %%xmm4, "TAN1" \n\t" \ - "psraw $6, %%xmm1 \n\t" \ - "psraw $6, "SREG2" \n\t" \ - "psraw $6, "TAN1" \n\t" \ - "psraw $6, %%xmm6 \n\t" \ - "movdqa %%xmm1, ("dct") \n\t" \ - "movdqa "TAN1", 3*16("dct") \n\t" \ - "movdqa %%xmm6, 4*16("dct") \n\t" \ - "movdqa "SREG2", 7*16("dct") \n\t" - -av_extern_inline void ff_xvid_idct_sse2(short *block) -{ - __asm__ volatile ( - "movq "MANGLE (m127) ", %%mm0 \n\t" - iMTX_MULT("(%0)", MANGLE(iTab1), ROUND(MANGLE(walkenIdctRounders)), PUT_EVEN(ROW0)) - iMTX_MULT("1*16(%0)", MANGLE(iTab2), ROUND("1*16+"MANGLE(walkenIdctRounders)), PUT_ODD(ROW1)) - iMTX_MULT("2*16(%0)", MANGLE(iTab3), ROUND("2*16+"MANGLE(walkenIdctRounders)), PUT_EVEN(ROW2)) - - TEST_TWO_ROWS("3*16(%0)", "4*16(%0)", "%%eax", "%%ecx", CLEAR_ODD(ROW3), CLEAR_EVEN(ROW4)) - JZ("%%eax", "1f") - iMTX_MULT("3*16(%0)", MANGLE(iTab4), ROUND("3*16+"MANGLE(walkenIdctRounders)), PUT_ODD(ROW3)) - - TEST_TWO_ROWS("5*16(%0)", "6*16(%0)", "%%eax", "%%edx", CLEAR_ODD(ROW5), CLEAR_EVEN(ROW6)) - TEST_ONE_ROW("7*16(%0)", "%%esi", CLEAR_ODD(ROW7)) - iLLM_HEAD - ".p2align 4 \n\t" - JNZ("%%ecx", "2f") - JNZ("%%eax", "3f") - JNZ("%%edx", "4f") - JNZ("%%esi", "5f") - iLLM_PASS_SPARSE("%0") - "jmp 6f \n\t" - "2: \n\t" - iMTX_MULT("4*16(%0)", MANGLE(iTab1), "#", PUT_EVEN(ROW4)) - "3: \n\t" - iMTX_MULT("5*16(%0)", MANGLE(iTab4), ROUND("4*16+"MANGLE(walkenIdctRounders)), PUT_ODD(ROW5)) - JZ("%%edx", "1f") - "4: \n\t" - iMTX_MULT("6*16(%0)", MANGLE(iTab3), ROUND("5*16+"MANGLE(walkenIdctRounders)), PUT_EVEN(ROW6)) - JZ("%%esi", "1f") - "5: \n\t" - iMTX_MULT("7*16(%0)", MANGLE(iTab2), ROUND("5*16+"MANGLE(walkenIdctRounders)), PUT_ODD(ROW7)) -#if ARCH_X86_32 - iLLM_HEAD -#endif - iLLM_PASS("%0") - "6: \n\t" - : "+r" (block) - : NAMED_CONSTRAINTS_ARRAY(m127,iTab1,walkenIdctRounders,iTab2,iTab3,iTab4,tan3,tan1,tan2,sqrt2) - : XMM_CLOBBERS("%xmm0", "%xmm1", "%xmm2", "%xmm3", - "%xmm4", "%xmm5", "%xmm6", "%xmm7", ) -#if ARCH_X86_64 - XMM_CLOBBERS("%xmm8", "%xmm9", "%xmm10", "%xmm11", - "%xmm12", "%xmm13", "%xmm14", ) -#endif - "%eax", "%ecx", "%edx", "%esi", "memory"); -} - -void ff_xvid_idct_sse2_put(uint8_t *dest, int line_size, short *block) -{ - ff_xvid_idct_sse2(block); - ff_put_pixels_clamped(block, dest, line_size); -} - -void ff_xvid_idct_sse2_add(uint8_t *dest, int line_size, short *block) -{ - ff_xvid_idct_sse2(block); - ff_add_pixels_clamped(block, dest, line_size); -} - -#endif /* HAVE_SSE2_INLINE */ -- cgit v1.2.3