Mercurial > libavcodec.hg
view i386/fdct_mmx.c @ 8362:156137b60026 libavcodec
Split filter_mb_dir() out of filter_mb().
1% overall decoding speed up for cathedral-beta2-400extra-crop-avc.mp4
no speed change for Aladin.mpg
Benchmarks done on Pentium dual
author | michael |
---|---|
date | Wed, 17 Dec 2008 02:35:14 +0000 |
parents | f44191520adf |
children |
line wrap: on
line source
/* * MMX optimized forward DCT * The gcc porting is Copyright (c) 2001 Fabrice Bellard. * cleanup/optimizations are Copyright (c) 2002-2004 Michael Niedermayer <michaelni@gmx.at> * SSE2 optimization is Copyright (c) 2004 Denes Balatoni. * * from fdctam32.c - AP922 MMX(3D-Now) forward-DCT * * Intel Application Note AP-922 - fast, precise implementation of DCT * http://developer.intel.com/vtune/cbts/appnotes.htm * * Also of inspiration: * a page about fdct at http://www.geocities.com/ssavekar/dct.htm * Skal's fdct at http://skal.planet-d.net/coding/dct.html * * 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/common.h" #include "libavcodec/dsputil.h" #define ATTR_ALIGN(align) __attribute__ ((__aligned__ (align))) ////////////////////////////////////////////////////////////////////// // // constants for the forward DCT // ----------------------------- // // Be sure to check that your compiler is aligning all constants to QWORD // (8-byte) memory boundaries! Otherwise the unaligned memory access will // severely stall MMX execution. // ////////////////////////////////////////////////////////////////////// #define BITS_FRW_ACC 3 //; 2 or 3 for accuracy #define SHIFT_FRW_COL BITS_FRW_ACC #define SHIFT_FRW_ROW (BITS_FRW_ACC + 17 - 3) #define RND_FRW_ROW (1 << (SHIFT_FRW_ROW-1)) //#define RND_FRW_COL (1 << (SHIFT_FRW_COL-1)) #define X8(x) x,x,x,x,x,x,x,x //concatenated table, for forward DCT transformation static const int16_t fdct_tg_all_16[24] ATTR_ALIGN(16) = { X8(13036), // tg * (2<<16) + 0.5 X8(27146), // tg * (2<<16) + 0.5 X8(-21746) // tg * (2<<16) + 0.5 }; static const int16_t ocos_4_16[8] ATTR_ALIGN(16) = { X8(23170) //cos * (2<<15) + 0.5 }; static const int16_t fdct_one_corr[8] ATTR_ALIGN(16) = { X8(1) }; static const int32_t fdct_r_row[2] ATTR_ALIGN(8) = {RND_FRW_ROW, RND_FRW_ROW }; static struct { const int32_t fdct_r_row_sse2[4] ATTR_ALIGN(16); } fdct_r_row_sse2 ATTR_ALIGN(16)= {{ RND_FRW_ROW, RND_FRW_ROW, RND_FRW_ROW, RND_FRW_ROW }}; //static const long fdct_r_row_sse2[4] ATTR_ALIGN(16) = {RND_FRW_ROW, RND_FRW_ROW, RND_FRW_ROW, RND_FRW_ROW}; static const int16_t tab_frw_01234567[] ATTR_ALIGN(8) = { // forward_dct coeff table 16384, 16384, 22725, 19266, 16384, 16384, 12873, 4520, 21407, 8867, 19266, -4520, -8867, -21407, -22725, -12873, 16384, -16384, 12873, -22725, -16384, 16384, 4520, 19266, 8867, -21407, 4520, -12873, 21407, -8867, 19266, -22725, 22725, 22725, 31521, 26722, 22725, 22725, 17855, 6270, 29692, 12299, 26722, -6270, -12299, -29692, -31521, -17855, 22725, -22725, 17855, -31521, -22725, 22725, 6270, 26722, 12299, -29692, 6270, -17855, 29692, -12299, 26722, -31521, 21407, 21407, 29692, 25172, 21407, 21407, 16819, 5906, 27969, 11585, 25172, -5906, -11585, -27969, -29692, -16819, 21407, -21407, 16819, -29692, -21407, 21407, 5906, 25172, 11585, -27969, 5906, -16819, 27969, -11585, 25172, -29692, 19266, 19266, 26722, 22654, 19266, 19266, 15137, 5315, 25172, 10426, 22654, -5315, -10426, -25172, -26722, -15137, 19266, -19266, 15137, -26722, -19266, 19266, 5315, 22654, 10426, -25172, 5315, -15137, 25172, -10426, 22654, -26722, 16384, 16384, 22725, 19266, 16384, 16384, 12873, 4520, 21407, 8867, 19266, -4520, -8867, -21407, -22725, -12873, 16384, -16384, 12873, -22725, -16384, 16384, 4520, 19266, 8867, -21407, 4520, -12873, 21407, -8867, 19266, -22725, 19266, 19266, 26722, 22654, 19266, 19266, 15137, 5315, 25172, 10426, 22654, -5315, -10426, -25172, -26722, -15137, 19266, -19266, 15137, -26722, -19266, 19266, 5315, 22654, 10426, -25172, 5315, -15137, 25172, -10426, 22654, -26722, 21407, 21407, 29692, 25172, 21407, 21407, 16819, 5906, 27969, 11585, 25172, -5906, -11585, -27969, -29692, -16819, 21407, -21407, 16819, -29692, -21407, 21407, 5906, 25172, 11585, -27969, 5906, -16819, 27969, -11585, 25172, -29692, 22725, 22725, 31521, 26722, 22725, 22725, 17855, 6270, 29692, 12299, 26722, -6270, -12299, -29692, -31521, -17855, 22725, -22725, 17855, -31521, -22725, 22725, 6270, 26722, 12299, -29692, 6270, -17855, 29692, -12299, 26722, -31521, }; static struct { const int16_t tab_frw_01234567_sse2[256] ATTR_ALIGN(16); } tab_frw_01234567_sse2 ATTR_ALIGN(16) = {{ //static const int16_t tab_frw_01234567_sse2[] ATTR_ALIGN(16) = { // forward_dct coeff table #define TABLE_SSE2 C4, C4, C1, C3, -C6, -C2, -C1, -C5, \ C4, C4, C5, C7, C2, C6, C3, -C7, \ -C4, C4, C7, C3, C6, -C2, C7, -C5, \ C4, -C4, C5, -C1, C2, -C6, C3, -C1, // c1..c7 * cos(pi/4) * 2^15 #define C1 22725 #define C2 21407 #define C3 19266 #define C4 16384 #define C5 12873 #define C6 8867 #define C7 4520 TABLE_SSE2 #undef C1 #undef C2 #undef C3 #undef C4 #undef C5 #undef C6 #undef C7 #define C1 31521 #define C2 29692 #define C3 26722 #define C4 22725 #define C5 17855 #define C6 12299 #define C7 6270 TABLE_SSE2 #undef C1 #undef C2 #undef C3 #undef C4 #undef C5 #undef C6 #undef C7 #define C1 29692 #define C2 27969 #define C3 25172 #define C4 21407 #define C5 16819 #define C6 11585 #define C7 5906 TABLE_SSE2 #undef C1 #undef C2 #undef C3 #undef C4 #undef C5 #undef C6 #undef C7 #define C1 26722 #define C2 25172 #define C3 22654 #define C4 19266 #define C5 15137 #define C6 10426 #define C7 5315 TABLE_SSE2 #undef C1 #undef C2 #undef C3 #undef C4 #undef C5 #undef C6 #undef C7 #define C1 22725 #define C2 21407 #define C3 19266 #define C4 16384 #define C5 12873 #define C6 8867 #define C7 4520 TABLE_SSE2 #undef C1 #undef C2 #undef C3 #undef C4 #undef C5 #undef C6 #undef C7 #define C1 26722 #define C2 25172 #define C3 22654 #define C4 19266 #define C5 15137 #define C6 10426 #define C7 5315 TABLE_SSE2 #undef C1 #undef C2 #undef C3 #undef C4 #undef C5 #undef C6 #undef C7 #define C1 29692 #define C2 27969 #define C3 25172 #define C4 21407 #define C5 16819 #define C6 11585 #define C7 5906 TABLE_SSE2 #undef C1 #undef C2 #undef C3 #undef C4 #undef C5 #undef C6 #undef C7 #define C1 31521 #define C2 29692 #define C3 26722 #define C4 22725 #define C5 17855 #define C6 12299 #define C7 6270 TABLE_SSE2 }}; #define S(s) AV_TOSTRING(s) //AV_STRINGIFY is too long #define FDCT_COL(cpu, mm, mov)\ static av_always_inline void fdct_col_##cpu(const int16_t *in, int16_t *out, int offset)\ {\ __asm__ volatile (\ #mov" 16(%0), %%"#mm"0 \n\t" \ #mov" 96(%0), %%"#mm"1 \n\t" \ #mov" %%"#mm"0, %%"#mm"2 \n\t" \ #mov" 32(%0), %%"#mm"3 \n\t" \ "paddsw %%"#mm"1, %%"#mm"0 \n\t" \ #mov" 80(%0), %%"#mm"4 \n\t" \ "psllw $"S(SHIFT_FRW_COL)", %%"#mm"0 \n\t" \ #mov" (%0), %%"#mm"5 \n\t" \ "paddsw %%"#mm"3, %%"#mm"4 \n\t" \ "paddsw 112(%0), %%"#mm"5 \n\t" \ "psllw $"S(SHIFT_FRW_COL)", %%"#mm"4 \n\t" \ #mov" %%"#mm"0, %%"#mm"6 \n\t" \ "psubsw %%"#mm"1, %%"#mm"2 \n\t" \ #mov" 16(%1), %%"#mm"1 \n\t" \ "psubsw %%"#mm"4, %%"#mm"0 \n\t" \ #mov" 48(%0), %%"#mm"7 \n\t" \ "pmulhw %%"#mm"0, %%"#mm"1 \n\t" \ "paddsw 64(%0), %%"#mm"7 \n\t" \ "psllw $"S(SHIFT_FRW_COL)", %%"#mm"5 \n\t" \ "paddsw %%"#mm"4, %%"#mm"6 \n\t" \ "psllw $"S(SHIFT_FRW_COL)", %%"#mm"7 \n\t" \ #mov" %%"#mm"5, %%"#mm"4 \n\t" \ "psubsw %%"#mm"7, %%"#mm"5 \n\t" \ "paddsw %%"#mm"5, %%"#mm"1 \n\t" \ "paddsw %%"#mm"7, %%"#mm"4 \n\t" \ "por (%2), %%"#mm"1 \n\t" \ "psllw $"S(SHIFT_FRW_COL)"+1, %%"#mm"2 \n\t" \ "pmulhw 16(%1), %%"#mm"5 \n\t" \ #mov" %%"#mm"4, %%"#mm"7 \n\t" \ "psubsw 80(%0), %%"#mm"3 \n\t" \ "psubsw %%"#mm"6, %%"#mm"4 \n\t" \ #mov" %%"#mm"1, 32(%3) \n\t" \ "paddsw %%"#mm"6, %%"#mm"7 \n\t" \ #mov" 48(%0), %%"#mm"1 \n\t" \ "psllw $"S(SHIFT_FRW_COL)"+1, %%"#mm"3 \n\t" \ "psubsw 64(%0), %%"#mm"1 \n\t" \ #mov" %%"#mm"2, %%"#mm"6 \n\t" \ #mov" %%"#mm"4, 64(%3) \n\t" \ "paddsw %%"#mm"3, %%"#mm"2 \n\t" \ "pmulhw (%4), %%"#mm"2 \n\t" \ "psubsw %%"#mm"3, %%"#mm"6 \n\t" \ "pmulhw (%4), %%"#mm"6 \n\t" \ "psubsw %%"#mm"0, %%"#mm"5 \n\t" \ "por (%2), %%"#mm"5 \n\t" \ "psllw $"S(SHIFT_FRW_COL)", %%"#mm"1 \n\t" \ "por (%2), %%"#mm"2 \n\t" \ #mov" %%"#mm"1, %%"#mm"4 \n\t" \ #mov" (%0), %%"#mm"3 \n\t" \ "paddsw %%"#mm"6, %%"#mm"1 \n\t" \ "psubsw 112(%0), %%"#mm"3 \n\t" \ "psubsw %%"#mm"6, %%"#mm"4 \n\t" \ #mov" (%1), %%"#mm"0 \n\t" \ "psllw $"S(SHIFT_FRW_COL)", %%"#mm"3 \n\t" \ #mov" 32(%1), %%"#mm"6 \n\t" \ "pmulhw %%"#mm"1, %%"#mm"0 \n\t" \ #mov" %%"#mm"7, (%3) \n\t" \ "pmulhw %%"#mm"4, %%"#mm"6 \n\t" \ #mov" %%"#mm"5, 96(%3) \n\t" \ #mov" %%"#mm"3, %%"#mm"7 \n\t" \ #mov" 32(%1), %%"#mm"5 \n\t" \ "psubsw %%"#mm"2, %%"#mm"7 \n\t" \ "paddsw %%"#mm"2, %%"#mm"3 \n\t" \ "pmulhw %%"#mm"7, %%"#mm"5 \n\t" \ "paddsw %%"#mm"3, %%"#mm"0 \n\t" \ "paddsw %%"#mm"4, %%"#mm"6 \n\t" \ "pmulhw (%1), %%"#mm"3 \n\t" \ "por (%2), %%"#mm"0 \n\t" \ "paddsw %%"#mm"7, %%"#mm"5 \n\t" \ "psubsw %%"#mm"6, %%"#mm"7 \n\t" \ #mov" %%"#mm"0, 16(%3) \n\t" \ "paddsw %%"#mm"4, %%"#mm"5 \n\t" \ #mov" %%"#mm"7, 48(%3) \n\t" \ "psubsw %%"#mm"1, %%"#mm"3 \n\t" \ #mov" %%"#mm"5, 80(%3) \n\t" \ #mov" %%"#mm"3, 112(%3) \n\t" \ : \ : "r" (in + offset), "r" (fdct_tg_all_16), "r" (fdct_one_corr), \ "r" (out + offset), "r" (ocos_4_16)); \ } FDCT_COL(mmx, mm, movq) FDCT_COL(sse2, xmm, movdqa) static av_always_inline void fdct_row_sse2(const int16_t *in, int16_t *out) { __asm__ volatile( #define FDCT_ROW_SSE2_H1(i,t) \ "movq " #i "(%0), %%xmm2 \n\t" \ "movq " #i "+8(%0), %%xmm0 \n\t" \ "movdqa " #t "+32(%1), %%xmm3 \n\t" \ "movdqa " #t "+48(%1), %%xmm7 \n\t" \ "movdqa " #t "(%1), %%xmm4 \n\t" \ "movdqa " #t "+16(%1), %%xmm5 \n\t" #define FDCT_ROW_SSE2_H2(i,t) \ "movq " #i "(%0), %%xmm2 \n\t" \ "movq " #i "+8(%0), %%xmm0 \n\t" \ "movdqa " #t "+32(%1), %%xmm3 \n\t" \ "movdqa " #t "+48(%1), %%xmm7 \n\t" #define FDCT_ROW_SSE2(i) \ "movq %%xmm2, %%xmm1 \n\t" \ "pshuflw $27, %%xmm0, %%xmm0 \n\t" \ "paddsw %%xmm0, %%xmm1 \n\t" \ "psubsw %%xmm0, %%xmm2 \n\t" \ "punpckldq %%xmm2, %%xmm1 \n\t" \ "pshufd $78, %%xmm1, %%xmm2 \n\t" \ "pmaddwd %%xmm2, %%xmm3 \n\t" \ "pmaddwd %%xmm1, %%xmm7 \n\t" \ "pmaddwd %%xmm5, %%xmm2 \n\t" \ "pmaddwd %%xmm4, %%xmm1 \n\t" \ "paddd %%xmm7, %%xmm3 \n\t" \ "paddd %%xmm2, %%xmm1 \n\t" \ "paddd %%xmm6, %%xmm3 \n\t" \ "paddd %%xmm6, %%xmm1 \n\t" \ "psrad %3, %%xmm3 \n\t" \ "psrad %3, %%xmm1 \n\t" \ "packssdw %%xmm3, %%xmm1 \n\t" \ "movdqa %%xmm1, " #i "(%4) \n\t" "movdqa (%2), %%xmm6 \n\t" FDCT_ROW_SSE2_H1(0,0) FDCT_ROW_SSE2(0) FDCT_ROW_SSE2_H2(64,0) FDCT_ROW_SSE2(64) FDCT_ROW_SSE2_H1(16,64) FDCT_ROW_SSE2(16) FDCT_ROW_SSE2_H2(112,64) FDCT_ROW_SSE2(112) FDCT_ROW_SSE2_H1(32,128) FDCT_ROW_SSE2(32) FDCT_ROW_SSE2_H2(96,128) FDCT_ROW_SSE2(96) FDCT_ROW_SSE2_H1(48,192) FDCT_ROW_SSE2(48) FDCT_ROW_SSE2_H2(80,192) FDCT_ROW_SSE2(80) : : "r" (in), "r" (tab_frw_01234567_sse2.tab_frw_01234567_sse2), "r" (fdct_r_row_sse2.fdct_r_row_sse2), "i" (SHIFT_FRW_ROW), "r" (out) ); } static av_always_inline void fdct_row_mmx2(const int16_t *in, int16_t *out, const int16_t *table) { __asm__ volatile ( "pshufw $0x1B, 8(%0), %%mm5 \n\t" "movq (%0), %%mm0 \n\t" "movq %%mm0, %%mm1 \n\t" "paddsw %%mm5, %%mm0 \n\t" "psubsw %%mm5, %%mm1 \n\t" "movq %%mm0, %%mm2 \n\t" "punpckldq %%mm1, %%mm0 \n\t" "punpckhdq %%mm1, %%mm2 \n\t" "movq (%1), %%mm1 \n\t" "movq 8(%1), %%mm3 \n\t" "movq 16(%1), %%mm4 \n\t" "movq 24(%1), %%mm5 \n\t" "movq 32(%1), %%mm6 \n\t" "movq 40(%1), %%mm7 \n\t" "pmaddwd %%mm0, %%mm1 \n\t" "pmaddwd %%mm2, %%mm3 \n\t" "pmaddwd %%mm0, %%mm4 \n\t" "pmaddwd %%mm2, %%mm5 \n\t" "pmaddwd %%mm0, %%mm6 \n\t" "pmaddwd %%mm2, %%mm7 \n\t" "pmaddwd 48(%1), %%mm0 \n\t" "pmaddwd 56(%1), %%mm2 \n\t" "paddd %%mm1, %%mm3 \n\t" "paddd %%mm4, %%mm5 \n\t" "paddd %%mm6, %%mm7 \n\t" "paddd %%mm0, %%mm2 \n\t" "movq (%2), %%mm0 \n\t" "paddd %%mm0, %%mm3 \n\t" "paddd %%mm0, %%mm5 \n\t" "paddd %%mm0, %%mm7 \n\t" "paddd %%mm0, %%mm2 \n\t" "psrad $"S(SHIFT_FRW_ROW)", %%mm3 \n\t" "psrad $"S(SHIFT_FRW_ROW)", %%mm5 \n\t" "psrad $"S(SHIFT_FRW_ROW)", %%mm7 \n\t" "psrad $"S(SHIFT_FRW_ROW)", %%mm2 \n\t" "packssdw %%mm5, %%mm3 \n\t" "packssdw %%mm2, %%mm7 \n\t" "movq %%mm3, (%3) \n\t" "movq %%mm7, 8(%3) \n\t" : : "r" (in), "r" (table), "r" (fdct_r_row), "r" (out)); } static av_always_inline void fdct_row_mmx(const int16_t *in, int16_t *out, const int16_t *table) { //FIXME reorder (I do not have an old MMX-only CPU here to benchmark ...) __asm__ volatile( "movd 12(%0), %%mm1 \n\t" "punpcklwd 8(%0), %%mm1 \n\t" "movq %%mm1, %%mm2 \n\t" "psrlq $0x20, %%mm1 \n\t" "movq 0(%0), %%mm0 \n\t" "punpcklwd %%mm2, %%mm1 \n\t" "movq %%mm0, %%mm5 \n\t" "paddsw %%mm1, %%mm0 \n\t" "psubsw %%mm1, %%mm5 \n\t" "movq %%mm0, %%mm2 \n\t" "punpckldq %%mm5, %%mm0 \n\t" "punpckhdq %%mm5, %%mm2 \n\t" "movq 0(%1), %%mm1 \n\t" "movq 8(%1), %%mm3 \n\t" "movq 16(%1), %%mm4 \n\t" "movq 24(%1), %%mm5 \n\t" "movq 32(%1), %%mm6 \n\t" "movq 40(%1), %%mm7 \n\t" "pmaddwd %%mm0, %%mm1 \n\t" "pmaddwd %%mm2, %%mm3 \n\t" "pmaddwd %%mm0, %%mm4 \n\t" "pmaddwd %%mm2, %%mm5 \n\t" "pmaddwd %%mm0, %%mm6 \n\t" "pmaddwd %%mm2, %%mm7 \n\t" "pmaddwd 48(%1), %%mm0 \n\t" "pmaddwd 56(%1), %%mm2 \n\t" "paddd %%mm1, %%mm3 \n\t" "paddd %%mm4, %%mm5 \n\t" "paddd %%mm6, %%mm7 \n\t" "paddd %%mm0, %%mm2 \n\t" "movq (%2), %%mm0 \n\t" "paddd %%mm0, %%mm3 \n\t" "paddd %%mm0, %%mm5 \n\t" "paddd %%mm0, %%mm7 \n\t" "paddd %%mm0, %%mm2 \n\t" "psrad $"S(SHIFT_FRW_ROW)", %%mm3 \n\t" "psrad $"S(SHIFT_FRW_ROW)", %%mm5 \n\t" "psrad $"S(SHIFT_FRW_ROW)", %%mm7 \n\t" "psrad $"S(SHIFT_FRW_ROW)", %%mm2 \n\t" "packssdw %%mm5, %%mm3 \n\t" "packssdw %%mm2, %%mm7 \n\t" "movq %%mm3, 0(%3) \n\t" "movq %%mm7, 8(%3) \n\t" : : "r" (in), "r" (table), "r" (fdct_r_row), "r" (out)); } void ff_fdct_mmx(int16_t *block) { int64_t align_tmp[16] ATTR_ALIGN(8); int16_t * block1= (int16_t*)align_tmp; const int16_t *table= tab_frw_01234567; int i; fdct_col_mmx(block, block1, 0); fdct_col_mmx(block, block1, 4); for(i=8;i>0;i--) { fdct_row_mmx(block1, block, table); block1 += 8; table += 32; block += 8; } } void ff_fdct_mmx2(int16_t *block) { int64_t align_tmp[16] ATTR_ALIGN(8); int16_t *block1= (int16_t*)align_tmp; const int16_t *table= tab_frw_01234567; int i; fdct_col_mmx(block, block1, 0); fdct_col_mmx(block, block1, 4); for(i=8;i>0;i--) { fdct_row_mmx2(block1, block, table); block1 += 8; table += 32; block += 8; } } void ff_fdct_sse2(int16_t *block) { int64_t align_tmp[16] ATTR_ALIGN(16); int16_t * const block1= (int16_t*)align_tmp; fdct_col_sse2(block, block1, 0); fdct_row_sse2(block1, block); }