From 8cca8db811c7888058c6093d2f1ae5ae4f00cccc Mon Sep 17 00:00:00 2001 From: "Steinar H. Gunderson" Date: Sat, 16 Sep 2017 15:12:35 +0200 Subject: [PATCH] Add the missing DCT code. --- fdct.cpp | 314 +++++++++++++++++++++++++++++++++++++++++++++++++++ fdct.h | 50 +++++++++ idct.cpp | 337 +++++++++++++++++++++++++++++++++++++++++++++++++++++++ 3 files changed, 701 insertions(+) create mode 100644 fdct.cpp create mode 100644 fdct.h create mode 100644 idct.cpp diff --git a/fdct.cpp b/fdct.cpp new file mode 100644 index 0000000..964c9a5 --- /dev/null +++ b/fdct.cpp @@ -0,0 +1,314 @@ +/***************************************************************************** + * + * XVID MPEG-4 VIDEO CODEC + * - Forward DCT - + * + * Copyright (C) 2006-2011 Xvid Solutions GmbH + * + * This program is free software ; you can redistribute it and/or modify + * it under the terms of the GNU General Public License as published by + * the Free Software Foundation ; either version 2 of the License, or + * (at your option) any later version. + * + * This program 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 General Public License for more details. + * + * You should have received a copy of the GNU General Public License + * along with this program ; if not, write to the Free Software + * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA + * + * $Id: fdct.c 1986 2011-05-18 09:07:40Z Isibaar $ + * + ****************************************************************************/ + +/* + * Authors: Skal + * + * "Fast and precise" LLM implementation of FDCT/IDCT, where + * rotations are decomposed using: + * tmp = (x+y).cos t + * x' = tmp + y.(sin t - cos t) + * y' = tmp - x.(sin t + cos t) + * + * See details at the end of this file... + * + * Reference (e.g.): + * Loeffler C., Ligtenberg A., and Moschytz C.S.: + * Practical Fast 1D DCT Algorithm with Eleven Multiplications, + * Proc. ICASSP 1989, 988-991. + * + * IEEE-1180-like error specs for FDCT: + * Peak error: 1.0000 + * Peak MSE: 0.0340 + * Overall MSE: 0.0200 + * Peak ME: 0.0191 + * Overall ME: -0.0033 + * + ********************************************************/ + +#include + +//#include "fdct.h" + +/* function pointer */ +//fdctFuncPtr fdct; + +/* +////////////////////////////////////////////////////////// +*/ + +#define BUTF(a, b, tmp) \ + (tmp) = (a)+(b); \ + (b) = (a)-(b); \ + (a) = (tmp) + +#define LOAD_BUTF(m1, m2, a, b, tmp, S) \ + (m1) = (S)[(a)] + (S)[(b)]; \ + (m2) = (S)[(a)] - (S)[(b)] + +#define ROTATE(m1,m2,c,k1,k2,tmp,Fix,Rnd) \ + (tmp) = ( (m1) + (m2) )*(c); \ + (m1) *= k1; \ + (m2) *= k2; \ + (tmp) += (Rnd); \ + (m1) = ((m1)+(tmp))>>(Fix); \ + (m2) = ((m2)+(tmp))>>(Fix); + +#define ROTATE2(m1,m2,c,k1,k2,tmp) \ + (tmp) = ( (m1) + (m2) )*(c); \ + (m1) *= k1; \ + (m2) *= k2; \ + (m1) = (m1)+(tmp); \ + (m2) = (m2)+(tmp); + +#define ROTATE0(m1,m2,c,k1,k2,tmp) \ + (m1) = ( (m2) )*(c); \ + (m2) = (m2)*k2+(m1); + +#define SHIFTL(x,n) ((x)<<(n)) +#define SHIFTR(x, n) ((x)>>(n)) +#define HALF(n) (1<<((n)-1)) + +#define IPASS 3 +#define FPASS 2 +#define FIX 16 + +#if 1 + +#define ROT6_C 35468 +#define ROT6_SmC 50159 +#define ROT6_SpC 121095 +#define ROT17_C 77062 +#define ROT17_SmC 25571 +#define ROT17_SpC 128553 +#define ROT37_C 58981 +#define ROT37_SmC 98391 +#define ROT37_SpC 19571 +#define ROT13_C 167963 +#define ROT13_SmC 134553 +#define ROT13_SpC 201373 + +#else + +#include +#define FX(x) ( (int)floor((x)*(1<0; --i) + { + int mm0, mm1, mm2, mm3, mm4, mm5, mm6, mm7, Spill; + + // even + + LOAD_BUTF(mm1,mm6, 1, 6, mm0, pIn); + LOAD_BUTF(mm2,mm5, 2, 5, mm0, pIn); + LOAD_BUTF(mm3,mm4, 3, 4, mm0, pIn); + LOAD_BUTF(mm0,mm7, 0, 7, Spill, pIn); + + BUTF(mm1, mm2, Spill); + BUTF(mm0, mm3, Spill); + + ROTATE(mm3, mm2, ROT6_C, ROT6_SmC, -ROT6_SpC, Spill, FIX-FPASS, HALF(FIX-FPASS)); + pIn[2] = mm3; + pIn[6] = mm2; + + BUTF(mm0, mm1, Spill); + pIn[0] = SHIFTL(mm0, FPASS); + pIn[4] = SHIFTL(mm1, FPASS); + + + // odd + + mm3 = mm5 + mm7; + mm2 = mm4 + mm6; + ROTATE(mm2, mm3, ROT17_C, -ROT17_SpC, -ROT17_SmC, mm0, FIX-FPASS, HALF(FIX-FPASS)); + ROTATE(mm4, mm7, -ROT37_C, ROT37_SpC, ROT37_SmC, mm0, FIX-FPASS, HALF(FIX-FPASS)); + mm7 += mm3; + mm4 += mm2; + pIn[1] = mm7; + pIn[7] = mm4; + + ROTATE(mm5, mm6, -ROT13_C, ROT13_SmC, ROT13_SpC, mm0, FIX-FPASS, HALF(FIX-FPASS)); + mm5 += mm3; + mm6 += mm2; + pIn[3] = mm6; + pIn[5] = mm5; + + pIn += 8; + } + + pIn = In; + for(i=8; i>0; --i) + { + int mm0, mm1, mm2, mm3, mm4, mm5, mm6, mm7, Spill; + + // even + + LOAD_BUTF(mm1,mm6, 1*8, 6*8, mm0, pIn); + LOAD_BUTF(mm2,mm5, 2*8, 5*8, mm0, pIn); + BUTF(mm1, mm2, mm0); + + LOAD_BUTF(mm3,mm4, 3*8, 4*8, mm0, pIn); + LOAD_BUTF(mm0,mm7, 0*8, 7*8, Spill, pIn); + BUTF(mm0, mm3, Spill); + + ROTATE(mm3, mm2, ROT6_C, ROT6_SmC, -ROT6_SpC, Spill, 0, HALF(FIX+FPASS+3)); + pIn[2*8] = (int16_t)SHIFTR(mm3,FIX+FPASS+3); + pIn[6*8] = (int16_t)SHIFTR(mm2,FIX+FPASS+3); + + mm0 += HALF(FPASS+3) - 1; + BUTF(mm0, mm1, Spill); + pIn[0*8] = (int16_t)SHIFTR(mm0, FPASS+3); + pIn[4*8] = (int16_t)SHIFTR(mm1, FPASS+3); + + // odd + + mm3 = mm5 + mm7; + mm2 = mm4 + mm6; + + ROTATE(mm2, mm3, ROT17_C, -ROT17_SpC, -ROT17_SmC, mm0, 0, HALF(FIX+FPASS+3)); + ROTATE2(mm4, mm7, -ROT37_C, ROT37_SpC, ROT37_SmC, mm0); + mm7 += mm3; + mm4 += mm2; + pIn[7*8] = (int16_t)SHIFTR(mm4,FIX+FPASS+3); + pIn[1*8] = (int16_t)SHIFTR(mm7,FIX+FPASS+3); + + ROTATE2(mm5, mm6, -ROT13_C, ROT13_SmC, ROT13_SpC, mm0); + mm5 += mm3; + mm6 += mm2; + pIn[5*8] = (int16_t)SHIFTR(mm5,FIX+FPASS+3); + pIn[3*8] = (int16_t)SHIFTR(mm6,FIX+FPASS+3); + + pIn++; + } +} +#undef FIX +#undef FPASS +#undef IPASS + +#undef BUTF +#undef LOAD_BUTF +#undef ROTATE +#undef ROTATE2 +#undef SHIFTL +#undef SHIFTR + +////////////////////////////////////////////////////////// +// - Data flow schematics for FDCT - +// Output is scaled by 2.sqrt(2) +// Initial butterflies (in0/in7, etc.) are not fully depicted. +// Note: Rot6 coeffs are multiplied by sqrt(2). +////////////////////////////////////////////////////////// +/* + <---------Stage1 =even part=-----------> + + in3 mm3 +_____.___-___________.____* out6 + x \ / | + in4 mm4 \ | + / \ | + in0 mm0 +_____o___+__.___-___ | ___* out4 + x \ / | + in7 mm7 \ (Rot6) + / \ | + in1 mm1 +_____o___+__o___+___ | ___* out0 + x \ / | + in6 mm6 / | + / \ | + in2 mm2 +_____.___-___________o____* out2 + x + in5 mm5 + + <---------Stage2 =odd part=----------------> + + mm7*___._________.___-___[xSqrt2]___* out3 + | \ / + (Rot3) \ + | / \ + mm5*__ | ___o____o___+___.___-______* out7 + | | \ / + | (Rot1) \ + | | / \ + mm6*__ |____.____o___+___o___+______* out1 + | \ / + | / + | / \ + mm4*___o_________.___-___[xSqrt2]___* out5 + + + + Alternative schematics for stage 2: + ----------------------------------- + + mm7 *___[xSqrt2]____o___+____o_______* out1 + \ / | + / (Rot1) + / \ | + mm6 *____o___+______.___-___ | __.___* out5 + \ / | | + / | | + / \ | | + mm5 *____.___-______.___-____.__ | __* out7 + \ / | + \ (Rot3) + / \ | + mm4 *___[xSqrt2]____o___+________o___* out3 + +*/ diff --git a/fdct.h b/fdct.h new file mode 100644 index 0000000..93293ce --- /dev/null +++ b/fdct.h @@ -0,0 +1,50 @@ +/***************************************************************************** + * + * XVID MPEG-4 VIDEO CODEC + * - Forward DCT header - + * + * Copyright(C) 2001-2011 Michael Militzer + * + * This program is free software ; you can redistribute it and/or modify + * it under the terms of the GNU General Public License as published by + * the Free Software Foundation ; either version 2 of the License, or + * (at your option) any later version. + * + * This program 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 General Public License for more details. + * + * You should have received a copy of the GNU General Public License + * along with this program ; if not, write to the Free Software + * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA + * + * $Id: fdct.h 1986 2011-05-18 09:07:40Z Isibaar $ + * + ****************************************************************************/ + +#ifndef _FDCT_H_ +#define _FDCT_H_ + +#include "../portab.h" + +typedef void (fdctFunc) (short *const block); +typedef fdctFunc *fdctFuncPtr; + +extern fdctFuncPtr fdct; + +fdctFunc fdct_int32; + +#if defined(ARCH_IS_IA32) || defined(ARCH_IS_X86_64) +fdctFunc fdct_mmx_ffmpeg; +fdctFunc fdct_xmm_ffmpeg; +fdctFunc fdct_mmx_skal; +fdctFunc fdct_xmm_skal; +fdctFunc fdct_sse2_skal; +#endif + +#ifdef ARCH_IS_IA64 +fdctFunc fdct_ia64; +#endif + +#endif /* _FDCT_H_ */ diff --git a/idct.cpp b/idct.cpp new file mode 100644 index 0000000..0bc9712 --- /dev/null +++ b/idct.cpp @@ -0,0 +1,337 @@ +/***************************************************************************** + * + * XVID MPEG-4 VIDEO CODEC + * - Inverse DCT - + * + * Copyright (C) 2006-2011 Xvid Solutions GmbH + * + * This program is free software ; you can redistribute it and/or modify + * it under the terms of the GNU General Public License as published by + * the Free Software Foundation ; either version 2 of the License, or + * (at your option) any later version. + * + * This program 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 General Public License for more details. + * + * You should have received a copy of the GNU General Public License + * along with this program ; if not, write to the Free Software + * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA + * + * $Id: idct.c 1986 2011-05-18 09:07:40Z Isibaar $ + * + ****************************************************************************/ + +/* + * Authors: Skal + * + * Walken IDCT + * Alternative idct implementations for decoding compatibility + * + * NOTE: this "C" version is not the original one, + * but is modified to yield the same error profile + * than the MMX version. + * + ************************************************************************/ + +#include + +//#include "idct.h" + +/* function pointer */ +//idctFuncPtr idct; + +#define XVID_DSP_CLIP_255(x) ( ((x)&~255) ? ((-(x)) >> (8*sizeof((x))-1))&0xff : (x) ) + +#define ROW_SHIFT 11 +#define COL_SHIFT 6 + +// #define FIX(x) (int)((x) * (1<> ROW_SHIFT; + In[1] = (a1 + b1) >> ROW_SHIFT; + In[2] = (a2 + b2) >> ROW_SHIFT; + In[3] = (a3 + b3) >> ROW_SHIFT; + In[4] = (a3 - b3) >> ROW_SHIFT; + In[5] = (a2 - b2) >> ROW_SHIFT; + In[6] = (a1 - b1) >> ROW_SHIFT; + In[7] = (a0 - b0) >> ROW_SHIFT; + } + else + { + const int a0 = K >> ROW_SHIFT; + if (a0) { + In[0] = In[1] = In[2] = In[3] = + In[4] = In[5] = In[6] = In[7] = a0; + } + else return 0; + } + } + else if (!(Left|Right)) + { + const int a0 = (Rnd + C4*(In[0]+In[4])) >> ROW_SHIFT; + const int a1 = (Rnd + C4*(In[0]-In[4])) >> ROW_SHIFT; + + In[0] = a0; + In[3] = a0; + In[4] = a0; + In[7] = a0; + In[1] = a1; + In[2] = a1; + In[5] = a1; + In[6] = a1; + } + else + { + const int K = C4*In[0] + Rnd; + const int a0 = K + C2*In[2] + C4*In[4] + C6*In[6]; + const int a1 = K + C6*In[2] - C4*In[4] - C2*In[6]; + const int a2 = K - C6*In[2] - C4*In[4] + C2*In[6]; + const int a3 = K - C2*In[2] + C4*In[4] - C6*In[6]; + + const int b0 = C1*In[1] + C3*In[3] + C5*In[5] + C7*In[7]; + const int b1 = C3*In[1] - C7*In[3] - C1*In[5] - C5*In[7]; + const int b2 = C5*In[1] - C1*In[3] + C7*In[5] + C3*In[7]; + const int b3 = C7*In[1] - C5*In[3] + C3*In[5] - C1*In[7]; + + In[0] = (a0 + b0) >> ROW_SHIFT; + In[1] = (a1 + b1) >> ROW_SHIFT; + In[2] = (a2 + b2) >> ROW_SHIFT; + In[3] = (a3 + b3) >> ROW_SHIFT; + In[4] = (a3 - b3) >> ROW_SHIFT; + In[5] = (a2 - b2) >> ROW_SHIFT; + In[6] = (a1 - b1) >> ROW_SHIFT; + In[7] = (a0 - b0) >> ROW_SHIFT; + } + return 1; +} + +#define Tan1 0x32ec +#define Tan2 0x6a0a +#define Tan3 0xab0e +#define Sqrt2 0x5a82 + +#define MULT(c,x, n) ( ((c) * (x)) >> (n) ) +// 12b version => #define MULT(c,x, n) ( (((c)>>3) * (x)) >> ((n)-3) ) +// 12b zero-testing version: + +#define BUTF(a, b, tmp) \ + (tmp) = (a)+(b); \ + (b) = (a)-(b); \ + (a) = (tmp) + +#define LOAD_BUTF(m1, m2, a, b, tmp, S) \ + (m1) = (S)[(a)] + (S)[(b)]; \ + (m2) = (S)[(a)] - (S)[(b)] + +static void Idct_Col_8(short * const In) +{ + int mm0, mm1, mm2, mm3, mm4, mm5, mm6, mm7, Spill; + + // odd + + mm4 = (int)In[7*8]; + mm5 = (int)In[5*8]; + mm6 = (int)In[3*8]; + mm7 = (int)In[1*8]; + + mm0 = MULT(Tan1, mm4, 16) + mm7; + mm1 = MULT(Tan1, mm7, 16) - mm4; + mm2 = MULT(Tan3, mm5, 16) + mm6; + mm3 = MULT(Tan3, mm6, 16) - mm5; + + mm7 = mm0 + mm2; + mm4 = mm1 - mm3; + mm0 = mm0 - mm2; + mm1 = mm1 + mm3; + mm6 = mm0 + mm1; + mm5 = mm0 - mm1; + mm5 = 2*MULT(Sqrt2, mm5, 16); // 2*sqrt2 + mm6 = 2*MULT(Sqrt2, mm6, 16); // Watch out: precision loss but done to match + // the pmulhw used in mmx/sse versions + + // even + + mm1 = (int)In[2*8]; + mm2 = (int)In[6*8]; + mm3 = MULT(Tan2,mm2, 16) + mm1; + mm2 = MULT(Tan2,mm1, 16) - mm2; + + LOAD_BUTF(mm0, mm1, 0*8, 4*8, Spill, In); + + BUTF(mm0, mm3, Spill); + BUTF(mm0, mm7, Spill); + In[8*0] = (int16_t) (mm0 >> COL_SHIFT); + In[8*7] = (int16_t) (mm7 >> COL_SHIFT); + BUTF(mm3, mm4, mm0); + In[8*3] = (int16_t) (mm3 >> COL_SHIFT); + In[8*4] = (int16_t) (mm4 >> COL_SHIFT); + + BUTF(mm1, mm2, mm0); + BUTF(mm1, mm6, mm0); + In[8*1] = (int16_t) (mm1 >> COL_SHIFT); + In[8*6] = (int16_t) (mm6 >> COL_SHIFT); + BUTF(mm2, mm5, mm0); + In[8*2] = (int16_t) (mm2 >> COL_SHIFT); + In[8*5] = (int16_t) (mm5 >> COL_SHIFT); +} + +static void Idct_Col_4(short * const In) +{ + int mm0, mm1, mm2, mm3, mm4, mm5, mm6, mm7, Spill; + + // odd + + mm0 = (int)In[1*8]; + mm2 = (int)In[3*8]; + + mm1 = MULT(Tan1, mm0, 16); + mm3 = MULT(Tan3, mm2, 16); + + mm7 = mm0 + mm2; + mm4 = mm1 - mm3; + mm0 = mm0 - mm2; + mm1 = mm1 + mm3; + mm6 = mm0 + mm1; + mm5 = mm0 - mm1; + mm6 = 2*MULT(Sqrt2, mm6, 16); // 2*sqrt2 + mm5 = 2*MULT(Sqrt2, mm5, 16); + + // even + + mm0 = mm1 = (int)In[0*8]; + mm3 = (int)In[2*8]; + mm2 = MULT(Tan2,mm3, 16); + + BUTF(mm0, mm3, Spill); + BUTF(mm0, mm7, Spill); + In[8*0] = (int16_t) (mm0 >> COL_SHIFT); + In[8*7] = (int16_t) (mm7 >> COL_SHIFT); + BUTF(mm3, mm4, mm0); + In[8*3] = (int16_t) (mm3 >> COL_SHIFT); + In[8*4] = (int16_t) (mm4 >> COL_SHIFT); + + BUTF(mm1, mm2, mm0); + BUTF(mm1, mm6, mm0); + In[8*1] = (int16_t) (mm1 >> COL_SHIFT); + In[8*6] = (int16_t) (mm6 >> COL_SHIFT); + BUTF(mm2, mm5, mm0); + In[8*2] = (int16_t) (mm2 >> COL_SHIFT); + In[8*5] = (int16_t) (mm5 >> COL_SHIFT); +} + +static void Idct_Col_3(short * const In) +{ + int mm0, mm1, mm2, mm3, mm4, mm5, mm6, mm7, Spill; + + // odd + + mm7 = (int)In[1*8]; + mm4 = MULT(Tan1, mm7, 16); + + mm6 = mm7 + mm4; + mm5 = mm7 - mm4; + mm6 = 2*MULT(Sqrt2, mm6, 16); // 2*sqrt2 + mm5 = 2*MULT(Sqrt2, mm5, 16); + + // even + + mm0 = mm1 = (int)In[0*8]; + mm3 = (int)In[2*8]; + mm2 = MULT(Tan2,mm3, 16); + + BUTF(mm0, mm3, Spill); + BUTF(mm0, mm7, Spill); + In[8*0] = (int16_t) (mm0 >> COL_SHIFT); + In[8*7] = (int16_t) (mm7 >> COL_SHIFT); + BUTF(mm3, mm4, mm0); + In[8*3] = (int16_t) (mm3 >> COL_SHIFT); + In[8*4] = (int16_t) (mm4 >> COL_SHIFT); + + BUTF(mm1, mm2, mm0); + BUTF(mm1, mm6, mm0); + In[8*1] = (int16_t) (mm1 >> COL_SHIFT); + In[8*6] = (int16_t) (mm6 >> COL_SHIFT); + BUTF(mm2, mm5, mm0); + In[8*2] = (int16_t) (mm2 >> COL_SHIFT); + In[8*5] = (int16_t) (mm5 >> COL_SHIFT); +} + +#undef Tan1 +#undef Tan2 +#undef Tan3 +#undef Sqrt2 + +#undef ROW_SHIFT +#undef COL_SHIFT + +////////////////////////////////////////////////////////// + +void idct_int32(short *const In) +{ + int i, Rows = 0x07; + + Idct_Row(In + 0*8, Tab04, Rnd0); + Idct_Row(In + 1*8, Tab17, Rnd1); + Idct_Row(In + 2*8, Tab26, Rnd2); + if (Idct_Row(In + 3*8, Tab35, Rnd3)) Rows |= 0x08; + if (Idct_Row(In + 4*8, Tab04, Rnd4)) Rows |= 0x10; + if (Idct_Row(In + 5*8, Tab35, Rnd5)) Rows |= 0x20; + if (Idct_Row(In + 6*8, Tab26, Rnd6)) Rows |= 0x40; + if (Idct_Row(In + 7*8, Tab17, Rnd7)) Rows |= 0x80; + + if (Rows&0xf0) { + for(i=0; i<8; i++) + Idct_Col_8(In + i); + } + else if (Rows&0x08) { + for(i=0; i<8; i++) + Idct_Col_4(In + i); + } + else { + for(i=0; i<8; i++) + Idct_Col_3(In + i); + } +} -- 2.39.2