Mercurial > libavcodec.hg
view bfin/fdct_bfin.S @ 12197:fbf4d5b1b664 libavcodec
Remove FF_MM_SSE2/3 flags for CPUs where this is generally not faster than
regular MMX code. Examples of this are the Core1 CPU. Instead, set a new flag,
FF_MM_SSE2/3SLOW, which can be checked for particular SSE2/3 functions that
have been checked specifically on such CPUs and are actually faster than
their MMX counterparts.
In addition, use this flag to enable particular VP8 and LPC SSE2 functions
that are faster than their MMX counterparts.
Based on a patch by Loren Merritt <lorenm AT u washington edu>.
| author | rbultje |
|---|---|
| date | Mon, 19 Jul 2010 22:38:23 +0000 |
| parents | 8327c5b4df9b |
| children |
line wrap: on
line source
/* * fdct BlackFin * * Copyright (C) 2007 Marc Hoffman <marc.hoffman@analog.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 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA */ /* void ff_bfin_fdct (DCTELEM *buf); This implementation works only for 8x8 input. The range of input must be -256 to 255 i.e. 8bit input represented in a 16bit data word. The original data must be sign extended into the 16bit data words. Chen factorization of 8 X(m) = sum (x(n) * cos ((2n+1)*m*pi/16)) n=0 C4 0 --*-------------*0+7---*-----*0+3-------*-*-------------------> 0 \ / \ / X S4,S4 1 --*-\---------/-*1+6---*-\-/-*1+2-------*-*-------------------> 4 \ / \ -C4 C3 2 --*---\-----/---*2+5---*-/-\-*1-2---------------*-*-----------> 2 \ / / \ X S3,-S3 3 --*-----\-/-----*3+4---*-----*0-3---------------*-*-----------> 6 / C7 C3 4 --*-----/-\-----*3-4------------*-*4+5--*-----*---------------> 1 / \ -C4 X \ /S7 C3 5 --*---/-----\---*2-5---*-*------*=*4-5----\-/------*-*--------> 5 / \ X S4,S4 / X S3,-S3 6 --*-/---------\-*1-6---*-*------*=*7-6----/-\------*-*--------> 3 / \ C4 X / \-S7 C3 --*-------------*0-7------------*-*7+6--*-----*---------------> 7 C7 Notation Cn = cos(n*pi/8) used throughout the code. Registers used: R0, R1, R2, R3, R4, R5, R6,R7, P0, P1, P2, P3, P4, P5, A0, A1. Other registers used: I0, I1, I2, I3, B0, B2, B3, M0, M1, L3 registers and LC0. Input - r0 - pointer to start of DCTELEM *block Output - The DCT output coefficients in the DCTELEM *block Register constraint: This code is called from jpeg_encode. R6, R5, R4 if modified should be stored and restored. Performance: (Timer version 0.6.33) Code Size : 240 Bytes. Memory Required : Input Matrix : 8 * 8 * 2 Bytes. Coefficients : 16 Bytes Temporary matrix: 8 * 8 * 2 Bytes. Cycle Count :26+{18+8*(14+2S)}*2 where S -> Stalls (7.45 c/pel) ----------------------------------------- | Size | Forward DCT | Inverse DCT | ----------------------------------------- | 8x8 | 284 Cycles | 311 Cycles | ----------------------------------------- Ck = int16(cos(k/16*pi)*32767+.5)/2 #define C4 23170 #define C3 13623 #define C6 6270 #define C7 3196 Sk = int16(sin(k/16*pi)*32767+.5)/2 #define S4 11585 #define S3 9102 #define S6 15137 #define S7 16069 the coefficients are ordered as follows: short dct_coef[] C4,S4, C6,S6, C7,S7, S3,C3, ----------------------------------------------------------- FFMPEG conformance testing results ----------------------------------------------------------- dct-test: modified with the following dct_error("BFINfdct", 0, ff_bfin_fdct, fdct, test); produces the following output: root:/u/ffmpeg/bhead/libavcodec> ./dct-test ffmpeg DCT/IDCT test 2 -131 -6 -48 -36 33 -83 24 34 52 -24 -15 5 92 57 143 -67 -43 -1 74 -16 5 -71 32 -78 106 92 -34 -38 81 20 -18 7 -62 40 2 -15 90 -62 -83 -83 1 -104 -13 43 -19 7 11 -63 31 12 -29 83 72 21 10 -17 -63 -15 73 50 -91 159 -14 DCT BFINfdct: err_inf=2 err2=0.16425938 syserr=0.00795000 maxout=2098 blockSumErr=27 DCT BFINfdct: 92.1 kdct/s root:/u/ffmpeg/bhead/libavcodec> */ #include "config.h" #include "config_bfin.h" #if defined(__FDPIC__) && CONFIG_SRAM .section .l1.data.B,"aw",@progbits #else .data #endif .align 4; dct_coeff: .short 0x5a82, 0x2d41, 0x187e, 0x3b21, 0x0c7c, 0x3ec5, 0x238e, 0x3537; #if defined(__FDPIC__) && CONFIG_SRAM .section .l1.data.A,"aw",@progbits #endif .align 4 vtmp: .space 128 .text DEFUN(fdct,mL1, (DCTELEM *block)): [--SP] = (R7:4, P5:3); // Push the registers onto the stack. b0 = r0; RELOC(r0, P3, dct_coeff); b3 = r0; RELOC(r0, P3, vtmp); b2 = r0; L3 = 16; // L3 is set to 16 to make the coefficient // array Circular. //---------------------------------------------------------------------------- /* * I0, I1, and I2 registers are used to read the input data. I3 register is used * to read the coefficients. P0 and P1 registers are used for writing the output * data. */ M0 = 12 (X); // All these initializations are used in the M1 = 16 (X); // modification of address offsets. M2 = 128 (X); P2 = 16; P3 = 32 (X); P4 = -110 (X); P5 = -62 (X); P0 = 2(X); // Prescale the input to get the correct precision. i0=b0; i1=b0; lsetup (.0, .1) LC0 = P3; r0=[i0++]; .0: r1=r0<<3 (v) || r0=[i0++] ; .1: [i1++]=r1; /* * B0 points to the "in" buffer. * B2 points to "temp" buffer in the first iteration. */ lsetup (.2, .3) LC0 = P0; .2: I0 = B0; // I0 points to Input Element (0, 0). I1 = B0; // Element 1 and 0 is read in R0. I1 += M0 || R0 = [I0++]; // I1 points to Input Element (0, 6). I2 = I1; // Element 6 is read into R3.H. I2 -= 4 || R3.H = W[I1++]; // I2 points to Input Element (0, 4). I3 = B3; // I3 points to Coefficients. P0 = B2; // P0 points to temporary array Element // (0, 0). P1 = B2; // P1 points to temporary array. R7 = [P1++P2] || R2 = [I2++]; // P1 points to temporary array // Element (1, 0). // R7 is a dummy read. X4,X5 // are read into R2. R3.L = W[I1--]; // X7 is read into R3.L. R1.H = W[I0++]; // X2 is read into R1.H. /* * X0 = (X0 + X7) / 2. * X1 = (X1 + X6) / 2. * X6 = (X1 - X6) / 2. * X7 = (X0 - X7) / 2. * It reads the data 3 in R1.L. */ R0 = R0 +|+ R3, R3 = R0 -|- R3 || R1.L = W[I0++] || NOP; /* * X2 = (X2 + X5) / 2. * X3 = (X3 + X4) / 2. * X4 = (X3 - X4) / 2. * X5 = (X2 - X5) / 2. * R7 = C4 = cos(4*pi/16) */ R1 = R1 +|+ R2, R2 = R1 -|- R2 (CO) || NOP || R7 = [I3++]; /* * At the end of stage 1 R0 has (1,0), R1 has (2,3), R2 has (4, 5) and * R3 has (6,7). * Where the notation (x, y) represents uper/lower half pairs. */ /* * X0 = X0 + X3. * X1 = X1 + X2. * X2 = X1 - X2. * X3 = X0 - X3. */ R0 = R0 +|+ R1, R1 = R0 -|- R1; lsetup (.row0, .row1) LC1 = P2 >> 1; // 1d dct, loops 8x .row0: /* * This is part 2 computation continued..... * A1 = X6 * cos(pi/4) * A0 = X6 * cos(pi/4) * A1 = A1 - X5 * cos(pi/4) * A0 = A0 + X5 * cos(pi/4). * The instruction W[I0] = R3.L is used for packing it to R2.L. */ A1=R3.H*R7.l, A0=R3.H*R7.l || I1+=M1 || W[I0] = R3.L; R4.H=(A1-=R2.L*R7.l), R4.L=(A0+=R2.L*R7.l) || I2+=M0 || NOP; /* R0 = (X1,X0) R1 = (X2,X3) R4 = (X5, X6). */ /* * A1 = X0 * cos(pi/4) * A0 = X0 * cos(pi/4) * A1 = A1 - X1 * cos(pi/4) * A0 = A0 + X1 * cos(pi/4) * R7 = (C2,C6) */ A1=R0.L*R7.h, A0=R0.L*R7.h || NOP || R3.H=W[I1++]; R5.H=(A1-=R0.H*R7.h),R5.L=(A0+=R0.H*R7.h) || R7=[I3++] || NOP; /* * A1 = X2 * cos(3pi/8) * A0 = X3 * cos(3pi/8) * A1 = A1 + X3 * cos(pi/8) * A0 = A0 - X2 * cos(pi/8) * R3 = cos(pi/4) * R7 = (cos(7pi/8),cos(pi/8)) * X4 = X4 + X5. * X5 = X4 - X5. * X6 = X7 - X6. * X7 = X7 + X6. */ A1=R1.H*R7.L, A0=R1.L*R7.L || W[P0++P3]=R5.L || R2.L=W[I0]; R2=R2+|+R4, R4=R2-|-R4 || I0+=4 || R3.L=W[I1--]; R6.H=(A1+=R1.L*R7.H),R6.L=(A0 -= R1.H * R7.H) || I0+=4 || R7=[I3++]; /* R2 = (X4, X7) R4 = (X5,X6) R5 = (X1, X0) R6 = (X2,X3). */ /* * A1 = X4 * cos(7pi/16) * A0 = X7 * cos(7pi/16) * A1 = A1 + X7 * cos(pi/16) * A0 = A0 - X4 * cos(pi/16) */ A1=R2.H*R7.L, A0=R2.L*R7.L || W[P0++P3]=R6.H || R0=[I0++]; R2.H=(A1+=R2.L*R7.H),R2.L=(A0-=R2.H*R7.H) || W[P0++P3]=R5.H || R7=[I3++]; /* * A1 = X5 * cos(3pi/16) * A0 = X6 * cos(3pi/16) * A1 = A1 + X6 * cos(5pi/16) * A0 = A0 - X5 * cos(5pi/16) * The output values are written. */ A1=R4.H*R7.H, A0=R4.L*R7.H || W[P0++P2]=R6.L || R1.H=W[I0++]; R4.H=(A1+=R4.L*R7.L),R4.L=(A0-=R4.H*R7.L) || W[P0++P4]=R2.L || R1.L=W[I0++]; /* Beginning of next stage, **pipelined** + drain and store the rest of the column store. */ R0=R0+|+R3,R3=R0-|-R3 || W[P1++P3]=R2.H || R2=[I2++]; R1=R1+|+R2,R2=R1-|-R2 (CO) || W[P1++P3]=R4.L || R7=[I3++]; .row1: R0=R0+|+R1,R1=R0-|-R1 || W[P1++P5]=R4.H || NOP; // Exchange input with output. B1 = B0; B0 = B2; .3: B2 = B1; L3=0; (r7:4,p5:3) = [sp++]; RTS; DEFUN_END(fdct)