yading@10: /*
yading@10:  * Blackfin Pixel Operations
yading@10:  * Copyright (C) 2007 Marc Hoffman <marc.hoffman@analog.com>
yading@10:  *
yading@10:  * This file is part of FFmpeg.
yading@10:  *
yading@10:  * FFmpeg is free software; you can redistribute it and/or
yading@10:  * modify it under the terms of the GNU Lesser General Public
yading@10:  * License as published by the Free Software Foundation; either
yading@10:  * version 2.1 of the License, or (at your option) any later version.
yading@10:  *
yading@10:  * FFmpeg is distributed in the hope that it will be useful,
yading@10:  * but WITHOUT ANY WARRANTY; without even the implied warranty of
yading@10:  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
yading@10:  * Lesser General Public License for more details.
yading@10:  *
yading@10:  * You should have received a copy of the GNU Lesser General Public
yading@10:  * License along with FFmpeg; if not, write to the Free Software
yading@10:  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
yading@10:  */
yading@10: #include "config_bfin.h"
yading@10: 
yading@10: /*
yading@10:   motion compensation
yading@10:   primitives
yading@10: 
yading@10:      * Halfpel motion compensation with rounding (a+b+1)>>1.
yading@10:      * This is an array[4][4] of motion compensation funcions for 4
yading@10:      * horizontal blocksizes (8,16) and the 4 halfpel positions<br>
yading@10:      * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
yading@10:      * @param block destination where the result is stored
yading@10:      * @param pixels source
yading@10:      * @param line_size number of bytes in a horizontal line of block
yading@10:      * @param h height
yading@10: 
yading@10: */
yading@10: 
yading@10: DEFUN(put_pixels8uc,mL1,
yading@10:         (uint8_t *block, const uint8_t *s0, const uint8_t *s1,
yading@10:                  int dest_size, int line_size, int h)):
yading@10:         i3=r0;        // dest
yading@10:         i0=r1;        // src0
yading@10:         i1=r2;        // src1
yading@10:         r0=[sp+12];   // dest_size
yading@10:         r2=[sp+16];   // line_size
yading@10:         p0=[sp+20];   // h
yading@10:         [--sp] = (r7:6);
yading@10:         r0+=-4;
yading@10:         m3=r0;
yading@10:         r2+=-8;
yading@10:         m0=r2;
yading@10:         LSETUP(pp8$0,pp8$1) LC0=P0;
yading@10:         DISALGNEXCPT                || R0 = [I0++]  || R2  =[I1++];
yading@10: 
yading@10: pp8$0:  DISALGNEXCPT                || R1 = [I0++]  || R3  =[I1++];
yading@10:         R6 = BYTEOP1P(R1:0,R3:2)    || R0 = [I0++M0]|| R2  =[I1++M0];
yading@10:         R7 = BYTEOP1P(R1:0,R3:2)(R) || R0 = [I0++]  || [I3++] = R6 ;
yading@10: pp8$1:  DISALGNEXCPT                || R2 = [I1++]  || [I3++M3] = R7;
yading@10: 
yading@10:         (r7:6) = [sp++];
yading@10:         RTS;
yading@10: DEFUN_END(put_pixels8uc)
yading@10: 
yading@10: DEFUN(put_pixels16uc,mL1,
yading@10:         (uint8_t *block, const uint8_t *s0, const uint8_t *s1,
yading@10:                  int dest_size, int line_size, int h)):
yading@10:         link 0;
yading@10:         [--sp] = (r7:6);
yading@10:         i3=r0;        // dest
yading@10:         i0=r1;        // src0
yading@10:         i1=r2;        // src1
yading@10:         r0=[fp+20];   // dest_size
yading@10:         r2=[fp+24];   // line_size
yading@10:         p0=[fp+28];   // h
yading@10: 
yading@10: 
yading@10:         r0+=-12;
yading@10:         m3=r0;        // line_size
yading@10:         r2+=-16;
yading@10:         m0=r2;
yading@10: 
yading@10:         LSETUP(pp16$0,pp16$1) LC0=P0;
yading@10:          DISALGNEXCPT                || R0 = [I0++]   || R2  =[I1++];
yading@10: 
yading@10: pp16$0:  DISALGNEXCPT                || R1 = [I0++]   || R3  =[I1++];
yading@10:          R6 = BYTEOP1P(R1:0,R3:2)    || R0 = [I0++]   || R2  =[I1++];
yading@10:          R7 = BYTEOP1P(R1:0,R3:2)(R) || R1 = [I0++]   || R3  =[I1++];
yading@10:          [I3++] = R6;
yading@10:          R6 = BYTEOP1P(R1:0,R3:2)    || R0 = [I0++M0] || R2  =[I1++M0];
yading@10:          R7 = BYTEOP1P(R1:0,R3:2)(R) || R0 = [I0++]   || [I3++] = R7 ;
yading@10:          [I3++] = R6;
yading@10: pp16$1:  DISALGNEXCPT                || R2 = [I1++]   || [I3++M3] = R7;
yading@10: 
yading@10:         (r7:6) = [sp++];
yading@10:         unlink;
yading@10:         RTS;
yading@10: DEFUN_END(put_pixels16uc)
yading@10: 
yading@10: 
yading@10: 
yading@10: 
yading@10: 
yading@10: 
yading@10: DEFUN(put_pixels8uc_nornd,mL1,
yading@10:         (uint8_t *block, const uint8_t *s0, const uint8_t *s1,
yading@10:                  int line_size, int h)):
yading@10:         i3=r0;        // dest
yading@10:         i0=r1;        // src0
yading@10:         i1=r2;        // src1
yading@10:         r2=[sp+12];   // line_size
yading@10:         p0=[sp+16];   // h
yading@10:         [--sp] = (r7:6);
yading@10:         r2+=-4;
yading@10:         m3=r2;
yading@10:         r2+=-4;
yading@10:         m0=r2;
yading@10:         LSETUP(pp8$2,pp8$3) LC0=P0;
yading@10:         DISALGNEXCPT                || R0 = [I0++]  || R2  =[I1++];
yading@10: 
yading@10: pp8$2:  DISALGNEXCPT                || R1 = [I0++]  || R3  =[I1++];
yading@10:         R6 = BYTEOP1P(R1:0,R3:2)(T)  || R0 = [I0++M0]|| R2  =[I1++M0];
yading@10:         R7 = BYTEOP1P(R1:0,R3:2)(T,R) || R0 = [I0++]  || [I3++] = R6 ;
yading@10: pp8$3:  DISALGNEXCPT                || R2 = [I1++]  || [I3++M3] = R7;
yading@10: 
yading@10:         (r7:6) = [sp++];
yading@10:         RTS;
yading@10: DEFUN_END(put_pixels8uc_nornd)
yading@10: 
yading@10: DEFUN(put_pixels16uc_nornd,mL1,
yading@10:         (uint8_t *block, const uint8_t *s0, const uint8_t *s1,
yading@10:                  int line_size, int h)):
yading@10:         i3=r0;        // dest
yading@10:         i0=r1;        // src0
yading@10:         i1=r2;        // src1
yading@10:         r2=[sp+12];   // line_size
yading@10:         p0=[sp+16];   // h
yading@10: 
yading@10:         [--sp] = (r7:6);
yading@10:         r2+=-12;
yading@10:         m3=r2;        // line_size
yading@10:         r2+=-4;
yading@10:         m0=r2;
yading@10: 
yading@10:         LSETUP(pp16$2,pp16$3) LC0=P0;
yading@10:         DISALGNEXCPT                || R0 = [I0++]   || R2  =[I1++];
yading@10: 
yading@10: pp16$2:
yading@10:         DISALGNEXCPT                || R1 = [I0++]   || R3  =[I1++];
yading@10:         R6 = BYTEOP1P(R1:0,R3:2)(T)    || R0 = [I0++]   || R2  =[I1++];
yading@10:         R7 = BYTEOP1P(R1:0,R3:2)(T,R) || R1 = [I0++]   || R3  =[I1++];
yading@10:         [I3++] = R6;
yading@10: 
yading@10:         R6 = BYTEOP1P(R1:0,R3:2)(T)    || R0 = [I0++M0] || R2  =[I1++M0];
yading@10:         R7 = BYTEOP1P(R1:0,R3:2)(T,R) || R0 = [I0++]   || [I3++] = R7 ;
yading@10:         [I3++] = R6;
yading@10: pp16$3: DISALGNEXCPT                || R2 = [I1++]   || [I3++M3] = R7;
yading@10: 
yading@10:         (r7:6) = [sp++];
yading@10: 
yading@10:         RTS;
yading@10: DEFUN_END(put_pixels16uc_nornd)
yading@10: 
yading@10: DEFUN(z_put_pixels16_xy2,mL1,
yading@10:         (uint8_t *block, const uint8_t *s0,
yading@10:                  int dest_size, int line_size, int h)):
yading@10:         link 0;
yading@10:         [--sp] = (r7:4);
yading@10:         i3=r0;        // dest
yading@10:         i0=r1;        // src0--> pixels
yading@10:         i1=r1;        // src1--> pixels + line_size
yading@10:         r2+=-12;
yading@10:         m2=r2;        // m2=dest_width-4
yading@10:         r2=[fp+20];
yading@10:         m3=r2;        // line_size
yading@10:         p0=[fp+24];   // h
yading@10:         r2+=-16;
yading@10:         i1+=m3;       /* src1 + line_size */
yading@10:         m0=r2;        /* line-size - 20 */
yading@10: 
yading@10:         B0 = I0;
yading@10:         B1 = I1;
yading@10:         B3 = I3;
yading@10: 
yading@10:         DISALGNEXCPT                       || R0 = [I0++] || R2  =[I1++];
yading@10: 
yading@10:         LSETUP(LS$16E,LE$16E) LC0=P0;
yading@10: LS$16E: DISALGNEXCPT                       || R1 = [I0++] || R3  =[I1++];
yading@10:         R4 = BYTEOP2P (R3:2,R1:0) (RNDL)   || R0 = [I0++] || R2  =[I1++];
yading@10:         R5 = BYTEOP2P (R3:2,R1:0) (RNDL,R) || R1 = [I0++] || [I3++] = R4 ;
yading@10:         DISALGNEXCPT                       || R3 = [I1++] || [I3++] = R5;
yading@10:         R4 = BYTEOP2P (R3:2,R1:0) (RNDL)   || R0 = [I0++M0]|| R2  = [I1++M0];
yading@10:         R5 = BYTEOP2P (R3:2,R1:0) (RNDL,R) || R0 = [I0++] || [I3++] = R4 ;
yading@10: LE$16E: DISALGNEXCPT                       || R2 = [I1++] || [I3++M2] = R5;
yading@10: 
yading@10:         M1 = 1;
yading@10:         I3 = B3;
yading@10:         I1 = B1;
yading@10:         I0 = B0;
yading@10: 
yading@10:         I0 += M1;
yading@10:         I1 += M1;
yading@10: 
yading@10:         DISALGNEXCPT                       || R0 = [I0++] || R2  =[I1++];
yading@10:         LSETUP(LS$16O,LE$16O) LC0=P0;
yading@10: LS$16O: DISALGNEXCPT                       || R1 = [I0++] || R3  =[I1++];
yading@10:         R4 = BYTEOP2P (R3:2,R1:0) (RNDH)   || R0 = [I0++] || R2  =[I1++];
yading@10:         R5 = BYTEOP2P (R3:2,R1:0) (RNDH,R) || R1 = [I0++] || R6  =[I3++];
yading@10:         R4 = R4 +|+ R6                       || R7 = [I3--];
yading@10:         R5 = R5 +|+ R7                       || [I3++] = R4;
yading@10:         DISALGNEXCPT                       || R3  =[I1++] || [I3++] = R5;
yading@10:         R4 = BYTEOP2P (R3:2,R1:0) (RNDH)   || R0 = [I0++M0]|| R2  = [I1++M0];
yading@10:         R5 = BYTEOP2P (R3:2,R1:0) (RNDH,R) || R0 = [I0++] || R6 = [I3++];
yading@10:         R4 = R4 +|+ R6                       || R7 = [I3--];
yading@10:         R5 = R5 +|+ R7                       || [I3++] = R4;
yading@10: LE$16O: DISALGNEXCPT                       || R2 = [I1++] || [I3++M2] = R5;
yading@10: 
yading@10:         (r7:4) = [sp++];
yading@10:         unlink;
yading@10:         rts;
yading@10: DEFUN_END(z_put_pixels16_xy2)
yading@10: 
yading@10: DEFUN(put_pixels16_xy2_nornd,mL1,
yading@10:         (uint8_t *block, const uint8_t *s0,
yading@10:                  int line_size, int h)):
yading@10:         link 0;
yading@10:         [--sp] = (r7:4);
yading@10:         i3=r0;        // dest
yading@10:         i0=r1;        // src0--> pixels
yading@10:         i1=r1;        // src1--> pixels + line_size
yading@10:         m3=r2;
yading@10:         r2+=-12;
yading@10:         m2=r2;
yading@10:         r2+=-4;
yading@10:         i1+=m3;       /* src1 + line_size */
yading@10:         m0=r2;        /* line-size - 20 */
yading@10:         p0=[fp+20];   // h
yading@10: 
yading@10:         B0=I0;
yading@10:         B1=I1;
yading@10:         B3=I3;
yading@10: 
yading@10:         DISALGNEXCPT                       || R0 = [I0++] || R2  =[I1++];
yading@10: 
yading@10:         LSETUP(LS$16ET,LE$16ET) LC0=P0;
yading@10: LS$16ET:DISALGNEXCPT                       || R1 = [I0++] || R3  =[I1++];
yading@10:         R4 = BYTEOP2P (R3:2,R1:0) (TL)     || R0 = [I0++] || R2  =[I1++];
yading@10:         R5 = BYTEOP2P (R3:2,R1:0) (TL,R)   || R1 = [I0++] || [I3++] = R4 ;
yading@10:         DISALGNEXCPT                       || R3 = [I1++] || [I3++] = R5;
yading@10:         R4 = BYTEOP2P (R3:2,R1:0) (TL)     || R0 = [I0++M0]|| R2  = [I1++M0];
yading@10:         R5 = BYTEOP2P (R3:2,R1:0) (TL,R)   || R0 = [I0++] || [I3++] = R4 ;
yading@10: LE$16ET:DISALGNEXCPT                       || R2 = [I1++] || [I3++M2] = R5;
yading@10: 
yading@10:         M1 = 1;
yading@10:         I3=B3;
yading@10:         I1=B1;
yading@10:         I0=B0;
yading@10: 
yading@10:         I0 += M1;
yading@10:         I1 += M1;
yading@10: 
yading@10:         DISALGNEXCPT                       || R0 = [I0++] || R2  =[I1++];
yading@10:         LSETUP(LS$16OT,LE$16OT) LC0=P0;
yading@10: LS$16OT:DISALGNEXCPT                       || R1 = [I0++] || R3  =[I1++];
yading@10:         R4 = BYTEOP2P (R3:2,R1:0) (TH)     || R0 = [I0++] || R2  =[I1++];
yading@10:         R5 = BYTEOP2P (R3:2,R1:0) (TH,R)   || R1 = [I0++] || R6  =[I3++];
yading@10:         R4 = R4 +|+ R6                                    || R7 = [I3--];
yading@10:         R5 = R5 +|+ R7                                    || [I3++] = R4;
yading@10:         DISALGNEXCPT                       || R3  =[I1++] || [I3++] = R5;
yading@10:         R4 = BYTEOP2P (R3:2,R1:0) (TH)     || R0 = [I0++M0]|| R2  = [I1++M0];
yading@10:         R5 = BYTEOP2P (R3:2,R1:0) (TH,R)   || R0 = [I0++] || R6 = [I3++];
yading@10:         R4 = R4 +|+ R6                                    || R7 = [I3--];
yading@10:         R5 = R5 +|+ R7                                    || [I3++] = R4;
yading@10: LE$16OT:DISALGNEXCPT                       || R2 = [I1++] || [I3++M2] = R5;
yading@10: 
yading@10:         (r7:4) = [sp++];
yading@10:         unlink;
yading@10:         rts;
yading@10: DEFUN_END(put_pixels16_xy2_nornd)
yading@10: 
yading@10: DEFUN(z_put_pixels8_xy2,mL1,
yading@10:         (uint8_t *block, const uint8_t *s0,
yading@10:                  int dest_size, int line_size, int h)):
yading@10:         link 0;
yading@10:         [--sp] = (r7:4);
yading@10:         i3=r0;        // dest
yading@10:         i0=r1;        // src0--> pixels
yading@10:         i1=r1;        // src1--> pixels + line_size
yading@10:         r2+=-4;
yading@10:         m2=r2;        // m2=dest_width-4
yading@10:         r2=[fp+20];
yading@10:         m3=r2;        // line_size
yading@10:         p0=[fp+24];   // h
yading@10:         r2+=-8;
yading@10:         i1+=m3;       /* src1 + line_size */
yading@10:         m0=r2;        /* line-size - 20 */
yading@10: 
yading@10:         b0 = I0;
yading@10:         b1 = I1;
yading@10:         b3 = I3;
yading@10: 
yading@10:         LSETUP(LS$8E,LE$8E) LC0=P0;
yading@10:         DISALGNEXCPT                       || R0 = [I0++]   || R2  =[I1++];
yading@10: LS$8E:  DISALGNEXCPT                       || R1 = [I0++]   || R3  =[I1++];
yading@10:         R4 = BYTEOP2P (R3:2,R1:0) (RNDL)   || R0 = [I0++M0] || R2  =[I1++M0];
yading@10:         R5 = BYTEOP2P (R3:2,R1:0) (RNDL,R) || R0 = [I0++]   || [I3++] = R4 ;
yading@10: LE$8E:  DISALGNEXCPT                       || R2 = [I1++]   || [I3++M2] = R5;
yading@10: 
yading@10:         M1 = 1;
yading@10:         I3 = b3;
yading@10:         I1 = b1;
yading@10:         I0 = b0;
yading@10: 
yading@10:         I0 += M1;
yading@10:         I1 += M1;
yading@10: 
yading@10:         LSETUP(LS$8O,LE$8O) LC0=P0;
yading@10:         DISALGNEXCPT                       || R0 = [I0++]   || R2  =[I1++];
yading@10: LS$8O:  DISALGNEXCPT                       || R1 = [I0++]   || R3  =[I1++];
yading@10:         R4 = BYTEOP2P (R3:2,R1:0) (RNDH)   || R0 = [I0++M0] || R2  =[I1++M0];
yading@10:         R5 = BYTEOP2P (R3:2,R1:0) (RNDH,R) || R0 = [I0++]   || R6  =[I3++];
yading@10:         R4 = R4 +|+ R6                                      || R7 = [I3--];
yading@10:         R5 = R5 +|+ R7                                      || [I3++] = R4;
yading@10: LE$8O:  DISALGNEXCPT                       || R2  =[I1++]   || [I3++M2] = R5;
yading@10: 
yading@10:         (r7:4) = [sp++];
yading@10:         unlink;
yading@10:         rts;
yading@10: DEFUN_END(z_put_pixels8_xy2)
yading@10: 
yading@10: DEFUN(put_pixels8_xy2_nornd,mL1,
yading@10:         (uint8_t *block, const uint8_t *s0, int line_size, int h)):
yading@10:         link 0;
yading@10:         [--sp] = (r7:4);
yading@10:         i3=r0;        // dest
yading@10:         i0=r1;        // src0--> pixels
yading@10:         i1=r1;        // src1--> pixels + line_size
yading@10:         m3=r2;
yading@10:         r2+=-4;
yading@10:         m2=r2;
yading@10:         r2+=-4;
yading@10:         i1+=m3;       /* src1 + line_size */
yading@10:         m0=r2;        /* line-size - 20 */
yading@10:         p0=[fp+20];   // h
yading@10: 
yading@10: 
yading@10:         b0 = I0;
yading@10:         b1 = I1;
yading@10:         b3 = I3;
yading@10: 
yading@10:         LSETUP(LS$8ET,LE$8ET) LC0=P0;
yading@10:         DISALGNEXCPT                       || R0 = [I0++]   || R2  =[I1++];
yading@10: 
yading@10: LS$8ET: DISALGNEXCPT                       || R1 = [I0++]   || R3 = [I1++];
yading@10:         R4 = BYTEOP2P (R3:2,R1:0) (TL)     || R0 = [I0++M0] || R2 = [I1++M0];
yading@10:         R5 = BYTEOP2P (R3:2,R1:0) (TL,R)   || R0 = [I0++]   || [I3++] = R4 ;
yading@10: LE$8ET: DISALGNEXCPT                       || R2 = [I1++]   || [I3++M2] = R5;
yading@10: 
yading@10:         M1 = 1;
yading@10:         I3 = b3;
yading@10:         I1 = b1;
yading@10:         I0 = b0;
yading@10: 
yading@10:         I0 += M1;
yading@10:         I1 += M1;
yading@10: 
yading@10:         LSETUP(LS$8OT,LE$8OT) LC0=P0;
yading@10:         DISALGNEXCPT                       || R0 = [I0++]   || R2 = [I1++];
yading@10: 
yading@10: LS$8OT: DISALGNEXCPT                       || R1 = [I0++]   || R3 = [I1++];
yading@10:         R4 = BYTEOP2P (R3:2,R1:0) (TH)     || R0 = [I0++M0] || R2 = [I1++M0];
yading@10:         R5 = BYTEOP2P (R3:2,R1:0) (TH,R)   || R0 = [I0++]   || R6 = [I3++];
yading@10:         R4 = R4 +|+ R6                                      || R7 = [I3--];
yading@10:         R5 = R5 +|+ R7                                      || [I3++] = R4;
yading@10: LE$8OT: DISALGNEXCPT                       || R2  =[I1++]   || [I3++M2] = R5;
yading@10: 
yading@10:         (r7:4) = [sp++];
yading@10:         unlink;
yading@10:         rts;
yading@10: DEFUN_END(put_pixels8_xy2_nornd)