📄 pixels_bfin.s
字号:
M1 = 1; I3 = b3; I1 = b1; I0 = b0; I0 += M1; I1 += M1; LSETUP(LS$8O,LE$8O) LC0=P0; DISALGNEXCPT || R0 = [I0++] || R2 =[I1++];LS$8O: DISALGNEXCPT || R1 = [I0++] || R3 =[I1++]; R4 = BYTEOP2P (R3:2,R1:0) (RNDH) || R0 = [I0++M0] || R2 =[I1++M0]; R5 = BYTEOP2P (R3:2,R1:0) (RNDH,R) || R0 = [I0++] || R6 =[I3++]; R4 = R4 +|+ R6 || R7 = [I3--]; R5 = R5 +|+ R7 || [I3++] = R4;LE$8O: DISALGNEXCPT || R2 =[I1++] || [I3++M2] = R5; (r7:4) = [sp++]; unlink; rts;DEFUN_END(z_put_pixels8_xy2)DEFUN(put_pixels8_xy2_nornd,mL1, (uint8_t *block, const uint8_t *s0, int line_size, int h)): link 0; [--sp] = (r7:4); i3=r0; // dest i0=r1; // src0--> pixels i1=r1; // src1--> pixels + line_size m3=r2; r2+=-4; m2=r2; r2+=-4; i1+=m3; /* src1 + line_size */ m0=r2; /* line-size - 20 */ p0=[fp+20]; // h b0 = I0; b1 = I1; b3 = I3; LSETUP(LS$8ET,LE$8ET) LC0=P0; DISALGNEXCPT || R0 = [I0++] || R2 =[I1++];LS$8ET: DISALGNEXCPT || R1 = [I0++] || R3 = [I1++]; R4 = BYTEOP2P (R3:2,R1:0) (TL) || R0 = [I0++M0] || R2 = [I1++M0]; R5 = BYTEOP2P (R3:2,R1:0) (TL,R) || R0 = [I0++] || [I3++] = R4 ;LE$8ET: DISALGNEXCPT || R2 = [I1++] || [I3++M2] = R5; M1 = 1; I3 = b3; I1 = b1; I0 = b0; I0 += M1; I1 += M1; LSETUP(LS$8OT,LE$8OT) LC0=P0; DISALGNEXCPT || R0 = [I0++] || R2 = [I1++];LS$8OT: DISALGNEXCPT || R1 = [I0++] || R3 = [I1++]; R4 = BYTEOP2P (R3:2,R1:0) (TH) || R0 = [I0++M0] || R2 = [I1++M0]; R5 = BYTEOP2P (R3:2,R1:0) (TH,R) || R0 = [I0++] || R6 = [I3++]; R4 = R4 +|+ R6 || R7 = [I3--]; R5 = R5 +|+ R7 || [I3++] = R4;LE$8OT: DISALGNEXCPT || R2 =[I1++] || [I3++M2] = R5; (r7:4) = [sp++]; unlink; rts;DEFUN(diff_pixels,mL1, (DCTELEM *block, uint8_t *s1, uint8_t *s2, int stride)): link 0; [--sp] = (r7:4); p0=8; i3=r0; // block i0=r1; // s1 i1=r2; // s2 r2=[fp+20]; // stride r2+=-8; m0=r2; LSETUP(.LS0,.LE0) LC0=P0; DISALGNEXCPT || R0 = [I0++] || R2 =[I1++];.LS0: DISALGNEXCPT || R1 = [I0++] || R3 = [I1++]; (R5,R4) = BYTEOP16M (R1:0,R3:2) || R0 = [I0++M0] || R2 = [I1++M0]; (R7,R6) = BYTEOP16M (R1:0,R3:2) (R)|| R0 = [I0++] || [I3++] = R4; DISALGNEXCPT || R2 = [I1++] || [I3++] = R5; [i3++]=r6;.LE0: [i3++]=r7; (r7:4) = [sp++]; unlink; rts;DEFUN_END(put_pixels8_xy2_nornd)/* for (i = 0; i < 16; i++) { for (j = 0; j < 16; j++) { sum += pix[j]; } pix += line_size; }*/DEFUN(pix_sum,mL1, (uint8_t *p, int stride)): link 0; [--sp] = (r7:4); p0=8; i0=r0; // s1 i1=r0; m1=r1; r1=r1+r1; r1+=-16; // stride m0=r1; i1+=m1; r6=0; LSETUP(LS$PS,LE$PS) LC0=P0; DISALGNEXCPT || R0 = [I0++] || R2 =[I1++];LS$PS: DISALGNEXCPT || R1 = [I0++] || R3 = [I1++]; (R5,R4) = BYTEOP16P (R3:2,R1:0) || R0 = [I0++] || R2 = [I1++]; r6=r6+|+r5; r6=r6+|+r4; (R5,R4) = BYTEOP16P (R3:2,R1:0) (R)|| R1 = [I0++] || R3 = [I1++]; r6=r6+|+r5; r6=r6+|+r4; (R5,R4) = BYTEOP16P (R3:2,R1:0) || R0 = [I0++m0] || R2 = [I1++m0]; r6=r6+|+r5; r6=r6+|+r4; (R5,R4) = BYTEOP16P (R3:2,R1:0) (R)|| R0 = [I0++] || R2 = [I1++]; r6=r6+|+r5;LE$PS: r6=r6+|+r4; r0.l=r6.l+r6.h; r0.h=0; (r7:4) = [sp++]; unlink; rts;DEFUN_END(pix_sum)DEFUN(get_pixels,mL1, (DCTELEM *restrict block, const uint8_t *pixels, int line_size)): [--sp] = (r7:4); i3=r0; // dest i0=r1; // src0 p0=8; r2+=-8; m0=r2; LSETUP(gp8$0,gp8$1) LC0=P0; DISALGNEXCPT || R0 = [I0++]; DISALGNEXCPT || R1 = [I0++];gp8$0: (R7,R6) = byteunpack R1:0 || R0 = [I0++M0]; (R5,R4) = byteunpack R1:0 (R) || R0 = [I0++] || [I3++]=R6; DISALGNEXCPT || R1 = [I0++] || [I3++]=R7; [I3++]=R4;gp8$1: [I3++]=R5 (r7:4) = [sp++]; RTS;DEFUN_END(get_pixels)/* sad = sad16x16 (ubyte *mb, ubyte *refwin, srcwidth, refwinwidth, h) *//* 91 cycles */DEFUN(z_sad16x16,mL1, (uint8_t *blk1, uint8_t *blk2, int dsz, int line_size, int h)): link 0; I0 = R0; I1 = R1; A1 = A0 = 0; R0 = [sp+20]; // rwidth P2 = [sp+24]; // height R3 = 16; R0 = R0 - R3; R3 = R2 - R3; M1 = R0; M0 = R3; DISALGNEXCPT || R0 = [I0++] || R2 = [I1++]; LSETUP (s$16, e$16) LC0=P2;s$16: DISALGNEXCPT || R1 = [I0++] || R3 = [I1++]; SAA (R1:0,R3:2) || R0 = [I0++] || R2 = [I1++]; SAA (R1:0,R3:2) (R) || R1 = [I0++] || R3 = [I1++]; SAA (R1:0,R3:2) || R0 = [I0++M0] || R2 = [I1++M1];e$16: SAA (R1:0,R3:2) (R) || R0 = [I0++] || R2 = [I1++]; R3=A1.L+A1.H, R2=A0.L+A0.H ; R0 = R2 + R3 ; unlink; RTS;DEFUN_END(z_sad16x16)/* sad = sad8x8 (ubyte *mb, ubyte *refwin, int srcwidth, int refwinwidth, int h) *//* 36 cycles */DEFUN(z_sad8x8,mL1, (uint8_t *blk1, uint8_t *blk2, int dsz, int line_size, int h)): I0 = R0; I1 = R1; A1 = A0 = 0; r0 = [sp+12]; // rwidth P2 = [sp+16]; //height R3 = 8; R0 = R0 - R3; R3 = R2 - R3; M0 = R3; M1 = R0; LSETUP (s$8, e$8) LC0=P2; DISALGNEXCPT || R0 = [I0++] || R2 = [I1++]; DISALGNEXCPT || R1 = [I0++] || R3 = [I1++];s$8: SAA (R1:0,R3:2) || R0 = [I0++M0] || R2 = [I1++M1]; SAA (R1:0,R3:2) (R) || R0 = [I0++] || R2 = [I1++];e$8: DISALGNEXCPT || R1 = [I0++] || R3 = [I1++]; R3=A1.L+A1.H, R2=A0.L+A0.H ; R0 = R2 + R3 ; RTS;DEFUN_END(z_sad8x8)DEFUN(pix_norm1,mL1, (uint8_t * pix, int line_size)): [--SP]=(R7:4,P5:3); // Fetch the input arguments. P1 = R0; // pix P0 = R1; // line_size P5 = 16; // loop ctr. P0 -= P5; M0 = P0; // M0 = line_size-16; // Now for the real work. A1 = A0 = 0; lsetup(_pix_norm1_blkfn_loopStart, _pix_norm1_blkfn_loopEnd) LC1 = P5; I0 = P1; DISALGNEXCPT || r0 = [i0++];_pix_norm1_blkfn_loopStart: // following unpacks pix1[0..15] pix1+line_size[0..15] DISALGNEXCPT || r1 = [i0++]; (r5, r4) = byteunpack r1:0 || r0 = [i0++]; a1 += r5.h * r5.h, a0 += r5.l * r5.l (is); a1 += r4.h * r4.h, a0 += r4.l * r4.l (is); (r5, r4) = byteunpack r1:0(r) || r1 = [i0++]; a1 += r5.h * r5.h, a0 += r5.l * r5.l (is); a1 += r4.h * r4.h, a0 += r4.l * r4.l (is); (r5, r4) = byteunpack r1:0 || r0 = [i0++M0]; a1 += r5.h * r5.h, a0 += r5.l * r5.l (is); a1 += r4.h * r4.h, a0 += r4.l * r4.l (is); (r5, r4) = byteunpack r1:0(r) || r0 = [i0++]; a1 += r5.h * r5.h, a0 += r5.l * r5.l (is);_pix_norm1_blkfn_loopEnd: a1 += r4.h * r4.h, a0 += r4.l * r4.l (is);// Clean up at the end: R2 = A0, R3 = A1; R0 = R2 + R3 (S); (R7:4,P5:3)=[SP++]; RTS;DEFUN_END(pix_norm1)DEFUN(sse4,mL1, (void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)): link 0; [--sp] = (r7:6); p0=[fp+24]; // h i0=r1; // pix1 i1=r2; // pix2 r2=[fp+20]; // line_size r2+=-4; m0=r2; a0=a1=0; LSETUP(.S40,.E40) LC0=P0; DISALGNEXCPT || R0 = [I0++] || R2 =[I1++];.S40: DISALGNEXCPT || R1 = [I0++M0] || R3 = [I1++M0]; (R7,R6) = BYTEOP16M (R1:0,R3:2); a0 += r7.l * r7.l, a1 += r7.h * r7.h (is);.E40: a0 += r6.l * r6.l, a1 += r6.h * r6.h (is); a0 += a1; r0 = a0; (r7:6) = [sp++]; unlink; rts;DEFUN_END(sse4)DEFUN(sse8,mL1, (void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)): link 0; [--sp] = (r7:6); p0=[fp+24]; // h i0=r1; // pix1 i1=r2; // pix2 r2=[fp+20]; // line_size r2+=-8; m0=r2; a0=a1=0; LSETUP(.S80,.E80) LC0=P0; DISALGNEXCPT || R0 = [I0++] || R2 =[I1++];.S80: DISALGNEXCPT || R1 = [I0++] || R3 = [I1++]; (R7,R6) = BYTEOP16M (R1:0,R3:2) || R0 = [I0++M0] || R2 = [I1++M0]; a0 += r7.l * r7.l, a1 += r7.h * r7.h (is); a0 += r6.l * r6.l, a1 += r6.h * r6.h (is); (R7,R6) = BYTEOP16M (R1:0,R3:2) (R)|| R0 = [I0++] || R2 = [I1++]; a0 += r7.l * r7.l, a1 += r7.h * r7.h (is);.E80: a0 += r6.l * r6.l, a1 += r6.h * r6.h (is); a0 += a1; r0 = a0; (r7:6) = [sp++]; unlink; rts;DEFUN_END(sse8)DEFUN(sse16,mL1, (void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)): link 0; [--sp] = (r7:6); p0=[fp+24]; // h i0=r1; // pix1 i1=r2; // pix2 r2=[fp+20]; // line_size r2+=-16; m0=r2; a0=a1=0; DISALGNEXCPT || R0 = [I0++] || R2 =[I1++]; LSETUP(.S160,.E160) LC0=P0;.S160: DISALGNEXCPT || R1 = [I0++] || R3 = [I1++]; (R7,R6) = BYTEOP16M (R1:0,R3:2) || R0 = [I0++] || R2 = [I1++]; a0 += r7.l * r7.l, a1 += r7.h * r7.h (is); a0 += r6.l * r6.l, a1 += r6.h * r6.h (is); (R7,R6) = BYTEOP16M (R1:0,R3:2) (R)|| R1 = [I0++] || R3 = [I1++]; a0 += r7.l * r7.l, a1 += r7.h * r7.h (is); a0 += r6.l * r6.l, a1 += r6.h * r6.h (is); (R7,R6) = BYTEOP16M (R1:0,R3:2) || R0 = [I0++M0] || R2 = [I1++M0]; a0 += r7.l * r7.l, a1 += r7.h * r7.h (is); a0 += r6.l * r6.l, a1 += r6.h * r6.h (is); (R7,R6) = BYTEOP16M (R1:0,R3:2) (R)|| R0 = [I0++] || R2 = [I1++]; a0 += r7.l * r7.l, a1 += r7.h * r7.h (is);.E160: a0 += r6.l * r6.l, a1 += r6.h * r6.h (is); a0 += a1; r0 = a0; (r7:6) = [sp++]; unlink; rts;DEFUN_END(sse16)
⌨️ 快捷键说明
复制代码
Ctrl + C
搜索代码
Ctrl + F
全屏模式
F11
切换主题
Ctrl + Shift + D
显示快捷键
?
增大字号
Ctrl + =
减小字号
Ctrl + -