//******************************************************************************* //* * //* functions quant_inter and dequant_inter have been softwarepipelined * //* use was made of the pmpyshr2 instruction * //* * //* by Christian Engel and Hans-Joachim Daniels * //* christian.engel@ira.uka.de hans-joachim.daniels@ira.uka.de * //* * //* This was made for the ia64 DivX laboratory (yes, it was really called * //* this way, originally OpenDivX was intendet, but died shortly before our * //* work started (you will probably already know ...)) * //* at the Universitat Karlsruhe (TH) held between April and July 2002 * //* http://www.info.uni-karlsruhe.de/~rubino/ia64p/ * //* * //******************************************************************************* .file "quant_h263_ia64.s" .pred.safe_across_calls p1-p5,p16-p63 .section .rodata .align 4 .type multipliers#,@object .size multipliers#,128 multipliers: data4 0 data4 32769 data4 16385 data4 10923 data4 8193 data4 6554 data4 5462 data4 4682 data4 4097 data4 3641 data4 3277 data4 2979 data4 2731 data4 2521 data4 2341 data4 2185 data4 2049 data4 1928 data4 1821 data4 1725 data4 1639 data4 1561 data4 1490 data4 1425 data4 1366 data4 1311 data4 1261 data4 1214 data4 1171 data4 1130 data4 1093 data4 1058 .global __divdi3# .text .align 16 .global quant_intra_ia64# .proc quant_intra_ia64# quant_intra_ia64: .prologue .save ar.pfs, r38 alloc r38 = ar.pfs, 4, 3, 2, 0 adds r16 = -8, r12 .fframe 32 adds r12 = -32, r12 mov r17 = ar.lc addl r14 = @ltoff(multipliers#), gp ld2 r15 = [r33] ;; .savesp ar.lc, 24 st8 [r16] = r17, 8 ld8 r14 = [r14] sxt2 r15 = r15 ;; .save.f 0x1 stf.spill [r16] = f2 .save rp, r37 mov r37 = b0 .body dep.z r36 = r34, 1, 15 dep.z r16 = r34, 2, 32 cmp4.ge p6, p7 = 0, r15 ;; add r16 = r16, r14 ;; ld4 r16 = [r16] ;; setf.sig f2 = r16 (p6) br.cond.dptk .L8 extr r39 = r35, 1, 31 sxt4 r40 = r35 ;; add r39 = r39, r15 br .L21 ;; .L8: extr r39 = r35, 1, 31 sxt4 r40 = r35 ;; sub r39 = r15, r39 ;; .L21: sxt4 r39 = r39 br.call.sptk.many b0 = __divdi3# ;; addl r14 = 62, r0 st2 [r32] = r8 addl r19 = 1, r0 ;; mov ar.lc = r14 ;; .L20: dep.z r17 = r19, 1, 32 ;; add r15 = r17, r33 adds r19 = 1, r19 ;; ld2 r14 = [r15] ;; sxt2 r14 = r14 ;; mov r16 = r14 mov r18 = r14 ;; sub r15 = r0, r16 cmp4.le p8, p9 = r36, r16 cmp4.le p6, p7 = r0, r16 ;; sxt2 r14 = r15 (p6) br.cond.dptk .L14 ;; mov r16 = r14 add r18 = r17, r32 ;; setf.sig f6 = r16 cmp4.le p6, p7 = r36, r16 mov r15 = r18 ;; xma.l f6 = f6, f2, f0 (p7) st2 [r18] = r0 ;; getf.sig r14 = f6 ;; extr r14 = r14, 16, 16 ;; sub r14 = r0, r14 ;; (p6) st2 [r15] = r14 br .L12 .L14: .pred.rel "mutex", p8, p9 setf.sig f6 = r18 add r16 = r17, r32 ;; xma.l f6 = f6, f2, f0 mov r15 = r16 (p9) st2 [r16] = r0 ;; getf.sig r14 = f6 ;; extr r14 = r14, 16, 16 ;; (p8) st2 [r15] = r14 .L12: br.cloop.sptk.few .L20 adds r18 = 24, r12 ;; ld8 r19 = [r18], 8 mov ar.pfs = r38 mov b0 = r37 ;; mov ar.lc = r19 ldf.fill f2 = [r18] .restore sp adds r12 = 32, r12 br.ret.sptk.many b0 .endp quant_intra_ia64# .common quant_intra#,8,8 .common dequant_intra#,8,8 .align 16 .global dequant_intra_ia64# .proc dequant_intra_ia64# dequant_intra_ia64: .prologue ld2 r14 = [r33] andcm r15 = 1, r34 setf.sig f8 = r35 ;; sxt2 r14 = r14 sub r15 = r34, r15 addl r16 = -2048, r0 ;; setf.sig f6 = r14 setf.sig f7 = r15 shladd r34 = r34, 1, r0 ;; xma.l f8 = f6, f8, f0 .save ar.lc, r2 mov r2 = ar.lc ;; .body getf.sig r14 = f8 setf.sig f6 = r34 ;; sxt2 r15 = r14 st2 [r32] = r14 ;; cmp4.le p6, p7 = r16, r15 ;; (p7) st2 [r32] = r16 (p7) br.cond.dptk .L32 addl r14 = 2047, r0 ;; cmp4.ge p6, p7 = r14, r15 ;; (p7) st2 [r32] = r14 .L32: addl r14 = 62, r0 addl r19 = 1, r0 addl r22 = 2048, r0 addl r21 = -2048, r0 addl r20 = 2047, r0 ;; mov ar.lc = r14 ;; .L56: dep.z r16 = r19, 1, 32 ;; add r14 = r16, r33 add r17 = r16, r32 adds r19 = 1, r19 ;; ld2 r15 = [r14] ;; sxt2 r15 = r15 ;; cmp4.ne p6, p7 = 0, r15 cmp4.le p8, p9 = r0, r15 ;; (p7) st2 [r17] = r0 (p7) br.cond.dpnt .L36 add r18 = r16, r32 sub r17 = r0, r15 ;; mov r14 = r18 (p8) br.cond.dptk .L40 setf.sig f8 = r17 ;; xma.l f8 = f6, f8, f7 ;; getf.sig r15 = f8 ;; cmp4.lt p6, p7 = r22, r15 sub r16 = r0, r15 ;; (p7) st2 [r14] = r16 (p6) st2 [r14] = r21 br .L36 .L40: setf.sig f8 = r15 ;; xma.l f8 = f6, f8, f7 ;; getf.sig r15 = f8 ;; cmp4.le p6, p7 = r20, r15 ;; (p6) mov r14 = r20 (p7) mov r14 = r15 ;; st2 [r18] = r14 .L36: br.cloop.sptk.few .L56 ;; mov ar.lc = r2 br.ret.sptk.many b0 .endp dequant_intra_ia64# //uint32_t quant_inter_ia64(int16_t *coeff, const int16_t *data, const uint32_t quant) .common quant_inter#,8,8 .align 16 .global quant_inter_ia64# .proc quant_inter_ia64# quant_inter_ia64: //******************************************************* //* * //* const uint32_t mult = multipliers[quant]; * //* const uint16_t quant_m_2 = quant //* const uint16_t quant_d_2 = quant >> 1; * //* int sum = 0; * //* uint32_t i; * //* int16_t acLevel,acL; * //* * //*******************************************************/ LL=3 // LL = load latency //if LL is changed, you'll also have to change the .pred.rel... parts below! .prologue addl r14 = @ltoff(multipliers#), gp dep.z r15 = r34, 2, 32 .save ar.lc, r2 mov r2 = ar.lc ;; .body alloc r9=ar.pfs,0,24,0,24 mov r17 = ar.ec mov r10 = pr ld8 r14 = [r14] extr.u r16 = r34, 1, 16 //r16 = quant_d_2 dep.z r20 = r34, 1, 15 //r20 = quant_m_2 ;; add r15 = r15, r14 mov r21 = r16 //r21 = quant_d_2 mov r8 = r0 //r8 = sum = 0 mov pr.rot = 0 //p16-p63 = 0 ;; ld4 r15 = [r15] addl r14 = 63, r0 mov pr.rot = 1 ;; mov ar.lc = r14 mov ar.ec = LL+9 mov r29 = r15 ;; mov r15 = r33 //r15 = data mov r18 = r32 //r18 = coeff ;; .rotr ac1[LL+3], ac2[8], ac3[2] .rotp p[LL+9], cmp1[8], cmp1neg[8],cmp2[5], cmp2neg[2] //******************************************************************************* //* * //* for (i = 0; i < 64; i++) { * //* acL=acLevel = data[i]; * //* acLevel = ((acLevel < 0)?-acLevel:acLevel) - quant_d_2; * //* if (acLevel < quant_m_2){ * //* acLevel = 0; * //* } * //* acLevel = (acLevel * mult) >> SCALEBITS; * //* sum += acLevel; * //* coeff[i] = ((acL < 0)?-acLevel:acLevel); * //* } * //* * //*******************************************************************************/ .explicit .L58: .pred.rel "clear", p29, p37 .pred.rel "mutex", p29, p37 //pipeline stage {.mmi (p[0]) ld2 ac1[0] = [r15],2 // 0 acL=acLevel = data[i]; (p[LL+1]) sub ac2[0] = r0, ac1[LL+1] // LL+1 ac2=-acLevel (p[LL]) sxt2 ac1[LL] = ac1[LL] // LL } {.mmi (p[LL+1]) cmp4.le cmp1[0], cmp1neg[0] = r0, ac1[LL+1] // LL+1 cmp1 = (0 (p[LL+4]) cmp4.le cmp2[0], cmp2neg[0] = r20, ac2[3] // LL+4 cmp2 = (quant_m_2 < acLevel) ; cmp2neg = !(quant_m_2 < acLevel) (cmp1[1]) sub ac2[1] = ac1[LL+2], r21 // LL+2 acLevel = acLevel - quant_d_2; } {.mmi (cmp2neg[1]) mov ac2[4] = r0 // LL+5 if (acLevel < quant_m_2) acLevel=0; (cmp1neg[1]) sub ac2[1] = ac2[1], r21 // LL+2 acLevel = ac2 - quant_d_2; (p[LL+3]) sxt2 ac2[2] = ac2[2] // LL+3 } {.mmi .pred.rel "mutex", p34, p42 (cmp1[6]) mov ac3[0] = ac2[6] // LL+7 ac3 = acLevel; (cmp1neg[6]) sub ac3[0] = r0, ac2[6] // LL+7 ac3 = -acLevel; (p[LL+6]) pmpyshr2.u ac2[5] = r29, ac2[5], 16 // LL+6 acLevel = (acLevel * mult) >> SCALEBITS; } {.mib (p[LL+8]) st2 [r18] = ac3[1] , 2 // LL+8 coeff[i] = ac3; (cmp2[4]) add r8 = r8, ac2[7] // LL+8 sum += acLevel; br.ctop.sptk.few .L58 ;; } .pred.rel "clear", p29, p37 .default mov ar.ec = r17 ;; mov ar.lc = r2 mov pr = r10, -1 mov ar.pfs = r9 br.ret.sptk.many b0 .endp quant_inter_ia64# // void dequant_inter_ia64(int16_t *data, const int16_t *coeff, const uint32_t quant) .common dequant_inter#,8,8 .align 16 .global dequant_inter_ia64# .proc dequant_inter_ia64# dequant_inter_ia64: //*********************************************************************** //* * //* const uint16_t quant_m_2 = quant //* const uint16_t quant_add = (quant & 1 ? quant : quant - 1); * //* uint32_t i; * //* * //*********************************************************************** .prologue andcm r14 = 1, r34 dep.z r29 = r34, 1, 15 alloc r9=ar.pfs,0,32,0,32 .save ar.lc, r2 mov r2 = ar.lc ;; .body sub r15 = r34, r14 // r15 = quant addl r14 = 63, r0 addl r21 = -2048, r0 addl r20 = 2047, r0 mov r16 = ar.ec mov r17 = pr ;; zxt2 r15 = r15 mov ar.lc = r14 mov pr.rot = 0 ;; adds r14 = 0, r33 // r14 = coeff mov r18 = r32 // r18 = data mov ar.ec = LL+10 mov pr.rot = 1 ;; //******************************************************************************* //* * //*for (i = 0; i < 64; i++) { * //* int16_t acLevel = coeff[i]; * //* * //* if (acLevel == 0) * //* { * //* data[i] = 0; * //* } * //* else if (acLevel < 0) * //* { * //* acLevel = acLevel * quant_m_2 - quant_add; * //* data[i] = (acLevel >= -2048 ? acLevel : -2048); * //* } * //* else // if (acLevel > 0) * //* { * //* acLevel = acLevel * quant_m_2 + quant_add; * //* data[i] = (acLevel //* } * //* } * //* * //*******************************************************************************/ LL=2 // LL := load latency //if LL is changed, you'll also have to change the .pred.rel... parts below! .rotr ac1[LL+10], x[5], y1[3], y2[3] .rotp p[LL+10] , cmp1neg[8], cmp2[5], cmp2neg[5],cmp3[2], cmp3neg[2] .explicit //pipeline stage .L60: .pred.rel "clear", p36 .pred.rel "mutex", p47, p49 .pred.rel "mutex", p46, p48 .pred.rel "mutex", p40, p45 .pred.rel "mutex", p39, p44 .pred.rel "mutex", p38, p43 .pred.rel "mutex", p37, p42 .pred.rel "mutex", p36, p41 {.mmi (p[0])ld2 ac1[0] = [r14] ,2 // 0 acLevel = coeff[i]; (p[LL+1])cmp4.ne p6, cmp1neg[0] = 0, ac1[LL+1] // LL+1 (p[LL])sxt2 ac1[LL] = ac1[LL] // LL } {.mmi (p[LL+1])cmp4.le cmp2[0], cmp2neg[0] = r0, ac1[LL+1] // LL+1 (cmp2[1]) mov x[0] = r20 // LL+2 (p[LL+2])pmpyshr2.u ac1[LL+2] = r29, ac1[LL+2], 0 // LL+2 } {.mmi (cmp2neg[1]) mov x[0] = r21 // LL+2 (cmp2[2]) add ac1[LL+3] = ac1[LL+3], r15 // LL+3 (cmp2neg[2]) sub ac1[LL+3] = ac1[LL+3], r15 // LL+3 } {.mmi (cmp2neg[4]) mov y1[0] = ac1[LL+5] // LL+5 (cmp2neg[4]) mov y2[0] = x[3] // LL+5 (p[LL+4])sxt2 ac1[LL+4] = ac1[LL+4] // LL+4 } {.mmi (cmp2[4]) mov y1[0] = x[3] // LL+5 (cmp2[4]) mov y2[0] = ac1[LL+5] // LL+5 (p[LL+6])cmp4.le cmp3[0], cmp3neg[0] = x[4], ac1[LL+6] // LL+6 } {.mmi (cmp3[1]) mov ac1[LL+7] = y1[2] // LL+7 (cmp3neg[1]) mov ac1[LL+7] = y2[2] // LL+7 (cmp1neg[7]) mov ac1[LL+8] = r0 // LL+8 } {.mbb (p[LL+9])st2 [r18] = ac1[LL+9] ,2 // LL+9 nop.b 0x0 br.ctop.sptk.few .L60 ;; } .pred.rel "clear", p36 .default mov ar.lc = r2 mov ar.pfs = r9 mov ar.ec = r16 mov pr = r17, -1 ;; mov ar.lc = r2 br.ret.sptk.many b0 .endp dequant_inter_ia64# .ident "GCC: (GNU) 2.96 20000731 (Red Hat Linux 7.1 2.96-85)"
