; ; Measure P6 optimized DAXPY execution times ; From: Micheal Cranford ; ; This program requires TASM (under DOS) ; ; ASM.BAT file is as follows ; ; TASM /c /r /la /ml /m5 /w2 /zi %1 %2 %3 %4 ; ; LNK.BAT file is as follows ; ; TLINK /3 /c /C /d /l /m /s /v %1 %2 %3 %4 %5 %6 ; ; Results (from mcranfor@ichips.intel.com) ; ; Processor Points / Loop Total Cycles Cycles / DP MFLOP MFLOPS @ freq ; ; Pentium 5 826182430 3.2 72 @ 233 MHz ;Pentium Pro 5 395092362 1.5 194 @ 300 MHz .MODEL NEARSTACK SMALL TextLineSize EQU 80 ; number of ASCII characters to print FullSize EQU 25 ; number of full size steps Passes EQU 1000000 ; number of passes to make .STACK 100 .DATA Xarray DQ 128 DUP (1.0) Yarray DQ 128 DUP (2.0) Constant DQ 3.0 Corrections DD 0 ; RDTSC cycle time overhead CycleCounts DD 4 DUP (0) ; enough for four cycle times PackedCount DB 10 DUP (0) ; temporary packed BCD value OutgoingText DB 80 DUP (0) ; output character buffer .CODE .586 LOCALS RDTSC MACRO DB 0FH,31H ENDM ; call ReadTSC for each slot twice for a start time and a finish ; time and the resulting slot will have the corrected cycle time ReadTSC MACRO TimeSlot RDTSC ; get current TSC cycle count SUB EAX,DS:[Corrections] ; subtract overhead cycles SUB EAX,DS:[CycleCounts+TimeSlot*4] ; subtract last MOV DS:[CycleCounts+TimeSlot*4],EAX ; update cycles ENDM ; call InitTSC only once at the very beginning before using ReadTSC InitTSC MACRO XOR EAX,EAX ; reset cycle count MOV DS:[Corrections],EAX MOV DS:[CycleCounts+00],EAX MOV DS:[CycleCounts+04],EAX MOV DS:[CycleCounts+08],EAX MOV DS:[CycleCounts+12],EAX ReadTSC -1 ; get current TSC cycle count RDTSC ; get ReadTSC overhead SUB EAX,DS:[Corrections] ; subtract first MOV DS:[Corrections],EAX ; update cycles ENDM EVEN DAXPYP6: PUSH BP MOV BP,SP MOV AX,@data ; setup cycle count segments MOV DS,AX MOV ES,AX InitTSC ReadTSC 0 ; get starting cycle count FNINIT ; start from known FPU state FLD QWORD PTR DS:[Constant] ; get multiplier constant MOV ECX,Passes ; initialize loop counter OuterLoop: LEA SI,DS:[Xarray] ; get pointer to X array LEA DI,DS:[Yarray] ; get pointer to Y array MOV AX,FullSize ; get full steps counter InnerLoop: ; C PPPPP F7 F6 F5 F4 F3 F2 F1 F0 FLD QWORD PTR DS:[SI+00] ; get X0 0 K . . . . . . . FMUL ST(0),ST(1) ; T0:=K*X0 1 a K X0 . . . . . . FLD QWORD PTR DS:[SI+08] ; get X1 0 K t0 . . . . . . FMUL ST(0),ST(2) ; T1:=K*X1 1 ab K t0 X1 . . . . . FLD QWORD PTR DS:[SI+16] ; get X2 0 K t0 t1 . . . . . FMUL ST(0),ST(3) ; T2:=K*X2 1 abc K t0 t1 X2 . . . . FLD QWORD PTR DS:[SI+24] ; get X3 0 K t0 t1 t2 . . . . FMUL ST(0),ST(4) ; T3:=K*X3 1 abcd K t0 t1 t2 X3 . . . FXCH ST(2) ; T3 <=> T1 0 K t0 t1 t2 t3 . . . FLD QWORD PTR DS:[SI+32] ; get X4 0 K t0 t3 t2 t1 . . . FMUL ST(0),ST(5) ; T4:=K*X4 1 abcde K t0 t3 t2 t1 X4 . . FXCH ST(4) ; T4 <=> T0 0 K T0 t3 t2 t1 t4 . . FLD QWORD PTR DS:[DI+00] ; get Y0 0 K t4 t3 t2 t1 T0 . . FADDP ST(1),ST(0) ; Y0:=T0+Y0 1 fbcde K t4 t3 t2 t1 T0 Y0 . FLD QWORD PTR DS:[DI+08] ; get Y1 0 K t4 t3 t2 T1 y0 . . FADDP ST(2),ST(0) ; Y1:=T1+Y1 1 fgcde K t4 t3 t2 T1 y0 Y1 . FLD QWORD PTR DS:[DI+16] ; get Y2 0 K t4 t3 T2 y1 y0 . . FADDP ST(3),ST(0) ; Y2:=T2+Y2 1 fghde K t4 t3 T2 y1 y0 Y2 . FLD QWORD PTR DS:[DI+24] ; get Y3 0 K t4 T3 y2 y1 y0 . . FADDP ST(4),ST(0) ; Y3:=T3+Y3 1 fghie K t4 T3 y2 y1 y0 Y3 . FLD QWORD PTR DS:[DI+32] ; get Y4 0 K T4 y3 y2 y1 y0 . . FADDP ST(5),ST(0) ; Y4:=T4+Y4 1 fghij K T4 y3 y2 y1 y0 Y4 . FSTP QWORD PTR DS:[DI+00] ; store Y0 1 ghij K y4 y3 y2 y1 Y0 . . FSTP QWORD PTR DS:[DI+08] ; store Y1 1 hij K y4 y3 y2 Y1 . . . FSTP QWORD PTR DS:[DI+16] ; store Y2 1 ij K y4 y3 Y2 . . . . FSTP QWORD PTR DS:[DI+24] ; store Y3 1 j K y4 Y3 . . . . . FSTP QWORD PTR DS:[DI+32] ; store Y4 1 K Y4 . . . . . . ADD SI,40 ; update array X index ADD DI,40 ; update array Y index DEC AX ; update loop counter JNZ SHORT InnerLoop ; if more points FLD QWORD PTR DS:[SI+00] ; get X0 0 K . . . . . . . FMUL ST(0),ST(1) ; T0:=K*X0 1 a K X0 . . . . . . FLD QWORD PTR DS:[SI+08] ; get X1 0 K t0 . . . . . . FMUL ST(0),ST(1) ; T1:=K*X1 1 ab K t0 X1 . . . . . FLD QWORD PTR DS:[SI+16] ; get X2 0 K t0 t1 . . . . . FMUL ST(0),ST(2) ; T2:=K*X2 1 abc K t0 t1 X2 . . . . FXCH ST(2) ; T2 <=> T0 0 K t0 t1 t2 . . . . FLD QWORD PTR DS:[DI+00] ; get Y0 0 K t2 t1 t0 . . . . FADDP ST(1),ST(0) ; Y0:=T0+Y0 1 bcd K t2 t1 t0 Y0 . . . FLD QWORD PTR DS:[DI+08] ; get Y1 0 K t2 t1 y0 . . . . FADDP ST(2),ST(0) ; Y1:=T1+Y1 1 cde K t2 t1 y0 Y1 . . . FLD QWORD PTR DS:[DI+16] ; get Y2 0 K t2 y1 y0 . . . . FADDP ST(3),ST(0) ; Y2:=T2+Y2 1 f de K t2 y1 y0 Y2 . . . FSTP QWORD PTR DS:[DI+00] ; store Y0 f e K y2 y1 y0 . . . . FSTP QWORD PTR DS:[DI+08] ; store Y1 f K y2 y1 . . . . . FSTP QWORD PTR DS:[DI+16] ; store Y2 K y2 . . . . . . DEC ECX ; update pass counter JNZ SHORT OuterLoop ; if not finished FSTP ST(0) ; discard constant K . . . . . . . ReadTSC 0 ; compute total cycle count FinishedTest: LEA ESI,DS:[CycleCounts] ; get cycle counts pointer LEA EDI,DS:[OutgoingText] ; get ASCII buffer pointer MOV BX,4 ; get number of 32 bit words to convert NextFullWord: FILD DWORD PTR DS:[SI] ; get next 32 bit integer cycle count FBSTP TBYTE PTR DS:[PackedCount] ; save next packed 18 digit BCD MOV ECX,9 ; 18 packed BCD digits = 9 total bytes XOR EDX,EDX ; get packed BCD cycle count offset MOV WORD PTR DS:[DI+18],' ' ; spaces between BCD cycle counts NextBCDDigit: MOV AL,DS:[PackedCount+EDX] ; ASCIIfy next nibble pair MOV AH,AL SHR AL,4 AND AH,00FH ADD AX,03030H MOV DS:[EDI+ECX*2-2],AX ; save next pair ASCII digits INC DX ; update packed BCD cycle count offset DEC CX JNZ NextBCDDigit ; if more BCD digits to convert ADD SI,4 ; point to next 32 bit cycle count ADD DI,20 ; point to next ASCII output string DEC BX JNZ NextFullWord ; if more cycle counts to convert MOV CX,TextLineSize ; how many ASCII characters to print LEA DX,DS:[OutgoingText] ; get pointer to ASCII output buffer MOV BX,00001H ; ascii terminal display file handle MOV AH,040H ; print ASCII cycle counts INT 021H POP BP ; cleanup and exit MOV AX,04C00H INT 21H END DAXPYP6