19:56 <mardinauskas> I dunno why I described things another way around. In full length of the pipeline there are two modes, round-robin and greedy-than-oldest. In the short pipeline still it works the other way around, if one schedules tex and alus back to back, it will stay on the same wfid until there is one that does not issue. But the colesky is very good example, this is where i was not wrong.

20:12 <mardinauskas> algorithms-by-block version of cholesky, is really nice way to express what chip does using maths, bit it will need a jik version of cholesky, some drift away modified-ordering in the matrix to manage also branching via async instructions inside the queues.

20:12 <mardinauskas> *blocks

20:20 <mardinauskas> when you did not understand than cholesky is a method of lowering symmetric matrixes to submatrices with triangular ordering applied on the indices, but arbitrary operations can be used. Matrices was something i understood in maths at uni.

20:24 <mardinauskas> derivatives on the other hand was bit of freak show that time for me in the uni, today i am capable of understanding those as well, as well as any math

20:30 <mardinauskas> VLIW has more slots when register file is going larger, because the scoreboard is expensive in hw, hence they can allow doing longer queues and more alus in hw. with 32wavefronts hence they have 32*16*5 slots 2560 when GCN has 1600, 40*40

20:31 <mardinauskas> and there is a certain difference, when VLIW5 executes 5instructions from the same wavefront, then GCN takes 5instructions from different wavefronts

20:33 <mardinauskas> hence what i was saying, all the way i consider VLIW5 and VLIW4 to be superior hw

20:47 <mardinauskas> the difference is quit big as you see, 2560/4=640 for four different classes of instructions 1600/4=400 for each

20:48 <mardinauskas> hence the guy who murdered another one about graphics cards, which is insane but, he was technology wise on the correct side of the debate

20:49 <mardinauskas> at that time amd cards were VLIW and actually they are stronger then NVIDIAs true SIMD cards

20:55 <mardinauskas> and about this conspiring i was very madly treated for 20years in row by group of entire outsider nutters, for me they diserve to be manhandled

20:59 <mardinauskas> this was very muddily and intentionally coordinated terror towards, me which left me with highly big issues, almost death, and lots allready died to, i am not about to forget this

20:59 <mardinauskas> to/too

21:02 <mardinauskas> i acceot apolease only when there would had been an accident, but such row of terror can't be one, it will follow the pattern fellony and a punishment

21:10 <mardinauskas> so this is it by me, i am entering the driver coding quite soon, and now i will be away, when i finish that, i am going hand against hand with some of the fellons, i can not handle all of them, but the most important scamming terrorists are tried to be punished.

21:10 <mardinauskas> bye