пятница, 14 марта 2025 г.

nvidia sass disassembler, part 3

It looks like this rabbit hole goes much deeper

Some const banks does not have ConstBankAddressX:
CX:Sb[UniformRegister:URb][UImm(16)*:Sb_offset]
BITS_6_37_32_Ra_URb=URb
BITS_14_53_40_Sb_offset=Sb_offset SCALE 4 
 
Btw there is no encoding for field Sb

Next they have desc memory:
DESC:memoryDescriptor[UniformRegister:Ra_URb][Register:Ra /ONLY64:input_reg_sz_64_dist + SImm(24/0)*:Ra_offset]

genuine nvdisasm shows them like LDG.E.U16.CONSTANT R10, desc[UR8][R2.64], my disasm as LDG.E.U16.CONSTANT ,R10,desc[UR8][R2.64 + 0x0]

And finally we also have:

A:srcAttr[ UniformRegister:URa + SImm(11/0)*:URa_offset ]

GMMA:gdesc[ UniformRegister:URb ]

TMA:desc[ UniformRegister:URe ]

TTU:ttuAddr[ UImm(16)*:ImmU16 ]

RF:indexURb[UniformRegister:URb] ','UImm(4/0xf)*:PixMaskU04 

TMEMA:tmemA[ UniformRegister:URa ] 

TMEM perhaps means "tensor memory" and I have no idea about the rest of the prefixes

Комментариев нет:

Отправить комментарий