The OpenNET Project / Index page

[ новости /+++ | форум | теги | ]

форумы
правила/FAQ
поиск
регистрация
вход
слежка
RSS


Вариант для распечатки  
Пред. тема | След. тема 
Форум Разговоры, обсуждение новостей
Режим отображения отдельной подветви беседы [ Отслеживать ]

Оглавление

Взломавшие NVIDIA потребовали от компании перевести драйверы в разряд Open Source, opennews (ok), 02-Мрт-22, (0) [смотреть все]

Сообщения [Сортировка по времени | RSS]


308. "Взломавшие NVIDIA потребовали от компании перевести драйверы..."  +1 +/–
Сообщение от john_erohin (?), 05-Мрт-22, 10:15 
матерные комментарии и украинские фамилии в коде есть ?

ps: я скачал, но руки не доходят.

Ответить | Правка | Наверх | Cообщить модератору

310. "Взломавшие NVIDIA потребовали от компании перевести драйверы..."  +/–
Сообщение от Аноним (310), 05-Мрт-22, 12:17 
Фамилии не знаю как искать, мат есть :-D

                // Note: I've tried a manual reduction and not using it because:
                // i ] I see no benefit. (slightly more expensive than the driver optimized atomic)
                // ii] huge chance that OCG will fuck up and reorder the LDS/STS due to
                //     not being able to infer data dependencies between LDS/STS reads/writes.
                //     (see bug 840564)

            //  printf("fuck STS, pq[0]=%f\n",__half2float(*((__half*)(&((*(int4*)(&smem_pool[0])).x)))));
            //  printf("fuck STS, pq[1]=%f\n",__half2float(*((__half*)(&((*(int4*)(&smem_pool[512])).x)))));
            //  printf("fuck STS, pq[2]=%f\n",__half2float(*((__half*)(&((*(int4*)(&smem_pool[1024])).x)))));
            //  printf("fuck STS, pq[3]=%f\n",__half2float(*((__half*)(&((*(int4*)(&smem_pool[1536])).x)))));


        # XXXgth: this is FUCKING BROKEN!!!

    // Send methods setting up the SM version, which DirectAModel on Kepler is
    // keenly interested in.  Hardware, like honey badger, don't give a shit.

Ответить | Правка | Наверх | Cообщить модератору

312. "Взломавшие NVIDIA потребовали от компании перевести драйверы..."  +/–
Сообщение от Аноним (310), 05-Мрт-22, 13:08 
                ** Damn!  We are screwed.  This is a bad spot for an
                ** out of memory error.  It is also bloody unlikely,
                ** because we just freed up some memory.

    // Remove NOP; no longer needed. And I don't want it cluttering up nvInst dumps
    // and making people ask why there are so many damn NOPs.  

    // ugh... got a nice dual-linked list and we've still got to potentially walk the whole damn thing...

    // Damn well better be in a system memory pool if they preallocated memory
    // for us.

                else
                {
                    // oh crap...

        // Only bother with the bits the driver actually uses -- there's LOTS of random irrelevant
        // obsolete crap in this struct.


    // Maxwell: Divergent LDC are still crap, 24+10*#replays (replays are PER unique 32bit value!!!)
    // TODO: revisit on Volta

Мата мало, откровенно говоря. Всё-таки код корпоративный. Это вам не Линукс :-)

Ответить | Правка | К родителю #308 | Наверх | Cообщить модератору

314. "Взломавшие NVIDIA потребовали от компании перевести драйверы..."  +/–
Сообщение от Аноним (310), 05-Мрт-22, 14:50 
Весь список:

            //    1) there was a screw-up in computing an 'empty' slot during initialization
    // and making people ask why there are so many damn NOPs.
    // animations is the road to hell, and becomes something that is not easily maintained in the future.
                    applyShaderModificationsBS0(); // normalize crap that should have been removed by the app
        // approximation will screw things up.
    // a shit about them.
    // bail out.  Hopefully this won't screw up the DisplayPort library too
    // because it would screw up the relocation finding logic.
    // Bloody hack for GM10x A01. Will remove once everyone move to A02. Bug:1351868
    // Blow up if the parameters in the uniform information look like crap.
                    break; // Fix me! Spilling code can screw up the pairing information.  Until this is fixed we must exit here.
        // BUGBUG: precision goes to hell here... but I think
            // but not in the beginning/middle or else it can screw up with this logic.
            cmb->control.in[6], crap);
            // Crap - could not set things up.  Fall back
                            //Crap - fixup again. MS rotate structure does not appear to be correct...
            //crap in the registry not matching any of known types, we asked not to show all types we know,
    // Crap, MS DXGK hard limits total node count under 64.  So we have to trim nodes, especailly under 4/8 way SLI
                ** Damn!  We are screwed.  This is a bad spot for an
        // Damn well better be in a system memory pool if they preallocated memory
    // Damn well better be in a system memory pool if they preallocated memory
    // EMIT/ENDPRIM are boring as all hell!
        // emulator hell...
            // ETW can adjust event names on the fly instead of doing this crap...
    // fbo_modeset1 and fbo_validation_hell fail if this line is removed.
    // fbo_validation_hell tests for this problem.
    // FIXME -- a double to a float without the optimizer screwing it
                        // flush each line of text. screw perf, we are already debugging something using text files.
        GLubyte crap;
    // have to screw with it.
//   hell and the OS generates paging operations (thus causing >500us ctxsw'ing events), too.
    // hell of it.  They get reset to these values after every EMIT
//   hell (uncached read/writes) and the OS generates paging operations
// ideally we should have some meta-state for this but there are just WAY too many dependencies, so screw 'em.
    // if there's no perfect match, try the closet one
// Implementing this function because the Microsoft AMMD64 library seems to screw up on occasion
#include "arse.h"
#include "t30\arse.h"
    // keenly interested in.  Hardware, like honey badger, don't give a shit.
        // lut: has to be after .start(). otherwise we screw up the reserved PB space for commands above.
            // Mapping with 4KB VM pages for vidmem is a galactic screw job.
    // Mapping with 4KB VM pages for vidmem is a galactic screw job.
    // Maxwell: Divergent LDC are still crap, 24+10*#replays (replays are PER unique 32bit value!!!)
    m_nFrameIndex0[vidPnId] = 0;                // In normal case, we would be having AFR->Bridgeless transition. This may screw whole synchronization logic
        // MS treats texkil input reg as a "dest". crap crap!!
    // - Never delete dead arguments, we might screw up linking with our runtime.
// NV0080_CTRL_CMD_DMA_FLUSH, but using that directly will screw with the BC
        // obsolete crap in this struct.
                    // oh crap...
// -- one day I'll burn in hell for this...
            // or don't know how to talk to us, so screw them.
        // or don't know how to talk to us, so screw them.
                ** out of memory error.  It is also bloody unlikely,
        // Outputs are initialized to (0,0,0,1) for the hell of it.
    // Outputs are initialized to (0,0,0,1) for the hell of it.
    // parameters in the uniform information look like crap.
        // Paranoid check for staging bugger existence, WAR for NULL pointer write in case of driver reloading, bug 2629498
    // parseasm doesn't include wgf2um/internals/utils/util.cpp (dependency hell...) and thus
    //pretty damn sure that every time we recreate a Z-Buffer we have to re-clear.  tag bits are pointless
* ReplaceArgWithMove() - Replace the arg with a mov of the arg??? Does a lot of other crap as
    // Reset the results array back to (0,0,0,1) for the hell of it.  The
                    // resource arbiter or some other perfstrat didn't screw up
    //setting CurrentClockFrequency to null so that I2C cock should be reinitialzed in subsequent transaction.
    // shit about them.
            // Slow as hell, should switch to a vector supported memcopy
        // Slow as hell, should switch to a vector supported memcopy
        //   so skipping patch entries won't screw up what the GPU sees like it would under basic.
  // stepping on each other's toes.
                            // the damn nt kernel doesn't support floats - approx with the integer part
        // the next entry this warp will write to will be on the toes of another segment,
                // these actually do exist for TCIndex>8 in a different bank of renderstates, but for now, screw it. cylwrap should die anyhow.
  // This can happen if the user is screwing with us.
    // !!! this is crap.  we need to validate a lot, but need to figure out
            // !!! This would be a good thing for the expert driver to bitch about.
        // those smarts actually screw up the dimensions for fcall's FP_REG operand (at least,
                // to not completely screw over developers with old cubins now
// UGH, because of the dependency hell incurred by nvHash64 (among others) there
// UGH, because of the dependency hell incurred by nvHash64 there are a bunch
                     * UnbindUniform would screw up royally */
        // well, crap... a CPU read lock on a vidmem IB is going to really slow
    /* What is this crap?
*    where the hell x0 is and it found x0 in a live register.  The debugger
        // will screw up everything.
            // will screw up the heuristics that decide the optimal locations for buffer objects.
        // will screw up the heuristics that decide the optimal locations for buffer objects.
    // will screw up the heuristics that decide the optimal locations for buffer objects.

Учитывая, что это гигабайты исходного кода - очень очень мало.

Screw up - очень условный мат.

// b.

Ответить | Правка | К родителю #308 | Наверх | Cообщить модератору

315. "Взломавшие NVIDIA потребовали от компании перевести драйверы..."  –1 +/–
Сообщение от Аноним (310), 05-Мрт-22, 14:56 
Количество fixme в *.c *.cpp: 9101

Драйвера для видео - это по сложности полный песец.

// b.

Ответить | Правка | К родителю #308 | Наверх | Cообщить модератору

Архив | Удалить

Рекомендовать для помещения в FAQ | Индекс форумов | Темы | Пред. тема | След. тема




Партнёры:
PostgresPro
Inferno Solutions
Hosting by Hoster.ru
Хостинг:

Закладки на сайте
Проследить за страницей 		Created 1996-2024 by Maxim Chirkov
Добавить, Поддержать, Вебмастеру