[Mesa-dev] Question regarding cs_gem_write_reloc in radeon_cs_gem.c

[Miguel Ramos]

2012/6/15 Marek Olšák <maraeo at gmail.com>:
> On Fri, Jun 15, 2012 at 3:15 PM, Miguel Lopes Santos Ramos
> <org.mesa3d at miguel.ramos.name> wrote:
>> Hi all,
>>
>> I'm trying to learn to program radeons in assembly in Linux and
>> already got help once in this mailing list.
>>
>> I'm programming in C++ using the kernel DRM ioctls directly because
>> libdrm appears to be a thin wrapper around these (for my purpose,
>> which is using the GPU for computing), but in doing so, I need to
>> understand what libdrm is doing.
>>
>> Now, the cs_gem_write_reloc function, emits two dwords which encode a
>> relocation and adds information on that relocation to an array which
>> is passed to the kernel.
>> The function goes to some length to make sure a single
>> drm_radeon_cs_reloc structure is passed to the kernel for each buffer
>> object.
>>
>> Now, I'm trying to understand what are the constraints, I don't know
>> what is the desired behaviour.
>>
>> This function begins with:
>>
>>  if ((read_domain && write_domain) || (!read_domain && !write_domain)) {
>>        /* in one CS a bo can only be in read or write domain but not
>>         * in read & write domain at the same sime
>>         */
>>
>> Now, the comment seems to say one thing whereas the if-condition seems
>> to say the opposite. Am I mistaken? Which is the desired behaviour,
>> that of the comment or that of the code?
>>
>> Then, the code block within if (reloc->handle == bo->handle) { ... }
>> is even more confusing to me.
>> For instance, the following line:
>>
>> reloc->flags |= (flags & reloc->flags);
>>
>> This seems to be a no-op. Anyway, I'm not so worried about these flags
>> which appear never to be used...
>> But what are really the constraints for this read_domains and write_domain?
>
> I no longer remember how exactly the libdrm CS code works. Radeon
> gallium drivers don't use it, because it's inefficient and awful.
>
> To answer your question: the kernel uses write_domain if it's
> non-zero, otherwise it uses read_domains. That's it.

Thanks, that is clear.

> The Mesa code that replaces libdrm resides in the Mesa repository in:
> src/gallium/winsys/radeon/drm/

Hum, yes, it looks better, good formating and I see you are also using
a hash table.
I think this is all I need from DRM, now I only need to dig into the registers.

BTW, not that anyone cares, it's not useful, but since I'm reading so
much code from the people related to this list, for the sake of
attribution, I'm working here:
https://github.com/migle/trials/tree/master/radeon

-- 
Miguel Ramos <mail at miguel.ramos.name>
PGP A006A14C