Nouveau - Apr 2016 - [PATCH] nouveau: codegen: Take src swizzle into account on loads

Hi,

On 08-04-16 17:02, Ilia Mirkin wrote:
> On Fri, Apr 8, 2016 at 5:27 AM, Hans de Goede <hdegoede at
redhat.com> wrote:
>> Hi,
>>
>> On 07-04-16 15:58, Ilia Mirkin wrote:
>>>
>>> That's wrong.
>>
>>
>> It used to work with the old RES[] code and if one cannot specify
>> a source swizzle, then how can I do something like
>>
>> LOAD TEMP[0].y, MEMORY[0], address
>>
>> And get the data at absolute global memory address "address"
into TEMP[0].y
>> ?
>>
>> This is a must-have for llvm to be able to generate working TGSI code,
>> I do not see any way around this.
>>
>> AFAIK this is exactly what src-swizzling is for. Also note that
>> this commit does not change anything if no src-swizzling is specified,
>> in that case things work exactly as before.
>>
>>> The spec for the instruction needs to be clarified...
>>>
>>> The current nouveau impl is correct - only the .x of the address
>>> should be loaded, with up to 16 bytes read into the destination.
>>
>>
>> Ah note this is not about swizzling on the address, that indeed
>> makes no sense given how the addressing works for BUFFERS / MEMORY,
>> no this is about adding a swizlling postfix to the buffer / memory
>> resource specification, for example:
>>
>> LOAD TEMP[0].y, MEMORY[0].xxxx, TEMP[0]
>>
>> See the swizzling is done on the resource, not on the address, so
>> the swizzling specifies swizzling of the up to 16 bytes read from
>> address, it does not influence the address handling at all.
>>
>> I now see I made an error in my commit msg, it gives the following
>> example:
>>
>> LOAD TEMP[0].y, MEMORY[0].xxxx, TEMP[0].x
>>
>> This clearly is wrong, the last TEMP[0].x is not even valid TGSI,
>> the correct example would be:
>>
>> LOAD TEMP[0].y, MEMORY[0].xxxx, TEMP[0]
>
> I stand by my comment of "working as intended". But that
doesn't mean
> the intent can't be changed :)
>
> For memory/buffers, LOAD takes the address at TEMP[0].x and loads 16
> bytes (4 words), and sticks them into the destination's .xyzw. If you
> happen to have a writemask, then only some of those are written out.
>
> It seems that you're trying to add additional meaning to the swizzle
> on the "memory" argument. However I don't believe that such a
thing is
> defined. (And definitely not used anywhere, at least not on purpose.)
>
> Why does this cause you issues with LLVM-generated TGSI?
When dealing with non vector variables the llvm register allocator
will use TEMP[0].x then TEMP[0].y, etc.

When loading something from a global buffer it will calculate the
address to use, and store that in say TEMP[0].x, so it ends up
generating:

LOAD TEMP[0].y, MEMORY[0], TEMP[0]

Expecting the contents of TEMP[0].y to become the 32 bits of data
to which TEMP[0].x is pointing. But instead it will get the 32 bits of
data at address (TEMP[0].x + 4).

With the old RES[32767] code one could generate the following TGSI:

LOAD TEMP[0].y, RES[32767].xxxx, TEMP[0]

And things would work fine since the .xxxx swizzling postfix would
be honored and when storing to y (the only component set in the dest-mask)
the x component at address (TEMP[0].x) would be loaded, rather then the
y component at (TEMP[0].y)

Note that another approach would be to not increment the address by
a 32 bit word for skipped (not set in destmask) components.

The way I see it either:

1) We see that LOAD does not deal with vectors, but with flat memory,
in which case skipping 4 bytes because x is not set in the destmask
does not make sense, as that is a vector thing todo.

2) LOAD is vector layout aware in which case supporting swizzling
makes sense.

Currently we have a weird hybrid which is rather cumbersome to
work with from a compiler pov.

Regards,

Hans












>
>    -ilia
>

On Fri, Apr 8, 2016 at 11:28 AM, Hans de Goede <hdegoede at redhat.com>
wrote:
> When dealing with non vector variables the llvm register allocator
> will use TEMP[0].x then TEMP[0].y, etc.
>
> When loading something from a global buffer it will calculate the
> address to use, and store that in say TEMP[0].x, so it ends up
> generating:
>
> LOAD TEMP[0].y, MEMORY[0], TEMP[0]
>
> Expecting the contents of TEMP[0].y to become the 32 bits of data
> to which TEMP[0].x is pointing. But instead it will get the 32 bits of
> data at address (TEMP[0].x + 4).
>
> With the old RES[32767] code one could generate the following TGSI:
>
> LOAD TEMP[0].y, RES[32767].xxxx, TEMP[0]
>
> And things would work fine since the .xxxx swizzling postfix would
> be honored and when storing to y (the only component set in the dest-mask)
> the x component at address (TEMP[0].x) would be loaded, rather then the
> y component at (TEMP[0].y)
>
> Note that another approach would be to not increment the address by
> a 32 bit word for skipped (not set in destmask) components.
>
> The way I see it either:
>
> 1) We see that LOAD does not deal with vectors, but with flat memory,
> in which case skipping 4 bytes because x is not set in the destmask
> does not make sense, as that is a vector thing todo.
>
> 2) LOAD is vector layout aware in which case supporting swizzling
> makes sense.
>
> Currently we have a weird hybrid which is rather cumbersome to
> work with from a compiler pov.
And I guess LLVM never ends up generating any of the other "funny"
instructions like LIT and the such. Well, I have no problem adding the
swizzling logic, i.e. the way that LOAD will now work (logically) is
that it will

(a) fetch 4 values from the coordinates provided (4 sequential dwords
from src1.x in the case of buffer/memory, RGBA colors from src1.xyz in
the case of images)
(b) swizzle them according to the swizzle on the MEMORY/BUFFER/IMAGE argument
(c) store that swizzled result into the destination based on the writemask

That would sound reasonable to me, and if I understand correctly, is
option 2 of your proposal. We'd need some docs updates and buy-in from
the other gallium driver developers.

STORE remains unchanged, as the MEMORY/etc is in the destination,
where there is a writemask, which is presently used and will remain
effective.

  -ilia

Hi,

On 08-04-16 17:45, Ilia Mirkin wrote:
> On Fri, Apr 8, 2016 at 11:28 AM, Hans de Goede <hdegoede at
redhat.com> wrote:
>> When dealing with non vector variables the llvm register allocator
>> will use TEMP[0].x then TEMP[0].y, etc.
>>
>> When loading something from a global buffer it will calculate the
>> address to use, and store that in say TEMP[0].x, so it ends up
>> generating:
>>
>> LOAD TEMP[0].y, MEMORY[0], TEMP[0]
>>
>> Expecting the contents of TEMP[0].y to become the 32 bits of data
>> to which TEMP[0].x is pointing. But instead it will get the 32 bits of
>> data at address (TEMP[0].x + 4).
>>
>> With the old RES[32767] code one could generate the following TGSI:
>>
>> LOAD TEMP[0].y, RES[32767].xxxx, TEMP[0]
>>
>> And things would work fine since the .xxxx swizzling postfix would
>> be honored and when storing to y (the only component set in the
dest-mask)
>> the x component at address (TEMP[0].x) would be loaded, rather then the
>> y component at (TEMP[0].y)
>>
>> Note that another approach would be to not increment the address by
>> a 32 bit word for skipped (not set in destmask) components.
>>
>> The way I see it either:
>>
>> 1) We see that LOAD does not deal with vectors, but with flat memory,
>> in which case skipping 4 bytes because x is not set in the destmask
>> does not make sense, as that is a vector thing todo.
>>
>> 2) LOAD is vector layout aware in which case supporting swizzling
>> makes sense.
>>
>> Currently we have a weird hybrid which is rather cumbersome to
>> work with from a compiler pov.
>
> And I guess LLVM never ends up generating any of the other
"funny"
> instructions like LIT and the such. Well, I have no problem adding the
> swizzling logic, i.e. the way that LOAD will now work (logically) is
> that it will
>
> (a) fetch 4 values from the coordinates provided (4 sequential dwords
> from src1.x in the case of buffer/memory, RGBA colors from src1.xyz in
> the case of images)
> (b) swizzle them according to the swizzle on the MEMORY/BUFFER/IMAGE
argument
> (c) store that swizzled result into the destination based on the writemask
>
> That would sound reasonable to me, and if I understand correctly, is
> option 2 of your proposal.
Yes that is option 2, and is basically what the patch which started this
thread does. So that would work for me :)
> We'd need some docs updates and buy-in from the other gallium driver
developers.
What docs would need updating ? The TGSI docs I'm aware of are at:

http://gallium.readthedocs.org/en/latest/tgsi.html

I assume those have a source in the mesa src somewhere (I've not looked),
but those mostly just look quite incomplete in general when it comes to TGSI
(I've had to revert to figuring what things do from the mesa srcs quite
often)

Have I been looking at the wrong docs perhaps ?

Note that them being incomplete is not intended as an excuse to not document
this, I'm all for better documentation.
> STORE remains unchanged, as the MEMORY/etc is in the destination,
> where there is a writemask, which is presently used and will remain
> effective.
Right and note that the first src operand of STORE already takes swizzling
into account, so the proposed option 2 will actually make the 2 more inline.

Regards,

Hans