Skip to contentSkip to conversation

Trying to understand the URB code

DISCUSSION Kenneth Graunke 3 participants 3 messages Apr 7, 2011 – Apr 8, 2011

lore
Share & tools
lore
#0 lore raw

Hi Nanhai,

I'm trying to understand how the Gen6 URB setup works, and I had some questions...

   if (IS_GT1(intel->intelScreen->deviceID)) {
        urb_size = 32 * 1024;
        max_urb_entry = 128;
   } else {
        urb_size = 64 * 1024;
        max_urb_entry = 256;
   }

I see in vol5c.5 that GT1 has 32kB of URB space and GT2 has 64kB, so urb_size must be the total size of the URB. But what is max_urb_entry?
Where do 128 and 256 come from?

   brw->urb.vs_size = MAX2(brw->vs.prog_data->urb_entry_size, 1);

It looks like brw->vs.prog_data->urb_entry_size is the size of a single VUE, which depends on the number of input/outputs in the particular vertex shader being used.

So, brw->urb.vs_size is also the size of a VUE, but at least 1.
What are the units here? The number of 1024-bit blocks? (I'm looking at 3DSTATE_URB in vol2a of the bspec...)

   brw->urb.nr_vs_entries = max_urb_entry;
   brw->urb.nr_gs_entries = max_urb_entry;

   if (2 * brw->urb.vs_size * brw->urb.nr_vs_entries > urb_size)
           brw->urb.nr_vs_entries = brw->urb.nr_gs_entries =
                (urb_size ) / (2 * brw->urb.vs_size);

Here it looks like you're trying to allocate half of the URB to the VS, and half to the GS. I'm confused by the units, though: if vs_size is in 1024-bit (128-byte) blocks and urb_size is in bytes, don't we need to multiply vs_size by 128?

I think the above code could be simplified to:

int urb_entries = urb_size / (2 * brw->urb.vs_size * 128); brw->urb.nr_vs_entries = brw->urb.nr_gs_entries = MIN2(urb_entries, max_urb_entry);

What do you think?

Thanks for the help.
--Kenneth

#1 lore raw

Hi Ken,
URB allocation on gen6 is different than previous gens.
On previous gen, there is a total size urb for many stages of VS GS CLIP SF. So driver has to decide how much urb to allocate for each stage.

On gen6 only GS and VS will share an urb of 64k(32k on GT1)

urb_entry_number also has a limit, check 3DSTATE_URB command.

The more urb entry number we have, the more current threads can spawn.

But
We should make sure
VS urb size = single_urb_size * gs_urb_entry
+
GS urb size = single_urb_size * vs_urb_entry
< total_urb_size

In most case, this is ok, unless we have a huge single_urb_size.

Since we are always using GS pass through mode for now.
The number for GS urb dose not matter.
So allocate more urb size for VS, less for GS may be better for huge urb_size

But consider we will support GS shader later and huge urb_size case is very rare, I separate urb equally for GS and VS.

Thanks
Zou Nan hai

Quoted 48 lines
Quoted 48 lines

-----Original Message-----
From: Kenneth Graunke [mailto:[email protected]]
Sent: 2011年4月7日 13:01
To: Zou, Nanhai
Cc: [email protected]
Subject: Trying to understand the URB code

Hi Nanhai,

I'm trying to understand how the Gen6 URB setup works, and I had some questions...

   if (IS_GT1(intel->intelScreen->deviceID)) {
        urb_size = 32 * 1024;
        max_urb_entry = 128;
   } else {
        urb_size = 64 * 1024;
        max_urb_entry = 256;
   }

I see in vol5c.5 that GT1 has 32kB of URB space and GT2 has 64kB, so urb_size must be the total size of the URB. But what is max_urb_entry?
Where do 128 and 256 come from?

   brw->urb.vs_size = MAX2(brw->vs.prog_data->urb_entry_size, 1);

It looks like brw->vs.prog_data->urb_entry_size is the size of a single VUE, which depends on the number of input/outputs in the particular vertex shader being used.

So, brw->urb.vs_size is also the size of a VUE, but at least 1. What are the units here? The number of 1024-bit blocks? (I'm looking at 3DSTATE_URB in vol2a of the bspec...)

   brw->urb.nr_vs_entries = max_urb_entry;
   brw->urb.nr_gs_entries = max_urb_entry;

   if (2 * brw->urb.vs_size * brw->urb.nr_vs_entries > urb_size)
           brw->urb.nr_vs_entries = brw->urb.nr_gs_entries =
                (urb_size ) / (2 * brw->urb.vs_size);

Here it looks like you're trying to allocate half of the URB to the VS, and half to the GS. I'm confused by the units, though: if vs_size is in 1024-bit (128-byte) blocks and urb_size is in bytes, don't we need to multiply vs_size by 128?

I think the above code could be simplified to:

Quoted 8 lines
Quoted 8 lines

int urb_entries = urb_size / (2 * brw->urb.vs_size * 128); brw->urb.nr_vs_entries = brw->urb.nr_gs_entries = MIN2(urb_entries, max_urb_entry);

What do you think?

Thanks for the help.
--Kenneth

_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx

#2 lore raw

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

Quoted 8 lines · Nanhai

Hi Ken,
URB allocation on gen6 is different than previous gens.
On previous gen, there is a total size urb for many stages of VS GS CLIP SF. So driver has to decide how much urb to allocate for each stage.

On gen6 only GS and VS will share an urb of 64k(32k on GT1)

urb_entry_number also has a limit, check 3DSTATE_URB command.

Ken and I did some research on this yesterday. The bspec says the valid range for the VS is [24, 256] on GT2 and [24, 128] on GT1. For the GS it says the valid range is [0, 256] on GT2 and [0, 254] on GT1. I can't believe that 254 is correct.

Quoted 9 lines

The more urb entry number we have, the more current threads can spawn.

But
We should make sure
VS urb size = single_urb_size * gs_urb_entry
+
GS urb size = single_urb_size * vs_urb_entry
< total_urb_size

Quoted 8 lines

In most case, this is ok, unless we have a huge single_urb_size.

Since we are always using GS pass through mode for now.
The number for GS urb dose not matter.
So allocate more urb size for VS, less for GS may be better for huge urb_size

But consider we will support GS shader later and huge urb_size case is very rare, I separate urb equally for GS and VS.

-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.11 (GNU/Linux)
Comment: Using GnuPG with Fedora - http://enigmail.mozdev.org/

iEYEARECAAYFAk2fXAYACgkQX1gOwKyEAw+RSQCdFiIxdyeKmytlovCEO3w3RIIA uN4An1POHfJ3J4fzKZnGWbHLb4eZBwnM
=xpu8
-----END PGP SIGNATURE-----