GitLab

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

Drops dependency on NIR etc.
Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

They're not quite VERTEX shaders.
Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

Snoop GPU ID and pass it on to pandecode

See merge request !18

Signed-off-by: Tomeu Vizoso <tomeu.vizoso@collabora.com>

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

Since it's all blackmagic anyway, I don't know why I bothered, but
that's okay.
Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

Rather than dumping memory to the filesystem for analysis by an external
tool, we can link directly with the pandecode library built with mesa to
trace directly, which is significantly faster and more reliable.
Unfortunately, it makes building rather... complex... since we're still
out-of-tree.
Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

This fix got lost during a rebase, whoops.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>

Thank you Ryan for the reference code here.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>

Miniwrap

See merge request !17

Initial Bifrost decoding support

See merge request !16

Rather than hardcoding it

Also, document the reg_preload field of shader_meta.

It turns out that compute jobs reuse the same format as vertex jobs, so
this also lets us decode some previously unknown fields in
mali_vertex_tiler_prefix.

In order to properly support 64-bit jobs as well as Bifrost, a number of
structures need to be moved around, split, etc. which requires a lot of
shuffling things around. I think this would probably be hard to untangle
from the actual new Bifrost bits, so I've left everything together in
one commit. The list of things changed includes:

- To save some space due to the larger pointers, vertex and tiler jobs
  no longer have the same layout (i.e. common fields like shader_meta,
  etc. no longer have the same offset). A few of the tiler-only fields
  and vertex-only fields now overlap. The tiler job structure is much
  larger than the vertex job structure, although most of the new fields
  are zero for the things I've tried so far. It's not clear whether this
  also also happens on Midgard 64-bit jobs. In addition, there is a new
  Bifrost-only fused job type, where the common fields in the beginning
  are shared. The end result is that there is now a common prefix,
  followed by tiler-only or vertex-only fields (or both, for Midgard),
  and then a common postfix.
- For Bifrost, we no longer have to make a special incomplete FBD just
  for vertex and tiler jobs. Instead, the relevant fields (tiler_meta
  and the scratchpad) have been moved out into separate structures, and
  removed from the FBD when not needed for fragment jobs.
- There is no Bifrost driver that uses the SFBD, so we need to decode
  the MFBD now.

Bifrost assembler fixes

See merge request !15

It turns out that when using the constant quadword format, we sometimes
need to swap the first and second constants in the quadword, or else
some of the bits are corrupted when loading either constant. Doing this
is tricky, since it has to happen before encoding the instructions
(because we need to know which slot to load constants from) and hence we
have to duplicate some of the logic for encoding immediates to figure
out ahead of time which constants get grouped together. We also have to
insert dummy constants earlier, since a dummy constant may have to be
swapped with a non-dummy one.

We wouldn't assign port 2 in that case.