IREE AIE #14
IREE AIE #14
Conversation
makslevental
force-pushed
the
makslevental/aie-tests
branch
13 times, most recently
from
1aa93eb
to
127c9a5
Compare
makslevental
force-pushed
the
amd-staging
branch
from
7697356
to
c664d92
Compare
makslevental
force-pushed
the
iree-aie
branch
from
53faf46
to
773afe8
Compare
|
Before merging this PR I'd like for us to explore if we can remove the need for a new AMD ext function (change 3d8273b) for setting HW CTX config. I think this more properly belongs as a parameter that should be set in the dispatch pkt. If that doesn't work, we can leverage the ext I added. But, we should have a really good reason for proposing a new ext so want to think on it some more. |
Just FYI my current theory of operations in this repo is not to merge this branch but rather use it as a dev branch from which to send PRs upstream that then wind up in I say current because I'm not 100% sure if this is the right arrangement (will probably discuss with @MaheshRavishankar this week). |
makslevental
force-pushed
the
iree-aie
branch
from
b913e4a
to
6160643
Compare
Adds support for initialzing the XDNA driver so that a hardware context can be created for an AIE queue. Right now this initializes the device heap in the driver, gets the relevant tile parameters for the AIE agent, and creates a hardware context that backs the AIE queue. Change-Id: Ib90e1bc67a8637f6db3ff2bebe34677843796417
Adds support for the packet interface for interacting with the Embedded Runtime (ERT) on AIE agents. The ERT is what interprets command packets send to the AIE agent work queues. Change-Id: Id28fb98056b2c046354c446bdc9568d74385bea1
Change-Id: If180bdbcb3eb659f0d05a710526864494316d7a9
Change-Id: I799a8223d695ec5c0ea2eaea012bc1b5d877e103
This adds an AMD extension API for configuring a queue's hardware context. For AIEs the embedded runtime (ERT) creates a hardware context for a queue that has various configurable parameters, notably CUs can be configured. This API could also be used for future GPU HW context configuration (e.g., it could be used as a general way to set the CU masks). Change-Id: I7d1ce36d7ad6830f1b0c867bb2646e6e09cf4845
Adds support for AllocateMemoryOnly inside XDNA driver. Move the IsLocalMemory() check inside the KFD driver since the XDNA driver can, and needs to, create handles on system memory buffer objects. Changed handle variable name from thunk_handle to user_mode_driver_handle, which is more representative if we support non-GPU drivers. Change-Id: I95db9d575afd1ab0ff2de74cea5175d9a12a721b
Storing the correct dev heap size in the memory region. Change-Id: I14b053330c187da1d7d0213256625e50795b9902
* Add missing info cases to AIEAgent * Following the single value set from GPUAgent * AIEs have no caches, no cacheline is reported
* Adding a sideband interface to get a handle from a virtual address. This is just temporary and should be removed when we move to the vmem API * Removing unused queue structure * Removing the queue from the arguments of hsa_amd_get_handle_from_vaddr * Free XDNA BOs --------- Co-authored-by: Yiannis Papadopoulos <[email protected]>
#19) * Using BOs of type BO_SHMEM instead of BO_CMD for kernarg memory region * Changing test to use kernarg memory region for operand allocation
makslevental
force-pushed
the
iree-aie
branch
from
6160643
to
c817704
Compare
…er of packets the queue supports. (#23) * Fixed workarounds in the AIE soft queue regarding the size of the command chain and the individual commands. * Added the functionality to aie_hsa_dispatch_test.cc to query the size of the AIE queue and issue the maximum number of packets it supports. * Some additional small fixes on aie_hsa_dispatch_test.cc test. * Adding links to driver source
* Freeing the commands and the command chain created during dispatch * Unmaping and closing cmd BOs as well as freeing the queue ring buffer
No description provided.