Use correct offset for subBuffer created with clCreateSubBuffer
Make calls to clGetMemObjectInfo for CL_MEM_OFFSET and
CL_MEM_ASSOCIATED_MEMOBJECT return values like with pool buffer disabled
Related-To: NEO-7332
Signed-off-by: Dominik Dabek <dominik.dabek@intel.com>
Fix in copy only command lists for reset event call.
First fix is to perform reset on all active events.
Second is to add signal all packets call for copy only branch
Related-To: NEO-7490
Signed-off-by: Zbigniew Zdanowicz <zbigniew.zdanowicz@intel.com>
This fix reposition signal commands before possible staling commands
like post sync pipe control or multi tile barrier
Related-To: NEO-7490
Signed-off-by: Zbigniew Zdanowicz <zbigniew.zdanowicz@intel.com>
First fix is to calculate remaining packets only when available.
Second fix is to use multi-tile signal when number of remaining packets
does match partition count - this is to resolve subdevice - root device
possible conflicts.
Related-To: NEO-7490
Signed-off-by: Zbigniew Zdanowicz <zbigniew.zdanowicz@intel.com>
Observed about 50MB reduction in overall binaries size (directory build))
when building all targets
with MSVC (Visual Studio 2022 17.3.0 preview 6)
using Debug 64 configuration.
Signed-off-by: Kamil Kopryk <kamil.kopryk@intel.com>
To use, applications need to set EnableSetPair=1 explicitly.
When disabled, implicit scaling allocations require two IPC handles
to be exchanged with other processes using the zexMemGetIpcHandles
APIs.
When enabled, implicit scaling allocations only require one IPC
handle to be exchanged with other process using the zeMemGetIpcHandle
APIs. This is only available when allocation is imported in a different
device than the one in the exporter.
Signed-off-by: Jaime Arteaga <jaime.a.arteaga.molina@intel.com>
When env variable is set, then copies are always done on CPU.
Change the logic of CPU copy to make sure we lock if targeting device memory.
Related-To: NEO-7564
Signed-off-by: Michal Mrozek <michal.mrozek@intel.com>
Enable resolving dependencies by pipecontrol on same CSR, IOQ on DG2 by
default.
Related-To: NEO-7321
Signed-off-by: Dominik Dabek <dominik.dabek@intel.com>
This change has two issues fixed.
First fix assures event must not download all allocations sent to GPU
when event is not ready.
Second fix performs page walk on event allocation before event allocation
can be downloaded, as download before page walk is not supported scenario
in TBX mode.
Signed-off-by: Zbigniew Zdanowicz <zbigniew.zdanowicz@intel.com>
Earlier implementation of sysman events API was based on file
creation in the filesystem. Whenever a uevent for some event
which needs to be monitored arrive, at that time a file was
created in the filesystem based on some preinstalled udev rules.
This approach was inefficient as it heavily depends over file
system and second with this approach losing events is always a
possibility.
Now with this change, we are removing our dependency over file
creation in filesystem. Rather we will be using libudev library
to monitor the uevents. This approach could also be extended,
when we want to listen to all the uevents for all the gpu
devices present in the system.
Related-To: LOCI-2140
Signed-off-by: Jitendra Sharma <jitendra.sharma@intel.com>
we use tag allocation for multiple purposes, therefore we should define
all offsets in one place
Resolves: NEO-7559
Signed-off-by: Mateusz Jablonski <mateusz.jablonski@intel.com>
Observed about 4MB reduction in overall binaries size (directory bin)
when building unit_tests target
with MSVC (Visual Studio 2022 17.3.0 preview 6)
using Debug configuration.
Signed-off-by: Kamil Kopryk <kamil.kopryk@intel.com>
