Current flow will be to have one synchronization point
config.file. Read remains unblocking, only write(caching)
operation will be blocking (lock on config.file)
Related-To: NEO-4262
Signed-off-by: Diedrich, Kamil <kamil.diedrich@intel.com>
Command list batch buffers should be chained when no dynamic or global preamble
is present in command queue.
Return to command queue, when preamble is required.
Chain last command list to the command queue epilog.
Provide first command list batch buffer to KMD/ULLS when no command queue
preamble.
Related-To: NEO-7807
Signed-off-by: Zbigniew Zdanowicz <zbigniew.zdanowicz@intel.com>
The memadvise with preferred location for kmd-migrated shared allocation
is set to device associated with cmd list by default to migrate data
to lmem on non-atomic gpu page fault as well (for performance reasons).
Related-To: NEO-7252
Signed-off-by: Milczarek, Slawomir <slawomir.milczarek@intel.com>
Single command list object can be passed multiple times to the execution
command list.
Not all command list instances might require dynamic preamble, as it depends
what state is before particular command list instance.
Correctly assign the particular instance of command list to state transition.
Related-To: NEO-7828
Signed-off-by: Zbigniew Zdanowicz <zbigniew.zdanowicz@intel.com>
Global init flag is useful only for once per context initialization.
Correctly set the flag can save the visits to these once per context
calls.
Debugger programming is active not only when queue type allows it,
but also when commands state is dirty and debugger class available.
Related-To: NEO-7828
Signed-off-by: Zbigniew Zdanowicz <zbigniew.zdanowicz@intel.com>
Unify fp16/fp32/fp64 across all platforms. The capabilities indicated by
those flags now refer to both emulated and native-supported (HW) ones:
- Global/local atomic load: HW support on all platforms (handled by native
i16 atomic or) for FP16, FP32 and FP64.
- Global/local atomic store: HW support on all platforms (handled by
native i16 atomic exchange) for FP16, FP32 and FP64.
- Global/local atomic compare/exchange: HW support on all platforms
for FP32.
- Global/local atomic min/max: Emulation support on all platforms for
FP64, HW support on all platforms for FP32, HW support on XE+ platforms
and emulation support on all others for FP16.
- Global atomic add: HW support for PVC+ platforms, emulation support on
all other platforms for FP64, HW support on XE+ platforms and emulation
support on all other platforms for FP32.
- Local atomic add: Emulation on all platforms for both FP64 and FP32.
Signed-off-by: Kacper Nowak <kacper.nowak@intel.com>
Related-To: NEO-7734
Immediate command list can use internal command queue.
Immediate command list then uses variable start offset and it does not
work with primary batch buffer.
Related-To: NEO-7807
Signed-off-by: Zbigniew Zdanowicz <zbigniew.zdanowicz@intel.com>
Before state transition was done twice, 1st time for estimation, 2nd time for
dispatch.
Now state transitions only during estimation and required state is saved then.
Commands are dispatched only when command list and property are marked to
dispatch.
During regular workload submission transition is performed only once and it
should be benefitial to reduce host overhead.
Related-To: NEO-7828
Signed-off-by: Zbigniew Zdanowicz <zbigniew.zdanowicz@intel.com>
Add "DumpZEBin" debug flag. When this flag is enabled, Zebin will be
dumped to a .elf file (with appropiate suffix, in case such file has
been dumped before).
Signed-off-by: Kacper Nowak <kacper.nowak@intel.com>
Related-To: NEO-7895
Enable cpu copy for USM device to USM host transfer in level zero
immediate cmdlist.
Related-To: NEO-7553
Signed-off-by: Dominik Dabek <dominik.dabek@intel.com>
1. separate front end programing when tracking is enabled and disabled, it will
limit number of conditional checks.
2. setup command list front end properties only when front end state is dirty.
3. instanced context id should be set once, as this is one time per context
property.
Related-To: NEO-7828
Signed-off-by: Zbigniew Zdanowicz <zbigniew.zdanowicz@intel.com>
When getting residency count for all command lists, driver is able to
reallocate container only once and not per each command list.
Add non-zero initial value for command queue residual allocations.
Related-To: NEO-7828
Signed-off-by: Zbigniew Zdanowicz <zbigniew.zdanowicz@intel.com>
- group same implementation into dedicated inl files
- remove double implementations for the similiar hw generations
Signed-off-by: Zbigniew Zdanowicz <zbigniew.zdanowicz@intel.com>
In the case of mtl+ platforms, the returned config value
should equal the hardware ip version value.
This change fixes situations where some config has not been
added and in this case we returned an unknown value.
Signed-off-by: Daria Hinz <daria.hinz@intel.com>
Related-To: NEO-7738
For primary batch buffer command list driver should not use return point.
Return points are useful when batch buffers are dispatched as secondary,
for primary buffers, patching of front end command is more desirable option.
Related-To: NEO-7807
Signed-off-by: Zbigniew Zdanowicz <zbigniew.zdanowicz@intel.com>
Implicit Scaling barrier have the same requirements as kernel.
It must dispach bb start command with the same level as the command list
is dispatched.
Related-To: NEO-7807
Signed-off-by: Zbigniew Zdanowicz <zbigniew.zdanowicz@intel.com>
Extended the regkey ForceMemoryPrefetchForKmdMigratedSharedAllocations
to force meory prefetch of kmd-migrated shared allocation
in clEnqueueNDRangeKernel(), clEnqueueMemFillINTEL, ...
Related-To: NEO-7841
Signed-off-by: Milczarek, Slawomir <slawomir.milczarek@intel.com>
Related-To: LOCI-4174
- Call zelSetDriverTeardown during L0 Driver teardown to prevent users
from calling into destroyed functions and encountering crashes
during teardown.
Signed-off-by: Neil R Spruit <neil.r.spruit@intel.com>
This fix is most important for multi command list execution use cases.
It is also benefitial for single command list execution, as driver saves
on loop enters and exits.
Methods handling single command list instead of array of objects are simpler.
Removed loops were at:
- CommandListExecutionContext constructor
- estimateLinearStreamSizeInitial method
- computePreemptionSize method
- collectPrintfContentsFromAllCommandsLists method
Related-To: NEO-7828
Signed-off-by: Zbigniew Zdanowicz <zbigniew.zdanowicz@intel.com>
This change adds space reservation in command list for returning batch buffer
start hw command.
Primary batch buffer can be run from direct submission or from KMD call and
must be aligned to required size.
Ending patch for batch buffer start must be in the last command buffer of the
command list.
Related-To: NEO-7807
Signed-off-by: Zbigniew Zdanowicz <zbigniew.zdanowicz@intel.com>
Plenty of calls require hw command programming only once per context.
There is no need to visit every method of them every execute call.
Set global init flag only if any of them is true and then visit all of them.
But for regular command list execution it can save time when there is single
global check.
Related-To: NEO-7828
Signed-off-by: Zbigniew Zdanowicz <zbigniew.zdanowicz@intel.com>
This is number of small tweaks to state cache invalidation:
1. Invalidate if heap was actually created.
2. Check if os context was actually initialized.
3. Heap allocation was actually submitted, as it might attain zero task count
value, when allocation is stored in csr internal storage, as csr wasn't used,
but the csr task count being zero is assigned to heap allocation when stored.
Related-To: NEO-5055
Signed-off-by: Zbigniew Zdanowicz <zbigniew.zdanowicz@intel.com>
There is no need to reset all fields and load support flags every reset call.
Add dedicated calls that will reset values and dirty flags.
Call virtual methods only once at init time.
Related-To: NEO-7828
Signed-off-by: Zbigniew Zdanowicz <zbigniew.zdanowicz@intel.com>
Globals surface allocation via USM manager will have correct allocation
type set (instead of just BUFFER) and will use cpu copy when possible.
Related-To: NEO-7796
Signed-off-by: Dominik Dabek <dominik.dabek@intel.com>
total cache size calculation should use subdevice count
rather than MultiTileArchInfo to ensure correct size when subdevices
are queried
Related-to: LOCI-4217
Signed-off-by: Brandon Yates <brandon.yates@intel.com>
- 3D btd command should be programed only once per context
- Add conditional pipe control command prior dispatching 3D btd command
- share 3D btd state between immediate and regular command lists
- add pipe control after ray tracing kernel to invalidate state cache
Related-To: NEO-5055
Signed-off-by: Zbigniew Zdanowicz <zbigniew.zdanowicz@intel.com>
Add the regkey ForceMemoryPrefetchForKmdMigratedSharedAllocations
to force meory prefetch of kmd-migrated shared allocation
in zeCommandQueueExecuteCommandLists().
Related-To: NEO-7841
Signed-off-by: Milczarek, Slawomir <slawomir.milczarek@intel.com>
Include an offset for the appendMemoryCopyBlitRegion
parameters when an source/destination pointer is offseted within
allocation.
Lack of this offset could result in an invalid pointer when source and
destination are on the same allocation or they do not point the start of
allocation.
Related-To: NEO-7694
Signed-off-by: Naklicki, Mateusz <mateusz.naklicki@intel.com>
