- change the way we handle blocked commands.
- instead of allocating CPU pointer and populating it with commands, create
real IndirectHeap that may be later submitted to the GPU
- that removes a lot of copy operations that were happening on submit time
- for device enqueue, this requires dsh & shh to be passed directly to the
underlying commands, in that scenario device queue buffers are not used
Change-Id: I1124a8edbb46777ea7f7d3a5946f302e7fdf9665
* adding support for map/unmap
* adding support for origin/region validation with mipmaps
* fixing slices returned in map/unmap
* removing ambiguity around mipLevel naming
* enabling cl_khr_mipmap_image in current shape
* enabling cl_khr_mipmap_image_writes in current shape
* fixing CompileProgramWithReraFlag test
Change-Id: I0c9d83028c5c376f638e45151755fd2c7d0fb0ab
- While estimating the required size of Indirect Object Heap we were not
handling properly the lack of local ids case
- In such case we should allocate one GRF per HW thread that will be unused
Change-Id: Ibcd359e431e3ffd9d55628ac7cf7eeefad72e7ba
- cpu virtual address was used instead of gpu va
- this caused incorrect behaviour of TBX server when
special heap allocator assigning GPU addresses was used
Change-Id: I2328cf2441be797311fd6a3c7b331b0fff79d4fc
Any CPU related updates such as clEnqueueMapBuffer or similar
need to trigger a re-dump of memory prior to the next clEnqueue call.
Change-Id: I7b31e559278e92ff55b6ebab8ef4190caef1ebc0
- Do not obtain pattern allocation from reusable pool.
- This is due to the fact that it may contain allocations from internal
heap, which cannot be used for arguments declared as kernel argument.
Change-Id: I6c73445c409edc4ce25f8d8eba966f512dfd6cc9
- This removes Instruction Heap allocation from enqueue path
- Blocked path is handled as well
- Heap is no longer allocated on demand it is bind to kernelInfo.
Change-Id: I54545beceed3404ee0330a8bac2b0934944cac30
- Switch to internal heap for kernel ISA allocations.
- remove IH from various functions
- remove IHState from CSR , IH is never dirty
- ISA is no longer copied on enqueue calls.
Change-Id: I0099cf2a9ebab6192ea03a74dd35f7da963fd5a5
- Make sure that blocks ISA is made resident
- both blocked & non blocked path
- fix a bug where private surface was not made resident in blocked path.
Change-Id: Ie564595b176b94ecc7c79d7efeae20598c5874fb
-This is to make sure those functions are not called when gtpin is not used
-This preserves CPU instruction cache pollution.
-Our enqueue path needs to be as thin as possible, even with this small change
there is visible gain in ULT execution time.
Change-Id: I44cc2144754cda95ca1fe058184cd8a151b8d35c
- KmdNotifyProperties struct for CapabilityTable that can be extended by
incoming KmdNotify related optimizations
- Quick KMD sleep optimization that is called from async events handler
- Optimization makes a taskCount check in busy loop with much smaller
delay than basic version of KMD Notify optimization
Change-Id: I60c851c59895f0cf9de1e1f21e755a8b4c2fe900
- Previously only taskCount was updated
- This improves KMD notify usage for Events handled asynchronously
Change-Id: I283982890579254033557de0e1cef2239c0035e2
- Also refactor debug manager tests , they now check for default value
in igdrcl.config file
- There is no need to write dedicated tests now , so I remove them.
Change-Id: Ib338ca05b6059302c29469c673239e7886dc4b9b
- read only memory cannot be used for allocation,
Oses cannot create graphics alocation for such memory
- if memory allocation fails for host_ptr passed
to enqueueWrite calls, then try doing new allocation
and copy host_ptr on cpu
Change-Id: I415a4673ae1319ea8f77e53bd8fba7489fe85218
Add macro to add all subdirectories
Add macro to create project source tree based on target sources
Small cleanup runtime/CMakeLists.txt
Change-Id: I9b99145c544f648c4c3fe7421752d0c5d9504edf
- Due to use cases where one shared buffer may be mapped to multiple CL
buffers we need to flush DC between enqueues.
Change-Id: I05d7f844afe31d52a0004f5e2e5efa776f9dadbe
- Dont make cpu/gpu writes on read-only unmap
- Read/Write on limited map range only
- Overlaps checks for non read-only maps
- Fixed cmd type on returned event
Change-Id: I98ca542e8d369d2426a87279f86cadb0bf3db299
When queue is blocked on non-blocking call, map operation is added to
waitlist dependencies. Returning slice/row pitch for map image was skipped
Change-Id: I46f97590315e7aee7fbbfbdb615f383cdb666307
- Introducing MapInfo struct which will be used as container for multiple
map operations
- Unified mapped offset and size for Buffers and Images
- Fixed incorrect map params for CPU and GPU path
- Missing API level checks
Change-Id: Ib4077c9e2c0c333b131ffd5ccbc4a1404920eb5b
- Curently each non-zerocopy CPU operation on map/unmap make a full copy
using hostPtr
- This commit adds functionality to select specific range of copy
- Multiple mapping with different size is not supported yet,
so copy will be made on full range for now. This is for future usage.
Change-Id: I7652e85482ba6fffb2474169447baf9b080dcd1e
-For in order queue application can have fine grain granularity of completion
-For out of order queue application wants to execute workloads concurrently
-This change disables pipe control nooping for ioq calls when event returned.
Change-Id: Iaeaf677f768f7434b2efa1842b50653ab80777ad
- This change enabled multiple independent command queues to execute
concurrently without stalling pipe controls in between
- This change removes L3 flushes between kernels
- Dependencies between commands are resolved via task level mechanism
- Out of order queues are not changing task level between submissions
- In order queues are increasing task level between submissions
- Whenever task level changes there is pipe control with cs stall emitted
between GPGPU_WALKERs
Change-Id: I558653b296424e4775d060df3072e2a50684b715
- Some of the paths were made only for ULTs
- Params like mappedPtr were ignored
- Improve confusing method names
- Fix for memory leak in map shared buffer path (not tested code)
Change-Id: I8a69035f1d1c340f2d131a6f8d7e13116e3ddabc
- Instruction heap is currently heavily used as every kernel copies ISA into
it.
- It dries out very fast and each change to new heap requires whole pipeline
drain that prevents concurrency
- Problem is even larger when sip kernel is used as it limits the total heap
size
- In order to maximize heap re-use and to limit the count of pipeline drains
this change introduces new minimal size for instruction heap 512 KB.
Change-Id: Ic54e9ef4448b1d35dab01b084ee1d59b509642cb
- In various scenarios code was not programming the max heap size correctly
- It was possible for SSH to overcome the limit
- Size was programmed smaller then it really was, which resulted in smaller
reuse, which led to SBA reprogramming which led to lower performance in ooq
scenarios
- This change fixes the heap size programming by always utilizing full
allocation size and always limiting SSH at proper value
Change-Id: Ib703d2b0709ed8227a293def3a454bf1bb516dfd
Fixing InterfaceDescriptor programming for
blocked commands when MidThread preemption is
enabled
Additionally, fixing couple of tests that block
global preemption enabling in ULTs
Change-Id: I454c9608f8606f23d7446785ac24c7c7d8701ae0
- It should use thread count not EU count.
- change variable name to reflect that we work on sublices.
- fix test description, add missing test
- change hasBarrier variable to be boolean
Change-Id: I627bdf17b661d2f9b5eb3d8cd6ca53eba5d46b81
- Call waitForTaskCountAndCleanAllocationList with latest flushed task count
to reflect what was actually sent to HW.
- refactor cleanAllocationList to waitForTaskCountAndCleanAllocationList
Change-Id: I5301185c5fce212e39eb017b952b43c279559cf4
- Fix SUPPORT_X and TESTS_X flags in cmake
Previously it was possible to enable TESTS_X with disabled SUPPORT_X
- Remove genX dirs from include paths
Change-Id: I70aa6ee4df6cc1ca641cce2e99b66f92f99fd4e1
- Prevents destruction of MemObj while it may still be in use.
- Add UNRECOVERABLE to check whether object is deleted while having
dependencies, fix all problems is tests due to that fact.
- Fix special queue setting, clean interfaces.
Change-Id: I2a467e80df00ea1650decdcfa6866acf10b441f8
- due to the fact that device mutex was obtained to prevent threaded access to
image there was a problem when other thread was also doing readImage call
That thread got read Image kernel mutex first and then it was acquiring device
mutex, which was taken by other thread doing mapImage call.
- In current code device mutex is not taken to service mapImage call, instead
image is being guarded by its own mutex.
Change-Id: Ic4c5a019708d7ec5b240bc5b08c5a65173827392
- This causes event tree update if virtual event is holding commands or
callbacks
- That causes race between other threads that may be updating the tree
Change-Id: Ic80a8b71ed1e1c1deab8af1bc64f8ce81c21de1b
Mrozek, Michal