OpenCL: Skip SVM pointer annotation if possible#785
Conversation
pjaaskel
left a comment
pjaaskel
left a comment
There was a problem hiding this comment.
Looks quite. Small things.
| @@ -0,0 +1,14 @@ | |||
| // Check we detect that the device code does not have indirect global | |||
There was a problem hiding this comment.
Pls add the copyright/license texts to new files.
There was a problem hiding this comment.
Done. Is it fine to use "short-form" copyright text? And was it the MIT license the chipStar is using?
There was a problem hiding this comment.
Not sure what is the consensus. I've seen Apache 2 in the LLVM files at least.
There was a problem hiding this comment.
The existing LICENSE file in project root seems to match to the MIT license.
|
I see some failures that seem unrelated to this PR, let me investigate. |
| @@ -0,0 +1,14 @@ | |||
| // Check we detect that the device code does not have indirect global | |||
There was a problem hiding this comment.
Not sure what is the consensus. I've seen Apache 2 in the LLVM files at least.
pvelesko
force-pushed
the
skip-svm-annotation
branch
from
073b933 to
6fddd63
Compare
linehill
force-pushed
the
skip-svm-annotation
branch
2 times, most recently
from
2edc2d8 to
bb22981
Compare
|
Rebased now. |
|
Only this is failing now This is fixed in #791 so let's merge that first |
|
conflits |
linehill
force-pushed
the
skip-svm-annotation
branch
from
bb22981 to
82c4b08
Compare
... for hosting module level information.
Skip calling clSetKernelExecInfo() for kernels from modules for which we know they have no indirect accesses to SVM allocations.
Co-authored-by: Pekka Jääskeläinen <pekka.jaaskelainen@intel.com>
linehill
force-pushed
the
skip-svm-annotation
branch
from
1540b45 to
e2c57cd
Compare
|
Rebased. Should have compiler error resolved seen in the CI in the merge prior this rebase. |
3 participants
Calling clSetKernelExecInfo() for passing CL_KERNEL_EXEC_INFO_SVM_PTRS demonstrates notable overhead on PVC. For example, HeCBench/hybridsort-hip sees 4-16% slowdown.
This patch adds a rudimentary analysis for detecting whether device modules might have indirect buffer accesses. The OpenCL backend skips clSetKernelExecInfo() calls when it knows the module, the kernel belongs to, does not have any indirect accesses.