Skip to content

[EXP][CUDA] Enable Large cluster sizes and fix cluster dimensions being set for dimensions less than 3 #1765

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
wants to merge 11 commits into from
Closed

[EXP][CUDA] Enable Large cluster sizes and fix cluster dimensions being set for dimensions less than 3 #1765

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
11 commits
Select commit Hold shift + click to select a range
8f7f066
set attribute allowing cluster size greater than 8
AD2605 Jun 17, 2024
9369144
set property cudaFuncAttributeNonPortableClusterSizeAllowed only if c…
AD2605 Jun 17, 2024
a8c442d
set has_property_cluster_launch only if cluster property is used
AD2605 Jun 17, 2024
b91c582
fix cluster dimensions being set in accordance to grid dimensions
AD2605 Jun 17, 2024
3f2ed1c
fix ordering of cluster dims for workDim 2
AD2605 Jun 17, 2024
be8af68
fix compilation errors
AD2605 Jun 17, 2024
5d38fe1
review comments 1
AD2605 Jun 17, 2024
7c1c64e
review comments 1
AD2605 Jun 17, 2024
bf73617
Merge branch 'main' into atharva/allow_max_cluster_size
AD2605 Jun 19, 2024
4a23bb9
increase cluster size upon launch to check CU_FUNC_ATTRIBUTE_NON_PORT…
AD2605 Jun 21, 2024
7088393
Merge remote-tracking branch 'oneapi-ur/main' into atharva/allow_max_…
AD2605 Jun 27, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
61 changes: 41 additions & 20 deletions source/adapters/cuda/enqueue.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -530,6 +530,21 @@ UR_APIEXPORT ur_result_t UR_APICALL urEnqueueKernelLaunchCustomExp(
}

std::vector<CUlaunchAttribute> launch_attribute(numPropsInLaunchPropList);

// Early exit for zero size kernel
if (*pGlobalWorkSize == 0) {
return urEnqueueEventsWaitWithBarrier(hQueue, numEventsInWaitList,
phEventWaitList, phEvent);
}

// Set the number of threads per block to the number of threads per warp
// by default unless user has provided a better number
size_t ThreadsPerBlock[3] = {32u, 1u, 1u};
size_t BlocksPerGrid[3] = {1u, 1u, 1u};

uint32_t LocalSize = hKernel->getLocalSize();
CUfunction CuFunc = hKernel->get();

for (uint32_t i = 0; i < numPropsInLaunchPropList; i++) {
switch (launchPropList[i].id) {
case UR_EXP_LAUNCH_PROPERTY_ID_IGNORE: {
Expand All @@ -540,12 +555,32 @@ UR_APIEXPORT ur_result_t UR_APICALL urEnqueueKernelLaunchCustomExp(

launch_attribute[i].id = CU_LAUNCH_ATTRIBUTE_CLUSTER_DIMENSION;
// Note that cuda orders from right to left wrt SYCL dimensional order.
launch_attribute[i].value.clusterDim.x =
launchPropList[i].value.clusterDim[2];
launch_attribute[i].value.clusterDim.y =
launchPropList[i].value.clusterDim[1];
launch_attribute[i].value.clusterDim.z =
launchPropList[i].value.clusterDim[0];
if (workDim == 3) {
launch_attribute[i].value.clusterDim.x =
Comment on lines +558 to +559
Copy link
Contributor Author

@AD2605 AD2605 Jun 17, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Could use some help Here -

I was not able to figure out where this flipping of order happens,
I see it's being set in setKernelParams but how it flips it I was not able to understand it

launchPropList[i].value.clusterDim[2];
launch_attribute[i].value.clusterDim.y =
launchPropList[i].value.clusterDim[1];
launch_attribute[i].value.clusterDim.z =
launchPropList[i].value.clusterDim[0];
} else if (workDim == 2) {
launch_attribute[i].value.clusterDim.x =
launchPropList[i].value.clusterDim[1];
launch_attribute[i].value.clusterDim.y =
launchPropList[i].value.clusterDim[0];
launch_attribute[i].value.clusterDim.z =
launchPropList[i].value.clusterDim[2];
} else {
launch_attribute[i].value.clusterDim.x =
launchPropList[i].value.clusterDim[0];
launch_attribute[i].value.clusterDim.y =
launchPropList[i].value.clusterDim[1];
launch_attribute[i].value.clusterDim.z =
launchPropList[i].value.clusterDim[2];
}

UR_CHECK_ERROR(cuFuncSetAttribute(
CuFunc, CU_FUNC_ATTRIBUTE_NON_PORTABLE_CLUSTER_SIZE_ALLOWED, 1));

break;
}
case UR_EXP_LAUNCH_PROPERTY_ID_COOPERATIVE: {
Expand All @@ -560,20 +595,6 @@ UR_APIEXPORT ur_result_t UR_APICALL urEnqueueKernelLaunchCustomExp(
}
}

// Early exit for zero size kernel
if (*pGlobalWorkSize == 0) {
return urEnqueueEventsWaitWithBarrier(hQueue, numEventsInWaitList,
phEventWaitList, phEvent);
}

// Set the number of threads per block to the number of threads per warp
// by default unless user has provided a better number
size_t ThreadsPerBlock[3] = {32u, 1u, 1u};
size_t BlocksPerGrid[3] = {1u, 1u, 1u};

uint32_t LocalSize = hKernel->getLocalSize();
CUfunction CuFunc = hKernel->get();

// This might return UR_RESULT_ERROR_ADAPTER_SPECIFIC, which cannot be handled
// using the standard UR_CHECK_ERROR
if (ur_result_t Ret =
Expand Down
2 changes: 1 addition & 1 deletion test/conformance/exp_launch_properties/launch_properties.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -78,7 +78,7 @@ TEST_P(urEnqueueKernelLaunchCustomTest, Success) {
if (compute_capability >= 9.0) {
ur_exp_launch_property_t cluster_dims_prop;
cluster_dims_prop.id = UR_EXP_LAUNCH_PROPERTY_ID_CLUSTER_DIMENSION;
cluster_dims_prop.value.clusterDim[0] = 1;
cluster_dims_prop.value.clusterDim[0] = 16;
cluster_dims_prop.value.clusterDim[1] = 1;
cluster_dims_prop.value.clusterDim[2] = 1;

Expand Down