Skip to content

[L0] Add support for multi-device kernel compilation #1135

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

Merged
merged 3 commits into from
Jan 23, 2024
Merged

[L0] Add support for multi-device kernel compilation #1135

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
8868230
[L0] Add support for multi-device kernel compilation
nrspruit Nov 28, 2023
c2682e9
[L0] don't delete program handle in kernel release and check work dim…
nrspruit Jan 17, 2024
1b2cd5b
[L0] Remove Device Filtering with urProgramBuild Handling MultiDevice
nrspruit Jan 18, 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
111 changes: 78 additions & 33 deletions source/adapters/level_zero/kernel.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -41,6 +41,15 @@ UR_APIEXPORT ur_result_t UR_APICALL urEnqueueKernelLaunch(
*OutEvent ///< [in,out][optional] return an event object that identifies
///< this particular kernel execution instance.
) {
auto ZeDevice = Queue->Device->ZeDevice;

ze_kernel_handle_t ZeKernel{};
if (Kernel->ZeKernelMap.empty()) {
ZeKernel = Kernel->ZeKernel;
} else {
auto It = Kernel->ZeKernelMap.find(ZeDevice);
ZeKernel = It->second;
}
// Lock automatically releases when this goes out of scope.
std::scoped_lock<ur_shared_mutex, ur_shared_mutex, ur_shared_mutex> Lock(
Queue->Mutex, Kernel->Mutex, Kernel->Program->Mutex);
Expand All @@ -51,7 +60,7 @@ UR_APIEXPORT ur_result_t UR_APICALL urEnqueueKernelLaunch(
}

ZE2UR_CALL(zeKernelSetGlobalOffsetExp,
(Kernel->ZeKernel, GlobalWorkOffset[0], GlobalWorkOffset[1],
(ZeKernel, GlobalWorkOffset[0], GlobalWorkOffset[1],
GlobalWorkOffset[2]));
}

Expand All @@ -65,18 +74,22 @@ UR_APIEXPORT ur_result_t UR_APICALL urEnqueueKernelLaunch(
Queue->Device));
}
ZE2UR_CALL(zeKernelSetArgumentValue,
(Kernel->ZeKernel, Arg.Index, Arg.Size, ZeHandlePtr));
(ZeKernel, Arg.Index, Arg.Size, ZeHandlePtr));
}
Kernel->PendingArguments.clear();

ze_group_count_t ZeThreadGroupDimensions{1, 1, 1};
uint32_t WG[3]{};

// global_work_size of unused dimensions must be set to 1
UR_ASSERT(WorkDim == 3 || GlobalWorkSize[2] == 1,
UR_RESULT_ERROR_INVALID_VALUE);
UR_ASSERT(WorkDim >= 2 || GlobalWorkSize[1] == 1,
UR_RESULT_ERROR_INVALID_VALUE);
if (WorkDim >= 2) {
UR_ASSERT(WorkDim >= 2 || GlobalWorkSize[1] == 1,
UR_RESULT_ERROR_INVALID_VALUE);
if (WorkDim == 3) {
UR_ASSERT(WorkDim == 3 || GlobalWorkSize[2] == 1,
UR_RESULT_ERROR_INVALID_VALUE);
}
}
if (LocalWorkSize) {
// L0
UR_ASSERT(LocalWorkSize[0] < (std::numeric_limits<uint32_t>::max)(),
Expand All @@ -99,7 +112,7 @@ UR_APIEXPORT ur_result_t UR_APICALL urEnqueueKernelLaunch(
}
if (SuggestGroupSize) {
ZE2UR_CALL(zeKernelSuggestGroupSize,
(Kernel->ZeKernel, GlobalWorkSize[0], GlobalWorkSize[1],
(ZeKernel, GlobalWorkSize[0], GlobalWorkSize[1],
GlobalWorkSize[2], &WG[0], &WG[1], &WG[2]));
} else {
for (int I : {0, 1, 2}) {
Expand Down Expand Up @@ -175,7 +188,7 @@ UR_APIEXPORT ur_result_t UR_APICALL urEnqueueKernelLaunch(
return UR_RESULT_ERROR_INVALID_WORK_GROUP_SIZE;
}

ZE2UR_CALL(zeKernelSetGroupSize, (Kernel->ZeKernel, WG[0], WG[1], WG[2]));
ZE2UR_CALL(zeKernelSetGroupSize, (ZeKernel, WG[0], WG[1], WG[2]));

bool UseCopyEngine = false;
_ur_ze_event_list_t TmpWaitList;
Expand Down Expand Up @@ -227,18 +240,16 @@ UR_APIEXPORT ur_result_t UR_APICALL urEnqueueKernelLaunch(
Queue->CaptureIndirectAccesses();
// Add the command to the command list, which implies submission.
ZE2UR_CALL(zeCommandListAppendLaunchKernel,
(CommandList->first, Kernel->ZeKernel, &ZeThreadGroupDimensions,
ZeEvent, (*Event)->WaitList.Length,
(*Event)->WaitList.ZeEventList));
(CommandList->first, ZeKernel, &ZeThreadGroupDimensions, ZeEvent,
(*Event)->WaitList.Length, (*Event)->WaitList.ZeEventList));
} else {
// Add the command to the command list for later submission.
// No lock is needed here, unlike the immediate commandlist case above,
// because the kernels are not actually submitted yet. Kernels will be
// submitted only when the comamndlist is closed. Then, a lock is held.
ZE2UR_CALL(zeCommandListAppendLaunchKernel,
(CommandList->first, Kernel->ZeKernel, &ZeThreadGroupDimensions,
ZeEvent, (*Event)->WaitList.Length,
(*Event)->WaitList.ZeEventList));
(CommandList->first, ZeKernel, &ZeThreadGroupDimensions, ZeEvent,
(*Event)->WaitList.Length, (*Event)->WaitList.ZeEventList));
}

urPrint("calling zeCommandListAppendLaunchKernel() with"
Expand Down Expand Up @@ -363,23 +374,46 @@ UR_APIEXPORT ur_result_t UR_APICALL urKernelCreate(
return UR_RESULT_ERROR_INVALID_PROGRAM_EXECUTABLE;
}

ZeStruct<ze_kernel_desc_t> ZeKernelDesc;
ZeKernelDesc.flags = 0;
ZeKernelDesc.pKernelName = KernelName;

ze_kernel_handle_t ZeKernel;
ZE2UR_CALL(zeKernelCreate, (Program->ZeModule, &ZeKernelDesc, &ZeKernel));

try {
ur_kernel_handle_t_ *UrKernel =
new ur_kernel_handle_t_(ZeKernel, true, Program);
ur_kernel_handle_t_ *UrKernel = new ur_kernel_handle_t_(true, Program);
*RetKernel = reinterpret_cast<ur_kernel_handle_t>(UrKernel);
} catch (const std::bad_alloc &) {
return UR_RESULT_ERROR_OUT_OF_HOST_MEMORY;
} catch (...) {
return UR_RESULT_ERROR_UNKNOWN;
}

for (auto It : Program->ZeModuleMap) {
auto ZeModule = It.second;
ZeStruct<ze_kernel_desc_t> ZeKernelDesc;
ZeKernelDesc.flags = 0;
ZeKernelDesc.pKernelName = KernelName;

ze_kernel_handle_t ZeKernel;
ZE2UR_CALL(zeKernelCreate, (ZeModule, &ZeKernelDesc, &ZeKernel));

auto ZeDevice = It.first;

// Store the kernel in the ZeKernelMap so the correct
// kernel can be retrieved later for a specific device
// where a queue is being submitted.
(*RetKernel)->ZeKernelMap[ZeDevice] = ZeKernel;
(*RetKernel)->ZeKernels.push_back(ZeKernel);

// If the device used to create the module's kernel is a root-device
// then store the kernel also using the sub-devices, since application
// could submit the root-device's kernel to a sub-device's queue.
uint32_t SubDevicesCount = 0;
zeDeviceGetSubDevices(ZeDevice, &SubDevicesCount, nullptr);
std::vector<ze_device_handle_t> ZeSubDevices(SubDevicesCount);
zeDeviceGetSubDevices(ZeDevice, &SubDevicesCount, ZeSubDevices.data());
for (auto ZeSubDevice : ZeSubDevices) {
(*RetKernel)->ZeKernelMap[ZeSubDevice] = ZeKernel;
}
}

(*RetKernel)->ZeKernel = (*RetKernel)->ZeKernelMap.begin()->second;

UR_CALL((*RetKernel)->initialize());

return UR_RESULT_SUCCESS;
Expand All @@ -396,6 +430,8 @@ UR_APIEXPORT ur_result_t UR_APICALL urKernelSetArgValue(
) {
std::ignore = Properties;

UR_ASSERT(Kernel, UR_RESULT_ERROR_INVALID_NULL_HANDLE);

// OpenCL: "the arg_value pointer can be NULL or point to a NULL value
// in which case a NULL value will be used as the value for the argument
// declared as a pointer to global or constant memory in the kernel"
Expand All @@ -409,8 +445,11 @@ UR_APIEXPORT ur_result_t UR_APICALL urKernelSetArgValue(
}

std::scoped_lock<ur_shared_mutex> Guard(Kernel->Mutex);
ZE2UR_CALL(zeKernelSetArgumentValue,
(Kernel->ZeKernel, ArgIndex, ArgSize, PArgValue));
for (auto It : Kernel->ZeKernelMap) {
auto ZeKernel = It.second;
ZE2UR_CALL(zeKernelSetArgumentValue,
(ZeKernel, ArgIndex, ArgSize, PArgValue));
}

return UR_RESULT_SUCCESS;
}
Expand Down Expand Up @@ -596,16 +635,21 @@ UR_APIEXPORT ur_result_t UR_APICALL urKernelRelease(

auto KernelProgram = Kernel->Program;
if (Kernel->OwnNativeHandle) {
auto ZeResult = ZE_CALL_NOCHECK(zeKernelDestroy, (Kernel->ZeKernel));
// Gracefully handle the case that L0 was already unloaded.
if (ZeResult && ZeResult != ZE_RESULT_ERROR_UNINITIALIZED)
return ze2urResult(ZeResult);
for (auto &ZeKernel : Kernel->ZeKernels) {
auto ZeResult = ZE_CALL_NOCHECK(zeKernelDestroy, (ZeKernel));
// Gracefully handle the case that L0 was already unloaded.
if (ZeResult && ZeResult != ZE_RESULT_ERROR_UNINITIALIZED)
return ze2urResult(ZeResult);
}
}
Kernel->ZeKernelMap.clear();
if (IndirectAccessTrackingEnabled) {
UR_CALL(urContextRelease(KernelProgram->Context));
}
// do a release on the program this kernel was part of
UR_CALL(urProgramRelease(KernelProgram));
// do a release on the program this kernel was part of without delete of the
// program handle
KernelProgram->ur_release_program_resources(false);

delete Kernel;

return UR_RESULT_SUCCESS;
Expand Down Expand Up @@ -639,6 +683,7 @@ UR_APIEXPORT ur_result_t UR_APICALL urKernelSetExecInfo(
std::ignore = PropSize;
std::ignore = Properties;

auto ZeKernel = Kernel->ZeKernel;
std::scoped_lock<ur_shared_mutex> Guard(Kernel->Mutex);
if (PropName == UR_KERNEL_EXEC_INFO_USM_INDIRECT_ACCESS &&
*(static_cast<const ur_bool_t *>(PropValue)) == true) {
Expand All @@ -649,7 +694,7 @@ UR_APIEXPORT ur_result_t UR_APICALL urKernelSetExecInfo(
ZE_KERNEL_INDIRECT_ACCESS_FLAG_HOST |
ZE_KERNEL_INDIRECT_ACCESS_FLAG_DEVICE |
ZE_KERNEL_INDIRECT_ACCESS_FLAG_SHARED;
ZE2UR_CALL(zeKernelSetIndirectAccess, (Kernel->ZeKernel, IndirectFlags));
ZE2UR_CALL(zeKernelSetIndirectAccess, (ZeKernel, IndirectFlags));
} else if (PropName == UR_KERNEL_EXEC_INFO_CACHE_CONFIG) {
ze_cache_config_flag_t ZeCacheConfig{};
auto CacheConfig =
Expand All @@ -663,7 +708,7 @@ UR_APIEXPORT ur_result_t UR_APICALL urKernelSetExecInfo(
else
// Unexpected cache configuration value.
return UR_RESULT_ERROR_INVALID_VALUE;
ZE2UR_CALL(zeKernelSetCacheConfig, (Kernel->ZeKernel, ZeCacheConfig););
ZE2UR_CALL(zeKernelSetCacheConfig, (ZeKernel, ZeCacheConfig););
} else {
urPrint("urKernelSetExecInfo: unsupported ParamName\n");
return UR_RESULT_ERROR_INVALID_VALUE;
Expand Down
15 changes: 11 additions & 4 deletions source/adapters/level_zero/kernel.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -14,10 +14,8 @@
#include <unordered_set>

struct ur_kernel_handle_t_ : _ur_object {
ur_kernel_handle_t_(ze_kernel_handle_t Kernel, bool OwnZeHandle,
ur_program_handle_t Program)
: Context{nullptr}, Program{Program}, ZeKernel{Kernel},
SubmissionsCount{0}, MemAllocs{} {
ur_kernel_handle_t_(bool OwnZeHandle, ur_program_handle_t Program)
: Program{Program}, SubmissionsCount{0}, MemAllocs{} {
OwnNativeHandle = OwnZeHandle;
}

Expand All @@ -37,6 +35,15 @@ struct ur_kernel_handle_t_ : _ur_object {
// Level Zero function handle.
ze_kernel_handle_t ZeKernel;

// Map of L0 kernels created for all the devices for which a UR Program
// has been built. It may contain duplicated kernel entries for a root
// device and its sub-devices.
std::unordered_map<ze_device_handle_t, ze_kernel_handle_t> ZeKernelMap;

// Vector of L0 kernels. Each entry is unique, so this is used for
// destroying the kernels instead of ZeKernelMap
std::vector<ze_kernel_handle_t> ZeKernels;

// Counter to track the number of submissions of the kernel.
// When this value is zero, it means that kernel is not submitted for an
// execution - at this time we can release memory allocations referenced by
Expand Down
Loading