Skip to content

[SYCL][UR][L0 v2] Implement OOO queue #18551

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 5 commits into from
Closed

[SYCL][UR][L0 v2] Implement OOO queue #18551

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
3d6e38f
[SYCL][UR][L0 v2] Implement OOO queue
igchor May 16, 2025
4790e24
Implement eventsWait, etc by waiting on signal events
igchor May 20, 2025
183d528
XXX
igchor May 20, 2025
dd2eba5
Bench with higher NumKernels
igchor May 20, 2025
78eb05a
iterate
igchor May 21, 2025
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
2 changes: 1 addition & 1 deletion devops/scripts/benchmarks/benches/compute.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -367,7 +367,7 @@ def bin_args(self) -> list[str]:
"--iterations=100000",
"--Profiling=0",
"--NumKernels=10",
"--KernelExecTime=1",
"--KernelExecTime=20",
f"--UseEvents={self.UseEvents}",
]

Expand Down
1 change: 1 addition & 0 deletions unified-runtime/source/adapters/level_zero/CMakeLists.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -186,6 +186,7 @@ if(UR_BUILD_ADAPTER_L0_V2)
${CMAKE_CURRENT_SOURCE_DIR}/v2/queue_api.cpp
${CMAKE_CURRENT_SOURCE_DIR}/v2/queue_create.cpp
${CMAKE_CURRENT_SOURCE_DIR}/v2/queue_immediate_in_order.cpp
${CMAKE_CURRENT_SOURCE_DIR}/v2/queue_immediate_out_of_order.cpp
${CMAKE_CURRENT_SOURCE_DIR}/v2/usm.cpp
)
install_ur_library(ur_adapter_level_zero_v2)
Expand Down
41 changes: 20 additions & 21 deletions unified-runtime/source/adapters/level_zero/v2/command_buffer.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -68,8 +68,7 @@ ur_exp_command_buffer_handle_t_::ur_exp_command_buffer_handle_t_(
const ur_exp_command_buffer_desc_t *desc)
: commandListManager(
context, device,
std::forward<v2::raii::command_list_unique_handle>(commandList),
v2::EVENT_FLAGS_COUNTER, nullptr),
std::forward<v2::raii::command_list_unique_handle>(commandList)),
isUpdatable(desc ? desc->isUpdatable : false), context(context),
device(device) {}

Expand Down Expand Up @@ -246,7 +245,7 @@ ur_result_t urCommandBufferAppendKernelLaunchExp(
commandListLocked, hKernel, workDim, pGlobalWorkSize,
numKernelAlternatives, kernelAlternatives, command));
}
UR_CALL(commandListLocked->appendKernelLaunch(
UR_CALL(commandListLocked->enqueueKernelLaunch(
hKernel, workDim, pGlobalWorkOffset, pGlobalWorkSize, pLocalWorkSize, 0,
nullptr, nullptr));
return UR_RESULT_SUCCESS;
Expand All @@ -266,8 +265,8 @@ ur_result_t urCommandBufferAppendUSMMemcpyExp(

// Responsibility of UMD to offload to copy engine
auto commandListLocked = hCommandBuffer->commandListManager.lock();
UR_CALL(commandListLocked->appendUSMMemcpy(false, pDst, pSrc, size, 0,
nullptr, nullptr));
UR_CALL(commandListLocked->enqueueUSMMemcpy(false, pDst, pSrc, size, 0,
nullptr, nullptr));

return UR_RESULT_SUCCESS;
} catch (...) {
Expand All @@ -289,7 +288,7 @@ ur_result_t urCommandBufferAppendMemBufferCopyExp(
// sync mechanic can be ignored, because all lists are in-order
// Responsibility of UMD to offload to copy engine
auto commandListLocked = hCommandBuffer->commandListManager.lock();
UR_CALL(commandListLocked->appendMemBufferCopy(
UR_CALL(commandListLocked->enqueueMemBufferCopy(
hSrcMem, hDstMem, srcOffset, dstOffset, size, 0, nullptr, nullptr));

return UR_RESULT_SUCCESS;
Expand All @@ -312,8 +311,8 @@ ur_result_t urCommandBufferAppendMemBufferWriteExp(
// sync mechanic can be ignored, because all lists are in-order
// Responsibility of UMD to offload to copy engine
auto commandListLocked = hCommandBuffer->commandListManager.lock();
UR_CALL(commandListLocked->appendMemBufferWrite(hBuffer, false, offset, size,
pSrc, 0, nullptr, nullptr));
UR_CALL(commandListLocked->enqueueMemBufferWrite(hBuffer, false, offset, size,
pSrc, 0, nullptr, nullptr));

return UR_RESULT_SUCCESS;
} catch (...) {
Expand All @@ -334,8 +333,8 @@ ur_result_t urCommandBufferAppendMemBufferReadExp(
// the same issue as in urCommandBufferAppendKernelLaunchExp
// Responsibility of UMD to offload to copy engine
auto commandListLocked = hCommandBuffer->commandListManager.lock();
UR_CALL(commandListLocked->appendMemBufferRead(hBuffer, false, offset, size,
pDst, 0, nullptr, nullptr));
UR_CALL(commandListLocked->enqueueMemBufferRead(hBuffer, false, offset, size,
pDst, 0, nullptr, nullptr));

return UR_RESULT_SUCCESS;
} catch (...) {
Expand All @@ -359,7 +358,7 @@ ur_result_t urCommandBufferAppendMemBufferCopyRectExp(
// sync mechanic can be ignored, because all lists are in-order
// Responsibility of UMD to offload to copy engine
auto commandListLocked = hCommandBuffer->commandListManager.lock();
UR_CALL(commandListLocked->appendMemBufferCopyRect(
UR_CALL(commandListLocked->enqueueMemBufferCopyRect(
hSrcMem, hDstMem, srcOrigin, dstOrigin, region, srcRowPitch,
srcSlicePitch, dstRowPitch, dstSlicePitch, 0, nullptr, nullptr));

Expand All @@ -385,7 +384,7 @@ ur_result_t urCommandBufferAppendMemBufferWriteRectExp(

// Responsibility of UMD to offload to copy engine
auto commandListLocked = hCommandBuffer->commandListManager.lock();
UR_CALL(commandListLocked->appendMemBufferWriteRect(
UR_CALL(commandListLocked->enqueueMemBufferWriteRect(
hBuffer, false, bufferOffset, hostOffset, region, bufferRowPitch,
bufferSlicePitch, hostRowPitch, hostSlicePitch, pSrc, 0, nullptr,
nullptr));
Expand All @@ -412,7 +411,7 @@ ur_result_t urCommandBufferAppendMemBufferReadRectExp(

// Responsibility of UMD to offload to copy engine
auto commandListLocked = hCommandBuffer->commandListManager.lock();
UR_CALL(commandListLocked->appendMemBufferReadRect(
UR_CALL(commandListLocked->enqueueMemBufferReadRect(
hBuffer, false, bufferOffset, hostOffset, region, bufferRowPitch,
bufferSlicePitch, hostRowPitch, hostSlicePitch, pDst, 0, nullptr,
nullptr));
Expand All @@ -434,8 +433,8 @@ ur_result_t urCommandBufferAppendUSMFillExp(
ur_exp_command_buffer_command_handle_t * /*phCommand*/) try {

auto commandListLocked = hCommandBuffer->commandListManager.lock();
UR_CALL(commandListLocked->appendUSMFill(pMemory, patternSize, pPattern, size,
0, nullptr, nullptr));
UR_CALL(commandListLocked->enqueueUSMFill(pMemory, patternSize, pPattern,
size, 0, nullptr, nullptr));
return UR_RESULT_SUCCESS;
} catch (...) {
return exceptionToResult(std::current_exception());
Expand All @@ -454,7 +453,7 @@ ur_result_t urCommandBufferAppendMemBufferFillExp(

// the same issue as in urCommandBufferAppendKernelLaunchExp
auto commandListLocked = hCommandBuffer->commandListManager.lock();
UR_CALL(commandListLocked->appendMemBufferFill(
UR_CALL(commandListLocked->enqueueMemBufferFill(
hBuffer, pPattern, patternSize, offset, size, 0, nullptr, nullptr));
return UR_RESULT_SUCCESS;
} catch (...) {
Expand All @@ -475,8 +474,8 @@ ur_result_t urCommandBufferAppendUSMPrefetchExp(
// the same issue as in urCommandBufferAppendKernelLaunchExp

auto commandListLocked = hCommandBuffer->commandListManager.lock();
UR_CALL(commandListLocked->appendUSMPrefetch(pMemory, size, flags, 0, nullptr,
nullptr));
UR_CALL(commandListLocked->enqueueUSMPrefetch(pMemory, size, flags, 0,
nullptr, nullptr));

return UR_RESULT_SUCCESS;
} catch (...) {
Expand All @@ -496,7 +495,7 @@ ur_result_t urCommandBufferAppendUSMAdviseExp(
// the same issue as in urCommandBufferAppendKernelLaunchExp

auto commandListLocked = hCommandBuffer->commandListManager.lock();
UR_CALL(commandListLocked->appendUSMAdvise(pMemory, size, advice, nullptr));
UR_CALL(commandListLocked->enqueueUSMAdvise(pMemory, size, advice, nullptr));

return UR_RESULT_SUCCESS;
} catch (...) {
Expand Down Expand Up @@ -546,13 +545,13 @@ ur_result_t urCommandBufferAppendNativeCommandExp(
// Barrier on all commands before user defined commands.

auto commandListLocked = hCommandBuffer->commandListManager.lock();
UR_CALL(commandListLocked->appendBarrier(0, nullptr, nullptr));
UR_CALL(commandListLocked->enqueueEventsWaitWithBarrier(0, nullptr, nullptr));

// Call user-defined function immediately
pfnNativeCommand(pData);

// Barrier on all commands after user defined commands.
UR_CALL(commandListLocked->appendBarrier(0, nullptr, nullptr));
UR_CALL(commandListLocked->enqueueEventsWaitWithBarrier(0, nullptr, nullptr));

return UR_RESULT_SUCCESS;
}
Expand Down
Loading
Loading