Skip to content

Commit 4a89e1c

Browse files
igchorigchor
committed
[Spec] fix urKernelSuggestMaxCooperativeGroupCountExp
Add extra param: ur_device_handle_t It is necessary to implement this function on L0 for kernels that are build for multiple devices. Right now, the implementation only works when the kernel is created from a native handle. Ref: #2262
1 parent b5f4a93 commit 4a89e1c

File tree

16 files changed

+69
-17
lines changed

16 files changed

+69
-17
lines changed

‎include/ur_api.h

Lines changed: 3 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -9536,13 +9536,15 @@ urEnqueueCooperativeKernelLaunchExp(
95369536
/// - ::UR_RESULT_ERROR_ADAPTER_SPECIFIC
95379537
/// - ::UR_RESULT_ERROR_INVALID_NULL_HANDLE
95389538
/// + `NULL == hKernel`
9539+
/// + `NULL == hDevice`
95399540
/// - ::UR_RESULT_ERROR_INVALID_NULL_POINTER
95409541
/// + `NULL == pLocalWorkSize`
95419542
/// + `NULL == pGroupCountRet`
95429543
/// - ::UR_RESULT_ERROR_INVALID_KERNEL
95439544
UR_APIEXPORT ur_result_t UR_APICALL
95449545
urKernelSuggestMaxCooperativeGroupCountExp(
95459546
ur_kernel_handle_t hKernel, ///< [in] handle of the kernel object
9547+
ur_device_handle_t hDevice, ///< [in] handle of the device object
95469548
uint32_t workDim, ///< [in] number of dimensions, from 1 to 3, to specify the work-group
95479549
///< work-items
95489550
const size_t *pLocalWorkSize, ///< [in] pointer to an array of workDim unsigned values that specify the
@@ -11083,6 +11085,7 @@ typedef struct ur_kernel_set_specialization_constants_params_t {
1108311085
/// allowing the callback the ability to modify the parameter's value
1108411086
typedef struct ur_kernel_suggest_max_cooperative_group_count_exp_params_t {
1108511087
ur_kernel_handle_t *phKernel;
11088+
ur_device_handle_t *phDevice;
1108611089
uint32_t *pworkDim;
1108711090
const size_t **ppLocalWorkSize;
1108811091
size_t *pdynamicSharedMemorySize;

‎include/ur_ddi.h

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -651,6 +651,7 @@ typedef ur_result_t(UR_APICALL *ur_pfnGetKernelProcAddrTable_t)(
651651
/// @brief Function-pointer for urKernelSuggestMaxCooperativeGroupCountExp
652652
typedef ur_result_t(UR_APICALL *ur_pfnKernelSuggestMaxCooperativeGroupCountExp_t)(
653653
ur_kernel_handle_t,
654+
ur_device_handle_t,
654655
uint32_t,
655656
const size_t *,
656657
size_t,

‎include/ur_print.hpp

Lines changed: 6 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -13187,6 +13187,12 @@ inline std::ostream &operator<<(std::ostream &os, [[maybe_unused]] const struct
1318713187
ur::details::printPtr(os,
1318813188
*(params->phKernel));
1318913189

13190+
os << ", ";
13191+
os << ".hDevice = ";
13192+
13193+
ur::details::printPtr(os,
13194+
*(params->phDevice));
13195+
1319013196
os << ", ";
1319113197
os << ".workDim = ";
1319213198

‎scripts/core/exp-cooperative-kernels.yml

Lines changed: 3 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -78,6 +78,9 @@ params:
7878
- type: $x_kernel_handle_t
7979
name: hKernel
8080
desc: "[in] handle of the kernel object"
81+
- type: $x_device_handle_t
82+
name: hDevice
83+
desc: "[in] handle of the device object"
8184
- type: uint32_t
8285
name: workDim
8386
desc: "[in] number of dimensions, from 1 to 3, to specify the work-group work-items"

‎source/adapters/cuda/kernel.cpp

Lines changed: 5 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -190,10 +190,13 @@ UR_APIEXPORT ur_result_t UR_APICALL urKernelGetNativeHandle(
190190
}
191191

192192
UR_APIEXPORT ur_result_t UR_APICALL urKernelSuggestMaxCooperativeGroupCountExp(
193-
ur_kernel_handle_t hKernel, uint32_t workDim, const size_t *pLocalWorkSize,
194-
size_t dynamicSharedMemorySize, uint32_t *pGroupCountRet) {
193+
ur_kernel_handle_t hKernel, ur_device_handle_t hDevice, uint32_t workDim,
194+
const size_t *pLocalWorkSize, size_t dynamicSharedMemorySize,
195+
uint32_t *pGroupCountRet) {
195196
UR_ASSERT(hKernel, UR_RESULT_ERROR_INVALID_KERNEL);
196197

198+
std::ignore = hDevice;
199+
197200
size_t localWorkSize = pLocalWorkSize[0];
198201
localWorkSize *= (workDim >= 2 ? pLocalWorkSize[1] : 1);
199202
localWorkSize *= (workDim == 3 ? pLocalWorkSize[2] : 1);

‎source/adapters/hip/kernel.cpp

Lines changed: 4 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -169,9 +169,11 @@ urKernelGetNativeHandle(ur_kernel_handle_t, ur_native_handle_t *) {
169169
}
170170

171171
UR_APIEXPORT ur_result_t UR_APICALL urKernelSuggestMaxCooperativeGroupCountExp(
172-
ur_kernel_handle_t hKernel, uint32_t workDim, const size_t *pLocalWorkSize,
173-
size_t dynamicSharedMemorySize, uint32_t *pGroupCountRet) {
172+
ur_kernel_handle_t hKernel, ur_device_handle_t hDevice, uint32_t workDim,
173+
const size_t *pLocalWorkSize, size_t dynamicSharedMemorySize,
174+
uint32_t *pGroupCountRet) {
174175
std::ignore = hKernel;
176+
std::ignore = hDevice;
175177
std::ignore = workDim;
176178
std::ignore = pLocalWorkSize;
177179
std::ignore = dynamicSharedMemorySize;

‎source/adapters/level_zero/kernel.cpp

Lines changed: 6 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -1054,8 +1054,9 @@ ur_result_t urKernelGetNativeHandle(
10541054
}
10551055

10561056
ur_result_t urKernelSuggestMaxCooperativeGroupCountExp(
1057-
ur_kernel_handle_t hKernel, uint32_t workDim, const size_t *pLocalWorkSize,
1058-
size_t dynamicSharedMemorySize, uint32_t *pGroupCountRet) {
1057+
ur_kernel_handle_t hKernel, ur_device_handle_t hDevice, uint32_t workDim,
1058+
const size_t *pLocalWorkSize, size_t dynamicSharedMemorySize,
1059+
uint32_t *pGroupCountRet) {
10591060
(void)dynamicSharedMemorySize;
10601061
std::shared_lock<ur_shared_mutex> Guard(hKernel->Mutex);
10611062

@@ -1066,8 +1067,10 @@ ur_result_t urKernelSuggestMaxCooperativeGroupCountExp(
10661067
ZE2UR_CALL(zeKernelSetGroupSize, (hKernel->ZeKernel, WG[0], WG[1], WG[2]));
10671068

10681069
uint32_t TotalGroupCount = 0;
1070+
ze_kernel_handle_t ZeKernel;
1071+
UR_CALL(getZeKernel(hDevice->ZeDevice, hKernel, &ZeKernel));
10691072
ZE2UR_CALL(zeKernelSuggestMaxCooperativeGroupCount,
1070-
(hKernel->ZeKernel, &TotalGroupCount));
1073+
(ZeKernel, &TotalGroupCount));
10711074
*pGroupCountRet = TotalGroupCount;
10721075
return UR_RESULT_SUCCESS;
10731076
}

‎source/adapters/level_zero/ur_interface_loader.hpp

Lines changed: 3 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -691,8 +691,9 @@ ur_result_t urEnqueueCooperativeKernelLaunchExp(
691691
const size_t *pLocalWorkSize, uint32_t numEventsInWaitList,
692692
const ur_event_handle_t *phEventWaitList, ur_event_handle_t *phEvent);
693693
ur_result_t urKernelSuggestMaxCooperativeGroupCountExp(
694-
ur_kernel_handle_t hKernel, uint32_t workDim, const size_t *pLocalWorkSize,
695-
size_t dynamicSharedMemorySize, uint32_t *pGroupCountRet);
694+
ur_kernel_handle_t hKernel, ur_device_handle_t hDevice, uint32_t workDim,
695+
const size_t *pLocalWorkSize, size_t dynamicSharedMemorySize,
696+
uint32_t *pGroupCountRet);
696697
ur_result_t urEnqueueTimestampRecordingExp(
697698
ur_queue_handle_t hQueue, bool blocking, uint32_t numEventsInWaitList,
698699
const ur_event_handle_t *phEventWaitList, ur_event_handle_t *phEvent);

‎source/adapters/level_zero/v2/api.cpp

Lines changed: 3 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -560,8 +560,9 @@ ur_result_t urCommandBufferCommandGetInfoExp(
560560
}
561561

562562
ur_result_t urKernelSuggestMaxCooperativeGroupCountExp(
563-
ur_kernel_handle_t hKernel, uint32_t workDim, const size_t *pLocalWorkSize,
564-
size_t dynamicSharedMemorySize, uint32_t *pGroupCountRet) {
563+
ur_kernel_handle_t hKernel, ur_device_handle_t hDevice, uint32_t workDim,
564+
const size_t *pLocalWorkSize, size_t dynamicSharedMemorySize,
565+
uint32_t *pGroupCountRet) {
565566
logger::error("{} function not implemented!", __FUNCTION__);
566567
return UR_RESULT_ERROR_UNSUPPORTED_FEATURE;
567568
}

‎source/adapters/mock/ur_mockddi.cpp

Lines changed: 6 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -10057,6 +10057,7 @@ __urdlllocal ur_result_t UR_APICALL urEnqueueCooperativeKernelLaunchExp(
1005710057
/// @brief Intercept function for urKernelSuggestMaxCooperativeGroupCountExp
1005810058
__urdlllocal ur_result_t UR_APICALL urKernelSuggestMaxCooperativeGroupCountExp(
1005910059
ur_kernel_handle_t hKernel, ///< [in] handle of the kernel object
10060+
ur_device_handle_t hDevice, ///< [in] handle of the device object
1006010061
uint32_t
1006110062
workDim, ///< [in] number of dimensions, from 1 to 3, to specify the work-group
1006210063
///< work-items
@@ -10072,7 +10073,11 @@ __urdlllocal ur_result_t UR_APICALL urKernelSuggestMaxCooperativeGroupCountExp(
1007210073
ur_result_t result = UR_RESULT_SUCCESS;
1007310074

1007410075
ur_kernel_suggest_max_cooperative_group_count_exp_params_t params = {
10075-
&hKernel, &workDim, &pLocalWorkSize, &dynamicSharedMemorySize,
10076+
&hKernel,
10077+
&hDevice,
10078+
&workDim,
10079+
&pLocalWorkSize,
10080+
&dynamicSharedMemorySize,
1007610081
&pGroupCountRet};
1007710082

1007810083
auto beforeCallback = reinterpret_cast<ur_mock_callback_t>(

‎source/adapters/opencl/kernel.cpp

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -390,6 +390,7 @@ UR_APIEXPORT ur_result_t UR_APICALL urKernelGetNativeHandle(
390390

391391
UR_APIEXPORT ur_result_t UR_APICALL urKernelSuggestMaxCooperativeGroupCountExp(
392392
[[maybe_unused]] ur_kernel_handle_t hKernel,
393+
[[maybe_unused]] ur_device_handle_t hDevice,
393394
[[maybe_unused]] uint32_t workDim,
394395
[[maybe_unused]] const size_t *pLocalWorkSize,
395396
[[maybe_unused]] size_t dynamicSharedMemorySize,

‎source/loader/layers/tracing/ur_trcddi.cpp

Lines changed: 7 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -8633,6 +8633,7 @@ __urdlllocal ur_result_t UR_APICALL urEnqueueCooperativeKernelLaunchExp(
86338633
/// @brief Intercept function for urKernelSuggestMaxCooperativeGroupCountExp
86348634
__urdlllocal ur_result_t UR_APICALL urKernelSuggestMaxCooperativeGroupCountExp(
86358635
ur_kernel_handle_t hKernel, ///< [in] handle of the kernel object
8636+
ur_device_handle_t hDevice, ///< [in] handle of the device object
86368637
uint32_t
86378638
workDim, ///< [in] number of dimensions, from 1 to 3, to specify the work-group
86388639
///< work-items
@@ -8654,7 +8655,11 @@ __urdlllocal ur_result_t UR_APICALL urKernelSuggestMaxCooperativeGroupCountExp(
86548655
}
86558656

86568657
ur_kernel_suggest_max_cooperative_group_count_exp_params_t params = {
8657-
&hKernel, &workDim, &pLocalWorkSize, &dynamicSharedMemorySize,
8658+
&hKernel,
8659+
&hDevice,
8660+
&workDim,
8661+
&pLocalWorkSize,
8662+
&dynamicSharedMemorySize,
86588663
&pGroupCountRet};
86598664
uint64_t instance = getContext()->notify_begin(
86608665
UR_FUNCTION_KERNEL_SUGGEST_MAX_COOPERATIVE_GROUP_COUNT_EXP,
@@ -8664,7 +8669,7 @@ __urdlllocal ur_result_t UR_APICALL urKernelSuggestMaxCooperativeGroupCountExp(
86648669
logger.info(" ---> urKernelSuggestMaxCooperativeGroupCountExp\n");
86658670

86668671
ur_result_t result = pfnSuggestMaxCooperativeGroupCountExp(
8667-
hKernel, workDim, pLocalWorkSize, dynamicSharedMemorySize,
8672+
hKernel, hDevice, workDim, pLocalWorkSize, dynamicSharedMemorySize,
86688673
pGroupCountRet);
86698674

86708675
getContext()->notify_end(

‎source/loader/layers/validation/ur_valddi.cpp

Lines changed: 11 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -9656,6 +9656,7 @@ __urdlllocal ur_result_t UR_APICALL urEnqueueCooperativeKernelLaunchExp(
96569656
/// @brief Intercept function for urKernelSuggestMaxCooperativeGroupCountExp
96579657
__urdlllocal ur_result_t UR_APICALL urKernelSuggestMaxCooperativeGroupCountExp(
96589658
ur_kernel_handle_t hKernel, ///< [in] handle of the kernel object
9659+
ur_device_handle_t hDevice, ///< [in] handle of the device object
96599660
uint32_t
96609661
workDim, ///< [in] number of dimensions, from 1 to 3, to specify the work-group
96619662
///< work-items
@@ -9681,6 +9682,10 @@ __urdlllocal ur_result_t UR_APICALL urKernelSuggestMaxCooperativeGroupCountExp(
96819682
return UR_RESULT_ERROR_INVALID_NULL_HANDLE;
96829683
}
96839684

9685+
if (NULL == hDevice) {
9686+
return UR_RESULT_ERROR_INVALID_NULL_HANDLE;
9687+
}
9688+
96849689
if (NULL == pLocalWorkSize) {
96859690
return UR_RESULT_ERROR_INVALID_NULL_POINTER;
96869691
}
@@ -9695,8 +9700,13 @@ __urdlllocal ur_result_t UR_APICALL urKernelSuggestMaxCooperativeGroupCountExp(
96959700
getContext()->refCountContext->logInvalidReference(hKernel);
96969701
}
96979702

9703+
if (getContext()->enableLifetimeValidation &&
9704+
!getContext()->refCountContext->isReferenceValid(hDevice)) {
9705+
getContext()->refCountContext->logInvalidReference(hDevice);
9706+
}
9707+
96989708
ur_result_t result = pfnSuggestMaxCooperativeGroupCountExp(
9699-
hKernel, workDim, pLocalWorkSize, dynamicSharedMemorySize,
9709+
hKernel, hDevice, workDim, pLocalWorkSize, dynamicSharedMemorySize,
97009710
pGroupCountRet);
97019711

97029712
return result;

‎source/loader/ur_ldrddi.cpp

Lines changed: 5 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -8844,6 +8844,7 @@ __urdlllocal ur_result_t UR_APICALL urEnqueueCooperativeKernelLaunchExp(
88448844
/// @brief Intercept function for urKernelSuggestMaxCooperativeGroupCountExp
88458845
__urdlllocal ur_result_t UR_APICALL urKernelSuggestMaxCooperativeGroupCountExp(
88468846
ur_kernel_handle_t hKernel, ///< [in] handle of the kernel object
8847+
ur_device_handle_t hDevice, ///< [in] handle of the device object
88478848
uint32_t
88488849
workDim, ///< [in] number of dimensions, from 1 to 3, to specify the work-group
88498850
///< work-items
@@ -8871,9 +8872,12 @@ __urdlllocal ur_result_t UR_APICALL urKernelSuggestMaxCooperativeGroupCountExp(
88718872
// convert loader handle to platform handle
88728873
hKernel = reinterpret_cast<ur_kernel_object_t *>(hKernel)->handle;
88738874

8875+
// convert loader handle to platform handle
8876+
hDevice = reinterpret_cast<ur_device_object_t *>(hDevice)->handle;
8877+
88748878
// forward to device-platform
88758879
result = pfnSuggestMaxCooperativeGroupCountExp(
8876-
hKernel, workDim, pLocalWorkSize, dynamicSharedMemorySize,
8880+
hKernel, hDevice, workDim, pLocalWorkSize, dynamicSharedMemorySize,
88778881
pGroupCountRet);
88788882

88798883
return result;

‎source/loader/ur_libapi.cpp

Lines changed: 3 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -8935,12 +8935,14 @@ ur_result_t UR_APICALL urEnqueueCooperativeKernelLaunchExp(
89358935
/// - ::UR_RESULT_ERROR_ADAPTER_SPECIFIC
89368936
/// - ::UR_RESULT_ERROR_INVALID_NULL_HANDLE
89378937
/// + `NULL == hKernel`
8938+
/// + `NULL == hDevice`
89388939
/// - ::UR_RESULT_ERROR_INVALID_NULL_POINTER
89398940
/// + `NULL == pLocalWorkSize`
89408941
/// + `NULL == pGroupCountRet`
89418942
/// - ::UR_RESULT_ERROR_INVALID_KERNEL
89428943
ur_result_t UR_APICALL urKernelSuggestMaxCooperativeGroupCountExp(
89438944
ur_kernel_handle_t hKernel, ///< [in] handle of the kernel object
8945+
ur_device_handle_t hDevice, ///< [in] handle of the device object
89448946
uint32_t
89458947
workDim, ///< [in] number of dimensions, from 1 to 3, to specify the work-group
89468948
///< work-items
@@ -8961,7 +8963,7 @@ ur_result_t UR_APICALL urKernelSuggestMaxCooperativeGroupCountExp(
89618963
}
89628964

89638965
return pfnSuggestMaxCooperativeGroupCountExp(
8964-
hKernel, workDim, pLocalWorkSize, dynamicSharedMemorySize,
8966+
hKernel, hDevice, workDim, pLocalWorkSize, dynamicSharedMemorySize,
89658967
pGroupCountRet);
89668968
} catch (...) {
89678969
return exceptionToResult(std::current_exception());

‎source/ur_api.cpp

Lines changed: 2 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -7578,12 +7578,14 @@ ur_result_t UR_APICALL urEnqueueCooperativeKernelLaunchExp(
75787578
/// - ::UR_RESULT_ERROR_ADAPTER_SPECIFIC
75797579
/// - ::UR_RESULT_ERROR_INVALID_NULL_HANDLE
75807580
/// + `NULL == hKernel`
7581+
/// + `NULL == hDevice`
75817582
/// - ::UR_RESULT_ERROR_INVALID_NULL_POINTER
75827583
/// + `NULL == pLocalWorkSize`
75837584
/// + `NULL == pGroupCountRet`
75847585
/// - ::UR_RESULT_ERROR_INVALID_KERNEL
75857586
ur_result_t UR_APICALL urKernelSuggestMaxCooperativeGroupCountExp(
75867587
ur_kernel_handle_t hKernel, ///< [in] handle of the kernel object
7588+
ur_device_handle_t hDevice, ///< [in] handle of the device object
75877589
uint32_t
75887590
workDim, ///< [in] number of dimensions, from 1 to 3, to specify the work-group
75897591
///< work-items

0 commit comments

Comments
 (0)