Skip to content

Commit 006dc42

Browse files
author
Omar Ahmed
authored
[SYCL] Remove UR objects release checks on Windows (#15425)
These e2e tests started failing after PI removal and replacing it with UR ([PR](#14145)) However, they were not related to it. On Windows, dlls unloading is inconsistent and if we try to release these UR objects manually, inconsistent hangs happen due to a race between unloading the UR adapters dlls (in addition to their dependency dlls) and the releasing of these UR objects (The proxy loader have solved this problem partially [here](#15262)). So, we currently shutdown without releasing them and windows should handle the memory cleanup. This behaviour is the same old behaviour as before removing PI as on Investigations on this. This was only hidden before PI removal as it was calling the PI entry-point but doesn't make it to UR entry-point and Filecheck logs check for objects release would pass as it only check for the call to the PI entry-point without check that the call was a successful call. That was the PI call for `piContextRelease` before PI removal: ``` ---> piContextRelease( <unknown> : 0000023CC0EBA6C0 ) ---> API Called After Plugin Teardown, Functon Call ignored. ``` Fixes #14768 Fixes #14950 Fixes #14968
1 parent 91eaa05 commit 006dc42

File tree

5 files changed

+54
-31
lines changed

5 files changed

+54
-31
lines changed

‎sycl/test-e2e/Basic/queue/release.cpp‎

Lines changed: 13 additions & 7 deletions
Original file line numberDiff line numberDiff line change
@@ -1,8 +1,7 @@
11
// RUN: %{build} -o %t.out
2-
// RUN: env SYCL_UR_TRACE=2 %{run} %t.out | FileCheck %s
2+
// RUN: env SYCL_UR_TRACE=2 %{run} %t.out | FileCheck %s %if !windows %{--check-prefixes=CHECK-RELEASE%}
33
//
4-
// TODO: Reenable on Windows, see https://github.com/intel/llvm/issues/14768
5-
// XFAIL: hip_nvidia, windows
4+
// XFAIL: hip_nvidia
65

76
#include <sycl/detail/core.hpp>
87
int main() {
@@ -19,7 +18,14 @@ int main() {
1918
// specific queue workaround.
2019
// CHECK-DAG: <--- urEventRelease(
2120
// CHECK-DAG: <--- urQueueRelease(
22-
// CHECK: <--- urContextRelease(
23-
// CHECK: <--- urKernelRelease(
24-
// CHECK: <--- urProgramRelease(
25-
// CHECK: <--- urDeviceRelease(
21+
22+
// On Windows, dlls unloading is inconsistent and if we try to release these UR
23+
// objects manually, inconsistent hangs happen due to a race between unloading
24+
// the UR adapters dlls (in addition to their dependency dlls) and the releasing
25+
// of these UR objects. So, we currently shutdown without releasing them and
26+
// windows should handle the memory cleanup.
27+
28+
// CHECK-RELEASE: <--- urContextRelease(
29+
// CHECK-RELEASE: <--- urKernelRelease(
30+
// CHECK-RELEASE: <--- urProgramRelease(
31+
// CHECK-RELEASE: <--- urDeviceRelease(

‎sycl/test-e2e/KernelAndProgram/disable-caching.cpp‎

Lines changed: 13 additions & 10 deletions
Original file line numberDiff line numberDiff line change
@@ -5,10 +5,7 @@
55
// RUN: env ZE_DEBUG=-6 SYCL_UR_TRACE=2 SYCL_CACHE_IN_MEM=0 %{run} %t.out \
66
// RUN: | FileCheck %s
77
// RUN: env ZE_DEBUG=-6 SYCL_UR_TRACE=2 %{run} %t.out \
8-
// RUN: | FileCheck %s --check-prefixes=CHECK-CACHE
9-
10-
// TODO: Reenable on Windows, see https://github.com/intel/llvm/issues/14768
11-
// XFAIL: windows
8+
// RUN: | FileCheck %s --check-prefixes=CHECK-CACHE%if !windows %{,CHECK-RELEASE%}
129

1310
#include <sycl/detail/core.hpp>
1411

@@ -93,10 +90,16 @@ int main() {
9390
free(p, q);
9491
}
9592

93+
// On Windows, dlls unloading is inconsistent and if we try to release these UR
94+
// objects manually, inconsistent hangs happen due to a race between unloading
95+
// the UR adapters dlls (in addition to their dependency dlls) and the releasing
96+
// of these UR objects. So, we currently shutdown without releasing them and
97+
// windows should handle the memory cleanup.
98+
9699
// (Program cache releases)
97-
// CHECK-CACHE: <--- urKernelRelease
98-
// CHECK-CACHE: <--- urKernelRelease
99-
// CHECK-CACHE: <--- urKernelRelease
100-
// CHECK-CACHE: <--- urProgramRelease
101-
// CHECK-CACHE: <--- urProgramRelease
102-
// CHECK-CACHE: <--- urProgramRelease
100+
// CHECK-RELEASE: <--- urKernelRelease
101+
// CHECK-RELEASE: <--- urKernelRelease
102+
// CHECK-RELEASE: <--- urKernelRelease
103+
// CHECK-RELEASE: <--- urProgramRelease
104+
// CHECK-RELEASE: <--- urProgramRelease
105+
// CHECK-RELEASE: <--- urProgramRelease

‎sycl/test-e2e/Regression/context_is_destroyed_after_exception.cpp‎

Lines changed: 8 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -1,10 +1,9 @@
11
// REQUIRES: gpu
22

33
// RUN: %{build} -o %t.out
4-
// RUN: env SYCL_UR_TRACE=2 %{run} %t.out 2>&1 | FileCheck %s
4+
// RUN: env SYCL_UR_TRACE=2 %{run} %t.out %if !windows %{2>&1 | FileCheck %s %}
55
//
6-
// TODO: Reenable on Windows, see https://github.com/intel/llvm/issues/14768
7-
// XFAIL: hip_nvidia, windows
6+
// XFAIL: hip_nvidia
87

98
#include <sycl/detail/core.hpp>
109

@@ -32,4 +31,10 @@ int main() {
3231
return 0;
3332
}
3433

34+
// On Windows, dlls unloading is inconsistent and if we try to release these UR
35+
// objects manually, inconsistent hangs happen due to a race between unloading
36+
// the UR adapters dlls (in addition to their dependency dlls) and the releasing
37+
// of these UR objects. So, we currently shutdown without releasing them and
38+
// windows should handle the memory cleanup.
39+
3540
// CHECK: <--- urContextRelease(

‎sycl/test-e2e/Regression/pi_release.cpp‎

Lines changed: 7 additions & 4 deletions
Original file line numberDiff line numberDiff line change
@@ -1,9 +1,6 @@
11
// REQUIRES: opencl || level_zero || cuda
22
// RUN: %{build} -o %t.out
3-
// RUN: env SYCL_UR_TRACE=2 %{run} %t.out 2>&1 | FileCheck %s
4-
//
5-
// TODO: Reenable on Windows, see https://github.com/intel/llvm/issues/14768
6-
// XFAIL: windows
3+
// RUN: env SYCL_UR_TRACE=2 %{run} %t.out %if !windows %{2>&1 | FileCheck %s %}
74

85
#include <sycl/detail/core.hpp>
96

@@ -12,5 +9,11 @@ int main() {
129
return 0;
1310
}
1411

12+
// On Windows, dlls unloading is inconsistent and if we try to release these UR
13+
// objects manually, inconsistent hangs happen due to a race between unloading
14+
// the UR adapters dlls (in addition to their dependency dlls) and the releasing
15+
// of these UR objects. So, we currently shutdown without releasing them and
16+
// windows should handle the memory cleanup.
17+
1518
// CHECK: <--- urQueueRelease
1619
// CHECK: <--- urContextRelease

‎sycl/test-e2e/Scheduler/ReleaseResourcesTest.cpp‎

Lines changed: 13 additions & 7 deletions
Original file line numberDiff line numberDiff line change
@@ -1,8 +1,7 @@
11
// RUN: %{build} -Wno-error=unused-command-line-argument -fsycl-dead-args-optimization -o %t.out
2-
// RUN: env SYCL_UR_TRACE=2 %{run} %t.out 2>&1 | FileCheck %s
2+
// RUN: env SYCL_UR_TRACE=2 %{run} %t.out 2>&1 | FileCheck %s %if !windows %{--check-prefix=CHECK-RELEASE%}
33
//
4-
// TODO: Reenable on Windows, see https://github.com/intel/llvm/issues/14768
5-
// XFAIL: hip_nvidia, windows
4+
// XFAIL: hip_nvidia
65

76
//==------------------- ReleaseResourcesTests.cpp --------------------------==//
87
//
@@ -50,7 +49,14 @@ int main() {
5049
// CHECK: <--- urQueueCreate
5150
// CHECK: <--- urProgramCreate
5251
// CHECK: <--- urKernelCreate
53-
// CHECK: <--- urQueueRelease
54-
// CHECK: <--- urContextRelease
55-
// CHECK: <--- urKernelRelease
56-
// CHECK: <--- urProgramRelease
52+
53+
// On Windows, dlls unloading is inconsistent and if we try to release these UR
54+
// objects manually, inconsistent hangs happen due to a race between unloading
55+
// the UR adapters dlls (in addition to their dependency dlls) and the releasing
56+
// of these UR objects. So, we currently shutdown without releasing them and
57+
// windows should handle the memory cleanup.
58+
59+
// CHECK-RELEASE: <--- urQueueRelease
60+
// CHECK-RELEASE: <--- urContextRelease
61+
// CHECK-RELEASE: <--- urKernelRelease
62+
// CHECK-RELEASE: <--- urProgramRelease

0 commit comments

Comments
 (0)