[SYCL][E2E] bindless image - switch tests to immediate command list, etc (#22046)

Bindless image semaphores require immediate command lists. This updates
the docs and the tests to be aligned with this requirement. Updating the
tests with this change means that certain tests that used to fail now
pass, so they are are being enabled. Also, some of the failing tests
pass with the correct graphic driver, so that expectation is now more
explicit in the tests. And, lastly, a few of the tests tripped the
validator because they can forget to free a resource if an exception is
encountered - fixed that too.

---------

Signed-off-by: Chris Perkins <chris.perkins@intel.com>

Author: cperkinsintel

sycl/doc/extensions/experimental/sycl_ext_oneapi_bindless_images.asciidoc

@@ -2259,13 +2259,18 @@ memory resources handles can take different forms of structure and type
 depending on the API and operating system, so do external semaphore resource 
 handles.
 
-It is important to note, that the use of imported external semaphore objects
-within SYCL has the restriction in that imported external semaphores can only
-be used in conjuction with SYCL queues that have been constructed with the
-`property::queue::in_order` property. The semaphore synchronization mechanism
-is not supported for the default SYCL out-of-order queues. Use of the semaphore 
-synchronization mechanism with SYCL queues which were not constructed with the 
-`queue::in_order` property will result in undefined behaviour.
+It is important to note that the use of imported external semaphore objects
+within SYCL requires the SYCL queue to have been constructed with both of the
+following properties:
+
+* `sycl::property::queue::in_order` -- the semaphore synchronization mechanism
+  is not supported on the default out-of-order queues.
+* `sycl::ext::intel::property::queue::immediate_command_list` -- external
+  semaphore operations are only supported on queues backed by immediate
+  command lists. This restriction might be lifted in the future.
+
+Use of the semaphore synchronization mechanism with a SYCL queue that was not
+constructed with both of these properties will result in undefined behaviour.
 
 External semaphore import is facilitated through the following proposed 
 descriptor struct.

sycl/test-e2e/bindless_images/dx11_interop/read_write_unsampled.cpp

@@ -10,6 +10,7 @@
 #include "dx11_interop.h"
 
 #include <sycl/ext/oneapi/bindless_images.hpp>
+#include <sycl/properties/queue_properties.hpp>
 
 #ifdef TEST_SEMAPHORE_IMPORT
 #include <d3d11_4.h> // Used for ID3D11Device5 / ID3D11DeviceContext4 / ID3D11Fence
@@ -435,8 +436,16 @@ int runTest(D3D11ProgramState &d3d11ProgramState, sycl::queue syclQueue,
 }
 
 int main() {
-  // Create SYCL queue, relying on SYCL device selection
+  // Create SYCL queue, relying on SYCL device selection.
+#ifdef TEST_SEMAPHORE_IMPORT
+  // External semaphore ops require an in-order queue backed by immediate
+  // command lists (see sycl_ext_oneapi_bindless_images.asciidoc).
+  sycl::queue syclQueue{
+      {sycl::property::queue::in_order{},
+       sycl::ext::intel::property::queue::immediate_command_list{}}};
+#else
   sycl::queue syclQueue;
+#endif
   sycl::device syclDevice = syclQueue.get_device();
 
   // Initialize D3D11 and create DX11 programs state from the SYCL device

sycl/test-e2e/bindless_images/dx12_interop/D3D12_sycl_buffer_timeline_semaphore.cpp

@@ -2,18 +2,7 @@
 // REQUIRES: aspect-ext_oneapi_external_semaphore_import
 // REQUIRES: windows
 
-// UNSUPPORTED: gpu-intel-dg2
-// UNSUPPORTED-TRACKER: GSD-12428
-// semaphores-do-not-work-dg2
-
-// UNSUPPORTED: gpu-intel-gen12
-// UNSUPPORTED-TRACKER: GSD-12427
-// Gen12-semaphores-work-but-this-test-hangs.
-
-// UNSUPPORTED: arch-intel_gpu_bmg_g21
-// UNSUPPORTED-TRACKER: GSD-12436
-// this test works on BMG, but if run in parallel with itself, or with  other
-// semaphore tests, it can hang.
+// REQUIRES-INTEL-DRIVER: lin: 38303 win: 101.9999
 
 // RUN: %{build} %link-directx -o %t.exe %if target-spir %{ -Wno-ignored-attributes %}
 // RUN: %{run} %t.exe --no-sem
@@ -50,6 +39,7 @@
 #include <string>
 #include <sycl/detail/core.hpp>
 #include <sycl/ext/oneapi/bindless_images.hpp>
+#include <sycl/properties/queue_properties.hpp>
 #include <vector>
 
 #define WIN32_LEAN_AND_MEAN
@@ -120,7 +110,15 @@ int main(int argc, char **argv) {
 
   // SYCL INTEROP
   try {
-    sycl::queue q;
+    // Bindless image interop requires an in-order queue (per spec). External
+    // semaphore ops additionally require immediate command lists; see
+    // sycl_ext_oneapi_bindless_images.asciidoc.
+    sycl::property_list qProps =
+        useSemaphores ? sycl::property_list{sycl::property::queue::in_order{},
+                                            sycl::ext::intel::property::queue::
+                                                immediate_command_list{}}
+                      : sycl::property_list{sycl::property::queue::in_order{}};
+    sycl::queue q{qProps};
     auto device = q.get_device();
     auto context = q.get_context();
 

sycl/test-e2e/bindless_images/dx12_interop/D3D12_sycl_interop_1D_read.cpp

@@ -74,7 +74,10 @@
 
 
 // Semaphore coverage tests
-// At this time, semaphores aren't working on DG2 (GSD-12428), and can hang on BMG if run in parallel (GSD-12436).
+// On Windows, we require driver 38303 or later to avoid semaphore issues, which the CI does not yet have. 
+// Rather than mark the WHOLE test as requiring 38303, which would mean no testing nowhere,
+// we are limiting it with R U N - I F 
+
 // RUN-IF: (!gpu-intel-dg2 && !arch-intel_gpu_bmg_g21), %{run} %t.exe --type float --channels 4 --semaphores 33x
 // RUN-IF: (!gpu-intel-dg2 && !arch-intel_gpu_bmg_g21), %{run} %t.exe --type half --channels 2 --semaphores 33x
 // RUN-IF: (!gpu-intel-dg2 && !arch-intel_gpu_bmg_g21), %{run} %t.exe --type int32 --channels 1 --semaphores 33x
@@ -98,6 +101,7 @@
 #include <sycl/detail/core.hpp>
 #include <sycl/ext/oneapi/bindless_images.hpp>
 #include <sycl/ext/oneapi/bindless_images_interop.hpp>
+#include <sycl/properties/queue_properties.hpp>
 
 namespace syclexp = sycl::ext::oneapi::experimental;
 
@@ -187,7 +191,15 @@ int runTest(
 
   // SYCL Import and Verification
   try {
-    sycl::queue q;
+    // Bindless image interop requires an in-order queue (per spec). External
+    // semaphore ops additionally require immediate command lists; see
+    // sycl_ext_oneapi_bindless_images.asciidoc.
+    sycl::property_list qProps =
+        useSemaphores ? sycl::property_list{sycl::property::queue::in_order{},
+                                            sycl::ext::intel::property::queue::
+                                                immediate_command_list{}}
+                      : sycl::property_list{sycl::property::queue::in_order{}};
+    sycl::queue q{qProps};
 
     syclexp::external_mem_descriptor<syclexp::resource_win32_handle> extMemDesc{
         imgRes.sharedHandle, syclexp::external_mem_handle_type::win32_nt_handle,

sycl/test-e2e/bindless_images/dx12_interop/D3D12_sycl_interop_1D_write_unsampled.cpp

@@ -7,7 +7,7 @@
 
 // clang-format off
 /*
-    clang++.exe -fsycl -o ds1w.exe D3D12_sycl_interop_1D_write.cpp -ld3d12 -ldxgi -ld3dcompiler
+    clang++.exe -fsycl -o ds1w.exe D3D12_sycl_interop_1D_write_unsampled.cpp -ld3d12 -ldxgi -ld3dcompiler
 
     FLAGS:
     --sampled      ERROR: Sampled image writes are not supported
@@ -42,12 +42,12 @@
 // RUN: %{run} %t.exe --type int8 --channels 1 33x
 // RUN: %{run} %t.exe --type int8 --channels 2 33x
 // RUN: %{run} %t.exe --type int8 --channels 4 33x
-// RUN: %{run} %t.exe --type unorm8 --channels 1 33x
-// RUN: %{run} %t.exe --type unorm8 --channels 2 33x
-// RUN: %{run} %t.exe --type unorm8 --channels 4 33x
 
 // Semaphore coverage tests
-// At this time, semaphores aren't working on DG2 (GSD-12428), and can hang on BMG if run in parallel (GSD-12436).
+// On Windows, we require driver 38303 or later to avoid semaphore issues, which the CI does not yet have. 
+// Rather than mark the WHOLE test as requiring 38303, which would mean no testing nowhere,
+// we are limiting it with R U N - I F 
+
 // RUN-IF: (!gpu-intel-dg2 && !arch-intel_gpu_bmg_g21), %{run} %t.exe --type float --channels 4 --semaphores 33x
 // RUN-IF: (!gpu-intel-dg2 && !arch-intel_gpu_bmg_g21), %{run} %t.exe --type float --channels 1 --semaphores 33x
 // RUN-IF: (!gpu-intel-dg2 && !arch-intel_gpu_bmg_g21), %{run} %t.exe --type half --channels 2 --semaphores 33x
@@ -57,7 +57,6 @@
 // RUN-IF: (!gpu-intel-dg2 && !arch-intel_gpu_bmg_g21), %{run} %t.exe --type uint16 --channels 4 --semaphores 33x
 // RUN-IF: (!gpu-intel-dg2 && !arch-intel_gpu_bmg_g21), %{run} %t.exe --type uint8 --channels 1 --semaphores 33x
 // RUN-IF: (!gpu-intel-dg2 && !arch-intel_gpu_bmg_g21), %{run} %t.exe --type int8 --channels 2 --semaphores 33x
-// RUN-IF: (!gpu-intel-dg2 && !arch-intel_gpu_bmg_g21), %{run} %t.exe --type unorm8 --channels 4 --semaphores 33x
 // RUN-IF: (!gpu-intel-dg2 && !arch-intel_gpu_bmg_g21), %{run} %t.exe --type half --channels 4 --semaphores 33x
 // RUN-IF: (!gpu-intel-dg2 && !arch-intel_gpu_bmg_g21), %{run} %t.exe --type uint32 --channels 2 --semaphores 33x
 
@@ -71,6 +70,7 @@
 #include <sycl/detail/core.hpp>
 #include <sycl/ext/oneapi/bindless_images.hpp>
 #include <sycl/ext/oneapi/bindless_images_interop.hpp>
+#include <sycl/properties/queue_properties.hpp>
 
 namespace syclexp = sycl::ext::oneapi::experimental;
 
@@ -148,7 +148,15 @@ int runTest(
   }
 
   try {
-    sycl::queue q;
+    // Bindless image interop requires an in-order queue (per spec). External
+    // semaphore ops additionally require immediate command lists; see
+    // sycl_ext_oneapi_bindless_images.asciidoc.
+    sycl::property_list qProps =
+        useSemaphores ? sycl::property_list{sycl::property::queue::in_order{},
+                                            sycl::ext::intel::property::queue::
+                                                immediate_command_list{}}
+                      : sycl::property_list{sycl::property::queue::in_order{}};
+    sycl::queue q{qProps};
 
     syclexp::external_mem_descriptor<syclexp::resource_win32_handle> extMemDesc{
         imgRes.sharedHandle, syclexp::external_mem_handle_type::win32_nt_handle,

sycl/test-e2e/bindless_images/dx12_interop/D3D12_sycl_interop_2D_arithmetic.cpp

@@ -2,10 +2,6 @@
 // REQUIRES: aspect-ext_oneapi_external_memory_import
 // REQUIRES: windows
 
-// UNSUPPORTED: arch-intel_gpu_bmg_g21
-// UNSUPPORTED-TRACKER: https://github.com/intel/llvm/issues/20384
-// also GSD-12429
-
 // RUN: %{build} -o %t.exe %link-directx
 // RUN: %{run} %t.exe --type float --channels 4 8x8
 
@@ -26,7 +22,6 @@
 
     DG2:
     - WORKS, including --sampled
-    - semaphores segfault
 
     DG2 $ sycl-ls
     [level_zero:gpu][level_zero:0] Intel(R) oneAPI Unified Runtime over
@@ -61,9 +56,6 @@
 // RUN: %{run} %t.exe --type int8 --channels 1 32x33
 // RUN: %{run} %t.exe --type int8 --channels 2 32x33
 // RUN: %{run} %t.exe --type int8 --channels 4 32x33
-// RUN-IF: !gpu-intel-bmg, %{run} %t.exe --type unorm8 --channels 1 32x33
-// RUN-IF: !gpu-intel-bmg, %{run} %t.exe --type unorm8 --channels 2 32x33
-// RUN-IF: !gpu-intel-bmg, %{run} %t.exe --type unorm8 --channels 4 32x33
 // RUN: %{run} %t.exe --type float --channels 1 --sampled 32x33
 // RUN: %{run} %t.exe --type float --channels 2 --sampled 32x33
 // RUN: %{run} %t.exe --type float --channels 4 --sampled 32x33
@@ -88,12 +80,12 @@
 // RUN: %{run} %t.exe --type int8 --channels 1 --sampled 32x33
 // RUN: %{run} %t.exe --type int8 --channels 2 --sampled 32x33
 // RUN: %{run} %t.exe --type int8 --channels 4 --sampled 32x33
-// RUN-IF: !gpu-intel-bmg, %{run} %t.exe --type unorm8 --channels 1 --sampled 32x33
-// RUN-IF: !gpu-intel-bmg, %{run} %t.exe --type unorm8 --channels 2 --sampled 32x33
-// RUN-IF: !gpu-intel-bmg, %{run} %t.exe --type unorm8 --channels 4 --sampled 32x33
 
 // Semaphore coverage tests
-// At this time, semaphores aren't working on DG2 (GSD-12428), and can hang on BMG if run in parallel (GSD-12436).
+// On Windows, we require driver 38303 or later to avoid semaphore issues, which the CI does not yet have. 
+// Rather than mark the WHOLE test as requiring 38303, which would mean no testing nowhere,
+// we are limiting it with R U N - I F 
+
 // RUN-IF: (!gpu-intel-dg2 && !arch-intel_gpu_bmg_g21), %{run} %t.exe --type float --channels 4 --semaphores 32x33
 // RUN-IF: (!gpu-intel-dg2 && !arch-intel_gpu_bmg_g21), %{run} %t.exe --type half --channels 2 --semaphores 32x33
 // RUN-IF: (!gpu-intel-dg2 && !arch-intel_gpu_bmg_g21), %{run} %t.exe --type int32 --channels 1 --semaphores 32x33
@@ -105,7 +97,6 @@
 // RUN-IF: (!gpu-intel-dg2 && !arch-intel_gpu_bmg_g21), %{run} %t.exe --type float --channels 4 --sampled --semaphores 32x33
 // RUN-IF: (!gpu-intel-dg2 && !arch-intel_gpu_bmg_g21), %{run} %t.exe --type half --channels 2 --sampled --semaphores 32x33
 // RUN-IF: (!gpu-intel-dg2 && !arch-intel_gpu_bmg_g21), %{run} %t.exe --type int32 --channels 1 --sampled --semaphores 32x33
-// RUN-IF: (!gpu-intel-dg2 && !arch-intel_gpu_bmg_g21), %{run} %t.exe --type unorm8 --channels 4 --sampled --semaphores 32x33
 
 // clang-format on
 
@@ -118,6 +109,7 @@
 #include <sycl/detail/core.hpp>
 #include <sycl/ext/oneapi/bindless_images.hpp>
 #include <sycl/ext/oneapi/bindless_images_interop.hpp>
+#include <sycl/properties/queue_properties.hpp>
 #include <vector>
 
 namespace syclexp = sycl::ext::oneapi::experimental;
@@ -415,7 +407,15 @@ int runTest(
     signalExportableFence(ctx, extFenceB);
 
   try {
-    sycl::queue q;
+    // Bindless image interop requires an in-order queue (per spec). External
+    // semaphore ops additionally require immediate command lists; see
+    // sycl_ext_oneapi_bindless_images.asciidoc.
+    sycl::property_list qProps =
+        useSemaphores ? sycl::property_list{sycl::property::queue::in_order{},
+                                            sycl::ext::intel::property::queue::
+                                                immediate_command_list{}}
+                      : sycl::property_list{sycl::property::queue::in_order{}};
+    sycl::queue q{qProps};
 
     auto extMemA = syclexp::import_external_memory(
         syclexp::external_mem_descriptor<syclexp::resource_win32_handle>{

sycl/test-e2e/bindless_images/dx12_interop/D3D12_sycl_interop_2D_read.cpp

@@ -2,10 +2,6 @@
 // REQUIRES: aspect-ext_oneapi_external_memory_import
 // REQUIRES: windows
 
-// UNSUPPORTED: arch-intel_gpu_bmg_g21
-// UNSUPPORTED-TRACKER: https://github.com/intel/llvm/issues/20384
-// also: GSD-12429
-
 // RUN: %{build} -o %t.exe %link-directx
 // RUN: %{run} %t.exe --type float --channels 4 32x33
 
@@ -23,9 +19,7 @@
     WxH            Set custom Width x Height (e.g. 8x4)
 
 
-    BMG:
-    - 1x1 works, nothing else. Suggesting offset/pitch issue.
-    - semaphores segfault. suggesting segfaulting semaphores.
+
 
     BMG $ sycl-ls
     [level_zero:gpu][level_zero:0] Intel(R) oneAPI Unified Runtime over
@@ -35,7 +29,6 @@
 
     DG2:
     - WORKS, including --sampled
-    - semaphores segfault
 
     DG2 $ sycl-ls
     [level_zero:gpu][level_zero:0] Intel(R) oneAPI Unified Runtime over
@@ -98,7 +91,10 @@
 // RUN: %{run} %t.exe --type int8 --channels 4 --sampled 32x33
 
 // Semaphore coverage tests
-// At this time, semaphores aren't working on DG2 (GSD-12428), and can hang on BMG if run in parallel (GSD-12436).
+// On Windows, we require driver 38303 or later to avoid semaphore issues, which the CI does not yet have. 
+// Rather than mark the WHOLE test as requiring 38303, which would mean no testing nowhere,
+// we are limiting it with R U N - I F 
+
 // RUN-IF: (!gpu-intel-dg2 && !arch-intel_gpu_bmg_g21), %{run} %t.exe --type float --channels 4 --semaphores 32x33
 // RUN-IF: (!gpu-intel-dg2 && !arch-intel_gpu_bmg_g21), %{run} %t.exe --type half --channels 2 --semaphores 32x33
 // RUN-IF: (!gpu-intel-dg2 && !arch-intel_gpu_bmg_g21), %{run} %t.exe --type int32 --channels 1 --semaphores 32x33
@@ -122,6 +118,7 @@
 #include <sycl/detail/core.hpp>
 #include <sycl/ext/oneapi/bindless_images.hpp>
 #include <sycl/ext/oneapi/bindless_images_interop.hpp>
+#include <sycl/properties/queue_properties.hpp>
 
 namespace syclexp = sycl::ext::oneapi::experimental;
 
@@ -208,7 +205,15 @@ int runTest(
 
   // SYCL Import and Verification
   try {
-    sycl::queue q;
+    // Bindless image interop requires an in-order queue (per spec). External
+    // semaphore ops additionally require immediate command lists; see
+    // sycl_ext_oneapi_bindless_images.asciidoc.
+    sycl::property_list qProps =
+        useSemaphores ? sycl::property_list{sycl::property::queue::in_order{},
+                                            sycl::ext::intel::property::queue::
+                                                immediate_command_list{}}
+                      : sycl::property_list{sycl::property::queue::in_order{}};
+    sycl::queue q{qProps};
 
     syclexp::external_mem_descriptor<syclexp::resource_win32_handle> extMemDesc{
         imgRes.sharedHandle, syclexp::external_mem_handle_type::win32_nt_handle,