Merge branch 'main' into yc-test-main/0120-on-shadow-per-type

Author: yingcong-wu

scripts/benchmarks/benches/compute.py

@@ -22,7 +22,7 @@ def setup(self):
         if options.sycl is None:
             return
 
-        repo_path = git_clone(self.directory, "compute-benchmarks-repo", "https://github.com/intel/compute-benchmarks.git", "d13e5b4d8dd3d28926a74ab7f67f78c10f708a01")
+        repo_path = git_clone(self.directory, "compute-benchmarks-repo", "https://github.com/intel/compute-benchmarks.git", "578a7ac6f9bc48f6c2b408ef64a19a2ef9a216e7")
         build_path = create_build_path(self.directory, 'compute-benchmarks-build')
 
         configure_command = [
@@ -74,6 +74,18 @@ def benchmarks(self) -> list[Benchmark]:
             MemcpyExecute(self, 10, 16, 1024, 10000, 0, 1, 1),
             MemcpyExecute(self, 4096, 1, 1024, 10, 0, 1, 0),
             MemcpyExecute(self, 4096, 4, 1024, 10, 0, 1, 0),
+            GraphApiSinKernelGraphSYCL(self, 0, 10),
+            GraphApiSinKernelGraphSYCL(self, 1, 10),
+            GraphApiSinKernelGraphSYCL(self, 0, 100),
+            GraphApiSinKernelGraphSYCL(self, 1, 100),
+            # Submit
+            GraphApiSubmitExecGraph(self, 0, 1, 10),
+            GraphApiSubmitExecGraph(self, 1, 1, 10),
+            GraphApiSubmitExecGraph(self, 1, 1, 100),
+            # Exec
+            GraphApiSubmitExecGraph(self, 0, 0, 10),
+            GraphApiSubmitExecGraph(self, 1, 0, 10),
+            GraphApiSubmitExecGraph(self, 1, 0, 100),
         ]
 
         if options.ur is not None:
@@ -357,3 +369,46 @@ def bin_args(self) -> list[str]:
             f"--SrcUSM={self.srcUSM}",
             f"--DstUSM={self.dstUSM}",
         ]
+
+class GraphApiSinKernelGraphSYCL(ComputeBenchmark):
+    def __init__(self, bench, withGraphs, numKernels):
+        self.withGraphs = withGraphs
+        self.numKernels = numKernels
+        super().__init__(bench, "graph_api_benchmark_sycl", "SinKernelGraph")
+
+    def explicit_group(self):
+        return "SinKernelGraph"
+
+    def name(self):
+        return f"graph_api_benchmark_sycl SinKernelGraph graphs:{self.withGraphs}, numKernels:{self.numKernels}"
+
+    def bin_args(self) -> list[str]:
+        return [
+            "--iterations=100",
+            f"--numKernels={self.numKernels}",
+            f"--withGraphs={self.withGraphs}",
+        ]
+        
+class GraphApiSubmitExecGraph(ComputeBenchmark):
+    def __init__(self, bench, ioq, submit, numKernels):
+        self.ioq = ioq
+        self.submit = submit
+        self.numKernels = numKernels
+        super().__init__(bench, "graph_api_benchmark_sycl", "SubmitExecGraph")
+
+    def name(self):
+        return f"graph_api_benchmark_sycl SubmitExecGraph ioq:{self.ioq}, submit:{self.submit}, numKernels:{self.numKernels}"
+
+    def explicit_group(self):
+        if self.submit:
+            return "SubmitGraph"
+        else:
+            return "ExecGraph"
+
+    def bin_args(self) -> list[str]:
+        return [
+            "--iterations=100",
+            f"--measureSubmit={self.submit}",
+            f"--ioq={self.ioq}",
+            f"--numKernels={self.numKernels}",
+        ]   

source/adapters/level_zero/image.cpp

@@ -517,6 +517,7 @@ ur_result_t bindlessImagesCreateImpl(ur_context_handle_t hContext,
     ZE2UR_CALL(zeContextMakeImageResident,
                (hContext->ZeContext, hDevice->ZeDevice, ZeImage));
   } else if (MemAllocProperties.type == ZE_MEMORY_TYPE_DEVICE ||
+             MemAllocProperties.type == ZE_MEMORY_TYPE_HOST ||
              MemAllocProperties.type == ZE_MEMORY_TYPE_SHARED) {
     ZeStruct<ze_image_pitched_exp_desc_t> PitchedDesc;
     PitchedDesc.ptr = reinterpret_cast<void *>(hImageMem);

source/loader/layers/sanitizer/asan/asan_buffer.hpp

@@ -29,7 +29,8 @@ struct MemBuffer {
 
   // Sub-buffer constructor
   MemBuffer(std::shared_ptr<MemBuffer> Parent, size_t Origin, size_t Size)
-      : Context(Parent->Context), Size(Size), SubBuffer{{Parent, Origin}} {}
+      : Context(Parent->Context), Size(Size),
+        SubBuffer{{std::move(Parent), Origin}} {}
 
   ur_result_t getHandle(ur_device_handle_t Device, char *&Handle);
 

test/adapters/level_zero/v2/CMakeLists.txt

@@ -63,25 +63,32 @@ add_adapter_test(level_zero_memory_residency
 )
 
 if(NOT WIN32)
-    add_adapter_test(level_zero_deferred_kernel
-        FIXTURE KERNELS
-        SOURCES
-            deferred_kernel.cpp
-        ENVIRONMENT
-            "UR_ADAPTERS_FORCE_LOAD=\"$<TARGET_FILE:ur_adapter_level_zero_v2>\""
-    )
+    if (NOT UR_DPCXX)
+        # Tests that require kernels can't be used if we aren't generating
+        # device binaries
+        message(WARNING
+            "UR_DPCXX is not defined, skipping some tests for level_zero_v2 adapter")
+    else()
+        add_adapter_test(level_zero_deferred_kernel
+            FIXTURE KERNELS
+            SOURCES
+                deferred_kernel.cpp
+            ENVIRONMENT
+                "UR_ADAPTERS_FORCE_LOAD=\"$<TARGET_FILE:ur_adapter_level_zero_v2>\""
+        )
 
-    set(backend level_zero)
-    add_adapter_memcheck_test(level_zero_deferred_kernel
-        FIXTURE KERNELS
-        SOURCES
-            deferred_kernel.cpp
-        ENVIRONMENT
-            "UR_ADAPTERS_FORCE_LOAD=\"$<TARGET_FILE:ur_adapter_level_zero_v2>\""
-    )
+        set(backend level_zero)
+        add_adapter_memcheck_test(level_zero_deferred_kernel
+            FIXTURE KERNELS
+            SOURCES
+                deferred_kernel.cpp
+            ENVIRONMENT
+                "UR_ADAPTERS_FORCE_LOAD=\"$<TARGET_FILE:ur_adapter_level_zero_v2>\""
+        )
 
-    target_link_libraries(test-adapter-level_zero_deferred_kernel PRIVATE
-        LevelZeroLoader
-        LevelZeroLoader-Headers
-    )
+        target_link_libraries(test-adapter-level_zero_deferred_kernel PRIVATE
+            LevelZeroLoader
+            LevelZeroLoader-Headers
+        )
+    endif()
 endif()