[SYCL][BENCHMARK] Enable velocityBench and SyclBench for cuda adapter (#17217)

Enable benchmarking sycl for cuda adapter through VelocityBench and
SyclBench in benchmarking scripts.

Author: Omar Ahmed

unified-runtime/scripts/benchmarks/benches/compute.py

@@ -58,6 +58,9 @@ def benchmarks(self) -> list[Benchmark]:
         if options.sycl is None:
             return []
 
+        if options.ur_adapter == "cuda":
+            return []
+
         benches = [
             SubmitKernelL0(self, 0),
             SubmitKernelL0(self, 1),

unified-runtime/scripts/benchmarks/benches/llamacpp.py

@@ -75,6 +75,9 @@ def benchmarks(self) -> list[Benchmark]:
         if options.sycl is None:
             return []
 
+        if options.ur_adapter == "cuda":
+            return []
+
         return [LlamaBench(self)]
 
 

unified-runtime/scripts/benchmarks/benches/syclbench.py

@@ -45,6 +45,11 @@ def setup(self):
             f"-DSYCL_IMPL=dpcpp",
         ]
 
+        if options.ur_adapter == "cuda":
+            configure_command += [
+                f"-DCMAKE_CXX_FLAGS=-fsycl -fsycl-targets=nvptx64-nvidia-cuda"
+            ]
+
         run(configure_command, add_sycl=True)
         run(f"cmake --build {build_path} -j", add_sycl=True)
 

unified-runtime/scripts/benchmarks/benches/velocity.py

@@ -41,6 +41,15 @@ def benchmarks(self) -> list[Benchmark]:
         if options.sycl is None:
             return []
 
+        if options.ur_adapter == "cuda":
+            return [
+                Hashtable(self),
+                Bitcracker(self),
+                CudaSift(self),
+                QuickSilver(self),
+                SobelFilter(self),
+            ]
+
         return [
             Hashtable(self),
             Bitcracker(self),
@@ -66,6 +75,8 @@ def download_deps(self):
         return
 
     def extra_cmake_args(self) -> list[str]:
+        if options.ur_adapter == "cuda":
+            return [f"-DUSE_NVIDIA_BACKEND=YES", f"-DUSE_SM=80"]
         return []
 
     def ld_libraries(self) -> list[str]:
@@ -358,6 +369,12 @@ def download_deps(self):
 
     def extra_cmake_args(self):
         oneapi = get_oneapi()
+        if options.ur_adapter == "cuda":
+            return [
+                f"-DUSE_NVIDIA_BACKEND=YES",
+                f"-DUSE_SM=80",
+                f"-DCMAKE_CXX_FLAGS=-O3 -fsycl -ffast-math -I{oneapi.dnn_include()} -I{oneapi.mkl_include()} -L{oneapi.dnn_lib()} -L{oneapi.mkl_lib()}",
+            ]
         return [
             f"-DCMAKE_CXX_FLAGS=-O3 -fsycl -ffast-math -I{oneapi.dnn_include()} -I{oneapi.mkl_include()} -L{oneapi.dnn_lib()} -L{oneapi.mkl_lib()}"
         ]
@@ -415,6 +432,12 @@ def download_deps(self):
 
     def extra_cmake_args(self):
         oneapi = get_oneapi()
+        if options.ur_adapter == "cuda":
+            return [
+                f"-DUSE_NVIDIA_BACKEND=YES",
+                f"-DUSE_SM=80",
+                f"-DCMAKE_CXX_FLAGS=-O3 -fsycl -ffast-math -I{oneapi.dnn_include()} -I{oneapi.mkl_include()} -L{oneapi.dnn_lib()} -L{oneapi.mkl_lib()}",
+            ]
         return [
             f"-DCMAKE_CXX_FLAGS=-O3 -fsycl -ffast-math -I{oneapi.dnn_include()} -I{oneapi.mkl_include()} -L{oneapi.dnn_lib()} -L{oneapi.mkl_lib()}"
         ]
@@ -452,6 +475,12 @@ def ld_libraries(self):
 
     def extra_cmake_args(self):
         oneapi = get_oneapi()
+        if options.ur_adapter == "cuda":
+            return [
+                f"-DUSE_NVIDIA_BACKEND=YES",
+                f"-DUSE_SM=80",
+                f"-DCMAKE_CXX_FLAGS=-O3 -fsycl -ffast-math -I{oneapi.dnn_include()} -I{oneapi.mkl_include()} -L{oneapi.dnn_lib()} -L{oneapi.mkl_lib()}",
+            ]
         return [
             f"-DCMAKE_CXX_FLAGS=-O3 -fsycl -ffast-math -I{oneapi.dnn_include()} -I{oneapi.mkl_include()} -L{oneapi.dnn_lib()} -L{oneapi.mkl_lib()}"
         ]

unified-runtime/scripts/benchmarks/main.py

@@ -411,6 +411,18 @@ def validate_and_parse_env_args(env_args):
         help="The name of the results which should be used as a baseline for metrics calculation",
         default=options.current_run_name,
     )
+    parser.add_argument(
+        "--cudnn_directory",
+        type=str,
+        help="Directory for cudnn library",
+        default=None,
+    )
+    parser.add_argument(
+        "--cublas_directory",
+        type=str,
+        help="Directory for cublas library",
+        default=None,
+    )
 
     args = parser.parse_args()
     additional_env_vars = validate_and_parse_env_args(args.env)
@@ -434,6 +446,8 @@ def validate_and_parse_env_args(env_args):
     options.iterations_stddev = args.iterations_stddev
     options.build_igc = args.build_igc
     options.current_run_name = args.relative_perf
+    options.cudnn_directory = args.cudnn_directory
+    options.cublas_directory = args.cublas_directory
 
     if args.build_igc and args.compute_runtime is None:
         parser.error("--build-igc requires --compute-runtime to be set")