[Bench] Add TorchLinearKernelSize benchmark

Author: lukaszstolarczuk

devops/scripts/benchmarks/benches/compute.py

@@ -385,6 +385,46 @@ def createTorchSlmSizeBench(variant_name: str, **kwargs):
                 ),
             ]
 
+        # Add TorchLinearKernelSize benchmarks
+        for runtime in filter(lambda x: x != RUNTIMES.UR, RUNTIMES):
+
+            def createTorchLinearKernelSizeBench(variant_name: str, **kwargs):
+                return TorchLinearKernelSize(
+                    self,
+                    runtime,
+                    variant_name,
+                    PROFILERS.TIMER,
+                    **kwargs,
+                )
+
+            benches += [
+                createTorchLinearKernelSizeBench(
+                    "array32",
+                    kernelBatchSize=512,
+                    kernelSize=32,
+                ),
+                createTorchLinearKernelSizeBench(
+                    "array128",
+                    kernelBatchSize=512,
+                    kernelSize=128,
+                ),
+                createTorchLinearKernelSizeBench(
+                    "array512",
+                    kernelBatchSize=512,
+                    kernelSize=512,
+                ),
+                createTorchLinearKernelSizeBench(
+                    "array1024",
+                    kernelBatchSize=512,
+                    kernelSize=1024,
+                ),
+                createTorchLinearKernelSizeBench(
+                    "array5120",
+                    kernelBatchSize=512,
+                    kernelSize=5120,
+                ),
+            ]
+
         # Add UR-specific benchmarks
         benches += [
             # TODO: multithread_benchmark_ur fails with segfault
@@ -916,6 +956,20 @@ def __init__(
         )
 
 
+class TorchLinearKernelSize(TorchBenchmark):
+    def __init__(
+        self, suite, runtime: RUNTIMES, variant_name: str, profiler_type, **kwargs
+    ):
+        super().__init__(
+            suite,
+            runtime,
+            "KernelSubmitLinearKernelSize",
+            variant_name,
+            profiler_type,
+            **kwargs,
+        )
+
+
 class QueueInOrderMemcpy(ComputeBenchmark):
     def __init__(self, bench, isCopyOnly, source, destination, size, profiler_type):
         self._is_copy_only = isCopyOnly

devops/scripts/benchmarks/tests/test_integration.py

@@ -199,6 +199,11 @@ def test_torch_l0(self):
             "KernelSubmitSlmSize small",
             {"pytorch", "L0"},
         )
+        self._checkCase(
+            "torch_benchmark_l0 kernelBatchSize 512, kernelSize 32",
+            "KernelSubmitLinearKernelSize array32",
+            {"pytorch", "L0"},
+        )
 
     def test_torch_sycl(self):
         self._checkCase(
@@ -211,6 +216,11 @@ def test_torch_sycl(self):
             "KernelSubmitSlmSize max",
             {"pytorch", "SYCL"},
         )
+        self._checkCase(
+            "torch_benchmark_sycl kernelBatchSize 512, kernelSize 5120",
+            "KernelSubmitLinearKernelSize array5120",
+            {"pytorch", "SYCL"},
+        )
 
     def test_torch_syclpreview(self):
         self._checkCase(
@@ -223,6 +233,11 @@ def test_torch_syclpreview(self):
             "KernelSubmitSlmSize medium",
             {"pytorch", "SYCL"},
         )
+        self._checkCase(
+            "torch_benchmark_syclpreview kernelBatchSize 512, kernelSize 512",
+            "KernelSubmitLinearKernelSize array512",
+            {"pytorch", "SYCL"},
+        )
 
 
 if __name__ == "__main__":