Added Performance tests

BackendBench/backends/llm.py

@@ -14,6 +14,7 @@
 
 import torch
 
+from BackendBench.eval import eval_performance
 from BackendBench.llm_client import LLMKernelGenerator
 from BackendBench.multiprocessing_eval import MultiprocessingEvaluator
 from BackendBench.utils import (
@@ -257,6 +258,34 @@ def test_kernel_correctness(
             feedback_info["summary"] = "Compilation failed"
             return False, feedback_info
 
+    def test_kernel_performance(
+        self, op, kernel_code: str, performance_tests: List, attempt: int = 1
+    ) -> tuple[float, List]:
+        """Test kernel performance return performance score with results."""
+
+        op_str = str(op)
+        op_name = extract_operator_name(op_str)
+        kernel_file = self._generate_kernel_file_path(op_name, attempt)
+
+        # Use compile_kernel_from_string for consistent loading
+        module_name = f"{op_name}_implementation_v{attempt}"
+        try:
+            kernel_impl = compile_kernel_from_string(
+                kernel_code=kernel_code,
+                op_name=op_name,
+                kernel_file_path=kernel_file,
+                expected_fn_name=op_name,
+                module_name=module_name,
+            )
+            performance_score, performance_results = eval_performance(
+                op, kernel_impl, performance_tests
+            )
+        except Exception as e:
+            logger.error(f"Performance evaluation failed: {str(e)}")
+            performance_score, performance_results = 0.0, []
+
+        return performance_score, performance_results
+
     def __getitem__(self, key):
         if key in self.compiled_kernels:
             return self.compiled_kernels[key]
@@ -303,11 +332,21 @@ def generate_kernels(self, suite, max_attempts=5):
 
             # Create feedback callback
             def feedback_callback(kernel_code: str, attempt: int) -> tuple[bool, Dict]:
-                # TODO: Add performance testing in addition to correctness testing
-                return self.test_kernel_correctness(
+                is_correct, feedback_info = self.test_kernel_correctness(
                     op, kernel_code, op_test.correctness_tests, attempt
                 )
 
+                if is_correct:
+                    perf_score, perf_results = self.test_kernel_performance(
+                        op, kernel_code, op_test.performance_tests, attempt
+                    )
+                    feedback_info["performance_score"] = perf_score
+                    feedback_info["performance_results"] = perf_results
+                else:
+                    feedback_info["performance_score"] = "N/A"
+                    feedback_info["performance_results"] = []
+                return is_correct, feedback_info
+
             # Generate kernel with iterative refinement
             kernel_code, attempts_used, success = self.llm_client.generate_kernel_with_retry(
                 op_name,