Add workaround in IPEX Triton bench (#3971)

* Add legacy triton spirv-target launcher support.

* Work around that the XPU event cannot capture the Triton kernel time.

intel_extension_for_pytorch/_inductor/xpu/utils.py

@@ -2,6 +2,15 @@
 
 import torch
 
+from datetime import datetime
+
+def synchronize():
+    import torch
+    if torch.cuda.is_available():
+        torch.cuda.synchronize()
+    elif torch.xpu.is_available():
+        torch.xpu.synchronize()
+
 
 def triton_do_bench(
     fn,
@@ -58,6 +67,8 @@ def triton_do_bench(
     n_repeat = max(1, int(rep / estimate_ms))
     start_event = [torch.xpu.Event(enable_timing=True) for i in range(n_repeat)]
     end_event = [torch.xpu.Event(enable_timing=True) for i in range(n_repeat)]
+    start_times = [datetime for i in range(n_repeat)]
+    end_times = [datetime for i in range(n_repeat)]
     # Warm-up
     for _ in range(n_warmup):
         fn()
@@ -71,15 +82,21 @@ def triton_do_bench(
                 x.grad = None
         # we clear the L2 cache before each run
         cache.zero_()
+        synchronize()
         # record time of `fn`
         start_event[i].record()
+        start_times[i] = datetime.now()
         fn()
         end_event[i].record()
+        synchronize()
+        end_times[i] = datetime.now()
     # Record clocks
     torch.xpu.synchronize()
-    times = torch.tensor(
-        [s.elapsed_time(e) for s, e in zip(start_event, end_event)], dtype=torch.float
-    )
+    # times = torch.tensor(
+    #     [s.elapsed_time(e) for s, e in zip(start_event, end_event)], dtype=torch.float
+    # )
+    times = torch.tensor([(e.timestamp() - s.timestamp()) * 1000 for s, e in zip(start_times, end_times)],
+                         dtype=torch.float)
     if quantiles is not None:
         ret = torch.quantile(times, torch.tensor(quantiles, dtype=torch.float)).tolist()
         if len(ret) == 1: