Add launch latency benchmarks for triton.CompiledKernel and inductor

Summary:
There are a number of more detailed views into launch latency that we can get, in addition to the path we get from `triton.JitFunction`:
- `triton.compiler.CompiledKernel`, which is the lowest-level interface used by
  triton
- Inductor's `CachingAutotuner.run`, which is the lowest-level lauch interface
  used by inductor
- launching a mostly-nop inductor kernel (can't be truly nop because inductor
  won't generate a kernel with nothing in it)

Reviewed By: xuzhao9, chenyang78

Differential Revision: D56073036

fbshipit-source-id: c72b80eb016a5c2ea27717664e8a1ff0f35c705a

torchbenchmark/operators/launch_latency/__init__.py

@@ -1,6 +1,9 @@
 import torch
 import triton
 import triton.language as tl
+from torch._C import _cuda_getCurrentRawStream as get_raw_stream
+from torch._inductor import triton_heuristics
+from torch._inductor.codecache import AsyncCompile
 
 from torchbenchmark.util.triton_op import (
     BenchmarkOperator,
@@ -40,6 +43,55 @@ def nop_with_args_kernel(
     pass
 
 
+@torch.compile
+def trivial_add_kernel(*args):
+    return sum([torch.tensor(1.0, device="cuda"), *args])
+
+
+async_compile = AsyncCompile()
+
+inductor_nop = async_compile.triton(
+    "inductor_nop",
+    """
+import triton
+import triton.language as tl
+from triton.compiler.compiler import AttrsDescriptor
+
+from torch._inductor import triton_heuristics
+
+@triton_heuristics.pointwise(
+    size_hints=[1],
+    triton_meta={'signature': {0: 'i32'}, 'device': 0, 'device_type': 'cuda', 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(), equal_to_1=())]},
+)
+@triton.jit
+def inductor_nop(x):
+    pass
+""",
+    device_str="cuda",
+)
+
+
+inductor_nop_args = async_compile.triton(
+    "inductor_nop_args",
+    """
+import triton
+import triton.language as tl
+from triton.compiler.compiler import AttrsDescriptor
+
+from torch._inductor import triton_heuristics
+
+@triton_heuristics.pointwise(
+    size_hints=[1],
+    triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: '*fp32', 5: 'i32', 6: 'i32', 7: 'i32', 8: 'i32', 9: 'i32', 10: 'i32', 11: 'i32', 12: 'i32', 13: 'i32'}, 'device': 0, 'device_type': 'cuda', 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4), equal_to_1=(5, 6, 7, 8, 9, 10, 11, 12, 13))]},
+)
+@triton.jit
+def inductor_nop_args(t1, t2, t3, t4, t5, i1, i2, i3, i4, i5, i6, i7, i8, i9):
+    pass
+""",
+    device_str="cuda",
+)
+
+
 class Operator(BenchmarkOperator):
     DEFAULT_METRICS = ["walltime"]
 
@@ -59,6 +111,41 @@ def nop_triton_kernel(self, *args):
             return lambda: nop_kernel[1,]()
         return lambda: nop_with_args_kernel[1,](*args)
 
+    @register_benchmark()
+    def nop_triton_compiled_kernel_run(self, *args):
+        if len(args) == 0:
+            bin = nop_kernel[1,]()
+
+        else:
+            bin = nop_with_args_kernel[1,](*args)
+            args = args[:-5]  # remove tl.constexpr args
+        function = bin.function
+        metadata = (
+            bin.packed_metadata if hasattr(bin, "packed_metadata") else bin.metadata
+        )
+        if hasattr(triton.compiler.CompiledKernel, "launch_metadata"):
+            return lambda: bin.run(
+                1, 1, 1, 0, function, metadata, None, None, None, *args
+            )
+        else:
+            return lambda: bin.run(
+                1, 1, 1, 1, 1, 1, 1, 1, 0, 0, function, None, None, metadata, *args
+            )
+
+    @register_benchmark()
+    def nop_inductor_kernel_run(self, *args):
+        stream = get_raw_stream(0)
+        grid = triton_heuristics.grid(1)
+
+        if len(args) == 0:
+            return lambda: inductor_nop.run(1, grid=grid, stream=stream)
+        args = args[:-5]
+        return lambda: inductor_nop_args.run(*args, grid=grid, stream=stream)
+
+    @register_benchmark()
+    def nop_inductor_kernel(self, *args):
+        return lambda: trivial_add_kernel(*args)
+
     @register_benchmark(baseline=True)
     def nop_python_function(self, *args):
         def nop():