[Bench] Add external baselines for common ops

benchmarks/benchmark_base.py

@@ -543,6 +543,17 @@ def workload_field_params(workloads: list, keys: tuple) -> list:
         )
     return params
 
+
+def torch_inductor_baseline(fn: Callable) -> Callable:
+    """Compile a benchmark-local PyTorch baseline with TorchInductor.
+
+    The baseline implementation itself stays in the benchmark file so tests do
+    not become performance oracles. This helper only centralizes the compile
+    policy used by common-op benchmark baselines.
+    """
+    return torch.compile(fn, fullgraph=True)
+
+
 class ManifestBenchmark(BenchmarkBase[ShapeDtypeWorkload]):
     """Generic benchmark that reads FLOP/memory counts from an Op instance.
 

benchmarks/ops/bench_ada_layer_norm.py

@@ -2,7 +2,7 @@
 import torch
 import torch.nn.functional as F
 
-from benchmarks.benchmark_base import BenchmarkReport, ManifestBenchmark
+from benchmarks.benchmark_base import BenchmarkReport, ManifestBenchmark, torch_inductor_baseline
 from tileops.manifest import load_workloads
 from tileops.ops.norm.ada_layer_norm import AdaLayerNormFwdOp
 from tileops.ops.norm.ada_layer_norm_zero import AdaLayerNormZeroFwdOp
@@ -39,8 +39,8 @@ def baseline_fn(x, scale, shift):
         normed = F.layer_norm(x, (n,), weight=None, bias=None, eps=test.eps)
         return scale * normed + shift
 
-    result_bl = bm.profile(baseline_fn, *inputs)
-    BenchmarkReport.record(op, locals(), result_bl, tag="torch-ref")
+    result_bl = bm.profile(torch_inductor_baseline(baseline_fn), *inputs)
+    BenchmarkReport.record(op, locals(), result_bl, tag="torch_inductor")
 
 
 @pytest.mark.parametrize("m, n, dtype", _to_params(load_workloads(_ADA_ZERO_OP_NAME)))
@@ -58,8 +58,8 @@ def baseline_fn(x, scale, shift, gate):
         normed = F.layer_norm(x, (n,), weight=None, bias=None, eps=test.eps)
         return gate * (scale * normed + shift)
 
-    result_bl = bm.profile(baseline_fn, *inputs)
-    BenchmarkReport.record(op, locals(), result_bl, tag="torch-ref")
+    result_bl = bm.profile(torch_inductor_baseline(baseline_fn), *inputs)
+    BenchmarkReport.record(op, locals(), result_bl, tag="torch_inductor")
 
 
 if __name__ == "__main__":

benchmarks/ops/bench_argreduce.py

@@ -9,7 +9,12 @@
 import pytest
 import torch
 
-from benchmarks.benchmark_base import BenchmarkReport, ManifestBenchmark, workloads_to_params
+from benchmarks.benchmark_base import (
+    BenchmarkReport,
+    ManifestBenchmark,
+    torch_inductor_baseline,
+    workloads_to_params,
+)
 from tileops.ops.reduction.argreduce import ArgmaxFwdOp, ArgminFwdOp
 from workloads.reduction import ArgmaxTest, ArgminTest
 
@@ -60,8 +65,8 @@ def test_argmax_bench(shape: tuple, dtype: torch.dtype, extra: dict) -> None:
     def baseline_fn(x):
         return x.argmax(dim=dim)
 
-    result_bl = bm.profile(baseline_fn, *inputs)
-    BenchmarkReport.record(op, locals(), result_bl, tag="torch")
+    result_bl = bm.profile(torch_inductor_baseline(baseline_fn), *inputs)
+    BenchmarkReport.record(op, locals(), result_bl, tag="torch_inductor")
 
 
 # Argmin benchmarks
@@ -93,8 +98,8 @@ def test_argmin_bench(shape: tuple, dtype: torch.dtype, extra: dict) -> None:
     def baseline_fn(x):
         return x.argmin(dim=dim)
 
-    result_bl = bm.profile(baseline_fn, *inputs)
-    BenchmarkReport.record(op, locals(), result_bl, tag="torch")
+    result_bl = bm.profile(torch_inductor_baseline(baseline_fn), *inputs)
+    BenchmarkReport.record(op, locals(), result_bl, tag="torch_inductor")
 
 
 if __name__ == "__main__":

benchmarks/ops/bench_batch_norm.py

@@ -110,7 +110,7 @@ def test_batch_norm_fwd_bench(N, C, spatial, dtype, training, tune):
     result_bl = bm.profile(
         lambda x, rm, rv, w, b: _torch_bn_fwd(x, w, b, rm, rv), *inputs,
     )
-    BenchmarkReport.record(op, locals(), result_bl, tag="torch-cudnn")
+    BenchmarkReport.record(op, locals(), result_bl, tag="torch_cudnn")
 
 
 @pytest.mark.parametrize("N, C, spatial, dtype", _manifest_bwd_params())
@@ -127,7 +127,7 @@ def test_batch_norm_bwd_bench(N, C, spatial, dtype):
     BenchmarkReport.record(op, locals(), result, tag="tileops")
 
     result_bl = bm.profile(_torch_bn_bwd, *inputs)
-    BenchmarkReport.record(op, locals(), result_bl, tag="torch-autograd")
+    BenchmarkReport.record(op, locals(), result_bl, tag="torch_cudnn")
 
 
 if __name__ == "__main__":

benchmarks/ops/bench_binary_elementwise.py

@@ -15,6 +15,7 @@
     BenchmarkBase,
     BenchmarkReport,
     ManifestBenchmark,
+    torch_inductor_baseline,
 )
 from tileops.kernels.elementwise import (
     GeluAndMulFwdKernel,
@@ -200,8 +201,8 @@ def test_binary_arith_bench(
     result = bm.profile(op, *inputs)
     BenchmarkReport.record(op, locals(), result, tag="tileops")
 
-    result_bl = bm.profile(baseline_fn, *inputs)
-    BenchmarkReport.record(op, locals(), result_bl, tag="torch")
+    result_bl = bm.profile(torch_inductor_baseline(baseline_fn), *inputs)
+    BenchmarkReport.record(op, locals(), result_bl, tag="torch_inductor")
 
 
 # Comparison ops (6)
@@ -245,8 +246,8 @@ def test_comparison_bench(
     result = bm.profile(op, *inputs)
     BenchmarkReport.record(op, locals(), result, tag="tileops")
 
-    result_bl = bm.profile(baseline_fn, *inputs)
-    BenchmarkReport.record(op, locals(), result_bl, tag="torch")
+    result_bl = bm.profile(torch_inductor_baseline(baseline_fn), *inputs)
+    BenchmarkReport.record(op, locals(), result_bl, tag="torch_inductor")
 
 
 # Logical ops (2)
@@ -281,8 +282,8 @@ def test_logical_bench(
 
     # Baseline uses bool tensors
     a_bool, b_bool = inputs[0].bool(), inputs[1].bool()
-    result_bl = bm.profile(baseline_fn, a_bool, b_bool)
-    BenchmarkReport.record(op, locals(), result_bl, tag="torch")
+    result_bl = bm.profile(torch_inductor_baseline(baseline_fn), a_bool, b_bool)
+    BenchmarkReport.record(op, locals(), result_bl, tag="torch_inductor")
 
 
 # Bitwise ops (3)
@@ -315,8 +316,8 @@ def test_bitwise_bench(
     result = bm.profile(op, *inputs)
     BenchmarkReport.record(op, locals(), result, tag="tileops")
 
-    result_bl = bm.profile(baseline_fn, *inputs)
-    BenchmarkReport.record(op, locals(), result_bl, tag="torch")
+    result_bl = bm.profile(torch_inductor_baseline(baseline_fn), *inputs)
+    BenchmarkReport.record(op, locals(), result_bl, tag="torch_inductor")
 
 
 # Fused gated ops (2)
@@ -380,8 +381,11 @@ def _profile_fused_gated(bm: ManifestBenchmark, op, test, baseline_key: str,
     inputs = test.gen_inputs()
     result = bm.profile(op, *inputs)
     BenchmarkReport.record(op, params, result, tag="tileops")
-    result_bl = bm.profile(_FUSED_BASELINES[baseline_key], *inputs)
-    BenchmarkReport.record(op, params, result_bl, tag="torch-ref")
+    result_bl = bm.profile(
+        torch_inductor_baseline(_FUSED_BASELINES[baseline_key]),
+        *inputs,
+    )
+    BenchmarkReport.record(op, params, result_bl, tag="torch_inductor")
 
 
 @SiluAndMulBenchFixture
@@ -475,8 +479,10 @@ def test_fused_gated_strategy_bench(
     BenchmarkReport.record(f"{op_name}_strategy", locals(), result, tag=f"tileops-{strategy}")
 
     baseline_fn = _FUSED_BASELINES[op_name]
-    result_bl = bm.profile(baseline_fn, *inputs)
-    BenchmarkReport.record(f"{op_name}_strategy", locals(), result_bl, tag="torch")
+    result_bl = bm.profile(torch_inductor_baseline(baseline_fn), *inputs)
+    BenchmarkReport.record(
+        f"{op_name}_strategy", locals(), result_bl, tag="torch_inductor",
+    )
 
 
 # Broadcast benchmark (bias-add pattern)
@@ -575,8 +581,10 @@ def test_broadcast_bench(
     result = bm.profile(op, *inputs)
     BenchmarkReport.record(f"{op_name}_bcast", locals(), result, tag="tileops")
 
-    result_bl = bm.profile(baseline_fn, *inputs)
-    BenchmarkReport.record(f"{op_name}_bcast", locals(), result_bl, tag="torch")
+    result_bl = bm.profile(torch_inductor_baseline(baseline_fn), *inputs)
+    BenchmarkReport.record(
+        f"{op_name}_bcast", locals(), result_bl, tag="torch_inductor",
+    )
 
 
 if __name__ == "__main__":

benchmarks/ops/bench_convolution.py

@@ -136,7 +136,7 @@ def _profile_conv(
     BenchmarkReport.record(op, params, result, tag="tileops")
 
     result_bl = bm.profile(baseline_fn, *inputs)
-    BenchmarkReport.record(op, params, result_bl, tag="torch")
+    BenchmarkReport.record(op, params, result_bl, tag="torch_cudnn")
 
 
 # Conv1d

benchmarks/ops/bench_cumulative.py

@@ -11,6 +11,7 @@
 from benchmarks.benchmark_base import (
     BenchmarkReport,
     ManifestBenchmark,
+    torch_inductor_baseline,
     workloads_to_params,
 )
 from workloads.workload_base import WorkloadBase
@@ -63,8 +64,8 @@ def test_cumsum_bench(shape: tuple, dtype: torch.dtype) -> None:
     result = bm.profile(op, *inputs)
     BenchmarkReport.record(op, locals(), result, tag="tileops")
 
-    result_bl = bm.profile(test.ref_program, *inputs)
-    BenchmarkReport.record(op, locals(), result_bl, tag="torch")
+    result_bl = bm.profile(torch_inductor_baseline(test.ref_program), *inputs)
+    BenchmarkReport.record(op, locals(), result_bl, tag="torch_inductor")
 
 
 @pytest.mark.parametrize("shape, dtype", workloads_to_params(_CUMPROD_OP))
@@ -77,8 +78,8 @@ def test_cumprod_bench(shape: tuple, dtype: torch.dtype) -> None:
     result = bm.profile(op, *inputs)
     BenchmarkReport.record(op, locals(), result, tag="tileops")
 
-    result_bl = bm.profile(test.ref_program, *inputs)
-    BenchmarkReport.record(op, locals(), result_bl, tag="torch")
+    result_bl = bm.profile(torch_inductor_baseline(test.ref_program), *inputs)
+    BenchmarkReport.record(op, locals(), result_bl, tag="torch_inductor")
 
 
 if __name__ == "__main__":

benchmarks/ops/bench_dropout.py

@@ -10,7 +10,12 @@
 import torch
 import torch.nn.functional as F
 
-from benchmarks.benchmark_base import BenchmarkReport, ManifestBenchmark, workloads_to_params
+from benchmarks.benchmark_base import (
+    BenchmarkReport,
+    ManifestBenchmark,
+    torch_inductor_baseline,
+    workloads_to_params,
+)
 from tileops.ops.dropout import DropoutOp
 from workloads.workload_base import WorkloadBase
 
@@ -41,8 +46,8 @@ def test_dropout_bench(shape: tuple, dtype: torch.dtype) -> None:
     result = bm.profile(op, x)
     BenchmarkReport.record(op, locals(), result, tag="tileops")
 
-    result_bl = bm.profile(test.ref_program, x)
-    BenchmarkReport.record(op, locals(), result_bl, tag="torch")
+    result_bl = bm.profile(torch_inductor_baseline(test.ref_program), x)
+    BenchmarkReport.record(op, locals(), result_bl, tag="torch_inductor")
 
 
 if __name__ == "__main__":

benchmarks/ops/bench_elementwise_manifest.py

@@ -12,7 +12,7 @@
 import torch
 import torch.nn.functional as F
 
-from benchmarks.benchmark_base import BenchmarkReport, ManifestBenchmark
+from benchmarks.benchmark_base import BenchmarkReport, ManifestBenchmark, torch_inductor_baseline
 from tileops.manifest import load_workloads
 from tileops.ops.elementwise import (
     AddFwdOp,
@@ -362,8 +362,8 @@ def _record_unary(
 ) -> None:
     result = bm.profile(op, *inputs)
     BenchmarkReport.record(op, _manifest_params(bm), result, tag="tileops")
-    result_bl = bm.profile(baseline_fn, *inputs)
-    BenchmarkReport.record(op, _manifest_params(bm), result_bl, tag="torch")
+    result_bl = bm.profile(torch_inductor_baseline(baseline_fn), *inputs)
+    BenchmarkReport.record(op, _manifest_params(bm), result_bl, tag="torch_inductor")
 
 
 def _record_binary(
@@ -374,8 +374,8 @@ def _record_binary(
 ) -> None:
     result = bm.profile(op, *inputs)
     BenchmarkReport.record(op, _manifest_params(bm), result, tag="tileops")
-    result_bl = bm.profile(baseline_fn, *inputs)
-    BenchmarkReport.record(op, _manifest_params(bm), result_bl, tag="torch")
+    result_bl = bm.profile(torch_inductor_baseline(baseline_fn), *inputs)
+    BenchmarkReport.record(op, _manifest_params(bm), result_bl, tag="torch_inductor")
 
 
 _RELU_OP = "ReluFwdOp"
@@ -548,8 +548,8 @@ def test_prelu_manifest_bench(
     bm = ManifestBenchmark(_PRELU_OP, op, test)
     result = bm.profile(op, x, weight)
     BenchmarkReport.record(op, locals(), result, tag="tileops")
-    result_bl = bm.profile(F.prelu, x, weight)
-    BenchmarkReport.record(op, locals(), result_bl, tag="torch")
+    result_bl = bm.profile(torch_inductor_baseline(F.prelu), x, weight)
+    BenchmarkReport.record(op, locals(), result_bl, tag="torch_inductor")
 
 
 _MASKED_FILL_OP = "MaskedFillFwdOp"
@@ -572,8 +572,13 @@ def test_masked_fill_tensor_manifest_bench(
     bm = ManifestBenchmark(_MASKED_FILL_OP, op, test)
     result = bm.profile(op, x, mask, value)
     BenchmarkReport.record(op, locals(), result, tag="tileops")
-    result_bl = bm.profile(lambda a, m, v: a.masked_fill(m, v), x, mask, value)
-    BenchmarkReport.record(op, locals(), result_bl, tag="torch")
+    result_bl = bm.profile(
+        torch_inductor_baseline(lambda a, m, v: a.masked_fill(m, v)),
+        x,
+        mask,
+        value,
+    )
+    BenchmarkReport.record(op, locals(), result_bl, tag="torch_inductor")
 
 
 @pytest.mark.parametrize(
@@ -590,8 +595,12 @@ def test_masked_fill_scalar_manifest_bench(
     bm = ManifestBenchmark(_MASKED_FILL_SCALAR_OP, op, test)
     result = bm.profile(op, x, mask)
     BenchmarkReport.record(op, locals(), result, tag="tileops")
-    result_bl = bm.profile(lambda a, m: a.masked_fill(m, -100.0), x, mask)
-    BenchmarkReport.record(op, locals(), result_bl, tag="torch")
+    result_bl = bm.profile(
+        torch_inductor_baseline(lambda a, m: a.masked_fill(m, -100.0)),
+        x,
+        mask,
+    )
+    BenchmarkReport.record(op, locals(), result_bl, tag="torch_inductor")
 
 
 _ADD_OP = "AddFwdOp"
@@ -841,8 +850,8 @@ def test_where_manifest_bench(shape: tuple[int, ...], dtype: torch.dtype) -> Non
     bm = ManifestBenchmark(_WHERE_OP, op, test)
     result = bm.profile(op, cond, x, other)
     BenchmarkReport.record(op, locals(), result, tag="tileops")
-    result_bl = bm.profile(torch.where, cond, x, other)
-    BenchmarkReport.record(op, locals(), result_bl, tag="torch")
+    result_bl = bm.profile(torch_inductor_baseline(torch.where), cond, x, other)
+    BenchmarkReport.record(op, locals(), result_bl, tag="torch_inductor")
 
 
 @pytest.mark.parametrize("shape, dtype", _shape_dtype_params(load_workloads(_LERP_TENSOR_OP)))
@@ -853,8 +862,8 @@ def test_lerp_tensor_manifest_bench(shape: tuple[int, ...], dtype: torch.dtype)
     bm = ManifestBenchmark(_LERP_TENSOR_OP, op, test)
     result = bm.profile(op, x, end, weight)
     BenchmarkReport.record(op, locals(), result, tag="tileops")
-    result_bl = bm.profile(torch.lerp, x, end, weight)
-    BenchmarkReport.record(op, locals(), result_bl, tag="torch")
+    result_bl = bm.profile(torch_inductor_baseline(torch.lerp), x, end, weight)
+    BenchmarkReport.record(op, locals(), result_bl, tag="torch_inductor")
 
 
 if __name__ == "__main__":

benchmarks/ops/bench_fft.py

@@ -37,7 +37,7 @@ def test_fft_bench(shape: tuple, dtype: torch.dtype) -> None:
     BenchmarkReport.record(op, locals(), result, tag="tileops")
 
     result_bl = bm.profile(test.ref_program, *inputs)
-    BenchmarkReport.record(op, locals(), result_bl, tag="torch-cufft")
+    BenchmarkReport.record(op, locals(), result_bl, tag="torch_cufft")
 
 
 if __name__ == "__main__":

benchmarks/ops/bench_group_norm.py

@@ -2,7 +2,7 @@
 import torch
 import torch.nn.functional as F
 
-from benchmarks.benchmark_base import BenchmarkReport, ManifestBenchmark
+from benchmarks.benchmark_base import BenchmarkReport, ManifestBenchmark, torch_inductor_baseline
 from tileops.manifest import load_workloads
 from tileops.ops.norm.group_norm import (
     GroupNormFwdOp,
@@ -52,8 +52,8 @@ def test_group_norm_bench(n: int, c: int, spatial: tuple, num_groups: int,
     def baseline_fn(x, weight, bias):
         return F.group_norm(x, num_groups, weight=weight, bias=bias, eps=1e-5)
 
-    result_bl = bm.profile(baseline_fn, x, weight, bias)
-    BenchmarkReport.record(op, locals(), result_bl, tag="torch")
+    result_bl = bm.profile(torch_inductor_baseline(baseline_fn), x, weight, bias)
+    BenchmarkReport.record(op, locals(), result_bl, tag="torch_inductor")
 
 
 @pytest.mark.parametrize("n, c, spatial, num_groups, dtype, tune",
@@ -72,8 +72,8 @@ def test_group_norm_no_affine_bench(n: int, c: int, spatial: tuple,
     def baseline_no_affine(x):
         return F.group_norm(x, num_groups, weight=None, bias=None, eps=1e-5)
 
-    result_bl = bm.profile(baseline_no_affine, x)
-    BenchmarkReport.record(op, locals(), result_bl, tag="torch")
+    result_bl = bm.profile(torch_inductor_baseline(baseline_no_affine), x)
+    BenchmarkReport.record(op, locals(), result_bl, tag="torch_inductor")
 
 
 if __name__ == "__main__":