[Relax] Pipeline file reorganization

docs/how_to/tutorials/customize_opt.py

@@ -103,7 +103,7 @@ def forward(self, x):
 
 
 # Import cublas pattern
-import tvm.relax.backend.contrib.cublas as _cublas
+import tvm.relax.backend.cuda.cublas as _cublas
 
 
 # Define a new pass for CUBLAS dispatch

python/tvm/relax/backend/__init__.py

@@ -16,26 +16,7 @@
 # under the License.
 """Relax backends"""
 
-from tvm.target import Target
-
-from . import contrib
+from . import contrib, cpu_generic, cuda, gpu_generic, metal, rocm
 from .dispatch_sampling import DispatchSampling
 from .dispatch_sort_scan import DispatchSortScan
 from .pattern_registry import get_pattern, get_patterns_with_prefix
-
-
-def get_default_pipeline(target: Target):
-    """Get the default Relax compilation pipeline for the given target."""
-    if target.kind.name == "cuda":
-        from . import cuda  # pylint: disable=import-outside-toplevel
-
-        return cuda.get_default_pipeline(target)
-    if target.kind.name == "llvm":
-        from . import cpu_generic  # pylint: disable=import-outside-toplevel
-
-        return cpu_generic.get_default_pipeline(target)
-    # Todo(tvm-team): support gpu-generic
-    raise ValueError(
-        f"Target {target} is not yet supported by default pipeline. "
-        "Please lower and build the IRModule manually."
-    )

python/tvm/relax/backend/gpu_generic/__init__.py

@@ -0,0 +1,23 @@
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements.  See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership.  The ASF licenses this file
+# to you under the Apache License, Version 2.0 (the
+# "License"); you may not use this file except in compliance
+# with the License.  You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, either express or implied.  See the License for the
+# specific language governing permissions and limitations
+# under the License.
+"""The Relax Metal backend compilation pipeline and other passes."""
+from .pipeline import (
+    finalize_passes,
+    get_default_pipeline,
+    legalize_passes,
+    library_dispatch_passes,
+)

python/tvm/relax/backend/gpu_generic/pipeline.py

@@ -0,0 +1,87 @@
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements.  See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership.  The ASF licenses this file
+# to you under the Apache License, Version 2.0 (the
+# "License"); you may not use this file except in compliance
+# with the License.  You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, either express or implied.  See the License for the
+# specific language governing permissions and limitations
+# under the License.
+"""The Relax generic GPU backend compilation pipeline and other passes."""
+import tvm
+from tvm import dlight as dl
+from tvm import relax
+
+
+def library_dispatch_passes(target: tvm.target.Target):  # pylint: disable=unused-argument
+    """The default library dispatch passes for generic GPU backend."""
+    return [
+        relax.backend.DispatchSampling(),
+        relax.backend.DispatchSortScan(),
+    ]
+
+
+def legalize_passes(target: tvm.target.Target):  # pylint: disable=unused-argument
+    """The default legalization passes for generic GPU backend."""
+    return [
+        tvm.relax.transform.LegalizeOps(),
+        tvm.relax.transform.AnnotateTIROpPattern(),
+        tvm.relax.transform.FoldConstant(),
+        tvm.relax.transform.FuseOps(),
+        tvm.relax.transform.FuseTIR(),
+        dl.ApplyDefaultSchedule(
+            dl.gpu.Matmul(),
+            dl.gpu.GEMV(),
+            dl.gpu.Reduction(),
+            dl.gpu.GeneralReduction(),
+            dl.gpu.Fallback(),
+        ),
+    ]
+
+
+def dataflow_lower_passes(target: tvm.target.Target):  # pylint: disable=unused-argument
+    """The default dataflow lowering passes for generic GPU backend."""
+    return [
+        relax.transform.RewriteDataflowReshape(),
+        relax.transform.ToNonDataflow(),
+        relax.transform.RemovePurityChecking(),
+        relax.transform.CallTIRRewrite(),
+    ]
+
+
+def finalize_passes(target: tvm.target.Target):  # pylint: disable=unused-argument
+    """The default finalization passes for generic GPU backend."""
+    return [
+        relax.transform.StaticPlanBlockMemory(),
+        relax.transform.LowerAllocTensor(),
+        relax.transform.KillAfterLastUse(),
+        relax.transform.LowerRuntimeBuiltin(),
+        relax.transform.ComputePrimValue(),
+        relax.transform.VMShapeLower(),
+        relax.transform.AttachGlobalSymbol(),
+    ]
+
+
+def get_default_pipeline(target: tvm.target.Target):
+    """Return the default compilation pipeline for generic GPU."""
+
+    @tvm.transform.module_pass(opt_level=0)
+    def _pipeline(mod: tvm.ir.IRModule, _ctx: tvm.transform.PassContext):
+        with target:
+            seq = tvm.transform.Sequential(
+                library_dispatch_passes(target)
+                + legalize_passes(target)
+                + dataflow_lower_passes(target)
+                + finalize_passes(target)
+            )
+            mod = seq(mod)
+        return mod
+
+    return _pipeline

python/tvm/relax/backend/metal/__init__.py

@@ -0,0 +1,17 @@
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements.  See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership.  The ASF licenses this file
+# to you under the Apache License, Version 2.0 (the
+# "License"); you may not use this file except in compliance
+# with the License.  You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, either express or implied.  See the License for the
+# specific language governing permissions and limitations
+# under the License.
+"""The Relax Metal backend compilation pipeline and other passes."""

python/tvm/relax/backend/rocm/__init__.py

@@ -0,0 +1,23 @@
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements.  See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership.  The ASF licenses this file
+# to you under the Apache License, Version 2.0 (the
+# "License"); you may not use this file except in compliance
+# with the License.  You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, either express or implied.  See the License for the
+# specific language governing permissions and limitations
+# under the License.
+"""The Relax ROCm backend compilation pipeline and other passes."""
+from .pipeline import (
+    finalize_passes,
+    get_default_pipeline,
+    legalize_passes,
+    library_dispatch_passes,
+)

python/tvm/relax/backend/rocm/pipeline.py

@@ -0,0 +1,87 @@
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements.  See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership.  The ASF licenses this file
+# to you under the Apache License, Version 2.0 (the
+# "License"); you may not use this file except in compliance
+# with the License.  You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, either express or implied.  See the License for the
+# specific language governing permissions and limitations
+# under the License.
+"""The Relax ROCm backend compilation pipeline and other passes."""
+import tvm
+from tvm import dlight as dl
+from tvm import relax
+
+
+def library_dispatch_passes(target: tvm.target.Target):  # pylint: disable=unused-argument
+    """The default library dispatch passes for ROCm backend."""
+    return [
+        relax.backend.DispatchSampling(),
+        relax.backend.DispatchSortScan(),
+    ]
+
+
+def legalize_passes(target: tvm.target.Target):  # pylint: disable=unused-argument
+    """The default legalization passes for ROCm backend."""
+    return [
+        tvm.relax.transform.LegalizeOps(),
+        tvm.relax.transform.AnnotateTIROpPattern(),
+        tvm.relax.transform.FoldConstant(),
+        tvm.relax.transform.FuseOps(),
+        tvm.relax.transform.FuseTIR(),
+        dl.ApplyDefaultSchedule(
+            dl.gpu.Matmul(),
+            dl.gpu.GEMV(),
+            dl.gpu.Reduction(),
+            dl.gpu.GeneralReduction(),
+            dl.gpu.Fallback(),
+        ),
+    ]
+
+
+def dataflow_lower_passes(target: tvm.target.Target):  # pylint: disable=unused-argument
+    """The default dataflow lowering passes for ROCm backend."""
+    return [
+        relax.transform.RewriteDataflowReshape(),
+        relax.transform.ToNonDataflow(),
+        relax.transform.RemovePurityChecking(),
+        relax.transform.CallTIRRewrite(),
+    ]
+
+
+def finalize_passes(target: tvm.target.Target):  # pylint: disable=unused-argument
+    """The default finalization passes for ROCm backend."""
+    return [
+        relax.transform.StaticPlanBlockMemory(),
+        relax.transform.LowerAllocTensor(),
+        relax.transform.KillAfterLastUse(),
+        relax.transform.LowerRuntimeBuiltin(),
+        relax.transform.ComputePrimValue(),
+        relax.transform.VMShapeLower(),
+        relax.transform.AttachGlobalSymbol(),
+    ]
+
+
+def get_default_pipeline(target: tvm.target.Target):
+    """Return the default compilation pipeline for ROCm."""
+
+    @tvm.transform.module_pass(opt_level=0)
+    def _pipeline(mod: tvm.ir.IRModule, _ctx: tvm.transform.PassContext):
+        with target:
+            seq = tvm.transform.Sequential(
+                library_dispatch_passes(target)
+                + legalize_passes(target)
+                + dataflow_lower_passes(target)
+                + finalize_passes(target)
+            )
+            mod = seq(mod)
+        return mod
+
+    return _pipeline

python/tvm/relax/pipeline.py

@@ -22,10 +22,11 @@
 """
 # pylint: disable=unused-argument
 from typing import Union
+
 import tvm
 from tvm import meta_schedule as ms
 
-from . import transform, backend
+from . import backend, transform
 
 
 def zero_pipeline(*, enable_warning: bool = False):
@@ -237,3 +238,76 @@ def _register(func):
         return func
 
     return _register
+
+
+def library_dispatch_passes(target: tvm.target.Target):
+    """Get the default library dispatch passes for the given target."""
+    if target.kind.name == "cuda":
+        return backend.cuda.library_dispatch_passes(target)
+    if target.kind.name == "rocm":
+        return backend.rocm.library_dispatch_passes(target)
+    if target.kind.name == "metal":
+        return backend.gpu_generic.library_dispatch_passes(target)
+    if target.kind.name == "llvm":
+        return backend.cpu_generic.library_dispatch_passes(target)
+    # Todo(tvm-team): support gpu-generic
+    raise ValueError(f"Target {target} is not yet supported by library dispatch passes.")
+
+
+def legalize_passes(target: tvm.target.Target):
+    """Get the default legalization passes for the given target."""
+    if target.kind.name == "cuda":
+        return backend.cuda.legalize_passes(target)
+    if target.kind.name == "rocm":
+        return backend.rocm.legalize_passes(target)
+    if target.kind.name == "metal":
+        return backend.gpu_generic.legalize_passes(target)
+    if target.kind.name == "llvm":
+        return backend.cpu_generic.legalize_passes(target)
+    # Todo(tvm-team): support gpu-generic
+    raise ValueError(f"Target {target} is not yet supported by library dispatch passes.")
+
+
+def dataflow_lower_passes(target: tvm.target.Target):
+    """Get the default legalization passes for the given target."""
+    if target.kind.name == "cuda":
+        return backend.cuda.dataflow_lower_passes(target)
+    if target.kind.name == "rocm":
+        return backend.rocm.dataflow_lower_passes(target)
+    if target.kind.name == "metal":
+        return backend.gpu_generic.dataflow_lower_passes(target)
+    if target.kind.name == "llvm":
+        return backend.cpu_generic.dataflow_lower_passes(target)
+    # Todo(tvm-team): support gpu-generic
+    raise ValueError(f"Target {target} is not yet supported by dataflow lowering passes.")
+
+
+def finalize_passes(target: tvm.target.Target):
+    """Get the default legalization passes for the given target."""
+    if target.kind.name == "cuda":
+        return backend.cuda.finalize_passes(target)
+    if target.kind.name == "rocm":
+        return backend.rocm.finalize_passes(target)
+    if target.kind.name == "metal":
+        return backend.gpu_generic.finalize_passes(target)
+    if target.kind.name == "llvm":
+        return backend.cpu_generic.finalize_passes(target)
+    # Todo(tvm-team): support gpu-generic
+    raise ValueError(f"Target {target} is not yet supported by finalization passes.")
+
+
+def get_default_pipeline(target: tvm.target.Target):
+    """Get the default Relax compilation pipeline for the given target."""
+    if target.kind.name == "cuda":
+        return backend.cuda.get_default_pipeline(target)
+    if target.kind.name == "rocm":
+        return backend.rocm.get_default_pipeline(target)
+    if target.kind.name == "metal":
+        return backend.gpu_generic.get_default_pipeline(target)
+    if target.kind.name == "llvm":
+        return backend.cpu_generic.get_default_pipeline(target)
+    # Todo(tvm-team): support gpu-generic
+    raise ValueError(
+        f"Target {target} is not yet supported by default pipeline. "
+        "Please lower and build the IRModule manually."
+    )

tests/python/relax/test_codegen_coreml.py

@@ -41,7 +41,7 @@ def _has_xcode():
 
 
 def verify(mod, inputs):
-    from tvm.relax.backend.contrib.coreml import partition_for_coreml
+    from tvm.relax.backend.metal.coreml import partition_for_coreml
 
     mod1 = partition_for_coreml(mod)
     mod1 = relax.transform.RunCodegen()(mod1)

tests/python/relax/test_codegen_cublas.py

@@ -21,7 +21,7 @@
 import tvm.testing
 import tvm.topi.testing
 from tvm import relax
-from tvm.relax.backend.contrib.cublas import partition_for_cublas
+from tvm.relax.backend.cuda.cublas import partition_for_cublas
 from tvm.relax.testing import get_relax_matmul_module
 from tvm.script import relax as R
 from tvm.script.ir_builder import IRBuilder