[Relay] A set of utilities that allows a model to be run efficiently …

…on tensorcores. (apache#6748)

python/tvm/relay/analysis/__init__.py

@@ -29,3 +29,6 @@
 # Feature
 from . import feature
 from . import sparse_dense
+
+# Utilities
+from .count_layers import count_layers

python/tvm/relay/analysis/count_layers.py

@@ -0,0 +1,68 @@
+# 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.
+"""Utilities that enable counting the number of layers in a graph."""
+import tvm
+from tvm import relay
+from ..expr_functor import ExprVisitor
+
+
+class LayerCounter(ExprVisitor):
+    """A visitor pass that computes the deepest chain of specified ops in graph."""
+
+    def __init__(self, valid_ops):
+        self.depth_count = 0
+        self.deepest_count = 0
+        self.valid_ops = [relay.op.get(op) for op in valid_ops]
+        super().__init__()
+
+    def visit_call(self, call):
+        if call.op in self.valid_ops:
+            self.depth_count += 1
+        current_count = self.depth_count
+        self.deepest_count = max(self.deepest_count, current_count)
+        for arg in call.args:
+            self.visit(arg)
+            self.depth_count = current_count
+
+    def count(self):
+        return self.deepest_count
+
+
+def count_layers(expr, valid_ops):
+    """Determine the number of layers of specified ops in a graph.
+    This pass computes only the deepest chain of ops rather than the
+    total number of ops in a graph. Thus, if there are two parallel
+    convolutions (for example), they would be considered a single layer.
+
+    Parameters
+    ----------
+    expr : tvm.relay.Expr, tvm.relay.Function, or tvm.ir.IRModule.
+        The input expression.
+
+    valid_ops: List[str]
+        A list of the operations that should be included in the count.
+
+    Returns
+    -------
+    layer_count : int
+        The number of layers of the specified operations found in the graph.
+    """
+    if isinstance(expr, tvm.ir.IRModule):
+        expr = expr["main"]
+    count_pass = LayerCounter(valid_ops)
+    count_pass.visit(expr)
+    return count_pass.count()

python/tvm/relay/op/nn/_nn.py

@@ -28,6 +28,7 @@
 from ..op import OpPattern
 from .._tensor import elemwise_shape_func
 from ..strategy.generic import is_depthwise_conv2d
+from ...transform import LayoutConfig
 
 # relu
 reg.register_broadcast_schedule("nn.relu")
@@ -164,6 +165,16 @@ def convert_conv2d(attrs, inputs, tinfos, desired_layouts):
     from tvm import relay
 
     data, weight = inputs
+
+    # First check if there is a LayoutConfig scope, and if so, whether
+    # it indicates we should ignore this layer or not.
+    layout_config = LayoutConfig.current
+    if layout_config is not None:
+        skip_layer = layout_config.check_skip()
+        if skip_layer:
+            return relay.nn.conv2d(data, weight, **attrs)
+
+    # Prepare new layout.
     new_attrs = dict(attrs)
     assert len(desired_layouts) == 2, "A desired layout is expected for both of nn.conv2d's inputs"
     desired_data_layout, desired_kernel_layout = map(str, desired_layouts)
@@ -192,6 +203,9 @@ def convert_conv2d(attrs, inputs, tinfos, desired_layouts):
         else:
             new_attrs["kernel_layout"] = "HWIO"
         return relay.nn.conv2d(data, weight, **new_attrs)
+    elif desired_data_layout == "HWNC":
+        new_attrs["kernel_layout"] = "HWOI"
+        return relay.nn.conv2d(data, weight, **new_attrs)
 
     raise ValueError("Layout %s is not yet supported." % desired_data_layout)
 

python/tvm/relay/op/strategy/cuda.py

@@ -219,8 +219,13 @@ def conv2d_strategy_cuda(attrs, inputs, out_type, target):
                 out_channels = oc_chunk * oc_block_factor
             else:
                 _, _, out_channels, _ = get_const_tuple(kernel.shape)
-            if topi.cuda.is_shape_tensorcore_direct_qualified(
-                batch=N, in_channels=in_channels, num_filter=out_channels, in_dtype=data.dtype
+
+            tensorcore_dtypes = ["int4", "uint4", "int8", "uint8"]
+            if (
+                (N % 16 == 0 and in_channels % 16 == 0 and out_channels % 16 == 0)
+                or (N % 8 == 0 and in_channels % 16 == 0 and out_channels % 32 == 0)
+                or (N % 32 == 0 and in_channels % 16 == 0 and out_channels % 8 == 0)
+                and (data.dtype in tensorcore_dtypes and kernel.dtype in tensorcore_dtypes)
             ):
                 strategy.add_implementation(
                     wrap_compute_conv2d(topi.cuda.conv2d_hwnc_tensorcore),

python/tvm/relay/transform/__init__.py

@@ -18,4 +18,5 @@
 """The Relay IR namespace containing transformations."""
 # transformation passes
 from .transform import *
+from .recast import recast
 from . import memory_alloc

python/tvm/relay/transform/recast.py

@@ -0,0 +1,139 @@
+# 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.
+"""Relay type recasting pass"""
+import tvm
+from tvm import relay
+from tvm.ir import IRModule
+from .transform import InferType
+from ..analysis import count_layers
+from ..expr_functor import ExprMutator, Call
+
+
+class RecastMutator(ExprMutator):
+    """Cast operations to the target type."""
+
+    def __init__(self, dtype, out_dtype, valid_ops, valid_op_count, skip_layers):
+        self.dtype = dtype
+        self.out_dtype = out_dtype
+        self.depth_count = 0
+        self.valid_ops = [relay.op.get(op) for op in valid_ops]
+        self.valid_op_count = valid_op_count
+        self.skip_layers = skip_layers
+        # Convert negative indices to positive ones.
+        for i, layer in enumerate(skip_layers):
+            if layer < 0:
+                skip_layers[i] = self.valid_op_count + layer
+        super().__init__()
+
+    def visit_call(self, call):
+        # Keep track of our current depth and layer count
+        # so we can know whether to skip this layer or not.
+        current_depth = self.depth_count
+        current_layer = self.valid_op_count - current_depth - 1
+        if call.op in self.valid_ops:
+            self.depth_count += 1
+        # Visit current call operation
+        new_fn = self.visit(call.op)
+        # Visit current arguments
+        args = []
+        for arg in call.args:
+            args.append(self.visit(arg))
+            self.depth_count = current_depth
+
+        # Downcast this op if its the correct type and not skipped.
+        if call.op in self.valid_ops and current_layer not in self.skip_layers:
+            # Recast inputs to specified type.
+            args = [self.visit(arg) for arg in call.args]
+            new_args = list()
+            for arg in args:
+                new_args.append(relay.cast(arg, dtype=self.dtype))
+
+            # If out_dtype is in the attributes, we need to update it.
+            orig_dtype = None
+            if "out_dtype" in call.attrs.keys():
+                new_attr_dict = {}
+                for attr in call.attrs.keys():
+                    attr_value = call.attrs[attr]
+                    if isinstance(attr_value, tvm.ir.container.Array):
+                        attr_value = tuple(attr_value)
+                    new_attr_dict[str(attr)] = attr_value
+                new_attr_dict["out_dtype"] = self.out_dtype
+                attr_type = str(call.attrs).split("(")[0]
+                new_attrs = tvm.ir.make_node(attr_type, **new_attr_dict)
+                if call.attrs["out_dtype"] != "":
+                    orig_dtype = call.attrs["out_dtype"]
+            else:
+                new_attrs = call.attrs
+
+            if orig_dtype is None:
+                # Perform type inference to determine the original type.
+                new_mod = IRModule.from_expr(call)
+                new_mod = InferType()(new_mod)
+                checked_arg = new_mod["main"].body
+                orig_dtype = checked_arg.checked_type.dtype
+            # Recast the output for compatibility with other graph operations.
+            return relay.cast(Call(new_fn, new_args, new_attrs), orig_dtype)
+
+        # Otherwise return the unchanged call.
+        return Call(new_fn, args, call.attrs)
+
+
+def recast(expr, dtype, out_dtype, ops=None, skip_layers=None):
+    """Convert the types of operations in a graph to a new value.
+    Note that this is primarily useful for testing performance of individual
+    operations at the new datatype. In a real setting, this pass will
+    almost certainly do a poor job converting from one datatype to another
+    as it just applies hard casting. For example, when recasting from float
+    to integer, many small values will simply be set to 0. Although this will
+    allow autotuning and benchmarking to produce proper timings at the new
+    data type, the output of the model will of course be heavily impacted.
+
+    Parameters
+    ---------
+    expr: tvm.relay.Expr, tvm.relay.Function, or tvm.ir.IRModule
+        The original function that will have its type changed.
+    dtype: str
+        The target type to cast to.
+    out_dtype: str
+        The output type to cast to.
+    ops: List[str]
+        A list of operations that should have their type changed,
+        others will be left as is.
+    skip_layers: List[int]
+        A list of integers indicating operations that should
+        not have their type changed, counted starting with the
+        first valid operation encountered. Negative indices are
+        allowed and indicate starting at the last layer.
+    Returns
+    -------
+    output_expr : tvm.relay.Expr, tvm.relay.Function, or tvm.ir.IRModule
+        The graph after recasting to the specified datatype.
+    """
+    return_mod = False
+    if isinstance(expr, tvm.ir.IRModule):
+        expr = expr["main"]
+        return_mod = True
+    if ops is None:
+        ops = ["nn.conv2d"]
+    if skip_layers is None:
+        skip_layers = []
+    layer_depth = count_layers(expr, ops)
+    recast_pass = RecastMutator(dtype, out_dtype, ops, layer_depth, skip_layers)
+    expr = recast_pass.visit(expr)
+    if return_mod:
+        return tvm.IRModule.from_expr(expr)
+    return expr

python/tvm/relay/transform/transform.py

@@ -386,6 +386,33 @@ def AlterOpLayout():
     return _ffi_api.AlterOpLayout()
 
 
+class LayoutConfig(object):
+    """A structure for customizing the ConvertLayout pass."""
+
+    current = None
+
+    def __init__(self, skip_layers=None):
+        self.skip_counter = 0
+        self.skip_layers = skip_layers if skip_layers is not None else []
+
+    def check_skip(self):
+        skip = self.skip_counter in self.skip_layers
+        self.skip_counter += 1
+        return skip
+
+    def reset(self):
+        self.skip_counter = 0
+        self.skip_layers = []
+
+    def __enter__(self):
+        self._old_manager = LayoutConfig.current
+        LayoutConfig.current = self
+        return self
+
+    def __exit__(self, ptype, value, trace):
+        LayoutConfig.current = self._old_manager
+
+
 def ConvertLayout(desired_layouts):
     """Given a dest layout, this pass transforms the expr such that most of the ops input data
     layout is changed to the dest layout. In ideal situation, there are only 2 layout transforms,

src/runtime/graph/graph_runtime.cc

@@ -310,7 +310,7 @@ void GraphRuntime::SetupStorage() {
     ICHECK_GE(storage_id, 0) << "Do not support runtime shape op";
     DLDataType t = vtype[i];
     size_t bits = t.bits * t.lanes;
-    ICHECK(bits % 8U == 0U || bits == 1U);
+    ICHECK(bits % 8U == 0U || bits == 1U || bits == 4U);
     size_t bytes = ((bits + 7U) / 8U) * size;
 
     uint32_t sid = static_cast<uint32_t>(storage_id);

tests/python/relay/test_layer_count.py

@@ -0,0 +1,34 @@
+# 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.
+from tvm.relay.testing import resnet
+from tvm.relay.analysis import count_layers
+
+
+def test_layer_count():
+    def verify(num_layers):
+        # Load a resnet with a known number of layers.
+        mod, _ = resnet.get_workload(num_layers=num_layers)
+        # Count the number of conv and dense layers.
+        count = count_layers(mod, valid_ops=["nn.conv2d", "nn.dense"])
+        assert count == num_layers
+
+    verify(18)
+    verify(50)
+
+
+if __name__ == "__main__":
+    test_layer_count()