[Relay] Dense alter layout fixed for packed input (apache#8669)

* clean up typerel

* add layout transform when input is 3D

* add test

* update doc to clarify that only 2D input data is supported

* add weight_layout attribute in dense

* remove explicit layout transform from dense_alter_op.py

* Add DensePackInferCorrectLayout to insert layout transform

* relax type rel

* revert type rel relax and add check on dim

* introduce DensePackAttrs to avoid breaking dense op

* try fixing arm compute lib test

* Update tests/python/contrib/test_arm_compute_lib/test_dense.py

Co-authored-by: lhutton1 <35535092+lhutton1@users.noreply.github.com>

* formatting

Co-authored-by: lhutton1 <35535092+lhutton1@users.noreply.github.com>

Author: 2 people

include/tvm/relay/attrs/nn.h

@@ -1003,6 +1003,25 @@ struct DenseAttrs : public tvm::AttrsNode<DenseAttrs> {
   }
 };
 
+/*! \brief Attributes for dense_pack operator */
+struct DensePackAttrs : public tvm::AttrsNode<DensePackAttrs> {
+  IndexExpr units;
+  DataType out_dtype;
+  tvm::String weight_layout;
+
+  TVM_DECLARE_ATTRS(DensePackAttrs, "relay.attrs.DensePackAttrs") {
+    TVM_ATTR_FIELD(units).describe("Number of hidden units of the dense transformation.");
+
+    // use 0 bits to indicate none.
+    TVM_ATTR_FIELD(out_dtype)
+        .set_default(NullValue<DataType>())
+        .describe("Output data type, set to explicit type under mixed precision setting");
+    TVM_ATTR_FIELD(weight_layout)
+        .set_default("NK")
+        .describe("Dimension ordering of weight. Packed layouts, such as NK8n, are possible.");
+  }
+};
+
 /*! \brief Attributes for batch matmul operator. */
 struct BatchMatmulAttrs : public tvm::AttrsNode<BatchMatmulAttrs> {
   DataType out_dtype;

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

@@ -1259,9 +1259,9 @@ def dense_shape_func(attrs, inputs, _):
 @script
 def _dense_pack_shape_func(data_shape, weight_shape):
     out = output_tensor((data_shape.shape[0],), "int64")
-    for i in const_range(out.shape[0] - 1):
-        out[i] = data_shape[i]
-    out[out.shape[0] - 1] = weight_shape[0] * weight_shape[2]
+    assert data_shape.shape[0] == 2, "Input data must be 2D"
+    out[0] = data_shape[0]
+    out[1] = weight_shape[0] * weight_shape[2]
 
     return out
 

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

@@ -1548,9 +1548,9 @@ def dense(data, weight, units=None, out_dtype=""):
     return _make.dense(data, weight, units, out_dtype)
 
 
-def contrib_dense_pack(data, weight, units=None, out_dtype=""):
+def contrib_dense_pack(data, weight, weight_layout="NK", units=None, out_dtype=""):
     """Dense operator.
-    Applies a linear transformation
+    Applies a linear transformation with packed weight
 
     .. math::
 
@@ -1560,25 +1560,27 @@ def contrib_dense_pack(data, weight, units=None, out_dtype=""):
     ----------
     data : tvm.relay.Expr
         The input data to the operator,
-        of shape `(d_1, d_2, ..., d_n, units_in)`.
+        of shape `(batch, units_in)`.
 
     weight : tvm.relay.Expr
         The transformed weight expressions, 3-D matrix,
         of shape `(units // pack_weight_tile, units_in, pack_weight_tile)`.
 
+    weight_layout: str
+        The layout of weight, such as "NK" or "NK8n".
+
     units : int, optional
         Number of hidden units of the dense transformation.
 
     out_dtype : str, optional
-        Specifies the output data type for mixed precision dense,
-        of shape `(d_1, d_2, ..., d_n, units)`.
+        Specifies the output data type for mixed precision dense.
 
     Returns
     -------
     result : tvm.relay.Expr
         The computed result.
     """
-    return _make.contrib_dense_pack(data, weight, units, out_dtype)
+    return _make.contrib_dense_pack(data, weight, weight_layout, units, out_dtype)
 
 
 def fifo_buffer(data, buffer, axis):

python/tvm/topi/x86/dense_alter_op.py

@@ -63,6 +63,7 @@ def _alter_dense_layout(attrs, inputs, tinfos, out_type):
         relay.op.get("nn.dense"), attrs, tinfos, out_type, target
     )
     workload = autotvm.task.get_workload(outs)
+
     if workload:
         cfg = dispatch_ctx.query(target, workload)
         topi_impl = workload[0]
@@ -86,7 +87,6 @@ def _alter_dense_layout(attrs, inputs, tinfos, out_type):
                 topi_impl,
             )
             dispatch_ctx.update(target, new_workload, cfg)
-            weight_transform = relay.layout_transform(inputs[1], "NK", weight_layout)
-            return relay.nn.contrib_dense_pack(inputs[0], weight_transform, None, out_dtype)
+            return relay.nn.contrib_dense_pack(inputs[0], inputs[1], weight_layout, None, out_dtype)
 
     return None

src/relay/op/nn/nn.cc

@@ -206,6 +206,13 @@ Expr MakeDense(Expr data, Expr weight, IndexExpr units, DataType out_dtype) {
   return Call(op, {data, weight}, Attrs(attrs), {});
 }
 
+InferCorrectLayoutOutput DenseInferCorrectLayout(const Attrs& attrs,
+                                                 const Array<Layout>& new_in_layouts,
+                                                 const Array<Layout>& old_in_layouts,
+                                                 const Array<tvm::relay::Type>& old_in_types) {
+  return InferCorrectLayoutOutput({"NC", "NK"}, {"NC"}, attrs);
+}
+
 TVM_REGISTER_GLOBAL("relay.op.nn._make.dense").set_body_typed(MakeDense);
 
 RELAY_REGISTER_OP("nn.dense")
@@ -221,35 +228,75 @@ RELAY_REGISTER_OP("nn.dense")
     .add_argument("data", "nD Tensor", "Input data.")
     .add_argument("weight", "2D Tensor", "Weight matrix.")
     .set_support_level(1)
+    .set_attr<FInferCorrectLayout>("FInferCorrectLayout", DenseInferCorrectLayout)
     .add_type_rel("Dense", MatmulRel<DenseAttrs>);
 // ------------------- relay.nn.dense
 
 // ------------------- relay.nn.contrib_dense_pack
+TVM_REGISTER_NODE_TYPE(DensePackAttrs);
+
 // Positional relay function to create dense_pack operator used by frontend FFI.
-Expr MakeDensePack(Expr data, Expr weight, IndexExpr units, DataType out_dtype) {
-  auto attrs = make_object<DenseAttrs>();
+Expr MakeDensePack(Expr data, Expr weight, tvm::String weight_layout, IndexExpr units,
+                   DataType out_dtype) {
+  auto attrs = make_object<DensePackAttrs>();
   attrs->units = units;
   attrs->out_dtype = out_dtype;
+  attrs->weight_layout = std::move(weight_layout);
   static const Op& op = Op::Get("nn.contrib_dense_pack");
   return Call(op, {data, weight}, Attrs(attrs), {});
 }
 
 TVM_REGISTER_GLOBAL("relay.op.nn._make.contrib_dense_pack").set_body_typed(MakeDensePack);
 
+bool DensePackRel(const Array<Type>& types, int num_inputs, const Attrs& attrs,
+                  const TypeReporter& reporter) {
+  ICHECK_EQ(types.size(), 3);
+  const auto* data = types[0].as<TensorTypeNode>();
+  const auto* weight = types[1].as<TensorTypeNode>();
+  if (data == nullptr || weight == nullptr) return false;
+
+  const DensePackAttrs* param = attrs.as<DensePackAttrs>();
+  ICHECK(param != nullptr);
+
+  ICHECK_EQ(data->shape.size(), 2) << "Only 2D data is supported";
+  ICHECK_EQ(weight->shape.size(), 3) << "Weight is not packed";
+
+  Array<tvm::PrimExpr> oshape = data->shape;
+  oshape.Set(1, weight->shape[0] * weight->shape[2]);
+
+  DataType out_dtype = param->out_dtype;
+  if (out_dtype.bits() == 0) {
+    out_dtype = data->dtype;
+  }
+  // assign output type
+  reporter->Assign(types[2], TensorType(oshape, out_dtype));
+  return true;
+}
+
+InferCorrectLayoutOutput DensePackInferCorrectLayout(const Attrs& attrs,
+                                                     const Array<Layout>& new_in_layouts,
+                                                     const Array<Layout>& old_in_layouts,
+                                                     const Array<tvm::relay::Type>& old_in_types) {
+  auto params = attrs.as<DensePackAttrs>();
+  ICHECK(params);
+  return InferCorrectLayoutOutput({"NC", params->weight_layout}, {"NC"}, attrs);
+}
+
 RELAY_REGISTER_OP("nn.contrib_dense_pack")
     .describe(R"code(Applies a linear transformation: :math:`Y = XW^T`.
 
-- **data**: `(x1, x2, ..., xn, input_dim)`
+- **data**: `(batch, input_dim)`
 - **weight**: `(units // pack_weight_tile, input_dim, pack_weight_tile)`
-- **out**: `(x1, x2, ..., xn, units)`.
+- **out**: `(batch, units)`.
 
 )code" TVM_ADD_FILELINE)
     .set_attrs_type<DenseAttrs>()
     .set_num_inputs(2)
-    .add_argument("data", "nD Tensor", "Input data.")
+    .add_argument("data", "2D Tensor", "Input data.")
     .add_argument("weight", "3D Tensor", "Packed weight matrix.")
     .set_support_level(10)
-    .add_type_rel("DensePack", DensePackRel<DenseAttrs>);
+    .set_attr<FInferCorrectLayout>("FInferCorrectLayout", DensePackInferCorrectLayout)
+    .add_type_rel("DensePack", DensePackRel);
 // ------------------- relay.nn.contrib_dense_pack
 
 // relay.leaky_relu
@@ -307,7 +354,6 @@ bool PReluRel(const Array<Type>& types, int num_inputs, const Attrs& attrs,
   return true;
 }
 
-template <typename T>
 InferCorrectLayoutOutput PReluInferCorrectLayout(const Attrs& attrs,
                                                  const Array<Layout>& new_in_layouts,
                                                  const Array<Layout>& old_in_layouts,
@@ -343,7 +389,7 @@ where :math:`*` is an channelwise multiplication for each sample in the batch.
     .add_argument("alpha", "Tensor", "Input channelwise alpha.")
     .set_support_level(3)
     .add_type_rel("PRelu", PReluRel)
-    .set_attr<FInferCorrectLayout>("FInferCorrectLayout", PReluInferCorrectLayout<PReluAttrs>)
+    .set_attr<FInferCorrectLayout>("FInferCorrectLayout", PReluInferCorrectLayout)
     .set_attr<FTVMCompute>("FTVMCompute", [](const Attrs& attrs, const Array<te::Tensor>& inputs,
                                              const Type& out_type) {
       const auto* param = attrs.as<PReluAttrs>();

src/relay/op/nn/nn.h

@@ -116,29 +116,6 @@ bool MatmulRel(const Array<Type>& types, int num_inputs, const Attrs& attrs,
   return true;
 }
 
-template <typename AttrType>
-bool DensePackRel(const Array<Type>& types, int num_inputs, const Attrs& attrs,
-                  const TypeReporter& reporter) {
-  ICHECK_EQ(types.size(), 3);
-  const auto* data = types[0].as<TensorTypeNode>();
-  const auto* weight = types[1].as<TensorTypeNode>();
-  if (data == nullptr || weight == nullptr) return false;
-
-  const AttrType* param = attrs.as<AttrType>();
-  ICHECK(param != nullptr);
-
-  Array<tvm::PrimExpr> oshape = data->shape;
-  oshape.Set((oshape.size() - 1), weight->shape[0] * weight->shape[2]);
-
-  DataType out_dtype = param->out_dtype;
-  if (out_dtype.bits() == 0) {
-    out_dtype = data->dtype;
-  }
-  // assign output type
-  reporter->Assign(types[2], TensorType(oshape, out_dtype));
-  return true;
-}
-
 template <typename AttrType>
 bool BatchMatmulRel(const Array<Type>& types, int num_inputs, const Attrs& attrs,
                     const TypeReporter& reporter) {

tests/python/contrib/test_arm_compute_lib/test_dense.py

@@ -150,11 +150,16 @@ def _get_expected_codegen(shape, weight_shape, units, dtype, has_bias=False):
 
     if has_bias:
         bias_dtype = "int32" if dtype == "uint8" else "float32"
+        bias_shape = (
+            [1, weight_shape[0]]
+            if dtype == "float32" and weight_shape[0] != 1
+            else [weight_shape[0]]
+        )
         inputs.append(
             {
                 "op": "const",
                 "name": "",
-                "attrs": {"shape": [[[weight_shape[0]]]], "dtype": [[bias_dtype]]},
+                "attrs": {"shape": [[bias_shape]], "dtype": [[bias_dtype]]},
             }
         )
 

tests/python/relay/test_pass_alter_op_layout.py

@@ -1315,7 +1315,9 @@ def expected():
         weight = relay.var("weight", shape=(48, 64))
         target_layout = "NK16n"
         weight_transform = relay.layout_transform(weight, "NK", target_layout)
-        y = relay.nn.contrib_dense_pack(x, weight_transform, units=None, out_dtype="float32")
+        y = relay.nn.contrib_dense_pack(
+            x, weight_transform, target_layout, units=None, out_dtype="float32"
+        )
         y = relay.Function(analysis.free_vars(y), y)
         return y
 
@@ -1353,6 +1355,49 @@ def alter_conv2d(attrs, inputs, tinfos, out_type):
         assert before.body.attrs.layout == "NCHW"
 
 
+def test_alter_op_dense_packed_data():
+    def before():
+        x = relay.var("x", shape=(1, 32, 8, 8))
+        weight = relay.var("conv2d_weight", shape=(32, 32, 3, 3))
+        conv = relay.nn.conv2d(x, weight, channels=32, kernel_size=(3, 3), padding=(1, 1))
+        pool = relay.nn.avg_pool2d(conv, pool_size=[8, 8], padding=[0, 0, 0, 0])
+        squeeze = relay.squeeze(pool, axis=[2, 3])
+        dense = relay.nn.dense(squeeze, relay.var("dense_weight", shape=(16, 32)))
+        return relay.Function(analysis.free_vars(dense), dense)
+
+    def expected():
+        x = relay.var("x", shape=(1, 32, 8, 8))
+        conv_weight = relay.var("conv2d_weight", shape=(32, 32, 3, 3))
+        dense_weight = relay.var("dense_weight", shape=(16, 32))
+        conv = relay.nn.contrib_conv2d_nchwc(
+            relay.layout_transform(x, "NCHW", "NCHW8c"),
+            relay.layout_transform(conv_weight, "OIHW", "OIHW8i8o"),
+            channels=32,
+            kernel_size=(3, 3),
+            padding=(1, 1),
+            data_layout="NCHW8c",
+            kernel_layout="OIHW8i8o",
+            out_layout="NCHW8c",
+        )
+        pool = relay.nn.avg_pool2d(conv, pool_size=[8, 8], padding=[0, 0, 0, 0], layout="NCHW8c")
+        squeeze = relay.squeeze(pool, axis=[2, 3])
+        dense = relay.nn.contrib_dense_pack(
+            relay.layout_transform(squeeze, "NC8c", "NC"),
+            relay.layout_transform(dense_weight, "NK", "NK16n"),
+            "NK16n",
+            out_dtype="float32",
+        )
+        return relay.Function(analysis.free_vars(dense), dense)
+
+    with tvm.target.Target("llvm"):
+        with TempOpAttr(
+            "nn.dense", "FTVMAlterOpLayout", topi.x86.dense_alter_op._alter_dense_layout
+        ):
+            a = run_opt_pass(before(), transform.AlterOpLayout())
+            b = run_opt_pass(expected(), transform.InferType())
+            assert tvm.ir.structural_equal(a, b)
+
+
 if __name__ == "__main__":
     test_alter_op()
     test_alter_return_none()
@@ -1377,3 +1422,4 @@ def alter_conv2d(attrs, inputs, tinfos, out_type):
     test_alter_op_dense()
     test_alter_layout_strided_slice_axes_nhwc()
     test_not_inplace_modify()
+    test_alter_op_dense_packed_data()