[VM] Per-input, data dependence specification for shape func (apache#…

…7210)

* made TShapeDataDependant array

* add stub

* dyn strided slice working

* reshape also working

* remove log

* works on maskrcnn

* lint fix

* fix cpp test

* remove stale pop back

* add more doc

* dependant -> dependent

* remove redundant check

* remove data_dependent_

include/tvm/relay/op_attr_types.h

@@ -83,9 +83,9 @@ using TOpIsStateful = bool;
 using TNonComputational = bool;
 
 /*!
- * \brief Mark the operator whether output shape is data dependant.
+ * \brief Mark the operator whether output shape is data dependent.
  */
-using TShapeDataDependant = bool;
+using TShapeDataDependent = Array<Integer>;
 
 /*!
  * \brief Computation description interface.

python/tvm/relay/op/dyn/_transform.py

@@ -32,11 +32,8 @@
 
 
 @script
-def _reshape_shape_func_input_data(data, newshape, ndim):
+def _reshape_shape_func_input_data(data_shape, newshape, ndim):
     out = output_tensor((ndim,), "int64")
-    data_shape = allocate((len(data.shape),), "int64")
-    for x in const_range(len(data.shape)):
-        data_shape[x] = int64(data.shape[x])
     src_idx = 0
     dst_idx = 0
     infer_idx = -1
@@ -87,7 +84,7 @@ def _reshape_shape_func_input_data(data, newshape, ndim):
     return out
 
 
-@_reg.register_shape_func("dyn.reshape", True)
+@_reg.register_shape_func("dyn.reshape", [False, True])
 def dynamic_reshape_shape_func(attrs, inputs, out_ndims):
     return [_reshape_shape_func_input_data(*inputs, out_ndims[0])]
 
@@ -150,36 +147,36 @@ def one_hot_shape_func(attrs, inputs, _):
 
 
 @script
-def _strided_slice_shape_func_input_data(data, begin, end, strides, slice_mode):
-    ndim = len(data.shape)
+def _strided_slice_shape_func_input_data(data_shape, begin, end, strides, slice_mode):
+    ndim = len(data_shape)
     out = output_tensor((ndim,), "int64")
     for i in const_range(ndim):
         cbegin = int64(0)
-        cend = int64(data.shape[i])
+        cend = int64(data_shape[i])
         cstride = int64(1)
         if strides.shape[0] > i:
             cstride = int64(strides[i])
         if begin.shape[0] > i:
             cbegin = int64(begin[i])
             if cbegin < 0:
-                cbegin += int64(data.shape[i])
+                cbegin += int64(data_shape[i])
         if end.shape[0] <= i:
-            cend = int64(data.shape[i])
+            cend = int64(data_shape[i])
         elif slice_mode != 0:
             cstride = int64(1)
             if end[i] < 0:
-                cend = int64(data.shape[i])
+                cend = int64(data_shape[i])
             else:
                 cend = cbegin + int64(end[i])
         else:
-            if end[i] > data.shape[i]:
-                cend = int64(data.shape[i])
-            elif end[i] < -data.shape[i]:
+            if end[i] > data_shape[i]:
+                cend = int64(data_shape[i])
+            elif end[i] < -data_shape[i]:
                 cend = int64(-1)
             else:
                 cend = int64(end[i])
                 if cend < 0:
-                    cend += int64(data.shape[i])
+                    cend += int64(data_shape[i])
         assert cstride != 0, "Strides can't be zero."
         if cstride < 0:
             slice_range = cbegin - cend
@@ -192,7 +189,7 @@ def _strided_slice_shape_func_input_data(data, begin, end, strides, slice_mode):
     return out
 
 
-@_reg.register_shape_func("dyn.strided_slice", True)
+@_reg.register_shape_func("dyn.strided_slice", [False, True, True, True])
 def strided_slice_shape_func(attrs, inputs, _):
     """
     Shape func for strided_slice

python/tvm/relay/op/op.py

@@ -356,16 +356,18 @@ def register_gradient(op_name, fgradient=None, level=10):
     return tvm.ir.register_op_attr(op_name, "FPrimalGradient", fgradient, level)
 
 
-def register_shape_func(op_name, data_dependant, shape_func=None, level=10):
+def register_shape_func(op_name, data_dependent, shape_func=None, level=10):
     """Register operator shape function for an op.
 
     Parameters
     ----------
     op_name : str
         The name of the op.
 
-    data_dependant : bool
-        Whether the shape function depends on input data.
+    data_dependent : bool or list of bool
+        Whether the shape function depends on input data. If this is a list of bool,
+        the length of the list must be the same as the number of arguments of this op.
+        The list specifies per-input data dependence of the op.
 
     shape_func : function (attrs: Attrs, inputs: List[Tensor], out_ndims: List[IndexExpr])
                  -> shape_tensors: List<Tensor>
@@ -374,7 +376,9 @@ def register_shape_func(op_name, data_dependant, shape_func=None, level=10):
     level : int
         The priority level
     """
-    get(op_name).set_attr("TShapeDataDependant", data_dependant, level)
+    if not isinstance(data_dependent, list):
+        data_dependent = [data_dependent]
+    get(op_name).set_attr("TShapeDataDependent", data_dependent, level)
     return tvm.ir.register_op_attr(op_name, "FShapeFunc", shape_func, level)
 
 

src/relay/analysis/util.cc

@@ -473,24 +473,27 @@ bool IsDynamic(const Type& ty) {
 
 TVM_REGISTER_GLOBAL("relay.ir.IsDynamic").set_body_typed(IsDynamic);
 
-bool IsDataDependant(const CallNode* call) {
-  static auto tshape_data_dependant = Op::GetAttrMap<TShapeDataDependant>("TShapeDataDependant");
+bool IsDataDependent(const CallNode* call) {
+  static auto tshape_data_dependent = Op::GetAttrMap<TShapeDataDependent>("TShapeDataDependent");
   Op op = Downcast<Op>(call->op);
 
-  if (!tshape_data_dependant.count(op)) {
+  if (!tshape_data_dependent.count(op)) {
     return false;
   }
 
   if (op->name == "strided_slice") {
     if (const auto* attrs = call->attrs.as<StridedSliceAttrs>()) {
       if (attrs->begin && attrs->end && attrs->strides) {
-        // not data dependant if begin, end and strides exist
+        // not data dependent if begin, end and strides exist
         return false;
       }
     }
   }
 
-  return tshape_data_dependant[op];
+  for (auto req : tshape_data_dependent[op]) {
+    if (req->value != 0) return true;
+  }
+  return false;
 }
 }  // namespace relay
 }  // namespace tvm

src/relay/backend/compile_engine.cc

@@ -435,9 +435,9 @@ class MakeShapeFunc : public backend::MemoizedExprTranslator<Array<te::Tensor>>
       LOG(FATAL) << "Free variable " << var->name_hint();
       return {};
     } else {
-      ICHECK(data_dependants_.size());
-      bool data_dependant = data_dependants_.back();
-      if (data_dependant) {
+      ICHECK(data_dependents_per_input_.size());
+      auto data_dependent = data_dependents_per_input_.back();
+      if (data_dependent) {
         param_states_[var] |= kNeedInputData;
         return param_data_[var];
       } else {
@@ -449,12 +449,12 @@ class MakeShapeFunc : public backend::MemoizedExprTranslator<Array<te::Tensor>>
 
   Array<te::Tensor> VisitExpr_(const ConstantNode* op) final {
     using tir::make_const;
-    ICHECK(data_dependants_.size());
-    bool data_dependant = data_dependants_.back();
+    ICHECK(data_dependents_per_input_.size());
+    bool data_dependent = data_dependents_per_input_.back();
     if (!op->is_scalar()) {
       // This is a constant weight, extract the shape of the weight tensor.
       // This can not be data dependent.
-      CHECK(!data_dependant);
+      CHECK(!data_dependent);
       auto ttype = op->checked_type().as<TensorTypeNode>();
       int ndim = static_cast<int>(ttype->shape.size());
       Array<PrimExpr> out_shape{ndim};
@@ -472,7 +472,7 @@ class MakeShapeFunc : public backend::MemoizedExprTranslator<Array<te::Tensor>>
       scalars_.push_back(value);
       return {value};
     }
-    if (data_dependant) {
+    if (data_dependent) {
       void* data = op->data->data;
       DataType dtype = DataType(op->data->dtype);
       auto value = tvm::te::compute(
@@ -507,27 +507,38 @@ class MakeShapeFunc : public backend::MemoizedExprTranslator<Array<te::Tensor>>
 
   Array<te::Tensor> VisitExpr_(const CallNode* call_node) final {
     static auto fshape_func = Op::GetAttrMap<FShapeFunc>("FShapeFunc");
-    static auto tshape_data_dependant = Op::GetAttrMap<TShapeDataDependant>("TShapeDataDependant");
+    static auto tshape_data_dependent = Op::GetAttrMap<TShapeDataDependent>("TShapeDataDependent");
     ICHECK(call_node->op.as<OpNode>()) << "Primitive function only allows call into primitive ops";
     Op op = Downcast<Op>(call_node->op);
-    ICHECK(data_dependants_.empty() || !data_dependants_.back())
+    ICHECK(data_dependents_per_input_.empty() || !data_dependents_per_input_.back())
         << "Error in op fusion: output of the shape func is fed to a "
-        << "data-dependant shape func";
+        << "data-dependent shape func";
     ICHECK_GT(fshape_func.count(op), 0) << "Internal error, cannot find ShapeFunc for " << op->name;
-    ICHECK_GT(tshape_data_dependant.count(op), 0)
-        << "Internal error, cannot find TShapeDataDependant for " << op->name;
+    ICHECK_GT(tshape_data_dependent.count(op), 0)
+        << "Internal error, cannot find TShapeDataDependent for " << op->name;
+
+    Array<Integer> dep_spec = tshape_data_dependent[op];
+    if (dep_spec.size() == 1) {
+      // This is for cases when data dependence is specified per op
+      // Replicate 0 or 1 flag to all arguments
+      for (size_t i = 1; i < call_node->args.size(); ++i) {
+        dep_spec.push_back(dep_spec[0]);
+      }
+    }
 
-    data_dependants_.push_back(IsDataDependant(call_node));
     // Visit all inputs
     Array<te::Tensor> inputs;
     int count_tuple = 0;
-    for (Expr arg : call_node->args) {
+    for (size_t i = 0; i < call_node->args.size(); ++i) {
+      Expr arg = call_node->args[i];
       if (arg->checked_type().as<TupleTypeNode>()) {
         ++count_tuple;
       }
+      data_dependents_per_input_.push_back(dep_spec[i]->value != 0);
       for (te::Tensor tensor : VisitExpr(arg)) {
         inputs.push_back(tensor);
       }
+      data_dependents_per_input_.pop_back();
     }
     if (count_tuple) {
       ICHECK_EQ(call_node->args.size(), 1U) << "Only allow function with a single tuple input";
@@ -549,7 +560,6 @@ class MakeShapeFunc : public backend::MemoizedExprTranslator<Array<te::Tensor>>
     }
     // Call shape function
     auto outputs = fshape_func[op](call_node->attrs, inputs, out_ndims);
-    data_dependants_.pop_back();
     readable_name_stream_ << "_" << op->name;
     return outputs;
   }
@@ -593,8 +603,8 @@ class MakeShapeFunc : public backend::MemoizedExprTranslator<Array<te::Tensor>>
   std::unordered_map<Expr, Array<te::Tensor>, ObjectPtrHash, ObjectPtrEqual> param_data_;
   /*! \brief Map from parameter to list of shape placeholder */
   std::unordered_map<Expr, Array<te::Tensor>, ObjectPtrHash, ObjectPtrEqual> param_shapes_;
-  /*! \brief Stack of data dependencies for shape function */
-  std::vector<bool> data_dependants_;
+  /*! \brief Stack of data dependencies for shape function, specified per each op input */
+  std::vector<bool> data_dependents_per_input_;
   /*! \brief Scalars used in the shape function */
   Array<te::Tensor> scalars_;
 };

src/relay/transforms/fuse_ops.cc

@@ -241,7 +241,7 @@ class IndexedForwardGraph::Creator : private ExprVisitor {
     OpPatternKind op_pattern = kOpaque;
     if (const OpNode* opnode = call->op.as<OpNode>()) {
       auto op = GetRef<Op>(opnode);
-      if (IsDynamic(call->checked_type()) && IsDataDependant(call)) {
+      if (IsDynamic(call->checked_type()) && IsDataDependent(call)) {
         // output of a shape func can't be fed to a data-dependent shape func
         op_pattern = kOpaque;
       } else {

src/relay/transforms/pass_utils.h

@@ -90,11 +90,11 @@ Expr TypeSubst(const Expr& expr, const tvm::Map<TypeVar, Type>& subst_map);
 bool IsDynamic(const Type& ty);
 
 /*!
- * \brief Check if call is data dependant.
+ * \brief Check if call is data dependent.
  * \param call The call to be checked.
- * \return Whether the call is data dependant.
+ * \return Whether the call is data dependent.
  */
-bool IsDataDependant(const CallNode* call);
+bool IsDataDependent(const CallNode* call);
 
 /*!
  * \brief Make arbitrary transformation preserve the out most function.

tests/cpp/relay_build_module_test.cc

@@ -105,7 +105,9 @@ TEST(Relay, BuildModule) {
   }
   auto fgeneric = GenericFunc::Get("test.strategy_generic").set_default(*fs);
   (*reg)("add", "FTVMStrategy", fgeneric, 10);
-  (*reg)("add", "TShapeDataDependant", false, 10);
+  Array<Integer> dep;
+  dep.push_back(0);
+  (*reg)("add", "TShapeDataDependent", dep, 10);
   // build
   auto pfb = tvm::runtime::Registry::Get("relay.build_module._BuildModule");
   tvm::runtime::Module build_mod = (*pfb)();