[Relay][Pass] Support combine multiple dense op just into dense

include/tvm/relay/transform.h

@@ -223,10 +223,11 @@ TVM_DLL Pass CombineParallelConv2D(uint64_t min_num_branches = 3);
  * `min_num_branch`.
  *
  * \param min_num_branches The minimun number of branches.
+ * \param to_batch Combine matmuls to batch matmul.
  *
  * \return The pass.
  */
-TVM_DLL Pass CombineParallelDense(uint64_t min_num_branches = 3);
+TVM_DLL Pass CombineParallelDense(uint64_t min_num_branches = 3, bool to_batch = true);
 
 /*!
  * \brief Combine parallel batch_matmul ops into a single batch_matmul

python/tvm/relay/transform/transform.py

@@ -277,7 +277,7 @@ def CombineParallelConv2D(min_num_branches=3):
     return _ffi_api.CombineParallelConv2D(min_num_branches)
 
 
-def CombineParallelDense(min_num_branches=3):
+def CombineParallelDense(min_num_branches=3, to_batch=True):
     """Combine multiple dense operators into one. For example:
 
     .. code-block
@@ -295,18 +295,29 @@ def CombineParallelDense(min_num_branches=3):
                 |
             batch_matmul+elemwise/bcast (2,2,2)
 
+    or (if to_batch=False)
+
+    .. code-block
+
+                data
+                |
+            dense+elemwise/bcast (2,2+2)
+
     Parameters
     ----------
     min_num_branches : int
         The minimum number of required parallel branches for performing this
         optimization.
 
+    to_batch : bool
+        Whether convert multiple dense into batch_matmul.
+
     Returns
     -------
     ret: tvm.transform.Pass
         The registered pass that combines parallel dense operators.
     """
-    return _ffi_api.CombineParallelDense(min_num_branches)
+    return _ffi_api.CombineParallelDense(min_num_branches, to_batch)
 
 def CombineParallelBatchMatmul(min_num_branches=3):
     """Combine multiple batch matmul operators into one. For example:

src/relay/backend/build_module.cc

@@ -277,7 +277,7 @@ class RelayBuildModule : public runtime::ModuleNode {
     });
     pass_seqs.push_back(transform::EliminateCommonSubexpr(fskip));
     pass_seqs.push_back(transform::CombineParallelConv2D(3));
-    pass_seqs.push_back(transform::CombineParallelDense(3));
+    pass_seqs.push_back(transform::CombineParallelDense(3, true));
     pass_seqs.push_back(transform::CombineParallelBatchMatmul(3));
     pass_seqs.push_back(transform::FoldConstant());
     pass_seqs.push_back(transform::FoldScaleAxis());

src/relay/backend/vm/compiler.cc

@@ -948,7 +948,7 @@ IRModule VMCompiler::OptimizeModule(const IRModule& mod, const TargetsMap& targe
   pass_seqs.push_back(transform::InlinePrimitives());
 
   pass_seqs.push_back(transform::CombineParallelConv2D(3));
-  pass_seqs.push_back(transform::CombineParallelDense(3));
+  pass_seqs.push_back(transform::CombineParallelDense(3, true));
   pass_seqs.push_back(transform::FoldConstant());
   pass_seqs.push_back(transform::FoldScaleAxis());
   pass_seqs.push_back(transform::CanonicalizeCast());

src/relay/transforms/combine_parallel_dense.cc

@@ -48,9 +48,12 @@
 namespace tvm {
 namespace relay {
 
-class ParallelDenseCombiner : public ParallelOpBatchCombiner {
+/*
+ * Class that find and combine parallel dense ops into batch_matmul.
+ */
+class ParallelDenseBatchCombiner : public ParallelOpBatchCombiner {
  public:
-  explicit ParallelDenseCombiner(uint64_t min_num_branches)
+  explicit ParallelDenseBatchCombiner(uint64_t min_num_branches)
       : ParallelOpBatchCombiner("nn.dense", "nn.batch_matmul", min_num_branches) {}
 
  protected:
@@ -68,17 +71,168 @@ class ParallelDenseCombiner : public ParallelOpBatchCombiner {
   }
 };
 
+/*
+ * Class that find and combine parallel dense ops into one dense op
+ * whose num of output units equals to sum of each sub-ops.
+ */
+class ParallelDenseFlatCombiner : public ParallelOpCombiner {
+ public:
+  explicit ParallelDenseFlatCombiner(uint64_t min_num_branches)
+      : ParallelOpCombiner("nn.dense", min_num_branches) {}
+
+ protected:
+  bool IsSupportedOp(const CallNode* n) { return true; }
+
+  bool CanOpsBeCombined(const CallNode* a, const CallNode* b) {
+    StructuralEqual eq;
+    const auto* attrs_a = a->attrs.as<DenseAttrs>();
+    const auto* attrs_b = b->attrs.as<DenseAttrs>();
+    const auto* weight_a = a->args[1]->type_as<TensorTypeNode>();
+    const auto* weight_b = b->args[1]->type_as<TensorTypeNode>();
+    CHECK(attrs_a != nullptr && attrs_b != nullptr && weight_a != nullptr && weight_b != nullptr);
+    // output dims (weight->shape[0]) can be different
+    return eq(attrs_a->out_dtype, attrs_b->out_dtype) && eq(weight_a->shape[1], weight_b->shape[1]);
+  }
+
+  Call MakeCombinedOp(const Group& branches) {
+    const Op& dense_op = Op::Get("nn.dense");
+    Expr input = branches[0][0]->args[0];
+    Expr new_weight;
+    IndexExpr new_output_dims;
+    // concat all weights into one
+    std::tie(new_weight, new_output_dims) = TransformWeight(branches);
+    const auto* origin_attrs = branches[0][0]->attrs.as<DenseAttrs>();
+    CHECK(origin_attrs);
+    const auto dense_attrs = make_object<DenseAttrs>();
+    dense_attrs->units = new_output_dims;
+    dense_attrs->out_dtype = origin_attrs->out_dtype;
+    return Call(dense_op, {input, new_weight}, Attrs{dense_attrs}, {});
+  }
+
+  bool IsArgCompatible(const CallNode* a, const CallNode* b, size_t index) {
+    StructuralEqual eq;
+    auto ta = a->args[index]->type_as<TensorTypeNode>();
+    auto tb = b->args[index]->type_as<TensorTypeNode>();
+    auto toutput_a = a->type_as<TensorTypeNode>();
+    auto toutput_b = b->type_as<TensorTypeNode>();
+    CHECK(ta != nullptr && tb != nullptr && toutput_a != nullptr && toutput_b != nullptr);
+
+    if (!eq(ta->dtype, tb->dtype) || ta->shape.size() != tb->shape.size()) {
+      return false;
+    }
+    if (toutput_a->shape.size() < ta->shape.size() || toutput_b->shape.size() < tb->shape.size()) {
+      return false;  // not broadcast/elemwise
+    }
+    if (ta->shape.size() > 0) {
+      for (size_t i = 0; i < ta->shape.size() - 1; i++) {
+        // shape dims must match except last dim
+        if (!eq(ta->shape[i], tb->shape[i])) return false;
+      }
+    }
+    return true;
+  }
+
+  Call MakeCombinedCallFromFollowingOps(const Expr& data, const Group& branches, size_t depth,
+                                        size_t parent_index) {
+    Array<Expr> new_args;
+    const CallNode* call = branches[0][depth];
+    for (size_t i = 0; i < call->args.size(); i++) {
+      if (i == parent_index) {
+        new_args.push_back(data);
+        continue;
+      }
+      size_t arg_ndim = call->args[i]->type_as<TensorTypeNode>()->shape.size();
+      size_t concat_axis = arg_ndim == 0 ? 0 : arg_ndim - 1;
+      Array<Expr> tuple;
+      for (const auto& branch : branches) {
+        auto parent = branch[depth]->args[parent_index];
+        auto& parent_shape = parent->type_as<TensorTypeNode>()->shape;
+        auto out_dim = tir::as_const_int(parent_shape[parent_shape.size() - 1]);
+        CHECK(out_dim != nullptr);
+
+        auto arg = branch[depth]->args[i];
+        auto& arg_shape = arg->type_as<TensorTypeNode>()->shape;
+        bool repeat_last_dim = false;
+        if (arg_ndim == 0) {
+          repeat_last_dim = true;
+          arg = MakeExpandDims(arg, -1, 1);
+        } else {
+          auto arg_last_dim = tir::as_const_int(arg_shape[arg_shape.size() - 1]);
+          CHECK(arg_last_dim != nullptr);
+          if (*out_dim > 1 && *arg_last_dim == 1) {
+            repeat_last_dim = true;
+          }
+        }
+        if (repeat_last_dim) {
+          // ensure broadcast is valid after concat args
+          arg = MakeRepeat(arg, *out_dim, concat_axis);
+        }
+        tuple.push_back(arg);
+      }
+      auto concat = MakeConcatenate(Tuple(tuple), concat_axis);
+      new_args.push_back(std::move(concat));
+    }
+    return Call(call->op, new_args, call->attrs, {});
+  }
+
+  void UpdateGroupOutput(const Expr& data, const Group& branches, size_t depth,
+                         ExprSubstMap* subst_map) {
+    int index = 0;
+    for (const auto& branch : branches) {
+      const CallNode* call = branch[depth];
+      auto& out_shape = call->type_as<TensorTypeNode>()->shape;
+      auto out_dims = tir::as_const_int(out_shape[out_shape.size() - 1]);
+      CHECK(out_dims != nullptr);
+      std::vector<int64_t> begin;
+      std::vector<int64_t> end;
+      std::vector<int64_t> strides;
+      for (size_t k = 0; k < out_shape.size() - 1; ++k) {
+        begin.push_back(0);
+        end.push_back(-1);
+        strides.push_back(1);
+      }
+      begin.push_back(index);
+      end.push_back(*out_dims);
+      strides.push_back(1);
+      index += *out_dims;
+      std::vector<int64_t> ndarray_shape = {static_cast<int64_t>(begin.size())};
+      Constant begin_const = MakeConstantTensor(DataType::Int(64), ndarray_shape, begin);
+      Constant end_const = MakeConstantTensor(DataType::Int(64), ndarray_shape, end);
+      Constant strides_const = MakeConstantTensor(DataType::Int(64), ndarray_shape, strides);
+      auto slice = MakeStridedSlice(data, begin_const, end_const, strides_const, "size");
+      subst_map->insert({GetRef<Expr>(branch[depth]), slice});
+    }
+  }
+
+ private:
+  std::tuple<Expr, IndexExpr> TransformWeight(const Group& branches) {
+    int64_t out_dims = 0;
+    Array<Expr> weights;
+    for (const auto& branch : branches) {
+      auto weight = branch[0]->args[1];
+      weights.push_back(weight);
+      out_dims += *tir::as_const_int(weight->type_as<TensorTypeNode>()->shape[0]);
+    }
+    return std::make_tuple(MakeConcatenate(Tuple(weights), 0),
+                           tir::make_const(DataType::Int(32), out_dims));
+  }
+};
+
 /*! \brief Combine parallel dense if number of branches >= min_num_branches */
-Expr CombineParallelDense(const Expr& expr, uint64_t min_num_branches) {
-  return ParallelDenseCombiner(min_num_branches).Combine(expr);
+Expr CombineParallelDense(const Expr& expr, uint64_t min_num_branches, bool to_batch) {
+  if (to_batch) {
+    return ParallelDenseBatchCombiner(min_num_branches).Combine(expr);
+  } else {
+    return ParallelDenseFlatCombiner(min_num_branches).Combine(expr);
+  }
 }
 
 namespace transform {
 
-Pass CombineParallelDense(uint64_t min_num_branches) {
+Pass CombineParallelDense(uint64_t min_num_branches, bool to_batch) {
   runtime::TypedPackedFunc<Function(Function, IRModule, PassContext)> pass_func =
       [=](Function f, IRModule m, PassContext pc) {
-        return Downcast<Function>(CombineParallelDense(f, min_num_branches));
+        return Downcast<Function>(CombineParallelDense(f, min_num_branches, to_batch));
       };
   return CreateFunctionPass(pass_func, 4, "CombineParallelDense", {"InferType"});
 }