feat(mge/imperative): add valid flag of `infer_output_attrs_fallible`

GitOrigin-RevId: b2b32774ee

imperative/src/impl/interpreter_impl.cpp

@@ -63,15 +63,17 @@ SmallVector<void*> ChannelImpl::apply_op(
         input_infos.push_back(info);
         input_descs.push_back(info->desc);
     }
-    auto output_descs = OpDef::infer_output_attrs_fallible(*op, input_descs);
+
+    auto [output_descs, validated] = OpDef::infer_output_attrs_fallible(*op, input_descs);
     ApplyOp cmd{std::move(op)};
     cmd.inputs = std::move(input_infos);
     cmd.outputs.reserve(output_descs.size());
     SmallVector<void*> outputs;
-    bool is_fallible = false;
+    // FIXME: remove this check when op check is correct
+    bool validated_bkp = true;
     for (auto&& desc : output_descs) {
         if (desc.layout.ndim == 0) {
-            is_fallible = true;
+            validated_bkp = false;
         }
         auto info = alloc();
         info->desc = desc;
@@ -80,8 +82,14 @@ SmallVector<void*> ChannelImpl::apply_op(
         outputs.push_back(info);
     }
     m_worker.add_task(std::move(cmd));
-    if (is_fallible && m_async_level <= 1) {
+    if (!(validated && validated_bkp) && m_async_level == 1) {
+        sync();
+    } else if (m_async_level == 0) {
         sync();
+        // check device error
+        for (auto&& oup : cmd.outputs) {
+            oup->ptr->comp_node().sync();
+        }
     }
     return outputs;
 }
@@ -194,6 +202,9 @@ ChannelImpl::~ChannelImpl() {
 void ChannelImpl::produce_tensor(TensorInfo* dest, TensorPtr ptr) {
     MGB_LOCK_GUARD(m_mutex);
     dest->value_fetched = ptr->value_fetched();
+    // update tensor desc for static infer
+    dest->desc.layout = ptr->layout();
+    dest->desc.comp_node = ptr->comp_node();
     dest->ptr = std::move(ptr);
     if (m_waitee == dest) {
         m_cv.notify_all();

imperative/src/impl/op_def.cpp

@@ -42,7 +42,7 @@ cg::OperatorNodeBase* OpDef::apply_on_var_node(
     return def.trait()->apply_on_var_node(def, inputs);
 }
 
-SmallVector<LogicalTensorDesc> OpDef::infer_output_attrs_fallible(
+std::tuple<SmallVector<LogicalTensorDesc>, bool> OpDef::infer_output_attrs_fallible(
     const OpDef& def,
     const SmallVector<LogicalTensorDesc>& inputs) {
     return def.trait()->infer_output_attrs_fallible(def, inputs);

imperative/src/impl/ops/backward_graph.cpp

@@ -24,12 +24,12 @@ BackwardGraph::InternalGraph::apply(
         inputs);
 }
 
-SmallVector<LogicalTensorDesc>
-BackwardGraph::InternalGraph::infer_attrs(
+std::tuple<SmallVector<LogicalTensorDesc>, bool> BackwardGraph::InternalGraph::infer_attrs(
         const SmallVector<LogicalTensorDesc>& inputs) const {
     using TensorAttr = LogicalTensorDesc;
     ThinHashMap<size_t, TensorAttr> node2attr;
     auto&& input_nodes = this->inputs;
+    auto&& output_nodes = this->outputs;
     mgb_assert(inputs.size() == input_nodes.size());
     for (size_t i = 0; i < inputs.size(); ++ i) {
         node2attr[input_nodes[i]] = inputs[i];
@@ -41,25 +41,29 @@ BackwardGraph::InternalGraph::infer_attrs(
             i.second->layout(), i.second->comp_node(),
             value->proxy_to_default_cpu()};
     }
+    bool validated = true;
     for (size_t i = 0; i < exprs.size(); ++ i) {
-        auto&& expr = exprs[i];
-        SmallVector<TensorAttr> inputs;
-        for (auto &&in : std::get<1>(expr)) {
-            inputs.push_back(node2attr.at(in));
+        auto&& [expr_op, expr_inps, expr_oups] = exprs[i];
+        SmallVector<TensorAttr> expr_input_descs;
+        for (auto &&inp : expr_inps) {
+            expr_input_descs.push_back(node2attr.at(inp));
         }
-        auto outputs = OpDef::infer_output_attrs_fallible(
-                *std::get<0>(expr), inputs);
-        auto output_nodes = std::get<2>(expr);
-        mgb_assert(outputs.size() == output_nodes.size());
-        for (size_t i = 0; i < outputs.size(); ++ i) {
-            node2attr[output_nodes[i]] = outputs[i];
+
+        auto[expr_output_descs, expr_validated] = OpDef::infer_output_attrs_fallible(
+            *expr_op, expr_input_descs);
+        validated = validated && expr_validated;
+
+        mgb_assert(expr_output_descs.size() == expr_oups.size());
+        for (size_t i = 0; i < expr_output_descs.size(); ++ i) {
+            node2attr[expr_oups[i]] = expr_output_descs[i];
         }
     }
+
     SmallVector<TensorAttr> ret;
-    for (auto &&i : outputs) {
+    for (auto &&i : output_nodes) {
         ret.push_back(node2attr.at(i));
     }
-    return ret;
+    return {ret, validated};
 }
 
 MGB_DYN_TYPE_OBJ_FINAL_IMPL(BackwardGraph);
@@ -72,11 +76,11 @@ SmallVector<TensorPtr> backward_impl(
             .graph().apply(tensors);
 }
 
-SmallVector<LogicalTensorDesc> infer_tensor_attrs(
+std::tuple<SmallVector<LogicalTensorDesc>, bool> infer_tensor_attrs(
     const OpDef& backward_graph,
     const SmallVector<LogicalTensorDesc> inputs) {
     return backward_graph.cast_final_safe<BackwardGraph>()
-            .graph().infer_attrs(inputs);
+        .graph().infer_attrs(inputs);
 }
 
 OP_TRAIT_REG(BackwardGraph, BackwardGraph)

imperative/src/impl/ops/batch_norm.cpp

@@ -44,7 +44,7 @@ cg::OperatorNodeBase* apply_on_var_node(
     }
 }
 
-SmallVector<LogicalTensorDesc> infer_output_attrs_fallible(
+std::tuple<SmallVector<LogicalTensorDesc>, bool> infer_output_attrs_fallible(
         const OpDef& def,
         const SmallVector<LogicalTensorDesc>& inputs) {
     auto&& op_def = def.cast_final_safe<BatchNorm>();
@@ -66,7 +66,7 @@ SmallVector<LogicalTensorDesc> infer_output_attrs_fallible(
         out_shapes[i] = {i1.layout, i1.comp_node};
     }
     out_shapes[nr_out-1] = {i0.layout, i0.comp_node};
-    return out_shapes;
+    return {out_shapes, true};
 }
 
 OP_TRAIT_REG(BatchNorm, BatchNorm, opr::BatchNorm)

imperative/src/impl/ops/broadcast.cpp

@@ -47,7 +47,7 @@ bool valid_broadcast(const TensorShape& src_shape,
     return true;
 }
 
-SmallVector<LogicalTensorDesc> infer_output_attrs_fallible(
+std::tuple<SmallVector<LogicalTensorDesc>, bool> infer_output_attrs_fallible(
         const OpDef& def,
         const SmallVector<LogicalTensorDesc>& inputs) {
     def.cast_final_safe<Broadcast>();
@@ -59,7 +59,7 @@ SmallVector<LogicalTensorDesc> infer_output_attrs_fallible(
     TensorLayout out_layout = src.layout;
     if (tshp.layout.ndim == 0 || tshp.value.empty()) {
         out_layout.ndim = 0;
-        return {{out_layout, src.comp_node}};
+        return {{{out_layout, src.comp_node}}, true};
     }
     mgb_assert(
         tshp.layout.ndim == 1, 
@@ -77,7 +77,7 @@ SmallVector<LogicalTensorDesc> infer_output_attrs_fallible(
                src.layout.TensorShape::to_string().c_str(),
                out_layout.TensorShape::to_string().c_str());
 
-    return {{out_layout, src.comp_node}};
+    return {{{out_layout, src.comp_node}}, true};
 }
 
 OP_TRAIT_REG(Broadcast, Broadcast, opr::Broadcast)

imperative/src/impl/ops/cond_take.cpp

@@ -25,7 +25,7 @@ namespace {
 
 class MegDNNDynOutMallocImpl final: public megdnn::DynOutMallocPolicy {
     using Output = std::array<TensorPtr, 2>;
-    
+
     CompNode m_cn;
     Output m_out;
 
@@ -110,14 +110,14 @@ SmallVector<TensorPtr> apply_on_physical_tensor(
     return out;
 }
 
-SmallVector<LogicalTensorDesc> infer_output_attrs_fallible(
+std::tuple<SmallVector<LogicalTensorDesc>, bool> infer_output_attrs_fallible(
     const OpDef& def,
     const SmallVector<LogicalTensorDesc>& inputs) {
     auto cn = inputs[0].comp_node;
-    return {
+    return {{
         {TensorLayout(inputs[0].layout.dtype), cn},
         {TensorLayout(dtype::Int32()), cn}
-    };
+    }, true};
 }
 
 OP_TRAIT_REG(CondTake, CondTake, opr::CondTake)
@@ -128,4 +128,4 @@ OP_TRAIT_REG(CondTake, CondTake, opr::CondTake)
 
 } // namespace
 
-} // namespace mgb::imperative
+} // namespace mgb::imperative

imperative/src/impl/ops/elemwise.cpp

@@ -29,7 +29,7 @@ cg::OperatorNodeBase* apply_on_var_node(
     return opr::Elemwise::make(inputs, elemwise_opr.mode).node()->owner_opr();
 }
 
-SmallVector<LogicalTensorDesc> infer_output_attrs_fallible(
+std::tuple<SmallVector<LogicalTensorDesc>, bool> infer_output_attrs_fallible(
         const OpDef& def,
         const SmallVector<LogicalTensorDesc>& inputs) {
     auto&& op_def = def.cast_final_safe<Elemwise>();
@@ -40,7 +40,7 @@ SmallVector<LogicalTensorDesc> infer_output_attrs_fallible(
     TensorShapeArray inp_shapes;
     DType out_dt;
     CompNode out_cn;
-    for (size_t i = 0; i < inputs.size(); ++ i) {        
+    for (size_t i = 0; i < inputs.size(); ++ i) {
         auto &&t = inputs[i];
         if (!i) {
             out_cn = t.comp_node;
@@ -55,12 +55,12 @@ SmallVector<LogicalTensorDesc> infer_output_attrs_fallible(
             TensorLayout out_layout;
             out_layout.ndim = 0;
             out_layout.dtype = out_dt;
-            return {{out_layout, out_cn}};
+            return {{{out_layout, out_cn}}, true};
         }
     }
 
     auto&& out_shape = opr::Elemwise::get_output_var_shape(op_def.mode, inp_shapes);
-    return {{TensorLayout(out_shape, out_dt, inputs[0].layout.format), out_cn}};
+    return {{{TensorLayout(out_shape, out_dt, inputs[0].layout.format), out_cn}}, true};
 }
 
 OP_TRAIT_REG(Elemwise, Elemwise, opr::Elemwise)

imperative/src/impl/ops/tensor_manip.cpp

@@ -40,21 +40,21 @@ SmallVector<TensorPtr> apply_on_physical_tensor(
     return {Tensor::make(std::move(hv))};
 }
 
-SmallVector<LogicalTensorDesc> infer_output_attrs_fallible(
+std::tuple<SmallVector<LogicalTensorDesc>, bool> infer_output_attrs_fallible(
         const OpDef& def,
         const SmallVector<LogicalTensorDesc>& inputs) {
     def.cast_final_safe<GetVarShape>();
     mgb_assert(inputs.size() == 1, "GetVarShape take 1 input, got %lu", inputs.size());
     auto&& desc = inputs[0];
     if (!desc.layout.ndim) {
-        return {{TensorLayout(dtype::Int32()), desc.comp_node}};
+        return {{{TensorLayout(dtype::Int32()), desc.comp_node}}, true};
     }
     DeviceTensorND value(CompNode::default_cpu(), {desc.layout.ndim}, dtype::Int32());
     auto* ptr = value.ptr<dt_int32>();
     for (size_t i = 0; i < desc.layout.ndim; ++i) {
         ptr[i] = desc.layout[i];
     }
-    return {{value.layout(), desc.comp_node, std::move(value)}};
+    return {{{value.layout(), desc.comp_node, std::move(value)}}, true};
 }
 
 std::shared_ptr<OpDef> make_from_op_node(cg::OperatorNodeBase* node_) {

imperative/src/impl/profiler.cpp

@@ -28,12 +28,13 @@ namespace {
 CompNode::UnorderedSet collect_comp_nodes(
         const OpDef& def, const SmallVector<TensorPtr>& inputs) {
     CompNode::UnorderedSet comp_nodes;
-    SmallVector<LogicalTensorDesc> descs;
+    SmallVector<LogicalTensorDesc> inp_descs;
     for (auto&& i : inputs) {
         comp_nodes.insert(i->comp_node());
-        descs.push_back({i->layout(), i->comp_node(), {}});
+        inp_descs.push_back({i->layout(), i->comp_node(), {}});
     }
-    for (auto&& output_attr : def.infer_output_attrs_fallible(def, descs)) {
+    SmallVector<LogicalTensorDesc> oup_descs = std::get<0>(def.infer_output_attrs_fallible(def, inp_descs));
+    for (auto&& output_attr : oup_descs) {
         comp_nodes.insert(output_attr.comp_node);
     }
     return comp_nodes;

imperative/src/impl/proxy_graph.cpp

@@ -14,6 +14,7 @@
 #include "megbrain/graph/static_infer.h"
 #include "megbrain/graph/operator_node.h"
 #include "megbrain/opr/io.h"
+#include "megbrain/opr/tensor_manip.h"
 #include "megbrain/opr/utility.h"
 #include "megbrain/imperative/ops/opr_attr.h"
 #include "megbrain/imperative/ops/backward_graph.h"
@@ -590,10 +591,9 @@ cg::OperatorNodeBase* ProxyGraph::get_proxy_opr(
         vinputs[i] = InputPlaceholder::make(*m_graph, *inputs[i]).node();
     }
     auto opr = OpDef::apply_on_var_node(opdef, vinputs);
-    mgb_assert(opr->dyn_typeinfo() != InputPlaceholder::typeinfo());
+    mgb_assert(!opr->same_type<InputPlaceholder>());
     for (auto &&i : opr->input()) {
-        mgb_assert(i->owner_opr()->dyn_typeinfo() ==
-                InputPlaceholder::typeinfo());
+        mgb_assert(i->owner_opr()->same_type<InputPlaceholder>());
     }
     return opr;
 }
@@ -605,17 +605,18 @@ size_t ProxyGraph::get_opr_output_size(const OpDef& opdef,
     return get_proxy_opr(opdef, inputs)->usable_output().size();
 }
 
-SmallVector<LogicalTensorDesc> ProxyGraph::infer_output_attrs_fallible(
+std::tuple<SmallVector<LogicalTensorDesc>, bool> ProxyGraph::infer_output_attrs_fallible(
         const OpDef& opdef,
         const SmallVector<LogicalTensorDesc>& inputs) {
     auto opr = get_proxy_opr(opdef, inputs);
     CUR_OPR_GUARD(opr);
-    do_shape_infer(false);
-    SmallVector<LogicalTensorDesc> ret;
+    SmallVector<LogicalTensorDesc> outputs;
+    bool validated = do_shape_infer(false);
     for (auto&& i : opr->usable_output()) {
-        ret.push_back({{i->shape(), i->dtype()}, i->comp_node()});
+        outputs.push_back({{i->shape(), i->dtype()}, i->comp_node()});
     }
-    return ret;
+    bool need_check = opr->same_type<opr::Reshape>();
+    return {outputs, validated && !need_check};
 }
 
 struct ProxyGraph::GradGraph {
@@ -811,16 +812,20 @@ VarNodeArray ProxyGraph::make_input_place_holders(const SmallVector<LogicalTenso
 
 /*********************** Common Impl ***********************/
 
-void ProxyGraph::do_shape_infer(bool sync_value) {
+bool ProxyGraph::do_shape_infer(bool sync_value) {
     m_static_infer_manager->update();
 
+    bool validated = true;
     for (auto* var : m_cur_opr->output()) {
         if (sync_value) {
             var->shape(m_static_infer_manager->infer_shape(var));
         } else if (auto* shape = m_static_infer_manager->infer_shape_fallible(var)) {
-            var->shape(*shape);
+                var->shape(*shape);
+        } else {
+            validated = false;
         }
     }
+    return validated;
 }
 
 TensorPtr ProxyGraph::as_tensor(cg::OperatorNodeBase* opr, bool share) {

imperative/src/impl/proxy_graph.h

@@ -48,7 +48,7 @@ public:
             const OpDef& opdef,
             const SmallVector<LogicalTensorDesc>& inputs);
 
-    SmallVector<LogicalTensorDesc> infer_output_attrs_fallible(
+    std::tuple<SmallVector<LogicalTensorDesc>, bool> infer_output_attrs_fallible(
             const OpDef& opdef,
             const SmallVector<LogicalTensorDesc>& inputs);
 
@@ -88,7 +88,7 @@ private:
 
     /********************** Common Helper **********************/
 
-    void do_shape_infer(bool sync_value);
+    bool do_shape_infer(bool sync_value);
 
     TensorPtr as_tensor(cg::OperatorNodeBase* opr, bool share=true);
 

imperative/src/impl/proxy_graph_detail.cpp

@@ -80,8 +80,7 @@ apply_on_physical_tensor(const OpDef& def,
     return outputs;
 }
 
-SmallVector<LogicalTensorDesc>
-infer_output_attrs_fallible(const OpDef& def,
+std::tuple<SmallVector<LogicalTensorDesc>, bool> infer_output_attrs_fallible(const OpDef& def,
         const SmallVector<LogicalTensorDesc>& inputs) {
     auto&& graph = ProxyGraph::get_default_graph();
     return graph->infer_output_attrs_fallible(def, inputs);
@@ -136,4 +135,4 @@ make_backward_graph(const OpDef& def,
 } // namespace imperative
 } // namespace mgb
 
-// vim: syntax=cpp.doxygen foldmethod=marker foldmarker=f{{{,f}}}
+// vim: syntax=cpp.doxygen foldmethod=marker foldmarker=f{{{,f}}}

imperative/src/impl/proxy_graph_detail.h

@@ -21,8 +21,7 @@ SmallVector<TensorPtr>
 apply_on_physical_tensor(const OpDef& def,
         const SmallVector<TensorPtr>& inputs);
 
-SmallVector<LogicalTensorDesc>
-infer_output_attrs_fallible(const OpDef& def,
+std::tuple<SmallVector<LogicalTensorDesc>, bool> infer_output_attrs_fallible(const OpDef& def,
         const SmallVector<LogicalTensorDesc>& inputs);
 
 BackwardGraphResult
@@ -35,4 +34,4 @@ make_backward_graph(const OpDef& def,
 } // namespace imperative
 } // namespace mgb
 
-// vim: syntax=cpp.doxygen foldmethod=marker foldmarker=f{{{,f}}}
+// vim: syntax=cpp.doxygen foldmethod=marker foldmarker=f{{{,f}}}

imperative/src/include/megbrain/imperative/op_def.h

@@ -44,7 +44,7 @@ public:
         const OpDef& def,
         const VarNodeArray& inputs);
 
-    static SmallVector<LogicalTensorDesc> infer_output_attrs_fallible(
+    static std::tuple<SmallVector<LogicalTensorDesc>, bool> infer_output_attrs_fallible(
         const OpDef& def,
         const SmallVector<LogicalTensorDesc>& inputs);
 

imperative/src/include/megbrain/imperative/ops/backward_graph.h

@@ -38,8 +38,8 @@ public:
         SmallVector<TensorPtr>
         apply(const SmallVector<TensorPtr>& inputs) const;
 
-        SmallVector<LogicalTensorDesc>
-        infer_attrs(const SmallVector<LogicalTensorDesc>& inputs) const;
+        std::tuple<SmallVector<LogicalTensorDesc>, bool> infer_attrs(
+            const SmallVector<LogicalTensorDesc>& inputs) const;
 
         template <typename T, typename F, typename C>
         SmallVector<T> interpret(F&& f, C&& c, const SmallVector<T>& inputs) const {