feat(opr): let Dot, MatrixMul and BatchedMatrixMul support empty input

GitOrigin-RevId: 10a3c5b106

imperative/python/test/unit/functional/test_functional.py

@@ -142,6 +142,26 @@ def test_matmul():
     )
 
 
+@pytest.mark.parametrize(
+    "shape_a, shape_b", [((0,), (0,)), ((10, 0), (0, 10)), ((3, 10, 0), (3, 0, 10)),],
+)
+@pytest.mark.parametrize("is_symbolic", [None, True, False])
+def test_matmul_empty_tensor(shape_a, shape_b, is_symbolic):
+    def func(a, b):
+        return F.matmul(a, b)
+
+    if is_symbolic is not None:
+        func = jit.trace(symbolic=is_symbolic)(func)
+
+    a = tensor(np.random.randn(*shape_a))
+    b = tensor(np.random.randn(*shape_b))
+    for _ in range(3):
+        out = func(a, b)
+        assert np.all(out.numpy() == 0)
+        if is_symbolic is None:
+            break
+
+
 def test_interpolate():
     def linear_interpolate():
         inp = tensor(np.arange(1, 3, dtype=np.float32).reshape(1, 1, 2))

src/opr/impl/blas.cpp

@@ -45,6 +45,7 @@ MatrixMul::MatrixMul(VarNode* a, VarNode* b, const Param& param,
     init_megdnn_opr(*this, param);
     m_policy = policy;
     add_input({a, b});
+    output(0)->add_flag(VarNode::Flag::ALLOW_EMPTY_SHAPE);
 }
 
 SymbolVar MatrixMul::make(SymbolVar a, SymbolVar b, const Param& param,
@@ -61,6 +62,15 @@ void MatrixMul::init_output_dtype() {
     output(0)->dtype(output_dtype);
 }
 
+MatrixMul::NodeProp* MatrixMul::do_make_node_prop() const {
+    auto ret = Super::do_make_node_prop();
+    ret->add_dep_type_existing_var(input(0),
+            NodeProp::DepType::VALUE_ALLOW_EMPTY);
+    ret->add_dep_type_existing_var(input(1),
+            NodeProp::DepType::VALUE_ALLOW_EMPTY);
+    return ret;
+}
+
 bool MatrixMul::check_layout(const TensorLayout& layout, int transpose) {
     mgb_assert(layout.ndim == 2, "input to MatrixMul must be 2-dim; got %s",
                layout.to_string().c_str());
@@ -138,6 +148,17 @@ void MatrixMul::scn_do_execute() {
     auto inp0 = input(0)->dev_tensor().as_megdnn(),
          inp1 = input(1)->dev_tensor().as_megdnn(),
          out = output(0)->dev_tensor().as_megdnn();
+    if ((inp0.layout.is_empty() || inp1.layout.is_empty())) {
+        if (!out.layout.is_empty()) {
+            if (!m_fill_opr) {
+                m_fill_opr = intl::get_megdnn_handle(comp_node())->
+                    create_operator<megdnn::Fill>();
+            }
+            m_fill_opr->param() = 0;
+            m_fill_opr->exec(out, {});
+        }
+        return;
+    }
     auto transpose = [](TensorLayout& layout, bool& trans) {
         if (!check_layout(layout, 0)) {
             mgb_assert(check_layout(layout, 1));
@@ -193,6 +214,7 @@ BatchedMatrixMul::BatchedMatrixMul(VarNode* a, VarNode* b, const Param& param,
     init_megdnn_opr(*this, param);
     m_policy = policy;
     add_input({a, b});
+    output(0)->add_flag(VarNode::Flag::ALLOW_EMPTY_SHAPE);
 }
 
 SymbolVar BatchedMatrixMul::make(SymbolVar a, SymbolVar b, const Param& param,
@@ -229,6 +251,15 @@ void BatchedMatrixMul::init_output_dtype() {
     output(0)->dtype(output_dtype);
 }
 
+BatchedMatrixMul::NodeProp* BatchedMatrixMul::do_make_node_prop() const {
+    auto ret = Super::do_make_node_prop();
+    ret->add_dep_type_existing_var(input(0),
+            NodeProp::DepType::VALUE_ALLOW_EMPTY);
+    ret->add_dep_type_existing_var(input(1),
+            NodeProp::DepType::VALUE_ALLOW_EMPTY);
+    return ret;
+}
+
 bool BatchedMatrixMul::check_layout(const TensorLayout& layout,
                                     bool transpose) {
     int lhs = (transpose) ? 2 : 1, rhs = (transpose) ? 1 : 2;
@@ -294,6 +325,17 @@ void BatchedMatrixMul::scn_do_execute() {
     auto inp0 = input(0)->dev_tensor().as_megdnn(),
          inp1 = input(1)->dev_tensor().as_megdnn(),
          out = output(0)->dev_tensor().as_megdnn();
+    if ((inp0.layout.is_empty() || inp1.layout.is_empty())) {
+        if (!out.layout.is_empty()) {
+            if (!m_fill_opr) {
+                m_fill_opr = intl::get_megdnn_handle(comp_node())->
+                    create_operator<megdnn::Fill>();
+            }
+            m_fill_opr->param() = 0;
+            m_fill_opr->exec(out, {});
+        }
+        return;
+    }
     auto transpose = [](TensorLayout& layout, bool& trans) {
         if (!check_layout(layout, false)) {
             mgb_assert(check_layout(layout, true));
@@ -354,6 +396,7 @@ Dot::Dot(VarNode *opr0, VarNode *opr1, const OperatorNodeConfig &config):
 {
     init_megdnn_opr(*this, {});
     add_input({opr0, opr1}, AddInputSortType::CUR_ADDED);
+    output(0)->add_flag(VarNode::Flag::ALLOW_EMPTY_SHAPE);
     static_assert(std::is_empty<Param>::value, "Dot param should be empty");
     mgb_assert(opr0->dtype().category() != DTypeCategory::QUANTIZED &&
                        opr1->dtype().category() != DTypeCategory::QUANTIZED,
@@ -406,10 +449,28 @@ void Dot::scn_do_execute() {
             i1.layout.stride[0] = 0;
         }
     }
+    if ((i0.layout.is_empty() || i1.layout.is_empty())) {
+        if (!m_fill_opr) {
+            m_fill_opr = intl::get_megdnn_handle(comp_node())->
+                create_operator<megdnn::Fill>();
+        }
+        m_fill_opr->param() = 0;
+        m_fill_opr->exec(output(0)->dev_tensor().as_megdnn(), {});
+        return;
+    }
     megdnn_opr()->exec(i0, i1, output(0)->dev_tensor().as_megdnn(),
             intl::get_megdnn_workspace_from_var(output(1)));
 }
 
+Dot::NodeProp* Dot::do_make_node_prop() const {
+    auto ret = Super::do_make_node_prop();
+    ret->add_dep_type_existing_var(input(0),
+            NodeProp::DepType::VALUE_ALLOW_EMPTY);
+    ret->add_dep_type_existing_var(input(1),
+            NodeProp::DepType::VALUE_ALLOW_EMPTY);
+    return ret;
+}
+
 void Dot::add_input_layout_constraint() {
     auto check = [](const TensorLayout &ly) {
         mgb_throw_if(ly.ndim != 1, GraphError,

src/opr/include/megbrain/opr/blas.h

@@ -17,6 +17,7 @@
 #include "megbrain/graph.h"
 #include "megbrain/opr/internal/megdnn_opr_wrapper.h"
 
+#include "megdnn/oprs/general.h"
 #include "megdnn/oprs/linalg.h"
 
 namespace mgb {
@@ -40,6 +41,7 @@ private:
     void add_input_layout_constraint() override;
     void scn_do_execute() override;
     void init_output_dtype() override;
+    NodeProp* do_make_node_prop() const override;
     size_t get_workspace_size_bytes(const TensorShapeArray& input_shapes,
                                     const TensorShapeArray& output_shapes)
             const override;
@@ -47,6 +49,7 @@ private:
 
     //! store the policy of all transpose situations
     megdnn::ExecutionPolicy m_cadidate_execution_policies[4];
+    std::unique_ptr<megdnn::Fill> m_fill_opr;
 };
 
 /*!
@@ -70,6 +73,7 @@ private:
     void add_input_layout_constraint() override;
     void init_output_dtype() override;
     void scn_do_execute() override;
+    NodeProp* do_make_node_prop() const override;
     size_t get_workspace_size_bytes(const TensorShapeArray& input_shapes,
                                     const TensorShapeArray& output_shapes)
             const override;
@@ -77,6 +81,7 @@ private:
     static bool check_layout(const TensorLayout& layout, bool transpose);
     //! store the policy of all transpose situations
     megdnn::ExecutionPolicy m_cadidate_execution_policies[4];
+    std::unique_ptr<megdnn::Fill> m_fill_opr;
 };
 
 /*!
@@ -101,7 +106,9 @@ MGB_DEFINE_OPR_CLASS(Dot, cg::SingleCNOperatorNodeBaseT<
         void add_input_layout_constraint() override;
         void scn_do_execute() override;
         void init_output_static_infer_desc() override;
+        NodeProp* do_make_node_prop() const override;
         void record_execute_deps(ExecDependencyArray &deps) override;
+        std::unique_ptr<megdnn::Fill> m_fill_opr;
 };
 
 MGB_DEFINE_MEGDNN_OPR_WRAPPER_FWD1(MatrixInverse);

src/opr/test/blas.cpp

@@ -94,7 +94,9 @@ void run_sgemm_test(bool transa, bool transb) {
     Checker(make_graph, fwd)
             .run({mkx(4, 6), mky(6, 2)}, opt)
             .run({mkx(2, 3), mky(3, 100)}, opt)
-            .run({mkx(20, 3), mky(3, 20)}, opt);
+            .run({mkx(20, 3), mky(3, 20)}, opt)
+            .run({mkx(10, 0), mky(0, 10)}, opt)
+            .run({mkx(0, 0), mky(0, 0)}, opt);
 }
 
 #define FWD_BATCH_GEMM(dt_src, dt_dst)                                       \
@@ -143,7 +145,9 @@ void run_batched_sgemm_test(bool transa, bool transb) {
     Checker(make_graph, fwd)
             .run({mkx(3, 5, 7), mky(3, 7, 2)}, opt)
             .run({mkx(64, 1, 2), mky(64, 2, 1)}, opt)
-            .run({mkx(1, 2, 3), mky(1, 3, 4)}, opt);
+            .run({mkx(1, 2, 3), mky(1, 3, 4)}, opt)
+            .run({mkx(3, 0, 2), mky(3, 2, 0)}, opt)
+            .run({mkx(64, 10, 0), mky(64, 0, 10)}, opt);
 }
 
 auto gen_fp16 = [](HostTensorND& dest) {
@@ -198,6 +202,7 @@ void run_batched_hgemm_test(bool transa, bool transb) {
 
     checker.run({mkx(3, 5, 7), mky(3, 7, 2)}, opt)
             .run({mkx(64, 1, 2), mky(64, 2, 1)}, opt)
+            .run({mkx(64, 10, 0), mky(64, 0, 10)}, opt)
             .run({mkx(1, 2, 3), mky(1, 3, 4)}, opt);
 }
 
@@ -236,6 +241,7 @@ void run_batched_igemm_test(bool transa, bool transb) {
 
     checker.run({mkx(3, 5, 7), mky(3, 7, 2)}, opt)
             .run({mkx(64, 1, 2), mky(64, 2, 1)}, opt)
+            .run({mkx(64, 10, 0), mky(64, 0, 10)}, opt)
             .run({mkx(1, 2, 3), mky(1, 3, 4)}, opt);
 }
 
@@ -650,7 +656,8 @@ TEST(TestOprBlas, Dot) {
             .run({TensorShape{15}, TensorShape{1}})
             .run({TensorShape{1}, TensorShape{16}})
             .run({TensorShape{23}, TensorShape{23}})
-            .run({TensorShape{1000}, TensorShape{1000}});
+            .run({TensorShape{1000}, TensorShape{1000}})
+            .run({TensorShape{0}, TensorShape{0}});
 }
 
 TEST(TestOprBlas, TransMatMul) {

test/src/autocheck.cpp

@@ -250,7 +250,6 @@ DEF_IMPL(void)::do_run(const ShapeInpArray& shapes, const RunOptions& opt) {
         for (size_t i = 0; i < nr_out; ++i) {
             if (m_outputs_allow_grad[i]) {
                 auto nr = m_outputs_truth[i].shape().total_nr_elems();
-                mgb_assert(nr, "got empty output");
                 if (opt.cont_loss_p) {
                     m_loss_p[i]->resize({nr});
                     auto ptr = m_loss_p[i]->template ptr<float>();

test/src/numerical_diff.cpp

@@ -36,7 +36,7 @@ std::vector<HostTensorND> mgb::numerical_diff_pt2(
             resize(cur_inp->shape());
         auto dptr = dest.ptr<float>();
 
-        mgb_assert(cur_inp->layout().is_contiguous());
+        mgb_assert(cur_inp->layout().is_contiguous() || cur_inp->layout().is_empty());
         auto cur_inp_ptr = cur_inp->ptr<float>();
 
         mgb::RealTimer timer;