!26811 Support CSRTensor in while loop and subgraph

Merge pull request !26811 from 杨林枫/csr_in_while

mindspore/ccsrc/backend/optimizer/pass/sparse_process.cc

@@ -26,6 +26,24 @@
 
 namespace mindspore {
 namespace opt {
+// Convert CSRTensor Parameter or ValueNode to Tuple by setting its abstract.
+void AbstractCSRToAbstractTuple(const AnfNodePtr &sparse) {
+  MS_EXCEPTION_IF_NULL(sparse);
+  if (!(sparse->isa<Parameter>() || sparse->isa<ValueNode>())) {
+    return;
+  }
+  auto param_abs = sparse->abstract();
+  MS_EXCEPTION_IF_NULL(param_abs);
+  if (param_abs->isa<abstract::AbstractCSRTensor>()) {
+    auto abs_sparse = param_abs->cast<abstract::AbstractCSRTensorPtr>();
+    std::vector<AbstractBasePtr> abstract_list{abs_sparse->indptr(), abs_sparse->indices(), abs_sparse->values(),
+                                               abs_sparse->dense_shape()};
+    auto abs_tuple = std::make_shared<abstract::AbstractTuple>(abstract_list);
+    abs_tuple->set_type(abs_tuple->BuildType());
+    sparse->set_abstract(abs_tuple);
+  }
+}
+
 const AnfNodePtr SparseProcess::Process(const FuncGraphPtr &func_graph, const AnfNodePtr &node,
                                         const EquivPtr &) const {
   MS_EXCEPTION_IF_NULL(func_graph);
@@ -60,14 +78,16 @@ const AnfNodePtr SparseProcess::Process(const FuncGraphPtr &func_graph, const An
   } else if (sparse_attr_map.find(prim_name) != sparse_attr_map.end()) {
     const auto &inputs = cnode->inputs();
     // Inputs should be [sparse_getattr, sparse]
+    if (inputs.size() <= 1) {
+      MS_LOG_EXCEPTION << "For SparseGetAttr, CNode must have 2 inputs (Prim, Sparse)";
+    }
     constexpr size_t sparse_index = 1;
-    AnfNodePtr sparse = inputs[sparse_index];
+    AbstractCSRToAbstractTuple(inputs[sparse_index]);
-    MS_EXCEPTION_IF_NULL(sparse);
     int64_t index = sparse_attr_map.at(prim_name);
     auto cons_node = NewValueNode(index);
     AbstractBasePtr aptr = std::make_shared<abstract::AbstractScalar>(std::make_shared<Int64Imm>(index));
     cons_node->set_abstract(aptr);
-    auto new_node = NewCNode({NewValueNode(prim::kPrimTupleGetItem), sparse, cons_node}, func_graph);
+    auto new_node = NewCNode({NewValueNode(prim::kPrimTupleGetItem), inputs[sparse_index], cons_node}, func_graph);
     new_node->set_abstract(node->abstract());
     return new_node;
   }

mindspore/ccsrc/backend/session/cpu_session.cc

@@ -170,7 +170,7 @@ void CPUSession::CreateOutputTensors(const GraphId &graph_id, const std::vector<
 void CPUSession::LoadInputData(const std::shared_ptr<KernelGraph> &kernel_graph,
                                const std::vector<tensor::TensorPtr> &inputs_const) const {
   MS_EXCEPTION_IF_NULL(kernel_graph);
-  auto &input_nodes = kernel_graph->inputs();
+  auto &input_nodes = kernel_graph->input_nodes();
   if (input_nodes.size() != inputs_const.size()) {
     MS_LOG(EXCEPTION) << "Input size " << inputs_const.size() << " is not equal to input node size "
                       << input_nodes.size();

mindspore/ccsrc/runtime/device/cpu/cpu_kernel_runtime.cc

@@ -130,7 +130,7 @@ void CPUKernelRuntime::AssignValueNodeAddress(session::KernelGraph *kernel_graph
 
 void CPUKernelRuntime::AssignInputNodeAddress(const session::KernelGraph *kernel_graph) {
   MS_EXCEPTION_IF_NULL(kernel_graph);
-  for (auto &item : kernel_graph->inputs()) {
+  for (auto &item : kernel_graph->input_nodes()) {
     MS_EXCEPTION_IF_NULL(item);
     if (item->isa<Parameter>()) {
       auto output_num = AnfAlgo::GetOutputTensorNum(item);
@@ -281,7 +281,7 @@ void CPUKernelRuntime::CreateOutputTensors(session::KernelGraph *kernel_graph,
   MS_EXCEPTION_IF_NULL(kernel_graph);
   MS_EXCEPTION_IF_NULL(outputs);
   MS_EXCEPTION_IF_NULL(tensor_to_node);
-  auto &input_nodes = kernel_graph->inputs();
+  auto &input_nodes = kernel_graph->input_nodes();
   if (input_nodes.size() != inputs.size()) {
     MS_LOG(EXCEPTION) << "Input size " << inputs.size() << " is not equal to input node size " << input_nodes.size();
   }
@@ -305,7 +305,7 @@ void CPUKernelRuntime::CreateOutputTensors(session::KernelGraph *kernel_graph,
 
 void CPUKernelRuntime::BindInputTensorAddressPtr(const session::KernelGraph &kernel_graph,
                                                  const std::vector<tensor::TensorPtr> &inputs) {
-  auto &input_nodes = kernel_graph.inputs();
+  auto &input_nodes = kernel_graph.input_nodes();
   if (input_nodes.size() != inputs.size()) {
     MS_LOG(EXCEPTION) << "Input size" << inputs.size() << " is not equal to input node size " << input_nodes.size();
   }

mindspore/core/abstract/abstract_value.h

@@ -1516,6 +1516,7 @@ class MS_CORE_API AbstractCSRTensor : public AbstractUndetermined {
   AbstractTensorPtr values_;
   AbstractTuplePtr dense_shape_;
 };
+using AbstractCSRTensorPtr = std::shared_ptr<AbstractCSRTensor>;
 
 class AbstractMonad : public AbstractBase {
  public:

tests/st/sparse/test_csr.py

@@ -15,9 +15,13 @@
 """smoke tests for CSR operations"""
 
 import pytest
+import numpy as np
+
 from mindspore import Tensor, CSRTensor, ms_function
 from mindspore.common import dtype as mstype
+from mindspore import nn, context
 
+context.set_context(mode=context.GRAPH_MODE)
 
 def compare_csr(csr1, csr2):
     assert isinstance(csr1, CSRTensor)
@@ -82,3 +86,62 @@ def test_csr_attr():
     csr1 = CSRTensor(*csr1_tuple)
     csr2 = CSRTensor(*csr2_tuple)
     compare_csr(csr1, csr2)
+
+
+@pytest.mark.level0
+@pytest.mark.platform_arm_ascend_training
+@pytest.mark.platform_x86_ascend_training
+@pytest.mark.platform_x86_gpu_training
+@pytest.mark.platform_x86_cpu
+@pytest.mark.env_onecard
+def test_csr_tensor_in_while():
+    """
+    Feature: Test CSRTensor in while loop.
+    Description: Test CSRTensor computation in while loop.
+    Expectation: Success.
+    """
+    class CSRTensorValuesDouble(nn.Cell):
+
+        def construct(self, x):
+            indptr = x.indptr
+            indices = x.indices
+            values = x.values * 2
+            shape = x.shape
+            return CSRTensor(indptr, indices, values, shape)
+
+    class CSRTensorValuesAdd2(nn.Cell):
+
+        def construct(self, x):
+            indptr = x.indptr
+            indices = x.indices
+            values = x.values + 2
+            shape = x.shape
+            return CSRTensor(indptr, indices, values, shape)
+
+    class CSRTensorWithControlWhile(nn.Cell):
+        def __init__(self, shape):
+            super().__init__()
+            self.op1 = CSRTensorValuesDouble()
+            self.op2 = CSRTensorValuesAdd2()
+            self.shape = shape
+
+        @ms_function
+        def construct(self, a, b, indptr, indices, values):
+            x = CSRTensor(indptr, indices, values, self.shape)
+            x = self.op2(x)
+            while a > b:
+                x = self.op1(x)
+                b = b + 1
+            return x
+    a = Tensor(3, mstype.int32)
+    b = Tensor(0, mstype.int32)
+    indptr = Tensor([0, 1, 2])
+    indices = Tensor([0, 1])
+    values = Tensor([1, 2], dtype=mstype.float32)
+    shape = (2, 6)
+    net = CSRTensorWithControlWhile(shape)
+    out = net(a, b, indptr, indices, values)
+    assert np.allclose(out.indptr.asnumpy(), indptr.asnumpy(), .0, .0)
+    assert np.allclose(out.indices.asnumpy(), indices.asnumpy(), .0, .0)
+    assert np.allclose((values.asnumpy() + 2) * 8, out.values.asnumpy(), .0, .0)
+    assert shape == out.shape