feat(test): add some comment and test in imperative

GitOrigin-RevId: c2363abe29

imperative/src/impl/interpreter/interpreter_impl.h

@@ -27,6 +27,11 @@ struct InterpreterImpl : Interpreter {
     std::unique_ptr<Channel> create_channel() override;
 };
 
+/*!
+ * \brief implement Channel to execute the commands asynchronously,
+ * almost commands are executed by the worker threads, commands are sent
+ * by the interface
+ */
 struct ChannelImpl : Interpreter::Channel, NonCopyableObj, NonMoveableObj {
     ChannelImpl();
     ~ChannelImpl() override;
@@ -304,3 +309,5 @@ private:
 };
 
 }  // namespace mgb::imperative::interpreter::intl
+
+// vim: syntax=cpp.doxygen foldmethod=marker foldmarker=f{{{,f}}}

imperative/src/include/megbrain/imperative/interpreter.h

@@ -7,6 +7,7 @@
 
 namespace mgb::imperative::interpreter {
 
+//! nested throw the stored exceptions
 struct AsyncError : std::nested_exception, std::exception {
     const char* what() const noexcept {
         try {
@@ -20,9 +21,16 @@ struct AsyncError : std::nested_exception, std::exception {
 };
 
 struct Interpreter {
+    /*!
+     * HandleImpl* just act as a key to represent TensorInfo
+     */
     struct HandleImpl {};
     using Handle = HandleImpl*;
 
+    /*!
+     * \brief the base command execution interface, Channel is similar to channel in
+     * golang, it executes the commands put into asynchronously.
+     */
     struct Channel {
         virtual ~Channel() = default;
 
@@ -66,3 +74,5 @@ protected:
 };
 
 }  // namespace mgb::imperative::interpreter
+
+// vim: syntax=cpp.doxygen foldmethod=marker foldmarker=f{{{,f}}}

imperative/src/test/interpreter_test.cpp

@@ -0,0 +1,81 @@
+#include "megbrain/imperative/interpreter.h"
+#include "../impl/interpreter/tensor_info.h"
+#include "./helper.h"
+#include "megbrain/comp_node_env.h"
+#include "megbrain/imperative/blob_manager.h"
+#include "megbrain/imperative/ops/opr_attr.h"
+#include "megbrain/imperative/physical_tensor.h"
+#include "megbrain/opr/basic_arith.h"
+#include "megbrain/opr/basic_arith_wrapper.h"
+#include "megbrain/opr/dnn/batch_norm.h"
+#include "megbrain/opr/dnn/convolution.h"
+#include "megbrain/opr/tensor_manip.h"
+#include "megbrain/opr/utility.h"
+
+using namespace mgb;
+using namespace cg;
+using namespace imperative;
+using namespace interpreter;
+
+TEST(TestImperative, InterpreterPut) {
+    HostTensorGenerator<> gen;
+    auto h0 = gen({3});
+    auto&& channel = Interpreter::inst().create_channel();
+    auto tensor_handle = channel->put(*h0, true);
+    auto tensor_info = reinterpret_cast<intl::TensorInfo*>(tensor_handle);
+    channel->sync();
+    ASSERT_TRUE(tensor_info->status == intl::TensorInfo::Produced);
+    //! because tensor elems is less than TensorShape::MAX_NDIM, stored it
+    //! directly
+    ASSERT_EQ(tensor_info->ptr->get_value().raw_ptr(), h0->raw_ptr());
+    auto shape = channel->get_shape(tensor_handle);
+    ASSERT_TRUE(shape.ndim == 1);
+    ASSERT_TRUE(shape.total_nr_elems() == 3);
+
+    auto h2 = gen({10});
+    auto tensor_handle2 = channel->put(*h2, false);
+    auto tensor_handle2_once = channel->put(*h2, false);
+    channel->sync();
+    ASSERT_NE(tensor_handle2_once, tensor_handle2);
+    auto finded = MultiCNConstTensorCache::inst().lookup(*h2);
+    ASSERT_TRUE(finded.get());
+    //! Device tensor ptr is not equal host tensor ptr
+    ASSERT_NE(finded->raw_ptr_not_for_readwrite(), h2->raw_ptr());
+    channel->del(tensor_handle);
+    channel->del(tensor_handle2);
+    channel->del(tensor_handle2_once);
+}
+
+TEST(TestImperative, InterpreterApplyOp) {
+    HostTensorGenerator<> gen;
+    size_t add = 2, dim0 = 5, dim1 = 10;
+    auto h0 = gen({1});
+    h0->ptr<float>()[0] = add;
+    auto h1 = gen({dim0, dim1});
+    for (size_t i = 0; i < dim0 * dim1; i++) {
+        h1->ptr<float>()[i] = i;
+    }
+    auto&& channel = Interpreter::inst().create_channel();
+    auto tensor_handle0 = channel->put(*h0, false);
+    auto tensor_handle1 = channel->put(*h1, false);
+    SmallVector<Interpreter::Handle> inputs{tensor_handle0, tensor_handle1};
+
+    auto op = OprAttr::make("Elemwise");
+    auto&& attr = op->cast_final_safe<OprAttr>();
+    using Param = opr::Elemwise::Param;
+    Param param{Param::Mode::ADD};
+    attr.param.write_pod(param);
+
+    auto outputs = channel->apply_op(op, inputs);
+    channel->sync();
+    auto out_tensor = reinterpret_cast<intl::TensorInfo*>(outputs[0])->ptr->get_value();
+    ASSERT_EQ(out_tensor.layout().ndim, 2);
+    ASSERT_EQ(out_tensor.shape(0), dim0);
+    ASSERT_EQ(out_tensor.shape(1), dim1);
+    float* output = out_tensor.ptr<float>();
+    for (size_t i = 0; i < dim0 * dim1; i++) {
+        ASSERT_EQ(output[i], i + add);
+    }
+}
+
+// vim: syntax=cpp.doxygen foldmethod=marker foldmarker=f{{{,f}}}