fix (frontend): 修改get输入输出，内存大小，增加compiler设置输入接口

src/02hardware/src/devices/nvidia/device.cc

@@ -23,7 +23,7 @@ namespace refactor::hardware {
 
         size_t free, total;
         CUDA_ASSERT(cudaMemGetInfo(&free, &total));
-        auto size = std::min(free, std::max(5ul << 30, total * 4 / 5));
+        auto size = free * 9 / 10;
         cudaDeviceProp prop;
         CUDA_ASSERT(cudaGetDeviceProperties(&prop, 0));
         size_t alignment = prop.textureAlignment;

src/09python_ffi/src/compiler.cc

@@ -19,7 +19,21 @@ namespace refactor::python_ffi {
     }
 
     void
-    Compiler::setInput(size_t index, int dataType, DimVec dims) {
+    Compiler::setInput(size_t index, pybind11::array data) {
+        ASSERT(index < _g.internal().topology.globalInputsCount(),
+               "Input {} not exist", index);
+
+        Shape shape(data.ndim(), DimExpr(1));
+        std::transform(std::execution::unseq,
+                       data.shape(), data.shape() + data.ndim(), shape.begin(),
+                       [](auto const &d) { return DimExpr(d); });
+        auto ans = Tensor::share(parseNumpyDType(data.dtype()), std::move(shape), {});
+        std::memcpy(ans->malloc(), data.data(), data.nbytes());
+        _g.internal().edges[index].tensor = std::move(ans);
+    }
+
+    void
+    Compiler::setInputInfo(size_t index, int dataType, DimVec dims) {
         ASSERT(index < _g.internal().topology.globalInputsCount(),
                "Input {} not exist", index);
 

src/09python_ffi/src/compiler.h

@@ -15,7 +15,8 @@ namespace refactor::python_ffi {
     public:
         explicit Compiler(frontend::Graph);
         void substitute(CStr, int64_t);
-        void setInput(size_t index, int dataType, DimVec dims);
+        void setInput(size_t index, pybind11::array);
+        void setInputInfo(size_t index, int dataType, DimVec dims);
         std::unordered_set<std::string> fillEdgeInfo(bool calculate);
         Arc<Executor> compileOn(
             Arc<hardware::Device> device,

src/09python_ffi/src/executor.cc

@@ -33,23 +33,23 @@ namespace refactor::python_ffi {
     }
 
     void Executor::setInput(count_t i, pybind11::array data) {
-        i = _stream.graph().topology.globalInputs().at(i);
+        i = _graph.internal().contiguous().topology.globalInputs().at(i);
 
         auto const &tensor = *_graph.internal().contiguous().edges[i].tensor;
         ASSERT(tensor.bytesSize() == static_cast<size_t>(data.nbytes()), "input size mismatch");
         _stream.setData(i, data.data(), data.nbytes());
     }
 
     void Executor::setInputBlob(count_t i, Arc<hardware::Device::Blob> blob) {
-        i = _stream.graph().topology.globalInputs().at(i);
+        i = _graph.internal().contiguous().topology.globalInputs().at(i);
 
         auto const &tensor = *_graph.internal().contiguous().edges[i].tensor;
         ASSERT(tensor.bytesSize() == blob->size(), "input size mismatch");
         _stream.setData(i, std::move(blob));
     }
 
     auto Executor::getOutput(count_t i) const -> pybind11::array {
-        i = _stream.graph().topology.globalOutputs().at(i);
+        i = _graph.internal().contiguous().topology.globalOutputs().at(i);
 
         auto const &tensor = *_graph.internal().contiguous().edges[i].tensor;
         auto ans = pybind11::array(buildNumpyDType(tensor.dataType), std::move(tensor.shape));
@@ -58,7 +58,7 @@ namespace refactor::python_ffi {
     }
 
     auto Executor::getOutputBlob(count_t i) const -> Arc<hardware::Device::Blob> {
-        i = _stream.graph().topology.globalOutputs().at(i);
+        i = _graph.internal().contiguous().topology.globalOutputs().at(i);
 
         return _stream.getData(i);
     }

src/09python_ffi/src/main.cpp

@@ -34,6 +34,7 @@ namespace refactor::python_ffi {
         py::class_<Compiler , Arc<Compiler>>(m, "Compiler" )
             .def("substitute"      , &Compiler::substitute       , return_::automatic )
             .def("set_input"       , &Compiler::setInput         , return_::automatic )
+            .def("set_input_info"  , &Compiler::setInputInfo     , return_::automatic )
             .def("check_variables" , &Compiler::fillEdgeInfo     , return_::move      )
             .def("zero_inputs"     , &Compiler::zeroInputs       , return_::move      )
             .def("get_tensor"      , &Compiler::getTensor        , return_::move      )