Merge from develop

CMakeLists.txt

@@ -145,6 +145,7 @@ rocm_create_package(
     DESCRIPTION "AMD's graph optimizer"
     MAINTAINER "Paul Fultz II <paul.fultz@amd.com>"
     LDCONFIG
+    PTH
     DEPENDS miopen-hip rocblas hip_hcc half
 )
 

Jenkinsfile

@@ -107,6 +107,10 @@ rocmtest tidy: rocmnode('rocmtest') { cmake_build ->
     stage('Clang Release') {
         cmake_build("hcc", "-DCMAKE_BUILD_TYPE=release")
     }
+
+    stage('Clang Release Python 3') {
+        cmake_build("hcc", "-DCMAKE_BUILD_TYPE=release -DPYTHON_EXECUTABLE=/usr/local/bin/python3")
+    }
 }, gcc5: rocmnode('rocmtest') { cmake_build ->
     stage('GCC 5 Debug') {
         cmake_build("g++-5", "-DCMAKE_BUILD_TYPE=debug")

requirements.txt

 google/protobuf -DCMAKE_POSITION_INDEPENDENT_CODE=On
-RadeonOpenCompute/rocm-cmake@ac45c6e --build
+RadeonOpenCompute/rocm-cmake@42f6740 --build
 ROCmSoftwarePlatform/rocBLAS@30a992ae02fda568688bcd190edd5e277d6674d9
 ROCmSoftwarePlatform/MIOpen@1.7.0
 blaze,https://bitbucket.org/blaze-lib/blaze/get/f0755dea0e03.tar.gz -X header -DHEADER_DIR=blaze

src/eliminate_contiguous.cpp

@@ -27,6 +27,13 @@ void eliminate_contiguous::apply(program& p) const
 {
     for(auto ins : iterator_for(p))
     {
+        // skip the reshape operator for now, since there is a bug
+        // for the transpose followed by a reshape
+        if(ins->name() == "reshape")
+        {
+            continue;
+        }
+
         // Make a copy so we can modify it while we iterate
         auto args = ins->inputs();
         for(auto arg : ins->inputs())

src/include/migraphx/literal.hpp

@@ -22,8 +22,8 @@ struct literal : raw_data<literal>
 {
     literal() {}
 
-    template <class U, class T = deduce<U>>
-    literal(U x) : buffer(make_shared_array<char>(sizeof(T))), m_shape(shape::get_type<T>{})
+    template <class U, class T = deduce<U>, shape::type_t ShapeType = shape::get_type<T>{}>
+    literal(U x) : buffer(make_shared_array<char>(sizeof(T))), m_shape(ShapeType)
     {
         static_assert(std::is_trivially_copyable<T>{}, "Literals can only be trivial types");
         *(reinterpret_cast<T*>(buffer.get())) = x;

src/include/migraphx/operators.hpp

@@ -757,43 +757,49 @@ struct gather
         // negative axis means counting dimensions from back
         int axis_index = (axis < 0) ? (n_dim + axis) : axis;
 
-        auto type        = inputs[0].type();
-        lens[axis_index] = inputs[1].elements();
-
-        return {type, lens};
-    }
+        auto type = inputs[0].type();
+        lens.erase(lens.begin() + axis_index);
+        if(!inputs[1].scalar())
+        {
+            auto ind_lens = inputs[1].lens();
+            lens.insert(lens.begin() + axis_index, ind_lens.begin(), ind_lens.end());
+        }
 
-    template <class T>
-    void compute_index(const T& out_idx,
-                       const int axis_index,
-                       const std::vector<std::size_t>& vec_indices,
-                       const std::size_t max_dim,
-                       T& in_idx) const
-    {
-        in_idx          = out_idx;
-        std::size_t idx = vec_indices.at(out_idx[axis_index]);
-        if(idx >= max_dim)
+        // for scalar output
+        if(lens.empty())
         {
-            MIGRAPHX_THROW("Gather: indices are out of range in input tensor");
+            return {type};
         }
-        in_idx[axis_index] = idx;
+
+        return {type, lens};
     }
 
     argument compute(const shape& output_shape, std::vector<argument> args) const
     {
         argument result{output_shape};
         // negative axis means counting dimensions from back
-        int axis_index = (axis < 0) ? (output_shape.lens().size() + axis) : axis;
+        int axis_index =
+            (axis < 0) ? static_cast<int>(args[0].get_shape().lens().size() + axis) : axis;
 
         // max dimension in axis
-        std::size_t max_dim = args[0].get_shape().lens()[axis_index];
-        std::vector<std::size_t> vec_indices;
-        args[1].visit([&](auto indices) { vec_indices.assign(indices.begin(), indices.end()); });
-        visit_all(result, args[0])([&](auto output, auto input) {
-            std::vector<std::size_t> in_idx;
-            shape_for_each(output.get_shape(), [&](const auto& idx) {
-                this->compute_index(idx, axis_index, vec_indices, max_dim, in_idx);
-                output(idx.begin(), idx.end()) = input(in_idx.begin(), in_idx.end());
+        visit_all(result, args[0])([&](auto output, auto data) {
+            args[1].visit([&](auto indices) {
+                if(output_shape.scalar())
+                {
+                    output[0] = data[indices.front()];
+                }
+                else
+                {
+                    auto out_lens        = data.get_shape().lens();
+                    out_lens[axis_index] = indices.get_shape().elements();
+                    migraphx::shape out_comp_shape{data.get_shape().type(), out_lens};
+                    shape_for_each(out_comp_shape, [&](const auto& out_idx) {
+                        auto data_idx        = out_idx;
+                        data_idx[axis_index] = indices[data_idx[axis_index]];
+                        output[out_comp_shape.index(out_idx.begin(), out_idx.end())] =
+                            data(data_idx.begin(), data_idx.end());
+                    });
+                }
             });
         });
 
@@ -1259,6 +1265,57 @@ struct gru
     }
 };
 
+struct lstm
+{
+    std::size_t hidden_size = 1;
+    std::vector<operation> actv_funcs{sigmoid{}, tanh{}, tanh{}};
+    rnn_direction direction = rnn_direction::forward;
+    float clip              = 0.0f;
+    int input_forget        = 0;
+
+    std::string name() const { return "lstm"; }
+    shape compute_shape(std::vector<shape> inputs) const
+    {
+        auto in_dims     = inputs[0].lens();
+        auto hidden_dims = inputs[2].lens();
+        if(hidden_size != hidden_dims[2])
+        {
+            MIGRAPHX_THROW("LSTM: hidden size mismatch in attribute and input");
+        }
+
+        std::size_t num_directions = 1;
+        if(direction == rnn_direction::bidirectional)
+        {
+            num_directions = 2;
+        }
+
+        if(num_directions != hidden_dims[0])
+        {
+            MIGRAPHX_THROW("LSTM: num_direction does not match the direction attribute");
+        }
+
+        std::vector<std::size_t> out_dims(in_dims);
+        out_dims.insert(out_dims.begin() + 1, num_directions);
+        out_dims.back() = hidden_size;
+
+        return {inputs[0].type(), out_dims};
+    }
+};
+
+struct lstm_last_cell_output
+{
+    std::string name() const { return "lstm_last_cell_output"; }
+    shape compute_shape(std::vector<shape> inputs) const
+    {
+        check_shapes{inputs, *this}.has(1);
+        auto dims = inputs[0].lens();
+
+        // remove the first dimension, remaing are output shape
+        dims.erase(dims.begin());
+        return {inputs[0].type(), dims};
+    }
+};
+
 struct undefined
 {
     std::string name() const { return "undefined"; }

src/include/migraphx/rewrite_rnn.hpp

@@ -45,6 +45,18 @@ struct rewrite_rnn
                                           const operation& actv_func2) const;
 
     std::vector<operation> gru_actv_funcs(instruction_ref ins) const;
+
+    // for lstm operators
+    void apply_lstm(program& prog, instruction_ref ins) const;
+    std::vector<instruction_ref> lstm_cell(bool is_forward,
+                                           program& prog,
+                                           instruction_ref ins,
+                                           std::vector<instruction_ref> inputs,
+                                           const operation& actv_func1,
+                                           const operation& actv_func2,
+                                           const operation& actv_func3) const;
+
+    std::vector<operation> lstm_actv_funcs(instruction_ref ins) const;
 };
 
 } // namespace MIGRAPHX_INLINE_NS

src/include/migraphx/tensor_view.hpp

@@ -124,6 +124,8 @@ struct tensor_view
             return m_data + this->size();
     }
 
+    std::vector<T> to_vector() const { return std::vector<T>(this->begin(), this->end()); }
+
     friend std::ostream& operator<<(std::ostream& os, const tensor_view<T>& x)
     {
         if(!x.empty())

src/onnx/onnx.cpp

@@ -89,6 +89,7 @@ struct onnx_parser
         add_mem_op("Transpose", &onnx_parser::parse_transpose);
         add_mem_op("RNN", &onnx_parser::parse_rnn);
         add_mem_op("GRU", &onnx_parser::parse_gru);
+        add_mem_op("LSTM", &onnx_parser::parse_lstm);
         add_mem_op("Pad", &onnx_parser::parse_pad);
 
         // init the activation function map
@@ -354,7 +355,9 @@ struct onnx_parser
         }
         if(args.size() == 2)
         {
-            literal s = args[1]->get_literal();
+            auto s = args[1]->eval();
+            if(s.empty())
+                MIGRAPHX_THROW("Dynamic shape is not supported.");
             s.visit([&](auto v) { copy(v, std::back_inserter(op.dims)); });
         }
         return prog.add_instruction(op, args[0]);
@@ -433,7 +436,15 @@ struct onnx_parser
                                    attribute_map attributes,
                                    const std::vector<instruction_ref>&)
     {
-        literal v = parse_value(attributes.at("value"));
+        literal v     = parse_value(attributes.at("value"));
+        auto dim_size = attributes.at("value").t().dims_size();
+        // if dim_size is 0, it is a scalar
+        if(dim_size == 0)
+        {
+            migraphx::shape scalar_shape{v.get_shape().type()};
+            return prog.add_literal(migraphx::literal{scalar_shape, v.data()});
+        }
+
         return prog.add_literal(v);
     }
 
@@ -460,6 +471,7 @@ struct onnx_parser
         {
             transb = parse_value(attributes.at("transB")).at<bool>();
         }
+
         std::vector<int64_t> perm = {1, 0};
         auto l1 = (transa) ? prog.add_instruction(op::transpose{perm}, args[0]) : args[0];
         auto l2 = (transb) ? prog.add_instruction(op::transpose{perm}, args[1]) : args[1];
@@ -480,7 +492,10 @@ struct onnx_parser
                 return add_broadcastable_binary_op(l3, l4, op::add{});
             }
         }
-        return prog.add_instruction(op::dot{alpha, beta}, l1, l2);
+
+        auto dot_res = prog.add_instruction(op::dot{alpha, beta}, l1, l2);
+
+        return dot_res;
     }
 
     instruction_ref
@@ -749,15 +764,17 @@ struct onnx_parser
         {
             auto names = attributes.at("activations").strings();
             vec_names.clear();
-            for_each(names.begin(), names.end(), [&](auto& fn) { vec_names.push_back(fn); });
+            vec_names.resize(names.size());
+            std::copy(names.begin(), names.end(), vec_names.begin());
         }
 
-        for_each(vec_names.begin(), vec_names.end(), [&](auto& fn) {
-            if(map_actv_funcs.count(fn) == 0)
-            {
-                MIGRAPHX_THROW("RNN: activation function " + std::string(fn) + " not supported");
-            }
+        auto name_it = std::find_if(vec_names.begin(), vec_names.end(), [&](auto& name) {
+            return (map_actv_funcs.count(name) == 0);
         });
+        if(name_it != vec_names.end())
+        {
+            MIGRAPHX_THROW("RNN: activation function " + std::string(*name_it) + " not supported");
+        }
 
         // bidirectional case should have two activation functions.
         // one is for forward, and the other is for reverse.
@@ -839,8 +856,7 @@ struct onnx_parser
             auto names = attributes.at("activations").strings();
             vec_names.clear();
             vec_names.resize(names.size());
-            std::transform(
-                names.begin(), names.end(), vec_names.begin(), [](auto& str) { return str; });
+            std::copy(names.begin(), names.end(), vec_names.begin());
         }
 
         // need 4 activation functions
@@ -878,12 +894,13 @@ struct onnx_parser
             }
         }
 
-        for_each(vec_names.begin(), vec_names.end(), [&](auto& name) {
-            if(map_actv_funcs.count(name) == 0)
-            {
-                MIGRAPHX_THROW("GRU: activation function " + std::string(name) + " not supported");
-            }
+        auto name_it = std::find_if(vec_names.begin(), vec_names.end(), [&](auto& name) {
+            return (map_actv_funcs.count(name) == 0);
         });
+        if(name_it != vec_names.end())
+        {
+            MIGRAPHX_THROW("GRU: activation function " + std::string(*name_it) + " not supported");
+        }
 
         std::vector<operation> vec_actv_funcs(vec_names.size());
         std::transform(vec_names.begin(), vec_names.end(), vec_actv_funcs.begin(), [&](auto& name) {
@@ -920,6 +937,178 @@ struct onnx_parser
         return {hidden_states, last_output};
     }
 
+    std::vector<instruction_ref>
+    parse_lstm(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
+    {
+        migraphx::shape input_shape = args[0]->get_shape();
+        std::size_t hidden_size     = args[2]->get_shape().lens()[2];
+
+        if(contains(attributes, "hidden_size"))
+        {
+            std::size_t hidden_size_att = parse_value(attributes.at("hidden_size")).at<int>();
+            if(hidden_size != hidden_size_att)
+            {
+                MIGRAPHX_THROW("LSTM: hidden size mismatch in input and attribute");
+            }
+        }
+
+        // Handling of direction to be added later
+        std::string direction{"forward"};
+        if(contains(attributes, "direction"))
+        {
+            direction = attributes.at("direction").s();
+        }
+
+        op::rnn_direction dirct = op::rnn_direction::forward;
+        if(direction == "bidirectional")
+        {
+            dirct = op::rnn_direction::bidirectional;
+        }
+        else if(direction == "reverse")
+        {
+            dirct = op::rnn_direction::reverse;
+        }
+        else if(direction == "forward")
+        {
+            dirct = op::rnn_direction::forward;
+        }
+        else
+        {
+            MIGRAPHX_THROW("LSTM: incorrect direction attribute");
+        }
+
+        std::vector<std::string> vec_names = {"sigmoid", "tanh", "tanh"};
+        if(contains(attributes, "activations"))
+        {
+            auto names = attributes.at("activations").strings();
+            vec_names.clear();
+            vec_names.resize(names.size());
+            std::copy(names.begin(), names.end(), vec_names.begin());
+        }
+
+        // need 6 activation functions for bidirectional directions
+        if(dirct == op::rnn_direction::bidirectional)
+        {
+            // 6 activation functions are used in the bidirectional
+            // scenario. No spec is provided in onnx::operator. we
+            // use the algorithm that: if 1 actv function is provided,
+            // repeat 1st six times. If 2 actv functins are provided,
+            // repeat 2nd once, then repeat all three once
+            // if 3 actv funcs are provide, repeat all three once.
+            // the same algorithm is used for 4, 5, and 6 actv funcions
+            // provided. This may need change later
+            switch(vec_names.size())
+            {
+            case 1:
+                vec_names = {vec_names.at(0),
+                             vec_names.at(0),
+                             vec_names.at(0),
+                             vec_names.at(0),
+                             vec_names.at(0),
+                             vec_names.at(0)};
+                break;
+
+            case 2:
+                // repeat the 2nd actv func once, then repeat all three another time
+                vec_names = {vec_names.at(0),
+                             vec_names.at(1),
+                             vec_names.at(1),
+                             vec_names.at(0),
+                             vec_names.at(1),
+                             vec_names.at(1)};
+                break;
+
+            case 3:
+                // repeat all three actv funcs once
+                vec_names = {vec_names.at(0),
+                             vec_names.at(1),
+                             vec_names.at(2),
+                             vec_names.at(0),
+                             vec_names.at(1),
+                             vec_names.at(2)};
+                break;
+
+            case 4:
+                vec_names = {vec_names.at(0),
+                             vec_names.at(1),
+                             vec_names.at(2),
+                             vec_names.at(3),
+                             vec_names.at(3),
+                             vec_names.at(3)};
+                break;
+
+            case 5:
+                vec_names = {vec_names.at(0),
+                             vec_names.at(1),
+                             vec_names.at(2),
+                             vec_names.at(3),
+                             vec_names.at(4),
+                             vec_names.at(4)};
+                break;
+
+            default: break;
+            }
+        }
+        else
+        {
+            switch(vec_names.size())
+            {
+            case 1: vec_names = {vec_names.at(0), vec_names.at(0), vec_names.at(0)}; break;
+
+            case 2:
+                // repeat the 2nd actv func once, so we have 3 actv funcs
+                vec_names = {vec_names.at(0), vec_names.at(1), vec_names.at(1)};
+                break;
+
+            default: break;
+            }
+        }
+
+        auto name_it = std::find_if(vec_names.begin(), vec_names.end(), [&](auto& name) {
+            return (map_actv_funcs.count(name) == 0);
+        });
+        if(name_it != vec_names.end())
+        {
+            MIGRAPHX_THROW("LSTM: activation function " + std::string(*name_it) + " not supported");
+        }
+
+        std::vector<operation> vec_actv_funcs(vec_names.size());
+        std::transform(vec_names.begin(), vec_names.end(), vec_actv_funcs.begin(), [&](auto& name) {
+            return map_actv_funcs[name];
+        });
+
+        float clip = 0.0;
+        if(contains(attributes, "clip"))
+        {
+            clip = parse_value(attributes.at("clip")).at<float>();
+        }
+
+        int input_forget = 0;
+        if(contains(attributes, "input_forget"))
+        {
+            input_forget = parse_value(attributes.at("input_forget")).at<int>();
+        }
+
+        // append undefined opeator to make 6 arguments
+        if(args.size() < 8)
+        {
+            auto ins = prog.add_instruction(op::undefined{});
+            args.insert(args.end(), 8 - args.size(), ins);
+        }
+
+        // first output for concatenation of hidden states
+        auto hidden_states = prog.add_instruction(
+            op::lstm{hidden_size, vec_actv_funcs, dirct, clip, input_forget}, std::move(args));
+
+        // second output for last lstm output
+        auto last_output = prog.add_instruction(op::rnn_last_output{}, hidden_states);
+
+        // third output for last cell output
+        auto last_cell_output = prog.add_instruction(op::lstm_last_cell_output{}, hidden_states);
+
+        return {hidden_states, last_output, last_cell_output};
+    }
+
     void parse_from(std::istream& is)
     {
         onnx::ModelProto model;
@@ -1133,7 +1322,15 @@ struct onnx_parser
         case onnx::TensorProto::BOOL:
             return literal{{shape::int32_type, dims}, t.int32_data().begin(), t.int32_data().end()};
         case onnx::TensorProto::FLOAT16:
-            return literal{{shape::half_type, dims}, t.float_data().begin(), t.float_data().end()};
+        {
+            std::vector<uint16_t> data_uint16(t.int32_data().begin(), t.int32_data().end());
+            std::vector<half> data_half;
+            std::transform(data_uint16.begin(),
+                           data_uint16.end(),
+                           std::back_inserter(data_half),
+                           [](uint16_t raw_val) { return *reinterpret_cast<half*>(&raw_val); });
+            return literal{{shape::half_type, dims}, data_half.begin(), data_half.end()};
+        }
         case onnx::TensorProto::DOUBLE:
             return literal{
                 {shape::double_type, dims}, t.double_data().begin(), t.double_data().end()};

src/opt/memory_coloring_impl.cpp

@@ -207,9 +207,8 @@ void memory_coloring_impl::rewrite()
 
             if(is_allocate(ins))
             {
-                assert(!ins->inputs().empty());
                 p_program->replace_instruction(
-                    ins, op::load{ins->inputs().at(0)->get_shape(), offset}, scratch_param);
+                    ins, op::load{ins->get_shape(), offset}, scratch_param);
             }
             else if(is_literal(ins))
             {

src/program.cpp

@@ -2,6 +2,7 @@
 #include <migraphx/stringutils.hpp>
 #include <migraphx/instruction.hpp>
 #include <migraphx/operators.hpp>
+#include <migraphx/target.hpp>
 #include <migraphx/env.hpp>
 #include <migraphx/ranges.hpp>
 #include <migraphx/time.hpp>
@@ -353,13 +354,17 @@ argument generic_eval(const program& p,
         }
         else if(ins->name() == "@param")
         {
-            results.emplace(ins, trace(ins, [&] {
-                                auto param_name =
-                                    any_cast<builtin::param>(ins->get_operator()).parameter;
-                                if(not contains(params, param_name))
-                                    MIGRAPHX_THROW("Parameter not found: " + param_name);
-                                return params.at(param_name);
-                            }));
+            results.emplace(
+                ins, trace(ins, [&] {
+                    auto param_name = any_cast<builtin::param>(ins->get_operator()).parameter;
+                    if(not contains(params, param_name))
+                        MIGRAPHX_THROW("Parameter not found: " + param_name);
+                    auto param = params.at(param_name);
+                    if(param.get_shape() != ins->get_shape())
+                        MIGRAPHX_THROW("Incorrect shape {" + to_string(param.get_shape()) +
+                                       "} for parameter: " + param_name);
+                    return param;
+                }));
         }
         else if(ins->name() == "@outline")
         {

src/py/CMakeLists.txt

 
 option(MIGRAPHX_ENABLE_PYTHON "Enable python bindings" ON)
 if(MIGRAPHX_ENABLE_PYTHON)
-find_program(DEFAULT_PYTHON_EXE python)
-if(DEFAULT_PYTHON_EXE)
-set(PYTHON_EXECUTABLE ${DEFAULT_PYTHON_EXE} CACHE PATH "Path to python executable")
-endif()
-find_package(pybind11 REQUIRED)
-pybind11_add_module(migraphx_py migraphx_py.cpp)
-set_target_properties(migraphx_py PROPERTIES 
-    OUTPUT_NAME migraphx
-    C_VISIBILITY_PRESET hidden
-    CXX_VISIBILITY_PRESET hidden
-)
-target_link_libraries(migraphx_py PRIVATE migraphx migraphx_onnx migraphx_cpu)
-if(MIGRAPHX_ENABLE_GPU)
-target_link_libraries(migraphx_py PRIVATE migraphx_gpu)
-target_compile_definitions(migraphx_py PRIVATE -DHAVE_GPU)
-endif()
+    find_program(DEFAULT_PYTHON_EXE python)
+    if(DEFAULT_PYTHON_EXE)
+        set(PYTHON_EXECUTABLE ${DEFAULT_PYTHON_EXE} CACHE PATH "Path to python executable")
+    endif()
+    find_package(pybind11 REQUIRED)
+    pybind11_add_module(migraphx_py migraphx_py.cpp)
+    set_target_properties(migraphx_py PROPERTIES 
+        OUTPUT_NAME migraphx
+        C_VISIBILITY_PRESET hidden
+        CXX_VISIBILITY_PRESET hidden
+    )
+    target_link_libraries(migraphx_py PRIVATE migraphx migraphx_onnx migraphx_cpu)
+    if(MIGRAPHX_ENABLE_GPU)
+        target_link_libraries(migraphx_py PRIVATE migraphx_gpu)
+        target_compile_definitions(migraphx_py PRIVATE -DHAVE_GPU)
+    endif()
+    rocm_install_targets(TARGETS migraphx_py)
 endif()

src/py/migraphx_py.cpp

@@ -28,6 +28,11 @@ struct throw_half
     {
         throw std::runtime_error("Half not supported in python yet.");
     }
+
+    void operator()(migraphx::tensor_view<migraphx::half>) const
+    {
+        throw std::runtime_error("Half not supported in python yet.");
+    }
 };
 
 template <class F>
@@ -42,6 +47,8 @@ struct skip_half
     }
 
     void operator()(migraphx::shape::as<migraphx::half>) const {}
+
+    void operator()(migraphx::tensor_view<migraphx::half>) const {}
 };
 
 template <class F>
@@ -50,6 +57,12 @@ void visit_type(const migraphx::shape& s, F f)
     s.visit_type(throw_half<F>{f});
 }
 
+template <class T, class F>
+void visit(const migraphx::raw_data<T>& x, F f)
+{
+    x.visit(throw_half<F>{f});
+}
+
 template <class F>
 void visit_types(F f)
 {
@@ -60,6 +73,9 @@ template <class T>
 py::buffer_info to_buffer_info(T& x)
 {
     migraphx::shape s = x.get_shape();
+    auto strides      = s.strides();
+    std::transform(
+        strides.begin(), strides.end(), strides.begin(), [&](auto i) { return i * s.type_size(); });
     py::buffer_info b;
     visit_type(s, [&](auto as) {
         b = py::buffer_info(x.data(),
@@ -67,7 +83,7 @@ py::buffer_info to_buffer_info(T& x)
                             py::format_descriptor<decltype(as())>::format(),
                             s.lens().size(),
                             s.lens(),
-                            s.strides());
+                            strides);
     });
     return b;
 }
@@ -75,11 +91,20 @@ py::buffer_info to_buffer_info(T& x)
 migraphx::shape to_shape(const py::buffer_info& info)
 {
     migraphx::shape::type_t t;
+    std::size_t n = 0;
     visit_types([&](auto as) {
         if(info.format == py::format_descriptor<decltype(as())>::format())
+        {
             t = as.type_enum();
+            n = sizeof(as());
+        }
+
     });
-    return migraphx::shape{t, info.shape, info.strides};
+    auto strides = info.strides;
+    std::transform(strides.begin(), strides.end(), strides.begin(), [&](auto i) -> std::size_t {
+        return n > 0 ? i / n : 0;
+    });
+    return migraphx::shape{t, info.shape, strides};
 }
 
 PYBIND11_MODULE(migraphx, m)
@@ -108,6 +133,13 @@ PYBIND11_MODULE(migraphx, m)
                  py::buffer_info info = b.request();
                  new(&x) migraphx::argument(to_shape(info), info.ptr);
              })
+        .def("get_shape", &migraphx::argument::get_shape)
+        .def("tolist",
+             [](migraphx::argument& x) {
+                 py::list l{x.get_shape().elements()};
+                 visit(x, [&](auto data) { l = py::cast(data.to_vector()); });
+                 return l;
+             })
         .def("__eq__", std::equal_to<migraphx::argument>{})
         .def("__ne__", std::not_equal_to<migraphx::argument>{})
         .def("__repr__", [](const migraphx::argument& x) { return migraphx::to_string(x); });

src/shape.cpp

@@ -19,7 +19,7 @@ struct shape_impl
 
     shape_impl() : m_type(shape::float_type), m_standard(false) {}
 
-    shape_impl(shape::type_t t) : m_type(t), m_lens({1}), m_strides({1}), m_standard(true) {}
+    shape_impl(shape::type_t t) : m_type(t), m_lens({1}), m_strides({0}), m_standard(true) {}
     shape_impl(shape::type_t t, std::vector<std::size_t> l)
         : m_type(t), m_lens(std::move(l)), m_standard(true)
     {

src/targets/cpu/include/migraphx/cpu/target.hpp

@@ -7,6 +7,7 @@
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
+struct pass;
 namespace cpu {
 
 struct target

src/targets/cpu/target.cpp

 
 #include <migraphx/cpu/target.hpp>
 #include <migraphx/cpu/lowering.hpp>
+#include <migraphx/pass.hpp>
 #include <migraphx/auto_contiguous.hpp>
 #include <migraphx/rewrite_rnn.hpp>
 #include <migraphx/dead_code_elimination.hpp>

src/targets/gpu/abs.cpp

 #include <migraphx/gpu/abs.hpp>
-#include <migraphx/operators.hpp>
-#include <migraphx/manage_ptr.hpp>
-#include <migraphx/gpu/miopen.hpp>
-#include <utility>
+#include <migraphx/gpu/context.hpp>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {

src/targets/gpu/batchnorm.cpp

 #include <migraphx/gpu/batchnorm.hpp>
-#include <migraphx/operators.hpp>
-#include <migraphx/manage_ptr.hpp>
-#include <migraphx/gpu/miopen.hpp>
-#include <utility>
+#include <migraphx/gpu/context.hpp>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {

src/targets/gpu/concat.cpp

 #include <migraphx/gpu/concat.hpp>
-#include <migraphx/operators.hpp>
-#include <migraphx/manage_ptr.hpp>
-#include <migraphx/gpu/miopen.hpp>
+#include <migraphx/gpu/context.hpp>
 #include <migraphx/gpu/device/concat.hpp>
-#include <utility>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {