Merge branch 'develop' into verify_fp16_tol

src/include/migraphx/op/scatter.hpp

@@ -66,7 +66,7 @@ struct scatter : op_name<Derived>
 
     shape normalize_compute_shape(std::vector<shape> inputs) const
     {
-        check_shapes{inputs, *this}.has(3).standard();
+        check_shapes{inputs, *this}.has(3);
         // If non-packed, this converts to a packed output while preserving permutation of tensor
         return inputs.front().with_lens(inputs.front().lens());
     }

src/include/migraphx/optional.hpp

@@ -29,6 +29,17 @@
 #if defined(CPPCHECK)
 #define MIGRAPHX_HAS_OPTIONAL 1
 #define MIGRAPHX_HAS_OPTIONAL_TS 1
+#elif defined(_WIN32)
+#if _MSC_VER >= 1920
+#define MIGRAPHX_HAS_OPTIONAL 1
+#define MIGRAPHX_HAS_OPTIONAL_TS 0
+#elif _MSC_VER >= 1900
+#define MIGRAPHX_HAS_OPTIONAL 0
+#define MIGRAPHX_HAS_OPTIONAL_TS 1
+#else
+#define MIGRAPHX_HAS_OPTIONAL 0
+#define MIGRAPHX_HAS_OPTIONAL_TS 0
+#endif
 #elif defined(__has_include)
 #if __has_include(<optional>) && __cplusplus >= 201703L
 #define MIGRAPHX_HAS_OPTIONAL 1

src/include/migraphx/run_loop.hpp

@@ -31,6 +31,7 @@
 #include <migraphx/module.hpp>
 #include <migraphx/config.hpp>
 #include <migraphx/ranges.hpp>
+#include <array>
 #include <string>
 
 namespace migraphx {

src/include/migraphx/source_location.hpp

@@ -24,6 +24,7 @@
 #ifndef MIGRAPHX_GUARD_MIGRAPHX_SOURCE_LOCATION_HPP
 #define MIGRAPHX_GUARD_MIGRAPHX_SOURCE_LOCATION_HPP
 
+#include <cstdint>
 #include <migraphx/config.hpp>
 
 #if defined(CPPCHECK)

src/include/migraphx/tmp_dir.hpp

@@ -34,6 +34,7 @@ struct MIGRAPHX_EXPORT tmp_dir
 {
     fs::path path;
     tmp_dir(const std::string& prefix = "");
+    tmp_dir(tmp_dir&&) = default;
 
     void execute(const std::string& exe, const std::string& args) const;
 

src/include/migraphx/type_name.hpp

@@ -34,7 +34,7 @@ template <class PrivateMigraphTypeNameProbe>
 std::string compute_type_name()
 {
     std::string name;
-#ifdef _MSC_VER
+#if defined(_MSC_VER) && !defined(__clang__)
     name = typeid(PrivateMigraphTypeNameProbe).name();
     name = name.substr(7);
 #else

src/onnx/onnx_parser.cpp

@@ -244,7 +244,7 @@ void onnx_parser::parse_from(std::istream& is, std::string name)
     this->filename   = std::move(name);
     auto parent_path = fs::path(this->filename).parent_path();
     if(not parent_path.empty())
-        this->path = parent_path;
+        this->path = parent_path.string();
 
     onnx::ModelProto model;
     if(model.ParseFromIstream(&is))

src/process.cpp

@@ -26,13 +26,23 @@
 #include <migraphx/env.hpp>
 #include <functional>
 #include <iostream>
+#include <optional>
+
+#ifdef _WIN32
+// cppcheck-suppress definePrefix
+#define WIN32_LEAN_AND_MEAN
+#include <Windows.h>
+#else
 #include <unistd.h>
+#endif
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 
 MIGRAPHX_DECLARE_ENV_VAR(MIGRAPHX_TRACE_CMD_EXECUTE)
 
+#ifndef _WIN32
+
 std::function<void(const char*)> redirect_to(std::ostream& os)
 {
     return [&](const char* x) { os << x; };
@@ -74,6 +84,155 @@ int exec(const std::string& cmd, std::function<void(process::writer)> std_in)
     });
 }
 
+#else
+
+constexpr std::size_t MIGRAPHX_PROCESS_BUFSIZE = 4096;
+
+class pipe
+{
+    public:
+    explicit pipe(bool inherit_handle = true)
+    {
+        SECURITY_ATTRIBUTES attrs;
+        attrs.nLength              = sizeof(SECURITY_ATTRIBUTES);
+        attrs.bInheritHandle       = inherit_handle ? TRUE : FALSE;
+        attrs.lpSecurityDescriptor = nullptr;
+
+        if(CreatePipe(&m_read, &m_write, &attrs, 0) == FALSE)
+            throw GetLastError();
+
+        if(SetHandleInformation(&m_read, HANDLE_FLAG_INHERIT, 0) == FALSE)
+            throw GetLastError();
+    }
+
+    pipe(const pipe&)            = delete;
+    pipe& operator=(const pipe&) = delete;
+
+    pipe(pipe&&) = default;
+
+    ~pipe()
+    {
+        CloseHandle(m_read);
+        m_read = nullptr;
+        CloseHandle(m_write);
+        m_write = nullptr;
+    }
+
+    std::optional<std::pair<bool, DWORD>> read(LPVOID buffer, DWORD length) const
+    {
+        DWORD bytes_read;
+        if(ReadFile(m_read, buffer, length, &bytes_read, nullptr) == FALSE)
+        {
+            DWORD error{GetLastError()};
+            if(error != ERROR_MORE_DATA)
+            {
+                return std::nullopt;
+            }
+            return {{true, bytes_read}};
+        }
+        return {{false, bytes_read}};
+    }
+
+    HANDLE get_read_handle() const { return m_read; }
+
+    bool write(LPCVOID buffer, DWORD length) const
+    {
+        DWORD bytes_written;
+        return WriteFile(m_write, buffer, length, &bytes_written, nullptr) == TRUE;
+    }
+
+    HANDLE get_write_handle() const { return m_write; }
+
+    private:
+    HANDLE m_write = nullptr, m_read = nullptr;
+};
+
+template <typename F>
+int exec(const std::string& cmd, F f)
+{
+    try
+    {
+        if(enabled(MIGRAPHX_TRACE_CMD_EXECUTE{}))
+            std::cout << cmd << std::endl;
+
+        STARTUPINFO info;
+        PROCESS_INFORMATION process_info;
+
+        pipe in{}, out{};
+
+        ZeroMemory(&info, sizeof(STARTUPINFO));
+        info.cb         = sizeof(STARTUPINFO);
+        info.hStdError  = out.get_write_handle();
+        info.hStdOutput = out.get_write_handle();
+        info.hStdInput  = in.get_read_handle();
+        info.dwFlags |= STARTF_USESTDHANDLES;
+
+        ZeroMemory(&process_info, sizeof(process_info));
+
+        if(CreateProcess(nullptr,
+                         const_cast<LPSTR>(cmd.c_str()),
+                         nullptr,
+                         nullptr,
+                         TRUE,
+                         0,
+                         nullptr,
+                         nullptr,
+                         &info,
+                         &process_info) == FALSE)
+        {
+            return GetLastError();
+        }
+
+        f(in, out);
+
+        WaitForSingleObject(process_info.hProcess, INFINITE);
+
+        DWORD status{};
+        GetExitCodeProcess(process_info.hProcess, &status);
+
+        CloseHandle(process_info.hProcess);
+        CloseHandle(process_info.hThread);
+
+        return static_cast<int>(status);
+    }
+    // cppcheck-suppress catchExceptionByValue
+    catch(DWORD last_error)
+    {
+        return last_error;
+    }
+}
+
+int exec(const std::string& cmd)
+{
+    TCHAR buffer[MIGRAPHX_PROCESS_BUFSIZE];
+    HANDLE std_out{GetStdHandle(STD_OUTPUT_HANDLE)};
+    return (std_out == nullptr or std_out == INVALID_HANDLE_VALUE)
+               ? GetLastError()
+               : exec(cmd, [&](const pipe&, const pipe& out) {
+                     for(;;)
+                     {
+                         if(auto result = out.read(buffer, MIGRAPHX_PROCESS_BUFSIZE))
+                         {
+                             auto [more_data, bytes_read] = *result;
+                             if(not more_data or bytes_read == 0)
+                                 break;
+                             DWORD written;
+                             if(WriteFile(std_out, buffer, bytes_read, &written, nullptr) == FALSE)
+                                 break;
+                         }
+                     }
+                 });
+}
+
+int exec(const std::string& cmd, std::function<void(process::writer)> std_in)
+{
+    return exec(cmd, [&](const pipe& in, const pipe&) {
+        std_in([&](const char* buffer, std::size_t n) { in.write(buffer, n); });
+    });
+}
+
+#endif
+
 struct process_impl
 {
     std::string command{};
@@ -119,7 +278,14 @@ process& process::cwd(const fs::path& p)
     return *this;
 }
 
-void process::exec() { impl->check_exec(impl->get_command(), redirect_to(std::cout)); }
+void process::exec()
+{
+#ifndef _WIN32
+    impl->check_exec(impl->get_command(), redirect_to(std::cout));
+#else
+    impl->check_exec(impl->get_command());
+#endif
+}
 
 void process::write(std::function<void(process::writer)> pipe_in)
 {

src/py/CMakeLists.txt

@@ -22,27 +22,24 @@
 # THE SOFTWARE.
 #####################################################################################
 
-option(MIGRAPHX_ENABLE_PYTHON "Enable python bindings" ON)
 add_library(migraphx_py py_loader.cpp)
 migraphx_generate_export_header(migraphx_py)
 target_include_directories(migraphx_py PRIVATE include)
 target_link_libraries(migraphx_py PUBLIC migraphx)
 rocm_install_targets(TARGETS migraphx_py INCLUDE include)
-if(MIGRAPHX_ENABLE_PYTHON)
-    include(PythonModules)
 
+include(PythonModules)
 
-    foreach(PYTHON_VERSION ${PYTHON_VERSIONS})
-        py_add_module(migraphx_pybind_${PYTHON_VERSION} migraphx_py.cpp PYTHON_VERSION ${PYTHON_VERSION} PYTHON_MODULE migraphx)
-        target_link_libraries(migraphx_pybind_${PYTHON_VERSION} PRIVATE migraphx migraphx_tf migraphx_onnx migraphx_all_targets)
-        rocm_install_targets(TARGETS migraphx_pybind_${PYTHON_VERSION})
-        add_dependencies(migraphx_py migraphx_pybind_${PYTHON_VERSION})
-        
-        add_library(migraphx_py_${PYTHON_VERSION} py.cpp)
-        target_include_directories(migraphx_py_${PYTHON_VERSION} PRIVATE include)
-        target_link_libraries(migraphx_py_${PYTHON_VERSION} PUBLIC migraphx)
-        target_link_libraries(migraphx_py_${PYTHON_VERSION} PRIVATE pybind11::pybind11 python${PYTHON_VERSION}::runtime)
-        rocm_install_targets(TARGETS migraphx_py_${PYTHON_VERSION})
-        add_dependencies(migraphx_py migraphx_py_${PYTHON_VERSION})
-    endforeach()
-endif()
+foreach(PYTHON_VERSION ${PYTHON_VERSIONS})
+    py_add_module(migraphx_pybind_${PYTHON_VERSION} migraphx_py.cpp PYTHON_VERSION ${PYTHON_VERSION} PYTHON_MODULE migraphx)
+    target_link_libraries(migraphx_pybind_${PYTHON_VERSION} PRIVATE migraphx migraphx_tf migraphx_onnx migraphx_all_targets)
+    rocm_install_targets(TARGETS migraphx_pybind_${PYTHON_VERSION})
+    add_dependencies(migraphx_py migraphx_pybind_${PYTHON_VERSION})
+
+    add_library(migraphx_py_${PYTHON_VERSION} py.cpp)
+    target_include_directories(migraphx_py_${PYTHON_VERSION} PRIVATE include)
+    target_link_libraries(migraphx_py_${PYTHON_VERSION} PUBLIC migraphx)
+    target_link_libraries(migraphx_py_${PYTHON_VERSION} PRIVATE pybind11::pybind11 python${PYTHON_VERSION}::runtime)
+    rocm_install_targets(TARGETS migraphx_py_${PYTHON_VERSION})
+    add_dependencies(migraphx_py migraphx_py_${PYTHON_VERSION})
+endforeach()

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

@@ -91,6 +91,19 @@ struct post_op : reflect_equality<post_op>, reflect_stream<post_op>
     }
 };
 
+template <class F>
+struct execute_wrapper
+{
+    F f;
+    argument operator()(context&, const std::vector<argument>& args) const { return f(args); }
+};
+
+template <class F>
+execute_wrapper<F> make_execute_wrapper(F f)
+{
+    return {std::move(f)};
+}
+
 template <class Derived, class Primitive>
 struct dnnl_op : auto_register_op<Derived>
 {
@@ -308,7 +321,7 @@ struct dnnl_op : auto_register_op<Derived>
 #ifndef NDEBUG
         auto prim_attr = get_primitive_attr(md);
 #endif
-        execute = [=](context&, const std::vector<argument>& args) {
+        execute = make_execute_wrapper([=](const std::vector<argument>& args) {
 #ifndef NDEBUG
             // Check that the memory descriptors have not changed
             auto debug_args = args;
@@ -379,7 +392,7 @@ struct dnnl_op : auto_register_op<Derived>
                 m[arg_lookup[i]] = to_dnnl_memory(md.at(arg_lookup[i]), args[i]);
             prim.execute(get_dnnl_context().stream, m);
             return args.back();
-        };
+        });
     }
     std::vector<shape> trim_post_op_inputs(const std::vector<shape>& inputs) const
     {

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

@@ -24,6 +24,7 @@
 #ifndef MIGRAPHX_GUARD_AMDMIGRAPHX_CPU_POINTWISE_HPP
 #define MIGRAPHX_GUARD_AMDMIGRAPHX_CPU_POINTWISE_HPP
 
+#include <array>
 #include <migraphx/config.hpp>
 #include <migraphx/context.hpp>
 #include <migraphx/check_shapes.hpp>

src/targets/gpu/compile_ops.cpp

@@ -185,8 +185,7 @@ struct compile_plan
             results.begin(), results.end(), std::back_inserter(times), [&](const auto& cr) {
                 if(not cr.has_value())
                     return std::numeric_limits<double>::max();
-                return time_op(*ctx, cr->replace.code_object, to_shapes(cr->ins->inputs()), 20)
-                    .first;
+                return time_op(*ctx, cr->replace.code_object, to_shapes(cr->ins->inputs()), 20);
             });
         auto i = std::distance(times.begin(), std::min_element(times.begin(), times.end()));
         std::cout << "Fastest solution: " << config->solutions.at(i) << std::endl;

src/targets/gpu/driver/compile_op.cpp

@@ -38,10 +38,8 @@ struct compile_op : action<compile_op>
         context ctx;
         auto inputs = p.parse_shapes(v.at("inputs"));
         auto op     = gpu::compile_op(v.at("name").to<std::string>(), ctx, inputs, v);
-        auto [host_time, device_time] = time_op(ctx, op, inputs, p.get(v, "iterations", 100));
-        std::cout << op << ": " << host_time << "ms";
-        if(device_time > 0)
-            std::cout << ", " << device_time << "ms";
+        auto t      = time_op(ctx, op, inputs, p.get(v, "iterations", 100));
+        std::cout << op << ": " << t << "ms";
         std::cout << std::endl;
     }
 };

src/targets/gpu/driver/run_op.cpp

@@ -43,8 +43,8 @@ struct run_op : action<run_op>
         auto op = make_op(name);
         if(v.contains("fields"))
             op.from_value(v.at("fields"));
-        auto [host_time, device_time] = time_op(ctx, op, inputs, p.get(v, "iterations", 100));
-        std::cout << op << ": " << host_time << "ms" << std::endl;
+        auto t = time_op(ctx, op, inputs, p.get(v, "iterations", 100));
+        std::cout << op << ": " << t << "ms" << std::endl;
     }
 };
 

src/targets/gpu/hiprtc/main.cpp

@@ -27,6 +27,7 @@
 #include <migraphx/msgpack.hpp>
 #include <migraphx/file_buffer.hpp>
 #include <migraphx/ranges.hpp>
+#include <array>
 #include <iostream>
 #include <cstring>
 

src/targets/gpu/include/migraphx/gpu/context.hpp

@@ -299,23 +299,6 @@ struct context
 
     any_ptr get_queue() { return get_stream().get(); }
 
-    void enable_perf_measurement(bool b = true)
-    {
-        if(b)
-        {
-            start_event = create_event_for_timing();
-            stop_event  = create_event_for_timing();
-            get_stream().record(start_event.get());
-            get_stream().record(stop_event.get());
-        }
-        else
-        {
-            start_event = nullptr;
-            stop_event  = nullptr;
-        }
-        measure_perf = b;
-    }
-
     std::pair<hipEvent_t, hipEvent_t> get_perf_events() const
     {
         if(measure_perf)
@@ -323,12 +306,12 @@ struct context
         return std::make_pair(nullptr, nullptr);
     }
 
-    float get_elapsed_ms() const
+    static float get_elapsed_ms(hipEvent_t start, hipEvent_t stop)
     {
         float result = 0;
-        if(start_event != nullptr and stop_event != nullptr)
+        if(start != nullptr and stop != nullptr)
         {
-            auto status = hipEventElapsedTime(&result, start_event.get(), stop_event.get());
+            auto status = hipEventElapsedTime(&result, start, stop);
             if(status != hipSuccess)
                 MIGRAPHX_THROW("Failed hipEventElapsedTime: " + hip_error(status));
         }

src/targets/gpu/include/migraphx/gpu/time_op.hpp

@@ -32,7 +32,7 @@ namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 namespace gpu {
 
-MIGRAPHX_GPU_EXPORT std::pair<double, double>
+MIGRAPHX_GPU_EXPORT double
 time_op(context& ictx, operation op, const std::vector<shape>& inputs, int n = 100);
 
 } // namespace gpu

src/targets/gpu/time_op.cpp

@@ -41,8 +41,7 @@ std::vector<argument> generate_arguments(const std::vector<shape>& shapes, unsig
 }
 
 using milliseconds = std::chrono::duration<double, std::milli>;
-std::pair<double, double>
-time_op(context& ictx, operation op, const std::vector<shape>& inputs, int n)
+double time_op(context& ictx, operation op, const std::vector<shape>& inputs, int n)
 {
 
     // TODO: Use std::ref
@@ -51,21 +50,19 @@ time_op(context& ictx, operation op, const std::vector<shape>& inputs, int n)
     auto output           = op.compute_shape(inputs);
     op.finalize(ctx, output, inputs);
     auto args = generate_arguments(inputs);
-    auto run  = [&] {
-        op.compute(ctx, output, args);
-        ctx.finish();
-    };
-    gctx.enable_perf_measurement();
+    auto start = context::create_event_for_timing();
+    auto stop  = context::create_event_for_timing();
+    auto run   = [&] { op.compute(ctx, output, args); };
     run();
-    double host_time   = 0.0;
-    double device_time = 0.0;
+    gctx.get_stream().record(start.get());
     for(auto i : range(n))
     {
         (void)i;
-        host_time += time<milliseconds>(run);
-        device_time += gctx.get_elapsed_ms();
+        run();
     }
-    return std::make_pair(host_time / n, device_time / n);
+    gctx.get_stream().record(stop.get());
+    gctx.finish();
+    return context::get_elapsed_ms(start.get(), stop.get()) / n;
 }
 
 } // namespace gpu

test/include/test.hpp

@@ -339,6 +339,8 @@ inline std::ostream& operator<<(std::ostream& os, const color& c)
     static const bool use_color = isatty(STDOUT_FILENO) != 0;
     if(use_color)
         return os << "\033[" << static_cast<std::size_t>(c) << "m";
+#else
+    (void)c;
 #endif
     return os;
 }

test/verify/run_verify.cpp

@@ -44,8 +44,7 @@ MIGRAPHX_DECLARE_ENV_VAR(MIGRAPHX_DUMP_TEST)
 
 // An improved async, that doesn't block
 template <class Function>
-std::future<typename std::result_of<Function()>::type> detach_async(Function&& f,
-                                                                    bool parallel = true)
+std::future<std::invoke_result_t<Function>> detach_async(Function&& f, bool parallel = true)
 {
     if(parallel)
     {