Merge branch 'develop' into tests

CMakeLists.txt

@@ -22,7 +22,7 @@ find_package(ROCM REQUIRED)
 
 include(ROCMSetupVersion)
 
-rocm_setup_version(VERSION 0.2)
+rocm_setup_version(VERSION 0.4)
 
 option( BUILD_SHARED_LIBS "Build as a shared library" ON )
 
@@ -32,15 +32,15 @@ if(CMAKE_CXX_COMPILER_ID STREQUAL "GNU")
     endif()
 endif()
 
-if(CMAKE_CXX_COMPILER MATCHES ".*hcc")
+include(CheckCXXCompilerFlag)
+check_cxx_compiler_flag("--cuda-host-only -x hip" HAS_HIP)
+if(HAS_HIP)
     message(STATUS "Enable miopen backend")
     set(MIGRAPHX_ENABLE_GPU On CACHE BOOL "")
 else()
     set(MIGRAPHX_ENABLE_GPU Off CACHE BOOL "")
 endif()
 
-set(MIGRAPHX_ENABLE_TF Off CACHE BOOL "")
-
 add_compile_options(-std=c++14)
 
 list(APPEND CMAKE_MODULE_PATH ${CMAKE_CURRENT_SOURCE_DIR}/cmake)
@@ -83,6 +83,7 @@ rocm_enable_clang_tidy(
         -modernize-use-override
         -modernize-pass-by-value
         -modernize-use-default-member-init
+        -modernize-use-trailing-return-type
         -modernize-use-transparent-functors
         -performance-type-promotion-in-math-fn
         -readability-braces-around-statements
@@ -124,6 +125,8 @@ rocm_enable_cppcheck(
         functionConst:*program.*
         shadowFunction
         shadowVar
+        shadowVariable
+        unsafeClassDivZero
         definePrefix:*test/include/test.hpp
     FORCE
     INCONCLUSIVE
@@ -153,6 +156,10 @@ rocm_create_package(
     DEPENDS miopen-hip rocblas hip_hcc half
 )
 
+set(CMAKE_LIBRARY_OUTPUT_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}/lib)
+set(CMAKE_ARCHIVE_OUTPUT_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}/lib)
+set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}/bin)
+
 add_subdirectory(src)
 add_subdirectory(doc)
 add_subdirectory(test)

Dockerfile

@@ -20,6 +20,7 @@ RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --allow-
     clang-format-5.0 \
     clang-tidy-5.0 \
     cmake \
+    comgr \
     curl \
     doxygen \
     g++-7 \
@@ -32,14 +33,16 @@ RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --allow-
     libncurses5-dev \
     libnuma-dev \
     libpthread-stubs0-dev \
+    libssl-dev \
     python \
     python-dev \
     python-pip \
+    rocm-device-libs \
     rocm-opencl \
     rocm-opencl-dev \
-    rocminfo \
     software-properties-common \
-    wget && \
+    wget \
+    zlib1g-dev && \
     apt-get clean && \
     rm -rf /var/lib/apt/lists/*
 
@@ -50,7 +53,7 @@ RUN pip install cget
 RUN pip install https://github.com/pfultz2/rclone/archive/master.tar.gz
 
 # Install hcc
-RUN rclone -b roc-2.3.x -c fd93baed7dcc4fe8019b5fdc90213bfe7c298245 https://github.com/RadeonOpenCompute/hcc.git /hcc
+RUN rclone -b roc-2.6.x -c 0f4c96b7851af2663a7f3ac16ecfb76c7c78a5bf https://github.com/RadeonOpenCompute/hcc.git /hcc
 RUN cget -p $PREFIX install hcc,/hcc
 
 # Use hcc

Jenkinsfile

@@ -99,13 +99,14 @@ rocmtest tidy: rocmnode('rocmtest') { cmake_build ->
             | xargs -n 1 -P 1 -I{} -t sh -c \'clang-format-5.0 -style=file {} | diff - {}\'
         '''
     }
-}, clang: rocmnode('vega') { cmake_build ->
+}, clang_debug: rocmnode('vega') { cmake_build ->
     stage('Clang Debug') {
-        // TODO: Enanle integer
+        // TODO: Enable integer
         def sanitizers = "undefined"
         def debug_flags = "-g -fno-omit-frame-pointer -fsanitize=${sanitizers} -fno-sanitize-recover=${sanitizers}"
         cmake_build("hcc", "-DCMAKE_BUILD_TYPE=debug -DMIGRAPHX_ENABLE_PYTHON=Off -DCMAKE_CXX_FLAGS_DEBUG='${debug_flags}'")
     }
+}, clang_release: rocmnode('vega') { cmake_build ->
     stage('Clang Release') {
         cmake_build("hcc", "-DCMAKE_BUILD_TYPE=release")
     }

cppcheck.rules

@@ -89,6 +89,15 @@
         <summary>Use manage pointer for resource management</summary>
     </message>
 </rule>
+<rule>
+    <tokenlist>raw</tokenlist>
+    <pattern><![CDATA[] (__device__ |__host__ )+(\(|{)]]></pattern>
+    <message>
+        <id>LambdaAttribute</id>
+        <severity>warning</severity>
+        <summary>Attributes to lambdas must come after parameter list.</summary>
+    </message>
+</rule>
 <rule>
     <tokenlist>raw</tokenlist>
     <pattern><![CDATA[hipLaunchKernelGGL \( (?!\( \w+ < \w+ > \))]]></pattern>

dev-requirements.txt

 pfultz2/rocm-recipes
-danmar/cppcheck@8aa68ee297c2d9ebadf5bcfd00c66ea8d9291e35 -DHAVE_RULES=1
-ROCm-Developer-Tools/HIP@e21df3058728ad8e73708bc99b82e0bdd3509c97
+danmar/cppcheck@ef714225bb31e9a76ac2484796763572386955ae -DHAVE_RULES=1
+ROCm-Developer-Tools/HIP@2490e42baa7d90458f0632fd9fbead2d395f41b9
 python/cpython@v3.6.6 -X autotools -H sha256:92aa914572c695c0aeb01b0a214813f414da4b51a371234df514a74761f2bb36
 -f requirements.txt

requirements.txt

-google/protobuf -DCMAKE_POSITION_INDEPENDENT_CODE=On
-RadeonOpenCompute/rocm-cmake@42f6740 --build
-ROCmSoftwarePlatform/rocBLAS@30a992ae02fda568688bcd190edd5e277d6674d9
-ROCmSoftwarePlatform/MIOpen@1.7.0
+google/protobuf@v3.8.0 -DCMAKE_POSITION_INDEPENDENT_CODE=On -X subdir -Dprotobuf_BUILD_TESTS=Off
+RadeonOpenCompute/rocm-cmake@78e54761251712362817e0cd076cae1c5c4e92f8 --build
+ROCmSoftwarePlatform/rocBLAS@7197df74e5a1ba64ff967065872e5f86a3516637
+ROCmSoftwarePlatform/MIOpen@2.0.0
 blaze,https://bitbucket.org/blaze-lib/blaze/get/f0755dea0e03.tar.gz -X header -DHEADER_DIR=blaze
 half,https://github.com/pfultz2/half/archive/1.12.0.tar.gz -X header -H sha256:0a08660b68abb176ebc2a0cdf8de46e3182a7f46c66443bb80dbfaaec98cf969
 pybind/pybind11@v2.2.4 -DPYBIND11_TEST=Off --build

src/CMakeLists.txt

@@ -4,7 +4,7 @@ include(ROCMPackageConfigHelpers)
 
 add_library(migraphx 
     auto_contiguous.cpp
-    common_subexpression_elimination.cpp
+    eliminate_common_subexpression.cpp
     propagate_constant.cpp
     dead_code_elimination.cpp
     eliminate_allocation.cpp
@@ -12,8 +12,9 @@ add_library(migraphx
     eliminate_concat.cpp
     eliminate_identity.cpp
     eliminate_pad.cpp
-    fwd_conv_batchnorm_rewrite.cpp
+    rewrite_batchnorm.cpp
     rewrite_rnn.cpp
+    rewrite_pooling.cpp
     env.cpp
     generate.cpp
     instruction.cpp
@@ -27,6 +28,7 @@ add_library(migraphx
     opt/memory_coloring.cpp
     opt/memory_coloring_impl.cpp
 )
+rocm_set_soversion(migraphx ${PROJECT_VERSION})
 rocm_clang_tidy_check(migraphx)
 rocm_install_targets(
   TARGETS migraphx
@@ -40,6 +42,7 @@ target_include_directories(migraphx SYSTEM PUBLIC $<BUILD_INTERFACE:${HALF_INCLU
 
 set(PACKAGE_DEPENDS)
 
+add_subdirectory(driver)
 add_subdirectory(onnx)
 add_subdirectory(tf)
 

src/driver/CMakeLists.txt

+
+add_executable(driver main.cpp verify.cpp perf.cpp)
+rocm_clang_tidy_check(driver)
+target_link_libraries(driver migraphx_cpu migraphx_onnx migraphx_tf)
+if(MIGRAPHX_ENABLE_GPU)
+target_link_libraries(driver migraphx_gpu)
+target_compile_definitions(driver PRIVATE -DHAVE_GPU)
+endif()

src/driver/argument_parser.hpp

+#ifndef MIGRAPHX_GUARD_RTGLIB_ARGUMENT_PARSER_HPP
+#define MIGRAPHX_GUARD_RTGLIB_ARGUMENT_PARSER_HPP
+
+#include <algorithm>
+#include <functional>
+#include <iostream>
+#include <set>
+#include <string>
+#include <sstream>
+#include <type_traits>
+#include <unordered_map>
+#include <utility>
+#include <vector>
+
+#include <migraphx/config.hpp>
+#include <migraphx/requires.hpp>
+#include <migraphx/type_name.hpp>
+#include <migraphx/functional.hpp>
+#include <migraphx/stringutils.hpp>
+
+namespace migraphx {
+namespace driver {
+inline namespace MIGRAPHX_INLINE_NS {
+
+#ifdef MIGRAPHX_USE_CLANG_TIDY
+#define MIGRAPHX_DRIVER_STATIC
+#else
+#define MIGRAPHX_DRIVER_STATIC static
+#endif
+
+template <class T>
+using bare = std::remove_cv_t<std::remove_reference_t<T>>;
+
+namespace detail {
+
+template <class T>
+auto is_container(int, T&& x) -> decltype(x.insert(x.end(), *x.begin()), std::true_type{});
+
+template <class T>
+std::false_type is_container(float, T&&);
+
+} // namespace detail
+
+template <class T>
+struct is_container : decltype(detail::is_container(int(0), std::declval<T>()))
+{
+};
+
+template <class T>
+using is_multi_value =
+    std::integral_constant<bool, (is_container<T>{} and not std::is_convertible<T, std::string>{})>;
+
+template <class T>
+struct value_parser
+{
+    template <MIGRAPHX_REQUIRES(not std::is_enum<T>{} and not is_multi_value<T>{})>
+    static T apply(const std::string& x)
+    {
+        T result;
+        std::stringstream ss;
+        ss.str(x);
+        ss >> result;
+        if(ss.fail())
+            throw std::runtime_error("Failed to parse: " + x);
+        return result;
+    }
+
+    template <MIGRAPHX_REQUIRES(std::is_enum<T>{} and not is_multi_value<T>{})>
+    static T apply(const std::string& x)
+    {
+        std::ptrdiff_t i;
+        std::stringstream ss;
+        ss.str(x);
+        ss >> i;
+        if(ss.fail())
+            throw std::runtime_error("Failed to parse: " + x);
+        return static_cast<T>(i);
+    }
+
+    template <MIGRAPHX_REQUIRES(is_multi_value<T>{} and not std::is_enum<T>{})>
+    static T apply(const std::string& x)
+    {
+        T result;
+        using value_type = typename T::value_type;
+        result.insert(result.end(), value_parser<value_type>::apply(x));
+        return result;
+    }
+};
+
+struct argument_parser
+{
+    struct argument
+    {
+        std::vector<std::string> flags;
+        std::function<bool(argument_parser&, const std::vector<std::string>&)> action{};
+        std::string type          = "";
+        std::string help          = "";
+        std::string metavar       = "";
+        std::string default_value = "";
+        unsigned nargs            = 1;
+    };
+
+    template <class T, MIGRAPHX_REQUIRES(is_multi_value<T>{})>
+    std::string as_string_value(const T& x)
+    {
+        return to_string_range(x);
+    }
+
+    template <class T, MIGRAPHX_REQUIRES(not is_multi_value<T>{})>
+    std::string as_string_value(const T& x)
+    {
+        return to_string(x);
+    }
+
+    template <class T, class... Fs>
+    void operator()(T& x, const std::vector<std::string>& flags, Fs... fs)
+    {
+        arguments.push_back({flags, [&](auto&&, const std::vector<std::string>& params) {
+                                 if(params.empty())
+                                     throw std::runtime_error("Flag with no value.");
+                                 x = value_parser<T>::apply(params.back());
+                                 return false;
+                             }});
+
+        argument& arg     = arguments.back();
+        arg.type          = migraphx::get_type_name<T>();
+        arg.default_value = as_string_value(x);
+        migraphx::each_args([&](auto f) { f(x, arg); }, fs...);
+    }
+
+    template <class... Fs>
+    void operator()(std::nullptr_t x, std::vector<std::string> flags, Fs... fs)
+    {
+        arguments.push_back({std::move(flags)});
+
+        argument& arg = arguments.back();
+        arg.type      = "";
+        arg.nargs     = 0;
+        migraphx::each_args([&](auto f) { f(x, arg); }, fs...);
+    }
+
+    MIGRAPHX_DRIVER_STATIC auto nargs(unsigned n = 1)
+    {
+        return [=](auto&&, auto& arg) { arg.nargs = n; };
+    }
+
+    template <class F>
+    MIGRAPHX_DRIVER_STATIC auto write_action(F f)
+    {
+        return [=](auto& x, auto& arg) {
+            arg.action = [&, f](auto& self, const std::vector<std::string>& params) {
+                f(self, x, params);
+                return false;
+            };
+        };
+    }
+
+    template <class F>
+    MIGRAPHX_DRIVER_STATIC auto do_action(F f)
+    {
+        return [=](auto&, auto& arg) {
+            arg.nargs  = 0;
+            arg.action = [&, f](auto& self, const std::vector<std::string>&) {
+                f(self);
+                return true;
+            };
+        };
+    }
+
+    MIGRAPHX_DRIVER_STATIC auto append()
+    {
+        return write_action([](auto&, auto& x, auto& params) {
+            using type = typename bare<decltype(params)>::value_type;
+            std::transform(params.begin(),
+                           params.end(),
+                           std::inserter(x, x.end()),
+                           [](std::string y) { return value_parser<type>::apply(y); });
+        });
+    }
+
+    MIGRAPHX_DRIVER_STATIC auto show_help(const std::string& msg = "")
+    {
+        return do_action([=](auto& self) {
+            for(auto&& arg : self.arguments)
+            {
+                std::cout << std::endl;
+                std::string prefix = "    ";
+                if(arg.flags.empty())
+                {
+                    std::cout << prefix;
+                    std::cout << arg.metavar;
+                }
+                for(const std::string& a : arg.flags)
+                {
+                    std::cout << prefix;
+                    std::cout << a;
+                    prefix = ", ";
+                }
+                if(not arg.type.empty())
+                {
+                    std::cout << " [" << arg.type << "]";
+                    if(not arg.default_value.empty())
+                        std::cout << " (Default: " << arg.default_value << ")";
+                }
+                std::cout << std::endl;
+                std::cout << "        " << arg.help << std::endl;
+            }
+            std::cout << std::endl;
+            if(not msg.empty())
+                std::cout << msg << std::endl;
+        });
+    }
+
+    MIGRAPHX_DRIVER_STATIC auto help(const std::string& help)
+    {
+        return [=](auto&, auto& arg) { arg.help = help; };
+    }
+
+    MIGRAPHX_DRIVER_STATIC auto metavar(const std::string& metavar)
+    {
+        return [=](auto&, auto& arg) { arg.metavar = metavar; };
+    }
+
+    template <class T>
+    MIGRAPHX_DRIVER_STATIC auto set_value(T value)
+    {
+        return [=](auto& x, auto& arg) {
+            arg.nargs  = 0;
+            arg.type   = "";
+            arg.action = [&, value](auto&, const std::vector<std::string>&) {
+                x = value;
+                return false;
+            };
+        };
+    }
+
+    bool parse(std::vector<std::string> args)
+    {
+        std::unordered_map<std::string, unsigned> keywords;
+        for(auto&& arg : arguments)
+        {
+            for(auto&& flag : arg.flags)
+                keywords[flag] = arg.nargs + 1;
+        }
+        auto arg_map =
+            generic_parse(std::move(args), [&](const std::string& x) { return keywords[x]; });
+        for(auto&& arg : arguments)
+        {
+            auto flags = arg.flags;
+            if(flags.empty())
+                flags = {""};
+            for(auto&& flag : flags)
+            {
+                if(arg_map.count(flag) > 0)
+                {
+                    if(arg.action(*this, arg_map[flag]))
+                        return true;
+                }
+            }
+        }
+        return false;
+    }
+
+    using string_map = std::unordered_map<std::string, std::vector<std::string>>;
+    template <class IsKeyword>
+    static string_map generic_parse(std::vector<std::string> as, IsKeyword is_keyword)
+    {
+        string_map result;
+
+        std::string flag;
+        bool clear = false;
+        for(auto&& x : as)
+        {
+            auto k = is_keyword(x);
+            if(k > 0)
+            {
+                flag = x;
+                result[flag]; // Ensure the flag exists
+                if(k == 1)
+                    flag = "";
+                else if(k == 2)
+                    clear = true;
+                else
+                    clear = false;
+            }
+            else
+            {
+                result[flag].push_back(x);
+                if(clear)
+                    flag = "";
+                clear = false;
+            }
+        }
+        return result;
+    }
+
+    private:
+    std::vector<argument> arguments;
+};
+
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace driver
+} // namespace migraphx
+
+#endif

src/driver/command.hpp

+#ifndef MIGRAPHX_GUARD_RTGLIB_COMMAND_HPP
+#define MIGRAPHX_GUARD_RTGLIB_COMMAND_HPP
+
+#include "argument_parser.hpp"
+
+#include <migraphx/config.hpp>
+#include <migraphx/type_name.hpp>
+#include <migraphx/stringutils.hpp>
+
+#include <unordered_map>
+#include <utility>
+#include <vector>
+
+namespace migraphx {
+namespace driver {
+inline namespace MIGRAPHX_INLINE_NS {
+
+inline auto& get_commands()
+{
+    static std::unordered_map<std::string, std::function<void(std::vector<std::string> args)>> m;
+    return m;
+}
+
+template <class T>
+std::string compute_command_name()
+{
+    static const std::string& tname = get_type_name<T>();
+    auto name                       = tname.substr(tname.rfind("::") + 2);
+    if(ends_with(name, "_command"))
+        name = name.substr(0, name.size() - 8);
+    if(ends_with(name, "_cmd"))
+        name = name.substr(0, name.size() - 4);
+    return name;
+}
+
+template <class T>
+const std::string& command_name()
+{
+    static const std::string& name = compute_command_name<T>();
+    return name;
+}
+
+template <class T>
+void run_command(std::vector<std::string> args, bool add_help = false)
+{
+    T x;
+    argument_parser ap;
+    if(add_help)
+        ap(nullptr, {"-h", "--help"}, ap.help("Show help"), ap.show_help());
+    x.parse(ap);
+    if(ap.parse(std::move(args)))
+        return;
+    x.run();
+}
+
+template <class T>
+int auto_register_command()
+{
+    auto& m              = get_commands();
+    m[command_name<T>()] = [](std::vector<std::string> args) { run_command<T>(args, true); };
+    return 0;
+}
+
+template <class T>
+struct command
+{
+    static int static_register;
+    // This typedef ensures that the static member will be instantiated if
+    // the class itself is instantiated
+    using static_register_type =
+        std::integral_constant<decltype(&static_register), &static_register>;
+};
+
+#ifdef __clang__
+#pragma clang diagnostic push
+#pragma clang diagnostic ignored "-Wglobal-constructors"
+#endif
+
+template <class T>
+int command<T>::static_register = auto_register_command<T>(); // NOLINT
+
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace driver
+} // namespace migraphx
+
+#endif

src/driver/main.cpp

+#include "argument_parser.hpp"
+#include "command.hpp"
+#include "verify.hpp"
+#include "perf.hpp"
+
+#include <migraphx/tf.hpp>
+#include <migraphx/onnx.hpp>
+#include <migraphx/stringutils.hpp>
+
+#include <migraphx/quantization.hpp>
+#include <migraphx/pass_manager.hpp>
+#include <migraphx/generate.hpp>
+#include <migraphx/dead_code_elimination.hpp>
+#include <migraphx/eliminate_identity.hpp>
+#include <migraphx/eliminate_pad.hpp>
+#include <migraphx/propagate_constant.hpp>
+#include <migraphx/simplify_algebra.hpp>
+#include <migraphx/simplify_reshapes.hpp>
+
+namespace migraphx {
+namespace driver {
+inline namespace MIGRAPHX_INLINE_NS {
+
+struct loader
+{
+    std::string file;
+    std::string file_type;
+    bool is_nhwc  = true;
+    unsigned trim = 0;
+    bool optimize = false;
+
+    void parse(argument_parser& ap)
+    {
+        ap(file, {}, ap.metavar("<input file>"));
+        ap(file_type, {"--onnx"}, ap.help("Load as onnx"), ap.set_value("onnx"));
+        ap(file_type, {"--tf"}, ap.help("Load as tensorflow"), ap.set_value("tf"));
+        ap(is_nhwc, {"--nhwc"}, ap.help("Treat tensorflow format as nhwc"), ap.set_value(true));
+        ap(is_nhwc, {"--nchw"}, ap.help("Treat tensorflow format as nchw"), ap.set_value(false));
+        ap(trim, {"--trim", "-t"}, ap.help("Trim instructions from the end"));
+        ap(optimize, {"--optimize"}, ap.help("Optimize when reading"), ap.set_value(true));
+    }
+
+    program load()
+    {
+        program p;
+        if(file_type.empty())
+        {
+            if(ends_with(file, ".onnx"))
+                file_type = "onnx";
+            else if(ends_with(file, ".pb"))
+                file_type = "tf";
+        }
+        std::cout << "Reading: " << file << std::endl;
+        if(file_type == "onnx")
+            p = parse_onnx(file);
+        else if(file_type == "tf")
+            p = parse_tf(file, is_nhwc);
+        if(trim > 0)
+        {
+            auto last = std::prev(p.end(), trim);
+            p.remove_instructions(last, p.end());
+        }
+        if(optimize)
+            migraphx::run_passes(p,
+                                 {
+                                     migraphx::eliminate_identity{},
+                                     migraphx::dead_code_elimination{},
+                                     migraphx::simplify_algebra{},
+                                     migraphx::dead_code_elimination{},
+                                     migraphx::simplify_reshapes{},
+                                     migraphx::dead_code_elimination{},
+                                     migraphx::propagate_constant{},
+                                     migraphx::dead_code_elimination{},
+                                     migraphx::eliminate_pad{},
+                                     migraphx::dead_code_elimination{},
+                                 });
+        return p;
+    }
+};
+
+struct compiler
+{
+    static const int q_fp16 = 1;
+    static const int q_int8 = 2;
+    loader l;
+    bool gpu     = true;
+    int quantize = 0;
+
+    std::vector<std::string> fill1;
+    void parse(argument_parser& ap)
+    {
+        l.parse(ap);
+        ap(gpu, {"--gpu"}, ap.help("Compile on the gpu"), ap.set_value(true));
+        ap(gpu, {"--cpu"}, ap.help("Compile on the cpu"), ap.set_value(false));
+        ap(quantize, {"--fp16"}, ap.help("Quantize for fp16"), ap.set_value(q_fp16));
+        ap(quantize, {"--int8"}, ap.help("Quantize for int8"), ap.set_value(q_int8));
+        ap(fill1, {"--fill1"}, ap.help("Fill parameter with 1s"), ap.append());
+    }
+
+    auto params(const program& p, bool use_gpu = true)
+    {
+        program::parameter_map m;
+        for(auto&& s : fill1)
+            m[s] = fill_argument(p.get_parameter_shape(s), 1);
+        fill_param_map(m, p, use_gpu && gpu);
+        return m;
+    }
+
+    program compile()
+    {
+        auto p = l.load();
+        auto t = get_target(gpu);
+        if(quantize == q_fp16)
+        {
+            quantize_fp16(p);
+        }
+        else if(quantize == q_int8)
+        {
+            quantize_int8(p, t, {params(p, false)});
+        }
+        p.compile(t);
+        return p;
+    }
+};
+
+struct read : command<read>
+{
+    loader l;
+    void parse(argument_parser& ap) { l.parse(ap); }
+
+    void run()
+    {
+        auto p = l.load();
+        std::cout << p << std::endl;
+    }
+};
+
+struct params : command<params>
+{
+    loader l;
+    void parse(argument_parser& ap) { l.parse(ap); }
+
+    void run()
+    {
+        auto p = l.load();
+        for(auto&& param : p.get_parameter_shapes())
+            std::cout << param.first << ": " << param.second << std::endl;
+    }
+};
+
+struct verify : command<verify>
+{
+    loader l;
+    double tolerance     = 80;
+    bool per_instruction = false;
+    bool reduce          = false;
+    void parse(argument_parser& ap)
+    {
+        l.parse(ap);
+        ap(tolerance, {"--tolerance"}, ap.help("Tolerance for errors"));
+        ap(per_instruction,
+           {"-i", "--per-instruction"},
+           ap.help("Verify each instruction"),
+           ap.set_value(true));
+        ap(reduce, {"-r", "--reduce"}, ap.help("Reduce program and verify"), ap.set_value(true));
+    }
+
+    void run()
+    {
+        auto p = l.load();
+        std::cout << p << std::endl;
+
+        if(per_instruction)
+        {
+            verify_instructions(p, tolerance);
+        }
+        else if(reduce)
+        {
+            verify_reduced_program(p, tolerance);
+        }
+        else
+        {
+            verify_program(l.file, p, tolerance);
+        }
+    }
+};
+
+struct compile : command<compile>
+{
+    compiler c;
+    void parse(argument_parser& ap) { c.parse(ap); }
+
+    void run()
+    {
+        std::cout << "Compiling ... " << std::endl;
+        auto p = c.compile();
+        std::cout << p << std::endl;
+    }
+};
+
+struct run_cmd : command<run_cmd>
+{
+    compiler c;
+    void parse(argument_parser& ap) { c.parse(ap); }
+
+    void run()
+    {
+        std::cout << "Compiling ... " << std::endl;
+        auto p = c.compile();
+        std::cout << "Allocating params ... " << std::endl;
+        auto m = c.params(p);
+        p.eval(m);
+        std::cout << p << std::endl;
+    }
+};
+
+struct perf : command<perf>
+{
+    compiler c;
+    unsigned n = 100;
+    void parse(argument_parser& ap)
+    {
+        c.parse(ap);
+        ap(n, {"--iterations", "-n"}, ap.help("Number of iterations to run for perf report"));
+    }
+
+    void run()
+    {
+        std::cout << "Compiling ... " << std::endl;
+        auto p = c.compile();
+        std::cout << "Allocating params ... " << std::endl;
+        auto m = c.params(p);
+        std::cout << "Running performance report ... " << std::endl;
+        p.perf_report(std::cout, n, m);
+    }
+};
+
+struct main_command
+{
+    static std::string get_command_help()
+    {
+        std::string result = "Commands:\n";
+        return std::accumulate(get_commands().begin(),
+                               get_commands().end(),
+                               result,
+                               [](auto r, auto&& p) { return r + "    " + p.first + "\n"; });
+    }
+    void parse(argument_parser& ap)
+    {
+        ap(nullptr, {"-h", "--help"}, ap.help("Show help"), ap.show_help(get_command_help()));
+    }
+
+    void run() {}
+};
+
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace driver
+} // namespace migraphx
+
+using namespace migraphx::driver; // NOLINT
+int main(int argc, const char* argv[])
+{
+    std::vector<std::string> args(argv + 1, argv + argc);
+    if(args.empty())
+        return 0;
+    auto&& m = get_commands();
+    auto cmd = args.front();
+    if(m.count(cmd) > 0)
+    {
+        m.at(cmd)({args.begin() + 1, args.end()});
+    }
+    else
+    {
+        run_command<main_command>(args);
+    }
+    return 0;
+}

src/driver/perf.cpp

+#include "perf.hpp"
+
+#include <migraphx/cpu/target.hpp>
+#include <migraphx/generate.hpp>
+#ifdef HAVE_GPU
+#include <migraphx/gpu/target.hpp>
+#include <migraphx/gpu/hip.hpp>
+#endif
+
+namespace migraphx {
+namespace driver {
+inline namespace MIGRAPHX_INLINE_NS {
+
+program::parameter_map fill_param_map(program::parameter_map& m, const program& p, bool gpu)
+{
+    for(auto&& x : p.get_parameter_shapes())
+    {
+        argument& arg = m[x.first];
+        if(arg.empty())
+            arg = generate_argument(x.second);
+#ifdef HAVE_GPU
+        if(gpu)
+            arg = gpu::to_gpu(arg);
+#else
+        (void)gpu;
+#endif
+    }
+    return m;
+}
+
+program::parameter_map create_param_map(const program& p, bool gpu)
+{
+    program::parameter_map m;
+    for(auto&& x : p.get_parameter_shapes())
+    {
+#ifdef HAVE_GPU
+        if(gpu)
+            m[x.first] = gpu::to_gpu(generate_argument(x.second));
+        else
+#else
+        (void)gpu;
+#endif
+            m[x.first] = generate_argument(x.second);
+    }
+    return m;
+}
+
+target get_target(bool gpu)
+{
+    if(gpu)
+    {
+#ifdef HAVE_GPU
+        return gpu::target{};
+#else
+        MIGRAPHX_THROW("Gpu not supported.");
+#endif
+    }
+    else
+    {
+        return cpu::target{};
+    }
+}
+
+void compile_program(program& p, bool gpu)
+{
+    if(gpu)
+    {
+#ifdef HAVE_GPU
+        p.compile(gpu::target{});
+#else
+        MIGRAPHX_THROW("Gpu not supported.");
+#endif
+    }
+    else
+    {
+        p.compile(cpu::target{});
+    }
+}
+
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace driver
+} // namespace migraphx

src/driver/perf.hpp

+#ifndef MIGRAPHX_GUARD_RTGLIB_PERF_HPP
+#define MIGRAPHX_GUARD_RTGLIB_PERF_HPP
+
+#include <migraphx/program.hpp>
+
+namespace migraphx {
+namespace driver {
+inline namespace MIGRAPHX_INLINE_NS {
+
+program::parameter_map fill_param_map(program::parameter_map& m, const program& p, bool gpu);
+program::parameter_map create_param_map(const program& p, bool gpu = true);
+target get_target(bool gpu);
+void compile_program(program& p, bool gpu = true);
+
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace driver
+} // namespace migraphx
+
+#endif

src/driver/verify.cpp

+#include "verify.hpp"
+
+#include <migraphx/cpu/target.hpp>
+#include <migraphx/generate.hpp>
+#include <migraphx/verify_args.hpp>
+#include <migraphx/instruction.hpp>
+
+#ifdef HAVE_GPU
+#include <migraphx/gpu/target.hpp>
+#include <migraphx/gpu/hip.hpp>
+#endif
+
+namespace migraphx {
+namespace driver {
+inline namespace MIGRAPHX_INLINE_NS {
+
+template <class T>
+auto get_hash(const T& x)
+{
+    return std::hash<T>{}(x);
+}
+
+argument run_cpu(program p)
+{
+    p.compile(cpu::target{});
+    program::parameter_map m;
+    for(auto&& x : p.get_parameter_shapes())
+    {
+        m[x.first] = generate_argument(x.second, get_hash(x.first));
+    }
+    auto out = p.eval(m);
+    std::cout << p << std::endl;
+    return out;
+}
+
+argument run_gpu(program p)
+{
+#ifdef HAVE_GPU
+    p.compile(gpu::target{});
+
+    program::parameter_map m;
+    for(auto&& x : p.get_parameter_shapes())
+    {
+        m[x.first] = gpu::to_gpu(generate_argument(x.second, get_hash(x.first)));
+    }
+    auto out = gpu::from_gpu(p.eval(m));
+    std::cout << p << std::endl;
+    return gpu::from_gpu(out);
+#else
+    (void)p;
+    MIGRAPHX_THROW("Gpu unsupported!");
+#endif
+}
+
+void verify_program(const std::string& name, const program& p, double tolerance)
+{
+    auto x = run_cpu(p);
+    auto y = run_gpu(p);
+    verify_args(name, x, y, tolerance);
+    // std::cout << "cpu: " << x << std::endl;
+    // std::cout << "gpu: " << y << std::endl;
+}
+
+void verify_instructions(const program& prog, double tolerance)
+{
+    for(auto&& ins : prog)
+    {
+        if(ins.name().front() == '@')
+            continue;
+        if(ins.name() == "broadcast")
+            continue;
+        if(ins.name() == "transpose")
+            continue;
+        if(ins.name() == "reshape")
+            continue;
+        program p;
+        std::vector<instruction_ref> inputs;
+        for(auto&& arg : ins.inputs())
+        {
+            if(arg->name() == "@literal")
+                inputs.push_back(p.add_literal(arg->get_literal()));
+            else
+                inputs.push_back(p.add_parameter(std::to_string(inputs.size()), arg->get_shape()));
+        }
+        p.add_instruction(ins.get_operator(), inputs);
+        try
+        {
+            std::cout << "Verify: " << ins.name() << std::endl;
+            std::cout << p << std::endl;
+            verify_program(ins.name(), p, tolerance);
+        }
+        catch(...)
+        {
+            std::cout << "Instruction " << ins.name() << " threw an exception." << std::endl;
+            throw;
+        }
+    }
+}
+
+void verify_reduced(program p, int n, double tolerance)
+{
+    auto last = std::prev(p.end(), n + 1);
+    p.remove_instructions(last, p.end());
+    std::cout << "Verify: " << std::endl;
+    std::cout << p << std::endl;
+    verify_program(std::to_string(n), p, tolerance);
+}
+
+void verify_reduced_program(const program& p, double tolerance)
+{
+    auto n = std::distance(p.begin(), p.end());
+    for(std::size_t i = 0; i < n; i++)
+    {
+        verify_reduced(p, i, tolerance);
+    }
+}
+
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace driver
+} // namespace migraphx

src/driver/verify.hpp

+#ifndef MIGRAPHX_GUARD_RTGLIB_DRIVER_VERIFY_HPP
+#define MIGRAPHX_GUARD_RTGLIB_DRIVER_VERIFY_HPP
+
+#include <migraphx/program.hpp>
+
+namespace migraphx {
+namespace driver {
+inline namespace MIGRAPHX_INLINE_NS {
+
+argument run_cpu(program p);
+argument run_gpu(program p);
+void verify_program(const std::string& name, const program& p, double tolerance = 100);
+void verify_instructions(const program& prog, double tolerance = 80);
+void verify_reduced_program(const program& p, double tolerance = 80);
+
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace driver
+} // namespace migraphx
+
+#endif

src/common_subexpression_elimination.cpp → src/eliminate_common_subexpression.cpp

-#include <migraphx/common_subexpression_elimination.hpp>
+#include <migraphx/eliminate_common_subexpression.hpp>
 #include <migraphx/program.hpp>
 #include <migraphx/instruction.hpp>
 #include <migraphx/iterator_for.hpp>
@@ -27,13 +27,17 @@ void cse_range(program& p, Range&& r)
             if(*eq != *ins)
                 continue;
             p.replace_instruction(ins, eq);
-            cse_range(p, eq->outputs());
+            auto outputs = eq->outputs();
+            std::sort(outputs.begin(), outputs.end(), [&](auto x, auto y) {
+                return std::distance(eq, x) < std::distance(eq, y);
+            });
+            cse_range(p, outputs);
         }
         instructions.emplace(ins->name(), ins);
     }
 }
 
-void common_subexpression_elimination::apply(program& p) const { cse_range(p, iterator_for(p)); }
+void eliminate_common_subexpression::apply(program& p) const { cse_range(p, iterator_for(p)); }
 
 } // namespace MIGRAPHX_INLINE_NS
 } // namespace migraphx

src/eliminate_contiguous.cpp

@@ -4,6 +4,8 @@
 #include <migraphx/iterator_for.hpp>
 #include <migraphx/ranges.hpp>
 #include <migraphx/stringutils.hpp>
+#include <migraphx/op/contiguous.hpp>
+#include <migraphx/op/identity.hpp>
 #include <utility>
 
 namespace migraphx {
@@ -82,6 +84,14 @@ void eliminate_contiguous::apply(program& p) const
                 {
                     instruction::replace_argument(ins, arg, prev);
                 }
+                else if(prev->can_eval())
+                {
+                    auto c = op::contiguous{};
+                    auto r = c.compute(c.compute_shape({prev->get_shape()}), {prev->eval()});
+
+                    auto l = p.add_literal(r.get_shape(), r.data());
+                    p.replace_instruction(arg, l);
+                }
             }
         }
     }

src/eliminate_pad.cpp

@@ -43,9 +43,7 @@ void eliminate_pad::update_op(T,
     std::vector<int64_t> pads = pad_op.pads;
     std::array<size_t, 2> new_pads{static_cast<size_t>(pads[2]), static_cast<size_t>(pads[3])};
 
-    T op = any_cast<T>(ins->get_operator());
-    if(op.padding_mode != op::padding_mode_t::default_)
-        return;
+    T op       = any_cast<T>(ins->get_operator());
     op.padding = new_pads;
 
     std::vector<instruction_ref> new_inputs{ins->inputs()};

src/generate.cpp

@@ -3,6 +3,17 @@
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 
+argument fill_argument(shape s, unsigned long value)
+{
+    argument result;
+    s.visit_type([&](auto as) {
+        using type = typename decltype(as)::type;
+        auto v     = fill_tensor_data<type>(s, value);
+        result     = {s, [v]() mutable { return reinterpret_cast<char*>(v.data()); }};
+    });
+    return result;
+}
+
 argument generate_argument(shape s, unsigned long seed)
 {
     argument result;

src/include/migraphx/argument.hpp

@@ -36,7 +36,7 @@ struct argument : raw_data<argument>
     }
 
     /// Provides a raw pointer to the data
-    std::function<char*()> data;
+    std::function<char*()> data = nullptr;
 
     /// Whether data is available
     bool empty() const { return not data; }