Merge branch 'develop' into mem_color_ordering_fix

.clang-tidy

 CheckOptions: 
   - key:             bugprone-unused-return-value.CheckedFunctions
     value:           '::std::async;::std::launder;::std::remove;::std::remove_if;::std::unique;::std::unique_ptr::release;::std::basic_string::empty;::std::vector::empty;::std::find;::std::find_if;::std::find_if_not;::std::all_of;::std::any_of;::std::none_of;::std::count;::std::count_if;::std::mismatch;::std::find_end;::std::find_first_of;::std::adjacent_find;::std::search;::std::search_n;::std::nth_element;::std::lower_bound;::std::upper_bound;::std::binary_search;::std::equal_range;::std::max;::std::max_element;::std::min;::std::min_element;::std::minmax;::std::minmax_element;::std::equal;::std::lexicographical_compare;::std::accumulate;::std::inner_product'
+  - key:             cppcoreguidelines-macro-usage.AllowedRegexp
+    value:           'DEBUG|FALLTHROUGH|_THROW|_REQUIRES|_DECLARE_|_VISIT_|_GENERATE_|_DETAIL_|_MANAGE_PTR|_MATCHER|DEVICE_SHARED'
   - key:             modernize-loop-convert.MinConfidence
     value:           risky   
   - key:             modernize-loop-convert.NamingStyle
     value:           lower_case
+  - key:             performance-unnecessary-copy-initialization.AllowedTypes
+    value:           'shape'
+  - key:             performance-unnecessary-value-param.AllowedTypes
+    value:           'shape'
   - key:             readability-function-size.BranchThreshold
     value:           '15'
   - key:             readability-function-size.LineThreshold
@@ -104,4 +110,4 @@ CheckOptions:
   # - key:             readability-identifier-naming.MacroDefinitionCase
   #   value:           UPPER_CASE
   # - key:             readability-identifier-naming.MacroDefinitionPrefix
-  #   value:           MIGRAPH_
+  #   value:           MIGRAPHX_

CMakeLists.txt

@@ -34,9 +34,9 @@ endif()
 
 if(CMAKE_CXX_COMPILER MATCHES ".*hcc")
     message(STATUS "Enable miopen backend")
-    set(MIGRAPH_ENABLE_GPU On CACHE BOOL "")
+    set(MIGRAPHX_ENABLE_GPU On CACHE BOOL "")
 else()
-    set(MIGRAPH_ENABLE_GPU Off CACHE BOOL "")
+    set(MIGRAPHX_ENABLE_GPU Off CACHE BOOL "")
 endif()
 
 add_compile_options(-std=c++14)
@@ -52,6 +52,7 @@ rocm_enable_clang_tidy(
         -clang-analyzer-optin.performance.Padding
         -clang-diagnostic-deprecated-declarations
         -clang-diagnostic-extern-c-compat
+        -clang-diagnostic-disabled-macro-expansion
         -clang-diagnostic-unused-command-line-argument
         -cppcoreguidelines-pro-bounds-array-to-pointer-decay
         -cppcoreguidelines-pro-bounds-constant-array-index
@@ -70,13 +71,12 @@ rocm_enable_clang_tidy(
         -hicpp-explicit-conversions
         -hicpp-no-array-decay
         -hicpp-special-member-functions
+        -hicpp-uppercase-literal-suffix
         -hicpp-use-override
         # This check is broken
-        -hicpp-use-auto
         -llvm-header-guard
         -llvm-include-order
         -misc-macro-parentheses
-        -modernize-use-auto
         -modernize-use-override
         -modernize-pass-by-value
         -modernize-use-default-member-init
@@ -84,7 +84,12 @@ rocm_enable_clang_tidy(
         -readability-braces-around-statements
         -readability-else-after-return
         -readability-named-parameter
+        -readability-uppercase-literal-suffix,
+        -*-avoid-c-arrays
         -*-explicit-constructor
+        -*-magic-numbers
+        -*-non-private-member-variables-in-classes
+        -*-use-auto
         -*-use-emplace
         -*-use-equals-default
     ERRORS
@@ -93,7 +98,8 @@ rocm_enable_clang_tidy(
     HEADER_FILTER
         ".*hpp"
     EXTRA_ARGS
-        -DMIGRAPH_USE_CLANG_TIDY
+        -DMIGRAPHX_USE_CLANG_TIDY
+        "-Dmain\\\\(...\\\\)=main\\\\(__VA_ARGS__\\\\) // NOLINT"
 
 )
 include(ROCMCppCheck)
@@ -111,7 +117,9 @@ rocm_enable_cppcheck(
         passedByValue
         unusedStructMember
         functionStatic
-        functionConst
+        functionConst:*program.*
+        shadowFunction
+        shadowVar
         definePrefix:*test/include/test.hpp
     FORCE
     INCONCLUSIVE
@@ -124,6 +132,7 @@ rocm_enable_cppcheck(
         ${CMAKE_CURRENT_SOURCE_DIR}/src/include
         ${CMAKE_CURRENT_SOURCE_DIR}/src/targets/cpu/include
         ${CMAKE_CURRENT_SOURCE_DIR}/src/targets/miopen/include
+        ${CMAKE_CURRENT_SOURCE_DIR}/test/include
     DEFINE
         CPPCHECK=1
 )
@@ -132,11 +141,11 @@ enable_testing()
 
 include(ROCMCreatePackage)
 rocm_create_package(
-    NAME MIGraph
+    NAME MIGraphX
     DESCRIPTION "AMD's graph optimizer"
     MAINTAINER "Paul Fultz II <paul.fultz@amd.com>"
     LDCONFIG
-    DEPENDS miopen-hip rocblas hip_hcc
+    DEPENDS miopen-hip rocblas hip_hcc half
 )
 
 add_subdirectory(src)

Dockerfile

@@ -50,7 +50,7 @@ RUN pip install cget
 RUN pip install https://github.com/pfultz2/rclone/archive/master.tar.gz
 
 # Install hcc
-RUN rclone -b sanitizer1 https://github.com/RadeonOpenCompute/hcc.git /hcc
+RUN rclone -b roc-2.0.x -c 757fb492517b80e7c86338af5fc1a43d63cb25a9 https://github.com/RadeonOpenCompute/hcc.git /hcc
 RUN cget -p $PREFIX install hcc,/hcc
 
 # Use hcc

Jenkinsfile

@@ -8,10 +8,16 @@ def rocmtestnode(variant, name, body) {
             mkdir build
             cd build
             CXX=${compiler} CXXFLAGS='-Werror -Wno-fallback' cmake ${flags} .. 
-            CTEST_PARALLEL_LEVEL=32 make -j32 all doc check
+            CTEST_PARALLEL_LEVEL=32 make -j32 generate all doc package check
         """
         echo cmd
         sh cmd
+        if (compiler == "hcc") {
+            // Only archive from master or develop
+            if (env.BRANCH_NAME == "develop" || env.BRANCH_NAME == "master") {
+                archiveArtifacts artifacts: "build/*.deb", allowEmptyArchive: true, fingerprint: true
+            }
+        }
     }
     node(name) {
         stage("checkout ${variant}") {
@@ -73,7 +79,7 @@ rocmtest tidy: rocmnode('rocmtest') { cmake_build ->
             rm -rf build
             mkdir build
             cd build
-            CXX='clang++-5.0' cmake .. 
+            CXX=hcc cmake .. 
             make -j8 -k analyze
         '''
     }
@@ -91,11 +97,12 @@ rocmtest tidy: rocmnode('rocmtest') { cmake_build ->
             | xargs -n 1 -P 1 -I{} -t sh -c \'clang-format-5.0 -style=file {} | diff - {}\'
         '''
     }
-}, clang: rocmnode('rocmtest') { cmake_build ->
+}, clang: rocmnode('vega') { cmake_build ->
     stage('Clang Debug') {
         // TODO: Enanle integer
         def sanitizers = "undefined"
-        cmake_build("hcc", "-DCMAKE_BUILD_TYPE=debug -DCMAKE_CXX_FLAGS_DEBUG='-g -fno-omit-frame-pointer -fsanitize=${sanitizers} -fno-sanitize-recover=${sanitizers}'")
+        def debug_flags = "-g -fno-omit-frame-pointer -fsanitize=${sanitizers} -fno-sanitize-recover=${sanitizers}"
+        cmake_build("hcc", "-DCMAKE_BUILD_TYPE=debug -DMIGRAPHX_ENABLE_PYTHON=Off -DCMAKE_CXX_FLAGS_DEBUG='${debug_flags}'")
     }
     stage('Clang Release') {
         cmake_build("hcc", "-DCMAKE_BUILD_TYPE=release")
@@ -113,6 +120,19 @@ rocmtest tidy: rocmnode('rocmtest') { cmake_build ->
         def cmake_linker_flags = "-DCMAKE_EXE_LINKER_FLAGS='${linker_flags}' -DCMAKE_SHARED_LINKER_FLAGS='${linker_flags}'"
         // TODO: Add bounds-strict
         def sanitizers = "undefined,address"
-        cmake_build("g++-7", "-DCMAKE_BUILD_TYPE=debug ${cmake_linker_flags} -DCMAKE_CXX_FLAGS_DEBUG='-g -fno-omit-frame-pointer -fsanitize-address-use-after-scope -fsanitize=${sanitizers} -fno-sanitize-recover=${sanitizers}'")
+        def debug_flags = "-g -fprofile-arcs -ftest-coverage -fno-omit-frame-pointer -fsanitize-address-use-after-scope -fsanitize=${sanitizers} -fno-sanitize-recover=${sanitizers}"
+        cmake_build("g++-7", "-DCMAKE_BUILD_TYPE=debug -DMIGRAPHX_ENABLE_PYTHON=Off ${cmake_linker_flags} -DCMAKE_CXX_FLAGS_DEBUG='${debug_flags}'")
+
+    }
+    stage('Codecov') {
+        env.CODECOV_TOKEN="8545af1c-f90b-4345-92a5-0d075503ca56"
+        sh '''
+            cd build
+            lcov --directory . --capture --output-file coverage.info
+            lcov --remove coverage.info '/usr/*' --output-file coverage.info
+            lcov --list coverage.info
+            curl -s https://codecov.io/bash | bash
+            echo "Uploaded"
+        '''
     }
 }

README.md

@@ -8,6 +8,7 @@ AMD's graph optimization engine.
 * [HIP](https://github.com/ROCm-Developer-Tools/HIP) for running on the GPU
 * [Protobuf](https://github.com/google/protobuf) for reading [onxx](https://github.com/onnx/onnx) files
 * [Half](http://half.sourceforge.net/) - IEEE 754-based half-precision floating point library
+* [pybind11](https://pybind11.readthedocs.io/en/stable/) - for python bindings
 
 ## Installing the dependencies
 

codecov.yml

+ignore:
+  - "test/"
+ 

cppcheck.rules

@@ -34,11 +34,19 @@
 </rule>
 <rule>
     <tokenlist>define</tokenlist>
-    <pattern>define (MIGRAP|[^H]{6})[^H][^_]</pattern>
+    <pattern>define (MIGRAPH|[^X]{7})[^X][^_]</pattern>
     <message>
         <id>definePrefix</id>
         <severity>style</severity>
-        <summary>Macros must be prefixed with MIGRAPH_</summary>
+        <summary>Macros must be prefixed with MIGRAPHX_</summary>
+    </message>
+</rule>
+<rule>
+    <pattern>mutable \w+</pattern>
+    <message>
+        <id>MutableVariable</id>
+        <severity>style</severity>
+        <summary>Do not create mutable variables.</summary>
     </message>
 </rule>
 <rule>
@@ -74,7 +82,7 @@
     </message>
 </rule>
 <rule>
-    <pattern>(fclose|free|hipFree) \(</pattern>
+    <pattern>\\W(fclose|free|hipFree|hipHostFree|hipFreeArray|hipMemFree|hipStreamDestroy|hipEventDestroy|hipArrayDestroy|hipCtxDestroy|hipDestroyTextureObject|hipDestroySurfaceObject) \(</pattern>
     <message>
         <id>useManagePointer</id>
         <severity>style</severity>

dev-requirements.txt

 pfultz2/rocm-recipes
-pcre
-danmar/cppcheck@f965e5873 -DHAVE_RULES=1
-ROCm-Developer-Tools/HIP@3c7f5dbce24802ec4237e615038daff2909a2e8e
-# python/cpython@v3.6.6 -X autotools -H sha256:92aa914572c695c0aeb01b0a214813f414da4b51a371234df514a74761f2bb36
+danmar/cppcheck@681cb7dd909d1bfe41796b7616e43265177b9464 -DHAVE_RULES=1
+ROCm-Developer-Tools/HIP@fc22ef991ce7cb15821c8ccb4f03cdfc3e7e43cf
+python/cpython@v3.6.6 -X autotools -H sha256:92aa914572c695c0aeb01b0a214813f414da4b51a371234df514a74761f2bb36
 -f requirements.txt

doc/CMakeLists.txt

@@ -32,7 +32,7 @@ add_doxygen_doc(
     PREDEFINED DOXYGEN
 )
 add_custom_target(remove_inline_ns 
-    sed -i "s/MIGRAPH_INLINE_NS:://g" *.xml 
+    sed -i "s/MIGRAPHX_INLINE_NS:://g" *.xml 
     WORKING_DIRECTORY ${DOXYGEN_OUTPUT}/xml)
 add_dependencies(remove_inline_ns doxygen)
 

doc/src/reference/data.rst

@@ -21,7 +21,7 @@ raw_data
 
 .. doxygenstruct:: migraphx::raw_data
 
-.. doxygenfunction:: migraphx::MIGRAPH_INLINE_NS::visit_all
+.. doxygenfunction:: migraphx::MIGRAPHX_INLINE_NS::visit_all
 
 
 tensor_view

doc/src/reference/program.rst

@@ -21,4 +21,4 @@ program
 parse_onnx
 ----------
 
-.. doxygenfunction:: migraphx::MIGRAPH_INLINE_NS::parse_onnx
+.. doxygenfunction:: migraphx::MIGRAPHX_INLINE_NS::parse_onnx

requirements.txt

 google/protobuf -DCMAKE_POSITION_INDEPENDENT_CODE=On
-RadeonOpenCompute/rocm-cmake@6240bb3 --build
-ROCmSoftwarePlatform/MIOpen@1.6.0
-ROCmSoftwarePlatform/rocBLAS@v14.3.0
+RadeonOpenCompute/rocm-cmake@ac45c6e --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
 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

@@ -11,6 +11,7 @@ add_library(migraphx
     eliminate_contiguous.cpp
     eliminate_concat.cpp
     fwd_conv_batchnorm_rewrite.cpp
+    rewrite_rnn.cpp
     env.cpp
     generate.cpp
     instruction.cpp
@@ -35,8 +36,9 @@ target_include_directories(migraphx SYSTEM PUBLIC $<BUILD_INTERFACE:${HALF_INCLU
 set(PACKAGE_DEPENDS)
 
 add_subdirectory(onnx)
+add_subdirectory(py)
 add_subdirectory(targets/cpu)
-if(MIGRAPH_ENABLE_GPU)
+if(MIGRAPHX_ENABLE_GPU)
 list(APPEND PACKAGE_DEPENDS MIOpen rocblas)
 add_subdirectory(targets/gpu)
 endif()

src/auto_contiguous.cpp

@@ -5,14 +5,14 @@
 #include <migraphx/iterator_for.hpp>
 
 namespace migraphx {
-inline namespace MIGRAPH_INLINE_NS {
+inline namespace MIGRAPHX_INLINE_NS {
 
 void auto_contiguous::apply(program& p) const
 {
     for(auto ins : iterator_for(p))
     {
         shape s = ins->get_shape();
-        if(not s.standard())
+        if(not s.standard() and s.elements() != 0)
         {
             auto c = p.insert_instruction(std::next(ins), op::contiguous{}, ins);
             p.replace_instruction(ins, c);
@@ -20,5 +20,5 @@ void auto_contiguous::apply(program& p) const
     }
 }
 
-} // namespace MIGRAPH_INLINE_NS
+} // namespace MIGRAPHX_INLINE_NS
 } // namespace migraphx

src/common_subexpression_elimination.cpp

@@ -8,7 +8,7 @@
 #include <unordered_set>
 
 namespace migraphx {
-inline namespace MIGRAPH_INLINE_NS {
+inline namespace MIGRAPHX_INLINE_NS {
 
 template <class Range>
 void cse_range(program& p, Range&& r)
@@ -35,5 +35,5 @@ void cse_range(program& p, Range&& r)
 
 void common_subexpression_elimination::apply(program& p) const { cse_range(p, iterator_for(p)); }
 
-} // namespace MIGRAPH_INLINE_NS
+} // namespace MIGRAPHX_INLINE_NS
 } // namespace migraphx

src/constant_propagate.cpp

@@ -4,7 +4,7 @@
 #include <migraphx/literal.hpp>
 
 namespace migraphx {
-inline namespace MIGRAPH_INLINE_NS {
+inline namespace MIGRAPHX_INLINE_NS {
 
 struct match_const_add
 {
@@ -13,7 +13,7 @@ struct match_const_add
         return match::name("add")(match::args(match::name("@literal"), match::name("@literal")));
     }
 
-    void apply(program& p, match::matcher_result r) const
+    void apply(program& p, const match::matcher_result& r) const
     {
         auto ins  = r.result;
         auto arg1 = ins->inputs().at(0)->get_literal();
@@ -26,5 +26,5 @@ struct match_const_add
 
 void constant_propagate::apply(program& p) const { match::find_matches(p, match_const_add{}); }
 
-} // namespace MIGRAPH_INLINE_NS
+} // namespace MIGRAPHX_INLINE_NS
 } // namespace migraphx

src/dead_code_elimination.cpp

@@ -6,7 +6,7 @@
 #include <migraphx/ranges.hpp>
 
 namespace migraphx {
-inline namespace MIGRAPH_INLINE_NS {
+inline namespace MIGRAPHX_INLINE_NS {
 
 template <class Range, class Iterator>
 std::ptrdiff_t bidistance(const Range& r, Iterator start, Iterator last)
@@ -41,8 +41,9 @@ void dead_code_elimination::apply(program& p) const
         // Skip the last instruction
         if(i == last)
             break;
-        // Skip instruction with empty shape as output unless its a builtin
-        if(i->get_shape().elements() == 0 and not(i->name().front() == '@'))
+        // Skip instruction with empty shape as output unless its a builtin or undefined
+        if(i->get_shape().elements() == 0 and not(i->name().front() == '@') and
+           not(i->name() == "undefined"))
             continue;
         assert(bidistance(p, i, last) > 0);
         fix([&](auto self, auto leaf) {
@@ -62,5 +63,5 @@ void dead_code_elimination::apply(program& p) const
     p.remove_instructions(std::next(last), p.end());
 }
 
-} // namespace MIGRAPH_INLINE_NS
+} // namespace MIGRAPHX_INLINE_NS
 } // namespace migraphx

src/eliminate_allocation.cpp

@@ -8,13 +8,11 @@
 #include <migraphx/pass_config.hpp>
 
 namespace migraphx {
-inline namespace MIGRAPH_INLINE_NS {
+inline namespace MIGRAPHX_INLINE_NS {
 
 void eliminate_allocation::apply(program& p) const
 {
     assert(alignment > 0);
-    if(!enabled(MIGRAPH_DISABLE_MEMORY_COLORING{}))
-        return;
 
     std::size_t n = 0;
     std::vector<std::pair<instruction_ref, std::size_t>> allocs;
@@ -27,15 +25,18 @@ void eliminate_allocation::apply(program& p) const
         std::size_t padding = (alignment - (size % alignment)) % alignment;
         n += size + padding;
     }
-    auto mem = p.add_parameter("memory", shape{shape::int8_type, {n}});
-    for(auto&& pp : allocs)
+    if(n > 0)
     {
-        auto ins    = pp.first;
-        auto s      = ins->get_shape();
-        auto offset = pp.second;
-        p.replace_instruction(ins, op::load{s, offset}, mem);
+        auto mem = p.add_parameter("memory", shape{shape::int8_type, {n}});
+        for(auto&& pp : allocs)
+        {
+            auto ins    = pp.first;
+            auto s      = ins->get_shape();
+            auto offset = pp.second;
+            p.replace_instruction(ins, op::load{s, offset}, mem);
+        }
     }
 }
 
-} // namespace MIGRAPH_INLINE_NS
+} // namespace MIGRAPHX_INLINE_NS
 } // namespace migraphx

src/eliminate_concat.cpp

@@ -7,7 +7,7 @@
 #include <migraphx/dfor.hpp>
 
 namespace migraphx {
-inline namespace MIGRAPH_INLINE_NS {
+inline namespace MIGRAPHX_INLINE_NS {
 void eliminate_concat::apply(program& p) const
 {
     for(auto ins : iterator_for(p))
@@ -36,14 +36,17 @@ void eliminate_concat::apply(program& p) const
             // Where are the allocations for the tensors to be concatenated?
             std::vector<instruction_ref> allocations;
 
-            for(auto ins2 = ins->inputs().begin(); ins2 != ins->inputs().end() - 1; ins2++)
-            {
-                auto last2 = (*ins2)->inputs().back();
-                if(last2->name() == concat_opt.allocate())
-                {
-                    allocations.push_back(last2);
-                }
-            }
+            std::transform(
+                ins->inputs().begin(),
+                std::prev(ins->inputs().end()),
+                std::back_inserter(allocations),
+                [&](instruction_ref x) { return instruction::get_output_alias(x, true); });
+
+            if(std::any_of(allocations.begin(), allocations.end(), [&](auto x) {
+                   return x->name() != concat_opt.allocate();
+               }))
+                continue;
+
             // Need to sort the allocations, so that we know where to
             // insert the "super"-allocation
             std::sort(
@@ -51,15 +54,15 @@ void eliminate_concat::apply(program& p) const
                     return std::distance(p.begin(), x) < std::distance(p.begin(), y);
                 });
             // Move "super" allocation to the front
-            auto first         = allocations.front();
-            auto super         = p.move_instruction(last, first);
+            auto first = allocations.front();
+            auto super = p.move_instruction(last, first);
+            // Replace each allocation with a load
             std::size_t offset = 0;
-            for(auto x : allocations)
+            for(auto alloc : allocations)
             {
-                migraphx::op::load op{x->get_shape(), offset};
-                // migraphx::op::load op{x->get_shape(), 0};
-                p.replace_instruction(x, op, {super});
-                offset += x->get_shape().bytes();
+                op::load op{alloc->get_shape(), offset};
+                p.replace_instruction(alloc, op, {super});
+                offset += alloc->get_shape().bytes();
             }
             std::vector<instruction_ref> args = {super};
             std::copy(ins->inputs().begin(), ins->inputs().end() - 1, std::back_inserter(args));
@@ -67,5 +70,5 @@ void eliminate_concat::apply(program& p) const
         }
     }
 }
-} // namespace MIGRAPH_INLINE_NS
+} // namespace MIGRAPHX_INLINE_NS
 } // namespace migraphx

src/eliminate_contiguous.cpp

@@ -8,7 +8,7 @@
 #include <utility>
 
 namespace migraphx {
-inline namespace MIGRAPH_INLINE_NS {
+inline namespace MIGRAPHX_INLINE_NS {
 
 bool try_compute_shape(const operation& op, const std::vector<instruction_ref>& args)
 {
@@ -47,5 +47,5 @@ void eliminate_contiguous::apply(program& p) const
     }
 }
 
-} // namespace MIGRAPH_INLINE_NS
+} // namespace MIGRAPHX_INLINE_NS
 } // namespace migraphx