Merge branch 'develop' into gpu-batch-gemm-bert

.github/workflows/ci.yaml

@@ -15,7 +15,8 @@ jobs:
 
     steps:
     - name: Free space
-      run: sudo rm -rf /usr/local/android /usr/share/dotnet /usr/local/share/boost /opt/ghc /usr/local/share/chrom* /usr/share/swift /usr/local/julia*
+      run: sudo rm -rf /usr/local/android /usr/share/dotnet /usr/local/share/boost /opt/ghc /usr/local/share/chrom* /usr/share/swift /usr/local/julia* /usr/local/lib/android
+
     - uses: actions/checkout@v2
 
     # In this step, this action saves a list of existing images,
@@ -63,7 +64,7 @@ jobs:
 
     steps:
     - name: Free space
-      run: sudo rm -rf /usr/local/android /usr/share/dotnet /usr/local/share/boost /opt/ghc /usr/local/share/chrom* /usr/share/swift /usr/local/julia*
+      run: sudo rm -rf /usr/local/android /usr/share/dotnet /usr/local/share/boost /opt/ghc /usr/local/share/chrom* /usr/share/swift /usr/local/julia* /usr/local/lib/android
     - uses: actions/checkout@v2
 
     # In this step, this action saves a list of existing images,
@@ -108,7 +109,7 @@ jobs:
 
     steps:
     - name: Free space
-      run: sudo rm -rf /usr/local/android /usr/share/dotnet /usr/local/share/boost /opt/ghc /usr/local/share/chrom* /usr/share/swift /usr/local/julia*
+      run: sudo rm -rf /usr/local/android /usr/share/dotnet /usr/local/share/boost /opt/ghc /usr/local/share/chrom* /usr/share/swift /usr/local/julia* /usr/local/lib/android
     - uses: actions/checkout@v2
 
     # In this step, this action saves a list of existing images,
@@ -143,7 +144,7 @@ jobs:
 
     steps:
     - name: Free space
-      run: sudo rm -rf /usr/local/android /usr/share/dotnet /usr/local/share/boost /opt/ghc /usr/local/share/chrom* /usr/share/swift /usr/local/julia*
+      run: sudo rm -rf /usr/local/android /usr/share/dotnet /usr/local/share/boost /opt/ghc /usr/local/share/chrom* /usr/share/swift /usr/local/julia* /usr/local/lib/android
     - uses: actions/checkout@v2
     - name: Set up Python
       uses: actions/setup-python@v2
@@ -182,7 +183,7 @@ jobs:
 
     steps:
     - name: Free space
-      run: sudo rm -rf /usr/local/android /usr/share/dotnet /usr/local/share/boost /opt/ghc /usr/local/share/chrom* /usr/share/swift /usr/local/julia*
+      run: sudo rm -rf /usr/local/android /usr/share/dotnet /usr/local/share/boost /opt/ghc /usr/local/share/chrom* /usr/share/swift /usr/local/julia* /usr/local/lib/android
     - uses: actions/checkout@v2
     - name: Set up Python
       uses: actions/setup-python@v2

CMakeLists.txt

@@ -93,11 +93,14 @@ rocm_enable_clang_tidy(
         modernize-*
         performance-*
         readability-*
-        -bugprone-signed-char-misuse
+        -bugprone-easily-swappable-parameters
+        -bugprone-implicit-widening-of-multiplication-result
         -bugprone-macro-parentheses
+        -bugprone-signed-char-misuse
         # Disable the aliased reserved identifiers
         -cert-dcl37-c
         -cert-dcl51-cpp
+        -cert-err33-c
         -cert-str34-c
         # Disable all alpha checks by default
         -clang-analyzer-alpha*
@@ -127,6 +130,7 @@ rocm_enable_clang_tidy(
         -cppcoreguidelines-pro-type-union-access
         -cppcoreguidelines-pro-type-vararg
         -cppcoreguidelines-special-member-functions
+        -cppcoreguidelines-virtual-class-destructor
         -google-readability-*
         -google-runtime-int
         -google-runtime-references
@@ -144,8 +148,10 @@ rocm_enable_clang_tidy(
         -readability-convert-member-functions-to-static
         -readability-else-after-return
         -readability-function-cognitive-complexity
+        -readability-identifier-length
         -readability-named-parameter
         -readability-redundant-string-init
+        -readability-suspicious-call-argument
         -readability-uppercase-literal-suffix
         -*-avoid-c-arrays
         -*-explicit-constructor
@@ -178,7 +184,7 @@ rocm_enable_cppcheck(
         style
         performance
         portability
-    SUPPRESS 
+    SUPPRESS
         ConfigurationNotChecked
         unmatchedSuppression
         unusedFunction
@@ -216,6 +222,7 @@ rocm_enable_cppcheck(
         CPPCHECK=1
         __device__=
         __host__=
+        __global__=
 )
 
 enable_testing()

Dockerfile

-FROM ubuntu:18.04
+FROM ubuntu:20.04
 
 ARG PREFIX=/usr/local
 
@@ -6,7 +6,7 @@ ARG PREFIX=/usr/local
 RUN dpkg --add-architecture i386
 
 # Add rocm repository
-RUN sh -c 'echo deb [arch=amd64 trusted=yes] http://repo.radeon.com/rocm/apt/4.5/ ubuntu main > /etc/apt/sources.list.d/rocm.list'
+RUN sh -c 'echo deb [arch=amd64 trusted=yes] http://repo.radeon.com/rocm/apt/5.0.2/ ubuntu main > /etc/apt/sources.list.d/rocm.list'
 
 # Install dependencies
 RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --allow-unauthenticated \
@@ -16,16 +16,12 @@ RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --allow-
     cmake \
     curl \
     doxygen \
-    g++-5 \
     g++-7 \
     gdb \
     git \
     lcov \
     locales \
     pkg-config \
-    python \
-    python-dev \
-    python-pip \
     python3 \
     python3-dev \
     python3-pip \

cmake/Embed.cmake

@@ -94,5 +94,6 @@ function(add_embed_library EMBED_NAME)
     generate_embed_source(${EMBED_NAME} SRC ${SRC_FILE} HEADER ${HEADER_FILE} OBJECTS ${OUTPUT_FILES} SYMBOLS ${SYMBOLS})
     add_library(${EMBED_NAME} STATIC ${OUTPUT_FILES} "${SRC_FILE}")
     target_include_directories(${EMBED_NAME} PUBLIC "${EMBED_DIR}/include")
+    target_compile_options(${EMBED_NAME} PRIVATE -Wno-reserved-identifier)
     set_target_properties(${EMBED_NAME} PROPERTIES POSITION_INDEPENDENT_CODE On)
 endfunction()

cmake/EnableCompilerWarnings.cmake

@@ -96,6 +96,7 @@ else()
                 -Wno-gnu-zero-variadic-macro-arguments
                 -Wno-missing-prototypes
                 -Wno-nested-anon-types
+                -Wno-option-ignored
                 -Wno-padded
                 -Wno-shorten-64-to-32
                 -Wno-sign-conversion

doc/src/reference/py.rst

@@ -146,6 +146,13 @@ module
     :param list[module] mod_args: optional list of module arguments to the operator.
     :rtype instruction
 
+.. py:method:: add_literal(data)
+
+    Adds constant or literal data of provided shape into the module from python buffer which includes numpy array.    
+
+    :param py::buffer data: Python buffer or numpy array 
+    :rtype instruction 
+
 .. py:method:: add_parameter(name, shape)
     
     Adds a parameter to the module with provided name and shape.

examples/nlp/python_bert_squad/BERT-Squad.ipynb

@@ -62,7 +62,7 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "!wget -nc https://github.com/onnx/models/raw/master/text/machine_comprehension/bert-squad/model/bertsquad-10.onnx"
+    "!wget -nc https://github.com/onnx/models/blob/main/text/machine_comprehension/bert-squad/model/bertsquad-10.onnx"
    ]
   },
   {

examples/nlp/python_bert_squad/README.md

@@ -23,7 +23,7 @@ unzip uncased_L-12_H-768_A-12.zip
 ```
 5) Get BERT ONNX model (bertsquad-10.onnx):
 ```
-wget https://github.com/onnx/models/raw/master/text/machine_comprehension/bert-squad/model/bertsquad-10.onnx
+wget https://github.com/onnx/models/blob/main/text/machine_comprehension/bert-squad/model/bertsquad-10.onnx
 ```
 6) Run the inference, it will compile and run the model on three questions and small data provided in `inputs.json`:
 ```

hip-clang.docker

-FROM ubuntu:18.04
+FROM ubuntu:20.04
 
 ARG PREFIX=/usr/local
 
@@ -6,7 +6,7 @@ ARG PREFIX=/usr/local
 RUN dpkg --add-architecture i386
 
 # Add rocm repository
-RUN sh -c 'echo deb [arch=amd64 trusted=yes] http://repo.radeon.com/rocm/apt/4.5/ ubuntu main > /etc/apt/sources.list.d/rocm.list'
+RUN sh -c 'echo deb [arch=amd64 trusted=yes] http://repo.radeon.com/rocm/apt/5.0.2/ ubuntu main > /etc/apt/sources.list.d/rocm.list'
 
 # Install dependencies
 RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --allow-unauthenticated \
@@ -20,9 +20,6 @@ RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --allow-
     git \
     lcov \
     pkg-config \
-    python \
-    python-dev \
-    python-pip \
     python3 \
     python3-dev \
     python3-pip \

src/adjust_allocation.cpp

@@ -8,9 +8,9 @@
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 
-void adjust_allocation::apply(module& p) const
+void adjust_allocation::apply(module& m) const
 {
-    for(auto ins : iterator_for(p))
+    for(auto ins : iterator_for(m))
     {
         // skip instruction with no input
         if(ins->inputs().empty())
@@ -27,13 +27,13 @@ void adjust_allocation::apply(module& p) const
         // of the instruction, reallocate and replace the previous one
         if(alias_ins->get_shape() == ins->get_shape())
             continue;
-        auto alloc_ins = p.insert_instruction(ins, model.allocate(ins->get_shape()));
-        p.replace_instruction(alias_ins, alloc_ins);
+        auto alloc_ins = m.insert_instruction(ins, model.allocate(ins->get_shape()));
+        m.replace_instruction(alias_ins, alloc_ins);
         // If the memory is an output parameter then copy the memory to the parameter
         if(alias_ins->name() == "@param")
         {
-            auto copy = p.insert_instruction(std::next(ins), make_op(model.copy()), ins, alias_ins);
-            auto tail = range(std::next(copy), p.end());
+            auto copy = m.insert_instruction(std::next(ins), make_op(model.copy()), ins, alias_ins);
+            auto tail = range(std::next(copy), m.end());
             for(auto i : iterator_for(tail))
             {
                 if(contains(i->inputs(), ins))

src/analyze_streams.cpp

@@ -14,31 +14,31 @@ bool happens_before(const std::vector<std::size_t>& e1, const std::vector<std::s
            not std::equal(e1.begin(), e1.end(), e2.begin(), e2.end(), std::greater_equal<>{});
 }
 
-std::vector<stream_race> analyze_streams(const module& p, const stream_model& m)
+std::vector<stream_race> analyze_streams(const module& m, const stream_model& strmm)
 {
     using vector_clock = std::vector<std::size_t>;
     std::vector<stream_race> races;
-    auto nstream = m.get_nstream();
+    auto nstream = strmm.get_nstream();
     std::vector<vector_clock> vclock(nstream, vector_clock(nstream));
     std::unordered_map<instruction_ref, vector_clock> timestamp;
     std::unordered_map<std::size_t, vector_clock> events;
-    for(auto ins : iterator_for(p))
+    for(auto ins : iterator_for(m))
     {
-        if(not m.has_stream(ins))
+        if(not strmm.has_stream(ins))
             continue;
-        std::size_t s = m.get_stream(ins);
+        std::size_t s = strmm.get_stream(ins);
         assert(s < nstream);
         assert(vclock.size() == nstream);
         assert(vclock[s].size() == nstream);
-        if(m.is_record(ins))
+        if(strmm.is_record(ins))
         {
             vclock[s][s]++;
-            auto event    = m.get_event_id(ins);
+            auto event    = strmm.get_event_id(ins);
             events[event] = vclock[s];
         }
-        else if(m.is_wait(ins))
+        else if(strmm.is_wait(ins))
         {
-            auto event = m.get_event_id(ins);
+            auto event = strmm.get_event_id(ins);
             if(not contains(events, event))
                 MIGRAPHX_THROW("Event is waited on before being recorded: " +
                                std::to_string(event));
@@ -57,21 +57,21 @@ std::vector<stream_race> analyze_streams(const module& p, const stream_model& m)
         }
         timestamp[ins] = vclock[s];
     }
-    for(auto ins : iterator_for(p))
+    for(auto ins : iterator_for(m))
     {
-        if(not m.has_stream(ins))
+        if(not strmm.has_stream(ins))
             continue;
         if(ins->inputs().empty())
             continue;
-        std::size_t s = m.get_stream(ins);
+        std::size_t s = strmm.get_stream(ins);
         // Find inputs from different streams
         std::vector<instruction_ref> inputs;
         fix([&](auto self, auto start) {
             for(auto input : start->inputs())
             {
-                if(not m.has_stream(input))
+                if(not strmm.has_stream(input))
                     self(input);
-                else if(m.get_stream(input) != s)
+                else if(strmm.get_stream(input) != s)
                     inputs.push_back(input);
             }
         })(ins);

src/api/api.cpp

@@ -1072,6 +1072,22 @@ migraphx_module_add_instruction_with_mod_args(migraphx_instruction_t* out,
     return api_error_result;
 }
 
+extern "C" migraphx_status migraphx_module_add_literal(migraphx_instruction_t* out,
+                                                       migraphx_module_t module,
+                                                       const_migraphx_shape_t shape,
+                                                       const char* buffer)
+{
+    auto api_error_result = migraphx::try_([&] {
+        if(module == nullptr)
+            MIGRAPHX_THROW(migraphx_status_bad_param, "Bad parameter module: Null pointer");
+        if(shape == nullptr)
+            MIGRAPHX_THROW(migraphx_status_bad_param, "Bad parameter shape: Null pointer");
+        *out = allocate<migraphx_instruction_t>(
+            (module->object).add_literal((shape->object), (buffer)));
+    });
+    return api_error_result;
+}
+
 extern "C" migraphx_status migraphx_module_add_parameter(migraphx_instruction_t* out,
                                                          migraphx_module_t module,
                                                          const char* name,

src/api/include/migraphx/migraphx.h

@@ -258,6 +258,11 @@ migraphx_status migraphx_module_add_instruction_with_mod_args(migraphx_instructi
                                                               migraphx_instructions_t args,
                                                               migraphx_modules_t module_refs);
 
+migraphx_status migraphx_module_add_literal(migraphx_instruction_t* out,
+                                            migraphx_module_t module,
+                                            const_migraphx_shape_t shape,
+                                            const char* buffer);
+
 migraphx_status migraphx_module_add_parameter(migraphx_instruction_t* out,
                                               migraphx_module_t module,
                                               const char* name,

src/api/include/migraphx/migraphx.hpp

@@ -314,6 +314,7 @@ struct interface_base : Base
                 T** y = reinterpret_cast<T**>(out);
                 T* x  = reinterpret_cast<T*>(input);
                 assert(x != nullptr and y != nullptr and *y == nullptr);
+                // cppcheck-suppress useSmartPointer
                 *y = new T(*x); // NOLINT
             });
         };
@@ -339,6 +340,7 @@ struct interface_base : Base
     template <class T, class Setter, class F>
     void set_auto_fp(Setter setter, F f)
     {
+        // cppcheck-suppress constParameter
         return set_fp<T>(setter, [=](T& obj, auto out, auto... xs) {
             auto_invoke(f, out, obj, auto_convert_param(rank<2>{}, xs)...);
         });
@@ -760,6 +762,15 @@ struct module
         return instruction(op_ins, own{});
     }
 
+    template <typename T>
+    instruction add_literal(const migraphx::shape& s, T* buffer)
+    {
+        migraphx_instruction_t literal_ins;
+        const auto* buffer_ptr = reinterpret_cast<const char*>(buffer);
+        call(&migraphx_module_add_literal, &literal_ins, mm.get(), s.get_handle_ptr(), buffer_ptr);
+        return instruction(literal_ins, own{});
+    }
+
     instruction add_parameter(const std::string& name, shape s)
     {
         migraphx_instruction_t param_ins;

src/api/migraphx.py

@@ -212,6 +212,9 @@ def module(h):
                         module_refs='std::vector<migraphx::module*>'),
              fname='add_instruction',
              returns='migraphx::instruction_ref')
+    h.method('add_literal',
+             api.params(shape='const migraphx::shape&', buffer='const char*'),
+             returns='migraphx::instruction_ref')
     h.method('add_parameter',
              api.params(name='const char*', shape='const migraphx::shape&'),
              returns='migraphx::instruction_ref')

src/auto_contiguous.cpp

@@ -8,10 +8,10 @@
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 
-void auto_contiguous::apply(module& p) const
+void auto_contiguous::apply(module& m) const
 {
     std::string key = "require_std_shape";
-    for(auto ins : reverse_iterator_for(p))
+    for(auto ins : reverse_iterator_for(m))
     {
         auto&& attr = ins->get_operator().attributes();
         if((attr.get(key, false)))
@@ -23,18 +23,18 @@ void auto_contiguous::apply(module& p) const
                 {
                     return in;
                 }
-                return p.insert_instruction(ins, make_op("contiguous"), in);
+                return m.insert_instruction(ins, make_op("contiguous"), in);
             });
 
             if(new_args != args)
             {
-                p.replace_instruction(ins, ins->get_operator(), new_args);
+                m.replace_instruction(ins, ins->get_operator(), new_args);
             }
         }
     }
 
-    auto last = std::prev(p.end());
-    for(auto ins : iterator_for(p))
+    auto last = std::prev(m.end());
+    for(auto ins : iterator_for(m))
     {
         // for last instruction that is NOT a return
         if(ins->outputs().empty() and ins != last)
@@ -42,8 +42,8 @@ void auto_contiguous::apply(module& p) const
         shape s = ins->get_shape();
         if(not s.standard() and s.elements() != 0)
         {
-            auto c = p.insert_instruction(std::next(ins), make_op("contiguous"), ins);
-            p.replace_instruction(ins, c);
+            auto c = m.insert_instruction(std::next(ins), make_op("contiguous"), ins);
+            m.replace_instruction(ins, c);
         }
     }
 }

src/dead_code_elimination.cpp

@@ -9,26 +9,6 @@
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 
-template <class Range, class Iterator>
-std::ptrdiff_t bidistance(const Range& r, Iterator start, Iterator last)
-{
-    auto start_forward   = start;
-    auto start_backwards = start;
-    std::size_t n        = 0;
-    while(start_forward != last and start_backwards != last)
-    {
-        n++;
-        if(start_forward != r.end())
-            start_forward++;
-        if(start_backwards != r.begin())
-            start_backwards--;
-    }
-    if(start_forward == last)
-        return n;
-    else
-        return -n;
-}
-
 void dead_code_elimination::apply(program& p) const { p.remove_unused_modules(); }
 
 void dead_code_elimination::apply(module& m) const
@@ -48,17 +28,21 @@ void dead_code_elimination::apply(module& m) const
         if(i->get_shape().elements() == 0 and i->name().front() != '@' and
            i->name() != "undefined" and i->name() != "identity")
             continue;
-        assert(bidistance(m, i, last) > 0);
+        assert(std::distance(m.begin(), i) <= std::distance(m.begin(), last));
+        std::unordered_set<instruction_ref> visited;
         fix([&](auto self, auto leaf) {
             if(not m.has_instruction(leaf))
                 return;
 
             if(leaf->outputs().empty())
             {
+                // Dont visit inputs twice
+                if(not visited.insert(leaf).second)
+                    return;
                 std::unordered_set<instruction_ref> args(leaf->inputs().begin(),
                                                          leaf->inputs().end());
                 leaf->clear_arguments();
-                assert(bidistance(m, last, leaf) < 0);
+                assert(std::distance(m.begin(), leaf) < std::distance(m.begin(), last));
                 assert(leaf != ins);
                 if(leaf->name() != "@param")
                     m.move_instruction(leaf, m.end());

src/driver/marker_roctx.cpp

@@ -17,7 +17,7 @@ class marker_roctx
     std::function<int(const char*)> sym_roctx_range_push;
     std::function<int()> sym_roctx_range_pop;
 
-    uint64_t range_id;
+    uint64_t range_id = 0;
 
     public:
     marker_roctx()

src/eliminate_allocation.cpp

@@ -13,13 +13,13 @@
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 
-void eliminate_allocation::apply(module& p) const
+void eliminate_allocation::apply(module& m) const
 {
     assert(alignment > 0);
 
     std::size_t n = 0;
     std::vector<std::pair<instruction_ref, std::size_t>> allocs;
-    for(auto ins : iterator_for(p))
+    for(auto ins : iterator_for(m))
     {
         if(ins->name() != allocation_op)
             continue;
@@ -30,13 +30,13 @@ void eliminate_allocation::apply(module& p) const
     }
     if(n > 0)
     {
-        auto mem = p.add_parameter("memory", shape{shape::int8_type, {n}});
+        auto mem = m.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(
+            m.replace_instruction(
                 ins, make_op("load", {{"shape", to_value(s)}, {"offset", offset}}), mem);
         }
     }

src/eliminate_common_subexpression.cpp

@@ -11,7 +11,7 @@ namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 
 template <class Range>
-void cse_range(module& p, Range&& r)
+void cse_range(module& m, Range&& r)
 {
     std::unordered_multimap<std::string, instruction_ref> instructions;
     std::unordered_set<instruction_ref> processed_ins;
@@ -30,24 +30,24 @@ void cse_range(module& p, Range&& r)
                 continue;
             if(*eq != *ins)
                 continue;
-            p.replace_instruction(ins, eq);
+            m.replace_instruction(ins, eq);
             processed_ins.emplace(ins);
             std::vector<instruction_ref> outputs;
             std::copy_if(eq->outputs().begin(),
                          eq->outputs().end(),
                          std::back_inserter(outputs),
-                         [&](auto x) { return p.has_instruction(x); });
+                         [&](auto x) { return m.has_instruction(x); });
 
             std::sort(outputs.begin(), outputs.end(), [&](auto x, auto y) {
                 return std::distance(eq, x) < std::distance(eq, y);
             });
-            cse_range(p, outputs);
+            cse_range(m, outputs);
         }
         instructions.emplace(ins->name(), ins);
     }
 }
 
-void eliminate_common_subexpression::apply(module& p) const { cse_range(p, iterator_for(p)); }
+void eliminate_common_subexpression::apply(module& m) const { cse_range(m, iterator_for(m)); }
 
 } // namespace MIGRAPHX_INLINE_NS
 } // namespace migraphx