Merge remote-tracking branch 'origin/develop' into bert-attention

CMakeLists.txt

@@ -203,6 +203,8 @@ rocm_enable_cppcheck(
         useSmartPointer:*make_shared_array.hpp
         constParameter:*src/targets/gpu/*.cpp
         constParameter:*src/targets/gpu/*.hpp
+        # Suppress mlir_conv.cpp since this file will be deleted
+        *:*src/targets/gpu/mlir_conv.cpp
     FORCE
     INCONCLUSIVE
     RULE_FILE

dev-requirements.txt

@@ -2,6 +2,6 @@ pfultz2/rocm-recipes
 facebook/zstd@v1.4.5 -X subdir -DCMAKE_DIR=build/cmake
 ccache@v4.1
 pcre,pfultz2/pcre@8.45 -H sha256:d6f7182602a775a7d500a0cedca6449af0400c6493951513046d17615ed0bf11
-danmar/cppcheck@2.6 -DHAVE_RULES=1
+danmar/cppcheck@2.8 -DHAVE_RULES=1
 RadeonOpenCompute/rocm-cmake@1ebf7e7bc61bb5e949c171562b421264065230a7 --build
 -f requirements.txt

examples/nlp/python_bert_squad/BERT-Squad.ipynb

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

examples/nlp/python_bert_squad/requirements_bertsquad.txt

-tensorflow==2.5.3
+tensorflow==2.6.4
 onnxruntime
 tokenizers
\ No newline at end of file

examples/vision/python_yolov4/yolov4_inference.ipynb

@@ -24,16 +24,16 @@
     "import os.path\n",
     "\n",
     "if not os.path.exists(\"./utilities/coco.names\"):\n",
-    "    !wget https://github.com/onnx/models/raw/master/vision/object_detection_segmentation/yolov4/dependencies/coco.names -P ./utilities/\n",
+    "    !wget https://github.com/onnx/models/raw/main/vision/object_detection_segmentation/yolov4/dependencies/coco.names -P ./utilities/\n",
     "if not os.path.exists(\"./utilities/yolov4_anchors.txt\"):\n",
-    "    !wget https://github.com/onnx/models/raw/master/vision/object_detection_segmentation/yolov4/dependencies/yolov4_anchors.txt -P ./utilities/\n",
+    "    !wget https://github.com/onnx/models/raw/main/vision/object_detection_segmentation/yolov4/dependencies/yolov4_anchors.txt -P ./utilities/\n",
     "if not os.path.exists(\"./utilities/input.jpg\"):\n",
     "    # The image used is from the COCO dataset (https://cocodataset.org/#explore)\n",
     "    # Other images can be tested by replacing the link below\n",
     "    image_link = \"https://farm3.staticflickr.com/2009/2306189268_88cc86b30f_z.jpg\"\n",
     "    !wget -O ./utilities/input.jpg $image_link\n",
     "if not os.path.exists(\"./utilities/yolov4.onnx\"):\n",
-    "    !wget https://github.com/onnx/models/raw/master/vision/object_detection_segmentation/yolov4/model/yolov4.onnx -P ./utilities/"
+    "    !wget https://github.com/onnx/models/raw/main/vision/object_detection_segmentation/yolov4/model/yolov4.onnx -P ./utilities/"
    ]
   },
   {

src/api/include/migraphx/migraphx.hpp

@@ -39,10 +39,7 @@ template <class T, class F, class... Ts>
 T* make(F f, Ts&&... xs)
 {
     T* result = nullptr;
-    // cppcheck-suppress redundantInitialization
-    // cppcheck-suppress redundantAssignment
-    // cppcheck-suppress unreadVariable
-    auto e = f(&result, std::forward<Ts>(xs)...);
+    auto e    = f(&result, std::forward<Ts>(xs)...);
     if(e != migraphx_status_success)
         throw std::runtime_error("Failed to call function");
     return result;
@@ -51,9 +48,6 @@ T* make(F f, Ts&&... xs)
 template <class F, class... Ts>
 void call(F f, Ts&&... xs)
 {
-    // cppcheck-suppress redundantInitialization
-    // cppcheck-suppress redundantAssignment
-    // cppcheck-suppress unreadVariable
     auto e = f(std::forward<Ts>(xs)...);
     if(e != migraphx_status_success)
         throw std::runtime_error("Failed to call function");
@@ -340,7 +334,6 @@ 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)...);
         });

src/argument.cpp

@@ -29,7 +29,6 @@ void argument::assign_buffer(std::function<char*()> d)
     // Collect all shapes
     std::unordered_map<std::size_t, shape> shapes;
     {
-        // cppcheck-suppress variableScope
         std::size_t i = 0;
         fix([&](auto self, auto ss) {
             if(ss.sub_shapes().empty())
@@ -60,7 +59,6 @@ void argument::assign_buffer(std::function<char*()> d)
     }
     assert(offset == s.bytes());
 
-    // cppcheck-suppress variableScope
     std::size_t i = 0;
     m_data        = fix<data_t>([&](auto self, auto ss) {
         data_t result;

src/include/migraphx/raw_data.hpp

@@ -207,8 +207,7 @@ auto visit_all_pack(const shape& s, V1&& v1)
 template <class T, class... Ts>
 auto visit_all(T&& x, Ts&&... xs)
 {
-    auto&& s = x.get_shape();
-    // cppcheck-suppress redundantInitialization
+    auto&& s                                   = x.get_shape();
     std::initializer_list<shape::type_t> types = {xs.get_shape().type()...};
     if(!std::all_of(types.begin(), types.end(), [&](shape::type_t t) { return t == s.type(); }))
         MIGRAPHX_THROW("Types must be the same");

src/include/migraphx/serialize.hpp

@@ -50,7 +50,6 @@ auto to_value_impl(rank<2>, const T& x) -> decltype(x.begin(), x.end(), value{})
     value result = value::array{};
     for(auto&& y : x)
     {
-        auto e = to_value(y);
         result.insert(to_value(y));
     }
     return result;

src/include/migraphx/tensor_view.hpp

@@ -120,10 +120,8 @@ struct tensor_view
         return m_data[m_shape.index(this->size() - 1)];
     }
 
-    // cppcheck-suppress functionConst
     iterator begin() { return {0, {this}}; }
 
-    // cppcheck-suppress functionConst
     iterator end() { return {this->size(), {this}}; }
 
     const_iterator begin() const { return {0, {this}}; }

src/include/migraphx/verify.hpp

@@ -168,7 +168,6 @@ bool verify_range(const R1& r1, const R2& r2, double tolerance = 80, double* out
 {
     double threshold = std::numeric_limits<range_value<R1>>::epsilon() * tolerance;
     auto error       = rms_range(r1, r2);
-    // cppcheck-suppress uninitvar
     if(out_error != nullptr)
         *out_error = error;
     return error <= threshold;

src/module.cpp

@@ -729,7 +729,6 @@ std::unordered_map<instruction_ref, std::string>
 module::print_cpp(std::ostream& os, std::unordered_map<instruction_ref, std::string> names) const
 {
     os << "migraphx::module p;" << std::endl;
-    // cppcheck-suppress variableScope
     unsigned long seed = 0;
     names              = this->print(
         [&](auto ins, auto ins_names) {

src/onnx/parse_mean.cpp

@@ -2,6 +2,7 @@
 #include <migraphx/onnx/checks.hpp>
 #include <migraphx/instruction.hpp>
 #include <migraphx/make_op.hpp>
+#include <migraphx/ranges.hpp>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
@@ -9,6 +10,9 @@ namespace onnx {
 
 struct parse_mean : op_parser<parse_mean>
 {
+    const std::set<shape::type_t> float_types = {
+        shape::float_type, shape::half_type, shape::double_type};
+
     std::vector<op_desc> operators() const { return {{"Mean"}}; }
 
     /// Calculates the element-wise mean of n>=1 input tensors
@@ -24,17 +28,29 @@ struct parse_mean : op_parser<parse_mean>
         auto divisor = info.add_literal(
             migraphx::literal{migraphx::shape{args[0]->get_shape().type()}, {num_data}});
 
-        // TODO: Only divide when using floating-point
-        return std::accumulate(args.begin() + 1,
-                               args.end(),
-                               info.add_broadcastable_binary_op("div", args[0], divisor),
-                               [&](auto mean, auto data_i) {
-                                   // Pre-divide each tensor element-wise by n to reduce risk of
-                                   // overflow during summation
-                                   auto div =
-                                       info.add_broadcastable_binary_op("div", data_i, divisor);
-                                   return info.add_broadcastable_binary_op("add", mean, div);
-                               });
+        if(contains(float_types, args[0]->get_shape().type()))
+        {
+            return std::accumulate(args.begin() + 1,
+                                   args.end(),
+                                   info.add_broadcastable_binary_op("div", args[0], divisor),
+                                   [&](auto mean, auto data_i) {
+                                       // Pre-divide each tensor element-wise by n to reduce risk of
+                                       // overflow during summation
+                                       auto div =
+                                           info.add_broadcastable_binary_op("div", data_i, divisor);
+                                       return info.add_broadcastable_binary_op("add", mean, div);
+                                   });
+        }
+        else
+        {
+            // Compute sum before division for integral types
+            auto sum = std::accumulate(
+                args.begin() + 1, args.end(), args[0], [&](auto accum, auto data_i) {
+                    return info.add_broadcastable_binary_op("add", accum, data_i);
+                });
+
+            return info.add_broadcastable_binary_op("div", sum, divisor);
+        }
     }
 };
 

src/onnx/parse_pooling.cpp

@@ -128,7 +128,7 @@ struct parse_pooling : op_parser<parse_pooling>
             std::fill_n(values["stride"].begin(), kdims, 1);
         }
         // used to calculate the supposed output shape
-        std::vector<int64_t> orig_padding(paddings.begin(), paddings.end());
+        std::vector<int64_t> orig_padding = paddings;
 
         std::vector<int64_t> slice_start;
         std::vector<int64_t> slice_end;

src/onnx/parse_squeeze.cpp

@@ -30,11 +30,11 @@ struct parse_squeeze : op_parser<parse_squeeze>
                           std::vector<instruction_ref> args) const
     {
         auto op = parser.load(opd.op_name, info);
-        std::vector<int64_t> axes;
         if(args.size() == 2)
         {
             auto arg_axes = args.at(1)->eval();
             check_arg_empty(arg_axes, "PARSE_" + opd.op_name + ": cannot handle variable axes!");
+            std::vector<int64_t> axes;
             arg_axes.visit([&](auto s) { axes.assign(s.begin(), s.end()); });
             op = assign_axes(op, axes);
         }

src/process.cpp

@@ -20,7 +20,6 @@ int exec(const std::string& cmd, const std::function<void(const char*)>& std_out
     int ec = 0;
     if(enabled(MIGRAPHX_TRACE_CMD_EXECUTE{}))
         std::cout << cmd << std::endl;
-    std::array<char, 128> buffer;
     auto closer = [&](FILE* stream) {
         auto status = pclose(stream);
         ec          = WIFEXITED(status) ? 0 : WEXITSTATUS(status); // NOLINT
@@ -30,6 +29,7 @@ int exec(const std::string& cmd, const std::function<void(const char*)>& std_out
         std::unique_ptr<FILE, decltype(closer)> pipe(popen(cmd.c_str(), "r"), closer); // NOLINT
         if(!pipe)
             MIGRAPHX_THROW("popen() failed: " + cmd);
+        std::array<char, 128> buffer;
         while(fgets(buffer.data(), buffer.size(), pipe.get()) != nullptr)
             std_out(buffer.data());
     }

src/propagate_constant.cpp

@@ -3,6 +3,7 @@
 #include <migraphx/matcher.hpp>
 #include <migraphx/literal.hpp>
 #include <migraphx/functional.hpp>
+#include <migraphx/par_for.hpp>
 #include <unordered_set>
 
 namespace migraphx {
@@ -20,33 +21,42 @@ bool skip_propogate(instruction_ref ins)
     return false;
 }
 
+bool is_const(instruction_ref ins) { return ins->can_eval() and not skip_propogate(ins); }
+
 void propagate_constant::apply(module& m) const
 {
+    std::unordered_set<instruction_ref> const_instrs;
+    auto last = std::prev(m.end());
+
+    // Find instructions that can be evaluated to a literal
     for(auto i : iterator_for(m))
     {
-        if(i->name() != "@literal")
+        if(is_const(i) and i != last)
             continue;
-        if(i->outputs().empty())
-            continue;
-        fix([&](auto self, auto ins) {
-            std::unordered_set<instruction_ref> children(ins->outputs().begin(),
-                                                         ins->outputs().end());
-            for(auto child : children)
-            {
-                if(child->name() == "@literal" or skip_propogate(child))
-                {
-                    self(child);
-                    continue;
-                }
-                auto r = child->eval();
-                if(not r.empty())
-                {
-                    assert(r.get_shape() == child->get_shape());
-                    auto l = m.add_literal(r.get_shape(), r.data());
-                    self(m.replace_instruction(child, l));
-                }
-            }
-        })(i);
+
+        std::copy_if(
+            i->inputs().begin(),
+            i->inputs().end(),
+            std::inserter(const_instrs, const_instrs.begin()),
+            [&](const instruction_ref ins) { return is_const(ins) and ins->name() != "@literal"; });
+    }
+
+    // Compute literals in parallel
+    std::vector<instruction_ref> const_instrs_vec{const_instrs.begin(), const_instrs.end()};
+    std::vector<argument> literals(const_instrs_vec.size());
+    par_for(const_instrs_vec.size(), 1, [&](const auto i) {
+        literals[i] = const_instrs_vec[i]->eval();
+    });
+
+    // Replace instructions in m
+    for(size_t i = 0; i < const_instrs_vec.size(); i++)
+    {
+        if(not literals[i].empty())
+        {
+            assert(literals[i].get_shape() == const_instrs_vec[i]->get_shape());
+            auto l = m.add_literal(literals[i].get_shape(), literals[i].data());
+            m.replace_instruction(const_instrs_vec[i], l);
+        }
     }
 }
 

src/reduce_dims.cpp

@@ -16,10 +16,8 @@ bool reduce_dim(std::vector<shape>& shapes, std::size_t n)
         auto bstride = s.strides()[n + 1];
         auto blen    = s.lens()[n + 1];
 
-        if(astride == bstride * blen)
-        {
+        if(astride == bstride * blen or alen == 1)
             new_lens.push_back(alen * blen);
-        }
     }
     if(new_lens.size() != shapes.size())
         return false;
@@ -37,10 +35,25 @@ bool reduce_dim(std::vector<shape>& shapes, std::size_t n)
     return true;
 }
 
+void reduce_dim1(std::vector<shape>& shapes)
+{
+    if(std::any_of(shapes.begin(), shapes.end(), [&](const auto& s) {
+           return s.lens().size() < 2 or s.lens().back() != 1;
+       }))
+        return;
+    for(auto& s : shapes)
+    {
+        auto lens    = s.lens();
+        auto strides = s.strides();
+        lens.pop_back();
+        strides.pop_back();
+        s = shape{s.type(), lens, strides};
+    }
+}
+
 std::size_t reduce_dim_all(std::vector<shape>& shapes, std::size_t n)
 {
     while(reduce_dim(shapes, n) and n < shapes.size()) {}
-
     return n + 1;
 }
 void reduce_dim_all(std::vector<shape>& shapes)
@@ -48,6 +61,7 @@ void reduce_dim_all(std::vector<shape>& shapes)
     std::size_t n = 0;
     while(n < shapes.front().lens().size() - 1)
         n = reduce_dim_all(shapes, n);
+    reduce_dim1(shapes);
 }
 
 std::vector<std::size_t> base_lens(const std::vector<shape>& shapes)

src/targets/cpu/copy.cpp

@@ -20,7 +20,6 @@ struct cpu_copy : reduce_dims_base, auto_register_op<cpu_copy>
         return inputs.at(1);
     }
     argument
-    // cppcheck-suppress constParameter
     compute(context& ctx, const shape& output_shape, const std::vector<argument>& args) const
     {
         argument result = get_arg(args, args.size() - 1);