Merge branch 'develop' of github.com:ROCmSoftwarePlatform/AMDMIGraphX into dyn_model_test

src/onnx/parse_resize.cpp

@@ -56,7 +56,7 @@ const auto& get_nearest_op(const std::string& mode)
              return static_cast<std::size_t>(std::ceil((val)));
          }}};
 
-    if(!contains(nearest_ops, mode))
+    if(not contains(nearest_ops, mode))
     {
         MIGRAPHX_THROW("PARSE_RESIZE: nearest_mode " + mode + " not supported!");
     }
@@ -86,7 +86,7 @@ const auto& get_original_idx_op(const std::string& mode)
              return (idx + 0.5) / scale;
          }}};
 
-    if(!contains(idx_ops, mode))
+    if(not contains(idx_ops, mode))
     {
         MIGRAPHX_THROW("PARSE_RESIZE: coordinate_transformation_mode " + mode + " not supported!");
     }

src/onnx/parse_reversesequence.cpp

@@ -31,7 +31,7 @@ namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 namespace onnx {
 
-//!  Parser for ReverseSequence ONNX operator.
+// Parser for ReverseSequence ONNX operator.
 /*!
   Reverses the data along the time axis for the batches along the batch axis.
   The sequence lengths can be given to reverse up to the given length for each batch, keeping the

src/opt/memory_coloring.cpp

@@ -29,7 +29,7 @@ inline namespace MIGRAPHX_INLINE_NS {
 
 void memory_coloring::apply(module& m) const
 {
-    if(!enabled(MIGRAPHX_DISABLE_MEMORY_COLORING{}))
+    if(not enabled(MIGRAPHX_DISABLE_MEMORY_COLORING{}))
     {
         memory_coloring_impl opt(&m, allocation_op, verify);
         opt.run();

src/opt/memory_coloring_impl.cpp

@@ -42,7 +42,7 @@ void memory_coloring_impl::run()
     {
         MIGRAPHX_DEBUG(dump_intervals());
         // Coloring
-        while(!alloc_queue.empty())
+        while(not alloc_queue.empty())
         {
             interval_ptr interval = alloc_queue.top();
             allocate(interval);
@@ -72,7 +72,7 @@ bool memory_coloring_impl::allocate(interval_ptr interval)
 
     if(conflict_table.find(vn) != conflict_table.end())
     {
-        std::set<int>& vn_set = conflict_table[vn];
+        const std::set<int>& vn_set = conflict_table[vn];
         for(const auto& iter : vn_set)
         {
             live_range* range = live_ranges[iter];
@@ -96,7 +96,7 @@ bool memory_coloring_impl::allocate(interval_ptr interval)
     }
 
     std::size_t offset = 0;
-    while(!conflict_queue.empty())
+    while(not conflict_queue.empty())
     {
         live_range* range       = conflict_queue.top();
         std::size_t iter_offset = range->offset;
@@ -149,7 +149,7 @@ void memory_coloring_impl::build()
         {
             def_interval = instr2_live[p_iter];
             bool is_lit  = is_literal(iter);
-            if(is_allocate(iter) || is_lit)
+            if(is_allocate(iter) or is_lit)
             {
                 live_range& range        = def_interval->segment;
                 def_interval->result     = iter->get_shape();
@@ -157,12 +157,12 @@ void memory_coloring_impl::build()
                 range.begin              = cur_points;
                 def_interval->def_point  = cur_points;
                 range.size               = (iter->get_shape()).bytes();
-                if(!is_lit || unify_literals)
+                if(not is_lit or unify_literals)
                     alloc_queue.push(def_interval);
                 live_set.erase(range.vn);
             }
         }
-        else if(!is_param(iter) && !is_outline(iter) && !is_check_context(iter))
+        else if(not is_param(iter) && not is_outline(iter) && not is_check_context(iter))
         {
             is_dead = true;
         }
@@ -179,7 +179,7 @@ void memory_coloring_impl::build()
             if(not p_mod->has_instruction(arg))
                 continue;
 
-            if(is_param(arg) || is_outline(arg))
+            if(is_param(arg) or is_outline(arg))
             {
                 if(is_output_param(arg))
                     is_dead = false;
@@ -235,7 +235,7 @@ void memory_coloring_impl::rewrite()
             if(interval->get_begin() == invalid_offset)
                 continue;
 
-            if(!unify_literals && interval->is_literal)
+            if(not unify_literals && interval->is_literal)
                 continue;
 
             std::size_t offset = 0;
@@ -267,12 +267,12 @@ void memory_coloring_impl::verify()
     {
         for(int i = 0; i < num_of_lives; ++i)
         {
-            live_interval& interval = live_intervals[i];
-            live_range& segment     = interval.segment;
+            const live_interval& interval = live_intervals[i];
+            const live_range& segment     = interval.segment;
 
             if(segment.begin == invalid_offset)
             {
-                // if(!interval.is_live_on_entry)
+                // if(not interval.is_live_on_entry)
                 // MIGRAPHX_THROW("interval is not live on entry");
                 continue;
             }
@@ -284,13 +284,13 @@ void memory_coloring_impl::verify()
             int vn = segment.vn;
             if(conflict_table.find(vn) != conflict_table.end())
             {
-                std::set<int>& vn_set = conflict_table[vn];
+                const std::set<int>& vn_set = conflict_table[vn];
                 for(const auto& iter : vn_set)
                 {
                     live_range* range = live_ranges[iter];
                     if(range->offset == invalid_offset)
                         continue;
-                    if(!is_disjoin(*range, segment))
+                    if(not is_disjoin(*range, segment))
                         MIGRAPHX_THROW("range and segment is not disjoined");
                 }
             }
@@ -319,8 +319,8 @@ void memory_coloring_impl::dump_intervals()
         {
             std::cout << " segment:" << i;
             std::cout << " =>";
-            std::set<int>& table = conflict_table[i];
-            for(auto& iter : table)
+            const std::set<int>& table = conflict_table[i];
+            for(const auto& iter : table)
             {
                 std::cout << (iter) << ",";
             }
@@ -357,7 +357,7 @@ void live_interval::dump()
     std::cout << "id:" << id;
     segment.dump();
     std::cout << " uses:";
-    for(auto& iter : use_points)
+    for(const auto& iter : use_points)
     {
         std::cout << " " << get_ins_enum(iter) << ",";
     }

src/opt/memory_coloring_impl.hpp

@@ -125,11 +125,11 @@ struct memory_coloring_impl
 
     static bool is_disjoin(const live_range& range1, const live_range& range2)
     {
-        if((range1.size == 0) || (range2.size == 0))
+        if((range1.size == 0) or (range2.size == 0))
             return false;
         auto end1 = range1.offset + range1.size - 1;
         auto end2 = range2.offset + range2.size - 1;
-        return ((end1 < range2.offset) || (end2 < range1.offset));
+        return ((end1 < range2.offset) or (end2 < range1.offset));
     }
     void verify();
 #ifdef MIGRAPHX_DEBUG_OPT

src/pad_calc.cpp

@@ -60,7 +60,7 @@ std::vector<std::size_t> calc_dyn_auto_pad(std::vector<std::size_t> tensor_lens,
 {
     std::vector<std::size_t> padding;
     padding.resize(2 * k_lens.size());
-    for(size_t i = 0; i < padding.size() / 2; i++)
+    for(std::size_t i = 0; i < padding.size() / 2; i++)
     {
         std::ptrdiff_t input_dim      = tensor_lens[i];
         std::ptrdiff_t stride         = strides[i];

src/process.cpp

@@ -50,7 +50,7 @@ int exec(const std::string& cmd, const std::function<void(const char*)>& std_out
     {
         // TODO: Use execve instead of popen
         std::unique_ptr<FILE, decltype(closer)> pipe(popen(cmd.c_str(), "r"), closer); // NOLINT
-        if(!pipe)
+        if(not pipe)
             MIGRAPHX_THROW("popen() failed: " + cmd);
         std::array<char, 128> buffer;
         while(fgets(buffer.data(), buffer.size(), pipe.get()) != nullptr)

src/program.cpp

@@ -37,6 +37,7 @@
 #include <migraphx/output_iterator.hpp>
 #include <migraphx/make_op.hpp>
 #include <migraphx/marker.hpp>
+#include <migraphx/supported_segments.hpp>
 #include <iostream>
 #include <sstream>
 #include <algorithm>
@@ -77,11 +78,11 @@ program& program::operator=(program p)
 
 void program::assign(const program& p)
 {
-    if(!impl)
+    if(not impl)
     {
         impl = std::make_unique<program_impl>();
     }
-    else if(!impl->modules.empty())
+    else if(not impl->modules.empty())
     {
         impl->modules.clear();
     }
@@ -167,13 +168,37 @@ target_assignments program::get_target_assignments(const std::vector<target>& ta
     target_assignments p;
 
     const auto* mod = get_main_module();
-    for(auto it : iterator_for(*mod))
+    std::vector<std::pair<target, supported_segments>> target_subgraphs;
+    target_subgraphs.reserve(targets.size());
+    std::transform(targets.begin(),
+                   targets.end(),
+                   std::back_inserter(target_subgraphs),
+                   [&](const auto& t) { return std::make_pair(t, t.find_supported(mod, m)); });
+
+    for(const auto ins : iterator_for(*mod))
     {
-        auto t = std::max_element(
-            targets.begin(), targets.end(), [it, m](const target& lhs, const target& rhs) {
-                return lhs.is_supported(it, m) < rhs.is_supported(it, m);
-            });
-        p.add_assignment(it, t->name());
+        if(contains(p, ins))
+        {
+            continue;
+        }
+
+        for(const auto& [target, subgraph] : target_subgraphs)
+        {
+            // can't pass a structured binding into lambda in C++17 so create a variable for it
+            const auto& t = target;
+            for(const auto& segment : subgraph)
+            {
+                const auto& instructions = segment.instructions;
+                if(not contains(instructions, ins))
+                {
+                    continue;
+                }
+                std::transform(instructions.begin(),
+                               instructions.end(),
+                               std::inserter(p, p.end()),
+                               [&](auto instr) { return std::make_pair(instr, t.name()); });
+            }
+        }
     }
     return p;
 }

src/quantization.cpp

@@ -70,7 +70,7 @@ void quantize_int8(program& prog,
 {
     std::set<std::string> op_names = {"convolution", "dot"};
     std::set<std::string> input_ins_names(ins_names.begin(), ins_names.end());
-    if(!std::includes(
+    if(not std::includes(
            op_names.begin(), op_names.end(), input_ins_names.begin(), input_ins_names.end()))
     {
         MIGRAPHX_THROW("QUANTIZE_INT8: only support DOT and CONVOLUTION operation");

src/replace_allocate.cpp

@@ -73,7 +73,7 @@ void insert_submod_allocations(instruction_ref ins, module& mod, const allocatio
         name_shapes.insert(ps.begin(), ps.end());
     }
 
-    for(auto& pn : name_shapes)
+    for(const auto& pn : name_shapes)
     {
         const auto& s = pn.second;
         instruction_ref output{};

src/rewrite_pooling.cpp

@@ -47,12 +47,12 @@ void rewrite_pooling::apply(module& m) const
         if(not s.standard())
             continue;
         auto&& op = any_cast<op::pooling>(ins->get_operator());
-        if(!std::all_of(op.padding.begin(), op.padding.end(), [](auto i) { return i == 0; }))
+        if(not std::all_of(op.padding.begin(), op.padding.end(), [](auto i) { return i == 0; }))
             continue;
-        if(!std::all_of(op.stride.begin(), op.stride.end(), [](auto i) { return i == 1; }))
+        if(not std::all_of(op.stride.begin(), op.stride.end(), [](auto i) { return i == 1; }))
             continue;
         auto lens = s.lens();
-        if(!std::equal(lens.begin() + 2, lens.end(), op.lengths.begin(), op.lengths.end()))
+        if(not std::equal(lens.begin() + 2, lens.end(), op.lengths.begin(), op.lengths.end()))
             continue;
         std::int64_t n = s.lens()[0];
         std::int64_t c = s.lens()[1];