Merge branch 'develop' into jit-vector-reduce

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`:
 ```

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/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/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

src/eliminate_concat.cpp

@@ -13,9 +13,9 @@
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
-void eliminate_concat::apply(module& p) const
+void eliminate_concat::apply(module& m) const
 {
-    for(auto ins : iterator_for(p))
+    for(auto ins : iterator_for(m))
     {
         // Look for the concat operator
         if(ins->name() != concat_opt.name())
@@ -64,22 +64,22 @@ void eliminate_concat::apply(module& p) const
             std::sort(sorted_allocations.begin(),
                       sorted_allocations.end(),
                       [&](instruction_ref x, instruction_ref y) {
-                          return std::distance(p.begin(), x) < std::distance(p.begin(), y);
+                          return std::distance(m.begin(), x) < std::distance(m.begin(), y);
                       });
             // Move "super" allocation to the front
             auto first = sorted_allocations.front();
-            auto super = p.move_instruction(last, first);
+            auto super = m.move_instruction(last, first);
             // Replace each allocation with a load
             std::size_t offset = 0;
             for(auto alloc : allocations)
             {
                 op::load op{alloc->get_shape(), offset};
-                p.replace_instruction(alloc, op, {super});
+                m.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));
-            p.replace_instruction(ins, migraphx::make_op("identity"), args);
+            m.replace_instruction(ins, migraphx::make_op("identity"), args);
         }
     }
 }

src/eliminate_contiguous.cpp

@@ -69,9 +69,9 @@ static bool try_compute_shape(instruction_ref ins,
     return try_compute_shape(ins, inputs, mods);
 }
 
-void eliminate_contiguous::apply(module& p) const
+void eliminate_contiguous::apply(module& m) const
 {
-    for(auto ins : iterator_for(p))
+    for(auto ins : iterator_for(m))
     {
         // return instruction should have inputs with standard shape
         if(ins->name() == "@return")
@@ -96,8 +96,8 @@ void eliminate_contiguous::apply(module& p) const
                     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);
+                    auto l = m.add_literal(r.get_shape(), r.data());
+                    m.replace_instruction(arg, l);
                 }
             }
         }

src/eliminate_identity.cpp

@@ -8,21 +8,21 @@
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 
-void eliminate_identity::apply(module& p) const
+void eliminate_identity::apply(module& m) const
 {
-    auto last = std::prev(p.end());
-    for(auto ins : iterator_for(p))
+    auto last = std::prev(m.end());
+    for(auto ins : iterator_for(m))
     {
         // Skip the first instruction, since we always process the previous
         // instruction
-        if(ins == p.begin())
+        if(ins == m.begin())
             continue;
         const auto i = std::prev(ins);
 
         if(i->name() == "identity")
         {
-            p.replace_instruction(i, i->inputs().front());
-            p.move_instruction(i, p.end());
+            m.replace_instruction(i, i->inputs().front());
+            m.move_instruction(i, m.end());
         }
         if(ins == last)
         {
@@ -31,7 +31,7 @@ void eliminate_identity::apply(module& p) const
                 const instruction_ref& identity_input = ins->inputs().front();
                 if(identity_input->outputs().size() == 1)
                 {
-                    p.move_instruction(identity_input, i);
+                    m.move_instruction(identity_input, i);
                     // since this is the last instruction, removing it only
                     // requires changing "last" and calling remove below
                     last = std::prev(last);
@@ -40,7 +40,7 @@ void eliminate_identity::apply(module& p) const
             break;
         }
     }
-    p.remove_instructions(std::next(last), p.end());
+    m.remove_instructions(std::next(last), m.end());
 }
 
 } // namespace MIGRAPHX_INLINE_NS

src/include/migraphx/adjust_allocation.hpp

@@ -13,7 +13,7 @@ struct adjust_allocation
 {
     allocation_model model;
     std::string name() const { return "adjust_allocation"; }
-    void apply(module& p) const;
+    void apply(module& m) const;
 };
 
 } // namespace MIGRAPHX_INLINE_NS

src/include/migraphx/analyze_streams.hpp

@@ -16,7 +16,7 @@ struct stream_race
     instruction_ref before;
 };
 
-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);
 
 } // namespace MIGRAPHX_INLINE_NS
 } // namespace migraphx

src/include/migraphx/auto_contiguous.hpp

@@ -13,7 +13,7 @@ struct module;
 struct auto_contiguous
 {
     std::string name() const { return "auto_contiguous"; }
-    void apply(module& p) const;
+    void apply(module& m) const;
 };
 
 } // namespace MIGRAPHX_INLINE_NS

src/include/migraphx/check_context.hpp

@@ -33,7 +33,7 @@ struct check_context
     };
 
     std::string name() const { return "check_context"; }
-    void apply(module& p) const { p.insert_instruction(p.begin(), op{}); }
+    void apply(module& m) const { m.insert_instruction(m.begin(), op{}); }
 };
 
 } // namespace MIGRAPHX_INLINE_NS

src/include/migraphx/eliminate_allocation.hpp

@@ -19,7 +19,7 @@ struct eliminate_allocation
     std::string allocation_op{};
     std::size_t alignment = 32;
     std::string name() const { return "eliminate_allocation"; }
-    void apply(module& p) const;
+    void apply(module& m) const;
 };
 
 } // namespace MIGRAPHX_INLINE_NS

src/include/migraphx/eliminate_common_subexpression.hpp

@@ -16,7 +16,7 @@ struct module;
 struct eliminate_common_subexpression
 {
     std::string name() const { return "eliminate_common_subexpression"; }
-    void apply(module& p) const;
+    void apply(module& m) const;
 };
 
 } // namespace MIGRAPHX_INLINE_NS

src/include/migraphx/eliminate_concat.hpp

@@ -18,7 +18,7 @@ struct eliminate_concat
 {
     concat_optimization concat_opt;
     std::string name() const { return "eliminate_concat"; }
-    void apply(module& p) const;
+    void apply(module& m) const;
 };
 
 } // namespace MIGRAPHX_INLINE_NS

src/include/migraphx/eliminate_contiguous.hpp

@@ -17,7 +17,7 @@ struct eliminate_contiguous
 {
     std::string op_name;
     std::string name() const { return "eliminate_contiguous"; }
-    void apply(module& p) const;
+    void apply(module& m) const;
 };
 
 } // namespace MIGRAPHX_INLINE_NS

src/include/migraphx/eliminate_identity.hpp

@@ -18,7 +18,7 @@ struct module;
 struct eliminate_identity
 {
     std::string name() const { return "eliminate_identity"; }
-    void apply(module& p) const;
+    void apply(module& m) const;
 };
 
 } // namespace MIGRAPHX_INLINE_NS