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

src/include/migraphx/rewrite_rnn.hpp

@@ -19,22 +19,22 @@ struct module;
 struct rewrite_rnn
 {
     std::string name() const { return "rewrite_rnn"; }
-    void apply(module& prog) const;
+    void apply(module& m) const;
 
     private:
     // for vanilla rnn operators
-    void apply_vanilla_rnn(module& prog, instruction_ref ins) const;
+    void apply_vanilla_rnn(module& m, instruction_ref ins) const;
     std::vector<instruction_ref> vanilla_rnn_cell(bool is_forward,
-                                                  module& prog,
+                                                  module& m,
                                                   instruction_ref ins,
                                                   std::vector<instruction_ref> inputs,
                                                   operation& actv_func) const;
     std::vector<operation> vanilla_rnn_actv_funcs(instruction_ref ins) const;
 
     // for gru operators
-    void apply_gru(module& prog, instruction_ref ins) const;
+    void apply_gru(module& m, instruction_ref ins) const;
     std::vector<instruction_ref> gru_cell(bool is_forward,
-                                          module& prog,
+                                          module& m,
                                           instruction_ref ins,
                                           std::vector<instruction_ref> inputs,
                                           int linear_before_reset,
@@ -44,9 +44,9 @@ struct rewrite_rnn
     std::vector<operation> gru_actv_funcs(instruction_ref ins) const;
 
     // for lstm operators
-    void apply_lstm(module& prog, instruction_ref ins) const;
+    void apply_lstm(module& m, instruction_ref ins) const;
     std::vector<instruction_ref> lstm_cell(bool is_forward,
-                                           module& prog,
+                                           module& m,
                                            instruction_ref ins,
                                            std::vector<instruction_ref> inputs,
                                            const operation& actv_func1,
@@ -55,24 +55,23 @@ struct rewrite_rnn
 
     std::vector<operation> lstm_actv_funcs(instruction_ref ins) const;
 
-    bool is_variable_seq_lens(const module& prog, instruction_ref seq_lens) const;
-    instruction_ref replace_last_hs_output(module& prog,
+    bool is_variable_seq_lens(const module& m, instruction_ref seq_lens) const;
+    instruction_ref replace_last_hs_output(module& m,
                                            instruction_ref ins,
                                            instruction_ref seq_lens,
                                            instruction_ref last_hs_output,
                                            op::rnn_direction dirct) const;
 
-    void replace_last_cell_output(module& prog,
+    void replace_last_cell_output(module& m,
                                   instruction_ref ins,
                                   instruction_ref seq_lens,
                                   instruction_ref cell_outputs,
                                   instruction_ref last_cell_output,
                                   op::rnn_direction dirct) const;
 
-    std::size_t
-    get_seq_len(const module& prog, instruction_ref input, instruction_ref seq_lens) const;
+    std::size_t get_seq_len(const module& m, instruction_ref input, instruction_ref seq_lens) const;
 
-    instruction_ref pad_hidden_states(module& prog,
+    instruction_ref pad_hidden_states(module& m,
                                       instruction_ref seq,
                                       instruction_ref seq_lens,
                                       instruction_ref hs) const;

src/include/migraphx/schedule.hpp

@@ -19,7 +19,7 @@ struct schedule
     schedule_model model{};
     bool enable = true;
     std::string name() const { return "schedule"; }
-    void apply(module& p) const;
+    void apply(module& m) const;
 };
 
 } // namespace MIGRAPHX_INLINE_NS

src/include/migraphx/simplify_algebra.hpp

@@ -15,7 +15,7 @@ struct module;
 struct simplify_algebra
 {
     std::string name() const { return "simplify_algebra"; }
-    void apply(module& p) const;
+    void apply(module& m) const;
 };
 
 } // namespace MIGRAPHX_INLINE_NS

src/include/migraphx/simplify_reshapes.hpp

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

src/make_op.cpp

@@ -5,20 +5,41 @@ namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 
 operation make_op(const std::string& name) { return load_op(name); }
-operation make_op(const std::string& name, const value& v)
+
+template <class F>
+operation make_op_generic(const std::string& name, F for_each)
 {
-    if(not(v.is_object() or (v.empty() and v.is_array())))
-        MIGRAPHX_THROW("Value is not an object");
     auto op = load_op(name);
     // Merge values
     value w = op.to_value();
-    for(auto&& x : v)
-    {
-        w.at(x.get_key()) = x.without_key();
-    }
+    for_each([&](const auto& key, const auto& x) {
+        if(not w.contains(key))
+            // NOLINTNEXTLINE(performance-inefficient-string-concatenation)
+            MIGRAPHX_THROW("No key '" + key + "' in " + name);
+        w.at(key) = x;
+    });
     op.from_value(w);
     return op;
 }
 
+operation make_op(const std::string& name,
+                  const std::initializer_list<std::pair<std::string, value>>& v)
+{
+    return make_op_generic(name, [&](auto f) {
+        for(auto&& [key, x] : v)
+            f(key, x);
+    });
+}
+
+operation make_op_from_value(const std::string& name, const value& v)
+{
+    if(not(v.is_object() or (v.empty() and v.is_array())))
+        MIGRAPHX_THROW("Value is not an object for make_op: " + name);
+    return make_op_generic(name, [&](auto f) {
+        for(auto&& x : v)
+            f(x.get_key(), x.without_key());
+    });
+}
+
 } // namespace MIGRAPHX_INLINE_NS
 } // namespace migraphx

src/onnx/parse_mean.cpp

@@ -24,14 +24,17 @@ struct parse_mean : op_parser<parse_mean>
         auto divisor = info.add_literal(
             migraphx::literal{migraphx::shape{args[0]->get_shape().type()}, {num_data}});
 
-        return std::accumulate(args.begin(), args.end(), args[0], [&](auto& mean, auto& data_i) {
-            // Pre-divide each tensor element-wise by n to reduce risk of overflow during summation
-            data_i = info.add_broadcastable_binary_op("div", data_i, divisor);
-
-            if(data_i != args[0])
-                return info.add_broadcastable_binary_op("add", mean, data_i);
-            return data_i;
-        });
+        // 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);
+                               });
     }
 };
 

src/opt/memory_coloring.cpp

@@ -4,11 +4,11 @@
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 
-void memory_coloring::apply(module& p) const
+void memory_coloring::apply(module& m) const
 {
     if(!enabled(MIGRAPHX_DISABLE_MEMORY_COLORING{}))
     {
-        memory_coloring_impl opt(&p, allocation_op, verify);
+        memory_coloring_impl opt(&m, allocation_op, verify);
         opt.run();
     }
 }

src/propagate_constant.cpp

@@ -20,9 +20,9 @@ bool skip_propogate(instruction_ref ins)
     return false;
 }
 
-void propagate_constant::apply(module& p) const
+void propagate_constant::apply(module& m) const
 {
-    for(auto i : iterator_for(p))
+    for(auto i : iterator_for(m))
     {
         if(i->name() != "@literal")
             continue;
@@ -42,8 +42,8 @@ void propagate_constant::apply(module& p) const
                 if(not r.empty())
                 {
                     assert(r.get_shape() == child->get_shape());
-                    auto l = p.add_literal(r.get_shape(), r.data());
-                    self(p.replace_instruction(child, l));
+                    auto l = m.add_literal(r.get_shape(), r.data());
+                    self(m.replace_instruction(child, l));
                 }
             }
         })(i);

src/py/migraphx_py.cpp

@@ -273,6 +273,14 @@ MIGRAPHX_PYBIND11_MODULE(migraphx, m)
             py::arg("op"),
             py::arg("args"),
             py::arg("mod_args") = std::vector<migraphx::module*>{})
+        .def(
+            "add_literal",
+            [](migraphx::module& mm, py::buffer data) {
+                py::buffer_info info = data.request();
+                auto literal_shape   = to_shape(info);
+                return mm.add_literal(literal_shape, reinterpret_cast<char*>(info.ptr));
+            },
+            py::arg("data"))
         .def(
             "add_parameter",
             [](migraphx::module& mm, const std::string& name, const migraphx::shape shape) {

src/rewrite_batchnorm.cpp

@@ -14,9 +14,9 @@
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 
-void rewrite_batchnorm::apply(module& p) const
+void rewrite_batchnorm::apply(module& m) const
 {
-    for(auto ins : iterator_for(p))
+    for(auto ins : iterator_for(m))
     {
         if(ins->name() != "batch_norm_inference")
             continue;
@@ -46,13 +46,13 @@ void rewrite_batchnorm::apply(module& p) const
             });
 
         auto broadcast   = op::broadcast{1, ins->get_shape().lens()};
-        auto a_ins       = p.add_literal({a.get_shape(), a.data()});
-        auto a_broadcast = p.insert_instruction(ins, broadcast, a_ins);
-        auto mul   = p.insert_instruction(ins, make_op("mul"), ins->inputs().front(), a_broadcast);
-        auto b_ins = p.add_literal({b.get_shape(), b.data()});
-        auto b_broadcast = p.insert_instruction(ins, broadcast, b_ins);
-        auto add         = p.insert_instruction(ins, make_op("add"), mul, b_broadcast);
-        p.replace_instruction(ins, add);
+        auto a_ins       = m.add_literal({a.get_shape(), a.data()});
+        auto a_broadcast = m.insert_instruction(ins, broadcast, a_ins);
+        auto mul   = m.insert_instruction(ins, make_op("mul"), ins->inputs().front(), a_broadcast);
+        auto b_ins = m.add_literal({b.get_shape(), b.data()});
+        auto b_broadcast = m.insert_instruction(ins, broadcast, b_ins);
+        auto add         = m.insert_instruction(ins, make_op("add"), mul, b_broadcast);
+        m.replace_instruction(ins, add);
     }
 }
 

src/rewrite_pooling.cpp

@@ -12,9 +12,9 @@
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 
-void rewrite_pooling::apply(module& prog) const
+void rewrite_pooling::apply(module& m) const
 {
-    for(auto ins : iterator_for(prog))
+    for(auto ins : iterator_for(m))
     {
         if(ins->name() != "pooling")
             continue;
@@ -33,26 +33,25 @@ void rewrite_pooling::apply(module& prog) const
             continue;
         std::int64_t n = s.lens()[0];
         std::int64_t c = s.lens()[1];
-        auto reshape   = prog.insert_instruction(
+        auto reshape   = m.insert_instruction(
             ins, make_op("reshape", {{"dims", {n * c, -1}}}), ins->inputs().front());
         instruction_ref pooling{};
 
         // average pooling
         if(op.mode == op::pooling_mode::average)
         {
-            pooling =
-                prog.insert_instruction(ins, make_op("reduce_mean", {{"axes", {1}}}), reshape);
+            pooling = m.insert_instruction(ins, make_op("reduce_mean", {{"axes", {1}}}), reshape);
         }
         // max pooling
         else
         {
-            pooling = prog.insert_instruction(ins, make_op("reduce_max", {{"axes", {1}}}), reshape);
+            pooling = m.insert_instruction(ins, make_op("reduce_max", {{"axes", {1}}}), reshape);
         }
 
         std::vector<int64_t> rsp_lens(lens.size(), 1);
         rsp_lens[0] = n;
         rsp_lens[1] = c;
-        prog.replace_instruction(ins, make_op("reshape", {{"dims", rsp_lens}}), pooling);
+        m.replace_instruction(ins, make_op("reshape", {{"dims", rsp_lens}}), pooling);
     }
 }
 

src/schedule.cpp

@@ -42,7 +42,7 @@ struct stream_info
     std::unordered_map<instruction_ref, std::size_t> iweights;
     ins_dep_map mod_implicit_deps;
 
-    void calc_implicit_deps(const module& p) { mod_implicit_deps = p.calc_implicit_deps(); }
+    void calc_implicit_deps(const module& m) { mod_implicit_deps = m.calc_implicit_deps(); }
 
     void accumulate_weights(instruction_ref last, const schedule_model& model)
     {
@@ -116,15 +116,15 @@ struct stream_info
         }
     };
 
-    std::size_t assign_streams(module& p, std::size_t n)
+    std::size_t assign_streams(module& m, std::size_t n)
     {
         assert(n > 0);
         partition critical;
         std::unordered_map<instruction_ref, std::deque<partition>> partitions;
         partitions.reserve(weights.size());
         fix([&](auto self, auto ins, auto& part) {
-            assert(not is_end(ins, p.end()));
-            if(not p.has_instruction(ins))
+            assert(not is_end(ins, m.end()));
+            if(not m.has_instruction(ins))
                 return;
             if(contains(partitions, ins))
                 return;
@@ -151,8 +151,8 @@ struct stream_info
                 }
             }
             // Sort instructions
-            p.move_instruction(ins, p.end());
-        })(std::prev(p.end()), critical);
+            m.move_instruction(ins, m.end());
+        })(std::prev(m.end()), critical);
 
         // Set the critical partition to stream 0
         set_stream(critical, 0);
@@ -197,13 +197,13 @@ struct stream_info
         }
     };
 
-    void sort(module& p, std::size_t)
+    void sort(module& m, std::size_t)
     {
         std::set<weight_ins, compare_weight_ins> children;
         std::unordered_map<instruction_ref, std::size_t> visited;
-        auto last      = std::prev(p.end());
+        auto last      = std::prev(m.end());
         auto mw        = this->weights.at(last);
-        auto nw        = mw / (p.size() + 1);
+        auto nw        = mw / (m.size() + 1);
         auto add_child = [&](auto ins) {
             auto x  = 1 + (mw - this->weights.at(ins)) / (nw + 1);
             auto w  = x * this->iweights.at(ins);
@@ -222,10 +222,10 @@ struct stream_info
             // Pop the first element
             auto top = children.begin()->second;
             children.erase(children.begin());
-            p.move_instruction(top, p.begin());
+            m.move_instruction(top, m.begin());
             for(auto ins : top->inputs())
             {
-                if(not p.has_instruction(ins))
+                if(not m.has_instruction(ins))
                     continue;
                 add_child(ins);
             }
@@ -234,7 +234,7 @@ struct stream_info
             {
                 for(auto ins : mod_implicit_deps.at(top))
                 {
-                    assert(p.has_instruction(ins));
+                    assert(m.has_instruction(ins));
                     add_child(ins);
                 }
             }
@@ -242,12 +242,12 @@ struct stream_info
 
         // move dangling parameter to the front so as not be removed
         auto ins = std::next(last);
-        while(ins != p.end())
+        while(ins != m.end())
         {
             auto next = std::next(ins);
             if(ins->name() == "@param")
             {
-                p.move_instruction(ins, p.begin());
+                m.move_instruction(ins, m.begin());
             }
             ins = next;
         }
@@ -364,18 +364,18 @@ struct stream_info
     }
 
     std::unordered_map<instruction_ref, std::vector<std::vector<instruction_ref>>>
-    find_concurrent_instructions(module& p) const
+    find_concurrent_instructions(module& m) const
     {
         std::unordered_map<instruction_ref, std::vector<std::vector<instruction_ref>>> result;
         std::unordered_map<instruction_ref, std::unordered_set<instruction_ref>> merge_from;
-        dominator_info di = compute_dominator(p);
-        result.reserve(p.size());
-        merge_from.reserve(p.size());
-        for(auto ins : reverse_iterator_for(p))
+        dominator_info di = compute_dominator(m);
+        result.reserve(m.size());
+        merge_from.reserve(m.size());
+        for(auto ins : reverse_iterator_for(m))
         {
             for(auto&& arg : ins->outputs())
             {
-                if(not p.has_instruction(arg))
+                if(not m.has_instruction(arg))
                     continue;
                 if(is_merge_point(arg))
                     merge_from[ins].insert(arg);
@@ -415,18 +415,18 @@ struct stream_info
     }
 
     std::unordered_map<instruction_ref, std::unordered_set<instruction_ref>>
-    get_conflicts(module& p)
+    get_conflicts(module& m)
     {
 
         using conflict_table_type =
             std::unordered_map<instruction_ref, std::unordered_set<instruction_ref>>;
         conflict_table_type conflict_table;
-        auto concur_ins = this->find_concurrent_instructions(p);
+        auto concur_ins = this->find_concurrent_instructions(m);
 
         // Compute an index for each instruction
         std::unordered_map<instruction_ref, std::size_t> ins2index;
         std::size_t index_total = 0;
-        for(auto ins : iterator_for(p))
+        for(auto ins : iterator_for(m))
             ins2index[ins] = index_total++;
 
         std::vector<conflict_table_type> thread_conflict_tables(
@@ -507,21 +507,21 @@ struct stream_info
     }
 };
 
-void schedule::apply(module& p) const
+void schedule::apply(module& m) const
 {
     if(not enable)
         return;
 
     stream_info si;
-    si.calc_implicit_deps(p);
-    auto last = std::prev(p.end());
+    si.calc_implicit_deps(m);
+    auto last = std::prev(m.end());
     si.accumulate_weights(last, model);
-    auto nstreams = si.assign_streams(p, model.concurrency());
-    si.sort(p, model.concurrency());
+    auto nstreams = si.assign_streams(m, model.concurrency());
+    si.sort(m, model.concurrency());
 
     if(enabled(MIGRAPHX_TRACE_COMPILE{}) or enabled(MIGRAPHX_TRACE_SCHEDULE{}))
     {
-        p.annotate(std::cout, [&](auto ins) {
+        m.annotate(std::cout, [&](auto ins) {
             if(ins->name() == "@param" and not contains(si.weights, ins))
                 return;
 
@@ -548,9 +548,9 @@ void schedule::apply(module& p) const
     std::unordered_map<instruction_ref, std::size_t> ins2wait;
     std::unordered_map<std::size_t, std::unordered_set<std::size_t>> waited_for;
     std::unordered_map<instruction_ref, std::unordered_set<std::size_t>> ins2waited;
-    ins2wait.reserve(p.size());
-    ins2waited.reserve(p.size());
-    for(auto ins : iterator_for(p))
+    ins2wait.reserve(m.size());
+    ins2waited.reserve(m.size());
+    for(auto ins : iterator_for(m))
     {
         // Only schedule instructions that have a stream
         if(not si.has_stream(ins))
@@ -559,7 +559,7 @@ void schedule::apply(module& p) const
         // Schedule instruction on the stream
         auto stream = si.get_stream(ins);
         assert(stream < model.concurrency());
-        model.sched(p, ins, stream);
+        model.sched(m, ins, stream);
         // Insert wait instructions
         if(si.is_merge_point(ins, stream))
         {
@@ -572,14 +572,14 @@ void schedule::apply(module& p) const
                 if(not contains(ins2wait, i))
                 {
                     ins2wait[i] = wait_id;
-                    model.record(p, i, wait_id);
+                    model.record(m, i, wait_id);
                     wait_id++;
                 }
                 auto w = ins2wait.at(i);
                 // If we already waited for the event on this stream then dont
                 // insert another wait event
                 if(not contains(waited_for[stream], w))
-                    model.wait(p, ins, w);
+                    model.wait(m, ins, w);
                 // Store the event as waited
                 waited_for[stream].insert(w);
                 // Store all wait events that have been waited on prior to the recorded instruction
@@ -594,7 +594,7 @@ void schedule::apply(module& p) const
     }
 
     // Add memory conflicts
-    auto conflict_table = si.get_conflicts(p);
+    auto conflict_table = si.get_conflicts(m);
     for(auto&& ip : conflict_table)
     {
         if(ip.second.empty())
@@ -602,7 +602,7 @@ void schedule::apply(module& p) const
         std::vector<instruction_ref> args;
         args.push_back(ip.first);
         args.insert(args.end(), ip.second.begin(), ip.second.end());
-        p.insert_instruction(std::next(ip.first), make_op("identity"), args);
+        m.insert_instruction(std::next(ip.first), make_op("identity"), args);
     }
 }
 

src/simplify_qdq.cpp

@@ -53,7 +53,7 @@ struct match_find_quantizable_ops
             match::arg(1)(dequantizelinear_op("x2", "scale2")));
     }
 
-    void apply(module& m, match::matcher_result r) const
+    void apply(module& m, const match::matcher_result& r) const
     {
         auto qop    = r.result;
         auto q1     = r.instructions["x1"];

src/simplify_reshapes.cpp

@@ -70,19 +70,19 @@ struct find_reshaper
             match::any_of[match::outputs()](match::name(reshaper_names())));
     }
 
-    void apply(module& p, const match::matcher_result& mr) const
+    void apply(module& m, const match::matcher_result& mr) const
     {
         auto ins = mr.result;
         std::vector<instruction_ref> reshapes{ins};
         while(is_reshaper(reshapes.back()))
         {
             assert(!reshapes.back()->inputs().empty());
-            assert(p.has_instruction(reshapes.back()->inputs().front()));
+            assert(m.has_instruction(reshapes.back()->inputs().front()));
             auto input = reshapes.back()->inputs().front();
             reshapes.push_back(input);
         }
 
-        std::pair<instruction_ref, instruction_ref> r{p.end(), p.end()};
+        std::pair<instruction_ref, instruction_ref> r{m.end(), m.end()};
         for(auto start : iterator_for(reshapes))
         {
             auto last = std::find_if(reshapes.rbegin(), reshapes.rend(), [&](auto&& i) {
@@ -96,7 +96,7 @@ struct find_reshaper
         }
         if(r.first != r.second)
         {
-            p.replace_instruction(r.first, r.second);
+            m.replace_instruction(r.first, r.second);
         }
     }
 };
@@ -117,10 +117,10 @@ struct find_nop_reshapes
         return match::name(reshapes)(match::same_shape(match::arg(0)));
     }
 
-    void apply(module& p, const match::matcher_result& mr) const
+    void apply(module& m, const match::matcher_result& mr) const
     {
         auto ins = mr.result;
-        p.replace_instruction(ins, ins->inputs().front());
+        m.replace_instruction(ins, ins->inputs().front());
     }
 };
 
@@ -132,7 +132,7 @@ struct find_transpose
             match::skip_output(match::name("contiguous"))(match::name("transpose"))));
     }
 
-    void apply(module& p, const match::matcher_result& mr) const
+    void apply(module& m, const match::matcher_result& mr) const
     {
         auto ins = mr.result;
         auto x   = ins;
@@ -149,11 +149,11 @@ struct find_transpose
             return;
         if(is_no_transpose(dims))
         {
-            p.replace_instruction(ins, t->inputs().front());
+            m.replace_instruction(ins, t->inputs().front());
         }
         else
         {
-            p.replace_instruction(
+            m.replace_instruction(
                 ins, make_op("transpose", {{"permutation", dims}}), t->inputs().front());
         }
     }
@@ -223,7 +223,7 @@ struct find_nested_slice
         return result;
     }
 
-    void apply(module& p, const match::matcher_result& mr) const
+    void apply(module& m, const match::matcher_result& mr) const
     {
         auto ins   = mr.result;
         auto slice = ins->inputs().front();
@@ -241,7 +241,7 @@ struct find_nested_slice
             op.starts.push_back(pp.second.first);
             op.ends.push_back(pp.second.second);
         }
-        p.replace_instruction(ins, op, input);
+        m.replace_instruction(ins, op, input);
     }
 };
 
@@ -252,7 +252,7 @@ struct find_concat_transpose
         return match::name("concat")(match::all_of[match::inputs()](match::transpose_shape()));
     }
 
-    void apply(module& p, const match::matcher_result& mr) const
+    void apply(module& m, const match::matcher_result& mr) const
     {
         auto ins          = mr.result;
         auto trans_inputs = ins->inputs();
@@ -279,14 +279,14 @@ struct find_concat_transpose
         std::vector<instruction_ref> inputs;
         std::transform(
             ins->inputs().begin(), ins->inputs().end(), std::back_inserter(inputs), [&](auto i) {
-                return p.insert_instruction(
+                return m.insert_instruction(
                     ins, make_op("transpose", {{"permutation", permutation}}), i);
             });
-        auto concat = p.insert_instruction(ins, op, inputs);
-        auto t      = p.insert_instruction(
+        auto concat = m.insert_instruction(ins, op, inputs);
+        auto t      = m.insert_instruction(
             ins, make_op("transpose", {{"permutation", ipermutation}}), concat);
         assert(ins->get_shape().lens() == t->get_shape().lens());
-        p.replace_instruction(ins, t);
+        m.replace_instruction(ins, t);
     }
 };
 
@@ -303,7 +303,7 @@ struct find_nested_concat
         return op.axis;
     }
 
-    void apply(module& p, const match::matcher_result& mr) const
+    void apply(module& m, const match::matcher_result& mr) const
     {
         auto ins  = mr.result;
         auto axis = get_axis(ins);
@@ -317,7 +317,7 @@ struct find_nested_concat
                     args.push_back(i);
             }
         })(ins->inputs());
-        p.replace_instruction(ins, ins->get_operator(), args);
+        m.replace_instruction(ins, ins->get_operator(), args);
     }
 };
 
@@ -329,7 +329,7 @@ struct find_resize
             match::args(match::name("reshape").bind("data"), match::is_constant().bind("ind")));
     }
 
-    void apply(module& p, match::matcher_result r) const
+    void apply(module& m, const match::matcher_result& r) const
     {
         auto ins     = r.result;
         auto ins_rsp = r.instructions["data"];
@@ -417,13 +417,13 @@ struct find_resize
         }
 
         auto in_rsp   = ins_rsp->inputs().front();
-        auto rsp_data = p.insert_instruction(
+        auto rsp_data = m.insert_instruction(
             ins_rsp, migraphx::make_op("reshape", {{"dims", in_dims}}), in_rsp);
-        auto mb_rsp = p.insert_instruction(
+        auto mb_rsp = m.insert_instruction(
             ins_rsp, migraphx::make_op("multibroadcast", {{"out_lens", out_dims}}), rsp_data);
-        auto std_mb = p.insert_instruction(ins, migraphx::make_op("contiguous"), mb_rsp);
+        auto std_mb = m.insert_instruction(ins, migraphx::make_op("contiguous"), mb_rsp);
         std::vector<int64_t> rsp_dims(out_lens.begin(), out_lens.end());
-        p.replace_instruction(ins, migraphx::make_op("reshape", {{"dims", rsp_dims}}), std_mb);
+        m.replace_instruction(ins, migraphx::make_op("reshape", {{"dims", rsp_dims}}), std_mb);
     }
 };
 
@@ -436,7 +436,7 @@ struct find_where_op
                         match::is_constant().bind("ind")));
     }
 
-    void apply(module& p, match::matcher_result r) const
+    void apply(module& m, const match::matcher_result& r) const
     {
         auto ins     = r.result;
         auto concat  = r.instructions["data"];
@@ -475,11 +475,11 @@ struct find_where_op
 
         if(val)
         {
-            p.replace_instruction(ins, inputs.at(0));
+            m.replace_instruction(ins, inputs.at(0));
         }
         else
         {
-            p.replace_instruction(ins, inputs.at(1));
+            m.replace_instruction(ins, inputs.at(1));
         }
     }
 };
@@ -496,7 +496,7 @@ struct find_reshape_cont
                 match::any()));
     }
 
-    void apply(module& p, match::matcher_result r) const
+    void apply(module& m, const match::matcher_result& r) const
     {
         auto ins      = r.result;
         auto ins_cont = r.instructions["cont"];
@@ -530,11 +530,11 @@ struct find_reshape_cont
             else
             {
                 inputs.push_back(
-                    p.insert_instruction(ins, make_op("reshape", {{"dims", dims}}), in));
+                    m.insert_instruction(ins, make_op("reshape", {{"dims", dims}}), in));
             }
         }
-        auto out = p.insert_instruction(ins, ins->get_operator(), inputs);
-        p.replace_instruction(ins, make_op("reshape", {{"dims", out_dims}}), out);
+        auto out = m.insert_instruction(ins, ins->get_operator(), inputs);
+        m.replace_instruction(ins, make_op("reshape", {{"dims", out_dims}}), out);
     }
 };
 
@@ -564,25 +564,25 @@ struct find_transpose_contiguous_reshaper_unary
                          match::args(match_transpose_contiguous_reshaper()));
     }
 
-    void apply(module& p, match::matcher_result r) const
+    void apply(module& m, const match::matcher_result& r) const
     {
         auto ins           = r.result;
         auto reshaper_ins  = r.instructions["reshaper_ins"];
         auto trans_ins     = r.instructions["trans_ins"];
         auto cont_ins      = r.instructions["cont_ins"];
         auto unary_op_name = ins->get_operator().name();
-        auto unary_ins     = p.insert_instruction(cont_ins, make_op(unary_op_name), trans_ins);
-        auto new_cont_ins  = p.insert_instruction(cont_ins, make_op("contiguous"), unary_ins);
+        auto unary_ins     = m.insert_instruction(cont_ins, make_op(unary_op_name), trans_ins);
+        auto new_cont_ins  = m.insert_instruction(cont_ins, make_op("contiguous"), unary_ins);
         // older cont and reshape are removed by deadcode elimination
-        p.replace_instruction(ins, reshaper_ins->get_operator(), new_cont_ins);
+        m.replace_instruction(ins, reshaper_ins->get_operator(), new_cont_ins);
     }
 };
 
-void simplify_reshapes::apply(module& p) const
+void simplify_reshapes::apply(module& m) const
 {
     for(int i = 0; i < 2; i++)
     {
-        match::find_matches(p,
+        match::find_matches(m,
                             find_where_op{},
                             find_resize{},
                             find_reshape_cont{},
@@ -594,7 +594,7 @@ void simplify_reshapes::apply(module& p) const
                             find_nested_slice{},
                             find_nested_concat{},
                             find_transpose_contiguous_reshaper_unary{});
-        dead_code_elimination{}.apply(p);
+        dead_code_elimination{}.apply(m);
     }
 }
 

src/targets/cpu/include/migraphx/cpu/parallel.hpp

@@ -7,7 +7,16 @@
 #ifdef MIGRAPHX_DISABLE_OMP
 #include <migraphx/par_for.hpp>
 #else
+
+#ifdef __clang__
+#pragma clang diagnostic push
+#pragma clang diagnostic ignored "-Wreserved-identifier"
+#endif
 #include <omp.h>
+#ifdef __clang__
+#pragma clang diagnostic pop
+#endif
+
 #endif
 
 namespace migraphx {

src/targets/cpu/include/migraphx/cpu/pointwise.hpp

@@ -319,7 +319,7 @@ struct cpu_unary : reduce_dims_base, auto_register_op<cpu_unary<Op>>
     shape compute_shape(const std::vector<shape>& inputs) const
     {
         check_shapes{inputs, *this}.has(2);
-        auto s = inputs.at(0);
+        const auto& s = inputs.at(0);
         return {s.type(), s.lens()};
     }
     argument
@@ -357,7 +357,7 @@ struct cpu_binary : reduce_dims_base, auto_register_op<cpu_binary<Op>>
     shape compute_shape(const std::vector<shape>& inputs) const
     {
         check_shapes{inputs, *this}.has(3);
-        auto s = inputs.at(0);
+        const auto& s = inputs.at(0);
         return {s.type(), s.lens()};
     }
 

src/targets/cpu/lowering.cpp

@@ -223,7 +223,7 @@ struct cpu_unary2 : auto_register_op<cpu_unary2<Op>>
     shape compute_shape(const std::vector<shape>& inputs) const
     {
         check_shapes{inputs, *this}.has(1);
-        auto s = inputs.at(0);
+        const auto& s = inputs.at(0);
         return {s.type(), s.lens()};
     }
 
@@ -352,7 +352,7 @@ struct cpu_apply
             std::transform(bind_inputs.begin(),
                            bind_inputs.end(),
                            std::back_inserter(inputs),
-                           [&](const auto& s) { return r.instructions.at(s); });
+                           [&](const auto& s) { return r.instructions[s]; });
             inputs.push_back(this->insert_allocation(ins, ins->get_shape()));
             modl->replace_instruction(ins, op, inputs);
         });

src/targets/gpu/CMakeLists.txt

@@ -93,7 +93,7 @@ add_library(migraphx_device
 )
 add_library(compile_for_gpu INTERFACE)
 target_compile_options(compile_for_gpu INTERFACE -std=c++17 -fno-gpu-rdc -Wno-cuda-compat -Wno-unused-command-line-argument -Xclang -fallow-half-arguments-and-returns)
-target_link_libraries(compile_for_gpu INTERFACE hip::device -fno-gpu-rdc -Wno-invalid-command-line-argument -Wno-unused-command-line-argument)
+target_link_libraries(compile_for_gpu INTERFACE hip::device -fno-gpu-rdc -Wno-invalid-command-line-argument -Wno-unused-command-line-argument -Wno-option-ignored)
 check_cxx_compiler_flag("--cuda-host-only -fhip-lambda-host-device -x hip" HAS_HIP_LAMBDA_HOST_DEVICE)
 if(HAS_HIP_LAMBDA_HOST_DEVICE)
   message(STATUS "Enable -fhip-lambda-host-device")
@@ -158,6 +158,7 @@ add_library(migraphx_gpu
     nonzero.cpp
     pack_args.cpp
     pack_int8_args.cpp
+    prefuse_ops.cpp
     pad.cpp
     pooling.cpp
     quant_convolution.cpp