Merge branch 'dyn_batch_pass' of github.com:ROCmSoftwarePlatform/AMDMIGraphX into dyn_test_runner

src/include/migraphx/op/where.hpp

 /*
  * The MIT License (MIT)
  *
- * Copyright (c) 2015-2022 Advanced Micro Devices, Inc. All rights reserved.
+ * Copyright (c) 2015-2023 Advanced Micro Devices, Inc. All rights reserved.
  *
  * Permission is hereby granted, free of charge, to any person obtaining a copy
  * of this software and associated documentation files (the "Software"), to deal
@@ -42,9 +42,17 @@ struct where
 
     shape compute_shape(std::vector<shape> inputs) const
     {
-        check_shapes{inputs, *this}.has(3).same_dims();
+        check_shapes{inputs, *this, true}.has(3).same_dims();
         auto s1 = inputs.at(1);
         auto s2 = inputs.at(2);
+        if(s1.dynamic() or s2.dynamic())
+        {
+            if(s1 == s2)
+                return s1;
+            MIGRAPHX_THROW("WHERE: dynamic input shapes must be the same");
+        }
+
+        // Compare two static shapes, returning a standard shape
         if(s1 == s2 and s1.packed())
         {
             return s1;
@@ -63,12 +71,12 @@ struct where
         }
     }
 
-    argument compute(const shape& output_shape, std::vector<argument> args) const
+    argument compute(const dyn_output& dyn_out, std::vector<argument> args) const
     {
-        argument result{output_shape};
+        argument result{dyn_out.computed_shape};
         visit_all(result, args[1], args[2])([&](auto output, const auto x, const auto y) {
             args[0].visit([&](const auto condition) {
-                par_for(output_shape.elements(),
+                par_for(dyn_out.computed_shape.elements(),
                         [&](auto i) { output[i] = condition[i] ? x[i] : y[i]; });
             });
         });

src/include/migraphx/operation.hpp

@@ -140,6 +140,8 @@ template <class T>
 auto compute_shape_op(rank<2>, const T& x, const std::vector<shape>& inputs)
     -> decltype(x.normalize_compute_shape(inputs))
 {
+    if(inputs.empty())
+        MIGRAPHX_THROW("At least one input is required for " + x.name());
     dependent_type<operation, T> y = x;
     normalize_attributes(y, inputs[0].max_lens());
     return any_cast<T>(y).normalize_compute_shape(inputs);

src/include/migraphx/optimize_module.hpp

+/*
+ * The MIT License (MIT)
+ *
+ * Copyright (c) 2015-2022 Advanced Micro Devices, Inc. All rights reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
+ * THE SOFTWARE.
+ */
+#ifndef MIGRAPHX_GUARD_RTGLIB_OPTIMIZE_MODULE_HPP
+#define MIGRAPHX_GUARD_RTGLIB_OPTIMIZE_MODULE_HPP
+
+#include <string>
+#include <migraphx/instruction_ref.hpp>
+#include <migraphx/config.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+
+struct module_pass_manager;
+
+/**
+ * Runs several passes in a loop
+ */
+struct optimize_module
+{
+    std::string name() const { return "optimize_module"; }
+    void apply(module_pass_manager& mpm) const;
+};
+
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx
+
+#endif

src/include/migraphx/program.hpp

@@ -130,6 +130,7 @@ struct program
 
     // module related api
     module* create_module(const std::string& name);
+    module* copy_module(const std::string& in_name, const std::string& out_name);
     module* get_module(const std::string& name);
     const module* get_module(const std::string& name) const;
 

src/include/migraphx/serialize.hpp

@@ -28,6 +28,7 @@
 #include <migraphx/value.hpp>
 #include <migraphx/reflect.hpp>
 #include <migraphx/requires.hpp>
+#include <migraphx/optional.hpp>
 #include <migraphx/rank.hpp>
 #include <type_traits>
 
@@ -60,11 +61,12 @@ value to_value_impl(rank<0>, const T&)
     return value::object{};
 }
 
-template <class T, class U>
-value to_value_impl(rank<1>, const std::pair<T, U>& x)
+template <class T>
+auto to_value_impl(rank<1>, const T& x) -> decltype(std::tuple_size<T>{}, value{})
 {
-
-    return {x.first, x.second};
+    value result = value::array{};
+    repeat_c<std::tuple_size<T>{}>([&](auto i) { result.push_back(to_value(std::get<i>(x))); });
+    return result;
 }
 
 template <class T>
@@ -86,46 +88,55 @@ value to_value_impl(rank<3>, const T& x)
     return result;
 }
 
+template <class T>
+auto to_value_impl(rank<4>, const optional<T>& x)
+{
+    value result{};
+    if(x.has_value())
+        return to_value(*x);
+    return result;
+}
+
 template <class T, MIGRAPHX_REQUIRES(std::is_signed<T>{})>
-value to_value_impl(rank<4>, const T& x)
+value to_value_impl(rank<5>, const T& x)
 {
     return std::int64_t{x};
 }
 
 template <class T, MIGRAPHX_REQUIRES(std::is_unsigned<T>{})>
-value to_value_impl(rank<5>, const T& x)
+value to_value_impl(rank<6>, const T& x)
 {
     return std::uint64_t{x};
 }
 
 template <class T, MIGRAPHX_REQUIRES(std::is_floating_point<T>{})>
-value to_value_impl(rank<6>, const T& x)
+value to_value_impl(rank<7>, const T& x)
 {
     return double{x};
 }
 
 template <class T, MIGRAPHX_REQUIRES(std::is_enum<T>{})>
-value to_value_impl(rank<7>, const T& x)
+value to_value_impl(rank<8>, const T& x)
 {
     return x;
 }
 
-inline value to_value_impl(rank<8>, const std::string& x) { return x; }
+inline value to_value_impl(rank<9>, const std::string& x) { return x; }
 
 template <class T>
-auto to_value_impl(rank<9>, const T& x) -> decltype(migraphx_to_value(x))
+auto to_value_impl(rank<10>, const T& x) -> decltype(migraphx_to_value(x))
 {
     return migraphx_to_value(x);
 }
 
 template <class T>
-auto to_value_impl(rank<10>, const T& x) -> decltype(x.to_value())
+auto to_value_impl(rank<11>, const T& x) -> decltype(x.to_value())
 {
     return x.to_value();
 }
 
 template <class T>
-auto to_value_impl(rank<11>, const T& x)
+auto to_value_impl(rank<12>, const T& x)
     -> decltype(migraphx_to_value(std::declval<value&>(), x), value{})
 {
     value v;
@@ -144,7 +155,14 @@ void from_value_impl(rank<0>, const value& v, T& x)
 }
 
 template <class T>
-auto from_value_impl(rank<1>, const value& v, T& x)
+auto from_value_impl(rank<1>, const value& v, T& x) -> decltype(std::tuple_size<T>{}, void())
+{
+    repeat_c<std::tuple_size<T>{}>(
+        [&](auto i) { std::get<i>(x) = from_value<std::tuple_element_t<i, T>>(v[i]); });
+}
+
+template <class T>
+auto from_value_impl(rank<2>, const value& v, T& x)
     -> decltype(x.insert(x.end(), *x.begin()), void())
 {
     x.clear();
@@ -153,7 +171,7 @@ auto from_value_impl(rank<1>, const value& v, T& x)
 }
 
 template <class T, MIGRAPHX_REQUIRES(std::is_arithmetic<typename T::value_type>{})>
-auto from_value_impl(rank<2>, const value& v, T& x)
+auto from_value_impl(rank<3>, const value& v, T& x)
     -> decltype(x.insert(x.end(), *x.begin()), void())
 {
     x.clear();
@@ -170,7 +188,7 @@ auto from_value_impl(rank<2>, const value& v, T& x)
 }
 
 template <class T>
-auto from_value_impl(rank<3>, const value& v, T& x) -> decltype(x.insert(*x.begin()), void())
+auto from_value_impl(rank<4>, const value& v, T& x) -> decltype(x.insert(*x.begin()), void())
 {
     x.clear();
     for(auto&& e : v)
@@ -178,7 +196,7 @@ auto from_value_impl(rank<3>, const value& v, T& x) -> decltype(x.insert(*x.begi
 }
 
 template <class T, MIGRAPHX_REQUIRES(is_reflectable<T>{})>
-void from_value_impl(rank<4>, const value& v, T& x)
+void from_value_impl(rank<5>, const value& v, T& x)
 {
     reflect_each(x, [&](auto& y, const std::string& name) {
         using type = std::decay_t<decltype(y)>;
@@ -187,28 +205,29 @@ void from_value_impl(rank<4>, const value& v, T& x)
     });
 }
 
-template <class T, MIGRAPHX_REQUIRES(std::is_arithmetic<T>{})>
-void from_value_impl(rank<5>, const value& v, T& x)
+template <class T>
+void from_value_impl(rank<6>, const value& v, optional<T>& x)
 {
-    x = v.to<T>();
+    if(not v.is_null())
+        x = from_value<T>(v);
 }
 
-template <class T, MIGRAPHX_REQUIRES(std::is_enum<T>{})>
-void from_value_impl(rank<6>, const value& v, T& x)
+template <class T, MIGRAPHX_REQUIRES(std::is_arithmetic<T>{} or std::is_enum<T>{})>
+void from_value_impl(rank<7>, const value& v, T& x)
 {
     x = v.to<T>();
 }
 
-inline void from_value_impl(rank<7>, const value& v, std::string& x) { x = v.to<std::string>(); }
+inline void from_value_impl(rank<8>, const value& v, std::string& x) { x = v.to<std::string>(); }
 
 template <class T>
-auto from_value_impl(rank<8>, const value& v, T& x) -> decltype(x.from_value(v), void())
+auto from_value_impl(rank<9>, const value& v, T& x) -> decltype(x.from_value(v), void())
 {
     x.from_value(v);
 }
 
 template <class T>
-auto from_value_impl(rank<9>, const value& v, T& x) -> decltype(migraphx_from_value(v, x), void())
+auto from_value_impl(rank<10>, const value& v, T& x) -> decltype(migraphx_from_value(v, x), void())
 {
     migraphx_from_value(v, x);
 }
@@ -218,13 +237,13 @@ auto from_value_impl(rank<9>, const value& v, T& x) -> decltype(migraphx_from_va
 template <class T>
 value to_value(const T& x)
 {
-    return detail::to_value_impl(rank<11>{}, x);
+    return detail::to_value_impl(rank<12>{}, x);
 }
 
 template <class T>
 void from_value(const value& v, T& x)
 {
-    detail::from_value_impl(rank<9>{}, v, x);
+    detail::from_value_impl(rank<10>{}, v, x);
 }
 
 } // namespace MIGRAPHX_INLINE_NS

src/include/migraphx/shape.hpp

@@ -243,6 +243,9 @@ struct shape
     /// Return true if the shape is dynamic
     bool dynamic() const;
 
+    /// Return true if this shape or any of the sub_shapes are dynamic
+    bool any_of_dynamic() const;
+
     shape normalize_standard() const;
 
     shape with_lens(type_t t, const std::vector<std::size_t>& l) const;

src/include/migraphx/split_single_dyn_dim.hpp

+/*
+ * The MIT License (MIT)
+ *
+ * Copyright (c) 2015-2023 Advanced Micro Devices, Inc. All rights reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
+ * THE SOFTWARE.
+ */
+#ifndef MIGRAPHX_GUARD_RTGLIB_SPLIT_SINGLE_DYN_DIM_HPP
+#define MIGRAPHX_GUARD_RTGLIB_SPLIT_SINGLE_DYN_DIM_HPP
+
+#include <string>
+#include <migraphx/program.hpp>
+#include <migraphx/instruction_ref.hpp>
+#include <migraphx/config.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+
+/**
+ * Split dynamic batch dimension over submodules if exactly one dimension in the parameter list
+ * is dynamic. Should only run on the main module.
+ */
+struct split_single_dyn_dim
+{
+    std::string name() const { return "split_single_dyn_dim"; }
+    void apply(module_pass_manager& p) const;
+};
+
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx
+
+#endif

src/memory_coloring.cpp

+/*
+ * The MIT License (MIT)
+ *
+ * Copyright (c) 2015-2022 Advanced Micro Devices, Inc. All rights reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
+ * THE SOFTWARE.
+ */
+#include <migraphx/memory_coloring.hpp>
+#include <migraphx/module.hpp>
+#include <migraphx/operators.hpp>
+#include <migraphx/instruction.hpp>
+#include <migraphx/iterator_for.hpp>
+#include <migraphx/functional.hpp>
+#include <migraphx/algorithm.hpp>
+#include <migraphx/ranges.hpp>
+#include <migraphx/stringutils.hpp>
+#include <unordered_set>
+#include <unordered_map>
+#include <map>
+#include <set>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+
+MIGRAPHX_DECLARE_ENV_VAR(MIGRAPHX_DEBUG_MEMORY_COLORING);
+
+using instruction_set     = std::unordered_set<instruction_ref>;
+using instruction_set_map = std::unordered_map<instruction_ref, instruction_set>;
+
+// This will do liveness analysis on the module, and it will call the
+// function `f` with the instruction and the set of the other instructions
+// that are live
+template <class F>
+void liveness(const module& m, F f)
+{
+    auto implicit_deps = m.calc_implicit_deps();
+    instruction_set live_set;
+    auto rp = reverse(m);
+    for(auto rins : iterator_for(rp)) // NOLINT
+    {
+        // The base iterator is one ahead, so we need to use the previous iterator
+        auto ins = std::prev(rins.base());
+        // Add live variables
+        auto add_live_variables = [&](const auto& inputs) {
+            for(auto input : inputs)
+            {
+                auto i = instruction::get_output_alias(input);
+                // Skip if variable comes from parent
+                if(not m.has_instruction(i))
+                    continue;
+                live_set.insert(i);
+            }
+        };
+        add_live_variables(ins->inputs());
+        add_live_variables(implicit_deps[ins]);
+        // Remove last usage
+        auto it = live_set.find(ins);
+        if(it != live_set.end())
+        {
+            live_set.erase(it);
+            f(ins, live_set);
+        }
+    }
+}
+
+// This will build the conflict table or interference graph. This is
+// essentially a map from one instruction to a set of instruction that are
+// used together. Each instruction will be the allocation instruction.
+instruction_set_map build_conflict_table(const module& m, std::string allocation_op)
+{
+    instruction_set_map conflict_table;
+    liveness(m, [&](auto ins, auto live_set) {
+        // Skip variables that aren't allocations
+        if(ins->name() != allocation_op)
+            return;
+        // Skip zero allocations
+        if(ins->get_shape().bytes() == 0)
+            return;
+        conflict_table[ins];
+        for(auto i : live_set)
+        {
+            if(i == ins)
+                continue;
+            // Skip variables that aren't allocations
+            if(i->name() != allocation_op)
+                continue;
+            // Skip zero allocations
+            if(i->get_shape().bytes() == 0)
+                continue;
+            conflict_table[i].insert(ins);
+            conflict_table[ins].insert(i);
+        }
+    });
+    assert(std::all_of(conflict_table.begin(), conflict_table.end(), [](auto&& pp) {
+        return pp.second.count(pp.first) == 0;
+    }));
+    return conflict_table;
+}
+
+// Check if intervals overlap
+bool is_overlap(std::pair<std::size_t, std::size_t> x, std::pair<std::size_t, std::size_t> y)
+{
+    return std::max(x.first, y.first) < std::min(x.second, y.second);
+}
+
+struct allocation_segment
+{
+    using segment = std::pair<std::size_t, std::size_t>;
+    std::unordered_map<instruction_ref, segment> ins2segment;
+
+    const segment* add_segment(instruction_ref ins, segment s) { return &(ins2segment[ins] = s); }
+
+    const segment* get_segment(instruction_ref ins) const
+    {
+        auto it = ins2segment.find(ins);
+        if(it == ins2segment.end())
+            return nullptr;
+        return &it->second;
+    }
+
+    // Remove segment for an instruction
+    void remove(instruction_ref ins)
+    {
+        auto it = ins2segment.find(ins);
+        if(it != ins2segment.end())
+        {
+            ins2segment.erase(it);
+        }
+    }
+
+    std::size_t max()
+    {
+        std::size_t n = 0;
+        for(auto&& pp : ins2segment)
+        {
+            auto seg = pp.second;
+            n        = std::max(n, seg.second);
+        }
+        return n;
+    }
+
+    template <class Iterator>
+    static bool overlaps(Iterator first, Iterator last, const segment& s)
+    {
+        return std::any_of(first, last, [&](auto&& t) { return is_overlap(s, t); });
+    }
+
+    static bool overlaps(const std::set<segment>& segments, const segment& s)
+    {
+        return overlaps(segments.begin(), segments.end(), s);
+    }
+
+    static auto find_gap(const std::set<segment>& segments, std::size_t n)
+    {
+        std::size_t max_end = 0;
+        return std::adjacent_find(segments.begin(), segments.end(), [&](segment x, segment y) {
+            if(x.second < max_end)
+                return false;
+            max_end = x.second;
+            if(is_overlap(x, y))
+                return false;
+            assert(y.first >= x.second);
+            auto k = y.first - x.second;
+            return (k >= n);
+        });
+    }
+
+    static std::size_t max_type_size(const shape& s)
+    {
+        return std::accumulate(
+            s.sub_shapes().begin(),
+            s.sub_shapes().end(),
+            s.type_size(),
+            [](auto size, const auto& sub) { return std::max(size, max_type_size(sub)); });
+    }
+
+    static std::size_t compute_alignment(instruction_ref ins)
+    {
+        auto alignment = max_type_size(ins->get_shape());
+        // A rough estimate for the total number of elements
+        auto n = ins->get_shape().bytes() / alignment;
+        // Check for vectorized alignment
+        if(n > 4)
+        {
+            auto d = n % 4;
+            if(d == 0)
+                alignment *= 4;
+            if(d == 2)
+                alignment *= 2;
+        }
+        return alignment;
+    }
+
+    static segment
+    next_segment(std::set<segment>& segments, instruction_ref ins, std::size_t alignment)
+    {
+        assert(ins->get_shape().bytes() > 0);
+        // Compute alignment
+        auto n = 1 + (ins->get_shape().bytes() - 1) / alignment;
+        assert(n > 0);
+        auto start = 0;
+        // Insert at end if it cant fit at the begining
+        if(segments.empty() or segments.begin()->first <= n)
+        {
+            auto it = find_gap(segments, n);
+            if(it == segments.end())
+                it = std::max_element(segments.begin(), segments.end(), [&](segment x, segment y) {
+                    return x.second < y.second;
+                });
+            if(it != segments.end())
+                start = it->second;
+        }
+        auto s = segment{start, start + n};
+        assert(not overlaps(segments, s));
+        segments.insert(s);
+        return s;
+    }
+
+    static std::unordered_map<instruction_ref, int>
+    create_allocation_index(const module& m, const instruction_set_map& conflict_table)
+    {
+        std::unordered_map<instruction_ref, int> result;
+        int i = 0;
+        for(auto ins : iterator_for(m))
+        {
+            if(not contains(conflict_table, ins))
+                continue;
+            result[ins] = i++;
+        }
+        return result;
+    }
+
+    // Build the allocation_color class from the conflict_table
+    static allocation_segment
+    build(const module& m, const instruction_set_map& conflict_table, std::size_t alignment)
+    {
+        allocation_segment as{};
+        std::vector<instruction_ref> conflict_queue;
+        // Add all allocations to the conflict_queue
+        std::transform(conflict_table.begin(),
+                       conflict_table.end(),
+                       std::back_inserter(conflict_queue),
+                       [](auto&& pp) { return pp.first; });
+
+        auto alloc_index = create_allocation_index(m, conflict_table);
+
+        // Sort the conflict queue so we process the allocation with the most
+        // number of adjacent allocations first
+        std::sort(conflict_queue.begin(), conflict_queue.end(), by(std::greater<>{}, [&](auto x) {
+                      return std::make_tuple(
+                          conflict_table.at(x).size(), x->get_shape().bytes(), alloc_index.at(x));
+                  }));
+        // Process the conflict_queue, we refer to the current allocation as
+        // the parent and the adjacent allocations as children
+        for(auto parent : conflict_queue)
+        {
+            // Sort children by size
+            std::vector<instruction_ref> children(conflict_table.at(parent).begin(),
+                                                  conflict_table.at(parent).end());
+            std::sort(children.begin(), children.end(), by(std::less<>{}, [&](auto x) {
+                          return std::make_tuple(x->get_shape().bytes(), alloc_index.at(x));
+                      }));
+            assert(not contains(children, parent));
+            // This set is to track the segments already processed
+            std::set<segment> segments;
+            // Add all segments for the children to the segments already processed
+            transform_if(
+                children.begin(),
+                children.end(),
+                std::inserter(segments, segments.begin()),
+                [&](auto child) { return as.get_segment(child); },
+                [&](auto child) { return *as.get_segment(child); });
+
+            assert(as.get_segment(parent) == nullptr);
+            as.add_segment(parent, next_segment(segments, parent, alignment));
+        }
+        // Reduce the number of segments
+        for(std::size_t n = 0; n < 3; n++)
+        {
+            for(auto parent : conflict_queue)
+            {
+                auto children = conflict_table.at(parent);
+                // This set is to track the segments already processed
+                std::set<segment> segments;
+                // Add all segments for the children to the segments already processed
+                transform_if(
+                    children.begin(),
+                    children.end(),
+                    std::inserter(segments, segments.begin()),
+                    [&](auto child) { return as.get_segment(child); },
+                    [&](auto child) { return *as.get_segment(child); });
+                // Get the segment for the parent
+                const auto* parent_segment = as.get_segment(parent);
+                assert(parent_segment != nullptr);
+
+                auto s = next_segment(segments, parent, alignment);
+                if(s != *parent_segment and s.second <= as.max())
+                {
+                    as.add_segment(parent, s);
+                }
+            }
+        }
+        return as;
+    }
+};
+
+static std::size_t find_max_alignment(const module& m, const std::string& allocation_op)
+{
+    std::size_t alignment = 1;
+    for(auto ins : iterator_for(m))
+    {
+        if(ins->name() != allocation_op)
+            continue;
+        alignment = std::max(allocation_segment::compute_alignment(ins), alignment);
+    }
+    return alignment;
+}
+
+void memory_coloring::apply(module& m) const
+{
+    const std::size_t alignment = find_max_alignment(m, allocation_op);
+    auto conflict_table         = build_conflict_table(m, allocation_op);
+    auto as                     = allocation_segment::build(m, conflict_table, alignment);
+
+    // All allocations should have a segment
+    assert(std::all_of(conflict_table.begin(), conflict_table.end(), [&](auto&& pp) {
+        return as.get_segment(pp.first);
+    }));
+
+    // Adjacent allocations should not have overlapping segments
+    assert(std::none_of(conflict_table.begin(), conflict_table.end(), [&](auto&& pp) {
+        auto* x = as.get_segment(pp.first);
+        return std::any_of(pp.second.begin(), pp.second.end(), [&](auto ins) {
+            auto* y = as.get_segment(ins);
+            assert(x and y);
+            return is_overlap(*x, *y);
+        });
+    }));
+
+    // Print out segments
+    if(enabled(MIGRAPHX_DEBUG_MEMORY_COLORING{}))
+    {
+        for(auto&& pp : conflict_table)
+        {
+            std::cout << "------- conflict -------" << std::endl;
+            auto s1 = as.ins2segment.at(pp.first);
+            std::cout << s1.first << ", " << s1.second << ": ";
+            m.debug_print(pp.first);
+            for(auto ins : pp.second)
+            {
+                auto s2 = as.ins2segment.at(ins);
+                std::cout << s2.first << ", " << s2.second << ": ";
+                m.debug_print(ins);
+            }
+        }
+    }
+
+    // Total memory
+    std::size_t n = as.max() * alignment;
+
+    // Replace allocations
+    auto mem = m.add_parameter("scratch", shape{shape::int8_type, {n}});
+    for(auto&& [ins, seg] : as.ins2segment)
+    {
+        assert(ins->name() == allocation_op);
+        auto s             = ins->get_shape();
+        std::size_t offset = seg.first * alignment;
+        assert(offset < n);
+        m.replace_instruction(ins, op::load{s, offset}, mem);
+    }
+
+    // Replace zero allocation
+    for(auto ins : iterator_for(m))
+    {
+        if(ins->name() != allocation_op)
+            continue;
+        assert(ins->get_shape().bytes() == 0);
+        m.replace_instruction(ins, op::load{ins->get_shape(), 0}, mem);
+    }
+
+    // Remove scratch parameter if its not used
+    if(mem->outputs().empty())
+    {
+        m.remove_instruction(mem);
+    }
+}
+
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx

src/module.cpp

@@ -822,7 +822,8 @@ static void print_make_op(std::ostream& os, const operation& op)
 
 static void print_py_shape(std::ostream& os, const migraphx::shape& s)
 {
-    os << "migraphx.shape(" << s.type_string() << ", lens=" << to_json_string(s.lens());
+    os << "migraphx.shape(type=" << to_json_string(s.type_string())
+       << ", lens=" << to_json_string(s.lens());
     if(not s.standard())
         os << ", strides=" << to_json_string(s.strides());
     os << ")";

src/normalize_attributes.cpp

@@ -30,13 +30,16 @@
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 
-// different attributes
-// 1) use_input(default)/use_output
-// 2) use_rank(default)/use_len
-// 3) clip_min(default)/not_clip_min
-//   3.1) include_min(default)/exclude_min
-// 4) clip_max(default)/not_clip_max
-//   4.1) exclude_max(default)/include_max
+/**
+ * Parameters:
+ * vec: the vector attribute to normalize
+ * axes: the operator's axes attribute if it exists, empty otherwise
+ * val: the normalize_axes key and options. Ex: normalize["axes"] =
+ * value::array{normalize_attribute::include_min}; lens: shape dimensions passed when calling
+ * normalize_attributes(op&, lens)
+ *
+ * See normalize_attribute.hpp for explaining the options.
+ */
 auto tune_attribute(const std::vector<int64_t>& vec,
                     const std::vector<int64_t>& axes,
                     const value& val,
@@ -151,6 +154,11 @@ auto tune_pad_attribute(const value& val)
     return result;
 }
 
+/**
+ * Assumptions:
+ *  Dimensions to pad start from the third dimension (index 2).
+ *  Called by compute_shape_op() with the `lens` of the first input.
+ */
 bool normalize_attributes(operation& op, const std::vector<std::size_t>& lens)
 {
     bool tuned = false;
@@ -160,7 +168,6 @@ bool normalize_attributes(operation& op, const std::vector<std::size_t>& lens)
     {
         auto padding       = val.at(attrs.at("normalize_padding").to<std::string>());
         auto padding_size  = padding.size();
-        // for now, assume the dimensions to pad start at dim 2
         auto padding_start = 2;
 
         if(padding_size == 2 * (lens.size() - padding_start))

src/onnx/include/migraphx/onnx/onnx_parser.hpp

@@ -113,7 +113,8 @@ struct onnx_parser
 
     void parse_from(std::istream& is, std::string name = "");
     void parse_from(const void* data, std::size_t size);
-    void parse_graph(module* mod, const onnx::GraphProto& graph);
+    std::vector<instruction_ref>
+    parse_graph(module* mod, const onnx::GraphProto& graph, bool inlining = false);
     literal parse_value(const onnx::AttributeProto& attr) const;
     literal parse_tensor(const onnx::TensorProto& t) const;
     shape parse_type(const onnx::TypeProto& t, const std::vector<std::size_t>& input_dims) const;

src/onnx/onnx_parser.cpp

@@ -220,7 +220,7 @@ void onnx_parser::parse_from(std::istream& is, std::string name)
 
         if(model.has_graph())
         {
-            this->parse_graph(mm, model.graph());
+            (void)this->parse_graph(mm, model.graph());
         }
     }
     else
@@ -240,7 +240,7 @@ void onnx_parser::parse_from(const void* data, std::size_t size)
 
         if(model.has_graph())
         {
-            this->parse_graph(mm, model.graph());
+            (void)this->parse_graph(mm, model.graph());
         }
     }
     else
@@ -264,7 +264,8 @@ int64_t onnx_parser::get_opset_version(const onnx::ModelProto& model)
     return version;
 }
 
-void onnx_parser::parse_graph(module* mod, const onnx::GraphProto& graph)
+std::vector<instruction_ref>
+onnx_parser::parse_graph(module* mod, const onnx::GraphProto& graph, bool inlining)
 {
     std::unordered_map<std::string, instruction_ref> mod_insts;
     for(auto&& f : graph.initializer())
@@ -372,11 +373,16 @@ void onnx_parser::parse_graph(module* mod, const onnx::GraphProto& graph)
                    std::back_inserter(output_ins),
                    [&](const auto& name) { return instructions[name]; });
 
+    if(not inlining)
+    {
         // add the return instuction
         mod->add_return(output_ins);
 
-    // remove instructions added in this mod
+        // Remove instructions added in module (this is turned off for subgraph inlining)
         erase_if(instructions, [&](auto&& p) { return mod->has_instruction(p.second); });
+    }
+
+    return output_ins;
 }
 
 literal onnx_parser::parse_value(const onnx::AttributeProto& attr) const

src/onnx/parse_if.cpp

@@ -51,6 +51,24 @@ struct parse_if : op_parser<parse_if>
                            " condition input can have only one element!");
         }
 
+        // Fold instruction if condition is constant thus can be evaled
+        // prior to inference
+        if(args.front()->can_eval())
+        {
+            auto cond_arg = args.front()->eval();
+            auto* mod     = info.mod;
+            // then branch
+            if(cond_arg.at<bool>())
+            {
+                return parser.parse_graph(mod, then_graph, true);
+            }
+            // else branch
+            else
+            {
+                return parser.parse_graph(mod, else_graph, true);
+            }
+        }
+
         std::string then_name = info.name + "_if";
         module_ref then_mdl   = parser.prog.create_module(then_name);
 
@@ -58,10 +76,10 @@ struct parse_if : op_parser<parse_if>
         module_ref else_mdl   = parser.prog.create_module(else_name);
 
         // parse the then sub_graph
-        parser.parse_graph(then_mdl, then_graph);
+        (void)parser.parse_graph(then_mdl, then_graph);
 
         // parse_the else sub_graph
-        parser.parse_graph(else_mdl, else_graph);
+        (void)parser.parse_graph(else_mdl, else_graph);
 
         auto then_out_shapes = then_mdl->get_output_shapes();
         auto else_out_shapes = else_mdl->get_output_shapes();

src/onnx/parse_loop.cpp

@@ -71,7 +71,7 @@ struct parse_loop : op_parser<parse_loop>
         module_ref sub_mod    = parser.prog.create_module(mod_name);
 
         // parse the sub_graph
-        parser.parse_graph(sub_mod, sub_graph);
+        (void)parser.parse_graph(sub_mod, sub_graph);
 
         auto ret = info.add_instruction(
             make_op("loop", {{"max_iterations", max_iterations}}), args, {sub_mod});

src/onnx/parse_slice.cpp

@@ -46,7 +46,7 @@ struct parse_slice : op_parser<parse_slice>
         std::vector<int64_t> steps;
 
         // slice can have up to 5 inputs, we first check the 5th one
-        // to decide whether MIGRAPHX can handle this slice
+        // to decide whether MIGRAPHX can handle this slice.
         if(args.size() == 5)
         {
             migraphx::argument step_arg = args.back()->eval();
@@ -90,9 +90,10 @@ struct parse_slice : op_parser<parse_slice>
             s.visit([&](auto v) { copy(v, std::back_inserter(op.starts)); });
         }
 
+        // If axes arg is not given, the default is all of them.
         if(op.axes.empty())
         {
-            std::vector<int64_t> axes(args[0]->get_shape().lens().size());
+            std::vector<int64_t> axes(args[0]->get_shape().ndim());
             std::iota(axes.begin(), axes.end(), int64_t{0});
             op.axes = axes;
         }
@@ -103,6 +104,7 @@ struct parse_slice : op_parser<parse_slice>
         assert(op.axes.size() == op.starts.size());
         assert(op.axes.size() == op.ends.size());
 
+        // If any axes have negative step, prepare to add a "reverse" op
         for(auto i : range(steps.size()))
         {
             if(steps[i] >= 0)
@@ -117,7 +119,10 @@ struct parse_slice : op_parser<parse_slice>
 
         auto ins = info.add_instruction(op, args[0]);
         if(not raxes.empty())
+        {
             ins = info.add_instruction(make_op("reverse", {{"axes", raxes}}), ins);
+        }
+        // If any steps are other than default 1, add a "steps" op
         if(std::any_of(steps.begin(), steps.end(), [](auto s) { return std::abs(s) != 1; }))
         {
             std::vector<int64_t> nsteps;

src/onnx/parse_trilu.cpp

+/*
+ * The MIT License (MIT)
+ *
+ * Copyright (c) 2015-2022 Advanced Micro Devices, Inc. All rights reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
+ * THE SOFTWARE.
+ */
+#include <migraphx/onnx/op_parser.hpp>
+#include <migraphx/onnx/checks.hpp>
+#include <migraphx/ranges.hpp>
+#include <migraphx/instruction.hpp>
+#include <migraphx/make_op.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace onnx {
+
+struct parse_trilu : op_parser<parse_trilu>
+{
+    std::vector<op_desc> operators() const { return {{"Trilu"}}; }
+
+    instruction_ref parse(const op_desc&,
+                          const onnx_parser&,
+                          const onnx_parser::node_info& info,
+                          std::vector<instruction_ref> args) const
+    {
+        auto input_shape = args[0]->get_shape();
+        assert(input_shape.ndim() >= 2);
+        auto input_lens = input_shape.lens();
+
+        size_t num_rows = *(input_lens.rbegin() + 1);
+        size_t num_cols = input_lens.back();
+        int k           = 0;
+        bool upper      = true;
+
+        if(args.size() > 1)
+        {
+            auto arg_k = args[1]->eval();
+            check_arg_empty(arg_k, "PARSE_TRILU: dynamic k not supported");
+            k = arg_k.at<int>();
+        }
+
+        if(k < 0)
+            MIGRAPHX_THROW("PARSE_TRILU: negative k values not supported");
+
+        if(contains(info.attributes, "upper"))
+        {
+            upper = static_cast<bool>(info.attributes.at("upper").i());
+        }
+
+        shape::type_t output_type = args[0]->get_shape().type();
+
+        // when creating the mask, if upper == 1,
+        // the inner triangle will have values set to 0
+        std::vector<bool> mask_mat(num_rows * num_cols, upper);
+        for(size_t i = 0; i < num_rows; i++)
+        {
+            for(size_t j = 0; j < std::min(k, static_cast<int>(num_cols)); j++)
+            {
+                mask_mat[i * num_cols + j] = not upper;
+            }
+            k++;
+        }
+
+        auto mask = info.add_literal(
+            migraphx::literal{migraphx::shape{output_type, {num_rows, num_cols}}, mask_mat});
+
+        return info.add_broadcastable_binary_op("mul", mask, args[0]);
+    }
+};
+
+} // namespace onnx
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx

src/onnx/parse_where.cpp

 /*
  * The MIT License (MIT)
  *
- * Copyright (c) 2015-2022 Advanced Micro Devices, Inc. All rights reserved.
+ * Copyright (c) 2015-2023 Advanced Micro Devices, Inc. All rights reserved.
  *
  * Permission is hereby granted, free of charge, to any person obtaining a copy
  * of this software and associated documentation files (the "Software"), to deal
@@ -40,6 +40,19 @@ struct parse_where : op_parser<parse_where>
                           const onnx_parser::node_info& info,
                           std::vector<instruction_ref> args) const
     {
+        // TODO: broadcasting for dynamic shapes is only implemented
+        // for binary ops at time of writing, not ternary ops.
+        //   When it becomes available, add multibroadcasting steps in the dynamic shape case.
+        // For now for dynamic shapes, just insert the Where op.  All shapes must be the
+        // same for it to succeed.
+        if(std::all_of(args.begin(), args.end(), [](auto v) { return v->get_shape().dynamic(); }))
+        {
+            return info.add_instruction(make_op("where"), args[0], args[1], args[2]);
+        }
+        else if(std::none_of(
+                    args.begin(), args.end(), [](auto v) { return v->get_shape().dynamic(); }))
+        {
+            // If shapes are static and any are broadcasted, insert multibroadcast ops
             auto lens =
                 compute_broadcasted_lens(args[0]->get_shape().lens(), args[1]->get_shape().lens());
             lens = compute_broadcasted_lens(lens, args[2]->get_shape().lens());
@@ -63,6 +76,9 @@ struct parse_where : op_parser<parse_where>
 
             return info.add_instruction(make_op("where"), args[0], args[1], args[2]);
         }
+        else
+            MIGRAPHX_THROW("PARSE_WHERE: doesn't support mixed static and dynamic shape inputs");
+    }
 };
 
 } // namespace onnx

src/optimize_module.cpp

+/*
+ * The MIT License (MIT)
+ *
+ * Copyright (c) 2015-2022 Advanced Micro Devices, Inc. All rights reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
+ * THE SOFTWARE.
+ */
+#include <migraphx/optimize_module.hpp>
+#include <migraphx/pass_manager.hpp>
+#include <migraphx/simplify_reshapes.hpp>
+#include <migraphx/simplify_algebra.hpp>
+#include <migraphx/eliminate_common_subexpression.hpp>
+#include <migraphx/dead_code_elimination.hpp>
+#include <migraphx/propagate_constant.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+
+void optimize_module::apply(module_pass_manager& mpm) const
+{
+    for(int i = 0; i < 2; i++)
+    {
+        mpm.run_pass(simplify_reshapes{});
+        mpm.run_pass(simplify_algebra{});
+        mpm.run_pass(eliminate_common_subexpression{});
+        mpm.run_pass(dead_code_elimination{});
+        mpm.run_pass(propagate_constant{});
+        mpm.run_pass(dead_code_elimination{});
+    }
+}
+
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx

src/pass_manager.cpp

@@ -39,6 +39,7 @@ namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 
 MIGRAPHX_DECLARE_ENV_VAR(MIGRAPHX_TRACE_PASSES);
+MIGRAPHX_DECLARE_ENV_VAR(MIGRAPHX_TIME_PASSES);
 
 void validate_pass(module& mod, const pass& p, tracer trace)
 {
@@ -94,19 +95,19 @@ struct module_pm : module_pass_manager
     virtual void run_pass(const pass& p) override
     {
         assert(mod);
-
-        timer ts{};
-        using seconds = std::chrono::duration<double>;
-
-        trace("Module: ", mod->name(), ", Pass: ", p.name());
-        const double t1 = ts.record<seconds>();
         assert(mod->validate() == mod->end());
+        if(enabled(MIGRAPHX_TIME_PASSES{}))
+        {
+            using milliseconds = std::chrono::duration<double, std::milli>;
+            auto ms            = time<milliseconds>([&] { p.apply(*this); });
+            std::cout << p.name() << ": " << ms << "ms\n";
+        }
+        else
+        {
             p.apply(*this);
+        }
         trace(*mod);
         validate_pass(*mod, p, *t);
-
-        const double t2 = ts.record<seconds>();
-        trace("Pass: ", p.name(), " completed in (s): ", (t2 - t1));
     }
 };
 

src/program.cpp

@@ -210,17 +210,15 @@ void program::compile(const target& t, compile_options options)
     assert(not this->is_compiled());
     this->impl->target_name = t.name();
     this->impl->ctx         = t.get_context();
+
     if(enabled(MIGRAPHX_TRACE_COMPILE{}))
         options.trace = tracer{std::cout};
 
     options.trace(*this);
     options.trace();
-
     auto&& passes = t.get_passes(this->impl->ctx, options);
     run_passes(*this, passes, options.trace);
-
     auto mods = this->get_modules();
-
     // Validate and finalize
     for(const auto& mod : reverse(mods))
     {
@@ -336,7 +334,8 @@ std::vector<argument> generic_eval(const module* mod,
                     if(not ins->get_shape().dynamic() and param.get_shape() != ins->get_shape())
                     {
                         MIGRAPHX_THROW("Incorrect shape {" + to_string(param.get_shape()) +
-                                       "} for parameter: " + param_name);
+                                       "} for parameter: " + param_name +
+                                       " should be: " + to_string(ins->get_shape()));
                     }
                     return param;
                 }));
@@ -380,7 +379,7 @@ std::vector<argument> generic_eval(const module* mod,
                             }));
         }
         assert(results.find(ins) != results.end());
-        if(not ins->get_shape().dynamic())
+        if(not ins->get_shape().any_of_dynamic())
         {
             assert(results.at(ins).get_shape() == ins->get_shape());
         }