Merge

src/replace_allocate.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/replace_allocate.hpp>
+#include <migraphx/instruction.hpp>
+#include <migraphx/program.hpp>
+#include <migraphx/make_op.hpp>
+#include <migraphx/iterator_for.hpp>
+#include <migraphx/ranges.hpp>
+#include <migraphx/op/allocate.hpp>
+#include <map>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+
+std::unordered_map<instruction_ref, std::string> create_output_names(const module& mod)
+{
+    std::unordered_map<instruction_ref, std::string> mod_output_names{};
+    auto last = std::prev(mod.end());
+    if(last->name() == "@return")
+    {
+        const auto& prog_outputs = last->inputs();
+        std::vector<instruction_ref> outputs_alias(prog_outputs.size());
+
+        std::transform(prog_outputs.begin(),
+                       prog_outputs.end(),
+                       outputs_alias.begin(),
+                       [](const auto& i) { return instruction::get_output_alias(i); });
+
+        std::size_t index = 0;
+        for(auto ins : outputs_alias)
+        {
+            mod_output_names[ins] = mod.name() + ":#output_" + std::to_string(index++);
+        }
+    }
+    else
+    {
+        auto ins              = instruction::get_output_alias(last);
+        mod_output_names[ins] = "output";
+    }
+    return mod_output_names;
+}
+
+void insert_submod_allocations(instruction_ref ins, module& mod, const allocation_model& model)
+{
+    std::vector<instruction_ref> inputs = ins->inputs();
+    std::vector<module_ref> mod_args    = ins->module_inputs();
+
+    std::map<std::string, shape> name_shapes;
+    for(const auto& smod : mod_args)
+    {
+        auto ps = smod->get_parameter_shapes();
+        name_shapes.insert(ps.begin(), ps.end());
+    }
+
+    for(auto& pn : name_shapes)
+    {
+        const auto& s = pn.second;
+        instruction_ref output{};
+        output = mod.insert_instruction(ins, model.allocate(s));
+        inputs.push_back(output);
+    }
+
+    mod.replace_instruction(ins, ins->get_operator(), inputs, mod_args);
+}
+
+void replace_allocate::apply(module& m) const
+{
+    auto mod_output_names  = create_output_names(m);
+    bool main_offload_copy = m.name() == "main" ? this->offload_copy : false;
+    for(auto ins : iterator_for(m))
+    {
+        auto op      = ins->get_operator();
+        auto op_name = op.name();
+
+        // check if allocations from submodules need to be inserted
+        // for now, only the "if" operator is affected
+        if(op_name == "if")
+        {
+            insert_submod_allocations(ins, m, model);
+            continue;
+        }
+        if(op_name != "allocate")
+            continue;
+
+        auto s = ins->get_shape();
+
+        if(not main_offload_copy and model.needs_out_params() and contains(mod_output_names, ins))
+        {
+
+            auto out_param = m.add_parameter(mod_output_names[ins], s);
+            m.replace_instruction(ins, out_param);
+            continue;
+        }
+
+        m.replace_instruction(
+            ins,
+            m.insert_instruction(ins,
+                                 make_op(model.name(), migraphx::value{{"shape", to_value(s)}})));
+    }
+}
+
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx

src/rewrite_batchnorm.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/rewrite_batchnorm.hpp>
 #include <migraphx/program.hpp>
 #include <migraphx/instruction.hpp>

src/rewrite_gelu.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/rewrite_gelu.hpp>
+#include <migraphx/make_op.hpp>
+#include <migraphx/matcher.hpp>
+#include <migraphx/match/gelu_erf.hpp>
+#include <migraphx/common.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+
+struct find_gelu_erf
+{
+    auto matcher() const { return match::gelu_erf(); }
+
+    void apply(module& m, const match::matcher_result& r) const
+    {
+        auto ins = r.result;
+        auto x   = r.instructions["x"];
+        if(x->get_shape().type() != migraphx::shape::half_type)
+            return;
+
+        auto lit = m.add_literal(literal{shape{x->get_shape().type()}, {1.702f}});
+        auto mul = insert_common_op(m, ins, make_op("mul"), {x, lit});
+        auto sig = m.insert_instruction(ins, make_op("neg"), mul);
+        sig      = m.insert_instruction(ins, make_op("exp"), sig);
+        auto one = m.add_literal(literal{shape{x->get_shape().type()}, {1.0f}});
+        sig      = insert_common_op(m, ins, make_op("add"), {sig, one});
+        sig      = m.insert_instruction(ins, make_op("div"), x, sig);
+        m.replace_instruction(ins, sig);
+    }
+};
+
+void rewrite_gelu::apply(module& m) const { match::find_matches(m, find_gelu_erf{}); }
+
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx

src/rewrite_pooling.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/rewrite_pooling.hpp>
 #include <migraphx/instruction.hpp>
 #include <migraphx/iterator_for.hpp>
@@ -24,12 +47,12 @@ void rewrite_pooling::apply(module& m) const
         if(not s.standard())
             continue;
         auto&& op = any_cast<op::pooling>(ins->get_operator());
-        if(!std::all_of(op.padding.begin(), op.padding.end(), [](auto i) { return i == 0; }))
+        if(not std::all_of(op.padding.begin(), op.padding.end(), [](auto i) { return i == 0; }))
             continue;
-        if(!std::all_of(op.stride.begin(), op.stride.end(), [](auto i) { return i == 1; }))
+        if(not std::all_of(op.stride.begin(), op.stride.end(), [](auto i) { return i == 1; }))
             continue;
         auto lens = s.lens();
-        if(!std::equal(lens.begin() + 2, lens.end(), op.lengths.begin(), op.lengths.end()))
+        if(not std::equal(lens.begin() + 2, lens.end(), op.lengths.begin(), op.lengths.end()))
             continue;
         std::int64_t n = s.lens()[0];
         std::int64_t c = s.lens()[1];

src/rewrite_quantization.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/rewrite_quantization.hpp>
 #include <migraphx/instruction.hpp>
 #include <migraphx/iterator_for.hpp>

src/rewrite_rnn.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/rewrite_rnn.hpp>
 #include <migraphx/program.hpp>
 #include <migraphx/instruction.hpp>
@@ -191,7 +214,7 @@ void rewrite_rnn::apply_vanilla_rnn(module& m, instruction_ref ins) const
             ih = m.add_literal(migraphx::literal{ih_shape, data});
         }
 
-        if(!is_forward and variable_seq_len)
+        if(not is_forward and variable_seq_len)
         {
             args[0] =
                 m.insert_instruction(ins, make_op("rnn_var_sl_shift_sequence"), args[0], seq_lens);
@@ -497,7 +520,7 @@ void rewrite_rnn::apply_gru(module& m, instruction_ref ins) const
             ih = m.add_literal(migraphx::literal{ih_shape, data});
         }
 
-        if(!is_forward and variable_seq_len)
+        if(not is_forward and variable_seq_len)
         {
             args[0] =
                 m.insert_instruction(ins, make_op("rnn_var_sl_shift_sequence"), args[0], seq_lens);
@@ -954,7 +977,7 @@ void rewrite_rnn::apply_lstm(module& m, instruction_ref ins) const
             pph = args[7];
         }
 
-        if(!is_forward and variable_seq_len)
+        if(not is_forward and variable_seq_len)
         {
             args[0] =
                 m.insert_instruction(ins, make_op("rnn_var_sl_shift_sequence"), args[0], seq_lens);
@@ -1271,11 +1294,11 @@ bool rewrite_rnn::is_variable_seq_lens(const module& m, instruction_ref seq_lens
             std::vector<int64_t> vec_lens;
             arg_lens.visit([&](auto l) { vec_lens.assign(l.begin(), l.end()); });
             int64_t l = 0;
-            if(!vec_lens.empty())
+            if(not vec_lens.empty())
             {
                 l = vec_lens[0];
             }
-            if(!std::all_of(vec_lens.begin(), vec_lens.end(), [&](auto v) { return v == l; }))
+            if(not std::all_of(vec_lens.begin(), vec_lens.end(), [&](auto v) { return v == l; }))
             {
                 is_var_lens = true;
             }
@@ -1295,7 +1318,7 @@ rewrite_rnn::get_seq_len(const module& m, instruction_ref input, instruction_ref
     bool is_var_lens = is_variable_seq_lens(m, seq_lens);
     auto input_shape = input->get_shape();
     auto length      = input_shape.lens()[0];
-    if(!is_var_lens and seq_lens != m.end())
+    if(not is_var_lens and seq_lens != m.end())
     {
         auto arg_len = seq_lens->eval();
         std::vector<std::size_t> vec_lens;
@@ -1364,7 +1387,7 @@ void rewrite_rnn::replace_last_cell_output(module& m,
 
     if(variable_seq_len)
     {
-        if(!ins_outputs.empty())
+        if(not ins_outputs.empty())
         {
             cell_outputs = m.insert_instruction(
                 std::next(ins),

src/schedule.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/schedule.hpp>
 #include <migraphx/program.hpp>
 #include <migraphx/instruction.hpp>

src/serialize.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/serialize.hpp>
 #include <migraphx/argument.hpp>
 #include <migraphx/literal.hpp>
@@ -13,7 +36,7 @@ void raw_data_to_value(value& v, const RawData& rd)
     result["shape"] = migraphx::to_value(rd.get_shape());
     if(rd.get_shape().type() == shape::tuple_type)
         result["sub"] = migraphx::to_value(rd.get_sub_objects());
-    else
+    else if(not rd.empty())
         result["data"] = migraphx::value::binary(rd.data(), rd.get_shape().bytes());
     v = result;
 }
@@ -33,7 +56,7 @@ void migraphx_from_value(const value& v, argument& a)
         literal l = migraphx::from_value<literal>(v);
         a         = l.get_argument();
     }
-    else
+    else if(v.contains("sub"))
     {
         a = migraphx::from_value<std::vector<argument>>(v.at("sub"));
     }

src/shape.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/shape.hpp>
 #include <migraphx/stringutils.hpp>
 #include <migraphx/serialize.hpp>
 #include <migraphx/permutation.hpp>
+#include <migraphx/ranges.hpp>
 #include <numeric>
 #include <algorithm>
 #include <functional>
@@ -38,19 +62,25 @@ struct shape_impl
     {
         assert(t != shape::tuple_type);
         assert(m_lens.size() == m_strides.size());
-        // assert(std::any_of(m_strides.begin(), m_strides.end(), [](auto x) { return x > 0; }) and
-        //        "At least one stride must be non-zero");
-        m_standard = this->elements() == this->element_space() and
+        m_standard = this->elements() == this->element_space() and not skips() and
                      std::is_sorted(m_strides.rbegin(), m_strides.rend());
     }
 
+    shape_impl(shape::type_t t, std::vector<shape::dynamic_dimension> dims)
+        : m_type(t), m_dyn_dims(std::move(dims))
+    {
+    }
+
     shape_impl(const std::vector<shape>& subs) : m_type(shape::tuple_type), m_shapes(subs) {}
+
     shape::type_t m_type;
     std::vector<std::size_t> m_lens    = {};
     std::vector<std::size_t> m_strides = {};
     std::vector<shape> m_shapes        = {};
     bool m_standard                    = false;
 
+    std::vector<shape::dynamic_dimension> m_dyn_dims = {};
+
     void calculate_strides()
     {
         m_strides.clear();
@@ -66,6 +96,12 @@ struct shape_impl
 
     std::size_t element_space() const
     {
+        if(not m_dyn_dims.empty())
+        {
+            auto maxes = max_lens();
+            return std::accumulate(maxes.begin(), maxes.end(), std::size_t{1}, std::multiplies<>());
+        }
+
         assert(m_lens.size() == m_strides.size());
         if(m_lens.empty())
             return 0;
@@ -80,6 +116,11 @@ struct shape_impl
 
     std::size_t elements() const
     {
+        if(not m_dyn_dims.empty())
+        {
+            MIGRAPHX_THROW("SHAPE: elements() called on dynamic shape");
+        }
+
         assert(m_lens.size() == m_strides.size());
         if(m_lens.empty())
             return 0;
@@ -87,6 +128,44 @@ struct shape_impl
             m_lens.begin(), m_lens.end(), std::size_t{1}, std::multiplies<std::size_t>());
     }
 
+    std::vector<std::size_t> min_lens() const
+    {
+        std::vector<std::size_t> ret(m_dyn_dims.size());
+        std::transform(m_dyn_dims.cbegin(),
+                       m_dyn_dims.cend(),
+                       ret.begin(),
+                       [](shape::dynamic_dimension x) { return x.min; });
+        return ret;
+    }
+
+    std::vector<std::size_t> max_lens() const
+    {
+        std::vector<std::size_t> ret(m_dyn_dims.size());
+        std::transform(m_dyn_dims.cbegin(),
+                       m_dyn_dims.cend(),
+                       ret.begin(),
+                       [](shape::dynamic_dimension x) { return x.max; });
+        return ret;
+    }
+
+    std::vector<std::size_t> opt_lens() const
+    {
+        std::vector<std::size_t> ret(m_dyn_dims.size());
+        std::transform(m_dyn_dims.cbegin(),
+                       m_dyn_dims.cend(),
+                       ret.begin(),
+                       [](shape::dynamic_dimension x) { return x.opt; });
+        return ret;
+    }
+    // Does the shape skip over elements?
+    bool skips() const
+    {
+        assert(m_lens.size() == m_strides.size());
+        if(elements() == 1)
+            return false;
+        return std::none_of(m_strides.begin(), m_strides.end(), [](auto x) { return x == 1; });
+    }
+
     std::shared_ptr<shape_impl> copy() const { return std::make_shared<shape_impl>(*this); }
 };
 
@@ -135,6 +214,16 @@ shape::shape(type_t t, std::vector<std::size_t> l, std::vector<std::size_t> s)
 {
 }
 
+shape::shape(type_t t, std::initializer_list<std::size_t> d)
+    : shape::shape(t, std::vector<std::size_t>{d.begin(), d.end()})
+{
+}
+
+shape::shape(type_t t, std::vector<shape::dynamic_dimension> dims)
+    : impl(std::make_shared<shape_impl>(t, std::move(dims)))
+{
+}
+
 shape::shape(const std::vector<shape>& subs) : impl(std::make_shared<shape_impl>(subs)) {}
 
 shape::shape(std::shared_ptr<shape_impl> pimpl) : impl(std::move(pimpl)) {}
@@ -150,9 +239,13 @@ shape shape::from_permutation(type_t t,
 }
 
 shape::type_t shape::type() const { return impl->m_type; }
+
 const std::vector<std::size_t>& shape::lens() const { return impl->m_lens; }
+
 const std::vector<std::size_t>& shape::strides() const { return impl->m_strides; }
+
 std::size_t shape::elements() const { return impl->elements(); }
+
 std::size_t shape::bytes() const
 {
     if(this->sub_shapes().empty())
@@ -169,6 +262,7 @@ std::size_t shape::bytes() const
                                [&](auto x, auto y) { return x + y.bytes(); });
     }
 }
+
 std::size_t shape::type_size() const
 {
     std::size_t n = 0;
@@ -176,20 +270,35 @@ std::size_t shape::type_size() const
         this->visit_type([&](auto as) { n = as.size(); });
     return n;
 }
+
 std::size_t shape::index(std::initializer_list<std::size_t> l) const
 {
+    if(this->dynamic())
+    {
+        MIGRAPHX_THROW("SHAPE: index() called on dynamic shape");
+    }
     assert(l.size() <= this->lens().size());
     assert(this->lens().size() == this->strides().size());
     return std::inner_product(l.begin(), l.end(), this->strides().begin(), std::size_t{0});
 }
+
 std::size_t shape::index(const std::vector<std::size_t>& l) const
 {
+    if(this->dynamic())
+    {
+        MIGRAPHX_THROW("SHAPE: index() called on dynamic shape");
+    }
     assert(l.size() <= this->lens().size());
     assert(this->lens().size() == this->strides().size());
     return std::inner_product(l.begin(), l.end(), this->strides().begin(), std::size_t{0});
 }
+
 std::size_t shape::index(std::size_t i) const
 {
+    if(this->dynamic())
+    {
+        MIGRAPHX_THROW("SHAPE: index() called on dynamic shape");
+    }
     assert(this->lens().size() == this->strides().size());
     if(this->standard())
         return i;
@@ -237,11 +346,20 @@ void shape::multi_copy(std::size_t i, std::size_t* start, const std::size_t* end
 
 bool shape::packed() const
 {
-    return this->sub_shapes().empty() and this->elements() == this->element_space();
+    if(this->dynamic())
+    {
+        return false;
+    }
+    return this->sub_shapes().empty() and not impl->skips() and
+           this->elements() == this->element_space();
 }
 
 bool shape::transposed() const
 {
+    if(this->dynamic())
+    {
+        return false;
+    }
     if(this->broadcasted())
     {
         // TODO: Use a filter_iterator instead
@@ -261,15 +379,21 @@ bool shape::transposed() const
 
 bool shape::broadcasted() const
 {
+    if(this->dynamic())
+    {
+        return false;
+    }
     assert(this->lens().size() == this->strides().size());
-    return std::accumulate(this->strides().begin(),
-                           this->strides().end(),
-                           std::size_t{1},
-                           std::multiplies<std::size_t>()) == 0;
+    return std::any_of(
+        this->strides().begin(), this->strides().end(), [](auto x) { return x == 0; });
 }
 
 bool shape::scalar() const
 {
+    if(this->dynamic())
+    {
+        return false;
+    }
     assert(this->lens().size() == this->strides().size());
     // if any stride > 0, then accumulate will return false
     return this->sub_shapes().empty() and
@@ -288,6 +412,10 @@ shape shape::normalize_standard() const
 
 shape shape::with_lens(type_t t, const std::vector<std::size_t>& l) const
 {
+    if(this->dynamic())
+    {
+        MIGRAPHX_THROW("SHAPE: with_lens() called on dynamic shape");
+    }
     assert(l.size() == this->lens().size());
     auto perm = find_permutation(*this);
     return shape::from_permutation(t, l, perm);
@@ -295,6 +423,10 @@ shape shape::with_lens(type_t t, const std::vector<std::size_t>& l) const
 
 shape shape::with_lens(const std::vector<std::size_t>& l) const
 {
+    if(this->dynamic())
+    {
+        MIGRAPHX_THROW("SHAPE: with_lens() called on dynamic shape");
+    }
     return this->with_lens(this->type(), l);
 }
 
@@ -309,21 +441,81 @@ std::size_t shape::element_space() const { return impl->element_space(); }
 
 std::string shape::type_string() const { return name(this->type()); }
 
+bool shape::dynamic() const { return not impl->m_dyn_dims.empty(); }
+
+const std::vector<shape::dynamic_dimension>& shape::dyn_dims() const { return impl->m_dyn_dims; }
+
+std::vector<std::size_t> shape::min_lens() const
+{
+    return this->dynamic() ? impl->min_lens() : this->lens();
+}
+
+std::vector<std::size_t> shape::max_lens() const
+{
+    return this->dynamic() ? impl->max_lens() : this->lens();
+}
+
+std::vector<std::size_t> shape::opt_lens() const
+{
+    return this->dynamic() ? impl->opt_lens() : this->lens();
+}
+
+bool shape::dynamic_dimension::is_fixed() const { return this->min == this->max; }
+
+bool shape::dynamic_dimension::has_optimal() const { return opt != 0; }
+
+template <class Self, class F>
+auto shape::dynamic_dimension::reflect(Self& self, F f)
+{
+    return pack(f(self.min, "min"), f(self.max, "max"), f(self.opt, "opt"));
+}
+
+bool operator==(const shape::dynamic_dimension& x, const shape::dynamic_dimension& y)
+{
+    return (x.min == y.min and x.max == y.max and x.opt == y.opt);
+}
+
+bool operator!=(const shape::dynamic_dimension& x, const shape::dynamic_dimension& y)
+{
+    return not(x == y);
+}
+std::ostream& operator<<(std::ostream& os, const shape::dynamic_dimension& x)
+{
+    os << "[" << x.min << ", " << x.max << ", " << x.opt << "]";
+    return os;
+}
+
 bool operator==(const shape& x, const shape& y)
 {
-    return x.impl == y.impl or (x.type() == y.type() and x.lens() == y.lens() and
+    if(x.dynamic() and y.dynamic())
+    {
+        return x.impl == y.impl or (x.type() == y.type() and x.dyn_dims() == y.dyn_dims() and
+                                    x.sub_shapes() == y.sub_shapes());
+    }
+    return x.impl == y.impl or
+           (x.dynamic() == y.dynamic() and x.type() == y.type() and x.lens() == y.lens() and
             x.strides() == y.strides() and x.sub_shapes() == y.sub_shapes());
 }
-bool operator!=(const shape& x, const shape& y) { return !(x == y); }
+
+bool operator!=(const shape& x, const shape& y) { return not(x == y); }
 
 std::ostream& operator<<(std::ostream& os, const shape& x)
 {
     if(x.sub_shapes().empty())
+    {
+        if(x.dynamic())
+        {
+            os << "dynamic, ";
+            os << x.type_string() << ", ";
+            os << "{" << to_string_range(x.dyn_dims()) << "}";
+        }
+        else
         {
             os << x.type_string() << ", ";
             os << "{" << to_string_range(x.lens()) << "}, ";
             os << "{" << to_string_range(x.strides()) << "}";
         }
+    }
     else
     {
         os << "[" << to_string_range(x.sub_shapes()) << "]";
@@ -350,8 +542,10 @@ void migraphx_to_value(value& v, const shape& s)
     result["lens"]               = migraphx::to_value(s.lens());
     result["strides"]            = migraphx::to_value(s.strides());
     result["sub_shapes"]         = migraphx::to_value(s.sub_shapes());
+    result["dynamic_dimensions"] = migraphx::to_value(s.dyn_dims());
     v                            = result;
 }
+
 void migraphx_from_value(const value& v, shape& s)
 {
     auto t = v.at("type").get_string();
@@ -360,11 +554,27 @@ void migraphx_from_value(const value& v, shape& s)
         s = shape{migraphx::from_value<std::vector<migraphx::shape>>(v.at("sub_shapes"))};
     }
     else
+    {
+        if(v.at("dynamic_dimensions").empty())
         {
             s = shape{shape::parse_type(t),
                       v.at("lens").to_vector<std::size_t>(),
                       v.at("strides").to_vector<std::size_t>()};
         }
+        else
+        {
+            auto v_dd = v.at("dynamic_dimensions");
+            std::vector<shape::dynamic_dimension> dyn_dims(v.at("dynamic_dimensions").size());
+            std::transform(v_dd.begin(), v_dd.end(), dyn_dims.begin(), [](migraphx::value x) {
+                auto x_min = x.at("min").template to<size_t>();
+                auto x_max = x.at("max").template to<size_t>();
+                auto x_opt = x.at("opt").template to<size_t>();
+                return shape::dynamic_dimension{x_min, x_max, x_opt};
+            });
+
+            s = shape{shape::parse_type(t), dyn_dims};
+        }
+    }
 }
 
 } // namespace MIGRAPHX_INLINE_NS

src/simplify_algebra.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/simplify_algebra.hpp>
 #include <migraphx/dead_code_elimination.hpp>
 #include <migraphx/program.hpp>
@@ -185,6 +208,42 @@ struct find_mul_add
     }
 };
 
+struct find_dot_add
+{
+    auto matcher() const
+    {
+        return match::name("dot")(match::either_arg(0, 1)(
+            match::name("add")(
+                match::either_arg(0, 1)(match::any().bind("x"),
+                                        match::any_of(match::is_constant()).bind("b")),
+                match::none_of(match::args(match::is_constant(), match::is_constant())),
+                match::used_once()),
+            match::is_constant().bind("a")));
+    }
+
+    void apply(module& m, const match::matcher_result& r) const
+    {
+        auto ins   = r.result;
+        auto a_ins = r.instructions["a"];
+        auto b_ins = r.instructions["b"];
+        auto x_ins = r.instructions["x"];
+        assert(x_ins != b_ins);
+
+        const bool flipped = a_ins == ins->inputs().back();
+
+        auto insert_dot = [&](auto x, auto y) {
+            if(flipped)
+                return m.insert_instruction(ins, make_op("dot"), y, x);
+            else
+                return m.insert_instruction(ins, make_op("dot"), x, y);
+        };
+
+        auto ax_ins = insert_dot(a_ins, x_ins);
+        auto ab_ins = insert_dot(a_ins, b_ins);
+        m.replace_instruction(ins, make_op("add"), ax_ins, ab_ins);
+    }
+};
+
 struct find_add_lit_broadcast
 {
     auto matcher() const
@@ -244,28 +303,26 @@ struct find_double_add_lit_broadcast
 
 struct find_inner_broadcast
 {
-    auto matcher() const
-    {
-        return pointwise(
-            match::nargs(2),
-            match::args(match::name("broadcast").bind("x"), match::name("broadcast").bind("y")));
-    }
+    auto matcher() const { return pointwise(match::all_of[match::inputs()](match::broadcast())); }
 
     void apply(module& m, const match::matcher_result& r) const
     {
         auto ins        = r.result;
-        auto x_ins = r.instructions["x"];
-        auto y_ins = r.instructions["y"];
-
-        auto xbroadcast = any_cast<op::broadcast>(x_ins->get_operator());
-        auto ybroadcast = any_cast<op::broadcast>(y_ins->get_operator());
-
-        if(xbroadcast.axis != ybroadcast.axis)
+        auto broadcasts = ins->inputs();
+        if(broadcasts.empty())
+            return;
+        std::vector<instruction_ref> inputs;
+        std::transform(broadcasts.begin(),
+                       broadcasts.end(),
+                       std::back_inserter(inputs),
+                       [](auto i) { return i->inputs().front(); });
+        if(std::any_of(inputs.begin(), inputs.end(), [&](auto i) {
+               return i->get_shape() != inputs.front()->get_shape();
+           }))
             return;
 
-        auto op = m.insert_instruction(
-            ins, ins->get_operator(), x_ins->inputs().front(), y_ins->inputs().front());
-        m.replace_instruction(ins, xbroadcast, op);
+        auto op = m.insert_instruction(ins, ins->get_operator(), inputs);
+        m.replace_instruction(ins, broadcasts.front()->get_operator(), op);
     }
 };
 
@@ -393,8 +450,9 @@ struct find_splits
 {
     auto matcher() const
     {
-        return match::any(match::any_of[match::outputs()](match::name("slice")(
-            match::any_of[match::outputs()](match::pointwise(), reduction()))));
+        return match::any(
+            match::any_of[match::outputs()](match::name("slice")(match::any_of[match::outputs()](
+                match::pointwise(match::any_of(match::nargs(1), match::nargs(2))), reduction()))));
     }
 
     static bool is_dependent(const module& m, instruction_ref ins1, instruction_ref ins2)
@@ -557,10 +615,9 @@ struct find_splits
                     auto outputs = i->outputs();
                     for(auto output : outputs)
                     {
-                        if(not contains({"reshape", "squeeze", "unsqueeze"}, output->name()))
+                        if(output->name() != "reshape")
                             continue;
-                        auto x =
-                            m.insert_instruction(output, make_op("contiguous"), output->inputs());
+                        auto x = m.insert_instruction(output, make_op("contiguous"), i);
                         m.replace_instruction(output, output->get_operator(), x);
                     }
 
@@ -730,7 +787,7 @@ MIGRAPHX_PRED_MATCHER(horiz_conv_dot, instruction_ref ins)
     };
     auto dots  = std::count_if(ins->outputs().begin(), ins->outputs().end(), pred("dot"));
     auto convs = std::count_if(ins->outputs().begin(), ins->outputs().end(), pred("convolution"));
-    return !(dots < 2 and convs < 2);
+    return not(dots < 2 and convs < 2);
 }
 
 struct find_conv_dot_horiz_fusion
@@ -750,7 +807,7 @@ struct find_conv_dot_horiz_fusion
             auto y = j->inputs()[1]->get_shape().lens();
             if(x.size() != y.size())
                 return false;
-            // Check that non-axises match
+            // Check that non-axes match
             int axis = 1;
             if(i->name() == "dot")
             {
@@ -786,13 +843,22 @@ struct find_conv_dot_horiz_fusion
 
             for(auto arg : args)
                 m.move_instructions(arg, input);
-            // TODO: Check if axises match
+            // TODO: Check if axes match
             auto concat =
                 m.insert_instruction(input, make_op("concat", {{"axis", concat_axis}}), args);
             auto fused     = m.insert_instruction(std::next(input), op, input, concat);
             int64_t offset = 0;
             for(auto arg : range(start, last))
             {
+                auto outputs = arg->outputs();
+                for(auto output : outputs)
+                {
+                    if(output->name() != "reshape")
+                        continue;
+                    auto x = m.insert_instruction(output, make_op("contiguous"), arg);
+                    m.replace_instruction(output, output->get_operator(), x);
+                }
+
                 int64_t len = arg->get_shape().lens()[axis];
                 m.replace_instruction(
                     arg,
@@ -903,7 +969,7 @@ struct find_split_reshape
 
         // all outputs are reshape and of the same shape
         auto dims = any_cast<op::reshape>(rsp->get_operator()).dims;
-        if(!same_ops(vec_rsp))
+        if(not same_ops(vec_rsp))
         {
             return;
         }
@@ -919,23 +985,42 @@ struct find_split_reshape
         auto rsp_lens    = rsp->get_shape().lens();
         auto rsp_strides = rsp->get_shape().strides();
         rsp_strides.insert(rsp_strides.begin(), rsp_strides[0] * rsp_lens[0]);
+
         auto ait     = std::find(rsp_strides.begin(), rsp_strides.end(), slc_dim_size);
+        int rsp_axis = -1;
         if(ait == rsp_strides.end())
         {
             return;
         }
-        int rsp_axis = std::distance(rsp_strides.begin(), ait);
-
+        else if(ait == rsp_strides.end() - 1)
+        {
+            // edge case
+            // slice_dim == 1, in that case it could match with last stride of 1.
+            // it should accumulate lengths from last dim in that case. discount 1 to avoid going
+            // out of bounds.
+            assert(slc_dim_size == 1);
+            rsp_axis = std::distance(rsp_strides.begin(), ait) - 1;
+        }
+        else
+        {
+            rsp_axis = std::distance(rsp_strides.begin(), ait);
+        }
         // calculate reshape output shape
         std::vector<int64_t> vec_dims(vec_rsp.size());
+
         std::transform(vec_rsp.begin(), vec_rsp.end(), vec_dims.begin(), [&](auto is) {
             return is->get_shape().lens()[rsp_axis];
         });
 
         std::vector<int64_t> rsp_out_lens(rsp_lens.begin(), rsp_lens.end());
+
         rsp_out_lens[rsp_axis] = std::accumulate(vec_dims.begin(), vec_dims.end(), std::int64_t{0});
 
-        // insert the reshape instruction
+        // insert the reshape instruction and add contiguous if needed
+        if(not input->get_shape().standard())
+        {
+            input = m.insert_instruction(std::next(input), make_op("contiguous"), input);
+        }
         auto rsp_ins = m.insert_instruction(
             std::next(input), make_op("reshape", {{"dims", rsp_out_lens}}), input);
 
@@ -982,7 +1067,7 @@ struct find_split_transpose
 
         // all transpose are the same
         auto perm = any_cast<op::transpose>(trans->get_operator()).dims;
-        if(!same_ops(vec_trans))
+        if(not same_ops(vec_trans))
         {
             return;
         }
@@ -1025,6 +1110,7 @@ void simplify_algebra::apply(module& m) const
                             find_mul_conv{},
                             find_mul_slice_conv{},
                             find_mul_add{},
+                            find_dot_add{},
                             find_div_const{},
                             find_sub_const{},
                             find_rsqrt{},

src/simplify_qdq.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/simplify_qdq.hpp>
 #include <migraphx/instruction.hpp>
 #include <migraphx/iterator_for.hpp>

src/simplify_reshapes.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 <iterator>
 #include <migraphx/simplify_reshapes.hpp>
 #include <migraphx/program.hpp>
@@ -76,7 +99,7 @@ struct find_reshaper
         std::vector<instruction_ref> reshapes{ins};
         while(is_reshaper(reshapes.back()))
         {
-            assert(!reshapes.back()->inputs().empty());
+            assert(not reshapes.back()->inputs().empty());
             assert(m.has_instruction(reshapes.back()->inputs().front()));
             auto input = reshapes.back()->inputs().front();
             reshapes.push_back(input);
@@ -128,8 +151,11 @@ struct find_transpose
 {
     auto matcher() const
     {
-        return match::name("transpose")(match::none_of(
-            match::skip_output(match::name("contiguous"))(match::name("transpose"))));
+        auto output_not_transpose =
+            match::none_of(match::skip_output(match::name("contiguous"))(match::name("transpose")));
+        auto input_has_transpose =
+            match::args(match::skip(match::name("contiguous"))(match::name("transpose")));
+        return match::name("transpose")(output_not_transpose, input_has_transpose);
     }
 
     void apply(module& m, const match::matcher_result& mr) const
@@ -245,11 +271,49 @@ struct find_nested_slice
     }
 };
 
+struct find_concat_multibroadcasts
+{
+    auto matcher() const
+    {
+        return match::name("concat")(match::all_of[match::inputs()](match::name("multibroadcast")));
+    }
+
+    void apply(module& m, const match::matcher_result& mr) const
+    {
+        auto ins        = mr.result;
+        auto op         = any_cast<op::concat>(ins->get_operator());
+        auto out_lens   = ins->get_shape().lens();
+        auto inputs     = ins->inputs();
+        auto in_strides = inputs.front()->get_shape().strides();
+
+        // Only apply when concat axis is not a broadcasted dimension
+        if(std::any_of(inputs.begin(), inputs.end(), [&](auto i) {
+               return i->get_shape().strides()[op.axis] == 0;
+           }))
+        {
+            return;
+        }
+
+        // Use inputs of multibroadcast ops as inputs to new concat op
+        std::transform(inputs.begin(), inputs.end(), inputs.begin(), [](auto i) {
+            return i->inputs().front();
+        });
+
+        // Reduce axis by number of leading broadcasted dimensions
+        if(inputs.front()->get_shape().lens().size() < out_lens.size())
+            op.axis -= std::count(in_strides.begin(), in_strides.begin() + op.axis, 0);
+
+        auto concat = m.insert_instruction(ins, op, inputs);
+        m.replace_instruction(
+            ins, migraphx::make_op("multibroadcast", {{"out_lens", out_lens}}), concat);
+    }
+};
+
 struct find_concat_transpose
 {
     auto matcher() const
     {
-        return match::name("concat")(match::all_of[match::inputs()](match::transpose_shape()));
+        return match::name("concat")(match::all_of[match::inputs()](match::name("transpose")));
     }
 
     void apply(module& m, const match::matcher_result& mr) const
@@ -262,7 +326,7 @@ struct find_concat_transpose
         auto permutation = find_permutation(s);
 
         // permutation should be the same for all inputs
-        if(!std::all_of(trans_inputs.begin(), trans_inputs.end(), [&](auto in) {
+        if(not std::all_of(trans_inputs.begin(), trans_inputs.end(), [&](auto in) {
                return (find_permutation(in->get_shape()) == permutation);
            }))
         {
@@ -578,9 +642,157 @@ struct find_transpose_contiguous_reshaper_unary
     }
 };
 
+struct find_slice_transpose
+{
+    auto matcher() const
+    {
+        return match::any(match::any_of[match::outputs()](
+            match::name("slice")(match::output(match::name("transpose")))));
+    }
+
+    static std::vector<int64_t> find_common_perm(const std::vector<instruction_ref>& transposes)
+    {
+        std::map<std::vector<int64_t>, int64_t> count;
+        for(auto t : transposes)
+        {
+            auto perm = t->get_operator().to_value()["permutation"].to_vector<int64_t>();
+            count[perm]++;
+        }
+        return std::max_element(
+                   count.begin(), count.end(), by(std::less<>{}, [](auto&& p) { return p.second; }))
+            ->first;
+    }
+
+    void apply(module& m, const match::matcher_result& r) const
+    {
+        auto ins = r.result;
+        std::vector<instruction_ref> splits;
+        std::copy_if(ins->outputs().begin(),
+                     ins->outputs().end(),
+                     std::back_inserter(splits),
+                     [&](instruction_ref out) {
+                         return out->name() == "slice" and out->outputs().size() == 1 and
+                                out->outputs().front()->name() == "transpose";
+                     });
+        if(splits.size() < 2)
+            return;
+        std::vector<instruction_ref> transposes;
+        std::transform(splits.begin(),
+                       splits.end(),
+                       std::back_inserter(transposes),
+                       [](auto split) { return split->outputs().front(); });
+        auto perm  = find_common_perm(transposes);
+        auto iperm = invert_permutation(perm);
+        auto pre   = m.insert_instruction(
+            std::next(ins), make_op("transpose", {{"permutation", perm}}), ins);
+        for(auto i : range(transposes.size()))
+        {
+            auto split = splits[i];
+            auto t     = transposes[i];
+            auto op    = any_cast<op::slice>(split->get_operator());
+            std::transform(op.axes.begin(), op.axes.end(), op.axes.begin(), [&](auto axis) {
+                return iperm[axis];
+            });
+            auto new_ins = m.insert_instruction(t, op, pre);
+            if(t->get_operator() != pre->get_operator())
+            {
+                auto curr = t->get_operator().to_value()["permutation"].to_vector<int64_t>();
+                new_ins   = m.insert_instruction(
+                    t, make_op("transpose", {{"permutation", reorder_dims(iperm, curr)}}), new_ins);
+            }
+            m.replace_instruction(t, new_ins);
+        }
+    }
+};
+
+struct find_transpose_slice
+{
+    auto matcher() const
+    {
+        return match::name("transpose")(match::all_of[match::outputs()](match::name("slice")));
+    }
+
+    static std::vector<int64_t> slice_distance(const op::slice& op)
+    {
+        assert(op.starts.size() == op.ends.size());
+        std::vector<int64_t> result(op.starts.size());
+        std::transform(
+            op.ends.begin(), op.ends.end(), op.starts.begin(), result.begin(), std::minus<>{});
+        return result;
+    }
+
+    void apply(module& m, const match::matcher_result& r) const
+    {
+        auto ins    = r.result;
+        auto slices = ins->outputs();
+        if(slices.empty())
+            return;
+        auto slice     = any_cast<op::slice>(slices.front()->get_operator());
+        auto sdistance = slice_distance(slice);
+        // Check all distances and axes are the same
+        if(std::any_of(slices.begin(), slices.end(), [&](auto sins) {
+               auto s = any_cast<op::slice>(sins->get_operator());
+               return s.axes != slice.axes or slice_distance(s) != sdistance;
+           }))
+            return;
+        // Check distances are divisible by lens of corresponding axes
+        auto mod_by_distance = [&](const auto& v, auto f) {
+            return std::inner_product(v.begin(),
+                                      v.end(),
+                                      sdistance.begin(),
+                                      0,
+                                      std::plus<>{},
+                                      [&](auto x, auto d) -> uint64_t {
+                                          if(d == 0)
+                                              return 1;
+                                          return f(x) % d;
+                                      });
+        };
+        if(mod_by_distance(slice.axes, [&](auto x) { return ins->get_shape().lens()[x]; }) != 0 or
+           mod_by_distance(slice.starts, id{}) != 0 or mod_by_distance(slice.ends, id{}) != 0)
+            return;
+        // TODO: Handle multiple axes
+        if(sdistance.size() != 1)
+            return;
+        auto axis = slice.axes.front();
+        // Skip if axis would be packed
+        if(std::all_of(ins->get_shape().lens().begin(),
+                       ins->get_shape().lens().begin() + axis,
+                       [](auto x) { return x == 1; }))
+            return;
+        // Compute axis before transpose to use for unsqueeze
+        auto perm    = ins->get_operator().to_value()["permutation"].to_vector<int64_t>();
+        auto preaxis = std::find(perm.begin(), perm.end(), axis) - perm.begin();
+        // Make unsqeeze
+        auto unsqueeze = m.insert_instruction(
+            ins, make_op("unsqueeze", {{"axes", {preaxis}}, {"steps", sdistance}}), ins->inputs());
+        // Make transpose
+        std::transform(perm.begin(), perm.end(), perm.begin(), [&](auto i) {
+            if(i > preaxis)
+                return i + 1;
+            return i;
+        });
+        perm.insert(perm.begin(), preaxis + 1);
+        auto transpose =
+            m.insert_instruction(ins, make_op("transpose", {{"permutation", perm}}), unsqueeze);
+        // Slice and squeeze
+        for(auto s : slices)
+        {
+            auto op        = any_cast<op::slice>(s->get_operator());
+            op.axes        = {0};
+            op.starts      = {op.starts.front() / sdistance.front()};
+            op.ends        = {op.ends.front() / sdistance.front()};
+            auto slice_ins = m.insert_instruction(ins, op, transpose);
+            auto squeeze =
+                m.insert_instruction(ins, make_op("squeeze", {{"axes", {0}}}), slice_ins);
+            m.replace_instruction(s, squeeze);
+        }
+    }
+};
+
 void simplify_reshapes::apply(module& m) const
 {
-    for(int i = 0; i < 2; i++)
+    for(int i = 0; i < 4; i++)
     {
         match::find_matches(m,
                             find_where_op{},
@@ -590,9 +802,12 @@ void simplify_reshapes::apply(module& m) const
                             find_reshaper{},
                             find_transpose{},
                             find_concat_transpose{},
+                            find_concat_multibroadcasts{},
                             find_nested_convert{},
                             find_nested_slice{},
                             find_nested_concat{},
+                            find_transpose_slice{},
+                            find_slice_transpose{},
                             find_transpose_contiguous_reshaper_unary{});
         dead_code_elimination{}.apply(m);
     }

src/sqlite.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/sqlite.hpp>
+#include <migraphx/manage_ptr.hpp>
+#include <migraphx/errors.hpp>
+#include <sqlite3.h>
+#include <algorithm>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+
+using sqlite3_ptr = MIGRAPHX_MANAGE_PTR(sqlite3*, sqlite3_close);
+
+struct sqlite_impl
+{
+    sqlite3* get() const { return ptr.get(); }
+    void open(const fs::path& p, int flags)
+    {
+        sqlite3* ptr_tmp = nullptr;
+        int rc           = sqlite3_open_v2(p.string().c_str(), &ptr_tmp, flags, nullptr);
+        ptr              = sqlite3_ptr{ptr_tmp};
+        if(rc != 0)
+            MIGRAPHX_THROW("error opening " + p.string() + ": " + error_message());
+    }
+
+    template <class F>
+    void exec(const char* sql, F f)
+    {
+        auto callback = [](void* obj, auto... xs) -> int {
+            try
+            {
+                const auto* g = static_cast<const F*>(obj);
+                (*g)(xs...);
+                return 0;
+            }
+            catch(...)
+            {
+                return -1;
+            }
+        };
+        int rc = sqlite3_exec(get(), sql, callback, &f, nullptr);
+        if(rc != 0)
+            MIGRAPHX_THROW(error_message());
+    }
+
+    std::string error_message() const
+    {
+        std::string msg = "sqlite3: ";
+        return msg + sqlite3_errmsg(get());
+    }
+    sqlite3_ptr ptr;
+};
+
+sqlite sqlite::read(const fs::path& p)
+{
+    sqlite r;
+    r.impl = std::make_shared<sqlite_impl>();
+    r.impl->open(p, SQLITE_OPEN_READONLY);
+    return r;
+}
+
+sqlite sqlite::write(const fs::path& p)
+{
+    sqlite r;
+    r.impl = std::make_shared<sqlite_impl>();
+    // Using '+' instead of bitwise '|' to avoid compilation warning
+    r.impl->open(p, SQLITE_OPEN_READWRITE + SQLITE_OPEN_CREATE);
+    return r;
+}
+
+std::vector<std::unordered_map<std::string, std::string>> sqlite::execute(const std::string& s)
+{
+    std::vector<std::unordered_map<std::string, std::string>> result;
+    impl->exec(s.c_str(), [&](int n, char** texts, char** names) {
+        std::unordered_map<std::string, std::string> row;
+        row.reserve(n);
+        std::transform(
+            names,
+            names + n,
+            texts,
+            std::inserter(row, row.begin()),
+            [&](const char* name, const char* text) { return std::make_pair(name, text); });
+        result.push_back(row);
+    });
+    return result;
+}
+
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx

src/targets/cpu/CMakeLists.txt

+#####################################################################################
+# 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(CheckCXXCompilerFlag)
 

src/targets/cpu/allocate.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/config.hpp>
 #include <migraphx/check_shapes.hpp>
 #include <migraphx/argument.hpp>

src/targets/cpu/allocation_model.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/cpu/allocation_model.hpp>
 #include <migraphx/make_op.hpp>
 

src/targets/cpu/binary.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/config.hpp>
 #include <migraphx/cpu/dnnl.hpp>
 

src/targets/cpu/concat.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/config.hpp>
 #include <migraphx/cpu/pointwise.hpp>
 #include <migraphx/op/concat.hpp>

src/targets/cpu/convolution.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/config.hpp>
 #include <migraphx/register_op.hpp>
 #include <migraphx/reflect.hpp>

src/targets/cpu/copy.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/config.hpp>
 #include <migraphx/cpu/pointwise.hpp>