Merge

doc/src/driver/compile.rst

@@ -32,6 +32,10 @@ Disable fast math optimization
 
 Perform an exhaustive search to find the fastest version of generated kernels for selected backend
 
+.. options:: --split-single-dyn-dim
+
+Enable the split single dynamic dimension pass
+
 .. option::  --fp16
 
 Quantize for fp16

doc/src/driver/read.rst

@@ -24,7 +24,7 @@ Load as MIGraphX JSON
 
 .. option::  --batch [unsigned int] (Default: 1)
 
-Set batch size for model
+For a static model, set batch size. For a dynamic batch model, sets the batch size at runtime.
 
 .. option::  --nhwc
 
@@ -46,6 +46,14 @@ Trim instructions from the end (Default: 0)
 
 Dim of a parameter (format: "@name d1 d2 dn")
 
+.. options:: --dyn-input-dim [std::vector<std::string>]
+
+Set dynamic dimensions of a parameter using JSON formatting (format "@name" "dynamic_dimension_json")
+
+.. options:: --default-dyn-dim
+
+Set the default dynamic dimension (format {min:x, max:y, optimals:[o1,o2,...]})
+
 .. option::  --optimize, -O
 
 Optimize when reading

examples/migraphx/migraphx_driver/README.md

@@ -29,7 +29,7 @@ See below for a comprehensive list of commands and option arguments, as well as
 | --tf                                     | Load file as a tensorflow graph                           |
 | --migraphx                               | Load file as a migraphx graph                             |
 | --migraphx-json                          | Load file as a migraphx JSON graph                        |
-| --batch                                  | Set batch size for the model                              |
+| --batch                                  | For a static model, set batch size. For a dynamic batch model, sets the batch size at runtime.|
 | --nhwc                                   | Treat tensorflow format as nhwc                           |
 | --nchw                                   | Treat tensorflow format as nchw                           |
 | --skip-unknown-operators                 | Skip unknown operators when parsing and continue to parse |
@@ -44,12 +44,16 @@ See below for a comprehensive list of commands and option arguments, as well as
 | --output \| -o                           | Output to file                                            |
 | --fill0                                  | Fill parameter with 0s                                    |
 | --fill1                                  | Fill parameter with 1s                                    |
+| --input-dim                              | Set static dimensions of a parameter                      |
+| --dyn-input-dim                          | Set dynamic dimensions of a parameter                     |
+| --default-dyn-dim                        | Set default dynamic dimension                             |
 | --gpu                                    | Compile on the gpu                                        |
 | --cpu                                    | Compile on the cpu                                        |
 | --ref                                    | Compile on the reference implementation                   |
 | --enable-offload-copy                    | Enable implicit offload copying                           |
 | --disable-fast-math                      | Disable fast math optimization                            |
 | --exhaustive-tune                        | Enable exhaustive search to find fastest kernel           |
+| --split-single-dyn-dim                   | Enable split_single_dyn_dim compiler pass                 |
 | --fp16                                   | Quantize for fp16                                         |
 | --int8                                   | Quantize for int8                                         |
 | --tolerance                              | Tolerance for errors                                      |

src/CMakeLists.txt

@@ -50,6 +50,7 @@ add_library(migraphx
     env.cpp
     file_buffer.cpp
     fuse_pointwise.cpp
+    fuse_reduce.cpp
     generate.cpp
     inline_module.cpp
     insert_pad.cpp
@@ -73,6 +74,7 @@ add_library(migraphx
     process.cpp
     program.cpp
     propagate_constant.cpp
+    promote_literals.cpp
     quantization.cpp
     quantize_fp16.cpp
     quantize_int8.cpp
@@ -91,6 +93,7 @@ add_library(migraphx
     shape.cpp
     simplify_algebra.cpp
     simplify_reshapes.cpp
+    split_single_dyn_dim.cpp
     tmp_dir.cpp
     value.cpp
     verify_args.cpp

src/common.cpp

@@ -89,8 +89,8 @@ std::vector<shape::dynamic_dimension> compute_broadcasted_dyn_dims(shape s0, sha
             }
             else if(a == 1 or b == 1)
             {
-                // setting opt to 0, may need to be changed
-                return shape::dynamic_dimension{std::max(a.min, b.min), std::max(a.max, b.max), 0};
+                // setting optimals to empty, may need to be changed
+                return shape::dynamic_dimension{std::max(a.min, b.min), std::max(a.max, b.max)};
             }
             else
             {
@@ -148,10 +148,8 @@ shape common_shape(const std::vector<shape>& shapes)
     return {compute_common_types(shapes), compute_common_lens(shapes)};
 }
 
-instruction_ref insert_common_op(module& m,
-                                 instruction_ref ins,
-                                 const operation& op,
-                                 std::vector<instruction_ref> inputs)
+std::vector<instruction_ref>
+insert_common_args(module& m, instruction_ref ins, std::vector<instruction_ref> inputs)
 {
     if(std::any_of(
            inputs.cbegin(), inputs.cend(), [](auto input) { return input->get_shape().dynamic(); }))
@@ -210,7 +208,20 @@ instruction_ref insert_common_op(module& m,
             return input;
         });
     }
-    return m.insert_instruction(ins, op, inputs);
+    return inputs;
+}
+
+std::vector<instruction_ref> add_common_args(module& m, std::vector<instruction_ref> inputs)
+{
+    return insert_common_args(m, m.end(), std::move(inputs));
+}
+
+instruction_ref insert_common_op(module& m,
+                                 instruction_ref ins,
+                                 const operation& op,
+                                 std::vector<instruction_ref> inputs)
+{
+    return m.insert_instruction(ins, op, insert_common_args(m, ins, std::move(inputs)));
 }
 
 instruction_ref add_common_op(module& m, const operation& op, std::vector<instruction_ref> inputs)

src/cpp_generator.cpp

@@ -106,6 +106,13 @@ cpp_generator::function& cpp_generator::function::set_generic_types(const module
     return *this;
 }
 
+cpp_generator::function& cpp_generator::function::add_generic_param(const std::string& pname)
+{
+    params.push_back({pname, "T" + pname});
+    tparams.push_back("class T" + pname);
+    return *this;
+}
+
 struct cpp_generator_impl
 {
     std::stringstream fs{};
@@ -182,7 +189,8 @@ std::string cpp_generator::generate_point_op(const operation& op,
 
 std::string cpp_generator::str() const { return impl->fs.str(); }
 
-cpp_generator::function cpp_generator::generate_module(const module& m)
+cpp_generator::function cpp_generator::generate_module(const module& m,
+                                                       const generate_module_callback& g)
 {
     function f;
     auto name = transform_string(m.name(), [](char c) {
@@ -195,13 +203,7 @@ cpp_generator::function cpp_generator::generate_module(const module& m)
             if(ins->name() == "@literal")
                 return shape::cpp_type(ins->get_shape().type()) + "(" +
                        ins->get_literal().to_string() + ")";
-            std::vector<std::string> args;
-            std::transform(ins->inputs().begin(),
-                           ins->inputs().end(),
-                           std::back_inserter(args),
-                           [&](auto i) { return names.at(i); });
-
-            auto s = this->generate_point_op(ins->get_operator(), args);
+            auto s = g(ins, names);
             if(impl->fresult)
                 return impl->fresult(ins->get_shape()) + '(' + s + ')';
             else
@@ -210,6 +212,24 @@ cpp_generator::function cpp_generator::generate_module(const module& m)
     return f;
 }
 
+std::vector<std::string>
+cpp_generator::to_args(const std::vector<instruction_ref>& inputs,
+                       const std::unordered_map<instruction_ref, std::string>& names)
+{
+    std::vector<std::string> args;
+    std::transform(inputs.begin(), inputs.end(), std::back_inserter(args), [&](auto i) {
+        return names.at(i);
+    });
+    return args;
+}
+
+cpp_generator::function cpp_generator::generate_module(const module& m)
+{
+    return this->generate_module(m, [&](auto ins, const auto& names) {
+        return this->generate_point_op(ins->get_operator(), to_args(ins->inputs(), names));
+    });
+}
+
 std::string cpp_generator::create_function(const cpp_generator::function& f)
 {
     impl->function_count++;

src/driver/argument_parser.hpp

@@ -148,13 +148,21 @@ struct value_parser
     template <MIGRAPHX_REQUIRES(not std::is_enum<T>{} and not is_multi_value<T>{})>
     static T apply(const std::string& x)
     {
-        T result;
-        std::stringstream ss;
-        ss.str(x);
-        ss >> result;
-        if(ss.fail())
-            throw std::runtime_error("Failed to parse '" + x + "' as " + type_name<T>::apply());
-        return result;
+        // handle whitespace in string
+        if constexpr(std::is_same<T, std::string>{})
+        {
+            return x;
+        }
+        else
+        {
+            T result;
+            std::stringstream ss;
+            ss.str(x);
+            ss >> result;
+            if(ss.fail())
+                throw std::runtime_error("Failed to parse '" + x + "' as " + type_name<T>::apply());
+            return result;
+        }
     }
 
     template <MIGRAPHX_REQUIRES(std::is_enum<T>{} and not is_multi_value<T>{})>

src/driver/main.cpp

@@ -33,6 +33,7 @@
 #include <migraphx/tf.hpp>
 #include <migraphx/onnx.hpp>
 #include <migraphx/stringutils.hpp>
+#include <migraphx/convert_to_json.hpp>
 #include <migraphx/load_save.hpp>
 #include <migraphx/json.hpp>
 #include <migraphx/version.h>
@@ -68,7 +69,9 @@ struct loader
     bool brief                  = false;
     std::string output_type;
     std::string output;
+    std::string default_dyn_dim;
     std::vector<std::string> param_dims;
+    std::vector<std::string> dyn_param_dims;
     std::vector<std::string> output_names;
 
     void parse(argument_parser& ap)
@@ -83,7 +86,11 @@ struct loader
         ap(file_type, {"--tf"}, ap.help("Load as tensorflow"), ap.set_value("tf"));
         ap(file_type, {"--migraphx"}, ap.help("Load as MIGraphX"), ap.set_value("migraphx"));
         ap(file_type, {"--migraphx-json"}, ap.help("Load as MIGraphX JSON"), ap.set_value("json"));
-        ap(batch, {"--batch"}, ap.help("Set batch size for model"));
+        ap(batch,
+           {"--batch"},
+           ap.help("For a static model, sets default_dim_value size (commonly batch size). For a "
+                   "dynamic batch model, sets the batch "
+                   "size at runtime."));
         ap(is_nhwc, {"--nhwc"}, ap.help("Treat tensorflow format as nhwc"), ap.set_value(true));
         ap(skip_unknown_operators,
            {"--skip-unknown-operators"},
@@ -96,7 +103,16 @@ struct loader
            ap.help("Dim of a parameter (format: \"@name d1 d2 dn\")"),
            ap.append(),
            ap.nargs(2));
-
+        ap(dyn_param_dims,
+           {"--dyn-input-dim"},
+           ap.help("Dynamic dimensions of a parameter (format: \"@name_1\" \"[{min:x, max:y, "
+                   "optimals:[o1,o2,...]}, dim2,dim3, ...]\", \"@name_2\", ... You can supply a "
+                   "single integer value for a dimension to specify it as fixed."),
+           ap.append(),
+           ap.nargs(2));
+        ap(default_dyn_dim,
+           {"--default-dyn-dim"},
+           ap.help("Default dynamic dimension (format: \"{min:x, max:y, optimals:[o1,o2]}\")."));
         ap(output_names,
            {"--output-names"},
            ap.help("Names of node output (format: \"name_1 name_2 name_n\")"),
@@ -147,6 +163,40 @@ struct loader
         return map_input_dims;
     }
 
+    static auto parse_dyn_dims_json(const std::string& dd_json)
+    {
+        // expecting a json string like "[{min:1,max:64,optimals:[1,2,4,8]},3,224,224]"
+        auto v = from_json_string(convert_to_json(dd_json));
+        std::vector<migraphx::shape::dynamic_dimension> dyn_dims;
+        std::transform(v.begin(), v.end(), std::back_inserter(dyn_dims), [&](auto x) {
+            if(x.is_object())
+                return from_value<migraphx::shape::dynamic_dimension>(x);
+            auto d = x.template to<std::size_t>();
+            return migraphx::shape::dynamic_dimension{d, d};
+        });
+        return dyn_dims;
+    }
+
+    static auto parse_dyn_dims_map(const std::vector<std::string>& param_dyn_dims)
+    {
+        // expecting vector of strings formatted like
+        // {"@param_name_0", "dd_json_0", "@param_name_1", "dd_json_1", ...}
+        std::unordered_map<std::string, std::vector<shape::dynamic_dimension>> map_dyn_input_dims;
+        std::string name = "";
+        for(auto&& x : param_dyn_dims)
+        {
+            if(x[0] == '@')
+            {
+                name = x.substr(1);
+            }
+            else
+            {
+                map_dyn_input_dims[name] = parse_dyn_dims_json(x);
+            }
+        }
+        return map_dyn_input_dims;
+    }
+
     static auto parse_output_names(const std::vector<std::string>& output_names_info)
     {
         std::vector<std::string> output_node_names;
@@ -158,13 +208,44 @@ struct loader
         return output_node_names;
     }
 
+    tf_options get_tf_options() const
+    {
+        auto map_input_dims    = parse_param_dims(param_dims);
+        auto output_node_names = parse_output_names(output_names);
+        tf_options options;
+        options.is_nhwc           = is_nhwc;
+        options.batch_size        = batch;
+        options.map_input_dims    = map_input_dims;
+        options.output_node_names = output_node_names;
+        return options;
+    }
+
+    onnx_options get_onnx_options() const
+    {
+        auto map_input_dims     = parse_param_dims(param_dims);
+        auto map_dyn_input_dims = parse_dyn_dims_map(dyn_param_dims);
+        onnx_options options;
+        if(default_dyn_dim.empty())
+        {
+            options.default_dim_value = batch;
+        }
+        else
+        {
+            auto v                        = from_json_string(convert_to_json(default_dyn_dim));
+            options.default_dyn_dim_value = from_value<migraphx::shape::dynamic_dimension>(v);
+        }
+        options.skip_unknown_operators = skip_unknown_operators;
+        options.print_program_on_error = true;
+        options.map_input_dims         = map_input_dims;
+        options.map_dyn_input_dims     = map_dyn_input_dims;
+        return options;
+    }
+
     program load()
     {
         program p;
         if(model.empty())
         {
-            auto map_input_dims    = parse_param_dims(param_dims);
-            auto output_node_names = parse_output_names(output_names);
             if(file_type.empty())
             {
                 if(ends_with(file, ".onnx"))
@@ -179,16 +260,11 @@ struct loader
             std::cout << "Reading: " << file << std::endl;
             if(file_type == "onnx")
             {
-                onnx_options options;
-                options.default_dim_value      = batch;
-                options.skip_unknown_operators = skip_unknown_operators;
-                options.print_program_on_error = true;
-                options.map_input_dims         = map_input_dims;
-                p                              = parse_onnx(file, options);
+                p = parse_onnx(file, get_onnx_options());
             }
             else if(file_type == "tf")
             {
-                p = parse_tf(file, tf_options{is_nhwc, batch, map_input_dims, output_node_names});
+                p = parse_tf(file, get_tf_options());
             }
             else if(file_type == "json")
             {
@@ -289,14 +365,21 @@ struct program_params
         ap(fill1, {"--fill1"}, ap.help("Fill parameter with 1s"), ap.append(), ap.nargs(2));
     }
 
-    auto generate(const program& p, const target& t, bool offload)
+    auto generate(const program& p, const target& t, bool offload, unsigned batch)
     {
         parameter_map m;
+        auto param_shapes = p.get_parameter_shapes();
+        std::unordered_map<std::string, shape> static_param_shapes;
+        std::transform(
+            param_shapes.cbegin(),
+            param_shapes.cend(),
+            std::inserter(static_param_shapes, static_param_shapes.end()),
+            [&](const auto& x) { return std::make_pair(x.first, x.second.to_static(batch)); });
         for(auto&& s : fill0)
-            m[s] = fill_argument(p.get_parameter_shape(s), 0);
+            m[s] = fill_argument(static_param_shapes.at(s), 0);
         for(auto&& s : fill1)
-            m[s] = fill_argument(p.get_parameter_shape(s), 1);
-        fill_param_map(m, p, t, offload);
+            m[s] = fill_argument(static_param_shapes.at(s), 1);
+        fill_param_map(m, static_param_shapes, t, offload);
         return m;
     }
 };
@@ -353,13 +436,18 @@ struct compiler
            {"--exhaustive-tune"},
            ap.help("Exhastively search for best tuning parameters for kernels"),
            ap.set_value(true));
+        ap(co.split_single_dyn_dim,
+           {"--split-single-dyn-dim"},
+           ap.help("If there is a single non-fixed dynamic dimension in the model, then split to "
+                   "static submodules"),
+           ap.set_value(true));
         ap(quantize, {"--fp16"}, ap.help("Quantize for fp16"), ap.set_value(precision::fp16));
         ap(quantize, {"--int8"}, ap.help("Quantize for int8"), ap.set_value(precision::int8));
     }
 
     auto params(const program& p)
     {
-        return parameters.generate(p, ct.get_target(), co.offload_copy);
+        return parameters.generate(p, ct.get_target(), co.offload_copy, l.batch);
     }
 
     program compile()
@@ -432,7 +520,7 @@ struct verify : command<verify>
         std::cout << p << std::endl;
 
         auto t = c.ct.get_target();
-        auto m = c.parameters.generate(p, t, true);
+        auto m = c.parameters.generate(p, t, true, c.l.batch);
 
         if(per_instruction)
         {

src/driver/perf.cpp

@@ -39,36 +39,25 @@ auto get_hash(const T& x)
     return std::hash<T>{}(x);
 }
 
-parameter_map fill_param_map(parameter_map& m, const program& p, const target& t, bool offload)
+parameter_map fill_param_map(parameter_map& m,
+                             const std::unordered_map<std::string, shape>& param_shapes,
+                             const target& t,
+                             bool offload)
 {
-    for(auto&& x : p.get_parameter_shapes())
+    for(auto&& x : param_shapes)
     {
         argument& arg = m[x.first];
         if(arg.empty())
+        {
+            assert(not x.second.dynamic());
             arg = generate_argument(x.second, get_hash(x.first));
+        }
         if(not offload)
             arg = t.copy_to(arg);
     }
     return m;
 }
 
-parameter_map fill_param_map(parameter_map& m, const program& p, bool gpu)
-{
-    for(auto&& x : p.get_parameter_shapes())
-    {
-        argument& arg = m[x.first];
-        if(arg.empty())
-            arg = generate_argument(x.second, get_hash(x.first));
-#ifdef HAVE_GPU
-        if(gpu)
-            arg = gpu::to_gpu(arg);
-#else
-        (void)gpu;
-#endif
-    }
-    return m;
-}
-
 parameter_map create_param_map(const program& p, const target& t, bool offload)
 {
     parameter_map m;

src/driver/perf.hpp

@@ -30,8 +30,10 @@ namespace migraphx {
 namespace driver {
 inline namespace MIGRAPHX_INLINE_NS {
 
-parameter_map
-fill_param_map(parameter_map& m, const program& p, const target& t, bool offload = false);
+parameter_map fill_param_map(parameter_map& m,
+                             const std::unordered_map<std::string, shape>& param_shapes,
+                             const target& t,
+                             bool offload = false);
 parameter_map create_param_map(const program& p, const target& t, bool offload = false);
 
 parameter_map fill_param_map(parameter_map& m, const program& p, bool gpu);

src/fuse_reduce.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/fuse_reduce.hpp>
+#include <migraphx/pass_manager.hpp>
+#include <migraphx/dead_code_elimination.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/check_shapes.hpp>
+#include <migraphx/matcher.hpp>
+#include <migraphx/register_op.hpp>
+#include <iterator>
+#include <map>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+
+struct fused_reduce
+{
+    std::vector<std::int64_t> axes{};
+
+    template <class Self, class F>
+    static auto reflect(Self& self, F f)
+    {
+        return pack(f(self.axes, "axes"));
+    }
+
+    shape compute_shape(const std::vector<shape>& inputs, std::vector<module_ref> mods) const
+    {
+        if(mods.size() != 1)
+            MIGRAPHX_THROW("should have one submodule.");
+        auto* sm = mods.front();
+        if(sm->get_output_shapes().size() != 1)
+            MIGRAPHX_THROW("Only one output supported");
+        auto names = sm->get_parameter_names();
+        check_shapes{inputs, *this}.has(names.size()).same_ndims();
+        std::sort(names.begin(), names.end());
+        auto shapes = sm->get_parameter_shapes();
+        // Check dimension matches for each input
+        if(not equal(names, inputs, [&](const auto& name, const auto& input) {
+               return shapes.at(name).lens() == input.lens();
+           }))
+            MIGRAPHX_THROW("Dimenstion does not match the submodule.");
+        const auto& s = inputs.at(0);
+        auto lens     = s.lens();
+        if(lens != sm->get_output_shapes().front().lens())
+        {
+            for(const auto& axis : axes)
+            {
+                lens[axis] = 1;
+            }
+        }
+
+        return shape::from_permutation(
+            sm->get_output_shapes().front().type(), lens, find_permutation(inputs));
+    }
+
+    std::string name() const { return "fused_reduce"; }
+};
+MIGRAPHX_REGISTER_OP(fused_reduce);
+
+static std::unordered_map<instruction_ref, instruction_ref>
+get_ins_param_map(const std::vector<instruction_ref>& inputs, const_module_ref sm)
+{
+    std::unordered_map<instruction_ref, instruction_ref> result;
+    auto names = sm->get_parameter_names();
+    std::sort(names.begin(), names.end());
+    assert(names.size() == inputs.size());
+    std::transform(names.begin(),
+                   names.end(),
+                   inputs.begin(),
+                   std::inserter(result, result.end()),
+                   [&](const auto& name, auto input) {
+                       return std::make_pair(input, sm->get_parameter(name));
+                   });
+    return result;
+}
+
+static void insert_params(module_ref sm,
+                          instruction_ref ins,
+                          std::unordered_map<instruction_ref, instruction_ref>& map_ins)
+{
+    auto n = sm->get_parameter_shapes().size();
+    for(auto input : ins->inputs())
+    {
+        if(contains(map_ins, input))
+            continue;
+        auto s         = shape{input->get_shape().type(), input->get_shape().lens()};
+        map_ins[input] = sm->add_parameter("x" + std::to_string(n++), s);
+    }
+}
+
+static auto insert_ins_in_submodule(module_ref sm,
+                                    instruction_ref ins,
+                                    std::unordered_map<instruction_ref, instruction_ref>& map_ins)
+{
+    insert_params(sm, ins, map_ins);
+    return sm->add_instructions({ins}, map_ins);
+}
+
+static auto insert_ins_in_submodule(module_ref sm, instruction_ref ins)
+{
+    std::unordered_map<instruction_ref, instruction_ref> map_ins;
+    return insert_ins_in_submodule(sm, ins, map_ins);
+}
+
+static auto
+insert_module_in_submodule(module_ref sm,
+                           instruction_ref ins,
+                           std::unordered_map<instruction_ref, instruction_ref>& map_ins)
+{
+    insert_params(sm, ins, map_ins);
+    auto* m        = ins->module_inputs().front();
+    auto param_map = get_ins_param_map(ins->inputs(), m);
+    for(auto&& [input, param] : param_map)
+    {
+        map_ins[param] = map_ins.at(input);
+    }
+    return sm->add_instructions(m, map_ins);
+}
+
+static std::vector<instruction_ref>
+find_inputs(module_ref sm,
+            const module& parent,
+            const std::unordered_map<instruction_ref, instruction_ref>& map_ins)
+{
+    std::vector<instruction_ref> result;
+    std::map<std::string, instruction_ref> names;
+    for(auto&& [input, param] : map_ins)
+    {
+        if(not sm->has_instruction(param))
+            continue;
+        if(param->name() != "@param")
+            continue;
+        if(not parent.has_instruction(input))
+            continue;
+        auto v      = param->get_operator().to_value();
+        auto name   = v.at("parameter").to<std::string>();
+        names[name] = input;
+    }
+    std::transform(names.begin(), names.end(), std::back_inserter(result), [](const auto& p) {
+        return p.second;
+    });
+    assert(result.size() == sm->get_parameter_shapes().size());
+    return result;
+}
+
+static void create_reduce_modules(module_pass_manager& mpm)
+{
+    std::size_t n = 0;
+    for(auto ins : iterator_for(mpm.get_module()))
+    {
+        if(not ins->get_operator().attributes().get("reduce", false))
+            continue;
+        if(ins->inputs().size() != 1)
+            continue;
+
+        auto* rm =
+            mpm.create_module(mpm.get_module().name() + ":" + ins->name() + std::to_string(n++));
+        rm->set_bypass();
+
+        rm->add_return(insert_ins_in_submodule(rm, ins));
+        auto v = ins->get_operator().to_value();
+        mpm.get_module().replace_instruction(
+            ins, make_op("fused_reduce", {{"axes", v["axes"]}}), ins->inputs(), {rm});
+    }
+}
+
+template <class... Ms>
+static auto match_broadcast(Ms... ms)
+{
+    return match::skip(match::name("contiguous"))(
+        match::name("multibroadcast")(match::arg(0)(ms...), match::used_once()).bind("broadcast"));
+}
+
+template <class... Ms>
+static auto any_input(Ms... ms)
+{
+    return match::any_of[match::inputs()](match::any(ms...).bind("input"));
+}
+
+static auto match_broadcastable_input(const std::string& op, const std::string& name)
+{
+    auto match_op                 = match::name(op)(match::used_once()).bind(name);
+    auto match_op_input           = any_input(match_op, match::used_once());
+    auto broadcast_match_op_input = any_input(match_broadcast(match_op), match::used_once());
+    return match::any_of(match_op_input, broadcast_match_op_input);
+}
+
+namespace {
+struct find_pointwise_reduce
+{
+    auto matcher() const
+    {
+        return match::name("fused_reduce")(match_broadcastable_input("pointwise", "pointwise"));
+    }
+
+    void apply(module_pass_manager& mpm, const match::matcher_result& r) const
+    {
+        auto reduce = r.result;
+        auto input  = r.instructions["pointwise"];
+
+        const auto* pm     = input->module_inputs().front();
+        const auto* old_rm = reduce->module_inputs().front();
+        auto* rm           = mpm.create_module(pm->name() + ":" + old_rm->name());
+        rm->set_bypass();
+
+        std::unordered_map<instruction_ref, instruction_ref> map_ins;
+        // Insert pointwise
+        auto rins      = insert_ins_in_submodule(rm, input, map_ins).front();
+        map_ins[input] = rins;
+
+        if(contains(r.instructions, "broadcast"))
+        {
+            auto broadcast     = r.instructions["broadcast"];
+            map_ins[broadcast] = insert_ins_in_submodule(rm, broadcast, map_ins).front();
+        }
+
+        // Insert fused_reduce
+        rm->add_return(insert_module_in_submodule(rm, reduce, map_ins));
+
+        auto new_inputs = find_inputs(rm, mpm.get_module(), map_ins);
+        mpm.get_module().replace_instruction(reduce, reduce->get_operator(), new_inputs, {rm});
+    }
+};
+
+struct find_reduce_pointwise
+{
+
+    auto matcher() const
+    {
+        return match::name("pointwise")(match_broadcastable_input("fused_reduce", "reduce"));
+    }
+
+    void apply(module_pass_manager& mpm, const match::matcher_result& r) const
+    {
+        auto pw     = r.result;
+        auto reduce = r.instructions["reduce"];
+        auto input  = r.instructions["input"];
+
+        const auto* pm     = pw->module_inputs().front();
+        const auto* old_rm = reduce->module_inputs().front();
+        auto* rm           = mpm.create_module(old_rm->name() + ":" + pm->name());
+        rm->set_bypass();
+        std::unordered_map<instruction_ref, instruction_ref> map_ins;
+        // Copy module instructions
+        insert_module_in_submodule(rm, reduce, map_ins);
+        if(contains(r.instructions, "broadcast"))
+        {
+            auto broadcast                       = r.instructions["broadcast"];
+            map_ins[broadcast->inputs().front()] = rm->get_returns().front();
+            auto bout                            = insert_ins_in_submodule(rm, broadcast, map_ins);
+            map_ins[input]                       = bout.front();
+        }
+        else
+        {
+            map_ins[input] = rm->get_returns().front();
+        }
+
+        auto out = insert_ins_in_submodule(rm, pw, map_ins);
+        rm->replace_return(out);
+
+        auto new_inputs = find_inputs(rm, mpm.get_module(), map_ins);
+        mpm.get_module().replace_instruction(pw, reduce->get_operator(), new_inputs, {rm});
+    }
+};
+
+struct find_reduce_reduce
+{
+    auto matcher() const
+    {
+        return match::name("fused_reduce")(match_broadcastable_input("fused_reduce", "reduce"));
+    }
+
+    void apply(module_pass_manager& mpm, const match::matcher_result& r) const
+    {
+        auto reduce1 = r.result;
+        auto reduce2 = r.instructions["reduce"];
+        auto input   = r.instructions["input"];
+
+        if(reduce1->get_operator() != reduce2->get_operator())
+            return;
+
+        const auto* rm1 = reduce1->module_inputs().front();
+        const auto* rm2 = reduce2->module_inputs().front();
+        auto* rm        = mpm.create_module(rm1->name() + ":" + rm2->name());
+        rm->set_bypass();
+
+        std::unordered_map<instruction_ref, instruction_ref> map_ins;
+        // Copy reduce1 instructions
+        insert_module_in_submodule(rm, reduce2, map_ins);
+        if(contains(r.instructions, "broadcast"))
+        {
+            auto broadcast                       = r.instructions["broadcast"];
+            map_ins[broadcast->inputs().front()] = rm->get_returns().front();
+            auto bout                            = insert_ins_in_submodule(rm, broadcast, map_ins);
+            map_ins[input]                       = bout.front();
+        }
+        else
+        {
+            map_ins[input] = rm->get_returns().front();
+        }
+
+        auto out = insert_module_in_submodule(rm, reduce1, map_ins);
+        rm->replace_return(out);
+
+        auto new_inputs = find_inputs(rm, mpm.get_module(), map_ins);
+        mpm.get_module().replace_instruction(reduce1, reduce1->get_operator(), new_inputs, {rm});
+    }
+};
+
+} // namespace
+
+void fuse_reduce::apply(module_pass_manager& mpm) const
+{
+    create_reduce_modules(mpm);
+    mpm.run_pass(dead_code_elimination{});
+    for(int i = 0; i < 4; i++)
+    {
+        match::find_matches(
+            mpm, find_reduce_pointwise{}, find_pointwise_reduce{}, find_reduce_reduce{});
+        mpm.run_pass(dead_code_elimination{});
+    }
+}
+
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx

src/include/migraphx/common.hpp

@@ -41,6 +41,11 @@ std::vector<shape::dynamic_dimension> compute_broadcasted_dyn_dims(shape s0, sha
 
 shape common_shape(const std::vector<shape>& shapes);
 
+std::vector<instruction_ref>
+insert_common_args(module& m, instruction_ref ins, std::vector<instruction_ref> inputs);
+
+std::vector<instruction_ref> add_common_args(module& m, std::vector<instruction_ref> inputs);
+
 instruction_ref insert_common_op(module& m,
                                  instruction_ref ins,
                                  const operation& op,

src/include/migraphx/compile_options.hpp

@@ -32,9 +32,17 @@ inline namespace MIGRAPHX_INLINE_NS {
 
 struct compile_options
 {
-    bool offload_copy    = false;
+    /**
+     * Have MIGX allocate memory for parameters and add instructions
+     * to copy parameters and output to/from an offload device like a GPU.
+     */
+    bool offload_copy = false;
+
     bool fast_math       = true;
     bool exhaustive_tune = false;
+
+    /// Use the split_single_dyn_dim pass
+    bool split_single_dyn_dim = false;
     tracer trace{};
 };
 

src/include/migraphx/convolution.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_CONVOLUTION_HPP
+#define MIGRAPHX_GUARD_RTGLIB_CONVOLUTION_HPP
+
+#include <migraphx/config.hpp>
+#include <migraphx/dfor.hpp>
+#include <migraphx/par_for.hpp>
+#include <migraphx/shape_for_each.hpp>
+#include <migraphx/tensor_view.hpp>
+#include <vector>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+
+template <class Output, class T, class Padding, class Stride>
+void convolution(Output output, T input, T weights, Padding padding, Stride stride, int group)
+{
+    auto output_shape = output.get_shape();
+    auto in_lens      = input.get_shape().lens();
+
+    auto wei_lens = weights.get_shape().lens();
+    auto wei_n    = wei_lens[0];
+    auto wei_c    = wei_lens[1];
+    std::vector<std::size_t> win_size(wei_lens.begin() + 1, wei_lens.end());
+
+    par_for(output_shape.elements(), [&](auto i) {
+        auto idx_o = output_shape.multi(i);
+        auto w     = idx_o[1];
+        auto n_dim = idx_o.size();
+
+        std::vector<std::ptrdiff_t> win_start;
+        for(std::size_t dim = 2; dim < n_dim; ++dim)
+        {
+            auto d_2 = dim - 2;
+            win_start.push_back(std::ptrdiff_t(idx_o[dim] * stride[d_2]) -
+                                std::ptrdiff_t(padding[d_2]));
+        }
+        const auto group_id = w / (wei_n / group);
+
+        shape win_shape{output_shape.type(), win_size};
+
+        double acc = 0.0;
+        shape_for_each(win_shape, [&](auto idx_win) {
+            auto k           = idx_win[0];
+            const auto in_ch = group_id * wei_c + k;
+            std::vector<std::ptrdiff_t> idx(idx_o.begin(), idx_o.end());
+            idx[1] = in_ch;
+            std::transform(idx_win.begin() + 1,
+                           idx_win.end(),
+                           win_start.begin(),
+                           idx.begin() + 2,
+                           [](std::ptrdiff_t ii, std::ptrdiff_t jj) { return ii + jj; });
+            std::vector<std::ptrdiff_t> idx_wei(idx_o.size());
+            idx_wei[0] = w;
+            std::copy(idx_win.begin(), idx_win.end(), idx_wei.begin() + 1);
+            if(std::all_of(idx.begin() + 2, idx.end(), [&](auto ii) { return ii >= 0; }) and
+               std::equal(idx.begin(),
+                          idx.end(),
+                          in_lens.begin(),
+                          in_lens.end(),
+                          std::less<std::ptrdiff_t>{}))
+            {
+                acc += input(idx.begin(), idx.end()) * weights(idx_wei.begin(), idx_wei.end());
+            }
+        });
+
+        output[i] = acc;
+    });
+}
+
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx
+
+#endif

src/include/migraphx/cpp_generator.hpp

@@ -77,6 +77,7 @@ struct cpp_generator
         function& set_types(const module& m);
         function& set_types(const module& m, const std::function<std::string(shape)>& parse);
         function& set_generic_types(const module& m);
+        function& add_generic_param(const std::string& pname);
     };
 
     cpp_generator();
@@ -105,6 +106,10 @@ struct cpp_generator
 
     std::string create_function(const function& f);
 
+    static std::vector<std::string>
+    to_args(const std::vector<instruction_ref>& inputs,
+            const std::unordered_map<instruction_ref, std::string>& names);
+
     private:
     std::unique_ptr<cpp_generator_impl> impl;
 };

src/include/migraphx/fuse_reduce.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_MIGRAPHX_FUSE_REDUCE_HPP
+#define MIGRAPHX_GUARD_MIGRAPHX_FUSE_REDUCE_HPP
+
+#include <migraphx/config.hpp>
+#include <string>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+
+struct module_pass_manager;
+
+struct fuse_reduce
+{
+    std::string name() const { return "fuse_reduce"; }
+    void apply(module_pass_manager& mpm) const;
+};
+
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx
+#endif // MIGRAPHX_GUARD_MIGRAPHX_FUSE_POINTWISE_HPP

src/include/migraphx/matcher.hpp

@@ -347,6 +347,7 @@ match::matcher_result find_match(module& modl, M&& m)
 }
 
 MIGRAPHX_DECLARE_ENV_VAR(MIGRAPHX_TRACE_MATCHES)
+MIGRAPHX_DECLARE_ENV_VAR(MIGRAPHX_VALIDATE_MATCHES)
 
 /// Find matches for an instruction in the module
 template <class Mod, class... Ms>
@@ -356,7 +357,11 @@ void find_matches(Mod& mod, instruction_ref ins, Ms&&... ms)
     const
 #endif
         int trace = value_of(MIGRAPHX_TRACE_MATCHES{});
-    bool match    = false;
+#if !defined(__GNUC__) || defined(__clang__) || __GNUC__ > 5
+    const
+#endif
+        bool validate = enabled(MIGRAPHX_VALIDATE_MATCHES{});
+    bool match        = false;
     each_args(
         [&](auto&& m) {
             if(match)
@@ -371,7 +376,20 @@ void find_matches(Mod& mod, instruction_ref ins, Ms&&... ms)
                 std::cout << "Matched by " << get_type_name(m) << std::endl;
                 get_module(mod).debug_print(ins);
             }
+            // If its already invalid dont validate it again
+            bool invalidated = validate and get_module(mod).validate() != get_module(mod).end();
             m.apply(mod, r);
+            if(validate and not invalidated)
+            {
+                auto invalid = get_module(mod).validate();
+                if(invalid != get_module(mod).end())
+                {
+                    std::cout << "Invalid program from match: " << get_type_name(m) << std::endl;
+                    std::cout << "Invalid instructions: " << std::endl;
+                    get_module(mod).debug_print(invalid->inputs());
+                    get_module(mod).debug_print(invalid);
+                }
+            }
             match = true;
         },
         ms...);

src/include/migraphx/module.hpp

@@ -178,6 +178,8 @@ struct module
 
     bool has_instruction(instruction_ref ins) const;
 
+    std::vector<instruction_ref> get_returns() const;
+
     std::size_t size() const;
     instruction_ref begin() const;
     instruction_ref end() const;

src/include/migraphx/msgpack.hpp

@@ -26,6 +26,7 @@
 
 #include <migraphx/config.hpp>
 #include <migraphx/value.hpp>
+#include <functional>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {

src/include/migraphx/onnx.hpp

@@ -37,7 +37,7 @@ struct onnx_options
     std::size_t default_dim_value = 0;
     /// Default dynamic dimension size (if both default_dim_value and default_dyn_dim_value set
     /// parser throws)
-    shape::dynamic_dimension default_dyn_dim_value = {1, 1, 0};
+    shape::dynamic_dimension default_dyn_dim_value = {1, 1};
     /// Explicitly specify the dims of an input
     std::unordered_map<std::string, std::vector<std::size_t>> map_input_dims = {};
     /// Explicitly specify dynamic dims of an input (if both map_input_dims and map_dyn_input_dims