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

src/include/migraphx/module.hpp

@@ -54,6 +54,10 @@ using ins_dep_map   = std::unordered_map<instruction_ref, std::unordered_set<ins
  */
 struct module
 {
+    // used by replace_allocate pass
+    // allocate memory in this module rather than using output parmaeters
+    bool use_local_alloc = false;
+
     module(const std::string& name = "");
 
     // move constructor

src/include/migraphx/op/select_module.hpp

@@ -26,8 +26,6 @@
 
 #include <migraphx/check_shapes.hpp>
 #include <migraphx/module.hpp>
-#include <migraphx/dyn_output.hpp>
-#include <set>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
@@ -35,98 +33,59 @@ namespace op {
 
 struct select_module
 {
-    // output shape of the dynamic model
-    shape output_dyn_shape;
-    int input_batch_index  = -1;
-    int output_batch_index = -1;
-    std::string dyn_batch_param_name;
+    shape output_dyn_shapes;
 
     template <class Self, class F>
     static auto reflect(Self& self, F f)
     {
-        return pack(f(self.output_dyn_shape, "output_dyn_shape"),
-                    f(self.input_batch_index, "input_batch_index"),
-                    f(self.output_batch_index, "output_batch_index"),
-                    f(self.dyn_batch_param_name, "dyn_batch_param_name"));
+        return pack(f(self.output_dyn_shapes, "output_dyn_shapes"));
     }
 
     std::string name() const { return "select_module"; }
 
-    // runs once during model compilation with dynamic shape input
-    // may run on each model evaluation with static shape input
-    shape compute_shape(std::vector<shape> inputs) const
+    shape compute_shape(const std::vector<shape>&, std::vector<module_ref>) const
     {
-        check_shapes{inputs, *this, true}.has(1);
-        auto s0 = inputs.at(0);
-        if(s0.dynamic())
-        {
-            // should we check that the submodules have the same parameters here?
-            // check that no more than one parameter is non-fixed?
-            // would need to use version of compute_shape with the parameter list
-            return shape{output_dyn_shape};
-        }
-        else
-        {
-            auto batch_size            = s0.lens().at(input_batch_index);
-            auto dds                   = output_dyn_shape.dyn_dims();
-            dds.at(output_batch_index) = {batch_size, batch_size};
-            std::vector<std::size_t> dims;
-            if(std::all_of(dds.begin(), dds.end(), [](auto dd) { return dd.is_fixed(); }))
-            {
-                std::transform(
-                    dds.begin(), dds.end(), std::back_inserter(dims), [](auto d) { return d.max; });
-                return {output_dyn_shape.type(), dims};
-            }
-            else
-            {
-                MIGRAPHX_THROW("SELECT_MODULE: more than one input dimension was non-fixed");
-            }
-        }
+        return shape{output_dyn_shapes};
     }
 
-    argument compute(const dyn_output& dyn_out,
+    argument compute(const shape&,
                      const std::vector<argument>& args,
                      const std::vector<module_ref>& submodule_list,
                      const std::function<std::vector<argument>(
                          module_ref&, const std::unordered_map<std::string, argument>&)>& run) const
     {
-        std::vector<module_ref> modules_to_run;
-        for(const auto& mod : submodule_list)
-        {
-            // find submodule with the same parameter shape as the input data
-            auto p_shape = mod->get_parameter_shape(dyn_batch_param_name);
-            if(p_shape == args.at(0).get_shape())
-            {
-                modules_to_run.push_back(mod);
-                break;
-            }
-        }
-        // TODO if an exact match is not found, assemble module list from binary base
+        // find submodule with input parameter shapes exactly the same as the input arguments
+        // assuming arguments are in the same order as the input parameters
+        auto module_iter =
+            std::find_if(submodule_list.cbegin(), submodule_list.cend(), [&](module_ref mr) {
+                auto param_names = mr->get_parameter_names();
+                assert(param_names.size() <= args.size());
+                return std::equal(param_names.cbegin(),
+                                  param_names.cend(),
+                                  args.cbegin(),
+                                  [&](auto p_name, auto a) {
+                                      return a.get_shape() == mr->get_parameter_shape(p_name);
+                                  });
+            });
 
-        if(modules_to_run.empty())
+        if(module_iter == submodule_list.end())
         {
-            MIGRAPHX_THROW("SELECT_MODULE: no compatible submodules found for input shape: " +
-                           migraphx::to_string(args.at(0).get_shape()));
+            MIGRAPHX_THROW("SELECT_MODULE: no compatible submodules found for given input shapes");
         }
-        std::set<std::string> pnames;
-        for(const auto& mod : modules_to_run)
-        {
-            // TODO If all the modules have the same parameters, this would only need to run once
-            auto names = mod->get_parameter_names();
-            pnames.insert(names.begin(), names.end());
-        }
-
-        assert(pnames.size() <= args.size());
+        auto module_to_run = *module_iter;
         std::unordered_map<std::string, argument> params;
-        std::transform(pnames.begin(),
-                       pnames.end(),
+
+        // add input parameters
+        auto param_names = module_to_run->get_parameter_names();
+        assert(param_names.size() <= args.size());
+        std::transform(param_names.begin(),
+                       param_names.end(),
                        args.begin(),
                        std::inserter(params, params.end()),
-                       [](auto&& name, auto&& arg) { return std::make_pair(name, arg); });
+                       [](auto&& name, auto&& a) { return std::make_pair(name, a); });
 
-        // TODO run multiple modules and split the parameter data to each batch size
-        auto results = run(modules_to_run.at(0), params);
-        return results.at(0);
+        auto results = run(module_to_run, params);
+        return argument{results};
     }
 };
 

src/include/migraphx/operation.hpp

@@ -140,9 +140,9 @@ template <class T>
 auto compute_shape_op(rank<2>, const T& x, const std::vector<shape>& inputs)
     -> decltype(x.normalize_compute_shape(inputs))
 {
-    dependent_type<operation, T> y = x;
     if(inputs.empty())
         MIGRAPHX_THROW("At least one input is required for " + x.name());
+    dependent_type<operation, T> y = x;
     normalize_attributes(y, inputs[0].max_lens());
     return any_cast<T>(y).normalize_compute_shape(inputs);
 }
@@ -168,7 +168,7 @@ shape compute_shape_op(const T& x, const std::vector<shape>& inputs)
 }
 
 template <class T>
-auto mod_compute_shape_op(rank<2>,
+auto mod_compute_shape_op(rank<1>,
                           const T& x,
                           const std::vector<shape>& inputs,
                           const std::vector<module_ref>& mod_args)
@@ -177,15 +177,6 @@ auto mod_compute_shape_op(rank<2>,
     return x.compute_shape(inputs, mod_args);
 }
 
-template <class T>
-auto mod_compute_shape_op(rank<1>,
-                          const T& x,
-                          const std::vector<shape>& inputs,
-                          const std::vector<module_ref>&) -> decltype(x.compute_shape(inputs))
-{
-    return x.compute_shape(inputs);
-}
-
 template <class T>
 shape mod_compute_shape_op(rank<0>,
                            const T& x,

src/include/migraphx/shape.hpp

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

src/program.cpp

@@ -379,7 +379,7 @@ std::vector<argument> generic_eval(const module* mod,
                             }));
         }
         assert(results.find(ins) != results.end());
-        if(not ins->get_shape().dynamic())
+        if(not ins->get_shape().any_of_dynamic())
         {
             assert(results.at(ins).get_shape() == ins->get_shape());
         }

src/replace_allocate.cpp

@@ -104,19 +104,17 @@ void replace_allocate::apply(module& m) const
             continue;
 
         auto s = ins->get_shape();
-
-        if(not main_offload_copy and model.needs_out_params() and contains(mod_output_names, ins))
+        if(not main_offload_copy and not(m.use_local_alloc) 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)}})));
+        else
+        {
+            m.replace_instruction(ins,
+                                  make_op(model.name(), migraphx::value{{"shape", to_value(s)}}));
+        }
     }
 }
 

src/shape.cpp

@@ -483,6 +483,17 @@ std::string shape::type_string() const { return name(this->type()); }
 
 bool shape::dynamic() const { return not impl->m_dyn_dims.empty(); }
 
+bool shape::any_of_dynamic() const
+{
+    if(this->dynamic())
+    {
+        return true;
+    }
+    return std::any_of(this->sub_shapes().cbegin(), this->sub_shapes().cend(), [](auto s) {
+        return s.any_of_dynamic();
+    });
+}
+
 const std::vector<shape::dynamic_dimension>& shape::dyn_dims() const { return impl->m_dyn_dims; }
 
 std::vector<std::size_t> shape::min_lens() const

src/targets/gpu/lowering.cpp

@@ -111,6 +111,7 @@ struct miopen_apply
         add_loop_op();
         add_neg_op();
         add_nms_op();
+        add_select_module_op();
     }
 
     void copy_params() const
@@ -358,6 +359,33 @@ struct miopen_apply
             return mod->replace_instruction(ins, gpu_out);
         });
     }
+
+    /**
+     * Turns on use_local_alloc in the select_module submodules.
+     * Changes the submodule returns to a hip::sync_stream.
+     */
+    void add_select_module_op()
+    {
+        apply_map.emplace("select_module", [=](instruction_ref ins) {
+            std::vector<instruction_ref> inputs = ins->inputs();
+            auto mod_args                       = ins->module_inputs();
+            for(auto smod : mod_args)
+            {
+                smod->use_local_alloc = true;
+                auto last_ins         = std::prev(smod->end());
+                if(last_ins->name() == "@return")
+                {
+                    for(auto out_ins : last_ins->inputs())
+                    {
+                        auto sync_out = smod->insert_instruction(
+                            last_ins, make_op("hip::sync_stream"), out_ins);
+                        smod->replace_return({sync_out});
+                    }
+                }
+            }
+            return ins;
+        });
+    }
 };
 
 void lowering::apply(module& m) const { miopen_apply{&m, this}.apply(); }

test/op_shape_test.cpp

@@ -2364,22 +2364,12 @@ TEST_CASE(rnn)
 TEST_CASE(select_module_dyn)
 {
     migraphx::shape input{migraphx::shape::float_type, {{1, 4}, {3, 3}, {255, 255}, {255, 255}}};
-    migraphx::shape out_attr = migraphx::shape{migraphx::shape::float_type, {{1, 4}, {1000, 1000}}};
+    std::vector<migraphx::shape> sub_shapes = {};
+    sub_shapes.push_back(migraphx::shape{migraphx::shape::float_type, {{1, 4}, {1000, 1000}}});
+    migraphx::shape out_attr = migraphx::shape{sub_shapes};
     expect_shape(
-        migraphx::shape{migraphx::shape::float_type, {{1, 4}, {1000, 1000}}},
-        migraphx::make_op("select_module", {{"output_dyn_shape", migraphx::to_value(out_attr)}}),
-        input);
-}
-
-TEST_CASE(select_module_static)
-{
-    migraphx::shape input{migraphx::shape::float_type, {3, 3, 255, 255}};
-    migraphx::shape out_attr = migraphx::shape{migraphx::shape::float_type, {{1, 4}, {1000, 1000}}};
-    expect_shape(migraphx::shape{migraphx::shape::float_type, {3, 1000}},
-                 migraphx::make_op("select_module",
-                                   {{"output_dyn_shape", migraphx::to_value(out_attr)},
-                                    {"output_batch_index", 0},
-                                    {"input_batch_index", 0}}),
+        out_attr,
+        migraphx::make_op("select_module", {{"output_dyn_shapes", migraphx::to_value(out_attr)}}),
         input);
 }
 

test/shape_test.cpp

@@ -238,6 +238,30 @@ TEST_CASE(test_shape_dynamic_serialize)
     EXPECT(s3 != s4);
 }
 
+TEST_CASE(any_of_dynamic_true)
+{
+    std::vector<migraphx::shape> sub_shapes = {};
+    sub_shapes.push_back(migraphx::shape{migraphx::shape::float_type, {{1, 4}, {4, 4}}});
+    sub_shapes.push_back(migraphx::shape{migraphx::shape::float_type, {3, 4, 5}});
+    migraphx::shape s0{sub_shapes};
+    EXPECT(s0.any_of_dynamic());
+
+    sub_shapes = {};
+    sub_shapes.push_back(migraphx::shape{migraphx::shape::float_type, {{1, 1}, {4, 4}}});
+    sub_shapes.push_back(migraphx::shape{migraphx::shape::float_type, {3, 4, 5}});
+    migraphx::shape s1{sub_shapes};
+    EXPECT(s1.any_of_dynamic());
+}
+
+TEST_CASE(any_of_dynamic_false)
+{
+    std::vector<migraphx::shape> sub_shapes = {};
+    sub_shapes.push_back(migraphx::shape{migraphx::shape::float_type, {1, 4}});
+    sub_shapes.push_back(migraphx::shape{migraphx::shape::float_type, {3, 4, 5}});
+    migraphx::shape s{sub_shapes};
+    EXPECT(not s.any_of_dynamic());
+}
+
 TEST_CASE(test_shape_packed)
 {
     migraphx::shape s{migraphx::shape::float_type, {2, 2}, {2, 1}};

test/verify/run_verify.cpp

@@ -184,9 +184,18 @@ void run_verify::verify(const std::string& name, const migraphx::program& p) con
         std::vector<std::pair<std::string, result_future>> results;
         migraphx::parameter_map m;
         for(auto&& x : p.get_parameter_shapes())
+        {
+            if(x.second.dynamic())
+            {
+                // create static shape using maximum dimensions
+                migraphx::shape static_shape{x.second.type(), x.second.max_lens()};
+                m[x.first] = migraphx::generate_argument(static_shape, get_hash(x.first));
+            }
+            else
             {
                 m[x.first] = migraphx::generate_argument(x.second, get_hash(x.first));
             }
+        }
 
         auto gold_f = detach_async([=] { return run_ref(p, m); });
         for(const auto& tname : target_names)

test/verify/test_select_module_add.cpp

+/*
+ * The MIT License (MIT)
+ *
+ * Copyright (c) 2015-2023 Advanced Micro Devices, Inc. All rights reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
+ * THE SOFTWARE.
+ */
+
+#include "verify_program.hpp"
+#include <migraphx/program.hpp>
+#include <migraphx/generate.hpp>
+#include <migraphx/make_op.hpp>
+
+struct test_select_module_add : verify_program<test_select_module_add>
+{
+    migraphx::program create_program() const
+    {
+        migraphx::program p;
+        auto* mm = p.get_main_module();
+        migraphx::shape lit_s{migraphx::shape{migraphx::shape::float_type, {1}}};
+        auto literal_ins = mm->add_literal(migraphx::literal{lit_s, {6}});
+
+        // create batch submodules
+        auto create_submodule = [&](std::size_t batch_size, std::string module_name) {
+            auto* submod = p.create_module(module_name);
+            migraphx::shape sm_shape{migraphx::shape::float_type, {batch_size, 4}};
+            auto sm_input = submod->add_parameter("data", sm_shape);
+            auto broadcast_lit =
+                submod->add_instruction(migraphx::make_op("multibroadcast"), literal_ins, sm_input);
+            auto add_ins =
+                submod->add_instruction(migraphx::make_op("add"), sm_input, broadcast_lit);
+            submod->add_return({add_ins});
+            return submod;
+        };
+        auto* batch1 = create_submodule(1, "batch_1");
+        auto* batch2 = create_submodule(2, "batch_2");
+        auto* batch3 = create_submodule(3, "batch_3");
+        auto* batch4 = create_submodule(4, "batch_4");
+
+        migraphx::shape s{migraphx::shape::float_type, {{1, 4}, {4, 4}}};
+        auto input                              = mm->add_parameter("data", s);
+        std::vector<migraphx::shape> sub_shapes = {};
+        sub_shapes.push_back(migraphx::shape{migraphx::shape::float_type, {{1, 4}, {4, 4}}});
+        migraphx::shape out_attr = migraphx::shape{sub_shapes};
+        auto sm_ins              = mm->add_instruction(
+            migraphx::make_op("select_module",
+                              {{"output_dyn_shapes", migraphx::to_value(out_attr)}}),
+            {input},
+            {batch1, batch2, batch3, batch4});
+        auto ret = mm->add_instruction(migraphx::make_op("get_tuple_elem", {{"index", 0}}), sm_ins);
+        mm->add_return({ret});
+
+        return p;
+    }
+};

test/verify/test_select_module_reduce.cpp

+/*
+ * The MIT License (MIT)
+ *
+ * Copyright (c) 2015-2023 Advanced Micro Devices, Inc. All rights reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
+ * THE SOFTWARE.
+ */
+
+#include "verify_program.hpp"
+#include <migraphx/program.hpp>
+#include <migraphx/generate.hpp>
+#include <migraphx/make_op.hpp>
+
+struct test_select_module_reduce : verify_program<test_select_module_reduce>
+{
+    migraphx::program create_program() const
+    {
+        migraphx::program p;
+
+        // create batch submodules
+        auto create_submodule = [&](std::size_t batch_size, std::string module_name) {
+            auto submod = p.create_module(module_name);
+            migraphx::shape sm_shape{migraphx::shape::float_type, {batch_size, 2, 2}};
+            auto sm_input = submod->add_parameter("data", sm_shape);
+            auto reduce_ins =
+                submod->add_instruction(migraphx::make_op("reduce_sum", {{"axes", {1}}}), sm_input);
+            auto squeeze_ins =
+                submod->add_instruction(migraphx::make_op("squeeze", {{"axes", {1}}}), reduce_ins);
+            submod->add_return({squeeze_ins});
+            return submod;
+        };
+        auto* batch1 = create_submodule(1, "batch_1");
+        auto* batch2 = create_submodule(2, "batch_2");
+        auto* batch3 = create_submodule(3, "batch_3");
+        auto* batch4 = create_submodule(4, "batch_4");
+
+        auto* mm = p.get_main_module();
+        migraphx::shape s{migraphx::shape::float_type, {{1, 4}, {2, 2}, {2, 2}}};
+        auto input                              = mm->add_parameter("data", s);
+        std::vector<migraphx::shape> sub_shapes = {};
+        sub_shapes.push_back(migraphx::shape{migraphx::shape::float_type, {{1, 4}, {2, 2}}});
+        migraphx::shape out_attr = migraphx::shape{sub_shapes};
+        auto sm_ins              = mm->add_instruction(
+            migraphx::make_op("select_module",
+                              {{"output_dyn_shapes", migraphx::to_value(out_attr)}}),
+            {input},
+            {batch1, batch2, batch3, batch4});
+        auto ret = mm->add_instruction(migraphx::make_op("get_tuple_elem", {{"index", 0}}), sm_ins);
+        mm->add_return({ret});
+
+        return p;
+    }
+};

tools/include/operation.hpp

@@ -168,7 +168,7 @@ shape compute_shape_op(const T& x, const std::vector<shape>& inputs)
 }
 
 template <class T>
-auto mod_compute_shape_op(rank<2>,
+auto mod_compute_shape_op(rank<1>,
                           const T& x,
                           const std::vector<shape>& inputs,
                           const std::vector<module_ref>& mod_args)
@@ -177,15 +177,6 @@ auto mod_compute_shape_op(rank<2>,
     return x.compute_shape(inputs, mod_args);
 }
 
-template <class T>
-auto mod_compute_shape_op(rank<1>,
-                          const T& x,
-                          const std::vector<shape>& inputs,
-                          const std::vector<module_ref>&) -> decltype(x.compute_shape(inputs))
-{
-    return x.compute_shape(inputs);
-}
-
 template <class T>
 shape mod_compute_shape_op(rank<0>,
                            const T& x,