Refactor

src/targets/gpu/target.cpp

@@ -18,10 +18,8 @@
 #include <migraphx/rewrite_rnn.hpp>
 #include <migraphx/eliminate_concat.hpp>
 #include <migraphx/gpu/concat_gpu_opt.hpp>
-#include <migraphx/pre_scheduling.hpp>
-#include <migraphx/gpu/machine_model.hpp>
-#include <migraphx/gpu/find_concur_gpu.hpp>
-#include <migraphx/gpu/insert_instruction_gpu.hpp>
+#include <migraphx/gpu/schedule_model.hpp>
+#include <migraphx/schedule.hpp>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
@@ -30,8 +28,6 @@ namespace gpu {
 std::vector<pass> target::get_passes(migraphx::context& gctx) const
 {
     auto& ctx = any_cast<context>(gctx);
-    std::function<std::pair<int, int>(const operation&)> weight_func = op_info();
-    int num_of_streams = ctx.get_current_device().nstreams();
     // clang-format off
     return
     {
@@ -57,8 +53,9 @@ std::vector<pass> target::get_passes(migraphx::context& gctx) const
         fuse_ops{&ctx},
         dead_code_elimination{},
         write_literals{&ctx},
-        pre_scheduling{weight_func, num_of_streams, insert_instruction_gpu{}},
-        memory_coloring{"hip::allocate", num_of_streams, find_concur_gpu{}},
+        schedule{gpu::schedule_model{ctx.get_current_device().nstreams()}},
+        memory_coloring{"hip::allocate"},
+        dead_code_elimination{},
         eliminate_workspace{},
         eliminate_allocation{"hip::allocate"},
         check_context<context>{},

test/gpu/miopen.cpp

@@ -130,9 +130,7 @@ migraphx::argument run_gpu(migraphx::program& p)
     EXPECT(is_shared(ctx, p.get_context()));
     p.dry_run(m);
     EXPECT(is_shared(ctx, p.get_context()));
-    auto eval    = p.eval(m);
-    auto ret_val = migraphx::gpu::from_gpu(eval);
-    return ret_val;
+    return migraphx::gpu::from_gpu(p.eval(m));
 }
 
 template <class V>

test/gpu/stream_execution.cpp

-#include <test.hpp>
-#include <basic_ops.hpp>
-#include <migraphx/program.hpp>
-#include <migraphx/instruction.hpp>
-#include <migraphx/generate.hpp>
-#include <migraphx/gpu/target.hpp>
-#include <migraphx/cpu/target.hpp>
-#include <migraphx/gpu/hip.hpp>
-#include <migraphx/verify_args.hpp>
-
-migraphx::program create_program()
-{
-    migraphx::program p;
-    auto in1 = p.add_parameter("0", migraphx::shape{migraphx::shape::float_type, {32, 64, 1, 1}});
-    auto in2 = p.add_parameter("1", migraphx::shape{migraphx::shape::float_type, {64, 64, 1, 1}});
-    auto p1  = p.add_instruction(migraphx::op::convolution{}, in1, in2);
-    auto in3 = p.add_parameter("2", migraphx::shape{migraphx::shape::float_type, {64, 64, 1, 1}});
-    auto p2  = p.add_instruction(migraphx::op::convolution{}, in1, in3);
-    p.add_instruction(migraphx::op::concat{1}, p1, p2);
-    return p;
-}
-
-migraphx::argument run_gpu()
-{
-    setenv("MIGRAPHX_DISABLE_NULL_STREAM", "1", 1);
-    migraphx::program p = create_program();
-    p.compile(migraphx::gpu::target{});
-    migraphx::program::parameter_map m;
-    for(auto&& x : p.get_parameter_shapes())
-    {
-        m[x.first] = migraphx::gpu::to_gpu(migraphx::generate_argument(x.second));
-    }
-    auto ret_val = migraphx::gpu::from_gpu(p.eval(m));
-    p.finish();
-    return ret_val;
-}
-
-migraphx::argument run_cpu()
-{
-    migraphx::program p = create_program();
-    p.compile(migraphx::cpu::target{});
-    migraphx::program::parameter_map m;
-    for(auto&& x : p.get_parameter_shapes())
-    {
-        m[x.first] = migraphx::generate_argument(x.second);
-    }
-    return p.eval(m);
-}
-
-void gpu_stream_execution_test()
-{
-    auto result1 = run_gpu();
-    auto result2 = run_cpu();
-    verify_args("test", result2, result1);
-}
-
-int main() { gpu_stream_execution_test(); }

test/include/test.hpp

@@ -36,6 +36,18 @@ inline std::ostream& operator<<(std::ostream& s, std::nullptr_t)
     return s;
 }
 
+template <class T>
+inline std::ostream& operator<<(std::ostream& s, const std::vector<T>& v)
+{
+    s << "{ ";
+    for(auto&& x : v)
+    {
+        s << x << ", ";
+    }
+    s << "}";
+    return s;
+}
+
 template <class T, class U, class Operator>
 struct expression
 {

test/memory_coloring_test.cpp

@@ -2,47 +2,15 @@
 #include <migraphx/operators.hpp>
 #include <migraphx/generate.hpp>
 #include <migraphx/instruction.hpp>
-#include <migraphx/dom_info.hpp>
-#include <migraphx/common_header.hpp>
-#include <migraphx/instruction.hpp>
 #include <basic_ops.hpp>
 #include <test.hpp>
 
-struct set_stream
-{
-    int stream = -1;
-    std::string name() const { return "gpu::set_stream"; }
-
-    migraphx::shape compute_shape(const std::vector<migraphx::shape>& inputs) const
-    {
-        if(inputs.empty())
-            return {};
-        else
-            return inputs.front();
-    }
-};
-
-struct find_concur
-{
-    void get_concur(
-        migraphx::program* p,
-        int num_of_streams,
-        std::unordered_map<const migraphx::instruction*,
-                           std::vector<std::vector<const migraphx::instruction*>>>& concur_instrs,
-        std::unordered_map<const migraphx::instruction*, int>& instr2_points) const
-    {
-        migraphx::dom_info info(p);
-        info.compute_dom(true);
-        info.propagate_splits(num_of_streams, concur_instrs, instr2_points);
-    }
-};
-
 struct memory_coloring_target
 {
     std::string name() const { return "memory_coloring"; }
     std::vector<migraphx::pass> get_passes(migraphx::context&) const
     {
-        return {migraphx::memory_coloring{"allocate", 4, find_concur{}, true}};
+        return {migraphx::memory_coloring{"allocate", true}};
     }
     migraphx::context get_context() const { return {}; }
 };
@@ -639,40 +607,4 @@ TEST_CASE(literal_test)
     CHECK(lit == result);
 }
 
-TEST_CASE(concurrent_test)
-{
-    migraphx::program p;
-    auto in = p.add_parameter("0", migraphx::shape{migraphx::shape::float_type, {40}});
-    auto a1 = add_alloc(p, {migraphx::shape::float_type, {40}});
-    auto p1 = p.add_instruction(pass_op{}, a1, in);
-    p.insert_instruction(p1, set_stream{0});
-    p1->set_stream(0);
-    auto a2 = add_alloc(p, {migraphx::shape::float_type, {40}});
-    auto p2 = p.add_instruction(pass_op{}, a2, p1);
-    p2->set_stream(0);
-    auto a4 = add_alloc(p, {migraphx::shape::float_type, {40}});
-    auto p4 = p.add_instruction(pass_op{}, a4, p2);
-    p4->set_stream(0);
-    auto a3 = add_alloc(p, {migraphx::shape::float_type, {40}});
-    auto p3 = p.add_instruction(pass_op{}, a3, p1);
-    p3->set_stream(1);
-    p.insert_instruction(p3, set_stream{1});
-    auto a5 = add_alloc(p, {migraphx::shape::float_type, {40}});
-    auto p5 = p.add_instruction(pass_op{}, a5, p3);
-    p5->set_stream(1);
-    auto a6 = add_alloc(p, {migraphx::shape::float_type, {40}});
-    auto p6 = p.add_instruction(pass_op{}, a6, p1);
-    p6->set_stream(2);
-    p.insert_instruction(p6, set_stream{2});
-    auto a7 = add_alloc(p, {migraphx::shape::float_type, {40}});
-    auto p7 = p.add_instruction(pass_op{}, a7, p6);
-    p7->set_stream(2);
-    auto a8 = add_alloc(p, {migraphx::shape::float_type, {40}});
-    auto p8 = p.add_instruction(migraphx::op::concat{0}, a8, p4, p5, p7);
-    p8->set_stream(0);
-    p.insert_instruction(p8, set_stream{0});
-    p.compile(memory_coloring_target{});
-    CHECK(p.get_parameter_shape("scratch").bytes() == 960);
-}
-
 int main(int argc, const char* argv[]) { test::run(argc, argv); }

test/stream_execution_test.cpp

-#include <migraphx/pre_scheduling.hpp>
-#include <migraphx/operators.hpp>
-#include <migraphx/generate.hpp>
-#include <migraphx/instruction.hpp>
-#include <migraphx/program.hpp>
-#include <basic_ops.hpp>
-#include <test.hpp>
-
-struct set_stream
-{
-    int stream = -1;
-    std::string name() const { return "set_stream"; }
-
-    migraphx::shape compute_shape(const std::vector<migraphx::shape>& inputs) const
-    {
-        if(inputs.empty())
-            return {};
-        else
-            return inputs.front();
-    }
-};
-
-struct create_events
-{
-    int num_of_events = 0;
-    std::string name() const { return "gpu::create_events"; }
-    migraphx::shape compute_shape(const std::vector<migraphx::shape>& inputs) const
-    {
-        if(inputs.empty())
-            return {};
-        else
-            return inputs.front();
-    }
-};
-
-struct weight_func
-{
-    weight_func()
-    {
-        weight_map["@param"]   = std::make_pair(1, 1);
-        weight_map["@literal"] = std::make_pair(1, 1);
-    };
-    std::pair<int, int> operator()(const migraphx::operation& op)
-    {
-        if(weight_map.find(op.name()) != weight_map.end())
-            return weight_map[op.name()];
-        else
-            return std::make_pair(1, 0);
-    }
-    std::unordered_map<std::string, std::pair<int, int>> weight_map;
-};
-
-struct insert_instruction
-{
-    void insert_stream(migraphx::program* p, migraphx::instruction_ref ins, int stream)
-    {
-
-        p->insert_instruction(ins, set_stream{stream});
-    }
-
-    void insert_create_events(migraphx::program*, migraphx::instruction_ref, int) {}
-    void insert_record_event(migraphx::program*, migraphx::instruction_ref, int) {}
-
-    void insert_wait_event(migraphx::program*, migraphx::instruction_ref, int) {}
-};
-
-struct stream_execution_target
-{
-    struct context
-    {
-        void finish() const {}
-        void set_stream(int) {}
-        void create_events(int) {}
-        void record_event(int) {}
-        void wait_event(int) {}
-    };
-    migraphx::context ctx = context{};
-    std::string name() const { return "stream_execution"; }
-    std::vector<migraphx::pass> get_passes(migraphx::context&) const
-    {
-        return {migraphx::pre_scheduling{weight_func(), 2, insert_instruction{}, true}};
-    }
-    migraphx::context get_context() const { return {ctx}; }
-};
-
-TEST_CASE(test1)
-{
-    migraphx::program p;
-    auto in1 =
-        p.add_parameter("0", migraphx::shape{migraphx::shape::float_type, {32, 256, 35, 35}});
-    auto l1 =
-        p.add_literal(migraphx::generate_literal({migraphx::shape::float_type, {64, 256, 1, 1}}));
-    auto p1 = p.add_instruction(migraphx::op::convolution{}, in1, l1);
-    auto l2 =
-        p.add_literal(migraphx::generate_literal({migraphx::shape::float_type, {48, 256, 1, 1}}));
-    auto p2 = p.add_instruction(migraphx::op::convolution{}, in1, l2);
-    p.add_instruction(migraphx::op::concat{1}, p1, p2);
-    p.compile(stream_execution_target{});
-    std::cout << p << std::endl;
-    CHECK(std::count_if(
-              p.begin(), p.end(), [](auto&& ins) { return ins.name() == "set_stream"; }) == 3);
-    CHECK(std::count_if(p.begin(), p.end(), [](auto&& ins) { return ins.get_stream() == 0; }) == 2);
-    CHECK(std::count_if(p.begin(), p.end(), [](auto&& ins) { return ins.get_stream() == 1; }) == 1);
-}
-
-int main(int argc, const char* argv[]) { test::run(argc, argv); }

tools/include/context.hpp

@@ -27,11 +27,12 @@ struct context
 
 <%
 interface('context',
-          virtual('finish', returns='void', const=True),
+    virtual('finish', returns='void', const=True)
 )
 %>
 
 #endif
+
 } // namespace MIGRAPHX_INLINE_NS
 } // namespace migraphx
 

tools/include/find_concur.hpp

-#ifndef MIGRAPHX_GUARD_FIND_CONCUR_HPP
-#define MIGRAPHX_GUARD_FIND_CONCUR_HPP
-
-#include <cassert>
-#include <string>
-#include <functional>
-#include <memory>
-#include <type_traits>
-#include <utility>
-#include <unordered_map>
-#include <vector>
-
-#include <migraphx/instruction.hpp>
-
-namespace migraphx {
-inline namespace MIGRAPHX_INLINE_NS {
-
-struct program;
-
-#ifdef DOXYGEN
-
-/// An interface for target-dependent analysis to find concurrent instructions
-/// executing in different streams.
-struct find_concur
-{
-    void get_concur(program* p,
-                    int num_of_streams,
-                    std::unordered_map<const instruction*,
-                                       std::vector<std::vector<const instruction*>>>& concur_instrs,
-                    std::unordered_map<const instruction*, int>& instr2_points);
-} const;
-
-#else
-
-<%
-interface('find_concur',
-          virtual('get_concur', returns='void', p = 'program*', num_of_stream = 'int', concur_instrs = 'std::unordered_map<const instruction*, std::vector<std::vector<const instruction*>>>&', input = 'std::unordered_map<const instruction*, int>&', const=True)
-)
-%>
-
-#endif
-
-} // namespace MIGRAPHX_INLINE_NS
-} // namespace migraphx
-
-#endif

tools/include/insert_instruction.hpp

-#ifndef MIGRAPHX_GUARD_INSERT_INSTRUCTION_HPP
-#define MIGRAPHX_GUARD_INSERT_INSTRUCTION_HPP
-
-#include <cassert>
-#include <string>
-#include <functional>
-#include <memory>
-#include <type_traits>
-#include <utility>
-
-#include <migraphx/instruction_ref.hpp>
-
-namespace migraphx {
-inline namespace MIGRAPHX_INLINE_NS {
-
-struct program;
-
-#ifdef DOXYGEN
-
-/// An interface for target-dependent instruction insertion.
-/// for multi-stream execution.
-struct insert_instruction
-{
-    void insert_create_events(program* p, instruction_ref ins, int num_of_events);
-    void insert_record_event(program* p, instruction_ref ins, int event);
-    void insert_wait_event(program* p, instruction_ref ins, int event);
-
-    void insert_stream(program* p, instruction_ref ins, int stream);
-};
-
-#else
-
-<%
-interface('insert_instruction',
-          virtual('insert_create_events', returns='void', p = 'program*', ins ='instruction_ref', input = 'int'),
-          virtual('insert_record_event', returns='void', p = 'program*', ins ='instruction_ref', input = 'int'),
-          virtual('insert_wait_event', returns='void', p = 'program*', ins = 'instruction_ref', input = 'int'),
-          virtual('insert_stream', returns='void', p = 'program*', ins ='instruction_ref', input = 'int')
-)
-%>
-
-#endif
-
-} // namespace MIGRAPHX_INLINE_NS
-} // namespace migraphx
-
-#endif

tools/include/schedule_model.hpp

+#ifndef MIGRAPHX_GUARD_SCHEDULE_MODEL_HPP
+#define MIGRAPHX_GUARD_SCHEDULE_MODEL_HPP
+
+#include <cassert>
+#include <string>
+#include <functional>
+#include <memory>
+#include <type_traits>
+#include <utility>
+
+#include <migraphx/config.hpp>
+#include <migraphx/instruction_ref.hpp>
+#include <vector>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+
+struct program;
+struct operation;
+
+#ifdef DOXYGEN
+
+/// An interface for target-dependent model for the scheduler
+struct schedule_model
+{
+    /// Get the number of concurrent instruction allowed
+    std::size_t concurrency() const;
+    /// Schedule a concurrent instruction
+    void sched(program& p, instruction_ref ins, std::size_t n) const;
+    // Insert necessary waits before an instruction
+    void wait(program& p, instruction_ref ins, std::size_t wait_id) const;
+    // Insert necessary records after an instruction
+    void record(program& p, instruction_ref ins, std::size_t wait_id) const;
+    /// Compute weights for an operation
+    std::size_t weight(const operation& op) const;
+};
+
+#else
+
+<%
+interface('schedule_model',
+    virtual('concurrency', returns='std::size_t', const=True),
+    virtual('sched', p='program&', ins='instruction_ref', n='std::size_t', const=True),
+    virtual('wait', p='program&', ins='instruction_ref', wait_id='std::size_t', const=True),
+    virtual('record', p='program&', ins='instruction_ref', wait_id='std::size_t', const=True),
+    virtual('weight', returns='std::size_t', op='const operation&', const=True)
+)
+%>
+
+#endif
+
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx
+
+#endif

tools/te.py

 import string, sys, re, os
 
+
+trivial = [
+    'std::size_t',
+    'instruction_ref'
+]
+
 headers = '''
 #include <algorithm>
 #include <cassert>
@@ -286,7 +292,7 @@ def convert_member(d, struct_name):
                     member['this'] = x
                     if 'const' in t:
                         member['member_const'] = 'const'
-                if t.endswith(('&', '*')):
+                if t.endswith(('&', '*')) or t in trivial:
                     if use_member: member_args.append(x)
                     args.append(arg_name)
                 else: