Merge pull request #46 from ROCmSoftwarePlatform/memory_coloring

Memory coloring

src/CMakeLists.txt

@@ -9,6 +9,8 @@ add_library(migraph
     program.cpp
     shape.cpp
     simplify_reshapes.cpp
+    opt/memory_coloring.cpp
+    opt/memory_coloring_impl.cpp
 )
 rocm_clang_tidy_check(migraph)
 target_include_directories(migraph PUBLIC $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/include>)
@@ -18,3 +20,9 @@ add_subdirectory(targets/cpu)
 if(MIGRAPH_ENABLE_GPU)
 add_subdirectory(targets/gpu)
 endif()
+
+#install (TARGETS migraph
+#   LIBRARY DESTINATION /opt/rocm/lib)
+
+#install (DIRECTORY include/migraph DESTINATION /opt/rocm/include)
+ 

src/include/migraph/memory_coloring.hpp

+#ifndef MIGRAPH_GUARD_RTGLIB_MEMORY_COLORING_HPP
+#define MIGRAPH_GUARD_RTGLIB_MEMORY_COLORING_HPP
+
+#include <string>
+#include <migraph/instruction_ref.hpp>
+
+namespace migraph {
+struct program;
+
+struct memory_coloring
+{
+    std::string name() const { return "memory coloring"; }
+    void apply(program& p) const;
+};
+} // namespace migraph
+
+#endif

src/include/migraph/operators.hpp

@@ -534,6 +534,27 @@ struct div : binary
     std::string name() const { return "div"; }
 };
 
+struct get_mem_ptr
+{
+    std::string name() const { return "get_mem_ptr:" + std::to_string(offset); }
+    shape compute_shape(std::vector<shape> inputs) const { return inputs.at(1); }
+    argument compute(context&, const shape& output_shape, const std::vector<argument>& args) const
+    {
+        return {std::move(output_shape), args.at(0).data() + offset};
+    }
+    std::size_t offset = 0;
+};
+
+struct write_literal
+{
+    std::string name() const { return "write_literal"; }
+    shape compute_shape(std::vector<shape> inputs) const { return inputs.at(2); }
+    argument compute(context&, const shape&, const std::vector<argument>&) const
+    {
+        MIGRAPH_THROW("not computable");
+    }
+};
+
 struct outline
 {
     shape s;
@@ -548,7 +569,6 @@ struct outline
         return {s, nullptr};
     }
 };
-
 } // namespace migraph
 
 #endif

src/include/migraph/program.hpp

@@ -100,7 +100,6 @@ struct program
     private:
     std::unique_ptr<program_impl> impl;
 };
-
 } // namespace migraph
 
 #endif

src/opt/common_header.hpp

+#ifndef MIGRAPH_GUARD_RTGLIB_COMMON_HEADER_HPP
+#define MIGRAPH_GUARD_RTGLIB_COMMON_HEADER_HPP
+#include <migraph/program.hpp>
+#include <migraph/stringutils.hpp>
+#include <migraph/instruction.hpp>
+#include <migraph/operators.hpp>
+#include <migraph/iterator_for.hpp>
+
+#include <set>
+#include <list>
+#include <vector>
+#include <queue>
+
+//#define MIGRAPH_DEBUG_OPT
+
+#ifdef MIGRAPH_DEBUG_OPT
+#define MIGRAPH_DEBUG(s) s
+#else
+#define MIGRAPH_DEBUG(s)
+#endif // MIGRAPH_DEBUG_OPT
+
+#endif // MIGRAPH_GUARD_RTGLIB_COMMON_HEADER_HPP

src/opt/memory_coloring.cpp

+#include <migraph/memory_coloring.hpp>
+#include "memory_coloring_impl.hpp"
+
+namespace migraph {
+
+void memory_coloring::apply(program& p) const
+{
+    memory_coloring_impl opt(&p);
+    opt.run();
+}
+} // namespace migraph

src/opt/memory_coloring_impl.cpp

+#include "memory_coloring_impl.hpp"
+
+namespace migraph {
+
+void memory_coloring_impl::run()
+{
+    build();
+    if(num_of_lives != 0)
+    {
+        MIGRAPH_DEBUG(dump("---Before memory coloring---"));
+        MIGRAPH_DEBUG(dump_program());
+        MIGRAPH_DEBUG(dump_intervals());
+        // Coloring
+        while(!alloc_queue.empty())
+        {
+            interval_ptr interval = alloc_queue.top();
+            allocate(interval);
+            alloc_queue.pop();
+        }
+        rewrite();
+        MIGRAPH_DEBUG(verify());
+    }
+}
+
+bool memory_coloring_impl::allocate(interval_ptr interval)
+{
+    shape s          = interval->result;
+    std::size_t size = s.bytes();
+    if(size == 0)
+        return false;
+    std::size_t element_size = size / s.elements();
+    live_range& segment      = interval->segment;
+    int vn                   = segment.vn;
+    std::priority_queue<live_range*, std::vector<live_range*>, ordering> conflict_queue;
+    std::unordered_map<long long, live_range*> offset2_live;
+    offset2_live.clear();
+
+    if(conflict_table.find(vn) != conflict_table.end())
+    {
+        std::set<int>& vn_set = conflict_table[vn];
+        for(auto& iter : vn_set)
+        {
+            live_range* range = live_ranges[iter];
+            long long offset  = range->offset;
+            if(offset != InvalidOffset)
+            {
+                conflict_queue.push(range);
+                if(offset2_live.find(offset) == offset2_live.end())
+                {
+                    offset2_live[offset] = range;
+                }
+                else
+                {
+                    live_range* prev = offset2_live[offset];
+                    assert(prev->offset == offset);
+                    if(prev->size < range->size)
+                        offset2_live[offset] = range;
+                }
+            }
+        }
+    }
+
+    long long offset = 0;
+    while(!conflict_queue.empty())
+    {
+        live_range* range    = conflict_queue.top();
+        long long cur_offset = range->offset;
+        if(offset2_live[cur_offset] == range)
+        {
+            if((cur_offset > offset) && (cur_offset - offset) >= size)
+            {
+                break;
+            }
+            offset = cur_offset + range->size;
+            if((offset % element_size) != 0)
+                offset += (element_size - (offset % element_size));
+        }
+        conflict_queue.pop();
+    }
+    segment.offset = offset;
+    MIGRAPH_DEBUG(segment.dump());
+    required_bytes = std::max(required_bytes, offset + segment.size);
+    return true;
+}
+
+void memory_coloring_impl::build()
+{
+    std::size_t num_of_instrs = p_program->size();
+    if(num_of_instrs == 0)
+        return;
+
+    int cur_points        = num_of_instrs * 2;
+    instruction_ref iter  = std::prev(p_program->end());
+    instruction_ref begin = p_program->begin();
+    std::vector<instruction_ref> dead_instrs;
+    std::set<int> live_set;
+    // Build live intervals.
+    live_intervals.resize(num_of_instrs);
+    do
+    {
+        const instruction* p_iter = &(*iter);
+        interval_ptr def_interval = nullptr;
+        bool is_dead              = false;
+        if(instr2_live.find(p_iter) != instr2_live.end())
+        {
+            def_interval = instr2_live[p_iter];
+            bool is_lit  = is_literal(iter);
+            if(is_allocate(iter) || is_lit)
+            {
+                live_range& range        = def_interval->segment;
+                def_interval->result     = iter->result;
+                def_interval->is_literal = is_lit;
+                alloc_queue.push(def_interval);
+                range.begin = cur_points;
+                range.size  = (iter->result).bytes();
+                live_set.erase(range.vn);
+            }
+        }
+        else if(!is_param(iter) && !is_outline(iter) && !is_check_context(iter))
+        {
+            is_dead = true;
+        }
+        int tie_ndx = get_input_tie_ndx(iter);
+        int cnt     = -1;
+        for(auto&& arg : iter->arguments)
+        {
+            cnt++;
+            if(is_param(arg) || is_outline(arg))
+            {
+                if(is_output_param(arg))
+                    is_dead = false;
+                continue;
+            }
+            const instruction* p_arg = &(*arg);
+            if(cnt == tie_ndx)
+            {
+                // input memory is used as this instruction's output.
+                // def is considered as use. Coalesce the live intervals.
+                assert(def_interval != nullptr);
+                def_interval->add_use(cur_points);
+                instr2_live[p_arg] = def_interval;
+            }
+            else if(instr2_live.find(p_arg) == instr2_live.end())
+            {
+                // First time see a use, create a live interval.
+                int id                = num_of_lives++;
+                interval_ptr interval = &(live_intervals[id]);
+                interval->id          = id;
+                interval->segment.end = cur_points;
+                interval->segment.vn  = ++max_value_number;
+                interval->add_use(cur_points);
+                instr2_live[p_arg] = interval;
+                add_conflicts(live_set, max_value_number);
+                live_set.insert(max_value_number);
+                live_ranges[max_value_number] = &(interval->segment);
+            }
+            else
+            {
+                interval_ptr interval = instr2_live[p_arg];
+                interval->add_use(cur_points);
+                assert(live_set.find(interval->id) != live_set.end());
+            }
+        }
+        if(is_dead)
+            dead_instrs.push_back(iter);
+        cur_points -= 2;
+        iter = std::prev(iter);
+    } while(iter != begin);
+}
+
+void memory_coloring_impl::rewrite()
+{
+    instruction_ref end           = p_program->end();
+    instruction_ref scratch_param = end;
+    std::vector<std::size_t> dims;
+    dims.push_back(required_bytes / sizeof(float));
+    shape s       = {shape::float_type, dims};
+    scratch_param = p_program->add_parameter("scratch", s);
+    for(auto ins : iterator_for(*p_program))
+    {
+        const instruction* p_iter = &(*ins);
+        if(instr2_live.find(p_iter) != instr2_live.end())
+        {
+            interval_ptr interval = instr2_live[p_iter];
+            if(interval->get_offset() == InvalidOffset)
+            {
+                assert((interval->get_begin() == InvalidOffset) || interval->result.bytes() == 0);
+                continue;
+            }
+            std::size_t offset = interval->get_offset();
+            if(is_allocate(ins))
+            {
+                p_program->replace_instruction(
+                    ins, get_mem_ptr{offset}, scratch_param, ins->arguments.at(0));
+            }
+            else if(is_literal(ins))
+            {
+                auto pre   = p_program->add_literal(ins->lit);
+                auto index = p_program->add_literal(offset);
+                p_program->replace_instruction(ins, write_literal{}, scratch_param, index, pre);
+            }
+        }
+    }
+    MIGRAPH_DEBUG(dump("---After rewrite---"));
+    MIGRAPH_DEBUG(dump_program());
+}
+
+#ifdef MIGRAPH_DEBUG_OPT
+void memory_coloring_impl::dump(const std::string str) { std::cout << str << std::endl; }
+
+void memory_coloring_impl::dump_program() { std::cout << *p_program << std::endl; }
+
+void memory_coloring_impl::dump_intervals()
+{
+    if(num_of_lives > 0)
+    {
+        std::cout << "---live intervals ---" << std::endl;
+        for(int i = 0; i < num_of_lives; ++i)
+        {
+            live_interval& interval = live_intervals[i];
+            interval.dump();
+        }
+        std::cout << "---conflict table---" << std::endl;
+        for(int i = 0; i <= max_value_number; ++i)
+        {
+            std::cout << " segment:" << i;
+            std::cout << " =>";
+            std::set<int>& table = conflict_table[i];
+            for(auto& iter : table)
+            {
+                std::cout << (iter) << ",";
+            }
+        }
+        std::cout << std::endl;
+    }
+}
+
+void memory_coloring_impl::verify()
+{
+    if(num_of_lives > 0)
+    {
+        for(int i = 0; i < num_of_lives; ++i)
+        {
+            live_interval& interval = live_intervals[i];
+            live_range& segment     = interval.segment;
+            if(segment.offset == InvalidOffset)
+                continue;
+            int vn = segment.vn;
+            if(conflict_table.find(vn) != conflict_table.end())
+            {
+                std::set<int>& vn_set = conflict_table[vn];
+                for(auto& iter : vn_set)
+                {
+                    live_range* range = live_ranges[iter];
+                    if(range->offset == InvalidOffset)
+                        continue;
+                    if(!is_disjoin(*range, segment))
+                        assert(false);
+                }
+            }
+        }
+    }
+}
+
+#define GET_INS_ENUM(x) (((x) > 0) ? (((x) >> 1) - 1) : InvalidOffset)
+
+void live_range::dump()
+{
+    std::cout << " segment:" << vn;
+    std::cout << " [" << GET_INS_ENUM(begin) << ", " << GET_INS_ENUM(end) << "]";
+    if(offset != InvalidOffset)
+    {
+        std::cout << " mem:";
+        std::cout << " [" << offset << "," << offset + size - 1 << "]";
+    }
+    std::cout << std::endl;
+}
+
+void live_interval::dump()
+{
+    std::cout << "id:" << id;
+    segment.dump();
+    std::cout << " uses:";
+    for(auto& iter : use_points)
+    {
+        std::cout << " " << GET_INS_ENUM(iter) << ",";
+    }
+
+    if(is_literal)
+        std::cout << " literal";
+    std::cout << " " << result;
+    std::cout << std::endl;
+}
+
+#endif
+} // namespace migraph

src/opt/memory_coloring_impl.hpp

+#ifndef MIGRAPH_GUARD_RTGLIB_MEMORY_COLORING_IMPL_HPP
+#define MIGRAPH_GUARD_RTGLIB_MEMORY_COLORING_IMPL_HPP
+#include "common_header.hpp"
+
+namespace migraph {
+
+static const int InvalidOffset = -1;
+
+struct live_range
+{
+    int begin;        // begin point in the instruction stream.
+    int end;          // end point in the instruction stream.
+    long long offset; // offset to base pointer of allocated memory trunk.
+    int vn;           // value number that identifies this live_range.
+    long long size;   // size of required memory in bytes
+#ifdef MIGRAPH_DEBUG_OPT
+    void dump();
+#endif
+};
+
+struct live_interval
+{
+    live_interval() : segment({-1, -1, InvalidOffset, -1, 0})
+    {
+        id         = -1;
+        is_literal = false;
+    }
+
+    void add_use(int use) { use_points.push_front(use); }
+    int get_begin() const { return segment.begin; }
+    int get_end() const { return segment.end; }
+    long long get_offset() const { return segment.offset; }
+
+#ifdef MIGRAPH_DEBUG_OPT
+    void dump();
+#endif
+
+    live_range segment;
+    int id;
+    std::list<int> use_points;
+    shape result;
+    bool is_literal;
+};
+
+typedef live_interval* interval_ptr;
+
+struct memory_coloring_impl
+{
+    memory_coloring_impl(program* p) : p_program(p)
+    {
+        instr2_live.clear();
+        live_ranges.clear();
+        conflict_table.clear();
+        num_of_lives     = 0;
+        max_value_number = -1;
+        required_bytes   = 0;
+    }
+    bool allocate(interval_ptr);
+    void add_conflicts(std::set<int>& live_set, int val)
+    {
+        for(auto& iter : live_set)
+        {
+            conflict_table[iter].insert(val);
+            conflict_table[val].insert(iter);
+        }
+    }
+    void build();
+    void run();
+    void rewrite();
+
+    private:
+    static bool is_param(const instruction_ref ins) { return ins->op.name() == "@param"; }
+    static bool is_output_param(const instruction_ref ins)
+    {
+        return is_param(ins) && any_cast<builtin::param>(ins->op).parameter == "output";
+    }
+    static bool is_allocate(const instruction_ref ins) { return ins->op.name() == "hip::allocate"; }
+    static bool is_outline(const instruction_ref ins) { return ins->op.name() == "@outline"; }
+    static bool is_literal(const instruction_ref ins) { return ins->op.name() == "@literal"; }
+    static bool is_check_context(const instruction_ref ins)
+    {
+        return ins->op.name() == "check_context";
+    }
+    static bool is_transpose(const instruction_ref ins) { return ins->op.name() == "transpose"; }
+    int get_input_tie_ndx(const instruction_ref ins)
+    {
+        if(is_transpose(ins))
+            return 0;
+        int cnt           = -1;
+        int last_allocate = -1;
+        for(auto&& arg : ins->arguments)
+        {
+            cnt++;
+            if(is_allocate(arg))
+                last_allocate = cnt;
+        }
+        return last_allocate;
+    }
+#ifdef MIGRAPH_DEBUG_OPT
+    static bool is_disjoin(live_range& range1, live_range& range2)
+    {
+        long long end1 = range1.offset + range1.size - 1;
+        long long end2 = range2.offset + range2.size - 1;
+        return ((end1 < range2.offset) || (end2 < range1.offset));
+    }
+    void dump(const std::string);
+    void dump_program();
+    void dump_intervals();
+    void verify();
+#endif
+    struct ordering
+    {
+        bool operator()(const interval_ptr i1, const interval_ptr i2) const
+        {
+            int len1 = i1->get_end() - i1->get_begin();
+            int len2 = i2->get_end() - i2->get_begin();
+            if(len1 != len2)
+            {
+                return (len1 < len2) ? true : false;
+            }
+            else if(i1->result.bytes() != i2->result.bytes())
+            {
+                return (i1->result.bytes() < i2->result.bytes()) ? true : false;
+            }
+            else
+            {
+                return i1->id > i2->id;
+            }
+        }
+        bool operator()(const live_range* i1, const live_range* i2) const
+        {
+            return (i1->offset > i2->offset);
+        }
+    };
+    program* p_program;
+    std::unordered_map<const instruction*, interval_ptr> instr2_live;
+    // universe of live intervals.
+    std::vector<live_interval> live_intervals;
+    // Map live range value number to live range.
+    std::unordered_map<int, live_range*> live_ranges;
+    // Map live range value number to a set of conflicting live ranges' value numbers.
+    std::unordered_map<int, std::set<int>> conflict_table;
+    // Priority queue for coloring.
+    std::priority_queue<interval_ptr, std::vector<interval_ptr>, ordering> alloc_queue;
+
+    int num_of_lives;
+    int max_value_number;
+    long long required_bytes;
+};
+} // namespace migraph
+#endif

src/program.cpp

@@ -425,5 +425,4 @@ std::ostream& operator<<(std::ostream& os, const program& p)
     print_program(os, p, [](auto&&...) {});
     return os;
 }
-
 } // namespace migraph

src/targets/cpu/CMakeLists.txt

@@ -13,3 +13,8 @@ target_link_libraries(migraph_cpu migraph Threads::Threads)
 target_include_directories(migraph_cpu PUBLIC $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/include>)
 target_include_directories(migraph_cpu PRIVATE ${BLAZE_INCLUDE})
 target_compile_definitions(migraph_cpu PRIVATE -DBLAZE_USE_CPP_THREADS)
+
+#install (TARGETS migraph_cpu
+#  LIBRARY DESTINATION /opt/rocm/lib)
+#install (DIRECTORY include/migraph DESTINATION /opt/rocm/include)
+

src/targets/gpu/CMakeLists.txt

@@ -33,3 +33,11 @@ add_library(migraph_gpu
 rocm_clang_tidy_check(migraph_gpu)
 target_link_libraries(migraph_gpu migraph MIOpen migraph_device roc::rocblas)
 target_include_directories(migraph_gpu PUBLIC $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/include>)
+
+#install (TARGETS migraph_gpu
+#  LIBRARY DESTINATION /opt/rocm/lib)
+#install (DIRECTORY include/migraph DESTINATION /opt/rocm/include)
+
+#install (TARGETS migraph_device
+#  LIBRARY DESTINATION /opt/rocm/lib)
+#install (DIRECTORY include/migraph DESTINATION /opt/rocm/include)

src/targets/gpu/hip.cpp

@@ -58,7 +58,9 @@ std::vector<T> read_from_gpu(const void* x, std::size_t sz)
 
 hip_ptr write_to_gpu(const void* x, std::size_t sz, bool host = false)
 {
+
     auto result = allocate_gpu(sz, host);
+    // gpu_sync();
     auto status = hipMemcpy(result.get(), x, sz, hipMemcpyHostToDevice);
     if(status != hipSuccess)
         MIGRAPH_THROW("Copy to gpu failed: " + hip_error(status));
@@ -90,6 +92,9 @@ argument from_gpu(argument arg)
 
 void gpu_sync() { hipDeviceSynchronize(); }
 
+void copy_to_gpu(char* dst, const char* src, std::size_t size)
+{
+    hipMemcpy(dst, src, size, hipMemcpyHostToDevice);
+}
 } // namespace gpu
-
 } // namespace migraph

src/targets/gpu/include/migraph/gpu/hip.hpp

@@ -14,6 +14,7 @@ migraph::argument to_gpu(migraph::argument arg, bool host = false);
 migraph::argument from_gpu(migraph::argument arg);
 
 void gpu_sync();
+void copy_to_gpu(char* dst, const char* src, std::size_t size);
 
 struct hip_allocate
 {
@@ -81,8 +82,21 @@ struct hip_write
     }
 };
 
+struct hip_memcpy
+{
+    std::string name() const { return "hip_memcpy"; }
+    shape compute_shape(std::vector<shape> inputs) const { return inputs.at(2); }
+    argument compute(context&, shape output_shape, std::vector<argument> args) const
+    {
+        std::size_t* p_data = reinterpret_cast<std::size_t*>(args.at(1).data());
+        char* dst           = args.at(0).data() + p_data[0];
+        const char* src     = args.at(2).data();
+        std::size_t size    = args.at(2).get_shape().bytes();
+        copy_to_gpu(dst, src, size);
+        return {output_shape, dst};
+    }
+};
 } // namespace gpu
-
 } // namespace migraph
 
 #endif

src/targets/gpu/lowering.cpp

@@ -108,7 +108,9 @@ struct miopen_convolution
         workspace_shape = shape{shape::int8_type, {workspace_size}};
 
         auto x         = to_gpu(generate_argument(inputs[0]->get_shape()));
+        gpu_sync();
         auto w         = to_gpu(generate_argument(inputs[1]->get_shape()));
+        gpu_sync();
         auto y         = to_gpu(generate_argument(output_shape));
         auto workspace = allocate_gpu(workspace_shape);
 

src/targets/gpu/target.cpp

 #include <migraph/gpu/target.hpp>
 #include <migraph/gpu/lowering.hpp>
+#include <migraph/memory_coloring.hpp>
 #include <migraph/gpu/write_literals.hpp>
 #include <migraph/gpu/context.hpp>
 #include <migraph/gpu/eliminate_workspace.hpp>
@@ -28,13 +29,16 @@ std::vector<pass> target::get_passes(migraph::context& gctx) const
         simplify_reshapes{},
         dead_code_elimination{},
         lowering{ctx},
+        memory_coloring{},
         fuse_ops{},
         dead_code_elimination{},
-        eliminate_workspace{},
+            //            eliminate_workspace{},
+
         eliminate_contiguous{},
         dead_code_elimination{},
+
         write_literals{&ctx},
-        eliminate_allocation{},
+            //        eliminate_allocation{},
         check_context<context>{},
         dead_code_elimination{}
     };

src/targets/gpu/write_literals.cpp

@@ -28,6 +28,7 @@ void write_literals::apply(program& p) const
     assert(ctx != nullptr);
     for(auto ins : iterator_for(p))
     {
+#if 0        
         if(ins->op.name() == "@literal")
         {
             argument a    = to_gpu(ins->lit.get_argument());
@@ -35,9 +36,13 @@ void write_literals::apply(program& p) const
             ctx->literals.push_back(a);
             p.replace_instruction(ins, hip_load_literal{a.get_shape(), n});
         }
+#else
+        if(ins->op.name() == "write_literal")
+        {
+            p.replace_instruction(ins, hip_memcpy{}, ins->arguments);
+        }
+#endif
     }
 }
-
 } // namespace gpu
-
 } // namespace migraph