staging

src/include/migraph/memory_coloring.hpp

@@ -12,8 +12,6 @@ struct memory_coloring
     std::string name() const { return "memory coloring"; }
     void apply(program& p) const;
 };
-    
 } // namespace migraph
 
-
 #endif

src/include/migraph/operators.hpp

@@ -537,12 +537,10 @@ struct div : binary
 struct get_mem_ptr
 {
     std::string name() const { return "get_mem_ptr:" + std::to_string(offset); }
-    shape compute_shape(std::vector<shape> inputs) const
+    shape compute_shape(std::vector<shape> inputs) const { return inputs.at(1); }
+    argument compute(context&, shape output_shape, std::vector<argument> args) const
     {
-        return inputs.at(1);
-    }
-    argument compute(context&, shape output_shape, std::vector<argument> args) const {
-        return {output_shape, args.at(0).data() + offset};
+        return {std::move(output_shape), args.at(0).data() + offset};
     }
     std::size_t offset = 0;
 };
@@ -550,11 +548,9 @@ struct get_mem_ptr
 struct write_literal
 {
     std::string name() const { return "write_literal"; }
-    shape compute_shape(std::vector<shape> inputs) const
+    shape compute_shape(std::vector<shape> inputs) const { return inputs.at(2); }
+    argument compute(context&, shape, std::vector<argument>) const
     {
-        return inputs.at(2);
-    }
-    argument compute(context&, shape output_shape, std::vector<argument> args) const {
         assert(false);
     }
 };
@@ -573,7 +569,6 @@ struct outline
         return {s, nullptr};
     }
 };
-
 } // namespace migraph
 
 #endif

src/include/migraph/program.hpp

@@ -102,7 +102,6 @@ struct program
     private:
     std::unique_ptr<program_impl> impl;
 };
-
 } // namespace migraph
 
 #endif

src/opt/common_header.hpp

@@ -5,6 +5,7 @@
 #include <migraph/instruction.hpp>
 #include <migraph/operators.hpp>
 #include <migraph/iterator_for.hpp>
+#include <migraph/manage_ptr.hpp>
 
 #include <set>
 #include <list>

src/opt/memory_coloring.cpp

@@ -2,10 +2,9 @@
 #include "memory_coloring_impl.hpp"
 
 namespace migraph {
-void memory_coloring::apply(program &p) const
+void memory_coloring::apply(program& p) const
 {
     memory_coloring_impl opt(&p);
     opt.run();
-    
 }
 } // namespace migraph

src/opt/memory_coloring_impl.cpp

@@ -4,29 +4,28 @@ namespace migraph {
 void memory_coloring_impl::run()
 {
     build();
-    if (num_of_lives != 0) {
+    if(num_of_lives != 0)
+    {
         DEBUG(dump("---Before memory coloring---"));
         DEBUG(dump_program());
         DEBUG(dump_intervals());
         // Coloring
-        while (!alloc_queue.empty()) {
-            T_live_interval* interval = alloc_queue.top();
+        while(!alloc_queue.empty())
+        {
+            interval_ptr interval = alloc_queue.top();
             allocate(interval);
             alloc_queue.pop();
         }
         rewrite();
         DEBUG(verify());
-        for (int i = 0; i < num_of_lives; ++i) {
-            free(live_intervals[i]);
-        }
     }
 }
 
-bool memory_coloring_impl::allocate(T_live_interval* interval)
+bool memory_coloring_impl::allocate(interval_ptr interval)
 {
     shape s          = interval->result;
     std::size_t size = s.bytes();
-    if (size == 0)
+    if(size == 0)
         return false;
     std::size_t element_size = size / s.elements();
     T_live_range& segment    = interval->segment;
@@ -35,19 +34,25 @@ bool memory_coloring_impl::allocate(T_live_interval* interval)
     std::unordered_map<long long, T_live_range*> offset2Live;
     offset2Live.clear();
 
-    if (conflict_table.find(vn) != conflict_table.end()) {
+    if(conflict_table.find(vn) != conflict_table.end())
+    {
         std::set<int>& vn_set = conflict_table[vn];
-        for (auto iter = vn_set.begin(), end = vn_set.end(); iter != end; ++iter) {
+        for(auto iter = vn_set.begin(), end = vn_set.end(); iter != end; ++iter)
+        {
             T_live_range* range = live_ranges[*iter];
             long long offset    = range->offset;
-            if (offset != InvalidOffset) {
+            if(offset != InvalidOffset)
+            {
                 conflict_queue.push(range);
-                if (offset2Live.find(offset) == offset2Live.end()) {
+                if(offset2Live.find(offset) == offset2Live.end())
+                {
                     offset2Live[offset] = range;
-                } else {
+                }
+                else
+                {
                     T_live_range* prev = offset2Live[offset];
                     assert(prev->offset == offset);
-                    if (prev->size < range->size)
+                    if(prev->size < range->size)
                         offset2Live[offset] = range;
                 }
             }
@@ -55,15 +60,18 @@ bool memory_coloring_impl::allocate(T_live_interval* interval)
     }
 
     long long offset = 0;
-    while (!conflict_queue.empty()) {
+    while(!conflict_queue.empty())
+    {
         T_live_range* range  = conflict_queue.top();
         long long cur_offset = range->offset;
-        if (offset2Live[cur_offset] == range) {
-            if ((cur_offset > offset) && (cur_offset - offset) >= size) {
+        if(offset2Live[cur_offset] == range)
+        {
+            if((cur_offset > offset) && (cur_offset - offset) >= size)
+            {
                 break;
             }
             offset = cur_offset + range->size;
-            if ((offset % element_size) != 0)
+            if((offset % element_size) != 0)
                 offset += (element_size - (offset % element_size));
         }
         conflict_queue.pop();
@@ -77,7 +85,7 @@ bool memory_coloring_impl::allocate(T_live_interval* interval)
 void memory_coloring_impl::build()
 {
     int num_of_instrs = p_program->get_size();
-    if (num_of_instrs == 0)
+    if(num_of_instrs == 0)
         return;
     int cur_points        = num_of_instrs * 2;
     instruction_ref iter  = std::prev(p_program->end());
@@ -85,45 +93,56 @@ void memory_coloring_impl::build()
     std::vector<instruction_ref> dead_instrs;
     std::set<int> live_set;
     // Build live intervals.
-    do {
+    do
+    {
         const instruction* p_iter     = &(*iter);
-        T_live_interval* def_interval = nullptr;
+        interval_ptr def_interval = nullptr;
         bool isDead                   = false;
-        if (instr2Live.find(p_iter) != instr2Live.end()) {
-            def_interval = instr2Live[p_iter];
+        if(instr2Live.find(p_iter) != instr2Live.end())
+        {
+            def_interval = std::move(instr2Live[p_iter]);
             bool isLit   = isLiteral(iter);
-            if (isAllocate(iter) || isLit) {
+            if(isAllocate(iter) || isLit)
+            {
                 T_live_range& range     = def_interval->segment;
                 def_interval->result    = iter->result;
                 def_interval->isLiteral = isLit;
-                alloc_queue.push(def_interval);
+                alloc_queue.push(std::move(def_interval));
                 range.begin = cur_points;
                 range.size  = (iter->result).bytes();
                 live_set.erase(range.vn);
             }
-        } else if (!isParam(iter) && !isOutline(iter) && !isCheckContext(iter)) {
+        }
+        else if(!isParam(iter) && !isOutline(iter) && !isCheckContext(iter))
+        {
             isDead = true;
         }
         int tieNdx = getInputTieNdx(iter);
-        if (!iter->arguments.empty()) {
+        if(!iter->arguments.empty())
+        {
             int cnt = -1;
-            for (auto&& arg : iter->arguments) {
+            for(auto&& arg : iter->arguments)
+            {
                 cnt++;
-                if (isParam(arg) || isOutline(arg)) {
-                    if (isOutputParam(arg))
+                if(isParam(arg) || isOutline(arg))
+                {
+                    if(isOutputParam(arg))
                         isDead = false;
                     continue;
                 }
                 const instruction* p_arg = &(*arg);
-                if (cnt == tieNdx) {
+                if(cnt == tieNdx)
+                {
                     // input memory is used as this instruction's output.
                     // def is considered as use. Coalesce the live intervals.
                     def_interval->addUse(cur_points);
                     instr2Live[p_arg] = def_interval;
-                } else if (instr2Live.find(p_arg) == instr2Live.end()) {
+                }
+                else if(instr2Live.find(p_arg) == instr2Live.end())
+                {
                     // First time see a use, create a live interval.
                     int id                    = num_of_lives++;
-                    T_live_interval* interval = new live_interval();
+                    interval_ptr interval(new live_interval());
                     interval->id              = id;
                     interval->segment.end     = cur_points;
                     interval->segment.vn      = ++max_value_number;
@@ -131,20 +150,22 @@ void memory_coloring_impl::build()
                     instr2Live[p_arg] = interval;
                     addConflicts(live_set, max_value_number);
                     live_set.insert(max_value_number);
-                    live_intervals[id] = interval;
+                    live_intervals[id]            = std::move(interval);
                     live_ranges[max_value_number] = &(interval->segment);
-                } else {
-                    T_live_interval* interval = instr2Live[p_arg];
+                }
+                else
+                {
+                    interval_ptr interval = instr2Live[p_arg];
                     interval->addUse(cur_points);
                     DEBUG(assert(live_set.find(interval->id) != live_set.end()));
                 }
             }
         }
-        if (isDead)
+        if(isDead)
             dead_instrs.push_back(iter);
         cur_points -= 2;
         iter = std::prev(iter);
-    } while (iter != begin);
+    } while(iter != begin);
 }
 
 void memory_coloring_impl::rewrite()
@@ -152,22 +173,29 @@ 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));
+    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)) {
+    for(auto ins : iterator_for(*p_program))
+    {
         const instruction* p_iter = &(*ins);
-        if (instr2Live.find(p_iter) != instr2Live.end()) {
-            T_live_interval* interval = instr2Live[p_iter];
-            if (interval->get_offset() == InvalidOffset) {
+        if(instr2Live.find(p_iter) != instr2Live.end())
+        {
+            interval_ptr interval = instr2Live[p_iter];
+            if(interval->get_offset() == InvalidOffset)
+            {
                 DEBUG(assert((interval->get_begin() == InvalidOffset) ||
                              interval->result.bytes() == 0));
                 continue;
             }
             std::size_t offset = interval->get_offset();
-            if (isAllocate(ins)) {
-                p_program->replace_instruction(ins, get_mem_ptr{offset}, scratch_param, ins->arguments.at(0));
-            } else if (isLiteral(ins)) {
+            if(isAllocate(ins))
+            {
+                p_program->replace_instruction(
+                    ins, get_mem_ptr{offset}, scratch_param, ins->arguments.at(0));
+            }
+            else if(isLiteral(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);
@@ -179,31 +207,28 @@ void memory_coloring_impl::rewrite()
 }
 
 #ifdef DEBUG_OPT
-void memory_coloring_impl::dump(std::string str)
-{
-    std::cout << str << std::endl;
-    
-}
+void memory_coloring_impl::dump(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_program() { std::cout << *p_program << std::endl; }
 
 void memory_coloring_impl::dump_intervals()
 {
-    if (num_of_lives > 0) {
+    if(num_of_lives > 0)
+    {
         std::cout << "---live intervals ---" << std::endl;
-        for (int i = 0; i < num_of_lives; ++i) {
-            T_live_interval* interval = live_intervals[i];
+        for(int i = 0; i < num_of_lives; ++i)
+        {
+            interval_ptr interval = live_intervals[i];
             interval->dump();
         }
         std::cout << "---conflict table---" << std::endl;
-        for (int i = 0; i <= max_value_number; ++i) {
+        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.begin(), end = table.end(); iter != end; ++iter) {
+            for(auto iter = table.begin(), end = table.end(); iter != end; ++iter)
+            {
                 std::cout << (*iter) << ",";
             }
         }
@@ -213,20 +238,24 @@ void memory_coloring_impl::dump_intervals()
 
 void memory_coloring_impl::verify()
 {
-    if (num_of_lives > 0) {
-        for (int i = 0; i < num_of_lives; ++i) {
-            T_live_interval* interval = live_intervals[i];
+    if(num_of_lives > 0)
+    {
+        for(int i = 0; i < num_of_lives; ++i)
+        {
+            interval_ptr interval = live_intervals[i];
             T_live_range& segment     = interval->segment;
-            if (segment.offset == InvalidOffset)
+            if(segment.offset == InvalidOffset)
                 continue;
             int vn = segment.vn;
-            if (conflict_table.find(vn) != conflict_table.end()) {
+            if(conflict_table.find(vn) != conflict_table.end())
+            {
                 std::set<int>& vn_set = conflict_table[vn];
-                for (auto iter = vn_set.begin(), end = vn_set.end(); iter != end; ++iter) {
+                for(auto iter = vn_set.begin(), end = vn_set.end(); iter != end; ++iter)
+                {
                     T_live_range* range = live_ranges[*iter];
-                    if (range->offset == InvalidOffset)
+                    if(range->offset == InvalidOffset)
                         continue;
-                    if (!isDisjoin(*range, segment))
+                    if(!isDisjoin(*range, segment))
                         assert(false);
                 }
             }
@@ -240,7 +269,8 @@ void live_range::dump()
 {
     std::cout << " segment:" << vn;
     std::cout << " [" << GET_INS_ENUM(begin) << ", " << GET_INS_ENUM(end) << "]";
-    if (offset != InvalidOffset) {
+    if(offset != InvalidOffset)
+    {
         std::cout << " mem:";
         std::cout << " [" << offset << "," << offset + size - 1 << "]";
     }
@@ -252,17 +282,17 @@ void live_interval::dump()
     std::cout << "id:" << id;
     segment.dump();
     std::cout << " uses:";
-    for (auto iter = use_points.begin(), end = use_points.end(); iter != end; ++iter) {
+    for(auto iter = use_points.begin(), end = use_points.end(); iter != end; ++iter)
+    {
         int& use = *iter;
         std::cout << " " << GET_INS_ENUM(use) << ",";
     }
 
-    if (isLiteral)
+    if(isLiteral)
         std::cout << " literal";
     std::cout << " " << result;
     std::cout << std::endl;
 }
 
 #endif
-    
 } // namespace migraph

src/opt/memory_coloring_impl.hpp

@@ -6,7 +6,9 @@ namespace migraph {
 
 #define InvalidOffset -1
 
-typedef struct live_range {
+
+typedef 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.
@@ -17,12 +19,15 @@ typedef struct live_range {
 #endif
 } T_live_range;
 
-typedef struct live_interval {
-    explicit live_interval() { init(); }
-    
-    void init() {
-        id = -1; isLiteral = false;
-        segment = { -1, -1, InvalidOffset, -1, 0};
+typedef struct live_interval
+{
+    live_interval() { init(); }
+    ~live_interval() {}
+    void init()
+    {
+        id        = -1;
+        isLiteral = false;
+        segment   = {-1, -1, InvalidOffset, -1, 0};
     }
     void addUse(int use) { use_points.push_front(use); }
     int get_begin() const { return segment.begin; }
@@ -41,9 +46,21 @@ typedef struct live_interval {
 
 } T_live_interval;
 
-struct memory_coloring_impl {
-    explicit memory_coloring_impl(program *p) : p_program(p)
+//   #define unique_interval_ptr std::unique_ptr<T_live_interval>
+#define interval_ptr T_live_interval*
+
+struct memory_coloring_impl
+{
+    memory_coloring_impl(){}
+    memory_coloring_impl(program* p) : p_program(p)
     {
+        init();
+    }
+    ~memory_coloring_impl() {
+        for(int i = 0; i < num_of_lives; ++i)
+            free(live_intervals[i]);
+    }
+    void init() {
         instr2Live.clear();
         live_intervals.clear();
         live_ranges.clear();
@@ -52,10 +69,11 @@ struct memory_coloring_impl {
         max_value_number = -1;
         required_bytes   = 0;
     }
-    bool allocate(T_live_interval*);
+    bool allocate(interval_ptr);
     void addConflicts(std::set<int>& live_set, int val)
     {
-        for (auto iter = live_set.begin(), end = live_set.end(); iter != end; ++ iter) {
+        for(auto iter = live_set.begin(), end = live_set.end(); iter != end; ++iter)
+        {
             conflict_table[*iter].insert(val);
             conflict_table[val].insert(*iter);
         }
@@ -63,6 +81,7 @@ struct memory_coloring_impl {
     void build();
     void run();
     void rewrite();
+
     private:
     bool isParam(const instruction_ref ins) { return ins->op.name() == "@param"; }
     bool isOutputParam(const instruction_ref ins)
@@ -78,19 +97,22 @@ struct memory_coloring_impl {
     bool isLiteral(const instruction_ref ins) { return ins->op.name() == "@literal"; }
     bool isCheckContext(const instruction_ref ins) { return ins->op.name() == "check_context"; }
     bool isTranspose(const instruction_ref ins) { return ins->op.name() == "transpose"; }
-    int getInputTieNdx(const instruction_ref ins) {
-        if (isTranspose(ins))
+    int getInputTieNdx(const instruction_ref ins)
+    {
+        if(isTranspose(ins))
             return 0;
         int cnt           = -1;
         int last_allocate = -1;
-        for (auto&& arg : ins->arguments) {
+        for(auto&& arg : ins->arguments)
+        {
             cnt++;
-            if (isAllocate(arg))
+            if(isAllocate(arg))
                 last_allocate = cnt;
         }
         return last_allocate;
     }
-    bool isDisjoin(T_live_range& range1, T_live_range& range2) {
+    bool isDisjoin(T_live_range& range1, T_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));
@@ -102,39 +124,44 @@ struct memory_coloring_impl {
     void dump_intervals();
     void verify();
 #endif
-    struct ordering { 
-        bool operator() (const T_live_interval* I1, const T_live_interval* I2) const
+    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) {
+            if(len1 != len2)
+            {
                 return (len1 < len2) ? true : false;
-            } else if (I1->result.bytes() != I2->result.bytes()) {
+            }
+            else if(I1->result.bytes() != I2->result.bytes())
+            {
                 return (I1->result.bytes() < I2->result.bytes()) ? true : false;
-            } else {
+            }
+            else
+            {
                 return I1->id > I2->id;
             }
         }
-        bool operator() (const T_live_range* I1, const T_live_range* I2) const
+        bool operator()(const T_live_range* I1, const T_live_range* I2) const
         {
             return (I1->offset > I2->offset);
         }
     };
     program* p_program;
-    std::unordered_map<const instruction*, T_live_interval*> instr2Live;
+    std::unordered_map<const instruction*, interval_ptr> instr2Live;
     // Map live interval Id to live interval.
-    std::unordered_map<int, T_live_interval*> live_intervals;
+    std::unordered_map<int, interval_ptr> live_intervals;
     // Map live range value number to live range.
     std::unordered_map<int, T_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<T_live_interval*, std::vector<T_live_interval*>, ordering> alloc_queue;
+    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/targets/gpu/hip.cpp

@@ -94,7 +94,5 @@ 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

@@ -85,12 +85,10 @@ struct hip_write
 struct hip_memcpy
 {
     std::string name() const { return "hip_memcpy"; }
-    shape compute_shape(std::vector<shape> inputs) const
+    shape compute_shape(std::vector<shape> inputs) const { return inputs.at(2); }
+    argument compute(context&, shape output_shape, std::vector<argument> args) 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());
+        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();
@@ -98,9 +96,7 @@ struct hip_memcpy
         return {output_shape, dst};
     }
 };
-
 } // namespace gpu
-
 } // namespace migraph
 
 #endif

src/targets/gpu/write_literals.cpp

@@ -37,13 +37,12 @@ void write_literals::apply(program& p) const
             p.replace_instruction(ins, hip_load_literal{a.get_shape(), n});
         }
 #else
-        if (ins->op.name() == "write_literal") {
+        if(ins->op.name() == "write_literal")
+        {
             p.replace_instruction(ins, hip_memcpy{}, ins->arguments);
         }
 #endif
     }
 }
-
 } // namespace gpu
-
 } // namespace migraph