!2754 optimize memory reuse according streams info

Merge pull request !2754 from laiyongqiang/mem_opt

mindspore/ccsrc/device/memory_manager.cc

@@ -99,6 +99,11 @@ uint8_t *MemoryManager::MallocStaticMem(size_t size, bool communication_mem) {
   } else {
     align_size = GetCommonAlignSize(size);
   }
+
+  MS_LOG(INFO) << "Malloc Memory for Static: total[" << device_mem_size_ << "](dynamic[" << total_dynamic_size_
+               << "] static[" << total_static_size_ << "])"
+               << " malloc [" << align_size << "] communication_mem: " << communication_mem;
+
   if (static_mem_offset_ < align_size) {
     MS_LOG(EXCEPTION) << "Out of memory!!! total[" << device_mem_size_ << "](dynamic[" << total_dynamic_size_
                       << "] static[" << total_static_size_ << "])"
@@ -126,6 +131,11 @@ uint8_t *MemoryManager::MallocDynamicMem(size_t size, bool communication_mem) {
   } else {
     align_size = GetCommonAlignSize(size);
   }
+
+  MS_LOG(INFO) << "Malloc Memory for Dynamic: total[" << device_mem_size_ << "](dynamic[" << total_dynamic_size_
+               << "] static[" << total_static_size_ << "])"
+               << " malloc [" << align_size << "] communication_mem: " << communication_mem;
+
   uint64_t offset = dynamic_mem_offset_;
   auto new_offset = dynamic_mem_offset_ + align_size;
   if (new_offset > static_mem_offset_) {

mindspore/ccsrc/pre_activate/mem_reuse/mem_reuse.cc

@@ -329,22 +329,25 @@ void MemReuseUtil::SetSummaryNodesRefCount() {
     return;
   }
 
+  size_t total_summary_size = 0;
   for (auto &node_item : summary_nodes) {
     auto node = node_item.second.first;
     size_t index = IntToSize(node_item.second.second);
-    MS_LOG(INFO) << "set summary node's ref count, node: " << node->fullname_with_scope() << " index: " << index;
     if (kernel_output_refs_.find(node.get()) != kernel_output_refs_.end()) {
       KernelRefCountPtr kernel_ref = kernel_output_refs_[node.get()][index];
       kernel_ref->ref_count_ = kMaxRefCount;
       kernel_ref->ref_count_dynamic_use_ = kMaxRefCount;
+      total_summary_size += kernel_ref->size_;
+      MS_LOG(INFO) << "Set summary node's ref count, node: " << node->fullname_with_scope() << " index: " << index;
     } else {
-      MS_LOG(WARNING) << "can't find summary node's kernel_def " << node->fullname_with_scope();
+      MS_LOG(WARNING) << "Can't find summary node's kernel_def " << node->fullname_with_scope() << " index: " << index;
     }
   }
 #ifdef MEM_REUSE_DEBUG
   auto graph = *graph_;
   MemReuseChecker::GetInstance().CheckMemReuseIR(total_refs_list_, kernel_def_ptr_list_, &graph);
 #endif
+  MS_LOG(INFO) << "Special Tensor total size: SummaryNodes: " << total_summary_size;
 }
 
 void MemReuseUtil::SetGraphOutputRefCount() {

mindspore/ccsrc/pre_activate/mem_reuse/mem_reuse_allocator.cc

@@ -17,6 +17,9 @@
 #include "pre_activate/mem_reuse/mem_reuse_allocator.h"
 #include "pre_activate/mem_reuse/mem_reuse.h"
 #include "pre_activate/mem_reuse/mem_reuse_checker.h"
+#ifdef ENABLE_D
+#include "device/ascend/ascend_stream_assign.h"
+#endif
 
 namespace mindspore {
 namespace memreuse {
@@ -34,6 +37,9 @@ void BestFitMemReuse::InitMemReuseInfo(const MemReuseUtil *mem_reuse_util_ptr) {
     wk->size_ = AlignMemorySize(wk->size_);
     wk->ref_count_ = 1;
   }
+#ifdef ENABLE_D
+  stream_groups_ = device::ascend::AscendStreamAssign::GetInstance().get_stream_group();
+#endif
 }
 
 void BestFitMemReuse::InitKernelDependence() {
@@ -63,21 +69,58 @@ void BestFitMemReuse::InitKernelDependence() {
   }
 }
 
-bool BestFitMemReuse::IsUsable(const KernelDefPtr &kernel_curr, const KernelDefPtr &kernel_prev) {
+bool BestFitMemReuse::IsUsable(const KernelDefPtr &kernel_curr, const MembufPtr &mem_buf) {
   // determine whether the kernel_curr can reuse kernel_prev's output tensor membuf
   MS_EXCEPTION_IF_NULL(kernel_curr);
+  MS_EXCEPTION_IF_NULL(mem_buf);
+  auto kernel_prev = mem_buf->used_kernel_;
   MS_EXCEPTION_IF_NULL(kernel_prev);
   auto curr_stream_id = kernel_curr->stream_id();
   auto prev_stream_id = kernel_prev->stream_id();
   if (curr_stream_id == prev_stream_id) {
+    mem_buf->type_ = IN_STREAM_REUSE;
+    return true;
+  }
+
+  bool reuse_between_streams = true;
+  for (auto &stream_group : stream_groups_) {
+    size_t cur_index = UINT32_MAX;
+    size_t prev_index = UINT32_MAX;
+    for (size_t index = 0; index < stream_group.size(); index++) {
+      if (curr_stream_id == stream_group[index]) {
+        cur_index = index;
+        continue;
+      }
+      if (prev_stream_id == stream_group[index]) {
+        prev_index = index;
+        continue;
+      }
+    }
+    if ((prev_index != UINT32_MAX) && (cur_index == UINT32_MAX || (prev_index > cur_index))) {
+      // previous stream and current stream are not in the same group can't be reused
+      // previous stream is behind current stream can't be reused
+      reuse_between_streams = false;
+      break;
+    }
+  }
+
+  if (reuse_between_streams) {
+    mem_buf->type_ = BETWEEN_STREAMS_REUSE;
     return true;
   }
+
   auto iter = kernel_front_map_.find(kernel_curr);
   if (iter == kernel_front_map_.end()) {
     MS_LOG(EXCEPTION) << kernel_curr->scope_full_name() << " is not init.";
   }
   auto kernel_curr_front = iter->second;
-  return kernel_curr_front.count(kernel_prev);
+  auto depend_count = kernel_curr_front.count(kernel_prev);
+  if (depend_count) {
+    mem_buf->type_ = KERNEL_DEPENDENCE_REUSE;
+    return true;
+  }
+
+  return false;
 }
 
 void BestFitMemReuse::AssignNodeOutputOffset() {
@@ -135,7 +178,7 @@ std::map<size_t, size_t> BestFitMemReuse::GetReusableMembufMap(size_t tensor_siz
     auto membuf = membuf_ptr_list_[i];
     auto index = i;
     bool is_membuf_ok = membuf->status_ == kUnused && membuf->size_ >= tensor_size;
-    if (is_membuf_ok && IsUsable(current_kernel_, membuf->used_kernel_)) {
+    if (is_membuf_ok && IsUsable(current_kernel_, membuf)) {
       (void)size_map.insert(std::make_pair(membuf->size_, index));
       break;
     }
@@ -163,8 +206,8 @@ void BestFitMemReuse::SplitMembuf(const KernelRefCount *tensor_desc, size_t memb
   auto bias = membuf->size_ - tensor_desc->size_;
   membuf->size_ = tensor_desc->size_;
   // to check if spilt membuf can be merge
-  auto new_membuf =
-    std::make_shared<Membuf>(kUnused, bias, membuf->offset_ + membuf->size_, kInvalidIndex, current_kernel_);
+  auto new_membuf = std::make_shared<Membuf>(kUnused, bias, membuf->offset_ + membuf->size_, kInvalidIndex,
+                                             membuf->type_, current_kernel_);
   (void)membuf_ptr_list_.insert(membuf_ptr_list_.begin() + SizeToInt(membuf_index + 1), new_membuf);
 }
 
@@ -176,7 +219,7 @@ void BestFitMemReuse::AddNewMembufPtr(KernelRefCount *tensor_desc, int flag) {
   }
   auto membuf_size = tensor_desc->size_;
   auto real_index = GetRealIndex(IntToSize(tensor_desc->index_), flag);
-  auto membuf = std::make_shared<Membuf>(kReused, membuf_size, membuf_offset, real_index, current_kernel_);
+  auto membuf = std::make_shared<Membuf>(kReused, membuf_size, membuf_offset, real_index, NEW, current_kernel_);
   membuf_ptr_list_.push_back(membuf);
   tensor_desc->offset_ = membuf_offset;
 }
@@ -242,7 +285,7 @@ void BestFitMemReuse::ReleaseMembuf(size_t tensor_index, int flag) {
     auto membuf_next = (*next_iter);
     MS_EXCEPTION_IF_NULL(membuf_next);
     if (membuf_next->status_ == kUnused) {
-      bool is_merge = IsUsable(current_kernel_, membuf_next->used_kernel_);
+      bool is_merge = IsUsable(current_kernel_, membuf_next);
       if (is_merge) {
         membuf->size_ += membuf_next->size_;
         (void)membuf_ptr_list_.erase(next_iter);
@@ -254,7 +297,7 @@ void BestFitMemReuse::ReleaseMembuf(size_t tensor_index, int flag) {
     auto membuf_prev = (*prev_iter);
     MS_EXCEPTION_IF_NULL(membuf_prev);
     if (membuf_prev->status_ == kUnused) {
-      bool is_merge = IsUsable(current_kernel_, membuf_prev->used_kernel_);
+      bool is_merge = IsUsable(current_kernel_, membuf_prev);
       if (is_merge) {
         membuf->size_ += membuf_prev->size_;
         membuf->offset_ = membuf_prev->offset_;

mindspore/ccsrc/pre_activate/mem_reuse/mem_reuse_allocator.h

@@ -40,11 +40,12 @@ static constexpr int kDynamicMem = -1;
 static constexpr int kWorkspaceMem = 1;
 static constexpr size_t kTotalSize = 0;
 enum Status { kUnused, kReused };
+enum MEMTYPE { NEW, IN_STREAM_REUSE, BETWEEN_STREAMS_REUSE, KERNEL_DEPENDENCE_REUSE };
 class Membuf {
  public:
   Membuf() = default;
-  Membuf(Status status, size_t size, size_t offset, int index, const KernelDefPtr &used_kernel)
-      : status_(status), size_(size), offset_(offset), index_(index), used_kernel_(used_kernel) {}
+  Membuf(Status status, size_t size, size_t offset, int index, MEMTYPE type, const KernelDefPtr &used_kernel)
+      : status_(status), size_(size), offset_(offset), index_(index), type_(type), used_kernel_(used_kernel) {}
   ~Membuf() = default;
   // Memory block status flags
   Status status_ = kUnused;
@@ -52,6 +53,7 @@ class Membuf {
   size_t offset_{0};
   // Store the tensor index stored in this memory block at a certain moment
   int index_{0};
+  MEMTYPE type_{NEW};
   KernelDefPtr used_kernel_;
 };
 using MembufPtr = std::shared_ptr<Membuf>;
@@ -122,10 +124,10 @@ class BestFitMemReuse {
   /**
    * determine if the kernel_curr can reuse the output tensor add of kernel_prev
    * @param kernel_curr, current kernel
-   * @param kernel_prev, the membuf used by this kernel
+   * @param mem_buf, the membuf
    * @return bool
    */
-  bool IsUsable(const KernelDefPtr &kernel_curr, const KernelDefPtr &kernel_prev);
+  bool IsUsable(const KernelDefPtr &kernel_curr, const MembufPtr &mem_buf);
   /**
    * init the dependence of all kernels in the graph
    */
@@ -150,6 +152,7 @@ class BestFitMemReuse {
   std::vector<MembufPtr> membuf_ptr_list_;
   // kernel_front_map_, key: the kernel_def, value: kernels before this kernel_def
   std::map<KernelDefPtr, std::set<KernelDefPtr>> kernel_front_map_;
+  std::vector<std::vector<uint32_t>> stream_groups_;
 };
 }  // namespace memreuse
 }  // namespace mindspore

mindspore/ccsrc/pre_activate/mem_reuse/mem_reuse_checker.cc

@@ -413,7 +413,8 @@ void MemReuseChecker::CheckNormalIR(const session::KernelGraph *graph) {
 void MemReuseChecker::SetMembuInfos(const KernelDef *op_def, const std::vector<MembufPtr> &membuf_ptr_list) {
   std::vector<MembufPtr> curr_mem_infos;
   for (const auto &mem : membuf_ptr_list) {
-    auto mem_checker = std::make_shared<Membuf>(mem->status_, mem->size_, mem->offset_, mem->index_, mem->used_kernel_);
+    auto mem_checker =
+      std::make_shared<Membuf>(mem->status_, mem->size_, mem->offset_, mem->index_, mem->type_, mem->used_kernel_);
     curr_mem_infos.push_back(mem_checker);
   }
   membuf_all_infos_.push_back(curr_mem_infos);
@@ -427,7 +428,8 @@ void MemReuseChecker::SetAddNewMembuInfos(const KernelDef *op_def, const std::ve
   std::vector<MembufPtr> add_new_curr_mem;
 
   for (const auto &mem : membuf_ptr_list) {
-    auto mem_checker = std::make_shared<Membuf>(mem->status_, mem->size_, mem->offset_, mem->index_, mem->used_kernel_);
+    auto mem_checker =
+      std::make_shared<Membuf>(mem->status_, mem->size_, mem->offset_, mem->index_, mem->type_, mem->used_kernel_);
     add_new_curr_mem.push_back(mem_checker);
   }
   add_new_mem_infos_.push_back(add_new_curr_mem);
@@ -451,6 +453,7 @@ void MemReuseChecker::ExportEachMembufInfo(std::ofstream &ofs) {
         << "mem_size\t"
         << "mem_head\t"
         << "mem_tail\t"
+        << "mem_type\t"
         << "used_kernel\n";
     size_t curr_used = 0;
     size_t curr_allocated = 0;
@@ -461,8 +464,8 @@ void MemReuseChecker::ExportEachMembufInfo(std::ofstream &ofs) {
           << "streamID[@" << membuf->used_kernel_->stream_id() << "]"
           << "\t"
           << "#" << static_cast<int>(membuf->status_) << "\t%" << membuf->index_ << "T"
-          << "\t" << membuf->size_ << "\t" << membuf->offset_ << "\t" << membuf->offset_ + membuf->size_ << "\t"
-          << GetSplitName(used_kernel) << "\n";
+          << "\t" << membuf->size_ << "\t" << membuf->offset_ << "\t\t" << membuf->offset_ + membuf->size_ << "\t"
+          << "\t" << static_cast<int>(membuf->type_) << "\t" << GetSplitName(used_kernel) << "\n";
       if (membuf->status_ == kReused) {
         curr_used += membuf->size_;
       }