!9958 fix opencl subgraph memory reuse

From: @ddwsky
Reviewed-by: 
Signed-off-by:

build.sh

@@ -542,15 +542,10 @@ gene_ocl_program() {
     echo "Compile SPIRV done"
 }
 
-build_opencl() {
+get_opencl() {
     cd ${BASEPATH}
     git submodule update --init third_party/OpenCL-Headers
     git submodule update --init third_party/OpenCL-CLHPP
-    if [[ "${OPENCL_OFFLINE_COMPILE}" == "on" ]]; then
-        gene_ocl_program
-    else
-        gene_clhpp
-    fi
 }
 
 
@@ -572,7 +567,8 @@ build_lite()
     fi
 
     if [ "${ENABLE_GPU}" == "on" ] && [ "${LITE_PLATFORM}" == "arm64" ] || [ $1 == "arm64" ]; then
-      echo "start build opencl"
+      echo "start get opencl"
+      get_opencl
     fi
     if [ "${ENABLE_NPU}" == "on" ]; then
       checkddk

mindspore/lite/nnacl/CMakeLists.txt

@@ -47,6 +47,8 @@ endif()
 string(REPLACE "-fvisibility=hidden" "-fvisibility=default" CMAKE_C_FLAGS "${CMAKE_C_FLAGS}")
 add_library(nnacl STATIC ${KERNEL_SRC} ${TRAIN_SRC} ${ASSEMBLY_SRC})
 add_library(nnacl_mid OBJECT ${KERNEL_SRC} ${TRAIN_SRC} ${ASSEMBLY_SRC})
+add_dependencies(nnacl fbs_src)
+add_dependencies(nnacl_mid fbs_src)
 
 ########################### arm64 build optimize library ########################
 if (PLATFORM_ARM64)

mindspore/lite/src/ops/CMakeLists.txt

@@ -6,3 +6,4 @@ file(GLOB OPS_SRC
     )
 
 add_library(cpu_ops_mid OBJECT ${OPS_SRC})
+add_dependencies(cpu_ops_mid fbs_src)

mindspore/lite/src/runtime/kernel/arm/CMakeLists.txt

@@ -14,6 +14,7 @@ set(KERNEL_SRC ${KERNEL_SRC} ${TRAIN_KERNEL_SRC})
 endif()
 
 add_library(cpu_kernel_mid OBJECT ${KERNEL_SRC})
+add_dependencies(cpu_kernel_mid fbs_src)
 
 if (PLATFORM_ARM64)
   file(GLOB FP16_KERNEL_SRC ${CMAKE_CURRENT_SOURCE_DIR}/fp16/*.cc)

mindspore/lite/src/runtime/kernel/opencl/CMakeLists.txt

@@ -1,2 +1,4 @@
 file(GLOB_RECURSE OPENCL_KERNEL_SRC ${CMAKE_CURRENT_SOURCE_DIR}/kernel/*.cc)
 add_library(opencl_kernel_mid OBJECT ${OPENCL_KERNEL_SRC})
+add_dependencies(opencl_kernel_mid fbs_src)
+

mindspore/lite/src/runtime/kernel/opencl/opencl_kernel.cc

@@ -210,6 +210,7 @@ int OpenCLKernel::DequantWeight() {
     void *dequant_weight{nullptr};
     bool set_flag{true};
     if (is_fp16) {
+#ifdef ENABLE_ARM64
       if (in_tensors_.at(kWeightIndex)->data_type() == kNumberTypeInt8) {
         dequant_weight = kernel::DequantUtil::DequantData<int8_t, float16_t>(weight_tensor);
       } else if (in_tensors_.at(kWeightIndex)->data_type() == kNumberTypeInt16) {
@@ -217,6 +218,9 @@ int OpenCLKernel::DequantWeight() {
       } else {
         set_flag = false;
       }
+#else
+      set_flag = false;
+#endif
     } else {
       if (in_tensors_.at(kWeightIndex)->data_type() == kNumberTypeInt8) {
         dequant_weight = kernel::DequantUtil::DequantData<int8_t, float>(weight_tensor);

mindspore/lite/src/runtime/opencl/opencl_allocator.cc

@@ -147,6 +147,11 @@ void *OpenCLAllocator::Malloc(size_t size, const std::vector<size_t> &img_size,
     UnLock();
     return host_ptr;
   }
+  total_size_ += size;
+  const uint64_t max_size = ocl_runtime_->GetGlobalMemSize();
+  if (total_size_ >= max_size) {
+    MS_LOG(ERROR) << "Mem pool out of max_size, total size: " << total_size_ << ", max size: " << max_size;
+  }
   cl::Buffer *buffer = nullptr;
   cl::Image2D *image = nullptr;
   cl_mem_flags flags = CL_MEM_READ_WRITE;
@@ -188,7 +193,8 @@ void *OpenCLAllocator::Malloc(size_t size, const std::vector<size_t> &img_size,
   UnLock();
   std::string type_name = img_size.empty() ? "buffer" : "Image2D";
   MS_LOG(DEBUG) << "Malloc a new " << type_name << ". size: " << mem_buf->size_ << ", host addr: " << mem_buf->host_ptr_
-                << ", device addr: " << mem_buf->device_ptr_ << ", image_addr: " << image;
+                << ", device addr: " << mem_buf->device_ptr_ << ", image_addr: " << image
+                << ", total size: " << total_size_;
   return host_ptr;
 }
 
@@ -250,10 +256,10 @@ void *OpenCLAllocator::GetBuffer(void *buffer) {
   return nullptr;
 }
 
-void OpenCLAllocator::Clear() {
-  Lock();
+template <typename T>
+void OpenCLAllocator::ClearMemList(T *list) {
   auto svm_capabilities = ocl_runtime_->GetSVMCapabilities();
-  for (auto it = allocated_list_.begin(); it != allocated_list_.end(); it++) {
+  for (auto it = list->begin(); it != list->end(); it++) {
     if (it->second->map_flags) {
       int ret = UnmapBuffer(it->second->host_ptr_);
       if (ret != RET_OK) {
@@ -278,29 +284,13 @@ void OpenCLAllocator::Clear() {
     }
     delete it->second;
   }
-  allocated_list_.clear();
+  list->clear();
+}
 
-  for (auto it = free_list_.begin(); it != free_list_.end(); it++) {
-    if (svm_capabilities) {
-      clSVMFree((*ocl_runtime_->Context())(), it->second->host_ptr_);
-      MS_LOG(DEBUG) << "OpenCL free svm buffer : " << it->second->host_ptr_;
-    } else {
-      cl::Buffer *buffer = static_cast<cl::Buffer *>(it->second->device_ptr_);
-      if (buffer != nullptr) {
-        MS_LOG(DEBUG) << "OpenCL free device buffer : " << buffer;
-        delete buffer;
-        it->second->device_ptr_ = nullptr;
-      }
-      cl::Image *image = static_cast<cl::Image *>(it->second->image_ptr_);
-      if (image != nullptr) {
-        MS_LOG(DEBUG) << "OpenCL free image : " << image;
-        delete image;
-        it->second->image_ptr_ = nullptr;
-      }
-    }
-    delete it->second;
-  }
-  free_list_.clear();
+void OpenCLAllocator::Clear() {
+  Lock();
+  ClearMemList<std::unordered_map<void *, MemBuf *>>(&allocated_list_);
+  ClearMemList<std::multimap<size_t, MemBuf *>>(&free_list_);
   UnLock();
 }
 

mindspore/lite/src/runtime/opencl/opencl_allocator.h

@@ -63,6 +63,12 @@ class OpenCLAllocator : public Allocator {
   void *CreateBuffer(size_t size, void *data, size_t flags, cl::Buffer **buffer);
   void *CreateImage2D(size_t size, const std::vector<size_t> &img_size, void *data, size_t flags, bool is_map,
                       cl::Buffer **buffer, cl::Image2D **image);
+  template <typename T>
+  void ClearMemList(T *list);
+
+ private:
+  OpenCLRuntime *ocl_runtime_{nullptr};
+  std::mutex lock;
   struct MemBuf {
     size_t size_;
     void *device_ptr_;
@@ -72,14 +78,13 @@ class OpenCLAllocator : public Allocator {
     bool map_flags{false};
   };
 
-  std::mutex lock;
   // <membuf->buf, membuf>
   std::unordered_map<void *, MemBuf *> allocated_list_;
   std::multimap<size_t, MemBuf *> free_list_;
+  uint64_t total_size_{0};
   // 6 is empirical value
   int shift_factor_ = 6;
   bool lock_flag_ = false;
-  OpenCLRuntime *ocl_runtime_{nullptr};
 };
 
 }  // namespace mindspore::lite::opencl

mindspore/lite/src/runtime/opencl/opencl_executor.cc

@@ -59,6 +59,10 @@ int OpenCLExecutor::RunOrTune(std::vector<Tensor *> &inputs, std::vector<Tensor
           return ret;
         }
         auto data_ptr = allocator_->Malloc(output->Size(), img_size);
+        if (data_ptr == nullptr) {
+          MS_LOG(ERROR) << "Malloc data failed";
+          return RET_ERROR;
+        }
         output->set_data(data_ptr);
       } else {
         ret = output->MallocData(allocator_);

mindspore/lite/src/runtime/opencl/opencl_runtime.cc

@@ -225,8 +225,12 @@ int OpenCLRuntime::Init() {
                    << "SVM_ATOMICS";
     }
   }
+  global_memery_size_ = device_->getInfo<CL_DEVICE_GLOBAL_MEM_SIZE>();
 
+  MS_LOG(INFO) << "Address space bits: " << device_->getInfo<CL_DEVICE_ADDRESS_BITS>();
+  MS_LOG(INFO) << "Global Mem Size: " << global_memery_size_;
   MS_LOG(INFO) << "Global Mem Cache Size: " << global_memery_cachesize_;
+  MS_LOG(INFO) << "Max Alloc Size: " << device_->getInfo<CL_DEVICE_MAX_MEM_ALLOC_SIZE>();
   MS_LOG(INFO) << "Compute Unit: " << compute_units_;
   MS_LOG(INFO) << "Clock Frequency: " << max_freq_ << " MHz";
 

mindspore/lite/src/runtime/opencl/opencl_runtime.h

@@ -61,6 +61,7 @@ class OpenCLRuntime {
   uint32_t DeviceMaxFreq() const;
   uint64_t GetMaxWorkGroupSize(const cl::Kernel &kernel);
   uint32_t GetSubGroupSize(const cl::Kernel &kernel, const cl::NDRange &range = cl::NullRange);
+  uint64_t GetGlobalMemSize() { return global_memery_size_; }
   GpuInfo GetGpuInfo();
   bool GetFp16Enable() const;
   bool SetFp16Enable(bool enable);
@@ -168,6 +169,7 @@ class OpenCLRuntime {
   std::map<std::string, cl::Program> program_map_;
   cl::Program binary_program_{0};
   uint64_t global_memery_cachesize_{0};
+  uint64_t global_memery_size_{0};
   int max_work_group_size_{1};
   uint32_t compute_units_{0};
   uint32_t max_freq_{0};