Merge branch 'fix-ws-bug' into 'master'

Add device_type to MemoryBlock and fix workspace preallocation bug.

See merge request !694

mace/core/net.cc

@@ -41,7 +41,7 @@ SerialNet::SerialNet(
   MACE_LATENCY_LOGGER(1, "Constructing SerialNet ", net_def->name());
   for (int idx = 0; idx < net_def->op_size(); ++idx) {
     const auto &operator_def = net_def->op(idx);
-    // TODO(liuqi): refactor based on PB
+    // TODO(liuqi): refactor to add device_type to OperatorDef
     const int op_device =
         ProtoArgHelper::GetOptionalArg<OperatorDef, int>(
             operator_def, "device", static_cast<int>(device_type_));

mace/core/workspace.cc

@@ -185,7 +185,7 @@ MaceStatus Workspace::CreateOutputTensorBuffer(const NetDef &net_def,
   // As DSP may have different data output type for each op,
   // we stick to the same concept.
   for (auto &op : net_def.op()) {
-    // TODO(liuqi): refactor based on PB
+    // TODO(liuqi): refactor to add device_type to OperatorDef
     const int op_device =
         ProtoArgHelper::GetOptionalArg<OperatorDef, int>(
             op, "device", static_cast<int>(device_type));
@@ -204,33 +204,38 @@ MaceStatus Workspace::CreateOutputTensorBuffer(const NetDef &net_def,
   // TODO(liyin): memory block should not have concept of type, but to be
   // consistent with gpu, all memory block use float/half as unit
   for (auto &mem_block : net_def.mem_arena().mem_block()) {
-    if (mem_block.mem_type() == MemoryType::CPU_BUFFER) {
-      std::unique_ptr<BufferBase> tensor_buf(
-          new Buffer(GetDeviceAllocator(DeviceType::CPU)));
-      MACE_RETURN_IF_ERROR(tensor_buf->Allocate(
-          mem_block.x() * GetEnumTypeSize(dtype)
-              + MACE_EXTRA_BUFFER_PAD_SIZE));
-      preallocated_allocator_.SetBuffer(mem_block.mem_id(),
-                                        std::move(tensor_buf));
-    } else if (mem_block.mem_type() == MemoryType::GPU_IMAGE) {
-      std::unique_ptr<BufferBase> image_buf(
-          new Image());
-      MACE_RETURN_IF_ERROR(image_buf->Allocate(
-          {mem_block.x(), mem_block.y()}, dtype));
-      preallocated_allocator_.SetBuffer(mem_block.mem_id(),
-                                        std::move(image_buf));
-    } else if (mem_block.mem_type() == MemoryType::GPU_BUFFER) {
-      std::unique_ptr<BufferBase> tensor_buf(
-          new Buffer(GetDeviceAllocator(DeviceType::GPU)));
-      MACE_RETURN_IF_ERROR(tensor_buf->Allocate(
-          mem_block.x() * GetEnumTypeSize(dtype)));
-      preallocated_allocator_.SetBuffer(mem_block.mem_id(),
-                                        std::move(tensor_buf));
+    if (mem_block.device_type() == device_type) {
+      VLOG(3) << "Preallocate memory block. id: " << mem_block.mem_id()
+              << ", device type: " << mem_block.device_type()
+              << ", memory type: " << mem_block.mem_type();
+      if (mem_block.mem_type() == MemoryType::CPU_BUFFER) {
+        std::unique_ptr<BufferBase> tensor_buf(
+            new Buffer(GetDeviceAllocator(DeviceType::CPU)));
+        MACE_RETURN_IF_ERROR(tensor_buf->Allocate(
+            mem_block.x() * GetEnumTypeSize(dtype)
+                + MACE_EXTRA_BUFFER_PAD_SIZE));
+        preallocated_allocator_.SetBuffer(mem_block.mem_id(),
+                                          std::move(tensor_buf));
+      } else if (mem_block.mem_type() == MemoryType::GPU_IMAGE) {
+        std::unique_ptr<BufferBase> image_buf(
+            new Image());
+        MACE_RETURN_IF_ERROR(image_buf->Allocate(
+            {mem_block.x(), mem_block.y()}, dtype));
+        preallocated_allocator_.SetBuffer(mem_block.mem_id(),
+                                          std::move(image_buf));
+      } else if (mem_block.mem_type() == MemoryType::GPU_BUFFER) {
+        std::unique_ptr<BufferBase> tensor_buf(
+            new Buffer(GetDeviceAllocator(DeviceType::GPU)));
+        MACE_RETURN_IF_ERROR(tensor_buf->Allocate(
+            mem_block.x() * GetEnumTypeSize(dtype)));
+        preallocated_allocator_.SetBuffer(mem_block.mem_id(),
+                                          std::move(tensor_buf));
+      }
     }
   }
   VLOG(3) << "Preallocate buffer to tensors";
   for (auto &op : net_def.op()) {
-    // TODO(liuqi): refactor based on PB
+    // TODO(liuqi): refactor to add device_type to OperatorDef
     const int op_device =
         ProtoArgHelper::GetOptionalArg<OperatorDef, int>(
             op, "device", static_cast<int>(device_type));

mace/proto/mace.proto

@@ -79,9 +79,10 @@ message OperatorDef {
 // for memory optimization
 message MemoryBlock {
   optional int32 mem_id = 1;
-  optional MemoryType mem_type = 2;
-  optional uint32 x = 3;
-  optional uint32 y = 4;
+  optional int32 device_type = 2;
+  optional MemoryType mem_type = 3;
+  optional uint32 x = 4;
+  optional uint32 y = 5;
 }
 message MemoryArena {
   repeated MemoryBlock mem_block = 1;

mace/python/tools/memory_optimizer.py

@@ -17,10 +17,22 @@ import operator
 from mace.proto import mace_pb2
 
 from mace.python.tools.converter_tool import base_converter as cvt
+from mace.python.tools.converter_tool.base_converter import DeviceType
 from mace.python.tools.convert_util import calculate_image_shape
 from mace.python.tools.convert_util import OpenCLBufferType
 
 
+def MemoryTypeToStr(mem_type):
+    if mem_type == mace_pb2.CPU_BUFFER:
+        return 'CPU_BUFFER'
+    elif mem_type == mace_pb2.GPU_BUFFER:
+        return 'GPU_BUFFER'
+    elif mem_type == mace_pb2.GPU_IMAGE:
+        return 'GPU_IMAGE'
+    else:
+        return 'UNKNOWN'
+
+
 class MemoryBlock(object):
     def __init__(self, mem_type, block):
         self._mem_type = mem_type
@@ -88,6 +100,7 @@ class MemoryOptimizer(object):
             arena = self.net_def.mem_arena
             block = arena.mem_block.add()
             block.mem_id = mem
+            block.device_type = DeviceType.CPU.value
             block.mem_type = self.mem_block[mem].mem_type
             block.x = self.mem_block[mem].block[0]
             block.y = 1
@@ -103,7 +116,8 @@ class MemoryOptimizer(object):
     def get_total_optimized_mem_size(self):
         optimized_mem_size = 0
         for mem in self.mem_block:
-            print mem, self.mem_block[mem].mem_type, self.mem_block[mem].block
+            print mem, MemoryTypeToStr(self.mem_block[mem].mem_type), \
+                self.mem_block[mem].block
             optimized_mem_size += self.mem_size(self.mem_block[mem])
         return optimized_mem_size
 
@@ -165,7 +179,7 @@ class MemoryOptimizer(object):
                             self.idle_mem.remove(mem_id)
 
                     if mem_id == -1:
-                        mem_id = self.mem_id_base() + self.total_mem_count
+                        mem_id = self.total_mem_count
                         self.total_mem_count += 1
                         self.mem_block[mem_id] = op_mem_block
 
@@ -198,9 +212,6 @@ class MemoryOptimizer(object):
               self.get_total_origin_mem_size(),
               self.get_total_optimized_mem_size()))
 
-    def mem_id_base(self):
-        return 0
-
 
 class GPUMemoryOptimizer(MemoryOptimizer):
     def op_need_optimize_memory(self, op):
@@ -256,6 +267,7 @@ class GPUMemoryOptimizer(MemoryOptimizer):
             arena = self.net_def.mem_arena
             block = arena.mem_block.add()
             block.mem_id = mem
+            block.device_type = DeviceType.GPU.value
             block.mem_type = self.mem_block[mem].mem_type
             block.x = self.mem_block[mem].block[0]
             block.y = self.mem_block[mem].block[1]
@@ -279,9 +291,6 @@ class GPUMemoryOptimizer(MemoryOptimizer):
         net_ocl_max_img_size_arg.ints[:] = [max_image_size_x,
                                             max_image_size_y]
 
-    def mem_id_base(self):
-        return 20000
-
 
 def optimize_gpu_memory(net_def):
     mem_optimizer = GPUMemoryOptimizer(net_def)

mace/python/tools/model.jinja2

@@ -129,6 +129,7 @@ void CreateMemoryArena(mace::MemoryArena *mem_arena) {
 
   mace::MemoryBlock* mem_block{{i}} = mem_arena->add_mem_block();
   mem_block{{i}}->set_mem_id({{net.mem_arena.mem_block[i].mem_id}});
+  mem_block{{i}}->set_device_type(static_cast<DeviceType>({{net.mem_arena.mem_block[i].device_type}}));
   mem_block{{i}}->set_mem_type(static_cast<MemoryType>({{net.mem_arena.mem_block[i].mem_type}}));
   mem_block{{i}}->set_x({{net.mem_arena.mem_block[i].x}});
   mem_block{{i}}->set_y({{net.mem_arena.mem_block[i].y}});

mace/test/mace_api_mt_test.cc

@@ -224,8 +224,7 @@ std::map<std::string, int> AddMemoryOptimization(
     const std::vector<std::vector<int64_t>> &output_shapes,
     NetDef *net_def) {
   std::map<std::string, int> res;
-  // TODO(liuqi) refactor based on PB
-  int mem_id = 20000;
+  int mem_id = 0;
   size_t input_shape_size = input_shapes.size();
   uint32_t in_mem_block_x = 0;
   uint32_t in_mem_block_y = 0;
@@ -244,6 +243,7 @@ std::map<std::string, int> AddMemoryOptimization(
   for (size_t i = 0; i < input_size; ++i) {
     MemoryBlock *mem_blk_ptr = mem_arena_ptr->add_mem_block();
     mem_blk_ptr->set_mem_id(mem_id);
+    mem_blk_ptr->set_device_type(DeviceType::GPU);
     mem_blk_ptr->set_mem_type(MemoryType::GPU_IMAGE);
     mem_blk_ptr->set_x(in_mem_block_x);
     mem_blk_ptr->set_y(in_mem_block_y);
@@ -264,6 +264,7 @@ std::map<std::string, int> AddMemoryOptimization(
   for (size_t i = 0; i < output_size; ++i) {
     MemoryBlock *mem_blk_ptr = mem_arena_ptr->add_mem_block();
     mem_blk_ptr->set_mem_id(mem_id);
+    mem_blk_ptr->set_device_type(DeviceType::GPU);
     mem_blk_ptr->set_mem_type(MemoryType::GPU_IMAGE);
     mem_blk_ptr->set_x(out_mem_block_x);
     mem_blk_ptr->set_y(out_mem_block_y);

mace/test/mace_api_test.cc

@@ -225,8 +225,7 @@ std::map<std::string, int> AddMemoryOptimization(
     const std::vector<std::vector<int64_t>> &output_shapes,
     NetDef *net_def) {
   std::map<std::string, int> res;
-  // TODO(liuqi) refactor based on PB
-  int mem_id = 20000;
+  int mem_id = 0;
   size_t input_shape_size = input_shapes.size();
   uint32_t in_mem_block_x = 0;
   uint32_t in_mem_block_y = 0;
@@ -245,6 +244,7 @@ std::map<std::string, int> AddMemoryOptimization(
   for (size_t i = 0; i < input_size; ++i) {
     MemoryBlock *mem_blk_ptr = mem_arena_ptr->add_mem_block();
     mem_blk_ptr->set_mem_id(mem_id);
+    mem_blk_ptr->set_device_type(DeviceType::GPU);
     mem_blk_ptr->set_mem_type(MemoryType::GPU_IMAGE);
     mem_blk_ptr->set_x(in_mem_block_x);
     mem_blk_ptr->set_y(in_mem_block_y);
@@ -265,6 +265,7 @@ std::map<std::string, int> AddMemoryOptimization(
   for (size_t i = 0; i < output_size; ++i) {
     MemoryBlock *mem_blk_ptr = mem_arena_ptr->add_mem_block();
     mem_blk_ptr->set_mem_id(mem_id);
+    mem_blk_ptr->set_device_type(DeviceType::GPU);
     mem_blk_ptr->set_mem_type(MemoryType::GPU_IMAGE);
     mem_blk_ptr->set_x(out_mem_block_x);
     mem_blk_ptr->set_y(out_mem_block_y);