[mkldnn-v1.0] Change MXNET_USE_MKLDNN from 100 back to 1

CMakeLists.txt

@@ -269,7 +269,7 @@ if(USE_MKLDNN)
 
   include_directories(3rdparty/mkldnn/include)
   include_directories(${PROJECT_BINARY_DIR}/3rdparty/mkldnn/include)
-  add_definitions(-DMXNET_USE_MKLDNN=100)
+  add_definitions(-DMXNET_USE_MKLDNN=1)
   list(APPEND mxnet_LINKER_LIBS mkldnn)
 endif()
 

Makefile

@@ -146,7 +146,7 @@ ifndef LINT_LANG
 endif
 
 ifeq ($(USE_MKLDNN), 1)
-	CFLAGS += -DMXNET_USE_MKLDNN=100
+	CFLAGS += -DMXNET_USE_MKLDNN=1
 	CFLAGS += -I$(ROOTDIR)/src/operator/nn/mkldnn/
 	CFLAGS += -I$(MKLDNNROOT)/include
 	LDFLAGS += -L$(MKLDNNROOT)/lib -L$(MKLDNNROOT)/lib64 -lmkldnn -Wl,-rpath,'$${ORIGIN}'

include/mxnet/base.h

@@ -551,7 +551,7 @@ inline std::ostream& operator<<(std::ostream &out, const Context &ctx) {
 #define ADD_FILELINE "\n\nDefined in " __FILE__ ":L" STRINGIZE(__LINE__)
 
 
-#if MXNET_USE_MKLDNN == 100
+#if MXNET_USE_MKLDNN == 1
 constexpr size_t kMKLDNNAlign = 64;
 #endif
 

include/mxnet/ndarray.h

@@ -37,7 +37,7 @@
 #include <algorithm>
 #include <memory>
 #include <algorithm>
-#if MXNET_USE_MKLDNN == 100
+#if MXNET_USE_MKLDNN == 1
 #include <mkldnn.hpp>
 #endif
 #include "./base.h"
@@ -699,7 +699,7 @@ class NDArray {
     ptr_->CheckAndAllocAuxData(i, aux_shape);
   }
 
-#if MXNET_USE_MKLDNN == 100
+#if MXNET_USE_MKLDNN == 1
   /*
    * Create NDArray from mkldnn memory.
    * mkldnn_mem The mkldnn memory to be managed.
@@ -825,7 +825,7 @@ class NDArray {
     */
     std::vector<Storage::Handle> aux_handles;
 
-#if MXNET_USE_MKLDNN == 100
+#if MXNET_USE_MKLDNN == 1
     /*! This is created when data is stored in MKLDNN format.
      */
     std::shared_ptr<MKLDNNMemory> mkl_mem_;
@@ -984,7 +984,7 @@ class NDArray {
     inline void CheckAndAlloc(void) {
       if (delay_alloc) {
         shandle = Storage::Get()->Alloc(shandle.size, shandle.ctx);
-#if MXNET_USE_MKLDNN == 100
+#if MXNET_USE_MKLDNN == 1
         mkl_mem_ = nullptr;
 #endif
         delay_alloc = false;
@@ -999,7 +999,7 @@ class NDArray {
       dbytes = std::max(dbytes, static_cast<uint64_t>(shandle.size));
       if (delay_alloc) {
         shandle = Storage::Get()->Alloc(dbytes, shandle.ctx);
-#if MXNET_USE_MKLDNN == 100
+#if MXNET_USE_MKLDNN == 1
         mkl_mem_ = nullptr;
 #endif
         delay_alloc = false;
@@ -1008,7 +1008,7 @@ class NDArray {
         Storage::Get()->Free(shandle);
         // init storage
         shandle = Storage::Get()->Alloc(dbytes, shandle.ctx);
-#if MXNET_USE_MKLDNN == 100
+#if MXNET_USE_MKLDNN == 1
         mkl_mem_ = nullptr;
 #endif
       }
@@ -1044,7 +1044,7 @@ class NDArray {
     // and allocate new storage
     void CheckAndAllocData(const mxnet::TShape &shape, int dtype);
 
-#if MXNET_USE_MKLDNN == 100
+#if MXNET_USE_MKLDNN == 1
     // Have MKL memory reference to the data in the default storage
     // or create memory for MKLDNN.
     void SetMKLMem(const mxnet::TShape &shape, int dtype);

src/c_api/c_api.cc

@@ -627,7 +627,7 @@ int MXNDArrayGetData(NDArrayHandle handle,
                      void **out_pdata) {
   API_BEGIN();
   NDArray *arr = static_cast<NDArray*>(handle);
-#if MXNET_USE_MKLDNN == 100
+#if MXNET_USE_MKLDNN == 1
   if (arr->IsMKLDNNData()) {
     arr->Reorder2DefaultAsync();
     arr->WaitToRead();

src/common/exec_utils.h

@@ -59,15 +59,15 @@ inline bool SetupDefaultBlobsIn(const std::vector<NDArray>& src,
   for (size_t i = 0; i < src.size(); i++) {
     auto& nd = src[i];
     bool is_default = nd.storage_type() == kDefaultStorage;
-#if MXNET_USE_MKLDNN == 100
+#if MXNET_USE_MKLDNN == 1
     // We have to make sure it's default storage and default layout.
     is_default = nd.IsDefaultData();
 #endif
     if (!is_default) {
       (*idx_map)[i] = temp_dst->size();
       NDArray temp = bufs != nullptr ? bufs->at(i) : NDArray(nd.shape(), nd.ctx(),
                                                              true, nd.dtype());
-#if MXNET_USE_MKLDNN == 100
+#if MXNET_USE_MKLDNN == 1
       CHECK(temp.IsDefaultData());
 #endif
       temp_src->emplace_back(nd);
@@ -91,7 +91,7 @@ inline bool SetupDefaultBlobsOut(const std::vector<NDArray>& src,
   for (size_t i = 0; i < src.size(); i++) {
     auto& nd = src[i];
     bool is_default = nd.storage_type() == kDefaultStorage;
-#if MXNET_USE_MKLDNN == 100
+#if MXNET_USE_MKLDNN == 1
     if (req->at(i) == kWriteInplace && nd.IsMKLDNNData())
       // If it's write inplace and the output array doesn't use the default
       // layout, we'll generate a temporary output array below, which means
@@ -102,7 +102,7 @@ inline bool SetupDefaultBlobsOut(const std::vector<NDArray>& src,
     is_default = nd.IsDefaultData();
 #endif
     if (!is_default) {
-#if MXNET_USE_MKLDNN == 100
+#if MXNET_USE_MKLDNN == 1
       NDArray temp;
       if (bufs != nullptr) {
         temp = bufs->at(i);

src/common/utils.h

@@ -47,7 +47,7 @@
 #include <limits>
 
 #include "../operator/mxnet_op.h"
-#if MXNET_USE_MKLDNN == 100
+#if MXNET_USE_MKLDNN == 1
 #include "../operator/nn/mkldnn/mkldnn_base-inl.h"
 #endif
 
@@ -484,7 +484,7 @@ inline void LogStorageFallback(const nnvm::NodeAttrs& attrs,
     "0 to suppress this warning.";
   os << "\nStorage type fallback detected:\n" << op_str << warning;
   LogOnce(os.str());
-#if MXNET_USE_MKLDNN == 100
+#if MXNET_USE_MKLDNN == 1
   if (!MKLDNNEnvSet()) common::LogOnce("MXNET_MKLDNN_ENABLED flag is off. "
                                        "You can re-enable by setting MXNET_MKLDNN_ENABLED=1");
   if (GetMKLDNNCacheSize() != -1) common::LogOnce("MXNET_MKLDNN_CACHE_NUM is set."

src/executor/attach_op_execs_pass.cc

@@ -116,7 +116,7 @@ class StatefulComputeExecutor : public StorageFallbackOpExecutor {
  public:
   void Run(RunContext rctx, bool is_gpu) override {
     op_ctx.run_ctx = rctx;
-#if MXNET_USE_MKLDNN == 100
+#if MXNET_USE_MKLDNN == 1
     InvalidateOutputs(out_array, req);
 #endif
     PreFCompute(is_gpu);
@@ -155,7 +155,7 @@ class StatefulComputeExExecutor : public OpExecutor {
  public:
   void Run(RunContext rctx, bool is_gpu) override {
     op_ctx.run_ctx = rctx;
-#if MXNET_USE_MKLDNN == 100
+#if MXNET_USE_MKLDNN == 1
     InvalidateOutputs(out_array, req);
     // TODO(alex): (MXNET-847) Remove this fallback feature after subgraph implemented
     const auto is_mkldnn = Op::GetAttr<bool>("TIsMKLDNN");
@@ -202,7 +202,7 @@ class FComputeExecutor : public StorageFallbackOpExecutor {
   void Run(RunContext rctx, bool is_gpu) override {
     using namespace common;
     op_ctx.run_ctx = rctx;
-#if MXNET_USE_MKLDNN == 100
+#if MXNET_USE_MKLDNN == 1
     InvalidateOutputs(out_array, req);
 #endif
     PreFCompute(is_gpu);
@@ -231,7 +231,7 @@ class FComputeExExecutor : public OpExecutor {
  public:
   void Run(RunContext rctx, bool is_gpu) override {
     op_ctx.run_ctx = rctx;
-#if MXNET_USE_MKLDNN == 100
+#if MXNET_USE_MKLDNN == 1
     InvalidateOutputs(out_array, req);
     // TODO(alex): (MXNET-847) Remove this fallback feature after subgraph implemented
     const auto is_mkldnn = Op::GetAttr<bool>("TIsMKLDNN");

src/executor/graph_executor.cc

@@ -42,7 +42,7 @@ namespace exec {
 using namespace mxnet::common;
 
 static const std::string GetDefaultSubgraphBackend() {
-#if MXNET_USE_MKLDNN == 100
+#if MXNET_USE_MKLDNN == 1
   return std::string("MKLDNN");
 #else
   return std::string();

src/imperative/imperative_utils.h

@@ -419,7 +419,7 @@ inline void PushFCompute(const FCompute& fn,
       std::vector<NDArray> pre_temp_src, pre_temp_dst, post_temp_dst, post_temp_src;
       // mapping from index in input_blobs to index in pre_temp_dst
       std::unordered_map<uint32_t, uint32_t> in_temp_idx_map;
-#if MXNET_USE_MKLDNN == 100
+#if MXNET_USE_MKLDNN == 1
       if (exec_type != ExecType::kCrossDeviceCopy) {
         // kCrossDeviceCopy is used for `_copy_to` operator, which doesn't compute immediately in
         // its FCcomputeEx, but AsyncPush the copy operation to engine.
@@ -468,7 +468,7 @@ inline void PushFComputeEx(const FComputeEx& fn,
   DerefInputOutput(p_inputs, p_outputs, &inputs, &outputs);
   const auto& run = [=](RunContext rctx) {
       OpContext opctx{need_grad, is_train, rctx, engine::CallbackOnComplete(), requested};
-#if MXNET_USE_MKLDNN == 100
+#if MXNET_USE_MKLDNN == 1
       if (exec_type != ExecType::kCrossDeviceCopy) {
         // kCrossDeviceCopy is used for `_copy_to` operator, which doesn't compute immediately in
         // its FCcomputeEx, but AsyncPush the copy operation to engine.
@@ -486,7 +486,7 @@ inline void PushFComputeEx(const FComputeEx& fn,
       } else {
 #endif
         fn(attrs, opctx, inputs, req, outputs);
-#if MXNET_USE_MKLDNN == 100
+#if MXNET_USE_MKLDNN == 1
       }
 #endif
       if (ctx.dev_mask() == gpu::kDevMask && exec_type == ExecType::kSync && !rctx.is_bulk) {
@@ -532,7 +532,7 @@ inline void PushOperator(const OpStatePtr& state,
     const auto& run = [=](RunContext rctx,
                           engine::CallbackOnComplete on_complete) {
       OpContext opctx{need_grad, is_train, rctx, on_complete, requested};
-#if MXNET_USE_MKLDNN == 100
+#if MXNET_USE_MKLDNN == 1
       if (exec_type != ExecType::kCrossDeviceCopy) {
         // kCrossDeviceCopy is used for `_copy_to` operator, which doesn't compute immediately in
         // its FCcomputeEx, but AsyncPush the copy operation to engine.
@@ -550,7 +550,7 @@ inline void PushOperator(const OpStatePtr& state,
       } else {
 #endif
         fcompute_ex(state, opctx, inputs, req, outputs);
-#if MXNET_USE_MKLDNN == 100
+#if MXNET_USE_MKLDNN == 1
       }
 #endif
       if (ctx.dev_mask() == gpu::kDevMask && exec_type == ExecType::kSync
@@ -588,7 +588,7 @@ inline void PushOperator(const OpStatePtr& state,
         std::vector<NDArray> pre_temp_src, pre_temp_dst, post_temp_dst, post_temp_src;
         // mapping from index in input_blobs to index in pre_temp_dst
         std::unordered_map<uint32_t, uint32_t> in_temp_idx_map;
-#if MXNET_USE_MKLDNN == 100
+#if MXNET_USE_MKLDNN == 1
       if (exec_type != ExecType::kCrossDeviceCopy) {
         // kCrossDeviceCopy is used for `_copy_to` operator, which doesn't compute immediately in
         // its FCcomputeEx, but AsyncPush the copy operation to engine.

src/ndarray/ndarray.cc

@@ -103,7 +103,7 @@ void NDArray::SetShapeFromChunk() {
 struct ChunkMem {
   Storage::Handle h;
   std::vector<Storage::Handle> aux_h;
-#if MXNET_USE_MKLDNN == 100
+#if MXNET_USE_MKLDNN == 1
   std::shared_ptr<MKLDNNMemory> mem;
 #endif
 };
@@ -113,14 +113,14 @@ NDArray::Chunk::~Chunk() {
   ChunkMem mem;
   mem.h = this->shandle;
   mem.aux_h = this->aux_handles;
-#if MXNET_USE_MKLDNN == 100
+#if MXNET_USE_MKLDNN == 1
   // We want to delete mkldnn memory after deleting the variable.
   mem.mem = this->mkl_mem_;
 #endif
   if (auto engine = engine_ref_.lock()) {
     engine->DeleteVariable([mem, skip_free](RunContext s) {
       if (skip_free == false) {
-#if MXNET_USE_MKLDNN == 100
+#if MXNET_USE_MKLDNN == 1
         if (mem.mem) {
           CHECK_LE(mem.mem->GetSize(), mem.h.size);
           CHECK_EQ(mem.mem->GetDataHandle(), mem.h.dptr);
@@ -144,7 +144,7 @@ void NDArray::Chunk::CheckAndAllocData(const mxnet::TShape &shape, int dtype) {
     Storage::Get()->Free(shandle);
     // init storage
     shandle = Storage::Get()->Alloc(dbytes, ctx);
-#if MXNET_USE_MKLDNN == 100
+#if MXNET_USE_MKLDNN == 1
     mkl_mem_ = nullptr;
 #endif
   }
@@ -172,7 +172,7 @@ nnvm::Symbol NDArray::get_autograd_symbol() const {
   return ret;
 }
 
-#if MXNET_USE_MKLDNN == 100
+#if MXNET_USE_MKLDNN == 1
 
 NDArray::NDArray(const mkldnn::memory::desc &md)
     : storage_type_(kDefaultStorage), entry_(nullptr) {
@@ -387,7 +387,7 @@ void NDArray::set_fresh_out_grad(bool state) const {
   info.fresh_out_grad = state;
 }
 
-#if MXNET_USE_MKLDNN == 100
+#if MXNET_USE_MKLDNN == 1
 
 bool NDArray::Chunk::IsMKLDNN() const {
   if (storage_type != kDefaultStorage)
@@ -740,7 +740,7 @@ void NDArray::SetTBlob() const {
   char *dptr = static_cast<char*>(ptr_->shandle.dptr);
   auto stype = storage_type();
   if (stype == kDefaultStorage) {
-#if MXNET_USE_MKLDNN == 100
+#if MXNET_USE_MKLDNN == 1
     CHECK(!IsMKLDNNData()) << "We can't generate TBlob for MKLDNN data. "
         << "Please use Reorder2Default() to generate a new NDArray first";
 #endif
@@ -1082,7 +1082,7 @@ inline void CopyFromToRspImpl(const NDArray& from, const NDArray& to, RunContext
 // Make a copy of a dense NDArray
 template<typename from_xpu, typename to_xpu>
 inline void CopyFromToDnsImpl(const NDArray& from, const NDArray& to, RunContext ctx) {
-#if MXNET_USE_MKLDNN == 100
+#if MXNET_USE_MKLDNN == 1
   // If neither is MKLDNN, we can copy data normally.
   if (!from.IsMKLDNNData() && !to.IsMKLDNNData()) {
 #endif
@@ -1091,7 +1091,7 @@ inline void CopyFromToDnsImpl(const NDArray& from, const NDArray& to, RunContext
     TBlob tmp = to.data();
     ndarray::Copy<from_xpu, to_xpu>(from.data(), &tmp,
                                     from.ctx(), to.ctx(), ctx);
-#if MXNET_USE_MKLDNN == 100
+#if MXNET_USE_MKLDNN == 1
   } else if (SupportMKLDNN(from.dtype(), from.shape())
              && SupportMKLDNN(to.dtype(), to.shape())
              && from.ctx().dev_mask() == cpu::kDevMask
@@ -1617,7 +1617,7 @@ void NDArray::Save(dmlc::Stream *strm) const {
   } else {
     this->WaitToRead();
     nd_cpu = *this;
-#if MXNET_USE_MKLDNN == 100
+#if MXNET_USE_MKLDNN == 1
     if (nd_cpu.IsMKLDNNData()) {
       nd_cpu = nd_cpu.Reorder2Default();
     }
@@ -2007,7 +2007,7 @@ void NDArray::SyncCopyToCPU(void *data, size_t size) const {
     this->WaitToRead();
     RunContext rctx{this->ctx(), nullptr, nullptr, false};
     NDArray src = *this;
-#if MXNET_USE_MKLDNN == 100
+#if MXNET_USE_MKLDNN == 1
     if (src.IsMKLDNNData()) {
       src = this->Reorder2Default();
     }