Added verbose oneDNN lib version (#29378)

paddle/fluid/operators/mkldnn/concat_mkldnn_op.cc

@@ -144,6 +144,7 @@ class ConcatMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
         platform::errors::InvalidArgument(
             "The axis is expected to be in range of [%d, %d), but got %d",
             -rank, rank, concat_axis));
+    platform::MKLDNNDeviceContext::tls().log_lib_version();
     if (concat_axis < 0) {
       concat_axis = concat_axis + rank;
     }

paddle/fluid/operators/mkldnn/fc_mkldnn_op.cc

@@ -572,6 +572,7 @@ class FCMKLDNNOpKernel : public framework::OpKernel<T_in> {
     PADDLE_ENFORCE_EQ(
         platform::is_cpu_place(ctx.GetPlace()), true,
         platform::errors::PreconditionNotMet("FC MKL-DNN must use CPUPlace."));
+    platform::MKLDNNDeviceContext::tls().log_lib_version();
     auto input = ctx.Input<LoDTensor>("Input");
     auto w = ctx.Input<Tensor>("W");
     auto bias = ctx.Input<Tensor>("Bias");

paddle/fluid/operators/mkldnn/matmul_mkldnn_op.cc

@@ -378,6 +378,7 @@ class DNNLMatMulKernel : public framework::OpKernel<T> {
                         platform::errors::Unimplemented(
                             "DNNL matmul doesn't support multiple heads."));
     }
+    platform::MKLDNNDeviceContext::tls().log_lib_version();
     ExecuteMatMul<T, T>(ctx);
   }
 };

paddle/fluid/operators/mkldnn/mul_mkldnn_op.cc

@@ -353,6 +353,7 @@ class MulMKLDNNKernel : public framework::OpKernel<XT> {
     PADDLE_ENFORCE_EQ(platform::is_cpu_place(ctx.GetPlace()), true,
                       paddle::platform::errors::PreconditionNotMet(
                           "Operator DNNL Mul must use CPUPlace"));
+    platform::MKLDNNDeviceContext::tls().log_lib_version();
     auto &dev_ctx = ctx.template device_context<MKLDNNDeviceContext>();
     const auto &mkldnn_engine = dev_ctx.GetEngine();
 

paddle/fluid/platform/device_context.cc

@@ -466,6 +466,15 @@ MKLDNNDeviceContextThreadLocals::Body::get_cur_paddle_data_layout(void) {
   return cur_paddle_data_layout;
 }
 
+void MKLDNNDeviceContextThreadLocals::Body::log_lib_version(void) {
+  if (!said_once) {
+    said_once = true;
+    auto dv = dnnl::version();
+    LOG(INFO) << "oneDNN v" << dv->major << "." << dv->minor << "."
+              << dv->patch;
+  }
+}
+
 void MKLDNNDeviceContext::ResetBlobMap() {
   std::lock_guard<decltype(*p_mutex_)> lock(*p_mutex_);
   if (!block_next_cache_clearing_) {

paddle/fluid/platform/device_context.h

@@ -466,6 +466,7 @@ class MKLDNNDeviceContextThreadLocals {
 
   typedef MKLDNNDeviceContextThreadLocals self;
   struct Body {
+    bool said_once = false;
     size_t cur_mkldnn_session_id;
     // Current data input shape string.
     // - For fixed-shape, it's a null string in default.
@@ -485,6 +486,7 @@ class MKLDNNDeviceContextThreadLocals {
     void set_cur_input_shape_cache_capacity(int input_shape_cache_capacity);
     void set_cur_paddle_data_layout(framework::DataLayout dl);
     framework::DataLayout get_cur_paddle_data_layout(void);
+    void log_lib_version(void);
   };
   MKLDNNDeviceContextThreadLocals() = default;
   MKLDNNDeviceContextThreadLocals(const MKLDNNDeviceContextThreadLocals& c) =

paddle/fluid/platform/mkldnn_reuse.h

@@ -45,7 +45,9 @@ class MKLDNNHandlerT {
         key_common_(base_key),
         key_(platform::ExtendKeyWithThreadInfoIfNeeded(dev_ctx, base_key)),
         fwd_pd_(nullptr),
-        bwd_pd_(nullptr) {}
+        bwd_pd_(nullptr) {
+    platform::MKLDNNDeviceContext::tls().log_lib_version();
+  }
 
   std::shared_ptr<TForward> AcquireForwardPrimitive() {
     const std::string key_p = key_ + "@fwd_p";
@@ -313,7 +315,9 @@ class MKLDNNHandler {
       : dev_ctx_(dev_ctx),
         engine_(engine),
         key_common_(base_key),
-        key_(platform::ExtendKeyWithThreadInfoIfNeeded(dev_ctx, base_key)) {}
+        key_(platform::ExtendKeyWithThreadInfoIfNeeded(dev_ctx, base_key)) {
+    platform::MKLDNNDeviceContext::tls().log_lib_version();
+  }
 
   std::shared_ptr<mkldnn::memory> AcquireSrcMemory(
       const mkldnn::memory::desc& md, void* ptr) {