Fix backend functions bug

shashwatup9k · Jul 4, 2019 · cb37661 · cb37661
1 parent d82bd9d
commit cb37661
Show file tree

Hide file tree

Showing 4 changed files with 27 additions and 18 deletions.
diff --git a/src/rescorer/rescorer.h b/src/rescorer/rescorer.h
@@ -69,8 +69,10 @@ class Rescore : public ModelTask {
     for(auto device : devices) {
       auto graph = New<ExpressionGraph>(true);
       graph->setDevice(device);
-      graph->getBackend()->setOptimized(options_->get<bool>("optimize"));
-      graph->getBackend()->setGemmType(options_->get<std::string>("gemm-type"));
+      if (device.type == DeviceType::cpu) {
+        graph->getBackend()->setOptimized(options_->get<bool>("optimize"));
+        graph->getBackend()->setGemmType(options_->get<std::string>("gemm-type"));
+      }
       graph->reserveWorkspaceMB(options_->get<size_t>("workspace"));
       graphs_.push_back(graph);
     }

diff --git a/src/tensors/backend.h b/src/tensors/backend.h
@@ -36,12 +36,12 @@ class Backend {
   virtual void setClip(float clipValue) { clipValue_ = clipValue; }
   float getClip() { return clipValue_; }
 
-  // for CPU & inference only, sets to use optimized code for inference.
-  // The program aborts if these are called from GPU device.
+  // for CPU, sets to use optimized code for inference.
+  // for GPU, this is always false.
   virtual void setOptimized(bool optimize) = 0;
   virtual bool isOptimized() = 0;
-  // for CPU only, selects different GEMM types for the inference.
-  // The program aborts if these are called from GPU device.
+  // for CPU, selects different GEMM types for the inference.
+  // for GPU, there's no gemm type. so, it does nothing.
   virtual void setGemmType(std::string gemmType) = 0;
   virtual GemmType getGemmType() = 0;
 };

diff --git a/src/tensors/gpu/backend.h b/src/tensors/gpu/backend.h
@@ -34,18 +34,21 @@ class Backend : public marian::Backend {
   cublasHandle_t getCublasHandle() { return cublasHandle_; }
   cusparseHandle_t getCusparseHandle() { return cusparseHandle_; }
 
-  // for CPU & inference only, sets to use optimized code for inference.
-  // The program aborts if these are called from GPU device.
-  void setOptimized(bool optimize) override { ABORT("Not supported for GPU_{}", optimize); }
+  // for CPU, sets to use optimized code for inference.
+  // for GPU, this is always false.
+  void setOptimized(bool optimize) override {
+    LOG(info, "Not supported for GPU_{}", optimize);
+  }
   bool isOptimized() override {
-    ABORT("Not supported for GPU");
     return false;
   }
-  // for CPU only, selects different GEMM types for the inference.
-  // The program aborts if these are called from GPU device.
-  void setGemmType(std::string gemmType) override { ABORT("Not supported for GPU_{}", gemmType); }
+  // for CPU, selects different GEMM types for the inference.
+  // for GPU, there's no gemm type. so, it does nothing.
+  void setGemmType(std::string gemmType) override {
+    LOG(info, "Not supported for GPU_{}", gemmType);
+  }
   GemmType getGemmType() override {
-    ABORT("Not supported for GPU");
+    LOG(info, "Not supported for GPU");
     return GemmType::Auto;
   }
 

diff --git a/src/translator/translator.h b/src/translator/translator.h
@@ -58,8 +58,10 @@ class Translate : public ModelTask {
         auto graph = New<ExpressionGraph>(true);
         graph->setDevice(device);
         graph->getBackend()->setClip(options_->get<float>("clip-gemm"));
-        graph->getBackend()->setOptimized(options_->get<bool>("optimize"));
-        graph->getBackend()->setGemmType(options_->get<std::string>("gemm-type"));
+        if (device.type == DeviceType::cpu) {
+          graph->getBackend()->setOptimized(options_->get<bool>("optimize"));
+          graph->getBackend()->setGemmType(options_->get<std::string>("gemm-type"));
+        }
         graph->reserveWorkspaceMB(options_->get<size_t>("workspace"));
         graphs_[id] = graph;
 
@@ -172,8 +174,10 @@ class TranslateService : public ModelServiceTask {
       auto graph = New<ExpressionGraph>(true);
       graph->setDevice(device);
       graph->getBackend()->setClip(options_->get<float>("clip-gemm"));
-      graph->getBackend()->setOptimized(options_->get<bool>("optimize"));
-      graph->getBackend()->setGemmType(options_->get<std::string>("gemm-type"));
+      if (device.type == DeviceType::cpu) {
+        graph->getBackend()->setOptimized(options_->get<bool>("optimize"));
+        graph->getBackend()->setGemmType(options_->get<std::string>("gemm-type"));
+      }
       graph->reserveWorkspaceMB(options_->get<size_t>("workspace"));
       graphs_.push_back(graph);