fix strategy only with hip

ginkgo-project · Nov 13, 2019 · c3075ac · c3075ac
1 parent 867e8ff
commit c3075ac
Showing 1 changed file with 68 additions and 34 deletions.
diff --git a/include/ginkgo/core/matrix/csr.hpp b/include/ginkgo/core/matrix/csr.hpp
@@ -228,26 +228,31 @@ class Csr : public EnableLinOp<Csr<ValueType, IndexType>>,
 
         int64_t clac_size(const int64_t nnz)
         {
-            int multiple = 8;
-            if (nnz >= 2000000) {
-                multiple = 128;
-            } else if (nnz >= 200000) {
-                multiple = 32;
-            }
+            if (warp_size_ > 0) {
+                int multiple = 8;
+                if (nnz >= 2000000) {
+                    multiple = 128;
+                } else if (nnz >= 200000) {
+                    multiple = 32;
+                }
 
 #if GINKGO_HIP_PLATFORM_HCC
-            if (!cuda_strategy_) {
-                multiple = 8;
-                if (nnz >= 10000000) {
-                    multiple = 64;
-                } else if (nnz >= 1000000) {
-                    multiple = 16;
+                if (!cuda_strategy_) {
+                    multiple = 8;
+                    if (nnz >= 10000000) {
+                        multiple = 64;
+                    } else if (nnz >= 1000000) {
+                        multiple = 16;
+                    }
                 }
-            }
 #endif  // GINKGO_HIP_PLATFORM_HCC
 
-            auto nwarps = nwarps_ * multiple;
-            return min(ceildiv(nnz, warp_size_), static_cast<int64_t>(nwarps));
+                auto nwarps = nwarps_ * multiple;
+                return min(ceildiv(nnz, warp_size_),
+                           static_cast<int64_t>(nwarps));
+            } else {
+                return 0;
+            }
         }
 
     private:
@@ -328,33 +333,62 @@ class Csr : public EnableLinOp<Csr<ValueType, IndexType>>,
 
     void convert_to(Csr<ValueType, IndexType> *result) const override
     {
+        auto executor = result->get_executor();
+        bool same_executor = this->get_executor() == executor;
+        auto strategy = this->get_strategy();
         EnableLinOp<Csr>::convert_to(result);
-        if (this->get_executor() != result->get_executor() &&
-            result->get_strategy()->get_name() == "load_balance") {
-            if (auto exec = std::dynamic_pointer_cast<const HipExecutor>(
-                    result->get_executor())) {
-                result->set_strategy(std::make_shared<load_balance>(exec));
-            } else if (auto exec =
-                           std::dynamic_pointer_cast<const CudaExecutor>(
-                               result->get_executor())) {
-                result->set_strategy(std::make_shared<load_balance>(exec));
+        if (!same_executor) {
+            // When strategy is load_balance or automatical, rebuild the
+            // strategy according to executor's property.
+            if (std::dynamic_pointer_cast<load_balance>(strategy)) {
+                if (auto exec = std::dynamic_pointer_cast<const HipExecutor>(
+                        executor)) {
+                    result->set_strategy(std::make_shared<load_balance>(exec));
+                } else if (auto exec =
+                               std::dynamic_pointer_cast<const CudaExecutor>(
+                                   executor)) {
+                    result->set_strategy(std::make_shared<load_balance>(exec));
+                }
+            } else if (std::dynamic_pointer_cast<automatical>(strategy)) {
+                if (auto exec = std::dynamic_pointer_cast<const HipExecutor>(
+                        executor)) {
+                    result->set_strategy(std::make_shared<automatical>(exec));
+                } else if (auto exec =
+                               std::dynamic_pointer_cast<const CudaExecutor>(
+                                   executor)) {
+                    result->set_strategy(std::make_shared<automatical>(exec));
+                }
             }
         }
     }
 
     void move_to(Csr<ValueType, IndexType> *result) override
     {
-        bool same_executor = this->get_executor() == result->get_executor();
+        auto executor = result->get_executor();
+        bool same_executor = this->get_executor() == executor;
+        auto strategy = this->get_strategy();
         EnableLinOp<Csr>::move_to(result);
-        if (!same_executor &&
-            result->get_strategy()->get_name() == "load_balance") {
-            if (auto exec = std::dynamic_pointer_cast<const HipExecutor>(
-                    result->get_executor())) {
-                result->set_strategy(std::make_shared<load_balance>(exec));
-            } else if (auto exec =
-                           std::dynamic_pointer_cast<const CudaExecutor>(
-                               result->get_executor())) {
-                result->set_strategy(std::make_shared<load_balance>(exec));
+        if (!same_executor) {
+            // When strategy is load_balance or automatical, rebuild the
+            // strategy according to executor's property.
+            if (std::dynamic_pointer_cast<load_balance>(strategy)) {
+                if (auto exec = std::dynamic_pointer_cast<const HipExecutor>(
+                        executor)) {
+                    result->set_strategy(std::make_shared<load_balance>(exec));
+                } else if (auto exec =
+                               std::dynamic_pointer_cast<const CudaExecutor>(
+                                   executor)) {
+                    result->set_strategy(std::make_shared<load_balance>(exec));
+                }
+            } else if (std::dynamic_pointer_cast<automatical>(strategy)) {
+                if (auto exec = std::dynamic_pointer_cast<const HipExecutor>(
+                        executor)) {
+                    result->set_strategy(std::make_shared<automatical>(exec));
+                } else if (auto exec =
+                               std::dynamic_pointer_cast<const CudaExecutor>(
+                                   executor)) {
+                    result->set_strategy(std::make_shared<automatical>(exec));
+                }
             }
         }
     }