removed SIMD for MSVC

CMakeLists.txt

@@ -125,11 +125,6 @@ if(APPLE)
       find_package(OpenMP REQUIRED)  
     endif()
 else()
-    # Set OpenMP runtime for MSVC
-    if (MSVC)
-        set(OpenMP_RUNTIME_MSVC experimental)
-        message(STATUS "OpenMP_RUNTIME_MSVC set to ${OpenMP_RUNTIME_MSVC}")
-    endif()
     find_package(OpenMP REQUIRED)
     set(OpenMP_CXX_FLAGS ${OpenMP_CXX_FLAGS})
     set(OpenMP_CXX_LIB_NAMES ${OpenMP_CXX_LIB_NAMES})
@@ -197,7 +192,8 @@ elseif (WIN32)
     target_link_libraries(gbrl_cpp PRIVATE OpenMP::OpenMP_CXX)
     if (USE_CUDA)
         set(cuda_lib_path "${CUDAToolkit_ROOT_DIR}/lib/x64")
-        target_link_libraries(gbrl_cpp PRIVATE ${cuda_lib_path}/cudart.lib)
+        # target_link_libraries(gbrl_cpp PRIVATE ${cuda_lib_path}/cudart.lib)
+        target_link_libraries(gbrl_cpp PRIVATE CUDA::cudart)
     endif()
 endif()
 

gbrl/src/cpp/CMakeLists.txt

@@ -33,8 +33,7 @@ if (CMAKE_SYSTEM_NAME STREQUAL "Linux")
 elseif (CMAKE_SYSTEM_NAME STREQUAL "Darwin")
     set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -O3 -std=c++14 ${OpenMP_C_FLAGS} -Wall -Wpedantic -Wextra")
 elseif (WIN32)
-    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /std:c++14 ${OpenMP_CXX_FLAGS} /O2 /W3 ")
-    message(STATUS ${OpenMP_CXX_FLAGS} )
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /O2 /std:c++14 ${OpenMP_CXX_FLAGS} /W3")
 endif()
 
 if (CMAKE_BUILD_TYPE STREQUAL "Debug")

gbrl/src/cpp/fitter.cpp

@@ -177,7 +177,9 @@ float Fitter::fit_cpu(dataSet *dataset, const float* targets, ensembleData *edat
         if (metadata->split_score_func == L2){
             float *mean_grads = calculate_mean(build_grads, batch_n_samples, output_dim, par_th);
             float *std = calculate_var_and_center(build_grads, mean_grads, batch_n_samples, output_dim, par_th);
-            #pragma omp simd
+#ifndef _MSC_VER
+    #pragma omp simd
+#endif
             for (int i = 0; i < output_dim; ++i)
                 std[i] = sqrtf(std[i]);
             divide_mat_by_vec_inplace(build_grads, std, batch_dataset.n_samples, metadata->output_dim, metadata->par_th);
@@ -527,7 +529,9 @@ void Fitter::calc_leaf_value(dataSet *dataset, ensembleData *edata, ensembleMeta
         }
         if (passed){
             idx = i*output_dim;
-            #pragma omp simd
+#ifndef _MSC_VER
+    #pragma omp simd
+#endif
             for (int d = 0; d < output_dim; ++d)
                 edata->values[leaf_idx*output_dim + d] += grads[idx + d];
             count += 1;

gbrl/src/cpp/loss.cpp

@@ -25,7 +25,9 @@ float MultiRMSE::get_loss_and_gradients(const float *raw_preds, const float *raw
         int thread_id = omp_get_thread_num();
         int start_idx = thread_id * elements_per_thread;
         int end_idx = (thread_id == n_threads - 1) ? n_elements : start_idx + elements_per_thread;
-        #pragma omp simd
+#ifndef _MSC_VER
+    #pragma omp simd
+#endif
         for (int i = start_idx; i < end_idx; ++i){
             row = i / output_dim;
             col = i % output_dim;
@@ -55,7 +57,9 @@ float MultiRMSE::get_loss(const float *raw_preds, const float *raw_targets, cons
             int end_idx = (thread_id == n_threads - 1) ? n_samples : start_idx + samples_per_thread;
             for (int sample_idx = start_idx; sample_idx < end_idx; ++sample_idx){
                 row = sample_idx * output_dim;
-                #pragma omp simd 
+#ifndef _MSC_VER
+    #pragma omp simd
+#endif 
                 for (int d = 0; d < output_dim; ++d){
                     grad_value = raw_preds[row + d] - raw_targets[row + d];
                     losses[thread_id] += (grad_value * grad_value);