diff --git a/paddle/fluid/operators/bilateral_slice_op.cu b/paddle/fluid/operators/bilateral_slice_op.cu
index e56a4be53d14968da60259b69e74e1e988caac8a..3c64ed1acc847d8f60ad39bf3437b22ad8f2bb4a 100644
--- a/paddle/fluid/operators/bilateral_slice_op.cu
+++ b/paddle/fluid/operators/bilateral_slice_op.cu
@@ -472,8 +472,8 @@ class BilateralSliceGradOpCUDAKernel : public framework::OpKernel<T> {
     grid_sizes.gw = gw;
     grid_sizes.input_chans = input_chans;
 
-    platform::GpuLaunchConfig config =
-        platform::GetGpuLaunchConfig1D(ctx.cuda_device_context(), grid_count);
+    platform::GpuLaunchConfig config = platform::GetGpuLaunchConfig1D(
+        ctx.cuda_device_context(), grid_count, 512);
 
     BilateralSliceCudaGridGradKernel<
         T><<<config.block_per_grid, config.thread_per_block, 0,
@@ -481,8 +481,8 @@ class BilateralSliceGradOpCUDAKernel : public framework::OpKernel<T> {
         grid_grad_data, output_grad_data, guide_data, input_data, grid_sizes,
         has_offset, grid_count, output_chans);
 
-    config =
-        platform::GetGpuLaunchConfig1D(ctx.cuda_device_context(), guide_count);
+    config = platform::GetGpuLaunchConfig1D(ctx.cuda_device_context(),
+                                            guide_count, 512);
 
     BilateralSliceCudaGuideGradKernel<
         T><<<config.block_per_grid, config.thread_per_block, 0,
@@ -490,8 +490,8 @@ class BilateralSliceGradOpCUDAKernel : public framework::OpKernel<T> {
         guide_grad_data, output_grad_data, grid_data, guide_data, input_data,
         grid_sizes, has_offset, guide_count, output_chans);
 
-    config =
-        platform::GetGpuLaunchConfig1D(ctx.cuda_device_context(), input_count);
+    config = platform::GetGpuLaunchConfig1D(ctx.cuda_device_context(),
+                                            input_count, 512);
 
     BilateralSliceCudaInputGradKernel<
         T><<<config.block_per_grid, config.thread_per_block, 0,
diff --git a/paddle/fluid/platform/gpu_launch_config.h b/paddle/fluid/platform/gpu_launch_config.h
index 57074452d88b259c5edc188260b7f2dfd605764a..920be1a8e497c37d5ade77ffa8e5a697f71a4310 100644
--- a/paddle/fluid/platform/gpu_launch_config.h
+++ b/paddle/fluid/platform/gpu_launch_config.h
@@ -36,7 +36,8 @@ struct GpuLaunchConfig {
 };
 
 inline GpuLaunchConfig GetGpuLaunchConfig1D(
-    const platform::CUDADeviceContext& context, int element_count) {
+    const platform::CUDADeviceContext& context, int element_count,
+    int max_threads = 1024) {
   PADDLE_ENFORCE_GT(element_count, 0,
                     platform::errors::InvalidArgument(
                         "element count should be greater than 0,"
@@ -53,7 +54,8 @@ inline GpuLaunchConfig GetGpuLaunchConfig1D(
       std::min(max_physical_threads, theory_thread_count);
 
   // Need get from device
-  const int thread_per_block = std::min(1024, context.GetMaxThreadsPerBlock());
+  const int thread_per_block =
+      std::min(max_threads, context.GetMaxThreadsPerBlock());
   const int block_count =
       std::min(DivUp(physical_thread_count, thread_per_block), sm);
 
diff --git a/tools/windows/run_unittests.sh b/tools/windows/run_unittests.sh
index 1399bf46995ef3f561f4cf8d9bdc76ab6ffa990e..b89ce54edf149f14ba71db80b60c9c4d1a72c3ca 100644
--- a/tools/windows/run_unittests.sh
+++ b/tools/windows/run_unittests.sh
@@ -53,7 +53,6 @@ diable_wingpu_test="^test_analysis_predictor$|\
 ^test_dataloader_unkeep_order$|\
 ^test_model$|\
 ^test_add_reader_dependency$|\
-^test_bilateral_slice_op$|\
 ^test_cholesky_op$|\
 ^test_dataloader_early_reset$|\
 ^test_decoupled_py_reader$|\