From 65d83e288f94765f6638d92ae4c82f0e1abbfe09 Mon Sep 17 00:00:00 2001
From: amdsc21 <96135754+amdsc21@users.noreply.github.com>
Date: Wed, 19 Apr 2023 19:53:26 +0200
Subject: [PATCH] fix device query

---
 src/tree/gpu_hist/histogram.cu | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/src/tree/gpu_hist/histogram.cu b/src/tree/gpu_hist/histogram.cu
index 426343901..c6f6b79b2 100644
--- a/src/tree/gpu_hist/histogram.cu
+++ b/src/tree/gpu_hist/histogram.cu
@@ -306,12 +306,14 @@ void BuildGradientHistogram(CUDAContext const* ctx, EllpackDeviceAccessor const&
     dh::safe_cuda(cudaDeviceGetAttribute(&n_mps, cudaDevAttrMultiProcessorCount, device));
     int n_blocks_per_mp = 0;
     dh::safe_cuda(cudaOccupancyMaxActiveBlocksPerMultiprocessor(&n_blocks_per_mp, kernel,
+                                                                kBlockThreads, smem_size));
 #elif defined(XGBOOST_USE_HIP)
     dh::safe_cuda(hipDeviceGetAttribute(&n_mps, hipDeviceAttributeMultiprocessorCount, device));
     int n_blocks_per_mp = 0;
     dh::safe_cuda(hipOccupancyMaxActiveBlocksPerMultiprocessor(&n_blocks_per_mp, kernel,
-#endif
                                                                 kBlockThreads, smem_size));
+#endif
+
     // This gives the number of blocks to keep the device occupied
     // Use this as the maximum number of blocks
     unsigned grid_size = n_blocks_per_mp * n_mps;