diff --git a/CMakeLists.txt b/CMakeLists.txt
index c58cfae..c16503e 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -31,7 +31,7 @@ FetchContent_MakeAvailable(cudawrappers)
 FetchContent_Declare(
   ccglib
   GIT_REPOSITORY https://git.astron.nl/RD/recruit/ccglib
-  GIT_TAG packing_complex_last) # temporary to support complex-last data
+  GIT_TAG main)
 FetchContent_MakeAvailable(ccglib)
 
 add_library(tcbf SHARED src/tcbf.cu)
diff --git a/src/tcbf.cu b/src/tcbf.cu
index 41104b3..4940d2d 100644
--- a/src/tcbf.cu
+++ b/src/tcbf.cu
@@ -35,9 +35,9 @@ Beamformer::Beamformer(const size_t pixels, const size_t frames, const size_t sa
   pack_rf_ = std::make_unique<ccglib::packing::Packing>(COMPLEX * frames_padded_ * samples_padded_, device_, stream_);
   transpose_rf_ = std::make_unique<ccglib::transpose::Transpose>(
       kBatchSize, frames_padded_, samples_padded_, kGEMMTileSize.y, kGEMMTileSize.z, kBitsPerSample, device_, stream_);
-  gemm_ =
-      std::make_unique<ccglib::mma::GEMM>(kBatchSize, pixels_padded_, frames_padded_, samples_padded_, kBitsPerSample,
-                                          device_, stream_, kGEMMPrecision, kGEMMVariant, ccglib::mma::col_major);
+  gemm_ = std::make_unique<ccglib::mma::GEMM>(kBatchSize, pixels_padded_, frames_padded_, samples_padded_,
+                                              kBitsPerSample, device_, stream_, kGEMMPrecision, kGEMMVariant,
+                                              ccglib::mma::complex_middle, ccglib::mma::col_major);
 }
 
 void Beamformer::read_A_matrix(const std::string path) {