diff --git a/ci/regression.sh b/ci/regression.sh
index 8adb5427..b6fa3c4f 100755
--- a/ci/regression.sh
+++ b/ci/regression.sh
@@ -11,15 +11,27 @@ make -s
 ./ci/test_opencl.sh
 ./ci/test_driver.sh  
 ./ci/test_simx.sh
-./ci/test_compiler.sh
+#./ci/test_compiler.sh
 
-# Blackbox tests
+# warp/threads configurations
+./ci/travis_run.py ./ci/blackbox.sh --driver=rtlsim --cores=1 --warps=2 --threads=2 --app=demo
+./ci/travis_run.py ./ci/blackbox.sh --driver=rtlsim --cores=1 --warps=2 --threads=8 --app=demo
+./ci/travis_run.py ./ci/blackbox.sh --driver=rtlsim --cores=1 --warps=8 --threads=2 --app=demo
+
+# cores clustering
+./ci/travis_run.py ./ci/blackbox.sh --driver=rtlsim --cores=1 --clusters=1 --app=demo --args="-n1"
+./ci/travis_run.py ./ci/blackbox.sh --driver=rtlsim --cores=2 --clusters=1 --app=demo --args="-n1"
+./ci/travis_run.py ./ci/blackbox.sh --driver=rtlsim --cores=2 --clusters=2 --app=demo --args="-n1"
+
+# L2/L3
+./ci/travis_run.py ./ci/blackbox.sh --driver=rtlsim --cores=2 --l2cache --app=demo --args="-n1"
+./ci/travis_run.py ./ci/blackbox.sh --driver=rtlsim --cores=2 --clusters=2 --l3cache --app=demo --args="-n1"
+./ci/travis_run.py ./ci/blackbox.sh --driver=rtlsim --cores=2 --clusters=2 --l2cache --l3cache --app=demo --args="-n1"
+
+# build flags
 ./ci/travis_run.py ./ci/blackbox.sh --driver=vlsim --cores=1 --perf --app=demo --args="-n1"
 ./ci/travis_run.py ./ci/blackbox.sh --driver=vlsim --cores=1 --debug --app=demo --args="-n1"
 ./ci/travis_run.py ./ci/blackbox.sh --driver=vlsim --cores=1 --scope --app=basic --args="-t0 -n1"
-./ci/travis_run.py ./ci/blackbox.sh --driver=rtlsim --cores=2 --clusters=2 --app=demo --args="-n1"
-./ci/travis_run.py ./ci/blackbox.sh --driver=rtlsim --cores=2 --l2cache --app=demo --args="-n1"
-./ci/travis_run.py ./ci/blackbox.sh --driver=rtlsim --cores=2 --clusters=2 --l2cache --l3cache --app=demo --args="-n1"
 
 # disabling M extension
 CONFIGS=-DEXT_M_DISABLE make -C hw/simulate
@@ -31,7 +43,7 @@ CONFIGS=-DEXT_F_DISABLE make -C hw/simulate
 CONFIGS=-DSM_ENABLE=0 make -C hw/simulate
 
 # using FPNEW core
-FPU_CORE=FPU_FPNEW ./ci/blackbox.sh --driver=rtlsim --cores=1 --app=demo
+FPU_CORE=FPU_FPNEW ./ci/blackbox.sh --driver=rtlsim --cores=1 --app=dogfood
 
 # test 128-bit MEM block
 CONFIGS=-DMEM_BLOCK_SIZE=16 ./ci/blackbox.sh --driver=vlsim --cores=1 --app=demo
diff --git a/hw/rtl/afu/VX_to_mem.v b/hw/rtl/afu/VX_to_mem.v
index 7e786e50..901aaa36 100644
--- a/hw/rtl/afu/VX_to_mem.v
+++ b/hw/rtl/afu/VX_to_mem.v
@@ -72,7 +72,7 @@ module VX_to_mem #(
         assign mem_req_valid_out  = mem_req_valid_in;
         assign mem_req_rw_out     = mem_req_rw_in;
         assign mem_req_byteen_out = DST_DATA_SIZE'(mem_req_byteen_in) << ((DST_LDATAW-3)'(req_idx) << (SRC_LDATAW-3));  
-        assign mem_req_data_out   = DST_DATA_WIDTH'(mem_req_data_in) << ((DST_LDATAW'(req_idx)) << DST_LDATAW);
+        assign mem_req_data_out   = DST_DATA_WIDTH'(mem_req_data_in) << ((DST_LDATAW'(req_idx)) << SRC_LDATAW);
         assign mem_req_tag_out    = DST_TAG_WIDTH'({mem_req_tag_in, req_idx});
         assign mem_req_ready_in   = mem_req_ready_out;