diff --git a/fpga/fpgactl b/fpga/fpgactl
index d5f478a..a0e9cf1 100755
--- a/fpga/fpgactl
+++ b/fpga/fpgactl
@@ -101,27 +101,59 @@ function create_qdma_queue() # {{{
 function do_system_reset() # {{{
 {
   # UartBootEn (bit2) + system reset (bit0)
-  dma-ctl qdma08000 reg write bar 2 0x0 0x0
+  dma-ctl qdma08000 reg write bar 2 0x0 0x0 > /dev/null
   sleep 0.2
   # Release system reset, we must wait until the memory is filled with 0s
-  dma-ctl qdma08000 reg write bar 2 0x0 0x1
+  dma-ctl qdma08000 reg write bar 2 0x0 0x1 > /dev/null
   #sleep 5
 } # }}}
 
 function do_system_release() # {{{
 {
   # Release Ariane's reset
-  dma-ctl qdma08000 reg write bar 2 0x0 0x3
+  dma-ctl qdma08000 reg write bar 2 0x0 0x3 > /dev/null
+} # }}}
+
+function copy_by_dma() # {{{
+{
+  ifile="$1"
+  address="$2"
+
+  ofile="/dev/qdma08000-MM-1"
+  bs=$((8*1024*1024)) # 8 MiB
+
+  total_size=$(stat --format "%s" "$ifile")
+
+  nblocks=$(( ($total_size + $bs - 1) / $bs ))
+  skip=0
+
+  # Using just a single dd command doesn't seem to work. My hypothesis is that
+  # the driver doesn't allow multiple writes without setting the fseek position.
+  # So we keep copying blocks until the end of the file, one at a time.
+
+  while [ "$skip" -lt "$nblocks" ]; do
+    dst=$(($address + $skip * $bs))
+    dd if="$ifile" skip=$skip count=1 bs=$bs of="$ofile" seek=$dst oflag=seek_bytes
+    let skip=$skip+1
+  done
+
 } # }}}
 
 function load_file_in_memory() # {{{
 {
   file="$1"
   address="$2"
+
+  # Previous tests...
   #strace -f dma-to-device -d /dev/qdma08000-MM-1 -a "$address" -s $((8*1024*1024)) -f "$file"
   #strace -f dd if="$file" bs=16M seek="${address}" oflag=seek_bytes of=/dev/qdma08000-MM-1 status=progress conv=sync
   #strace -f fpgakit/fpgadd -i "$file" -a "$address" -d /dev/qdma08000-MM-1 -c 1024 -s 1024
-  ID=08 ./load_image.sh "$file" "$address"
+  #ID=08 ./load_image.sh "$file" "$address"
+
+  # Now dd seems to work fine, but I will leave this as fallback:
+  #copy_by_dma "$file" "$address"
+
+  dd if="$file" bs=8M seek="${address}" oflag=seek_bytes of=/dev/qdma08000-MM-1 status=none
 } # }}}
 
 function do_boot_only() # {{{
@@ -190,9 +222,22 @@ function upload_bitstream_file()  # {{{
     exit 1
   fi
 
-  vivado -nolog -nojournal -mode batch -source load-bitstream.tcl -tclargs "$bitfile" "$fpgajtag"
+  script=$(mktemp vivado-XXXXXXXXXX.tcl)
+  cat > "$script" <<EOF
+open_hw_manager
+connect_hw_server -url localhost:3121
+current_hw_target "localhost:3121/xilinx_tcf/Xilinx/${fpgajtag}A"
+open_hw_target
+set dev [lindex [get_hw_devices] 0]
+current_hw_device \$dev
+set_property PROGRAM.FILE ${bitfile} \$dev
+program_hw_devices \$dev
+exit
+EOF
+
+  vivado -nolog -nojournal -mode batch -source "$script"
+  rm "$script"
 
-  echo "Bitstream   $bitfile   loaded."
   killall hw_server
 } # }}}
 
@@ -256,25 +301,25 @@ function load_bitstream()  # {{{
   create_qdma_queue
 } # }}}
 
-slotid="8"
 bitstream=
 bootloader=
 kernelfile=
 reloadfs=
 resetcpu=
+verbose=
 
-while getopts "s:w:k:r:b:" opt; do
+while getopts "s:w:k:r:b:v" opt; do
   case "${opt}" in
-    s) slotid="${OPTARG}" ;;
     w) bitstream="${OPTARG}" ;;
     b) bootloader="${OPTARG}"; resetcpu=1 ;;
     k) kernelfile="${OPTARG}"; resetcpu=1 ;;
     r) rootfs="${OPTARG}"; resetcpu=1 ;;
+    v) verbose=1 ;;
     *) usage ;;
   esac
 done
 
-set -x
+test "$verbose"     && set -x
 
 check_environment