diff --git a/src/finn/custom_op/fpgadataflow/templates.py b/src/finn/custom_op/fpgadataflow/templates.py
index f6114ea028817b4248b37572067ed80de8364712..4c38fea70d7e80f6cb17732f7f413bb2afb1b830 100644
--- a/src/finn/custom_op/fpgadataflow/templates.py
+++ b/src/finn/custom_op/fpgadataflow/templates.py
@@ -137,7 +137,9 @@ reg [31:0] config_d0 = 0;
 wire [31:0] config_q0;
 
 //multiple wire AXI Streams
-reg m_axis_0_afull = 0;
+reg m_axis_0_afull;
+// FIFO count to generate programmable full
+wire [5:0] fifo_0_count;
 wire m_axis_0_tready;
 wire m_axis_0_tvalid;
 wire $WEIGHT_RANGE$ m_axis_0_tdata;
@@ -257,12 +259,9 @@ mem
 
 );
 
-// two consecutive weight streamer FIFOs to provide the same functionality
-// as "programmable full"
 
-// weight streamer FIFO 1
 Q_srl #(
-.depth(16),
+.depth(32),
 .width($WEIGHT_WIDTH$)
 )
 $LAYER_NAME$_w_fifo_1
@@ -274,25 +273,10 @@ $LAYER_NAME$_w_fifo_1
  .i_r(m_axis_0_tready),
  .o_d(m_axis_0_tdata_q),
  .o_v(m_axis_0_tvalid_q),
- .o_r(m_axis_0_tready_q)
+ .o_r(m_axis_0_tready_q),
+ .count(fifo_0_count)
 );
 
-// weight streamer FIFO 2
-Q_srl #(
-.depth(16),
-.width($WEIGHT_WIDTH$)
-)
-$LAYER_NAME$_w_fifo_2
-(
- .clock(ap_clk),
- .reset(!ap_rst_n),
- .i_d(m_axis_0_tdata_q),
- .i_v(m_axis_0_tvalid_q),
- .i_r(m_axis_0_tready_q),
- .o_d(m_axis_0_tdata_q2),
- .o_v(m_axis_0_tvalid_q2),
- .o_r(m_axis_0_tready_q2)
-);
 
 //MVA_Stream_Unit
 
@@ -304,14 +288,16 @@ MVA_Stream_U
 .in0_V_V_TDATA(in0_V_V_TDATA),		//$IN_RANGE$ input
 .in0_V_V_TVALID(in0_V_V_TVALID),  	//input
 .in0_V_V_TREADY(in0_V_V_TREADY),	//output
-.weights_V_V_TDATA(m_axis_0_tdata_q2),	//$WEIGHT_RANGE$ input
-.weights_V_V_TVALID(m_axis_0_tvalid_q2),	//input
-.weights_V_V_TREADY(m_axis_0_tready_q2),	//output
+.weights_V_V_TDATA(m_axis_0_tdata_q),	//$WEIGHT_RANGE$ input
+.weights_V_V_TVALID(m_axis_0_tvalid_q),	//input
+.weights_V_V_TREADY(m_axis_0_tready_q),	//output
 .out_V_V_TDATA(out_V_V_TDATA),		//$OUT_RANGE$ output
 .out_V_V_TVALID(out_V_V_TVALID),	//output
 .out_V_V_TREADY(out_V_V_TREADY)		//input
 );
 
+// programmable full threshold at 16 elements
+assign m_axis_0_afull = (fifo_0_count > 16);
 
 endmodule
 """