--- XUSP3R.srcs/sources_1/ip/phy_ddr4/rtl/cal/ddr4_v2_2_cal_write.sv
@@ -157,9 +157,13 @@
 // used in pushing new write CAS commands into the wrQ shift register.  These values are
 // calculated for both casslot0 and casslot2.
 localparam FABRIC_CASSLOT0 = (   ALL_WR_LATENCY       / 4 ) - 2;
+localparam FABRIC_CASSLOT1 = ( ( ALL_WR_LATENCY + 1 ) / 4 ) - 2;
 localparam FABRIC_CASSLOT2 = ( ( ALL_WR_LATENCY + 2 ) / 4 ) - 2;
+localparam FABRIC_CASSLOT3 = ( ( ALL_WR_LATENCY + 3 ) / 4 ) - 2;
 localparam OFFSET_CASSLOT0 =   ( ALL_WR_LATENCY     ) % 4;
+localparam OFFSET_CASSLOT1 =   ( ALL_WR_LATENCY + 1 ) % 4;
 localparam OFFSET_CASSLOT2 =   ( ALL_WR_LATENCY + 2 ) % 4;
+localparam OFFSET_CASSLOT3 =   ( ALL_WR_LATENCY + 3 ) % 4;
 
 integer i;
 
@@ -192,10 +196,12 @@
 // Output enable shift register fabric load cycle for cal and mc.  mccasSlot2 is timing critical.
 reg  [3:0] oe_0_mux_cal;
 wire [3:0] oe_0_mux_cal_nxt = calDone    ? FABRIC_CASSLOT0 : FABRIC_CASSLOT2;
-wire [3:0] oe_fabric        = mccasSlot2 ? FABRIC_CASSLOT2 : oe_0_mux_cal;
+wire [3:0] oe_fabric        = mccasSlot2 ? (casSlot[0] ? FABRIC_CASSLOT3 : FABRIC_CASSLOT2) :
+                         (casSlot[0] ? FABRIC_CASSLOT1 : oe_0_mux_cal);
 
 // tCK offset for mc.  mccasSlot2 is timing critical.
-wire [1:0] tck_offset = mccasSlot2 ? OFFSET_CASSLOT2 : OFFSET_CASSLOT0;
+wire [1:0] tck_offset = mccasSlot2 ? (casSlot[0] ? OFFSET_CASSLOT3 : OFFSET_CASSLOT2) :
+                         (casSlot[0] ? OFFSET_CASSLOT1 : OFFSET_CASSLOT0);
 
 // Output enable shift register load value for cal and mc.  mcwrCAS and tck_offset are timing critical.
 wire [DBAW+4:0] wrQ_cal_load_value = {winBuf,   calRank, OFFSET_CASSLOT2[1:0], 1'b1};

--- XUSP3R.srcs/sources_1/ip/phy_ddr4_x8/rtl/cal/ddr4_v2_2_cal_write.sv
@@ -157,9 +157,13 @@
 // used in pushing new write CAS commands into the wrQ shift register.  These values are
 // calculated for both casslot0 and casslot2.
 localparam FABRIC_CASSLOT0 = (   ALL_WR_LATENCY       / 4 ) - 2;
+localparam FABRIC_CASSLOT1 = ( ( ALL_WR_LATENCY + 1 ) / 4 ) - 2;
 localparam FABRIC_CASSLOT2 = ( ( ALL_WR_LATENCY + 2 ) / 4 ) - 2;
+localparam FABRIC_CASSLOT3 = ( ( ALL_WR_LATENCY + 3 ) / 4 ) - 2;
 localparam OFFSET_CASSLOT0 =   ( ALL_WR_LATENCY     ) % 4;
+localparam OFFSET_CASSLOT1 =   ( ALL_WR_LATENCY + 1 ) % 4;
 localparam OFFSET_CASSLOT2 =   ( ALL_WR_LATENCY + 2 ) % 4;
+localparam OFFSET_CASSLOT3 =   ( ALL_WR_LATENCY + 3 ) % 4;
 
 integer i;
 
@@ -192,10 +196,12 @@
 // Output enable shift register fabric load cycle for cal and mc.  mccasSlot2 is timing critical.
 reg  [3:0] oe_0_mux_cal;
 wire [3:0] oe_0_mux_cal_nxt = calDone    ? FABRIC_CASSLOT0 : FABRIC_CASSLOT2;
-wire [3:0] oe_fabric        = mccasSlot2 ? FABRIC_CASSLOT2 : oe_0_mux_cal;
+wire [3:0] oe_fabric        = mccasSlot2 ? (casSlot[0] ? FABRIC_CASSLOT3 : FABRIC_CASSLOT2) :
+                         (casSlot[0] ? FABRIC_CASSLOT1 : oe_0_mux_cal);
 
 // tCK offset for mc.  mccasSlot2 is timing critical.
-wire [1:0] tck_offset = mccasSlot2 ? OFFSET_CASSLOT2 : OFFSET_CASSLOT0;
+wire [1:0] tck_offset = mccasSlot2 ? (casSlot[0] ? OFFSET_CASSLOT3 : OFFSET_CASSLOT2) :
+                         (casSlot[0] ? OFFSET_CASSLOT1 : OFFSET_CASSLOT0);
 
 // Output enable shift register load value for cal and mc.  mcwrCAS and tck_offset are timing critical.
 wire [DBAW+4:0] wrQ_cal_load_value = {winBuf,   calRank, OFFSET_CASSLOT2[1:0], 1'b1};

--- XUSP3R.srcs/sources_1/ip/phy_ddr4_x8_1R/rtl/cal/ddr4_v2_2_cal_write.sv
@@ -157,9 +157,13 @@
 // used in pushing new write CAS commands into the wrQ shift register.  These values are
 // calculated for both casslot0 and casslot2.
 localparam FABRIC_CASSLOT0 = (   ALL_WR_LATENCY       / 4 ) - 2;
+localparam FABRIC_CASSLOT1 = ( ( ALL_WR_LATENCY + 1 ) / 4 ) - 2;
 localparam FABRIC_CASSLOT2 = ( ( ALL_WR_LATENCY + 2 ) / 4 ) - 2;
+localparam FABRIC_CASSLOT3 = ( ( ALL_WR_LATENCY + 3 ) / 4 ) - 2;
 localparam OFFSET_CASSLOT0 =   ( ALL_WR_LATENCY     ) % 4;
+localparam OFFSET_CASSLOT1 =   ( ALL_WR_LATENCY + 1 ) % 4;
 localparam OFFSET_CASSLOT2 =   ( ALL_WR_LATENCY + 2 ) % 4;
+localparam OFFSET_CASSLOT3 =   ( ALL_WR_LATENCY + 3 ) % 4;
 
 integer i;
 
@@ -192,10 +196,12 @@
 // Output enable shift register fabric load cycle for cal and mc.  mccasSlot2 is timing critical.
 reg  [3:0] oe_0_mux_cal;
 wire [3:0] oe_0_mux_cal_nxt = calDone    ? FABRIC_CASSLOT0 : FABRIC_CASSLOT2;
-wire [3:0] oe_fabric        = mccasSlot2 ? FABRIC_CASSLOT2 : oe_0_mux_cal;
+wire [3:0] oe_fabric        = mccasSlot2 ? (casSlot[0] ? FABRIC_CASSLOT3 : FABRIC_CASSLOT2) :
+                         (casSlot[0] ? FABRIC_CASSLOT1 : oe_0_mux_cal);
 
 // tCK offset for mc.  mccasSlot2 is timing critical.
-wire [1:0] tck_offset = mccasSlot2 ? OFFSET_CASSLOT2 : OFFSET_CASSLOT0;
+wire [1:0] tck_offset = mccasSlot2 ? (casSlot[0] ? OFFSET_CASSLOT3 : OFFSET_CASSLOT2) :
+                         (casSlot[0] ? OFFSET_CASSLOT1 : OFFSET_CASSLOT0);
 
 // Output enable shift register load value for cal and mc.  mcwrCAS and tck_offset are timing critical.
 wire [DBAW+4:0] wrQ_cal_load_value = {winBuf,   calRank, OFFSET_CASSLOT2[1:0], 1'b1};
 
--- XUSP3R.srcs/sources_1/ip/phy_ddr4_udimm/rtl/cal/ddr4_v2_2_cal_write.sv
@@ -157,9 +157,13 @@
 // used in pushing new write CAS commands into the wrQ shift register.  These values are
 // calculated for both casslot0 and casslot2.
 localparam FABRIC_CASSLOT0 = (   ALL_WR_LATENCY       / 4 ) - 2;
+localparam FABRIC_CASSLOT1 = ( ( ALL_WR_LATENCY + 1 ) / 4 ) - 2;
 localparam FABRIC_CASSLOT2 = ( ( ALL_WR_LATENCY + 2 ) / 4 ) - 2;
+localparam FABRIC_CASSLOT3 = ( ( ALL_WR_LATENCY + 3 ) / 4 ) - 2;
 localparam OFFSET_CASSLOT0 =   ( ALL_WR_LATENCY     ) % 4;
+localparam OFFSET_CASSLOT1 =   ( ALL_WR_LATENCY + 1 ) % 4;
 localparam OFFSET_CASSLOT2 =   ( ALL_WR_LATENCY + 2 ) % 4;
+localparam OFFSET_CASSLOT3 =   ( ALL_WR_LATENCY + 3 ) % 4;
 
 integer i;
 
@@ -192,10 +196,12 @@
 // Output enable shift register fabric load cycle for cal and mc.  mccasSlot2 is timing critical.
 reg  [3:0] oe_0_mux_cal;
 wire [3:0] oe_0_mux_cal_nxt = calDone    ? FABRIC_CASSLOT0 : FABRIC_CASSLOT2;
-wire [3:0] oe_fabric        = mccasSlot2 ? FABRIC_CASSLOT2 : oe_0_mux_cal;
+wire [3:0] oe_fabric        = mccasSlot2 ? (casSlot[0] ? FABRIC_CASSLOT3 : FABRIC_CASSLOT2) :
+                         (casSlot[0] ? FABRIC_CASSLOT1 : oe_0_mux_cal);
 
 // tCK offset for mc.  mccasSlot2 is timing critical.
-wire [1:0] tck_offset = mccasSlot2 ? OFFSET_CASSLOT2 : OFFSET_CASSLOT0;
+wire [1:0] tck_offset = mccasSlot2 ? (casSlot[0] ? OFFSET_CASSLOT3 : OFFSET_CASSLOT2) :
+                         (casSlot[0] ? OFFSET_CASSLOT1 : OFFSET_CASSLOT0);
 
 // Output enable shift register load value for cal and mc.  mcwrCAS and tck_offset are timing critical.
 wire [DBAW+4:0] wrQ_cal_load_value = {winBuf,   calRank, OFFSET_CASSLOT2[1:0], 1'b1};