diff --git a/dsp/ph_acc.v b/dsp/ph_acc.v
index a696666b6..7d1172816 100644
--- a/dsp/ph_acc.v
+++ b/dsp/ph_acc.v
@@ -4,6 +4,43 @@
 // Tuned to allow 32-bit control, divided 20-bit high and 12-bit low,
 // which gets merged to 32-bit binary when modulo is zero.
 // But also supports non-binary frequencies: see the modulo input port.
+
+// Note that phase_step_h and phase_step_l combined fit in a 32-bit word.
+// This is intentional, to allow atomic updates of the two controls
+// in 32-bit systems.  Indeed, when modulo==0, those 32 bits can be considered
+// a single phase step increment
+
+// The phase generation algorithm
+// 0. The phase increments for dds are generated using a technique described
+// in these 2 places:
+// Section: PROGRAMMABLE MODULUS MODE in:
+//  https://www.analog.com/media/en/technical-documentation/data-sheets/ad9915.pdf
+//  (AND) https://en.wikipedia.org/wiki/Bresenham%27s_line_algorithm
+//  Basically, increment the phase step at a coarse resolution, accumulate the
+//  error on the side, and when that error accumulates to the lowest bit of
+//  the coarse counter, add an extra 1 to the following phase step.
+// 1. phase_step_h is the coarse (20 bit) integer truncated phase increment for
+//  the cordic. There is a 1-bit increment of phase that comes from
+//  accumulating residue phase.
+// 2. This residue phase is accumulating in steps of phase_step_l, in a 12-bit
+//  counter.
+// 3. However, there will be an extra residue even for this 12-bit counter,
+// which is the modulus, and this added as an offset when the counter crosses 0
+
+// 12-bit modulo supports largest known periodicity in a suggested LLRF system,
+// 1427 for JLab.  For more normal periodicities, use a multiple to get finer
+// granularity.
+// Note that the downloaded modulo control is the 2's complement of the
+// mathematical modulus.
+// e.g., SSRF IF/F_s ratio 8/11, use
+//     modulo = 4096 - 372*11 = 4
+//     phase_step_h = 2^20*8/11 = 762600
+//     phase_step_l = (2^20*8%11)*372 = 2976
+// e.g., Argonne RIA test IF/F_s ratio 9/13, use
+//     modulo = 4096 - 315*13 = 1
+//     phase_step_h = 2^20*9/13 = 725937
+//     phase_step_l = (2^20*9%13)*315 = 945
+
 module ph_acc(
 	input clk,  // Rising edge clock input; all logic is synchronous in this domain
 	input reset,  // Active high, synchronous with clk
diff --git a/dsp/rot_dds.v b/dsp/rot_dds.v
index 9ab9e12b0..613fd3713 100644
--- a/dsp/rot_dds.v
+++ b/dsp/rot_dds.v
@@ -10,40 +10,6 @@
 // 8/11 for SSRF
 // 9/13 for Argonne RIA
 //
-// The phase generation algorithm
-// 0. The phase increments for dds are generated using a technique described
-// in these 2 places:
-// Section: PROGRAMMABLE MODULUS MODE in:
-//  https://www.analog.com/media/en/technical-documentation/data-sheets/ad9915.pdf
-//  (AND) https://en.wikipedia.org/wiki/Bresenham%27s_line_algorithm
-//  Basically, increment the phase step at a coarse resolution, accumulate the
-//  error on the side, and when that error accumulates to the lowest bit of
-//  the coarse counter, add an extra 1 to the following phase step.
-// 1. phase_step_h is the coarse (20 bit) integer truncated phase increment for
-//  the cordic. There is a 1-bit increment of phase that comes from
-//  accumulating residue phase.
-// 2. This residue phase is accumulating in steps of phase_step_l, in a 12-bit
-//  counter.
-// 3. However, there will be an extra residue even for this 12-bit counter,
-// which is the modulus, and this added as an offset when the counter crosses 0
-
-// 12-bit modulo supports largest known periodicity in a suggested LLRF system,
-// 1427 for JLab.  For more normal periodicities, use a multiple to get finer
-// granularity.
-// Note that the downloaded modulo control is the 2's complement of the
-// mathematical modulus.
-// e.g., SSRF IF/F_s ratio 8/11, use
-//     modulo = 4096 - 372*11 = 4
-//     phase_step_h = 2^20*8/11 = 762600
-//     phase_step_l = (2^20*8%11)*372 = 2976
-// e.g., Argonne RIA test IF/F_s ratio 9/13, use
-//     modulo = 4096 - 315*13 = 1
-//     phase_step_h = 2^20*9/13 = 725937
-//     phase_step_l = (2^20*9%13)*315 = 945
-
-// TODO:
-// Potentially, isolate phase generation into a separate module.
-// Haha, turns out there is ph_acc.v (We should USE it).
 // Synthesizes to ??? slices at ??? MHz in XC3Sxxx-4 using XST-??
 //
 
@@ -65,19 +31,18 @@ parameter lo_amp = 18'd79590;
 // Sometimes we cheat and use slightly smaller values than above,
 // to make other computations fit better.
 
+wire [18:0] phase_acc;
+ph_acc ph_acc_i (
+  .clk(clk), .reset(reset), .en(1'b1), // input
+  .phase_acc(phase_acc), // output [18:0]
+  .phase_step_h(phase_step_h), // input [19:0]
+  .phase_step_l(phase_step_l), // input [11:0]
+  .modulo(modulo) // input [11:0]
+);
 // See rot_dds_config
 
-reg carry=0, reset_d=0;
-reg [19:0] phase_h=0, phase_step_hp=0;
-reg [11:0] phase_l=0;
-always @(posedge clk) begin
-	{carry, phase_l} <= reset ? 13'b0 : ((carry ? modulo : 12'b0) + phase_l + phase_step_l);
-	phase_step_hp <= phase_step_h;
-	reset_d <= reset;
-	phase_h <= reset_d ? 20'b0 : (phase_h + phase_step_hp + carry);
-end
 cordicg_b22 #(.nstg(20), .width(18), .def_op(0)) trig(.clk(clk), .opin(2'b00),
-	.xin(lo_amp), .yin(18'd0), .phasein(phase_h[19:1]),
+	.xin(lo_amp), .yin(18'd0), .phasein(phase_acc),
 // 2^17/1.64676 = 79594, use a smaller value to keep CORDIC round-off
 // from overflowing the output
 	.xout(cosa), .yout(sina));
diff --git a/dsp/rules.mk b/dsp/rules.mk
index 3bd71f498..ae8631912 100644
--- a/dsp/rules.mk
+++ b/dsp/rules.mk
@@ -30,7 +30,7 @@ bits: $(BITS_)
 	$(VVP) $< +trace
 	$(PYTHON) $(word 2, $^) -f $*
 
-rot_dds_tb: cordicg_b22.v
+rot_dds_tb: cordicg_b22.v ph_acc.v
 
 mon_12_tb: cordicg_b22.v