artiq/experimental-features/suservo_coherent.diff

diff --git a/artiq/coredevice/ad9910.py b/artiq/coredevice/ad9910.py
index 801b689ca0..bc19afe25c 100644
--- a/artiq/coredevice/ad9910.py
+++ b/artiq/coredevice/ad9910.py
@@ -277,6 +277,10 @@ def read32(self, addr: TInt32) -> TInt32:
 
         :param addr: Register address
         """
+        return self.read32_impl(addr)
+
+    @kernel
+    def read32_impl(self, addr):
         self.bus.set_config_mu(urukul.SPI_CONFIG, 8,
                                urukul.SPIT_DDS_WR, self.chip_select)
         self.bus.write((addr | 0x80) << 24)
@@ -981,7 +985,8 @@ def clear_smp_err(self):
 
     @kernel
     def tune_sync_delay(self,
-                        search_seed: TInt32 = 15) -> TTuple([TInt32, TInt32]):
+                        search_seed: TInt32 = 15,
+                        cpld_channel_idx: TInt32 = -1) -> TTuple([TInt32, TInt32]):
         """Find a stable SYNC_IN delay.
 
         This method first locates a valid SYNC_IN delay at zero validation
@@ -997,6 +1002,9 @@ def tune_sync_delay(self,
             Defaults to 15 (half range).
         :return: Tuple of optimal delay and window size.
         """
+        if cpld_channel_idx == -1:
+            cpld_channel_idx = self.chip_select - 4
+        assert 0 <= cpld_channel_idx < 4, "Invalid channel index"
         if not self.cpld.sync_div:
             raise ValueError("parent cpld does not drive SYNC")
         search_span = 31
@@ -1019,7 +1027,7 @@ def tune_sync_delay(self,
                 delay(100 * us)
                 err = urukul_sta_smp_err(self.cpld.sta_read())
                 delay(100 * us)  # slack
-                if not (err >> (self.chip_select - 4)) & 1:
+                if not (err >> cpld_channel_idx) & 1:
                     next_seed = in_delay
                     break
             if next_seed >= 0:  # valid delay found, scan next window
diff --git a/artiq/coredevice/suservo.py b/artiq/coredevice/suservo.py
index 1d0a72dad1..f7b516a4e7 100644
--- a/artiq/coredevice/suservo.py
+++ b/artiq/coredevice/suservo.py
@@ -1,19 +1,19 @@
 from artiq.language.core import kernel, delay, delay_mu, portable
 from artiq.language.units import us, ns
+from artiq.language import *
 from artiq.coredevice.rtio import rtio_output, rtio_input_data
 from artiq.coredevice import spi2 as spi
-from artiq.coredevice import urukul, sampler
+from artiq.coredevice import urukul, sampler, ad9910
+from math import ceil, log2
+from numpy import int32, int64
 
 
-COEFF_WIDTH = 18
+COEFF_WIDTH = 18  # Must match gateware IIRWidths.coeff
 Y_FULL_SCALE_MU = (1 << (COEFF_WIDTH - 1)) - 1
-COEFF_DEPTH = 10 + 1
-WE = 1 << COEFF_DEPTH + 1
-STATE_SEL = 1 << COEFF_DEPTH
-CONFIG_SEL = 1 << COEFF_DEPTH - 1
-CONFIG_ADDR = CONFIG_SEL | STATE_SEL
 T_CYCLE = (2*(8 + 64) + 2)*8*ns  # Must match gateware Servo.t_cycle.
-COEFF_SHIFT = 11
+COEFF_SHIFT = 11  # Must match gateware IIRWidths.shift
+PROFILE_WIDTH = 5  # Must match gateware IIRWidths.profile
+FINE_TS_WIDTH = 3  # Must match gateware IIRWidths.ioup_dly
 
 
 @portable
@@ -35,21 +35,21 @@ class SUServo:
     """Sampler-Urukul Servo parent and configuration device.
 
     Sampler-Urukul Servo is a integrated device controlling one
-    8-channel ADC (Sampler) and two 4-channel DDS (Urukuls) with a DSP engine
-    connecting the ADC data and the DDS output amplitudes to enable
+    8-channel ADC (Sampler) and any number of 4-channel DDS (Urukuls) with a
+    DSP engine connecting the ADC data and the DDS output amplitudes to enable
     feedback. SU Servo can for example be used to implement intensity
     stabilization of laser beams with an amplifier and AOM driven by Urukul
     and a photodetector connected to Sampler.
 
     Additionally SU Servo supports multiple preconfigured profiles per channel
-    and features like automatic integrator hold.
+    and features like automatic integrator hold and coherent phase tracking.
 
     Notes:
 
         * See the SU Servo variant of the Kasli target for an example of how to
           connect the gateware and the devices. Sampler and each Urukul need
           two EEM connections.
-        * Ensure that both Urukuls are AD9910 variants and have the on-board
+        * Ensure that all Urukuls are AD9910 variants and have the on-board
           dip switches set to 1100 (first two on, last two off).
         * Refer to the Sampler and Urukul documentation and the SU Servo
           example device database for runtime configuration of the devices
@@ -65,7 +65,9 @@ class SUServo:
     :param core_device: Core device name
     """
     kernel_invariants = {"channel", "core", "pgia", "cplds", "ddses",
-                         "ref_period_mu"}
+                         "ref_period_mu", "num_channels", "coeff_sel",
+                         "state_sel", "io_dly_addr", "config_addr",
+                         "write_enable"}
 
     def __init__(self, dmgr, channel, pgia_device,
                  cpld_devices, dds_devices,
@@ -83,9 +85,20 @@ def __init__(self, dmgr, channel, pgia_device,
             self.core.coarse_ref_period)
         assert self.ref_period_mu == self.core.ref_multiplier
 
+        # The width of parts of the servo memory address depends on the number
+        # of channels.
+        self.num_channels = 4 * len(dds_devices)
+        channel_width = ceil(log2(self.num_channels))
+        coeff_depth = PROFILE_WIDTH + channel_width + 3
+        self.io_dly_addr = 1 << (coeff_depth - 2)
+        self.state_sel = 2 << (coeff_depth - 2)
+        self.config_addr = 3 << (coeff_depth - 2)
+        self.coeff_sel = 1 << coeff_depth
+        self.write_enable = 1 << (coeff_depth + 1)
+
     @kernel
     def init(self):
-        """Initialize the servo, Sampler and both Urukuls.
+        """Initialize the servo, Sampler and all Urukuls.
 
         Leaves the servo disabled (see :meth:`set_config`), resets and
         configures all DDS.
@@ -111,8 +124,20 @@ def init(self):
             prev_cpld_cfg = cpld.cfg_reg
             cpld.cfg_write(prev_cpld_cfg | (0xf << urukul.CFG_MASK_NU))
             dds.init(blind=True)
+
+            if dds.sync_data.sync_delay_seed != -1:
+                for channel_idx in range(4):
+                    mask_nu_this = 1 << (urukul.CFG_MASK_NU + channel_idx)
+                    cpld.cfg_write(prev_cpld_cfg | mask_nu_this)
+                    delay(8 * us)
+                    dds.tune_sync_delay(dds.sync_data.sync_delay_seed,
+                                        cpld_channel_idx=channel_idx)
+                    delay(50 * us)
             cpld.cfg_write(prev_cpld_cfg)
 
+        self.set_io_update_delays(
+            [dds.sync_data.io_update_delay for dds in self.ddses])
+
     @kernel
     def write(self, addr, value):
         """Write to servo memory.
@@ -122,7 +147,7 @@ def write(self, addr, value):
         :param addr: Memory location address.
         :param value: Data to be written.
         """
-        addr |= WE
+        addr |= self.write_enable
         value &= (1 << COEFF_WIDTH) - 1
         value |= (addr >> 8) << COEFF_WIDTH
         addr = addr & 0xff
@@ -158,7 +183,7 @@ def set_config(self, enable):
             Disabling takes up to two servo cycles (~2.3 µs) to clear the
             processing pipeline.
         """
-        self.write(CONFIG_ADDR, enable)
+        self.write(self.config_addr, enable)
 
     @kernel
     def get_status(self):
@@ -179,7 +204,7 @@ def get_status(self):
         :return: Status. Bit 0: enabled, bit 1: done,
           bits 8-15: channel clip indicators.
         """
-        return self.read(CONFIG_ADDR)
+        return self.read(self.config_addr)
 
     @kernel
     def get_adc_mu(self, adc):
@@ -197,7 +222,8 @@ def get_adc_mu(self, adc):
         # State memory entries are 25 bits. Due to the pre-adder dynamic
         # range, X0/X1/OFFSET are only 24 bits. Finally, the RTIO interface
         # only returns the 18 MSBs (the width of the coefficient memory).
-        return self.read(STATE_SEL | (adc << 1) | (1 << 8))
+        return self.read(self.state_sel |
+                         (2 * adc + (1 << PROFILE_WIDTH) * self.num_channels))
 
     @kernel
     def set_pgia_mu(self, channel, gain):
@@ -236,6 +262,18 @@ def get_adc(self, channel):
         gain = (self.gains >> (channel*2)) & 0b11
         return adc_mu_to_volts(val, gain)
 
+    @kernel
+    def set_io_update_delays(self, dlys):
+        """Set IO_UPDATE pulse alignment delays.
+
+        :param dlys: List of delays for each Urukul
+        """
+        bits = 0
+        mask_fine_ts = (1 << FINE_TS_WIDTH) - 1
+        for i in range(len(dlys)):
+            bits |= (dlys[i] & mask_fine_ts) << (FINE_TS_WIDTH * i)
+        self.write(self.io_dly_addr, bits)
+
 
 class Channel:
     """Sampler-Urukul Servo channel
@@ -256,7 +294,7 @@ def __init__(self, dmgr, channel, servo_device):
         self.dds = self.servo.ddses[self.servo_channel // 4]
 
     @kernel
-    def set(self, en_out, en_iir=0, profile=0):
+    def set(self, en_out, en_iir=0, profile=0, en_pt=0):
         """Operate channel.
 
         This method does not advance the timeline. Output RF switch setting
@@ -270,9 +308,26 @@ def set(self, en_out, en_iir=0, profile=0):
         :param en_out: RF switch enable
         :param en_iir: IIR updates enable
         :param profile: Active profile (0-31)
+        :param en_pt: Coherent phase tracking enable
+            * en_pt=1: "coherent phase mode"
+            * en_pt=0: "continuous phase mode"
+            (see :func:`artiq.coredevice.ad9910.AD9910.set_phase_mode` for a
+            definition of the phase modes)
         """
         rtio_output(self.channel << 8,
-                    en_out | (en_iir << 1) | (profile << 2))
+                    en_out | (en_iir << 1) | (en_pt << 2) | (profile << 3))
+
+    @kernel
+    def set_reference_time(self):
+        """Set reference time for "coherent phase mode" (see :meth:`set`).
+
+        This method does not advance the timeline.
+        With en_pt=1 (see :meth:`set`), the tracked DDS output phase of
+        this channel will refer to the current timeline position.
+
+        """
+        fine_ts = now_mu() & ((1 << FINE_TS_WIDTH) - 1)
+        rtio_output(self.channel << 8 | 1, self.dds.sysclk_per_mu * fine_ts)
 
     @kernel
     def set_dds_mu(self, profile, ftw, offs, pow_=0):
@@ -285,10 +340,11 @@ def set_dds_mu(self, profile, ftw, offs, pow_=0):
         :param offs: IIR offset (17 bit signed)
         :param pow_: Phase offset word (16 bit)
         """
-        base = (self.servo_channel << 8) | (profile << 3)
+        base = self.servo.coeff_sel | (self.servo_channel <<
+                                       (3 + PROFILE_WIDTH)) | (profile << 3)
         self.servo.write(base + 0, ftw >> 16)
         self.servo.write(base + 6, (ftw & 0xffff))
-        self.set_dds_offset_mu(profile, offs)
+        self.servo.write(base + 4, offs)
         self.servo.write(base + 2, pow_)
 
     @kernel
@@ -319,7 +375,8 @@ def set_dds_offset_mu(self, profile, offs):
         :param profile: Profile number (0-31)
         :param offs: IIR offset (17 bit signed)
         """
-        base = (self.servo_channel << 8) | (profile << 3)
+        base = self.servo.coeff_sel | (self.servo_channel <<
+                                       (3 + PROFILE_WIDTH)) | (profile << 3)
         self.servo.write(base + 4, offs)
 
     @kernel
@@ -344,6 +401,30 @@ def dds_offset_to_mu(self, offset):
         """
         return int(round(offset * (1 << COEFF_WIDTH - 1)))
 
+    @kernel
+    def set_dds_phase_mu(self, profile, pow_):
+        """Set only POW in profile DDS coefficients.
+
+        See :meth:`set_dds_mu` for setting the complete DDS profile.
+
+        :param profile: Profile number (0-31)
+        :param pow_: Phase offset word (16 bit)
+        """
+        base = self.servo.coeff_sel | (self.servo_channel <<
+                                       (3 + PROFILE_WIDTH)) | (profile << 3)
+        self.servo.write(base + 2, pow_)
+
+    @kernel
+    def set_dds_phase(self, profile, phase):
+        """Set only phase in profile DDS coefficients.
+
+        See :meth:`set_dds` for setting the complete DDS profile.
+
+        :param profile: Profile number (0-31)
+        :param phase: DDS phase in turns
+        """
+        self.set_dds_phase_mu(profile, self.dds.turns_to_pow(phase))
+
     @kernel
     def set_iir_mu(self, profile, adc, a1, b0, b1, dly=0):
         """Set profile IIR coefficients in machine units.
@@ -378,7 +459,8 @@ def set_iir_mu(self, profile, adc, a1, b0, b1, dly=0):
         :param dly: IIR update suppression time. In units of IIR cycles
             (~1.2 µs, 0-255).
         """
-        base = (self.servo_channel << 8) | (profile << 3)
+        base = self.servo.coeff_sel | (self.servo_channel <<
+                                       (3 + PROFILE_WIDTH)) | (profile << 3)
         self.servo.write(base + 3, adc | (dly << 8))
         self.servo.write(base + 1, b1)
         self.servo.write(base + 5, a1)
@@ -470,7 +552,9 @@ def get_profile_mu(self, profile, data):
         :param profile: Profile number (0-31)
         :param data: List of 8 integers to write the profile data into
         """
-        base = (self.servo_channel << 8) | (profile << 3)
+        assert len(data) == 8
+        base = self.servo.coeff_sel | (self.servo_channel <<
+                                       (3 + PROFILE_WIDTH)) | (profile << 3)
         for i in range(len(data)):
             data[i] = self.servo.read(base + i)
             delay(4*us)
@@ -491,7 +575,8 @@ def get_y_mu(self, profile):
         :param profile: Profile number (0-31)
         :return: 17 bit unsigned Y0
         """
-        return self.servo.read(STATE_SEL | (self.servo_channel << 5) | profile)
+        return self.servo.read(self.servo.state_sel | (
+                self.servo_channel << PROFILE_WIDTH) | profile)
 
     @kernel
     def get_y(self, profile):
@@ -529,7 +614,8 @@ def set_y_mu(self, profile, y):
         """
         # State memory is 25 bits wide and signed.
         # Reads interact with the 18 MSBs (coefficient memory width)
-        self.servo.write(STATE_SEL | (self.servo_channel << 5) | profile, y)
+        self.servo.write(self.servo.state_sel | (
+                self.servo_channel << PROFILE_WIDTH) | profile, y)
 
     @kernel
     def set_y(self, profile, y):
@@ -552,3 +638,217 @@ def set_y(self, profile, y):
             raise ValueError("Invalid SUServo y-value!")
         self.set_y_mu(profile, y_mu)
         return y_mu
+
+
+class CPLD(urukul.CPLD):
+    """
+    This module contains a subclass of the Urukul driver class in artiq.coredevice
+    adapted to use CPLD read-back via half-duplex SPI. Only the 8 LSBs can be read
+    back as the read-back buffer on the CPLD is 8 bits wide.
+    """
+
+    def __init__(self, dmgr, spi_device, io_update_device=None,
+                 **kwargs):
+        # Separate IO_UPDATE TTL output device used by SUServo core,
+        # if active, else by artiq.coredevice.suservo.AD9910
+        # :meth:`measure_io_update_alignment`.
+        # The urukul.CPLD driver utilises the CPLD CFG register
+        # option instead for pulsing IO_UPDATE of masked DDSs.
+        self.io_update_ttl = dmgr.get(io_update_device)
+        urukul.CPLD.__init__(self, dmgr, spi_device, **kwargs)
+
+    @kernel
+    def enable_readback(self):
+        """
+        This method sets the RB_EN flag in the Urukul CPLD configuration
+        register. Once set, the CPLD expects an alternating sequence of
+        two SPI transactions:
+
+            * 1: Any transaction. If returning data, the 8 LSBs
+                of that will be stored in the CPLD.
+
+            * 2: One read transaction in half-duplex SPI mode shifting
+                out data from the CPLD over MOSI (use :meth:`readback`).
+
+        To end this protocol, call :meth:`disable_readback` during step 1.
+        """
+        self.cfg_write(self.cfg_reg | (1 << urukul.CFG_RB_EN))
+
+    @kernel
+    def disable_readback(self):
+        """
+        This method clears the RB_EN flag in the Urukul CPLD configuration
+        register. This marks the end of the readback protocol (see
+        :meth:`enable_readback`).
+        """
+        self.cfg_write(self.cfg_reg & ~(1 << urukul.CFG_RB_EN))
+
+    @kernel
+    def sta_read(self, full=False):
+        """
+        Read from status register
+
+        :param full: retrieve status register by concatenating data from
+            several readback transactions.
+        """
+        self.enable_readback()
+        self.sta_read_impl()
+        delay(16 * us)  # slack
+        r = self.readback() << urukul.STA_RF_SW
+        delay(16 * us)  # slack
+        if full:
+            self.enable_readback()  # dummy write
+            r |= self.readback(urukul.CS_RB_PLL_LOCK) << urukul.STA_PLL_LOCK
+            delay(16 * us)  # slack
+            self.enable_readback()  # dummy write
+            r |= self.readback(urukul.CS_RB_PROTO_REV) << urukul.STA_PROTO_REV
+            delay(16 * us)  # slack
+        self.disable_readback()
+        return r
+
+    @kernel
+    def proto_rev_read(self):
+        """Read 8 LSBs of proto_rev"""
+        self.enable_readback()
+        self.enable_readback()  # dummy write
+        r = self.readback(urukul.CS_RB_PROTO_REV)
+        self.disable_readback()
+        return r
+
+    @kernel
+    def pll_lock_read(self):
+        """Read PLL lock status"""
+        self.enable_readback()
+        self.enable_readback()  # dummy write
+        r = self.readback(urukul.CS_RB_PLL_LOCK)
+        self.disable_readback()
+        return r & 0xf
+
+    @kernel
+    def get_att_mu(self):
+        # Different behaviour to urukul.CPLD.get_att_mu: Here, the
+        # latch enable of the attenuators activates 31.5dB
+        # attenuation during the transactions.
+        att_reg = int32(0)
+        self.enable_readback()
+        for i in range(4):
+            self.core.break_realtime()
+            self.bus.set_config_mu(urukul.SPI_CONFIG | spi.SPI_END, 8,
+                                   urukul.SPIT_ATT_RD, urukul.CS_ATT)
+            self.bus.write(0)  # shift in zeros, shift out next 8 bits
+            r = self.readback() & 0xff
+            att_reg |= r << (8 * i)
+
+        delay(16 * us)  # slack
+        self.disable_readback()
+
+        self.att_reg = int32(att_reg)
+        delay(8 * us)  # slack
+        self.set_all_att_mu(self.att_reg)  # shift and latch current value again
+        return self.att_reg
+
+    @kernel
+    def readback(self, cs=urukul.CS_RB_LSBS):
+        """Read from the readback register in half-duplex SPI mode
+        See :meth:`enable_readback` for usage instructions.
+
+        :param cs: Select data to be returned from the readback register.
+             - urukul.CS_RB_LSBS does not modify the readback register upon readback
+             - urukul.CS_RB_PROTO_REV loads the 8 LSBs of proto_rev
+             - urukul.CS_PLL_LOCK loads the PLL lock status bits concatenated with the
+               IFC mode bits
+        :return: CPLD readback register.
+        """
+        self.bus.set_config_mu(
+            urukul.SPI_CONFIG | spi.SPI_END | spi.SPI_INPUT | spi.SPI_HALF_DUPLEX,
+            8, urukul.SPIT_CFG_RD, cs)
+        self.bus.write(0)
+        return int32(self.bus.read())
+
+
+class AD9910(ad9910.AD9910):
+    """
+    This module contains a subclass of the AD9910 driver class in artiq.coredevice
+    using CPLD read-back via half-duplex SPI.
+    """
+
+    # Re-declare set of kernel invariants to avoid warning about non-existent
+    # `sw` attribute, as the AD9910 (instance) constructor writes to the
+    # class attributes.
+    kernel_invariants = {
+        "chip_select", "cpld", "core", "bus", "ftw_per_hz", "sysclk_per_mu"
+    }
+
+    @kernel
+    def read32(self, addr):
+        """ Read from a 32-bit register
+
+        This method returns only the 8 LSBs of the return value.
+        """
+        self.cpld.enable_readback()
+        self.read32_impl(addr)
+        delay(12 * us)  # slack
+        r = self.cpld.readback()
+        delay(12 * us)  # slack
+        self.cpld.disable_readback()
+        return r
+
+    @kernel
+    def read64(self, addr):
+        # 3-wire SPI transactions consisting of multiple transfers are not supported.
+        raise NotImplementedError
+
+    @kernel
+    def read_ram(self, data):
+        # 3-wire SPI transactions consisting of multiple transfers are not supported.
+        raise NotImplementedError
+
+    @kernel
+    def measure_io_update_alignment(self, delay_start, delay_stop):
+        """Use the digital ramp generator to locate the alignment between
+        IO_UPDATE and SYNC_CLK.
+
+        Refer to `artiq.coredevice.ad9910` :meth:`measure_io_update_alignment`.
+        In order that this method can operate the io_update_ttl also used by the SUServo
+        core, deactivate the servo before (see :meth:`set_config`).
+        """
+        # set up DRG
+        self.set_cfr1(drg_load_lrr=1, drg_autoclear=1)
+        # DRG -> FTW, DRG enable
+        self.set_cfr2(drg_enable=1)
+        # no limits
+        self.write64(ad9910._AD9910_REG_RAMP_LIMIT, -1, 0)
+        # DRCTL=0, dt=1 t_SYNC_CLK
+        self.write32(ad9910._AD9910_REG_RAMP_RATE, 0x00010000)
+        # dFTW = 1, (work around negative slope)
+        self.write64(ad9910._AD9910_REG_RAMP_STEP, -1, 0)
+        # un-mask DDS
+        cfg_masked = self.cpld.cfg_reg
+        self.cpld.cfg_write(cfg_masked & ~(0xf << urukul.CFG_MASK_NU))
+        delay(70 * us)  # slack
+        # delay io_update after RTIO edge
+        t = now_mu() + 8 & ~7
+        at_mu(t + delay_start)
+        # assumes a maximum t_SYNC_CLK period
+        self.cpld.io_update_ttl.pulse(self.core.mu_to_seconds(16 - delay_start))  # realign
+        # re-mask DDS
+        self.cpld.cfg_write(cfg_masked)
+        delay(10 * us)  # slack
+        # disable DRG autoclear and LRR on io_update
+        self.set_cfr1()
+        delay(10 * us)  # slack
+        # stop DRG
+        self.write64(ad9910._AD9910_REG_RAMP_STEP, 0, 0)
+        delay(10 * us)  # slack
+        # un-mask DDS
+        self.cpld.cfg_write(cfg_masked & ~(0xf << urukul.CFG_MASK_NU))
+        at_mu(t + 0x20000 + delay_stop)
+        self.cpld.io_update_ttl.pulse_mu(16 - delay_stop)  # realign
+        # re-mask DDS
+        self.cpld.cfg_write(cfg_masked)
+        ftw = self.read32(ad9910._AD9910_REG_FTW)  # read out effective FTW
+        delay(100 * us)  # slack
+        # disable DRG
+        self.set_cfr2(drg_enable=0)
+        self.cpld.io_update.pulse_mu(16)
+        return ftw & 1
diff --git a/artiq/coredevice/urukul.py b/artiq/coredevice/urukul.py
index 2fd66bd65e..61fd476280 100644
--- a/artiq/coredevice/urukul.py
+++ b/artiq/coredevice/urukul.py
@@ -24,6 +24,7 @@
 CFG_RF_SW = 0
 CFG_LED = 4
 CFG_PROFILE = 8
+CFG_RB_EN = 11
 CFG_IO_UPDATE = 12
 CFG_MASK_NU = 13
 CFG_CLK_SEL0 = 17
@@ -51,18 +52,23 @@
 CS_DDS_CH1 = 5
 CS_DDS_CH2 = 6
 CS_DDS_CH3 = 7
+# chip selects for readback
+CS_RB_PROTO_REV = 1
+CS_RB_PLL_LOCK = 2
+CS_RB_LSBS = 3
 
 # Default profile
 DEFAULT_PROFILE = 7
 
 
 @portable
-def urukul_cfg(rf_sw, led, profile, io_update, mask_nu,
+def urukul_cfg(rf_sw, led, profile, rb_en, io_update, mask_nu,
                clk_sel, sync_sel, rst, io_rst, clk_div):
     """Build Urukul CPLD configuration register"""
     return ((rf_sw << CFG_RF_SW) |
             (led << CFG_LED) |
             (profile << CFG_PROFILE) |
+            (rb_en << CFG_RB_EN) |
             (io_update << CFG_IO_UPDATE) |
             (mask_nu << CFG_MASK_NU) |
             ((clk_sel & 0x01) << CFG_CLK_SEL0) |
@@ -191,7 +197,7 @@ def __init__(self, dmgr, spi_device, io_update_device=None,
             assert sync_div is None
             sync_div = 0
 
-        self.cfg_reg = urukul_cfg(rf_sw=rf_sw, led=0, profile=DEFAULT_PROFILE,
+        self.cfg_reg = urukul_cfg(rf_sw=rf_sw, led=0, profile=DEFAULT_PROFILE, rb_en=0,
                                   io_update=0, mask_nu=0, clk_sel=clk_sel,
                                   sync_sel=sync_sel,
                                   rst=0, io_rst=0, clk_div=clk_div)
@@ -226,6 +232,10 @@ def sta_read(self) -> TInt32:
 
         :return: The status register value.
         """
+        return self.sta_read_impl()
+
+    @kernel
+    def sta_read_impl(self):
         self.bus.set_config_mu(SPI_CONFIG | spi.SPI_END | spi.SPI_INPUT, 24,
                                SPIT_CFG_RD, CS_CFG)
         self.bus.write(self.cfg_reg << 8)
diff --git a/artiq/examples/kasli_suservo/device_db.py b/artiq/examples/kasli_suservo/device_db.py
index c52b82a947..8e9d875205 100644
--- a/artiq/examples/kasli_suservo/device_db.py
+++ b/artiq/examples/kasli_suservo/device_db.py
@@ -142,53 +142,66 @@
         "arguments": {"channel": 15},
     },
 
+    "ttl_urukul0_io_update": {
+        "type": "local",
+        "module": "artiq.coredevice.ttl",
+        "class": "TTLOut",
+        "arguments": {"channel": 16}
+    },
+    "ttl_urukul1_io_update": {
+        "type": "local",
+        "module": "artiq.coredevice.ttl",
+        "class": "TTLOut",
+        "arguments": {"channel": 17}
+    },
+
     "suservo0_ch0": {
         "type": "local",
         "module": "artiq.coredevice.suservo",
         "class": "Channel",
-        "arguments": {"channel": 16, "servo_device": "suservo0"}
+        "arguments": {"channel": 18, "servo_device": "suservo0"}
     },
     "suservo0_ch1": {
         "type": "local",
         "module": "artiq.coredevice.suservo",
         "class": "Channel",
-        "arguments": {"channel": 17, "servo_device": "suservo0"}
+        "arguments": {"channel": 19, "servo_device": "suservo0"}
     },
     "suservo0_ch2": {
         "type": "local",
         "module": "artiq.coredevice.suservo",
         "class": "Channel",
-        "arguments": {"channel": 18, "servo_device": "suservo0"}
+        "arguments": {"channel": 20, "servo_device": "suservo0"}
     },
     "suservo0_ch3": {
         "type": "local",
         "module": "artiq.coredevice.suservo",
         "class": "Channel",
-        "arguments": {"channel": 19, "servo_device": "suservo0"}
+        "arguments": {"channel": 21, "servo_device": "suservo0"}
     },
     "suservo0_ch4": {
         "type": "local",
         "module": "artiq.coredevice.suservo",
         "class": "Channel",
-        "arguments": {"channel": 20, "servo_device": "suservo0"}
+        "arguments": {"channel": 22, "servo_device": "suservo0"}
     },
     "suservo0_ch5": {
         "type": "local",
         "module": "artiq.coredevice.suservo",
         "class": "Channel",
-        "arguments": {"channel": 21, "servo_device": "suservo0"}
+        "arguments": {"channel": 23, "servo_device": "suservo0"}
     },
     "suservo0_ch6": {
         "type": "local",
         "module": "artiq.coredevice.suservo",
         "class": "Channel",
-        "arguments": {"channel": 22, "servo_device": "suservo0"}
+        "arguments": {"channel": 24, "servo_device": "suservo0"}
     },
     "suservo0_ch7": {
         "type": "local",
         "module": "artiq.coredevice.suservo",
         "class": "Channel",
-        "arguments": {"channel": 23, "servo_device": "suservo0"}
+        "arguments": {"channel": 25, "servo_device": "suservo0"}
     },
 
     "suservo0": {
@@ -196,7 +209,7 @@
         "module": "artiq.coredevice.suservo",
         "class": "SUServo",
         "arguments": {
-            "channel": 24,
+            "channel": 26,
             "pgia_device": "spi_sampler0_pgia",
             "cpld_devices": ["urukul0_cpld", "urukul1_cpld"],
             "dds_devices": ["urukul0_dds", "urukul1_dds"],
@@ -207,33 +220,37 @@
         "type": "local",
         "module": "artiq.coredevice.spi2",
         "class": "SPIMaster",
-        "arguments": {"channel": 25}
+        "arguments": {"channel": 27}
     },
 
     "spi_urukul0": {
         "type": "local",
         "module": "artiq.coredevice.spi2",
         "class": "SPIMaster",
-        "arguments": {"channel": 26}
+        "arguments": {"channel": 28}
     },
     "urukul0_cpld": {
         "type": "local",
-        "module": "artiq.coredevice.urukul",
+        "module": "artiq.coredevice.suservo",
         "class": "CPLD",
         "arguments": {
             "spi_device": "spi_urukul0",
+            "io_update_device": "ttl_urukul0_io_update",
+            "sync_device": "clkgen_dds_sync_in",
             "refclk": 100e6,
             "clk_sel": 0
         }
     },
     "urukul0_dds": {
         "type": "local",
-        "module": "artiq.coredevice.ad9910",
+        "module": "artiq.coredevice.suservo",
         "class": "AD9910",
         "arguments": {
             "pll_n": 40,
             "chip_select": 3,
             "cpld_device": "urukul0_cpld",
+            "io_update_delay": 0,
+            "sync_delay_seed": -1,
         }
     },
 
@@ -241,26 +258,40 @@
         "type": "local",
         "module": "artiq.coredevice.spi2",
         "class": "SPIMaster",
-        "arguments": {"channel": 27}
+        "arguments": {"channel": 29}
     },
     "urukul1_cpld": {
         "type": "local",
-        "module": "artiq.coredevice.urukul",
+        "module": "artiq.coredevice.suservo",
         "class": "CPLD",
         "arguments": {
             "spi_device": "spi_urukul1",
+            "io_update_device": "ttl_urukul1_io_update",
+            "sync_device": "clkgen_dds_sync_in",
             "refclk": 100e6,
             "clk_sel": 0
         }
     },
     "urukul1_dds": {
         "type": "local",
-        "module": "artiq.coredevice.ad9910",
+        "module": "artiq.coredevice.suservo",
         "class": "AD9910",
         "arguments": {
             "pll_n": 40,
             "chip_select": 3,
             "cpld_device": "urukul1_cpld",
+            "io_update_delay": 0,
+            "sync_delay_seed": -1,
+        }
+    },
+
+    "clkgen_dds_sync_in": {
+        "type": "local",
+        "module": "artiq.coredevice.ttl",
+        "class": "TTLClockGen",
+        "arguments": {
+            "channel": 30,
+            "acc_width": 4
         }
     },
 
diff --git a/artiq/frontend/artiq_ddb_template.py b/artiq/frontend/artiq_ddb_template.py
index b6d9294a37..93a74d46e7 100755
--- a/artiq/frontend/artiq_ddb_template.py
+++ b/artiq/frontend/artiq_ddb_template.py
@@ -424,6 +424,16 @@ def process_suservo(self, rtio_offset, peripheral):
         sampler_name = self.get_name("sampler")
         urukul_names = [self.get_name("urukul") for _ in range(2)]
         channel = count(0)
+        for urukul_name in urukul_names:
+            self.gen("""
+                device_db["ttl_{urukul_name}_io_update"] = {{
+                    "type": "local",
+                    "module": "artiq.coredevice.ttl",
+                    "class": "TTLOut",
+                    "arguments": {{"channel": 0x{ttl_channel:06x}}}
+                }}""",
+                urukul_name=urukul_name,
+                ttl_channel=rtio_offset+next(channel))
         for i in range(8):
             self.gen("""
                 device_db["{suservo_name}_ch{suservo_chn}"] = {{
@@ -472,17 +482,19 @@ def process_suservo(self, rtio_offset, peripheral):
                 }}
                 device_db["{urukul_name}_cpld"] = {{
                     "type": "local",
-                    "module": "artiq.coredevice.urukul",
+                    "module": "artiq.coredevice.suservo",
                     "class": "CPLD",
                     "arguments": {{
                         "spi_device": "spi_{urukul_name}",
+                        "io_update_device": "ttl_{urukul_name}_io_update",
+                        "sync_device": "clkgen_{suservo_name}_dds_sync_in",
                         "refclk": {refclk},
                         "clk_sel": {clk_sel}
                     }}
                 }}
                 device_db["{urukul_name}_dds"] = {{
                     "type": "local",
-                    "module": "artiq.coredevice.ad9910",
+                    "module": "artiq.coredevice.suservo",
                     "class": "AD9910",
                     "arguments": {{
                         "pll_n": {pll_n},
@@ -490,12 +502,25 @@ def process_suservo(self, rtio_offset, peripheral):
                         "cpld_device": "{urukul_name}_cpld"{pll_vco}
                     }}
                 }}""",
+                suservo_name=suservo_name,
                 urukul_name=urukul_name,
                 urukul_channel=rtio_offset+next(channel),
                 refclk=peripheral.get("refclk", self.master_description["rtio_frequency"]),
                 clk_sel=peripheral["clk_sel"],
                 pll_vco=",\n        \"pll_vco\": {}".format(pll_vco) if pll_vco is not None else "",
                 pll_n=peripheral["pll_n"])
+        self.gen("""
+            device_db["clkgen_{suservo_name}_dds_sync_in"] = {{
+                "type": "local",
+                "module": "artiq.coredevice.ttl",
+                "class": "TTLClockGen",
+                "arguments": {{
+                    "channel": 0x{clkgen_channel:06x},
+                    "acc_width": 4
+                }}
+            }}""",
+            suservo_name=suservo_name,
+            clkgen_channel=rtio_offset+next(channel))
         return next(channel)
 
     def process_zotino(self, rtio_offset, peripheral):
diff --git a/artiq/gateware/eem.py b/artiq/gateware/eem.py
index 467f3cae2e..c7ce7c5879 100644
--- a/artiq/gateware/eem.py
+++ b/artiq/gateware/eem.py
@@ -6,6 +6,7 @@
 from artiq.gateware.rtio.phy import spi2, ad53xx_monitor, dds, grabber
 from artiq.gateware.suservo import servo, pads as servo_pads
 from artiq.gateware.rtio.phy import servo as rtservo, fastino, phaser
+from artiq.gateware.rtio.phy import ttl_simple
 
 
 def _eem_signal(i):
@@ -536,17 +537,17 @@ def add_std(cls, target, eem, eem_aux=None, eem_aux2=None, ttl_out_cls=None,
 class SUServo(_EEM):
     @staticmethod
     def io(*eems, iostandard):
-        assert len(eems) in (4, 6)
-        io = (Sampler.io(*eems[0:2], iostandard=iostandard)
-                + Urukul.io_qspi(*eems[2:4], iostandard=iostandard))
-        if len(eems) == 6:  # two Urukuls
-            io += Urukul.io_qspi(*eems[4:6], iostandard=iostandard)
+        assert len(eems) >= 4 and len(eems) % 2 == 0
+        io = Sampler.io(*eems[0:2], iostandard=iostandard)
+        for i in range(len(eems) // 2 - 1):
+            io += Urukul.io_qspi(*eems[(2 * i + 2):(2 * i + 4)], iostandard=iostandard)
         return io
 
     @classmethod
     def add_std(cls, target, eems_sampler, eems_urukul,
                 t_rtt=4, clk=1, shift=11, profile=5,
-                iostandard=default_iostandard):
+                sync_gen_cls=ttl_simple.ClockGen,
+                iostandard=default_iostandard, sysclk_per_clk=8):
         """Add a 8-channel Sampler-Urukul Servo
 
         :param t_rtt: upper estimate for clock round-trip propagation time from
@@ -562,6 +563,8 @@ def add_std(cls, target, eems_sampler, eems_urukul,
             (default: 11)
         :param profile: log2 of the number of profiles for each DDS channel
             (default: 5)
+        :param sysclk_per_clk: DDS "sysclk" (4*refclk = 1GHz typ.) cycles per
+            FPGA "sys" clock (125MHz typ.) cycles (default: 8)
         """
         cls.add_extension(
             target, *(eems_sampler + sum(eems_urukul, [])),
@@ -573,27 +576,29 @@ def add_std(cls, target, eems_sampler, eems_urukul,
         urukul_pads = servo_pads.UrukulPads(
             target.platform, *eem_urukul)
         target.submodules += sampler_pads, urukul_pads
+        target.rtio_channels.extend(
+            rtio.Channel.from_phy(phy) for phy in urukul_pads.io_update_phys)
         # timings in units of RTIO coarse period
         adc_p = servo.ADCParams(width=16, channels=8, lanes=4, t_cnvh=4,
                                 # account for SCK DDR to CONV latency
                                 # difference (4 cycles measured)
                                 t_conv=57 - 4, t_rtt=t_rtt + 4)
         iir_p = servo.IIRWidths(state=25, coeff=18, adc=16, asf=14, word=16,
-                                accu=48, shift=shift, channel=3,
-                                profile=profile, dly=8)
-        dds_p = servo.DDSParams(width=8 + 32 + 16 + 16,
-                                channels=adc_p.channels, clk=clk)
+                                accu=48, shift=shift, profile=profile, dly=8)
+        dds_p = servo.DDSParams(width=8 + 32 + 16 + 16, sysclk_per_clk=sysclk_per_clk,
+                                channels=4*len(eem_urukul), clk=clk)
         su = servo.Servo(sampler_pads, urukul_pads, adc_p, iir_p, dds_p)
         su = ClockDomainsRenamer("rio_phy")(su)
         # explicitly name the servo submodule to enable the migen namer to derive
         # a name for the adc return clock domain
         setattr(target.submodules, "suservo_eem{}".format(eems_sampler[0]), su)
 
-        ctrls = [rtservo.RTServoCtrl(ctrl) for ctrl in su.iir.ctrl]
+        ctrls = [rtservo.RTServoCtrl(ctrl, ctrl_reftime)
+                 for ctrl, ctrl_reftime in zip(su.iir.ctrl, su.iir.ctrl_reftime)]
         target.submodules += ctrls
         target.rtio_channels.extend(
             rtio.Channel.from_phy(ctrl) for ctrl in ctrls)
-        mem = rtservo.RTServoMem(iir_p, su)
+        mem = rtservo.RTServoMem(iir_p, su, urukul_pads.io_update_phys)
         target.submodules += mem
         target.rtio_channels.append(rtio.Channel.from_phy(mem, ififo_depth=4))
 
@@ -603,27 +608,24 @@ def add_std(cls, target, eems_sampler, eems_urukul,
         target.submodules += phy
         target.rtio_channels.append(rtio.Channel.from_phy(phy, ififo_depth=4))
 
-        for i in range(2):
-            if len(eem_urukul) > i:
-                spi_p, spi_n = (
-                    target.platform.request("{}_spi_p".format(eem_urukul[i])),
-                    target.platform.request("{}_spi_n".format(eem_urukul[i])))
-            else:  # create a dummy bus
-                spi_p = Record([("clk", 1), ("cs_n", 1)])  # mosi, cs_n
-                spi_n = None
-
+        for eem_urukuli in eem_urukul:
+            spi_p, spi_n = (
+                target.platform.request("{}_spi_p".format(eem_urukuli)),
+                target.platform.request("{}_spi_n".format(eem_urukuli)))
             phy = spi2.SPIMaster(spi_p, spi_n)
             target.submodules += phy
             target.rtio_channels.append(rtio.Channel.from_phy(phy, ififo_depth=4))
 
-        for j, eem_urukuli in enumerate(eem_urukul):
-            pads = target.platform.request("{}_dds_reset_sync_in".format(eem_urukuli))
-            target.specials += DifferentialOutput(0, pads.p, pads.n)
+        if sync_gen_cls is not None:  # AD9910 variant and SYNC_IN from EEM
+            phy = sync_gen_cls(urukul_pads.dds_reset_sync_in, ftw_width=4)
+            target.submodules += phy
+            target.rtio_channels.append(rtio.Channel.from_phy(phy))
 
+        for j, eem_urukuli in enumerate(eem_urukul):
             for i, signal in enumerate("sw0 sw1 sw2 sw3".split()):
                 pads = target.platform.request("{}_{}".format(eem_urukuli, signal))
                 target.specials += DifferentialOutput(
-                    su.iir.ctrl[j*4 + i].en_out, pads.p, pads.n)
+                    su.iir.ctrl[j * 4 + i].en_out, pads.p, pads.n)
 
 
 class Mirny(_EEM):
diff --git a/artiq/gateware/rtio/phy/servo.py b/artiq/gateware/rtio/phy/servo.py
index 9fa6345211..0f7ebbf4b2 100644
--- a/artiq/gateware/rtio/phy/servo.py
+++ b/artiq/gateware/rtio/phy/servo.py
@@ -1,25 +1,32 @@
 from migen import *
-
 from artiq.gateware.rtio import rtlink
 
 
 class RTServoCtrl(Module):
     """Per channel RTIO control interface"""
-    def __init__(self, ctrl):
+    def __init__(self, ctrl, ctrl_reftime):
         self.rtlink = rtlink.Interface(
-            rtlink.OInterface(len(ctrl.profile) + 2))
+            rtlink.OInterface(
+                data_width=max(len(ctrl.profile) + 3,
+                               len(ctrl_reftime.sysclks_fine)),
+                address_width=1)
+            )
 
         # # #
 
+        sel_ref = self.rtlink.o.address[0]
         self.comb += [
-                ctrl.stb.eq(self.rtlink.o.stb),
-                self.rtlink.o.busy.eq(0)
+                ctrl.stb.eq(self.rtlink.o.stb & ~sel_ref),
+                self.rtlink.o.busy.eq(0),
+                ctrl_reftime.stb.eq(self.rtlink.o.stb & sel_ref),
         ]
+        ctrl_cases = {
+            0: Cat(ctrl.en_out, ctrl.en_iir, ctrl.en_pt, ctrl.profile).eq(
+                            self.rtlink.o.data),
+            1: ctrl_reftime.sysclks_fine.eq(self.rtlink.o.data),
+        }
         self.sync.rio_phy += [
-                If(self.rtlink.o.stb,
-                    Cat(ctrl.en_out, ctrl.en_iir, ctrl.profile).eq(
-                            self.rtlink.o.data)
-                )
+                If(self.rtlink.o.stb, Case(self.rtlink.o.address, ctrl_cases))
         ]
 
 
@@ -34,35 +41,45 @@ class RTServoMem(Module):
     """All-channel all-profile coefficient and state RTIO control
     interface.
 
+    The real-time interface exposes the following functions:
+      1. enable/disable servo iterations
+      2. read the servo status (including state of clip register)
+      3. access the IIR coefficient memory (set PI loop gains etc.)
+      4. access the IIR state memory (set offset and read ADC data)
+
+    The bit assignments for the servo address space are (from MSB):
+      * write-enable (1 bit)
+      * sel_coeff (1 bit)
+        If selected, the coefficient memory location is
+        addressed by all the lower bits excluding the LSB (high_coeff).
+          - high_coeff (1 bit) selects between the upper and lower halves of that
+            memory location.
+        Else (if ~sel_coeff), the following bits are:
+          - sel (2 bits) selects between the following memory locations:
+
+                 destination    |  sel  |  sel_coeff   |
+                ----------------|-------|--------------|
+                 IIR coeff mem  |   -   |       1      |
+                 DDS delay mem  |   1   |       0      |
+                 IIR state mem  |   2   |       0      |
+                 config (write) |   3   |       0      |
+                 status (read)  |   3   |       0      |
+
+          - IIR state memory address
+
     Servo internal addresses are internal_address_width wide, which is
     typically longer than the 8-bit RIO address space. We pack the overflow
     onto the RTIO data word after the data.
 
-    Servo address space (from LSB):
-      - IIR coefficient/state memory address, (w.profile + w.channel + 2) bits.
-        If the state memory is selected, the lower bits are used directly as
-        the memory address. If the coefficient memory is selected, the LSB
-        (high_coeff) selects between the upper and lower halves of the memory
-        location, which is two coefficients wide, with the remaining bits used
-        as the memory address.
-      - config_sel (1 bit)
-      - state_sel (1 bit)
-      - we (1 bit)
-
-     destination    | config_sel | state_sel
-    ----------------|------------|----------
-     IIR coeff mem  |    0       |   0
-     IIR coeff mem  |    1       |   0
-     IIR state mem  |    0       |   1
-     config (write) |    1       |   1
-     status (read)  |    1       |   1
+    The address layout reflects the fact that typically, the coefficient memory
+    address is 2 bits wider than the state memory address.
 
     Values returned to the user on the Python side of the RTIO interface are
     32 bit, so we sign-extend all values from w.coeff to that width. This works
     (instead of having to decide whether to sign- or zero-extend per address), as
     all unsigned values are less wide than w.coeff.
     """
-    def __init__(self, w, servo):
+    def __init__(self, w, servo, io_update_phys):
         m_coeff = servo.iir.m_coeff.get_port(write_capable=True,
                 mode=READ_FIRST,
                 we_granularity=w.coeff, clock_domain="rio")
@@ -71,6 +88,7 @@ def __init__(self, w, servo):
                 # mode=READ_FIRST,
                 clock_domain="rio")
         self.specials += m_state, m_coeff
+        w_channel = bits_for(len(servo.iir.dds) - 1)
 
         # just expose the w.coeff (18) MSBs of state
         assert w.state >= w.coeff
@@ -83,7 +101,7 @@ def __init__(self, w, servo):
         assert 8 + w.dly < w.coeff
 
         # coeff, profile, channel, 2 mems, rw
-        internal_address_width = 3 + w.profile + w.channel + 1 + 1
+        internal_address_width = 3 + w.profile + w_channel + 1 + 1
         rtlink_address_width = min(8, internal_address_width)
         overflow_address_width = internal_address_width - rtlink_address_width
         self.rtlink = rtlink.Interface(
@@ -99,7 +117,7 @@ def __init__(self, w, servo):
         # # #
 
         config = Signal(w.coeff, reset=0)
-        status = Signal(w.coeff)
+        status = Signal(len(self.rtlink.i.data))
         pad = Signal(6)
         self.comb += [
                 Cat(servo.start).eq(config),
@@ -109,15 +127,19 @@ def __init__(self, w, servo):
 
         assert len(self.rtlink.o.address) + len(self.rtlink.o.data) - w.coeff == (
                 1 +  # we
-                1 +  # state_sel
+                1 +  # sel_coeff
                 1 +  # high_coeff
                 len(m_coeff.adr))
-        # ensure that we can fit config/status into the state address space
+        # ensure that we can fit config/io_dly/status into the state address space
         assert len(self.rtlink.o.address) + len(self.rtlink.o.data) - w.coeff >= (
                 1 +  # we
-                1 +  # state_sel
-                1 +  # config_sel
+                1 +  # sel_coeff
+                2 +  # sel
                 len(m_state.adr))
+        # ensure that IIR state mem addresses are at least 2 bits less wide than
+        # IIR coeff mem addresses to ensure we can fit SEL after the state mem
+        # address and before the SEL_COEFF bit.
+        assert w.profile + w_channel >= 4
 
         internal_address = Signal(internal_address_width)
         self.comb += internal_address.eq(Cat(self.rtlink.o.address,
@@ -127,52 +149,60 @@ def __init__(self, w, servo):
         self.comb += coeff_data.eq(self.rtlink.o.data[:w.coeff])
 
         we = internal_address[-1]
-        state_sel = internal_address[-2]
-        config_sel = internal_address[-3]
+        sel_coeff = internal_address[-2]
+        sel1 = internal_address[-3]
+        sel0 = internal_address[-4]
         high_coeff = internal_address[0]
+        sel = Signal(2)
         self.comb += [
                 self.rtlink.o.busy.eq(0),
+                sel.eq(Mux(sel_coeff, 0, Cat(sel0, sel1))),
                 m_coeff.adr.eq(internal_address[1:]),
                 m_coeff.dat_w.eq(Cat(coeff_data, coeff_data)),
-                m_coeff.we[0].eq(self.rtlink.o.stb & ~high_coeff &
-                    we & ~state_sel),
-                m_coeff.we[1].eq(self.rtlink.o.stb & high_coeff &
-                    we & ~state_sel),
+                m_coeff.we[0].eq(self.rtlink.o.stb & ~high_coeff & we & sel_coeff),
+                m_coeff.we[1].eq(self.rtlink.o.stb & high_coeff & we & sel_coeff),
                 m_state.adr.eq(internal_address),
                 m_state.dat_w[w.state - w.coeff:].eq(self.rtlink.o.data),
-                m_state.we.eq(self.rtlink.o.stb & we & state_sel & ~config_sel),
+                m_state.we.eq(self.rtlink.o.stb & we & (sel == 2)),
         ]
         read = Signal()
-        read_state = Signal()
         read_high = Signal()
-        read_config = Signal()
+        read_sel = Signal(2)
         self.sync.rio += [
                 If(read,
                     read.eq(0)
                 ),
                 If(self.rtlink.o.stb,
                     read.eq(~we),
-                    read_state.eq(state_sel),
+                    read_sel.eq(sel),
                     read_high.eq(high_coeff),
-                    read_config.eq(config_sel),
                 )
         ]
+
+        # I/O update alignment delays
+        ioup_dlys = Cat(*[phy.fine_ts for phy in io_update_phys])
+        assert w.coeff >= len(ioup_dlys)
+
         self.sync.rio_phy += [
-                If(self.rtlink.o.stb & we & state_sel & config_sel,
+                If(self.rtlink.o.stb & we & (sel == 3),
                     config.eq(self.rtlink.o.data)
                 ),
-                If(read & read_config & read_state,
+                If(read & (read_sel == 3),
                     [_.clip.eq(0) for _ in servo.iir.ctrl]
-                )
+                ),
+                If(self.rtlink.o.stb & we & (sel == 1),
+                    ioup_dlys.eq(self.rtlink.o.data)
+                ),
         ]
+
+        # read return value by destination
+        read_acts = Array([
+                Mux(read_high, m_coeff.dat_r[w.coeff:], m_coeff.dat_r[:w.coeff]),
+                ioup_dlys,
+                m_state.dat_r[w.state - w.coeff:],
+                status
+        ])
         self.comb += [
                 self.rtlink.i.stb.eq(read),
-                _eq_sign_extend(self.rtlink.i.data,
-                    Mux(read_state,
-                        Mux(read_config,
-                            status,
-                            m_state.dat_r[w.state - w.coeff:]),
-                        Mux(read_high,
-                            m_coeff.dat_r[w.coeff:],
-                            m_coeff.dat_r[:w.coeff])))
+                _eq_sign_extend(self.rtlink.i.data, read_acts[read_sel]),
         ]
diff --git a/artiq/gateware/suservo/dds_ser.py b/artiq/gateware/suservo/dds_ser.py
index 38d1f6d946..cdccfcc98e 100644
--- a/artiq/gateware/suservo/dds_ser.py
+++ b/artiq/gateware/suservo/dds_ser.py
@@ -1,4 +1,5 @@
 import logging
+from collections import namedtuple
 
 from migen import *
 
@@ -6,11 +7,11 @@
 
 from . import spi
 
-
 logger = logging.getLogger(__name__)
 
-
-DDSParams = spi.SPIParams
+DDSParams = namedtuple("DDSParams", spi.SPIParams._fields + (
+    "sysclk_per_clk",  # DDS_CLK per FPGA system clock
+))
 
 
 class DDS(spi.SPISimple):
diff --git a/artiq/gateware/suservo/iir.py b/artiq/gateware/suservo/iir.py
index 0ec9bfa093..3fad77a6ea 100644
--- a/artiq/gateware/suservo/iir.py
+++ b/artiq/gateware/suservo/iir.py
@@ -1,6 +1,7 @@
 from collections import namedtuple
 import logging
 from migen import *
+from migen.genlib.coding import Encoder
 
 logger = logging.getLogger(__name__)
 
@@ -16,7 +17,6 @@
     "word",     # "word" size to break up DDS profile data (16)
     "asf",      # unsigned amplitude scale factor for DDS (14)
     "shift",    # fixed point scaling coefficient for a1, b0, b1 (log2!) (11)
-    "channel",  # channels (log2!) (3)
     "profile",  # profiles per channel (log2!) (5)
     "dly",      # the activation delay
 ])
@@ -99,14 +99,14 @@ class IIR(Module):
     This module implements a multi-channel IIR (infinite impulse response)
     filter processor optimized for synthesis on FPGAs.
 
-    The module is parametrized by passing a ``IIRWidths()`` object which
-    will be abbreviated W here.
+    The module is parametrized by passing a ``IIRWidths()`` object, and
+    two more objects which will be abbreviated W, W_O and W_I here.
 
-    It reads 1 << W.channels input channels (typically from an ADC)
+    It reads W_I.channels input channels (typically from an ADC)
     and on each iteration processes the data using a first-order IIR filter.
     At the end of the cycle each the output of the filter together with
     additional data (typically frequency tunning word and phase offset word
-    for a DDS) are presented at the 1 << W.channels outputs of the module.
+    for a DDS) are presented at the W_O.channels outputs of the module.
 
     Profile memory
     ==============
@@ -145,10 +145,10 @@ class IIR(Module):
     -------------
 
     The state memory holds all Y1 values (IIR processor outputs) for all
-    profiles of all channels in the lower half (1 << W.profile + W.channel
-    addresses) and the pairs of old and new ADC input values X1, and X0,
-    in the upper half (1 << W.channel addresses). Each memory location is
-    W.state bits wide.
+    profiles of all channels in the lower half (1 << W.profile)*W_O.channels
+    addresses, and the pairs of old and new ADC input values X1, and X0,
+    in the upper half (W_I.channels addresses).
+    Each memory location is W.state bits wide.
 
     Real-time control
     =================
@@ -157,15 +157,16 @@ class IIR(Module):
 
         * The active profile, PROFILE
         * Whether to perform IIR filter iterations, EN_IIR
+        * Whether to track the DDS phase coherently, EN_PT
         * The RF switch state enabling output from the channel, EN_OUT
 
     Delayed IIR processing
     ======================
 
-    The IIR filter iterations on a given channel are only performed all of the
-    following are true:
+    The IIR filter iterations on a given channel are only performed if all of
+    the following are true:
 
-        * PROFILE, EN_IIR, EN_OUT have not been updated in the within the
+        * PROFILE, EN_IIR, EN_OUT have not been updated within the
           last DLY cycles
         * EN_IIR is asserted
         * EN_OUT is asserted
@@ -176,9 +177,8 @@ class IIR(Module):
     Typical design at the DSP level. This does not include the description of
     the pipelining or the overall latency involved.
 
-    IIRWidths(state=25, coeff=18, adc=16,
-        asf=14, word=16, accu=48, shift=11,
-        channel=3, profile=5, dly=8)
+    IIRWidths(state=25, coeff=18, adc=16, asf=14,
+        word=16, accu=48, shift=11, profile=5, dly=8)
 
     X0 = ADC * 2^(25 - 1 - 16)
     X1 = X0 delayed by one cycle
@@ -213,39 +213,64 @@ class IIR(Module):
     --/--: signal with a given bit width always includes a sign bit
     -->--: flow is to the right and down unless otherwise indicated
     """
-    def __init__(self, w):
-        self.widths = w
-        for i, j in enumerate(w):
-            assert j > 0, (i, j, w)
+    def __init__(self, w, w_i, w_o, t_cycle):
+        for v in (w, w_i, w_o):
+            for i, j in enumerate(v):
+                assert j > 0, (i, j, v)
         assert w.word <= w.coeff  # same memory
         assert w.state + w.coeff + 3 <= w.accu
 
+        # Reference counter for coherent phase tracking (we assume this doesn't
+        # roll over – a good assumption, as the period is, for a typical clock
+        # frequency, 2^48 / 125 MHz = ~26 days).
+        self.t_running = Signal(48, reset_less=True)
+
+        # If true, internal DDS phase tracking state is reset, matching DDS
+        # chips with phase cleared (and zero FTW) before the start of the
+        # iteration. Automatically reset at the end of the iteration.
+        self.reset_dds_phase = Signal()
+
         # m_coeff of active profiles should only be accessed externally during
         # ~processing
         self.specials.m_coeff = Memory(
                 width=2*w.coeff,  # Cat(pow/ftw/offset, cfg/a/b)
-                depth=4 << w.profile + w.channel)
+                depth=(4 << w.profile) * w_o.channels)
         # m_state[x] should only be read externally during ~(shifting | loading)
         # m_state[y] of active profiles should only be read externally during
         # ~processing
         self.specials.m_state = Memory(
                 width=w.state,  # y1,x0,x1
-                depth=(1 << w.profile + w.channel) + (2 << w.channel))
+                depth=(1 << w.profile) * w_o.channels + 2 * w_i.channels)
         # ctrl should only be updated synchronously
         self.ctrl = [Record([
                 ("profile", w.profile),
                 ("en_out", 1),
                 ("en_iir", 1),
+                ("en_pt", 1),
                 ("clip", 1),
                 ("stb", 1)])
-                for i in range(1 << w.channel)]
+                for i in range(w_o.channels)]
+        # "Shadow copy" of phase accumulator in DDS accumulator for each output
+        # channel.
+        self.specials.m_accum_ftw = Memory(
+                width=2 * w.word,
+                depth=w_o.channels)
+        # ctrl_reftime should only be updated synchronously
+        self.ctrl_reftime = [Record([
+                ("sysclks_fine", bits_for(w_o.sysclk_per_clk - 1)),
+                ("stb", 1)])
+                for i in range(w_o.channels)]
+        # Reference time for each output channel.
+        self.specials.m_t_ref = Memory(
+                width=len(self.t_running),
+                depth=w_o.channels)
         # only update during ~loading
         self.adc = [Signal((w.adc, True), reset_less=True)
-                for i in range(1 << w.channel)]
+                for i in range(w_i.channels)]
         # Cat(ftw0, ftw1, pow, asf)
         # only read externally during ~processing
-        self.dds = [Signal(4*w.word, reset_less=True)
-                for i in range(1 << w.channel)]
+        self.dds = [Signal(4 * w.word, reset_less=True)
+                for i in range(w_o.channels)]
         # perform one IIR iteration, start with loading,
         # then processing, then shifting, end with done
         self.start = Signal()
@@ -265,8 +290,15 @@ def __init__(self, w):
         profiles = Array([ch.profile for ch in self.ctrl])
         en_outs = Array([ch.en_out for ch in self.ctrl])
         en_iirs = Array([ch.en_iir for ch in self.ctrl])
+        en_pts = Array([ch.en_pt for ch in self.ctrl])
         clips = Array([ch.clip for ch in self.ctrl])
 
+        # Sample of the reference counter at the start of the current iteration,
+        # such that a common reference time is used for phase calculations
+        # across all channels, in DDS sysclk units.
+        sysclks_to_iter_start = Signal(
+            len(self.t_running) + bits_for(w_o.sysclk_per_clk - 1))
+
         # Main state machine sequencing the steps of each servo iteration. The
         # module IDLEs until self.start is asserted, and then runs through LOAD,
         # PROCESS and SHIFT in order (see description of corresponding flags
@@ -281,7 +313,7 @@ def __init__(self, w):
         # using the (MSBs of) t_current_step, and, after all channels have been
         # covered, proceed once the pipeline has completely drained.
         self.submodules.fsm = fsm = FSM("IDLE")
-        t_current_step = Signal(w.channel + 2)
+        t_current_step = Signal(max=max(4 * (w_o.channels + 2), 2 * w_i.channels))
         t_current_step_clr = Signal()
 
         # pipeline group activity flags (SR)
@@ -293,12 +325,13 @@ def __init__(self, w):
                 self.done.eq(1),
                 t_current_step_clr.eq(1),
                 If(self.start,
+                    NextValue(sysclks_to_iter_start, self.t_running * w_o.sysclk_per_clk),
                     NextState("LOAD")
                 )
         )
         fsm.act("LOAD",
                 self.loading.eq(1),
-                If(t_current_step == (1 << w.channel) - 1,
+                If(t_current_step == w_i.channels - 1,
                     t_current_step_clr.eq(1),
                     NextValue(stages_active[0], 1),
                     NextState("PROCESS")
@@ -311,11 +344,12 @@ def __init__(self, w):
                 If(stages_active == 0,
                     t_current_step_clr.eq(1),
                     NextState("SHIFT"),
+                    NextValue(self.reset_dds_phase, 0)
                 )
         )
         fsm.act("SHIFT",
                 self.shifting.eq(1),
-                If(t_current_step == (2 << w.channel) - 1,
+                If(t_current_step == 2 * w_i.channels - 1,
                     NextState("IDLE")
                 )
         )
@@ -333,13 +367,13 @@ def __init__(self, w):
         # pipeline group channel pointer (SR)
         # for each pipeline stage, this is the channel currently being
         # processed
-        channel = [Signal(w.channel, reset_less=True) for i in range(3)]
+        channel = [Signal(max=w_o.channels, reset_less=True) for i in range(3)]
         self.comb += Cat(pipeline_phase, channel[0]).eq(t_current_step)
         self.sync += [
             If(pipeline_phase == 3,
                 Cat(channel[1:]).eq(Cat(channel[:-1])),
                 stages_active[1:].eq(stages_active[:-1]),
-                If(channel[0] == (1 << w.channel) - 1,
+                If(channel[0] == w_o.channels - 1,
                     stages_active[0].eq(0)
                 )
             )
@@ -393,13 +427,13 @@ def __init__(self, w):
 
         # selected adc and profile delay (combinatorial from dat_r)
         # both share the same coeff word (sel in the lower 8 bits)
-        sel_profile = Signal(w.channel)
+        sel_profile = Signal(max=w_i.channels)
         dly_profile = Signal(w.dly)
-        assert w.channel <= 8
+        assert w_o.channels < (1 << 8)
         assert 8 + w.dly <= w.coeff
 
         # latched adc selection
-        sel = Signal(w.channel, reset_less=True)
+        sel = Signal(max=w_i.channels, reset_less=True)
         # iir enable SR
         en = Signal(2, reset_less=True)
 
@@ -407,12 +441,12 @@ def __init__(self, w):
                 sel_profile.eq(m_coeff.dat_r[w.coeff:]),
                 dly_profile.eq(m_coeff.dat_r[w.coeff + 8:]),
                 If(self.shifting,
-                    m_state.adr.eq(t_current_step | (1 << w.profile + w.channel)),
+                    m_state.adr.eq(t_current_step + (1 << w.profile) * w_o.channels),
                     m_state.dat_w.eq(m_state.dat_r),
                     m_state.we.eq(t_current_step[0])
                 ),
                 If(self.loading,
-                    m_state.adr.eq((t_current_step << 1) | (1 << w.profile + w.channel)),
+                    m_state.adr.eq((t_current_step << 1) + (1 << w.profile) * w_o.channels),
                     m_state.dat_w[-w.adc - 1:-1].eq(Array(self.adc)[t_current_step]),
                     m_state.dat_w[-1].eq(m_state.dat_w[-2]),
                     m_state.we.eq(1)
@@ -424,9 +458,9 @@ def __init__(self, w):
                         # read old y
                         Cat(profile[0], channel[0]),
                         # read x0 (recent)
-                        0 | (sel_profile << 1) | (1 << w.profile + w.channel),
+                        0 | (sel_profile << 1) + (1 << w.profile) * w_o.channels,
                         # read x1 (old)
-                        1 | (sel << 1) | (1 << w.profile + w.channel),
+                        1 | (sel << 1) + (1 << w.profile) * w_o.channels,
                     ])[pipeline_phase]),
                     m_state.dat_w.eq(dsp.output),
                     m_state.we.eq((pipeline_phase == 0) & stages_active[2] & en[1]),
@@ -438,11 +472,9 @@ def __init__(self, w):
         #
 
         # internal channel delay counters
-        dlys = Array([Signal(w.dly)
-            for i in range(1 << w.channel)])
-        self._dlys = dlys  # expose for debugging only
+        dlys = Array([Signal(w.dly) for i in range(w_o.channels)])
 
-        for i in range(1 << w.channel):
+        for i in range(w_o.channels):
             self.sync += [
                     # (profile != profile_old) | ~en_out
                     If(self.ctrl[i].stb,
@@ -482,25 +514,81 @@ def __init__(self, w):
             }),
         ]
 
+        # Update coarse reference time from t_running upon ctrl_reftime strobe
+        ref_stb_encoder = Encoder(w_o.channels)
+        m_t_ref_stb = self.m_t_ref.get_port(write_capable=True)
+        self.specials += m_t_ref_stb
+        self.submodules += ref_stb_encoder
+        self.comb += [
+                ref_stb_encoder.i.eq(Cat([ch.stb for ch in self.ctrl_reftime])),
+                m_t_ref_stb.adr.eq(ref_stb_encoder.o),
+                m_t_ref_stb.we.eq(~ref_stb_encoder.n),
+                m_t_ref_stb.dat_w.eq(self.t_running),
+        ]
+
         #
-        # Update DDS profile with FTW/POW/ASF
-        # Stage 0 loads the POW, stage 1 the FTW, and stage 2 writes
-        # the ASF computed by the IIR filter.
+        # Update DDS profile with FTW/POW/ASF (including phase tracking, if
+        # enabled). Stage 0 loads the POW, stage 1 the FTW, and stage 2 writes
+        # the ASF computed by the IIR filter (and adds any phase correction).
         #
 
         # muxing
         ddss = Array(self.dds)
+        sysclks_ref_fine = Array([ch.sysclks_fine for ch in self.ctrl_reftime])
+
+        # registered copy of FTW on channel[1]
+        current_ftw = Signal(2 * w.word, reset_less=True)
+        # target effective DDS phase (accumulator + POW) at the coming io_update
+        target_dds_phase = Signal.like(current_ftw)
+        # DDS-internal phase accumulated until the coming io_update
+        accum_dds_phase = Signal.like(current_ftw)
+        # correction to add to the bare POW to yield a phase-coherent DDS output
+        correcting_pow = Signal(w.word, reset_less=True)
+        # sum of all FTWs on channel[1], updated with current FTW during the
+        # calculation
+        accum_ftw = Signal.like(current_ftw)
+        # sum of previous FTWs on channel[1] (or 0 on phase coherence reference
+        # reset)
+        prev_accum_ftw = Signal.like(current_ftw)
+        # time since reference time at coming io_update in DDS sysclk units
+        sysclks_to_ref = Signal.like(sysclks_to_iter_start)
+        # t_ref in DDS sysclk units
+        sysclks_ref_to_iter_start = Signal.like(sysclks_to_iter_start)
+
+        m_t_ref = self.m_t_ref.get_port()
+        m_accum_ftw = self.m_accum_ftw.get_port(write_capable=True, mode=READ_FIRST)
+        self.specials += m_accum_ftw, m_t_ref
+        prev_accum_ftw = Signal.like(accum_ftw)
+        self.comb += [
+            prev_accum_ftw.eq(Mux(self.reset_dds_phase, 0, m_accum_ftw.dat_r)),
+            m_accum_ftw.adr.eq(channel[1]),
+            m_accum_ftw.we.eq((pipeline_phase == 3) & stages_active[1]),
+            m_accum_ftw.dat_w.eq(accum_ftw),
+            m_t_ref.adr.eq(channel[0]),
+        ]
 
+        sysclks_per_iter = t_cycle * w_o.sysclk_per_clk
         self.sync += [
             Case(pipeline_phase, {
                 0: [
                     If(stages_active[1],
                         ddss[channel[1]][:w.word].eq(m_coeff.dat_r),  # ftw0
+                        current_ftw[:w.word].eq(m_coeff.dat_r),
+                        sysclks_ref_to_iter_start.eq(m_t_ref.dat_r * w_o.sysclk_per_clk),
+                    ),
+                    If(stages_active[2] & en_pts[channel[2]],
+                        # add pow correction if phase tracking enabled
+                        ddss[channel[2]][2*w.word:3*w.word].eq(
+                            ddss[channel[2]][2*w.word:3*w.word] + correcting_pow),
                     ),
                 ],
                 1: [
                     If(stages_active[1],
                         ddss[channel[1]][w.word:2 * w.word].eq(m_coeff.dat_r),  # ftw1
+                        current_ftw[w.word:].eq(m_coeff.dat_r),
+                        sysclks_to_ref.eq(sysclks_to_iter_start - (
+                            sysclks_ref_to_iter_start + sysclks_ref_fine[channel[1]])),
+                        accum_dds_phase.eq(prev_accum_ftw * sysclks_per_iter),
                     ),
                     If(stages_active[2],
                         ddss[channel[2]][3*w.word:].eq(  # asf
@@ -509,14 +597,40 @@ def __init__(self, w):
                 ],
                 2: [
                     If(stages_active[0],
-                        ddss[channel[0]][2*w.word:3*w.word].eq(m_coeff.dat_r),  # pow
+                        # Load bare POW from profile memory.
+                        ddss[channel[0]][2*w.word:3*w.word].eq(m_coeff.dat_r),
+                    ),
+                    If(stages_active[1],
+                        target_dds_phase.eq(current_ftw * sysclks_to_ref),
+                        accum_ftw.eq(prev_accum_ftw + current_ftw),
                     ),
                 ],
                 3: [
+                    If(stages_active[1],
+                        # Prepare most-significant word to add to POW from
+                        # profile for phase tracking.
+                        correcting_pow.eq(
+                            (target_dds_phase - accum_dds_phase)[w.word:]),
+                    ),
                 ],
             }),
         ]
 
+        # expose for simulation and debugging only
+        self.widths = w
+        self.widths_adc = w_i
+        self.widths_dds = w_o
+        self.t_cycle = t_cycle
+        self._state = t_current_step
+        self._stages = stages_active
+        self._dt_start = sysclks_to_iter_start
+        self._sysclks_to_ref = sysclks_to_ref
+        self._sysclks_ref_to_iter_start = sysclks_ref_to_iter_start
+        self._sysclks_ref_fine = sysclks_ref_fine
+        self._ph_acc = accum_dds_phase
+        self._ph_coh = target_dds_phase
+        self._dlys = dlys
+
     def _coeff(self, channel, profile, coeff):
         """Return ``high_word``, ``address`` and bit ``mask`` for the
         storage of coefficient name ``coeff`` in profile ``profile``
@@ -564,35 +678,45 @@ def get_coeff(self, channel, profile, coeff):
     def set_state(self, channel, val, profile=None, coeff="y1"):
         """Set a state value."""
         w = self.widths
+        w_o = self.widths_dds
         if coeff == "y1":
             assert profile is not None
             yield self.m_state[profile | (channel << w.profile)].eq(val)
         elif coeff == "x0":
             assert profile is None
-            yield self.m_state[(channel << 1) |
-                    (1 << w.profile + w.channel)].eq(val)
+            yield self.m_state[(channel << 1) +
+                    (1 << w.profile) * w_o.channels].eq(val)
         elif coeff == "x1":
             assert profile is None
-            yield self.m_state[1 | (channel << 1) |
-                    (1 << w.profile + w.channel)].eq(val)
+            yield self.m_state[1 | (channel << 1) +
+                    (1 << w.profile) * w_o.channels].eq(val)
         else:
             raise ValueError("no such state", coeff)
 
     def get_state(self, channel, profile=None, coeff="y1"):
         """Get a state value."""
         w = self.widths
+        w_o = self.widths_dds
         if coeff == "y1":
             val = yield self.m_state[profile | (channel << w.profile)]
         elif coeff == "x0":
-            val = yield self.m_state[(channel << 1) |
-                    (1 << w.profile + w.channel)]
+            val = yield self.m_state[(channel << 1) +
+                    (1 << w.profile) * w_o.channels]
         elif coeff == "x1":
-            val = yield self.m_state[1 | (channel << 1) |
-                    (1 << w.profile + w.channel)]
+            val = yield self.m_state[1 | (channel << 1) +
+                    (1 << w.profile) * w_o.channels]
         else:
             raise ValueError("no such state", coeff)
         return signed(val, w.state)
 
+    def get_accum_ftw(self, channel):
+        val = yield self.m_accum_ftw[channel]
+        return val
+
+    def get_t_ref(self, channel):
+        val = yield self.m_t_ref[channel]
+        return val
+
     def fast_iter(self):
         """Perform a single processing iteration."""
         assert (yield self.done)
@@ -607,6 +731,8 @@ def check_iter(self):
         """Perform a single processing iteration while verifying
         the behavior."""
         w = self.widths
+        w_i = self.widths_adc
+        w_o = self.widths_dds
 
         while not (yield self.done):
             yield
@@ -622,25 +748,33 @@ def check_iter(self):
 
         x0s = []
         # check adc loading
-        for i in range(1 << w.channel):
+        for i in range(w_i.channels):
             v_adc = signed((yield self.adc[i]), w.adc)
             x0 = yield from self.get_state(i, coeff="x0")
             x0s.append(x0)
-            assert v_adc << (w.state - w.adc - 1) == x0, (hex(v_adc), hex(x0))
             logger.debug("adc[%d] adc=%x x0=%x", i, v_adc, x0)
+            assert v_adc << (w.state - w.adc - 1) == x0, (hex(v_adc), hex(x0))
 
         data = []
         # predict output
-        for i in range(1 << w.channel):
+        for i in range(w_o.channels):
+            t0 = yield self._dt_start
+            dds_ftw_accu = yield from self.get_accum_ftw(i)
+            sysclks_ref = (yield from self.get_t_ref(i)) * self.widths_dds.sysclk_per_clk\
+                           + (yield self.ctrl_reftime[i].sysclks_fine)
+            logger.debug("dt_start=%d dt_ref=%d t_cycle=%d ftw_accu=%#x",
+                         t0, sysclks_ref, self.t_cycle, dds_ftw_accu)
+
             j = yield self.ctrl[i].profile
             en_iir = yield self.ctrl[i].en_iir
             en_out = yield self.ctrl[i].en_out
+            en_pt = yield self.ctrl[i].en_pt
             dly_i = yield self._dlys[i]
-            logger.debug("ctrl[%d] profile=%d en_iir=%d en_out=%d dly=%d",
-                    i, j, en_iir, en_out, dly_i)
+            logger.debug("ctrl[%d] profile=%d en_iir=%d en_out=%d en_pt=%d dly=%d",
+                    i, j, en_iir, en_out, en_pt, dly_i)
 
             cfg = yield from self.get_coeff(i, j, "cfg")
-            k_j = cfg & ((1 << w.channel) - 1)
+            k_j = cfg & ((1 << bits_for(w_i.channels - 1)) - 1)
             dly_j = (cfg >> 8) & 0xff
             logger.debug("cfg[%d,%d] sel=%d dly=%d", i, j, k_j, dly_j)
 
@@ -657,9 +791,13 @@ def check_iter(self):
 
             ftw0 = yield from self.get_coeff(i, j, "ftw0")
             ftw1 = yield from self.get_coeff(i, j, "ftw1")
-            pow = yield from self.get_coeff(i, j, "pow")
-            logger.debug("dds[%d,%d] ftw0=%#x ftw1=%#x pow=%#x",
-                    i, j, ftw0, ftw1, pow)
+            _pow = yield from self.get_coeff(i, j, "pow")
+            ph_coh = ((ftw0 | (ftw1 << w.word)) * (t0 - sysclks_ref))
+            ph_accu = dds_ftw_accu * self.t_cycle * self.widths_dds.sysclk_per_clk
+            ph = ph_coh - ph_accu
+            pow = (_pow + (ph >> w.word)) & 0xffff if en_pt else _pow
+            logger.debug("dds[%d,%d] ftw0=%#x ftw1=%#x ph_coh=%#x _pow=%#x pow=%#x",
+                    i, j, ftw0, ftw1, ph_coh, _pow, pow)
 
             y1 = yield from self.get_state(i, j, "y1")
             x1 = yield from self.get_state(k_j, coeff="x1")
@@ -681,6 +819,10 @@ def check_iter(self):
         # wait for output
         assert (yield self.processing)
         while (yield self.processing):
+            logger.debug("sysclks_to_ref=%d sysclks_ref_to_iter_start=%d",
+                         (yield self._sysclks_to_ref),
+                         (yield self._sysclks_ref_to_iter_start))
+            # logger.debug("%d %d %d %d", *[x for x in (yield self._sysclks_ref_fine)])
             yield
 
         assert (yield self.shifting)
@@ -694,7 +836,7 @@ def check_iter(self):
             logger.debug("adc[%d] x0=%x x1=%x", i, x0, x1)
 
         # check new state
-        for i in range(1 << w.channel):
+        for i in range(w_o.channels):
             j = yield self.ctrl[i].profile
             logger.debug("ch[%d] profile=%d", i, j)
             y1 = yield from self.get_state(i, j, "y1")
@@ -702,7 +844,7 @@ def check_iter(self):
             assert y1 == y0, (hex(y1), hex(y0))
 
         # check dds output
-        for i in range(1 << w.channel):
+        for i in range(w_o.channels):
             ftw0, ftw1, pow, y0, x1, x0 = data[i]
             asf = y0 >> (w.state - w.asf - 1)
             dds = (ftw0 | (ftw1 << w.word) |
diff --git a/artiq/gateware/suservo/pads.py b/artiq/gateware/suservo/pads.py
index 0ab7d352f1..bdae8ee35c 100644
--- a/artiq/gateware/suservo/pads.py
+++ b/artiq/gateware/suservo/pads.py
@@ -1,5 +1,7 @@
 from migen import *
 from migen.genlib.io import DifferentialOutput, DifferentialInput, DDROutput
+from artiq.gateware.rtio.phy import ttl_serdes_7series, ttl_serdes_generic
+from artiq.gateware.rtio import rtlink
 
 
 class SamplerPads(Module):
@@ -57,27 +59,85 @@ def __init__(self, platform, eem):
                 clk=dp.clkout, port=sdop)
 
 
+class OutIoUpdate_8X(Module):
+    def __init__(self, pad):
+        serdes = ttl_serdes_7series._OSERDESE2_8X()
+        self.submodules += serdes
+
+        self.passthrough = Signal()
+        self.data = Signal()
+        self.fine_ts = Signal(3)
+
+        self.rtlink = rtlink.Interface(
+            rtlink.OInterface(1, fine_ts_width=3))
+        self.probes = [serdes.o[-1]]
+        override_en = Signal()
+        override_o = Signal()
+        self.overrides = [override_en, override_o]
+
+        # # #
+
+        self.specials += Instance("IOBUFDS",
+                                  i_I=serdes.ser_out,
+                                  i_T=serdes.t_out,
+                                  io_IO=pad.p,
+                                  io_IOB=pad.n)
+
+        # Just strobe always in non-passthrough mode, as self.data is supposed
+        # to be always valid.
+        self.submodules += ttl_serdes_generic._SerdesDriver(
+            serdes.o,
+            Mux(self.passthrough, self.rtlink.o.stb, 1),
+            Mux(self.passthrough, self.rtlink.o.data, self.data),
+            Mux(self.passthrough, self.rtlink.o.fine_ts, self.fine_ts),
+            override_en, override_o)
+
+        self.comb += self.rtlink.o.busy.eq(~self.passthrough)
+
+
 class UrukulPads(Module):
     def __init__(self, platform, *eems):
         spip, spin = [[
                 platform.request("{}_qspi_{}".format(eem, pol), 0)
                 for eem in eems] for pol in "pn"]
-        ioup = [platform.request("{}_io_update".format(eem), 0)
-                for eem in eems]
+
         self.cs_n = Signal()
         self.clk = Signal()
         self.io_update = Signal()
+        self.passthrough = Signal()
+        self.dds_reset_sync_in = Signal(reset=0)  # sync_in phy (one for all)
+
+        # # #
+
+        self.io_update_phys = []
+        for eem in eems:
+            phy = OutIoUpdate_8X(platform.request("{}_io_update".format(eem), 0))
+            self.io_update_phys.append(phy)
+            setattr(self.submodules, "{}_io_update_phy".format(eem), phy)
+            self.comb += [
+                phy.data.eq(self.io_update),
+                phy.passthrough.eq(self.passthrough),
+            ]
+
+            sync_in_pads = platform.request("{}_dds_reset_sync_in".format(eem))
+            sync_in_r = Signal()
+            self.sync.rio_phy += sync_in_r.eq(self.dds_reset_sync_in)
+            sync_in_o = Signal()
+            self.specials += Instance("ODDR",
+                p_DDR_CLK_EDGE="SAME_EDGE",
+                i_C=ClockSignal("rio_phy"), i_CE=1, i_S=0, i_R=0,
+                i_D1=sync_in_r, i_D2=sync_in_r, o_Q=sync_in_o)
+            self.specials += DifferentialOutput(sync_in_o, sync_in_pads.p, sync_in_pads.n)
+
         self.specials += [(
                 DifferentialOutput(~self.cs_n, spip[i].cs, spin[i].cs),
-                DifferentialOutput(self.clk, spip[i].clk, spin[i].clk),
-                DifferentialOutput(self.io_update, ioup[i].p, ioup[i].n))
+                DifferentialOutput(self.clk, spip[i].clk, spin[i].clk))
                 for i in range(len(eems))]
-        for i in range(8):
+        for i in range(4 * len(eems)):
             mosi = Signal()
             setattr(self, "mosi{}".format(i), mosi)
-        for i in range(4*len(eems)):
             self.specials += [
-                DifferentialOutput(getattr(self, "mosi{}".format(i)),
+                DifferentialOutput(mosi,
                     getattr(spip[i // 4], "mosi{}".format(i % 4)),
                     getattr(spin[i // 4], "mosi{}".format(i % 4)))
             ]
diff --git a/artiq/gateware/suservo/servo.py b/artiq/gateware/suservo/servo.py
index 1aec95f027..15d31027e0 100644
--- a/artiq/gateware/suservo/servo.py
+++ b/artiq/gateware/suservo/servo.py
@@ -42,7 +42,7 @@ def __init__(self, adc_pads, dds_pads, adc_p, iir_p, dds_p):
         assert t_iir + 2*adc_p.channels < t_cycle, "need shifting time"
 
         self.submodules.adc = ADC(adc_pads, adc_p)
-        self.submodules.iir = IIR(iir_p)
+        self.submodules.iir = IIR(iir_p, adc_p, dds_p, t_cycle)
         self.submodules.dds = DDS(dds_pads, dds_p)
 
         # adc channels are reversed on Sampler
@@ -63,7 +63,6 @@ def __init__(self, adc_pads, dds_pads, adc_p, iir_p, dds_p):
         assert t_restart > 1
         cnt = Signal(max=t_restart)
         cnt_done = Signal()
-        active = Signal(3)
 
         # Indicates whether different steps (0: ADC, 1: IIR, 2: DDS) are
         # currently active (exposed for simulation only), with each bit being
@@ -71,6 +70,8 @@ def __init__(self, adc_pads, dds_pads, adc_p, iir_p, dds_p):
         # timing details of the different steps, any number can be concurrently
         # active (e.g. ADC read from iteration n, IIR computation from iteration
         # n - 1, and DDS write from iteration n - 2).
+        active = Signal(3)
+        self._active = active  # Exposed for debugging only.
 
         # Asserted once per cycle when the DDS write has been completed.
         self.done = Signal()
@@ -95,6 +96,17 @@ def __init__(self, adc_pads, dds_pads, adc_p, iir_p, dds_p):
                     cnt.eq(t_restart - 1)
                 )
         ]
+
+        # Count number of cycles since the servo was last started from idle.
+        self.sync += If(active == 0,
+            self.iir.t_running.eq(0),
+            self.iir.reset_dds_phase.eq(1)
+        ).Else(
+            self.iir.t_running.eq(self.iir.t_running + 1)
+        )
+
+        self.sync += dds_pads.passthrough.eq(active == 0)
+
         self.comb += [
                 cnt_done.eq(cnt == 0),
                 self.adc.start.eq(self.start & cnt_done),
diff --git a/artiq/gateware/test/suservo/__init__.py b/artiq/gateware/test/suservo/__init__.py
index e69de29bb2..7a1df77ac1 100644
--- a/artiq/gateware/test/suservo/__init__.py
+++ b/artiq/gateware/test/suservo/__init__.py
@@ -0,0 +1,10 @@
+"""Gateware implementation of the Sampler-Urukul (AD9910) DDS amplitude servo.
+
+General conventions:
+
+ - ``t_...`` signals and constants refer to time spans measured in the gateware
+   module's default clock (typically a 125 MHz RTIO clock).
+ - ``start`` signals cause modules to proceed with the next servo iteration iff
+   they are currently idle (i.e. their value is irrelevant while the module is
+   busy, so they are not necessarily one-clock-period strobes).
+"""
diff --git a/artiq/gateware/test/suservo/test_dds.py b/artiq/gateware/test/suservo/test_dds.py
index a666f14c56..d9a8167590 100644
--- a/artiq/gateware/test/suservo/test_dds.py
+++ b/artiq/gateware/test/suservo/test_dds.py
@@ -5,6 +5,9 @@
 
 from artiq.gateware.suservo.dds_ser import DDSParams, DDS
 
+class OutIoUpdateTB(Module):
+    def __init__(self):
+        self.fine_ts = Signal(3)
 
 class TB(Module):
     def __init__(self, p):
@@ -15,6 +18,12 @@ def __init__(self, p):
             setattr(self, "mosi{}".format(i), m)
         self.miso = Signal()
         self.io_update = Signal()
+        self.passthrough = Signal()
+
+        self.io_update_phys = []
+        for i in range(p.channels//4):
+            phy = OutIoUpdateTB()
+            self.io_update_phys.append(phy)
 
         clk0 = Signal()
         self.sync += clk0.eq(self.clk)
@@ -23,16 +32,19 @@ def __init__(self, p):
 
         self.ddss = []
         for i in range(p.channels):
-            dds = Record([("ftw", 32), ("pow", 16), ("asf", 16), ("cmd", 8)])
-            sr = Signal(len(dds))
+            dds = Record([("ftw", 32), ("pow", 16), ("asf", 16),
+                          ("cmd", 8), ("accu", 32), ("phase", 19)])
+            sr = Signal(32 + 16 + 16 + 8)
             self.sync += [
+                    dds.accu.eq(dds.accu + p.sysclk_per_clk * dds.ftw),
                     If(~self.cs_n & sample,
                         sr.eq(Cat(self.mosi[i], sr))
                     ),
                     If(self.io_update,
-                        dds.raw_bits().eq(sr)
+                        dds.raw_bits()[:len(sr)].eq(sr)
                     )
             ]
+            self.comb += dds.phase.eq((dds.pow << 3) + (dds.accu >> 13))
             self.ddss.append(dds)
 
     @passive
@@ -55,7 +67,7 @@ def log(self, data):
 
 
 def main():
-    p = DDSParams(channels=4, width=8 + 32 + 16 + 16, clk=1)
+    p = DDSParams(channels=4, width=8 + 32 + 16 + 16, clk=1, sysclk_per_clk=8)
     tb = TB(p)
     dds = DDS(tb, p)
     tb.submodules += dds
diff --git a/artiq/gateware/test/suservo/test_iir.py b/artiq/gateware/test/suservo/test_iir.py
index 919e7a6bf9..ab8a9a4a46 100644
--- a/artiq/gateware/test/suservo/test_iir.py
+++ b/artiq/gateware/test/suservo/test_iir.py
@@ -2,48 +2,67 @@
 import unittest
 
 from migen import *
-from artiq.gateware.suservo import iir
+from artiq.gateware.suservo import servo
+from collections import namedtuple
 
+logger = logging.getLogger(__name__)
+
+ADCParamsSim = namedtuple("ADCParams", ["channels"])
+DDSParamsSim = namedtuple("ADCParams", ["channels", "sysclk_per_clk"])
 
 def main():
-    w_kasli = iir.IIRWidths(state=25, coeff=18, adc=16,
-            asf=14, word=16, accu=48, shift=11,
-            channel=3, profile=5, dly=8)
-    w = iir.IIRWidths(state=17, coeff=16, adc=16,
-            asf=14, word=16, accu=48, shift=11,
-            channel=2, profile=1, dly=8)
+    w_kasli = servo.IIRWidths(state=25, coeff=18, adc=16, asf=14,
+            word=16, accu=48, shift=11, profile=5, dly=8)
+    p_adc = ADCParamsSim(channels=8)
+    p_dds = DDSParamsSim(channels=4, sysclk_per_clk=8)
+    w = servo.IIRWidths(state=17, coeff=16, adc=16, asf=14,
+            word=16, accu=48, shift=11, profile=2, dly=8)
 
+    t_iir = p_adc.channels + 4*p_dds.channels + 8 + 1
     def run(dut):
+        yield dut.t_running.eq(0)
         for i, ch in enumerate(dut.adc):
             yield ch.eq(i)
         for i, ch in enumerate(dut.ctrl):
             yield ch.en_iir.eq(1)
             yield ch.en_out.eq(1)
             yield ch.profile.eq(i)
-        for i in range(1 << w.channel):
+            yield ch.en_pt.eq(i)
+        for i, ch in enumerate(dut.ctrl_reftime):
+            yield ch.sysclks_fine.eq(i)
+            yield ch.stb.eq(1)
+            yield
+            yield dut.t_running.eq(dut.t_running + 1)
+            yield ch.stb.eq(0)
+            yield
+            yield dut.t_running.eq(dut.t_running + 1)
+        for i in range(p_adc.channels):
             yield from dut.set_state(i, i << 8, coeff="x1")
             yield from dut.set_state(i, i << 8, coeff="x0")
+        for i in range(p_dds.channels):
             for j in range(1 << w.profile):
                 yield from dut.set_state(i,
                         (j << 1) | (i << 8), profile=j, coeff="y1")
                 for k, l in enumerate("pow offset ftw0 ftw1".split()):
                     yield from dut.set_coeff(i, profile=j, coeff=l,
-                            value=(i << 12) | (j << 8) | (k << 4))
+                            value=(i << 10) | (j << 8) | (k << 4))
         yield
-        for i in range(1 << w.channel):
+        for i in range(p_dds.channels):
             for j in range(1 << w.profile):
-                for k, l in enumerate("cfg a1 b0 b1".split()):
+                for k, l in enumerate("a1 b0 b1".split()):
                     yield from dut.set_coeff(i, profile=j, coeff=l,
-                            value=(i << 12) | (j << 8) | (k << 4))
+                            value=(i << 10) | (j << 8) | (k << 4))
                 yield from dut.set_coeff(i, profile=j, coeff="cfg",
-                        value=(i << 0) | (j << 8))  # sel, dly
+                        value=(i % p_adc.channels) | (j << 8))  # sel, dly
         yield
-        for i in range(10):
+        for i in range(4):
+            logger.debug("check_iter {}".format(i))
             yield from dut.check_iter()
+            yield dut.t_running.eq((yield dut.t_running) + t_iir)
             yield
 
-    dut = iir.IIR(w)
-    run_simulation(dut, [run(dut)], vcd_name="iir.vcd")
+    dut = servo.IIR(w, p_adc, p_dds, t_iir)
+    run_simulation(dut, [run(dut)], vcd_name="servo.vcd")
 
 
 class IIRTest(unittest.TestCase):
diff --git a/artiq/gateware/test/suservo/test_servo.py b/artiq/gateware/test/suservo/test_servo.py
index cc1a73a2be..fe1708d033 100644
--- a/artiq/gateware/test/suservo/test_servo.py
+++ b/artiq/gateware/test/suservo/test_servo.py
@@ -1,5 +1,6 @@
 import logging
 import unittest
+import numpy as np
 
 from migen import *
 from migen.genlib import io
@@ -7,15 +8,17 @@
 from artiq.gateware.test.suservo import test_adc, test_dds
 from artiq.gateware.suservo import servo
 
+logger = logging.getLogger(__name__)
+
 
 class ServoSim(servo.Servo):
     def __init__(self):
         adc_p = servo.ADCParams(width=16, channels=8, lanes=4,
                 t_cnvh=4, t_conv=57 - 4, t_rtt=4 + 4)
         iir_p = servo.IIRWidths(state=25, coeff=18, adc=16, asf=14, word=16,
-                accu=48, shift=11, channel=3, profile=5, dly=8)
+                accu=48, shift=11, profile=5, dly=8)
         dds_p = servo.DDSParams(width=8 + 32 + 16 + 16,
-                channels=adc_p.channels, clk=1)
+                channels=4, clk=1, sysclk_per_clk=8)
 
         self.submodules.adc_tb = test_adc.TB(adc_p)
         self.submodules.dds_tb = test_dds.TB(dds_p)
@@ -23,37 +26,156 @@ def __init__(self):
         servo.Servo.__init__(self, self.adc_tb, self.dds_tb,
                 adc_p, iir_p, dds_p)
 
+        self.dds_output = []
+
+    def log_flow(self, cycle):
+        su_start = yield self.start
+        adc_start = yield self.adc.start
+        iir_start = yield self.iir.start
+        dds_start = yield self.dds.start
+        su_done = yield self.done
+        adc_done = yield self.adc.done
+        iir_done = yield self.iir.done
+        dds_done = yield self.dds.done
+        active = yield self._active
+        io_update = yield self.dds_tb.io_update
+        passthrough = yield self.dds_tb.passthrough
+        iir_loading = yield self.iir.loading
+        iir_processing = yield self.iir.processing
+        iir_shifting = yield self.iir.shifting
+        dt = yield self.iir.t_running
+        dt_iir = yield self.iir._dt_start
+        state = yield self.iir._state
+        stage0 = yield self.iir._stages[0]
+        stage1 = yield self.iir._stages[1]
+        stage2 = yield self.iir._stages[2]
+        logger.debug(
+            "cycle=%d "
+            #"start=[su=%d adc=%d iir=%d dds=%d] "
+            #"done=[su=%d adc=%d iir=%d dds=%d] "
+            "active=%s load_proc_shft=%d%d%d stages_active=%d%d%d "
+            "io_update=%d passthrough=%d "
+            "dt=%d dt_iir=%d state=%d",
+            cycle,
+            #su_start, adc_start, iir_start, dds_start,
+            #su_done, adc_done, iir_done, dds_done,
+            '{:03b}'.format(active), iir_loading, iir_processing, iir_shifting, stage0, stage1, stage2,
+            io_update, passthrough,
+            dt, dt_iir//8, state
+        )
+
+    def log_state(self, channel, profile, calls=[0]):
+        calls[0] += 1
+        # if not (yield self._active[1]):
+        #     return
+        yield from self.log_flow(calls[0] - 2)
+        return
+        cfg = yield from self.iir.get_coeff(channel, profile, "cfg")
+        sel = cfg & 0x7
+        x0 = yield from self.iir.get_state(sel, coeff="x0")
+        x1 = yield from self.iir.get_state(sel, coeff="x1")
+        y1 = yield from self.iir.get_state(channel, profile, coeff="y1")
+        _pow = yield from self.iir.get_coeff(channel, profile, "pow")
+        pow_iir = yield self.iir.dds[channel][2*self.iir.widths.word:3*self.iir.widths.word]
+        pow_dds = yield self.dds_tb.ddss[channel].pow
+        asf_dds = yield self.dds_tb.ddss[channel].asf
+        ftw_dds = yield self.dds_tb.ddss[channel].ftw
+        accu_dds = yield self.dds_tb.ddss[channel].accu
+        phase_dds = (yield self.dds_tb.ddss[channel].phase)
+        dds_output = np.cos(2*np.pi*phase_dds/2**19)
+        ph_coh = yield self.iir._ph_coh
+        ph_acc = yield self.iir._ph_acc
+        offset = yield from self.iir.get_coeff(channel, profile, "offset")
+        ftw0 = yield from self.iir.get_coeff(channel, profile, "ftw0")
+        ftw1 = yield from self.iir.get_coeff(channel, profile, "ftw1")
+        m_phase = yield from self.iir.get_accum_ftw(channel)
+        iir_adc = yield self.iir.adc[sel]
+        logger.debug("\t"
+                     "ch=%d pr=%d "
+                     # "x0=%d x1=%d adc=%d y1=%d sel=%d "
+                     "ftw=%#x pow_coeff=%#x ftw_accu=%#x "
+                     "ph_coh=%#x ph_acc=%#x "
+                     "pow_iir=%#x pow_dds=%#x ftw_dds=%#x asf_dds=%#x accu_dds=%#x phase_dds=%#x dds_output=%04.3f",
+                     channel, profile,
+                     # x0, x1, iir_adc, y1, sel,
+                     ftw0 | (ftw1 << 16), _pow, m_phase,
+                     ph_coh, ph_acc,
+                     pow_iir, pow_dds, ftw_dds, asf_dds, accu_dds, phase_dds >> 3, dds_output
+        )
+        self.dds_output.append(dds_output)
+        # yield from self.log_registers(profile)
+
+    def log_registers(self, profile):
+        adc_channels = self.iir.widths_adc.channels
+        dds_channels = self.iir.widths_dds.channels
+        x0s = [0]*adc_channels
+        x1s = [0]*adc_channels
+        y1s = [0]*dds_channels
+        for ch in range(adc_channels):
+            x0s[ch] = yield from self.iir.get_state(ch, coeff="x0")
+            x1s[ch] = yield from self.iir.get_state(ch, coeff="x1")
+        for ch in range(dds_channels):
+            y1s[ch] = yield from self.iir.get_state(ch, profile, coeff="y1")
+
+        logger.debug(("x0s = " + '{:05X} ' * adc_channels).format(*x0s))
+        logger.debug(("x1s = " + '{:05X} ' * adc_channels).format(*x1s))
+        logger.debug(("y1s = " + '{:05X} ' * dds_channels).format(*y1s))
+
     def test(self):
         assert (yield self.done)
 
-        adc = 1
+        adc = 7
         x0 = 0x0141
         yield self.adc_tb.data[-adc-1].eq(x0)
-        channel = 3
-        yield self.iir.adc[channel].eq(adc)
+        channel = 0
         yield self.iir.ctrl[channel].en_iir.eq(1)
         yield self.iir.ctrl[channel].en_out.eq(1)
-        profile = 5
+        yield self.iir.ctrl[channel].en_pt.eq(1)
+        profile = 31
         yield self.iir.ctrl[channel].profile.eq(profile)
         x1 = 0x0743
         yield from self.iir.set_state(adc, x1, coeff="x1")
         y1 = 0x1145
         yield from self.iir.set_state(channel, y1,
                 profile=profile, coeff="y1")
-        coeff = dict(pow=0x1333, offset=0x1531, ftw0=0x1727, ftw1=0x1929,
-                a1=0x0135, b0=0x0337, b1=0x0539, cfg=adc | (0 << 3))
+        coeff = dict(pow=0, offset=0x1531, ftw0=0xeb85, ftw1=0x51,
+                a1=0x0135, b0=0x0337, b1=0x0539, cfg=adc)
         for ks in "pow offset ftw0 ftw1", "a1 b0 b1 cfg":
             for k in ks.split():
                 yield from self.iir.set_coeff(channel, value=coeff[k],
                         profile=profile, coeff=k)
             yield
 
+        num_it = 1
+        num_proc_its = [0]*num_it # number of iterations while iir.processing
+        yield from self.log_state(channel, profile)
         yield self.start.eq(1)
         yield
-        yield self.start.eq(0)
-        while not (yield self.dds_tb.io_update):
-            yield
-        yield  # io_update
+        for i in range(num_it):
+            if i == 1:  # change ftw
+                yield from self.iir.set_coeff(channel,
+                    profile=profile, coeff='ftw0', value=coeff['ftw1'])
+                yield from self.iir.set_coeff(channel,
+                    profile=profile, coeff='ftw1', value=coeff['ftw0'])
+            if i == 2:  # change ftw back
+                yield from self.iir.set_coeff(channel,
+                    profile=profile, coeff='ftw0', value=coeff['ftw0'])
+                yield from self.iir.set_coeff(channel,
+                    profile=profile, coeff='ftw1', value=coeff['ftw1'])
+            logger.debug("iteration {}".format(i))
+            yield from self.log_state(channel, profile)
+            if i == num_it-1:
+                yield self.start.eq(0)
+            while not (yield self.dds_tb.io_update):
+                yield
+                if (yield self.iir.processing):
+                    num_proc_its[i] += 1
+                if (yield self.iir._stages) != 0:
+                    yield from self.log_state(channel, profile)
+            yield  # io_update
+        yield from self.log_state(channel, profile)
+        yield
+        yield from self.log_state(channel, profile)
 
         w = self.iir.widths
 
@@ -63,6 +185,8 @@ def test(self):
 
         offset = coeff["offset"] << (w.state - w.coeff - 1)
         a1, b0, b1 = coeff["a1"], coeff["b0"], coeff["b1"]
+
+        # works only for 1 iteration
         out = (
                 0*(1 << w.shift - 1) +  # rounding
                 a1*(y1 + 0) + b0*(x0 + offset) + b1*(x1 + offset)
@@ -76,8 +200,15 @@ def test(self):
         ftw = (coeff["ftw1"] << 16) | coeff["ftw0"]
         assert _ == ftw, (hex(_), hex(ftw))
 
+        t0 = yield self.iir._dt_start
+        # todo: include phase accumulator
+        ph = (ftw * t0) >> 16
+        if (yield self.iir.ctrl[channel].en_pt):
+            pow = (coeff["pow"] + ph) & 0xffff
+        else:
+            pow = coeff["pow"]
         _ = yield self.dds_tb.ddss[channel].pow
-        assert _ == coeff["pow"], (hex(_), hex(coeff["pow"]))
+        assert _ == pow, (hex(_), hex(pow))
 
         _ = yield self.dds_tb.ddss[channel].asf
         asf = y1 >> (w.state - w.asf - 1)
@@ -101,4 +232,5 @@ def test_run(self):
 
 
 if __name__ == "__main__":
+    logging.basicConfig(level=logging.DEBUG)
     main()