예제 #1
0
def gen_model(real_vals, sint_vals, uint_vals, addr_bits, sint_bits, uint_bits,
              real_type):
    # create mixed-signal model
    model = MixedSignalModel('model', build_dir=BUILD_DIR, real_type=real_type)
    model.add_digital_input('addr', width=addr_bits)
    model.add_digital_input('clk')
    model.add_analog_output('real_out')
    model.add_digital_output('sint_out', width=sint_bits)
    model.add_digital_output('uint_out', width=uint_bits)

    # create tables
    real_table = model.make_real_table(real_vals)
    sint_table = model.make_sint_table(sint_vals)
    uint_table = model.make_uint_table(uint_vals)

    # assign values
    model.set_from_sync_rom(model.real_out,
                            real_table,
                            model.addr,
                            clk=model.clk)
    model.set_from_sync_rom(model.sint_out,
                            sint_table,
                            model.addr,
                            clk=model.clk)
    model.set_from_sync_rom(model.uint_out,
                            uint_table,
                            model.addr,
                            clk=model.clk)

    # write the model
    return model.compile_to_file(VerilogGenerator())
예제 #2
0
    def __init__(self,  filename=None, **system_values):
        # set a fixed random seed for repeatability
        np.random.seed(2)

        module_name = Path(filename).stem
        build_dir   = Path(filename).parent

        print(f'Running model generator for {module_name}...')
        assert (all([req_val in system_values for req_val in self.required_values()])), \
            f'Cannot build {module_name}, Missing parameter in config file'

        m = MixedSignalModel(module_name, dt=system_values['dt'], build_dir=build_dir,
                             real_type=get_dragonphy_real_type())
        m.add_real_param('t_del', 0)
        m.add_digital_input('emu_rst')
        m.add_digital_input('emu_clk')
        m.add_analog_input('t_lo')
        m.add_analog_input('t_hi')
        m.add_analog_input('emu_dt')
        m.add_analog_output('dt_req', init=m.t_del)
        m.add_digital_output('clk_val')

        # determine if the request was granted
        m.bind_name('req_grant', m.dt_req == m.emu_dt)

        # update the clock value
        m.add_digital_state('prev_clk_val')
        m.set_next_cycle(m.prev_clk_val, m.clk_val, clk=m.emu_clk, rst=m.emu_rst)
        m.set_this_cycle(m.clk_val, if_(m.req_grant, ~m.prev_clk_val, m.prev_clk_val))

        # determine arguments for formating time steps
        # TODO: clean this up
        dt_fmt_kwargs = dict(
            range_=m.emu_dt.format_.range_,
            width=m.emu_dt.format_.width,
            exponent=m.emu_dt.format_.exponent
        )
        array_fmt_kwargs = deepcopy(dt_fmt_kwargs)
        array_fmt_kwargs['real_range_hint'] = m.emu_dt.format_.range_
        del array_fmt_kwargs['range_']

        # determine the next period
        dt_req_next_array = array([m.t_hi, m.t_lo], m.prev_clk_val, **array_fmt_kwargs)
        m.bind_name('dt_req_next', dt_req_next_array, **dt_fmt_kwargs)

        # increment the time request
        m.bind_name('dt_req_incr', m.dt_req - m.emu_dt, **dt_fmt_kwargs)

        # determine the next period
        dt_req_imm_array = array([m.dt_req_incr, m.dt_req_next], m.req_grant, **array_fmt_kwargs)
        m.bind_name('dt_req_imm', dt_req_imm_array, **dt_fmt_kwargs)
        m.set_next_cycle(m.dt_req, m.dt_req_imm, clk=m.emu_clk, rst=m.emu_rst)

        # generate the model
        m.compile_to_file(VerilogGenerator())

        self.generated_files = [filename]
예제 #3
0
파일: int_conv.py 프로젝트: xlchan/msdsl
def main():
    model = MixedSignalModel('model')
    model.add_analog_input('a')
    model.add_digital_input('b', width=8)
    model.add_digital_output('c', width=12)

    model.bind_name('d', model.a + model.b)

    clamped = clamp(to_sint(model.d, width=model.c.width + 1), 0, 1023)
    model.set_next_cycle(model.c, to_uint(clamped, width=model.c.width))

    model.compile_and_print(VerilogGenerator())
예제 #4
0
파일: comparator.py 프로젝트: xlchan/msdsl
def main():
    # define ports
    m = MixedSignalModel('comparator')
    m.add_analog_input('in_p')
    m.add_analog_input('in_n')
    m.add_digital_output('out', init=0)
    m.add_digital_input('clk')

    # define behavior
    m.immediate_assign('out_async', m.in_p > m.in_n)
    m.next_cycle_assign(m.out, m.out_async, clk=m.clk)

    # write model
    m.compile_and_print(VerilogGenerator())
예제 #5
0
def main():
    ctrl = ExampleControl()

    # define ports
    m = MixedSignalModel('comparator', dt=ctrl.dt)
    m.add_analog_input('in_p')
    m.add_analog_input('in_n')
    m.add_digital_output('out', init=0)
    m.add_digital_input('clk')

    # define behavior
    m.immediate_assign('out_async', m.in_p > m.in_n)
    m.next_cycle_assign(m.out, m.out_async, clk=m.clk, rst="1'b0")

    # write model
    ctrl.write_model(m)
예제 #6
0
def main():
    print('Running model generator...')

    # parse command line arguments
    parser = ArgumentParser()
    parser.add_argument('-o', '--output', type=str, default='build')
    parser.add_argument('--dt', type=float, default=0.1e-6)
    a = parser.parse_args()

    # create the model
    m = MixedSignalModel('osc', dt=a.dt)
    m.add_digital_input('emu_clk')
    m.add_digital_input('emu_rst')
    m.add_digital_output('dt_req', 32)
    m.add_digital_input('emu_dt', 32)
    m.add_digital_input('neg_emu_dt', 32)  # TODO: cleanup
    m.add_digital_output('clk_val')
    m.add_digital_input('t_lo', 32)
    m.add_digital_input('t_hi', 32)

    # determine if the request was granted
    m.bind_name('req_grant', m.dt_req == m.emu_dt)

    # update the clock value
    m.add_digital_state('prev_clk_val')
    m.set_next_cycle(m.prev_clk_val, m.clk_val, clk=m.emu_clk, rst=m.emu_rst)
    m.set_this_cycle(m.clk_val, if_(m.req_grant, ~m.prev_clk_val, m.prev_clk_val))

    # determine the next period
    m.bind_name('dt_req_next', if_(m.prev_clk_val, m.t_lo, m.t_hi))

    # increment the time request
    m.bind_name('dt_req_incr', m.dt_req + m.neg_emu_dt)

    # determine the next period
    m.bind_name('dt_req_imm', if_(m.req_grant, m.dt_req_next, m.dt_req_incr))
    m.set_next_cycle(m.dt_req, m.dt_req_imm[31:0], clk=m.emu_clk, rst=m.emu_rst)

    # determine the output filename
    filename = Path(a.output).resolve() / f'{m.module_name}.sv'
    print(f'Model will be written to: {filename}')

    # generate the model
    m.compile_to_file(VerilogGenerator(), filename)
예제 #7
0
def main():
    model = MixedSignalModel('model')

    model.add_analog_input('a_in')
    model.add_digital_output('d_out', width=8)

    model.add_analog_output('a_out')
    model.add_digital_input('d_in', width=8)

    # DAC from 0 to 1V as the input code varies from 0-255

    clamped = clamp(to_sint(model.a_in * 255, width=model.d_out.width + 1), 0,
                    255)
    model.set_this_cycle(model.d_out, to_uint(clamped,
                                              width=model.d_out.width))

    # ADC code goes from 0-255 as input voltage goes from 0 to 1V
    model.set_this_cycle(model.a_out, model.d_in / 255)

    model.compile_and_print(VerilogGenerator())
예제 #8
0
    def __init__(self, filename=None, **system_values):
        # set a fixed random seed for repeatability
        np.random.seed(3)

        module_name = Path(filename).stem
        build_dir = Path(filename).parent

        #This is a wonky way of validating this.. :(
        assert (all([req_val in system_values for req_val in self.required_values()])), \
            f'Cannot build {module_name}, Missing parameter in config file'

        m = MixedSignalModel(module_name,
                             dt=system_values['dt'],
                             build_dir=build_dir,
                             real_type=get_dragonphy_real_type())

        # Random number generator seed (default generated with random.org)
        m.add_digital_input('noise_seed', width=32)

        # main I/O: input, output, and clock
        m.add_analog_input('in_')
        m.add_digital_output('out_sgn')
        m.add_digital_output('out_mag', width=system_values['n'])
        m.add_digital_input('clk_val')

        # emulator clock and reset
        m.add_digital_input('emu_clk')
        m.add_digital_input('emu_rst')

        # Noise controls
        m.add_analog_input('noise_rms')

        # determine when sampling should happen
        m.add_digital_state('clk_val_prev')
        m.set_next_cycle(m.clk_val_prev,
                         m.clk_val,
                         clk=m.emu_clk,
                         rst=m.emu_rst)

        # detect a rising edge on the clock
        m.bind_name('pos_edge', m.clk_val & (~m.clk_val_prev))

        # delay the positive edge signal
        m.add_digital_state('pos_edge_prev', init=0)
        m.set_next_cycle(m.pos_edge_prev,
                         m.pos_edge,
                         clk=m.emu_clk,
                         rst=m.emu_rst)

        # add noise
        sample_noise = m.set_gaussian_noise('sample_noise',
                                            std=m.noise_rms,
                                            clk=m.emu_clk,
                                            rst=m.emu_rst,
                                            lfsr_init=m.noise_seed)
        in_plus_noise = m.bind_name('in_plus_noise', m.in_ + sample_noise)

        # determine out_sgn (note that the definition is opposite of the typical
        # meaning; "0" means negative)
        out_sgn = if_(in_plus_noise < 0, 0, 1)
        m.set_next_cycle(m.out_sgn,
                         out_sgn,
                         clk=m.emu_clk,
                         rst=m.emu_rst,
                         ce=m.pos_edge_prev)

        # determine out_mag
        vref, n = system_values['vref'], system_values['n']
        abs_val = if_(in_plus_noise < 0, -1.0 * in_plus_noise, in_plus_noise)
        code_real_unclamped = (abs_val / vref) * ((2**(n - 1)) - 1)
        code_real = clamp_op(code_real_unclamped, 0, (2**(n - 1)) - 1)
        code_sint = to_sint(code_real, width=n + 1)

        # TODO: clean this up -- since real ranges are not intervals, we need to tell MSDSL
        # that the range of the signed integer is smaller
        code_sint.format_ = SIntFormat(width=n + 1,
                                       min_val=0,
                                       max_val=(2**(n - 1)) - 1)

        code_uint = to_uint(code_sint, width=n)
        m.set_next_cycle(m.out_mag,
                         code_uint,
                         clk=m.emu_clk,
                         rst=m.emu_rst,
                         ce=m.pos_edge_prev)

        # generate the model
        m.compile_to_file(VerilogGenerator())

        self.generated_files = [filename]
예제 #9
0
    def __init__(self, filename=None, **system_values):
        # set a fixed random seed for repeatability
        np.random.seed(0)

        module_name = Path(filename).stem
        build_dir   = Path(filename).parent

        #This is a wonky way of validating this.. :(
        assert (all([req_val in system_values for req_val in self.required_values()])), \
            f'Cannot build {module_name}, Missing parameter in config file'

        m = MixedSignalModel(module_name, dt=system_values['dt'], build_dir=build_dir,
                             real_type=get_dragonphy_real_type())

        # Random number generator seeds (defaults generated with random.org)
        m.add_digital_input('jitter_seed', width=32)
        m.add_digital_input('noise_seed', width=32)

        # Chunk of bits from the history; corresponding delay is bit_idx/freq_tx delay
        m.add_digital_input('chunk', width=system_values['chunk_width'])
        m.add_digital_input('chunk_idx', width=int(ceil(log2(system_values['num_chunks']))))

        # Control code for the corresponding PI slice
        m.add_digital_input('pi_ctl', width=system_values['pi_ctl_width'])

        # Indicates sequencing of ADC slice within a bank (typically a static value)
        m.add_digital_input('slice_offset', width=int(ceil(log2(system_values['slices_per_bank']))))

        # Control codes that affect states in the slice
        m.add_digital_input('sample_ctl')
        m.add_digital_input('incr_sum')
        m.add_digital_input('write_output')

        # ADC sign and magnitude
        m.add_digital_output('out_sgn')
        m.add_digital_output('out_mag', width=system_values['n_adc'])

        # Emulator clock and reset
        m.add_digital_input('clk')
        m.add_digital_input('rst')

        # Noise controls
        m.add_analog_input('jitter_rms')
        m.add_analog_input('noise_rms')

        # Create "placeholder function" that can be updated
        # at runtime with the channel function
        chan_func = PlaceholderFunction(
                domain=system_values['func_domain'],
                order=system_values['func_order'],
                numel=system_values['func_numel'],
                coeff_widths=system_values['func_widths'],
                coeff_exps=system_values['func_exps']
        )

        # Check the function on a representative test case
        chan = Filter.from_file(get_file('build/chip_src/adapt_fir/chan.npy'))
        self.check_func_error(chan_func, chan.interp)

        # Add digital inputs that will be used to reconfigure the function at runtime
        wdata, waddr, we = add_placeholder_inputs(m=m, f=chan_func)

        # Sample the pi_ctl code
        m.add_digital_state('pi_ctl_sample', width=system_values['pi_ctl_width'])
        m.set_next_cycle(m.pi_ctl_sample, m.pi_ctl, clk=m.clk, rst=m.rst, ce=m.sample_ctl)

        # compute weights to apply to pulse responses
        weights = []
        for k in range(system_values['chunk_width']):
            # create a weight value for this bit
            weights.append(
                m.add_analog_state(
                    f'weights_{k}',
                    range_=system_values['vref_tx']
                )
            )

            # select a single bit from the chunk.  chunk_width=1 is unfortunately
            # a special case because some simulators don't support the bit-selection
            # syntax on a single-bit variable
            chunk_bit = m.chunk[k] if system_values['chunk_width'] > 1 else m.chunk

            # write the weight value
            m.set_next_cycle(
                weights[-1],
                if_(
                    chunk_bit,
                    system_values['vref_tx'],
                    -system_values['vref_tx']
                ),
                clk=m.clk,
                rst=m.rst
            )

        # Compute the delay due to the PI control code
        delay_amt_pre = m.bind_name(
            'delay_amt_pre',
            m.pi_ctl_sample / ((2.0**system_values['pi_ctl_width'])*system_values['freq_rx'])
        )

        # Add jitter to the sampling time
        if system_values['use_jitter']:
            # create a signal to represent jitter
            delay_amt_jitter = m.set_gaussian_noise(
                'delay_amt_jitter',
                std=m.jitter_rms,
                lfsr_init=m.jitter_seed,
                clk=m.clk,
                ce=m.sample_ctl,
                rst=m.rst
            )

            # add jitter to the delay amount (which might possibly yield a negative value)
            delay_amt_noisy = m.bind_name('delay_amt_noisy', delay_amt_pre + delay_amt_jitter)

            # make the delay amount non-negative
            delay_amt = m.bind_name('delay_amt', if_(delay_amt_noisy >= 0.0, delay_amt_noisy, 0.0))
        else:
            delay_amt = delay_amt_pre

        # Compute the delay due to the slice offset
        t_slice_offset = m.bind_name('t_slice_offset', m.slice_offset/system_values['freq_rx'])

        # Add the delay amount to the slice offset
        t_samp_new = m.bind_name('t_samp_new', t_slice_offset + delay_amt)

        # Determine if the new sampling time happens after the end of this period
        t_one_period = m.bind_name('t_one_period', system_values['slices_per_bank']/system_values['freq_rx'])
        exceeds_period = m.bind_name('exceeds_period', t_samp_new >= t_one_period)

        # Save the previous sample time
        t_samp_prev = m.add_analog_state('t_samp_prev', range_=system_values['slices_per_bank']/system_values['freq_rx'])
        m.set_next_cycle(t_samp_prev, t_samp_new-t_one_period, clk=m.clk, rst=m.rst, ce=m.sample_ctl)

        # Save whether the previous sample time exceeded one period
        prev_exceeded = m.add_digital_state('prev_exceeded')
        m.set_next_cycle(prev_exceeded, exceeds_period, clk=m.clk, rst=m.rst, ce=m.sample_ctl)

        # Compute the sample time to use for this period
        t_samp_idx = m.bind_name('t_samp_idx', concatenate([exceeds_period, prev_exceeded]))
        t_samp = m.bind_name(
            't_samp',
            array(
                [
                    t_samp_new,   # 0b00: exceeds_period=0, prev_exceeded=0
                    t_samp_new,   # 0b01: exceeds_period=0, prev_exceeded=1
                    0.0,          # 0b10: exceeds_period=1, prev_exceeded=0
                    t_samp_prev   # 0b11: exceeds_period=1, prev_exceeded=1
                ], t_samp_idx
            )
        )

        # Evaluate the step response function.  Note that the number of evaluation times is the
        # number of chunks plus one.
        f_eval = []
        for k in range(system_values['chunk_width']+1):
            # compute change time as an integer multiple of the TX period
            chg_idx = m.bind_name(
                f'chg_idx_{k}', (
                    system_values['slices_per_bank']*system_values['num_banks']
                    - (m.chunk_idx+1)*system_values['chunk_width']
                    + k
                )
            )

            # scale by TX period
            t_chg = m.bind_name(f't_chg_{k}', chg_idx/system_values['freq_tx'])

            # compute the kth evaluation time
            t_eval = m.bind_name(f't_eval_{k}', t_samp - t_chg)

            # evaluate the function (the last three inputs are used for updating the function contents)
            f_eval.append(m.set_from_sync_func(f'f_eval_{k}', chan_func, t_eval, clk=m.clk, rst=m.rst,
                                               wdata=wdata, waddr=waddr, we=we))

        # Compute the pulse responses for each bit
        pulse_resp = []
        for k in range(system_values['chunk_width']):
            pulse_resp.append(
                m.bind_name(
                    f'pulse_resp_{k}',
                    weights[k]*(f_eval[k] - f_eval[k+1])
                )
            )

        # sum up all of the pulse responses
        pulse_resp_sum = m.bind_name('pulse_resp_sum', sum_op(pulse_resp))

        # update the overall sample value
        sample_value_pre = m.add_analog_state('analog_sample_pre', range_=5*system_values['vref_rx'])
        m.set_next_cycle(
            sample_value_pre,
            if_(m.incr_sum, sample_value_pre + pulse_resp_sum, pulse_resp_sum),
            clk=m.clk,
            rst=m.rst
        )

        # add noise to the sample value
        if system_values['use_noise']:
            sample_noise = m.set_gaussian_noise(
                'sample_noise',
                std=m.noise_rms,
                clk=m.clk,
                rst=m.rst,
                ce=m.write_output,
                lfsr_init=m.noise_seed
            )
            sample_value = m.bind_name('sample_value', sample_value_pre + sample_noise)
        else:
            sample_value = sample_value_pre

        # there is a special case in which the output should not be updated:
        # when the previous cycle did not exceed the period, but this one did
        # in that case the sample value should be held constant
        should_write_output = m.bind_name('should_write_output',
                                          (prev_exceeded | (~exceeds_period)) & m.write_output)

        # determine out_sgn (note that the definition is opposite of the typical
        # meaning; "0" means negative)
        out_sgn = if_(sample_value < 0, 0, 1)
        m.set_next_cycle(m.out_sgn, out_sgn, clk=m.clk, rst=m.rst, ce=should_write_output)

        # determine out_mag
        vref_rx, n_adc = system_values['vref_rx'], system_values['n_adc']
        abs_val = m.bind_name('abs_val', if_(sample_value < 0, -1.0*sample_value, sample_value))
        code_real_unclamped = m.bind_name('code_real_unclamped', (abs_val / vref_rx) * ((2**(n_adc-1))-1))
        code_real = m.bind_name('code_real', clamp_op(code_real_unclamped, 0, (2**(n_adc-1))-1))

        # TODO: clean this up -- since real ranges are not intervals, we need to tell MSDSL
        # that the range of the signed integer is smaller
        code_sint = to_sint(code_real, width=n_adc+1)
        code_sint.format_ = SIntFormat(width=n_adc+1, min_val=0, max_val=(2**(n_adc-1))-1)
        code_sint = m.bind_name('code_sint', code_sint)

        code_uint = m.bind_name('code_uint', to_uint(code_sint, width=n_adc))
        m.set_next_cycle(m.out_mag, code_uint, clk=m.clk, rst=m.rst, ce=should_write_output)

        # generate the model
        m.compile_to_file(VerilogGenerator())

        self.generated_files = [filename]
예제 #10
0
    def __init__(self, filename=None, **system_values):
        # set a fixed random seed for repeatability
        np.random.seed(1)

        module_name = Path(filename).stem
        build_dir   = Path(filename).parent

        #This is a wonky way of validating this.. :(
        assert (all([req_val in system_values for req_val in self.required_values()])), \
            f'Cannot build {module_name}, Missing parameter in config file'

        # instantiate model
        m = MixedSignalModel(module_name, dt=system_values['dt'], build_dir=build_dir)

        # main I/O: delay code and gain
        m.add_digital_input('code', width=system_values['n_bits'])
        m.add_digital_input('clk_i_val')
        m.add_digital_output('clk_o_val')

        # timestep control: DT request and response
        m.add_analog_output('dt_req')
        m.add_analog_input('emu_dt')

        # emulator clock and reset
        m.add_digital_input('emu_clk')
        m.add_digital_input('emu_rst')

        # additional input: maximum timestep
        # TODO: clean this up
        m.add_analog_input('dt_req_max')

        # jitter control
        m.add_digital_input('jitter_seed', width=32)
        m.add_analog_input('jitter_rms')

        # compute the delay (with no jitter)
        m.bind_name('delay_amt_pre', m.code * (system_values['t_per'] / (2.0 ** system_values['n_bits'])))

        # add jitter to the delay amount (which might possibly yield a negative value)
        m.set_gaussian_noise('t_jitter', std=m.jitter_rms, clk=m.emu_clk,
                             rst=m.emu_rst, lfsr_init=m.jitter_seed)
        m.bind_name('delay_amt_noisy', m.delay_amt_pre + m.t_jitter)

        # make the delay amount non-negative
        m.bind_name('delay_amt', if_(m.delay_amt_noisy >= 0.0, m.delay_amt_noisy, 0.0))

        # determine when the clock value has changed
        m.add_digital_state('clk_i_val_prev')
        m.set_next_cycle(m.clk_i_val_prev, m.clk_i_val, clk=m.emu_clk, rst=m.emu_rst)
        m.bind_name('clk_edge', m.clk_i_val ^ m.clk_i_val_prev)

        # create pointer that advances each time there is a clock edge
        depth = system_values['depth']
        dbits = int(ceil(log2(depth)))
        m.add_digital_state('addr', width=dbits)
        m.add_digital_state('next_addr', width=dbits)
        m.set_this_cycle(m.next_addr, if_(m.addr == (depth-1), 0, m.addr+1))
        m.set_next_cycle(m.addr, m.next_addr, clk=m.emu_clk, rst=m.emu_rst, ce=m.clk_edge)

        # convenience function for formatting DT signals
        def add_dt_state(*args, range_=m.emu_dt.format_.range_, width=m.emu_dt.format_.width,
                         exponent=m.emu_dt.format_.exponent, **kwargs):
            return m.add_analog_state(*args, range_=range_, width=width, exponent=exponent, **kwargs)

        # convenience function for formatting DT signals
        def dt_array(*args, real_range_hint=m.emu_dt.format_.range_, width=m.emu_dt.format_.width,
                     exponent=m.emu_dt.format_.exponent, **kwargs):
            return array(*args, real_range_hint=real_range_hint, width=width, exponent=exponent,
                         **kwargs)

        # instantiate delay "units" that each keep track of one edge

        dt_req = []
        req_data = []
        req_grant = []
        req_valid = []

        for k in range(depth):
            # should load data if there is a clock edge and this slice is selected
            load_data = m.bind_name(f'load_data_{k}', m.clk_edge & (k == m.addr))

            # handle update of dt_req
            dt_req.append(add_dt_state(f'dt_req_{k}'))
            m.set_next_cycle(
                dt_req[-1],
                if_(load_data, m.delay_amt, dt_req[-1] - m.emu_dt),
                clk=m.emu_clk,
                rst=m.emu_rst
            )

            # handle update of req_data
            req_data.append(m.add_digital_state(f'req_data_{k}'))
            m.set_next_cycle(req_data[-1], m.clk_i_val, ce=load_data,
                             clk=m.emu_clk, rst=m.emu_rst)

            # handle update of req_grant
            req_grant.append(m.bind_name(f'req_grant_{k}', dt_req[-1] == m.emu_dt))

            # handle update of req_valid
            req_valid.append(m.add_digital_state(f'req_valid_{k}'))
            m.set_next_cycle(req_valid[-1], (req_valid[-1] & (~req_grant[-1])) | load_data,
                             clk=m.emu_clk, rst=m.emu_rst)

        # replace dt_req with dt_req_max for invalid requests

        dt_req_mux = []
        for k in range(depth):
            dt_req_mux.append(
                m.bind_name(
                    f'dt_req_mux_{k}',
                    dt_array(
                        [m.dt_req_max, dt_req[k]],
                        req_valid[k],
                    )
                )
            )

        # convenience function to find the minimum DT request using a tree structure

        counter = count()
        def tree_min(data):
            # check cases
            if len(data) == 0:
                raise Exception("This shouldn't happen...")
            elif len(data) == 1:
                return data[0]
            else:
                val0 = tree_min(data[:len(data)//2])
                val1 = tree_min(data[len(data)//2:])
                return m.bind_name(
                    f'dt_intern_{next(counter)}',
                    dt_array(
                        [val0, val1],
                        val1 < val0,
                    )
                )

        # set the "dt_req" output to the minimum time request (where invalid requests
        # are replaced with "dt_req_max"

        m.set_this_cycle(m.dt_req, tree_min(dt_req_mux))

        # determine if the output should be set or cleared

        set_out = req_grant[0] & req_valid[0] & req_data[0]
        clr_out = req_grant[0] & req_valid[0] & (~req_data[0])

        for k in range(1, depth):
            set_out = set_out | (req_grant[k] & req_valid[k] & req_data[k])
            clr_out = clr_out | (req_grant[k] & req_valid[k] & (~req_data[k]))

        m.bind_name('set_out', set_out)
        m.bind_name('clr_out', clr_out)

        # set output, clear output, or keep it the same

        m.add_digital_state('clk_o_val_prev')
        m.set_next_cycle(m.clk_o_val_prev, m.clk_o_val, clk=m.emu_clk, rst=m.emu_rst)
        m.set_this_cycle(m.clk_o_val, (m.clk_o_val_prev & (~m.clr_out)) | (m.set_out))

        # generate the model
        m.compile_to_file(VerilogGenerator())

        self.generated_files = [filename]