Brian2cuda INFO - error question

[sivanni]

Description of problem

I am building a retina simulator and successfully running brian2cuda in WSL2 under Windows 11, and the result seems ok. However, I am getting device -associated INFOs which include error lines (see below). I would like to understand the errors before continuing.

The error lines appear when executing the following expression:
b2.device.run(
run_args={
x_input: this_x_vec,
neuron_group.HS: this_HS,
neuron_group.TS: this_TS * b2u.ms,
}
)

Additional note: I needed np.ascontiguousarray method for my 2D views of a 3D array, which was not mentioned in the brian2cuda documentation.

Minimal code to reproduce problem

Original method:
“”"
def create_dynamic_temporal_signal(
self, tvec, svec, _dt, params, show_impulse=False, impulse_contrast=1.0
):
“”"
Dynamic temporal signal for midget units
“”"
# params is [n units, n_features, n_domains] array, domains are the [center, surround].
n_units = params.shape[0]
n_timepoints = tvec.shape[-1]

    if params.ndim != 3:  # params is [n_units, n_features, n_domains]
        raise ValueError("params array must be 3D for midget units, aborting...")

    if svec.size == n_units * n_timepoints * 2:
        # Visual stimulus
        svecs = svec
    elif svec.size == n_timepoints:
        # Impulse response
        svecs = np.tile(
            np.expand_dims(svec, -1), (params.shape[0], 1, params.shape[2])
        )
    else:
        raise ValueError(
            "svec size matches neither visual simulus or impulse, aborting..."
        )

    tvec = np.expand_dims(tvec, axis=0)
    lp = np.zeros((n_units, n_timepoints, 2))

    # parameter name order for midget ["NL", "NLTL", "TS", "HS", "D", "A"]
    for domain_idx in range(2):
        NL = np.int32(params[:, 0, domain_idx])
        TL = params[:, 1, domain_idx] / params[:, 0, domain_idx]  # TL = NLTL / NL
        NL = np.expand_dims(NL, axis=-1)
        TL = np.expand_dims(TL, axis=-1)

        lp[:, :, domain_idx] = self._create_lowpass_response(tvec, NL, TL)

    lp_total = np.sum(lp, axis=(-2))
    # h_cen and h_sur are [n_units, n_timepoints]
    h_cen = lp[:, :, 0] / np.expand_dims(lp_total[:, 0], -1)
    h_sur = lp[:, :, 1] / np.expand_dims(lp_total[:, 1], -1)

    _HS = params[:, 3]
    _TS = params[:, 2]
    D = params[:, 4]

    padding_size = n_timepoints - 1

    # Pad the stimulus
    svec_padded = np.pad(svecs, ((0, 0), (padding_size, 0), (0, 0)), mode="edge")

    # Initialize the output arrays
    x_vec = np.empty((n_units, n_timepoints, 2))
    yvecs = np.zeros([n_units, n_timepoints, 2])

    for domain_idx, h in enumerate([h_cen, h_sur]):
        # Convolve the stimulus with the low-pass kernel for all units
        for this_unit in range(n_units):
            x_vec[this_unit, :, domain_idx] = fftconvolve(
                svec_padded[this_unit, :, domain_idx], h[this_unit, :], mode="valid"
            )

    b2.set_device("cuda_standalone", build_on_run=False)

    # Define the equations for the high-pass stage
    eqs = """
    dy/dt = (-y + TS * x_derivative + (1 - HS) * x_input(t, i)) / TS : 1
    x_derivative = (x_input(t, i) - x_input(t-dt, i)) / dt : Hz
    # Parameters
    HS : 1
    TS : second
    """

    _dt = _dt * b2u.ms
    simulation_duration = n_timepoints * _dt
    x_input = b2.TimedArray(np.zeros_like(x_vec[:, :, 0].T), dt=_dt)
    neuron_group = b2.NeuronGroup(n_units, eqs, method="euler", dt=_dt)
    neuron_group.HS = _HS[:, 0]
    neuron_group.TS = _TS[:, 0] * b2u.ms
    state_monitor = b2.StateMonitor(neuron_group, ["y"], record=True)
    b2.run(simulation_duration)
    b2.device.build(directory="cuda_standalone", compile=True, run=False)
    for domain_idx, h in enumerate([h_cen, h_sur]):
        this_x_vec = np.ascontiguousarray(x_vec[:, :, domain_idx].T)
        this_HS = np.ascontiguousarray(_HS[:, domain_idx])
        this_TS = np.ascontiguousarray(_TS[:, domain_idx])

        b2.device.run(
            run_args={
                x_input: this_x_vec,
                neuron_group.HS: this_HS,
                neuron_group.TS: this_TS * b2u.ms,
            }
        )

        yvecs[:, :, domain_idx] = state_monitor.y

    # Sum center and surround responses
    yvecs = np.sum(yvecs, axis=-1)
    if show_impulse is True:
        return yvecs

    # Add delay
    delay_timepoints = np.int16(np.round(D / _dt))
    generator_potential = np.zeros((n_units, n_timepoints + delay_timepoints))
    generator_potential[:, delay_timepoints:] = yvecs

    return generator_potential

“”"

What you have aready tried

The files mentioned in the error code exist under cuda_standalone\static_arrays
They are binary and non-empty

Expected output (if relevant)

No error lines (see below)

Actual output (if relevant)

“”"
INFO CUDA installation directory found in standard location: /usr/local/cuda [brian2cuda.utils.gputools]
INFO Compiling device code for GPU 0 (NVIDIA GeForce RTX 3080 Ti Laptop GPU) [brian2cuda.utils.gputools]
INFO Compiling device code for compute capability 8.6 (compiler flags: [‘-arch=sm_86’]) [brian2cuda.device]
INFO Using the following preferences for CUDA standalone: [brian2cuda.device]
INFO devices.cuda_standalone.SM_multiplier = 1 [brian2cuda.device]
INFO devices.cuda_standalone.parallel_blocks = 1 [brian2cuda.device]
INFO devices.cuda_standalone.launch_bounds = False [brian2cuda.device]
INFO devices.cuda_standalone.syn_launch_bounds = False [brian2cuda.device]
INFO devices.cuda_standalone.calc_occupancy = True [brian2cuda.device]
INFO devices.cuda_standalone.extra_threshold_kernel = True [brian2cuda.device]
INFO devices.cuda_standalone.random_number_generator_type = CURAND_RNG_PSEUDO_DEFAULT [brian2cuda.device]
INFO devices.cuda_standalone.random_number_generator_ordering = False [brian2cuda.device]
INFO devices.cuda_standalone.push_synapse_bundles = True [brian2cuda.device]
INFO devices.cuda_standalone.threads_per_synapse_bundle = {max} [brian2cuda.device]
INFO devices.cuda_standalone.bundle_threads_warp_multiple = False [brian2cuda.device]
INFO devices.cuda_standalone.no_pre_references = False [brian2cuda.device]
INFO devices.cuda_standalone.no_post_references = False [brian2cuda.device]
INFO devices.cuda_standalone.default_functions_integral_convertion = <class ‘numpy.float64’> [brian2cuda.device]
INFO devices.cuda_standalone.use_atomics = True [brian2cuda.device]
INFO devices.cuda_standalone.profile_statemonitor_copy_to_host = None [brian2cuda.device]
INFO devices.cuda_standalone.cuda_backend.gpu_heap_size = 128 [brian2cuda.device]
INFO devices.cuda_standalone.cuda_backend.detect_gpus = True [brian2cuda.device]
INFO devices.cuda_standalone.cuda_backend.gpu_id = None [brian2cuda.device]
INFO devices.cuda_standalone.cuda_backend.extra_compile_args_nvcc = [‘-w’, ‘-use_fast_math’] [brian2cuda.device]
INFO devices.cuda_standalone.cuda_backend.compute_capability = None [brian2cuda.device]
INFO devices.cuda_standalone.cuda_backend.detect_cuda = True [brian2cuda.device]
INFO devices.cuda_standalone.cuda_backend.cuda_path = None [brian2cuda.device]
INFO devices.cuda_standalone.cuda_backend.cuda_runtime_version = None [brian2cuda.device]
INFO devices.cuda_standalone.cuda_backend.device_query_path = None [brian2cuda.device]
INFO: _init_arrays() took 0.002866s
Error reading ‘static_arrays/init_neurongroup_hs_044988eab6cb3bdced31d4879772c676.dat’: file size -1 does not match expected size 1624
Error reading ‘static_arrays/init_neurongroup_ts_2a79f1fc2454b46fa1072205f0f40dae.dat’: file size -1 does not match expected size 1624
INFO _run_kernel_statemonitor_codeobject
1 blocks
768 threads
12 registers per thread
0 bytes statically-allocated shared memory per block
0 bytes local memory per thread
504 bytes user-allocated constant memory
1.000 theoretical occupancy
INFO _run_kernel_neurongroup_stateupdater_codeobject
1 blocks
640 threads
42 registers per thread
0 bytes statically-allocated shared memory per block
0 bytes local memory per thread
504 bytes user-allocated constant memory
0.833 theoretical occupancy
INFO: _init_arrays() took 0.002349s
Error reading ‘static_arrays/init_neurongroup_hs_1ba5a51a89c69ee8012b4b83610e7327.dat’: file size -1 does not match expected size 1624
Error reading ‘static_arrays/init_neurongroup_ts_11d5277e27b4bf373e1731435da7d725.dat’: file size -1 does not match expected size 1624
INFO _run_kernel_statemonitor_codeobject
1 blocks
768 threads
12 registers per thread
0 bytes statically-allocated shared memory per block
0 bytes local memory per thread
504 bytes user-allocated constant memory
1.000 theoretical occupancy
INFO _run_kernel_neurongroup_stateupdater_codeobject
1 blocks
640 threads
42 registers per thread
0 bytes statically-allocated shared memory per block
0 bytes local memory per thread
504 bytes user-allocated constant memory
0.833 theoretical occupancy
“”"

[sivanni]

Hi,

I was too hasty. The error has been solved at:

Can't modify synaptic group attribute using run_args in standalone mode

The synaptic group attribute seems not being able to be modified when using run_args in standalone mode., although I don’t see the same behavior for neuron group. The following code demonstrate this issue. Is this a feature or can be bypassed? from brian2 import * set_device('cpp_standalone', build_on_run=False) eqs = ''' dv/dt = (I-v)/tau : 1 I : 1 tau : second ''' G = NeuronGroup(3, eqs, threshold='v>1', reset='v = 0', method='exact') G.I = [2, 0, 0] G.tau = [10, 100, 100]*ms S = Synapses…

It was an issue of capitals in variable names when using device.run(run_args {CAPITALS_IN_KEY : value})

Maybe it would be useful to mention this in the run_args example documentation at:
https://brian2.readthedocs.io/en/stable/user/computation.html

cordially,
Simo

[mstimberg]

Hi @sivanni

@sivanni

Additional note: I needed np.ascontiguousarray method for my 2D views of a 3D array, which was not mentioned in the brian2cuda documentation.

Could you please give some context for this?

@sivanni

It was an issue of capitals in variable names when using device.run(run_args {CAPITALS_IN_KEY : value})

Which version of Brian2/Brian2CUDA are you using? If I am not mistaken, this specific issue should have been fixed with Brian2 2.7.0/Brian2CUDA 1.0a5.

[sivanni]

Hi,
Version for brian2 is 2.7.1, and for brian2cuda 1.0a6

About the np memoryview issue.
When I do

this_x_vec = x_vec[:, :, domain_idx].T

instead of

this_x_vec = np.ascontiguousarray(x_vec[:, :, domain_idx].T)

in the loop below (w/ some context):

        b2.set_device("cuda_standalone", build_on_run=False)
        # breakpoint()
        # Define the equations for the high-pass stage
        eqs = """
        dy/dt = (-y + ts * x_derivative + (1 - hs) * x_input(t, i)) / ts : 1
        x_derivative = (x_input(t, i) - x_input(t-dt, i)) / dt : Hz
        # Parameters
        hs : 1
        ts : second
        """

        _dt = _dt * b2u.ms
        simulation_duration = n_timepoints * _dt
        x_input = b2.TimedArray(np.zeros_like(x_vec[:, :, 0].T), dt=_dt)
        neuron_group = b2.NeuronGroup(n_units, eqs, method="euler", dt=_dt)
        neuron_group.hs = _HS[:, 0]
        neuron_group.ts = _TS[:, 0] * b2u.ms
        state_monitor = b2.StateMonitor(neuron_group, ["y"], record=True)
        b2.run(simulation_duration)
        b2.device.build(directory="cuda_standalone", compile=True, run=False)
        for domain_idx in range(2):
            this_x_vec = np.ascontiguousarray(x_vec[:, :, domain_idx].T)
            this_HS = np.ascontiguousarray(_HS[:, domain_idx])
            this_TS = np.ascontiguousarray(_TS[:, domain_idx])

            b2.device.run(
                run_args={
                    x_input: this_x_vec,
                    neuron_group.hs: this_HS,
                    neuron_group.ts: this_TS * b2u.ms,
                }
            )

            yvecs[:, :, domain_idx] = state_monitor.y

I get the following error:

Traceback (most recent call last):
  File "/opt2/Git_Repos/MacaqueRetina/macaqueretina/project/project_conf_module.py", line 923, in <module>
    PM.simulate_retina.client(impulse=True)
  File "/opt2/Git_Repos/MacaqueRetina/macaqueretina/retina/simulate_retina_module.py", line 3580, in client
    director.run_impulse_response(contrasts)
  File "/opt2/Git_Repos/MacaqueRetina/macaqueretina/retina/simulate_retina_module.py", line 931, in run_impulse_response
    self.builder.get_impulse_response(contrasts)
  File "/opt2/Git_Repos/MacaqueRetina/macaqueretina/retina/simulate_retina_module.py", line 877, in get_impulse_response
    impulse_to_show = self.temporal_model.impulse_response(vs, gcs, contrasts)
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/opt2/Git_Repos/MacaqueRetina/macaqueretina/retina/simulate_retina_module.py", line 645, in impulse_response
    yvec = self.ganglion_cell.create_dynamic_temporal_signal(
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/opt2/Git_Repos/MacaqueRetina/macaqueretina/retina/simulate_retina_module.py", line 299, in create_dynamic_temporal_signal
    b2.device.run(
  File "/home/simo/.cache/pypoetry/virtualenvs/macaqueretina-mGLhXIth-py3.11/lib/python3.11/site-packages/brian2/devices/cpp_standalone/device.py", line 1264, in run
    ) = self._prepare_timed_array_run_arg(key, value)
        ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/simo/.cache/pypoetry/virtualenvs/macaqueretina-mGLhXIth-py3.11/lib/python3.11/site-packages/brian2/devices/cpp_standalone/device.py", line 1395, in _prepare_timed_array_run_arg
    value_name = f"init_{key.name}_values_{md5(value_ar.data).hexdigest()}.dat"
                                           ^^^^^^^^^^^^^^^^^^
BufferError: memoryview: underlying buffer is not C-contiguous

[mstimberg]

Thanks for the additional info, I will look into this in more detail soon.

[mstimberg]

OK, regarding the run_args with uppercase letters bug: this was fixed in Brian 2.7.0, but only for cpp_standalone mode – it is still present in Brian2CUDA

Regarding the contiguous array: many thanks for making us aware of this, I did not realize that the hash functions require a contiguous array. This should be easy to fix in Brian itself, but until then you can of course use your workaround. Note that your current code requires copying the array content – if you can re-arrange your array so that you can use x_vec[domain_idx, :, :] this would avoid this copy (by already being contiguous).