ECE 5760: Laboratory 2

Multiprocessor Drum Synthesis.


For this exercise, you will simulate the 2D wave equation on a square mesh in realtime to produce drum-like sounds.
This year we will add a nonlinear effect related to the instantaneous tension in the mesh.
We will also compare drum sizes that can be simulated in realtime using HPS and FPGA.


Read Study Notes on Numerical Solutions of the Wave Equation with the Finite Difference Method. The main result you will need to simulate is equation 2.18. A matlab program gives a sequential version of the algorithm and plots the Fourier modes of the drum. Another version is tuned to middle C (261 Hz). You can see in the figure below that the simulated sound spectrum (blue) matches the theoretical drum modes (red) up to about mode 8 or 9 (see Physical modeling with a 2D waveguide mesh for details) . The theoretical square drum mode frequencies follow the ratio sequence:
sqrt(m+n) where m,n=1,2,3,...
Where the first term (sqrt(2)) corresponds to the fundamental mode of the drum.
The first few modes are 1, sqrt(5)/sqrt(2), 2, sqrt(10)/sqrt(2), sqrt(13)/sqrt(2), sqrt(17)/sqrt(2).

Modifying the boundary conditions, damping, wave speed, drum size, and distrubution of input energy can modifiy the sound of the simulation from drum-like, to chime-like, to gong-like or bell-like. You can modify the program further to include frequency-dependent damping and other effects. This version simluates a long, thin bar struck at one end.

Adding tension modulation allows pitch bending observed in a real drum after a large amplitude input. The large amplitude means that the membrane is stretched more, and therefore the speed of propagation (and therefore pitch) is increased. This matlab code produces an exagerated pitch effect with initial high amplitude. See also PHYSICALLY-BASED SYNTHESIS OF NONLINEAR CIRCULAR MEMBRANES equation 10.

You will probably want to read

for ideas on parallelization.

The hardware audio interface is a Wolfson WM8731 codec which is controlled by an I2C interface.

Cyclone5 handbook describing available hardware, but here is a summary.

Verilog Testing:

  1. Simulate one node with four zero-value boundary conditions .
    Result should be simple harmonic motion. Remember that rho<0.5. Start with 0.25.
    Assuming1:17 fixed point, start with initial conditions, un=un-1 and amplitude about 1/8 full scale.
    Compare to matlab solution.
  2. Get generate statement running and simulate 10x10 array for demo in first lab.
    Compare to Matlab solution.
  3. Get 10x10 array running using the audio codec on the FPGA for demo in second lab.
    You will want to refer to the Bus Master page, Audio output bus_master example.
    Also refer to the DSP page.
  4. Scale up parallel processor to 24x24 or greater. Required audio output with nonlinear rho.

HPS Testing:

  1. A good place to start for audio output from HPS is to search for Using two ARM processors with IPC to display time while writing video and playing a tone on the DE1-SoC: University Computer page. The audio code example supports a finite difference drum scheme. The drum coded is a linear system with coefficients chosen so that cheap fixed-point shifts could be used to generate the 32-bit sound samples at 48Ksamples/sec. At -O3 optimization, I could just fit a 30x30 FDTD drum into the 20.8 microsecond systhesis time frame. With nonlinear rho, the maximum size was around 25x25 grid. Example generated by HPS on a 25x25 grid, with nonlinear rho.
  2. Modify the code for the nonlinear rho effect. This will require rewriting the solver to use fixed point multiplies.
  3. Get a finite difference scheme running using the audio codec on the FPGA for demo in second lab.
    Required audio output with nonlinear rho.
  4. Scale up HPS solution until it is no longer realtime.

Student examples running on FPGA:

Weekly check points

  1. Week 1: demonstrate the PDE solver running in simulation for a 10x10 (or greater) grid to produce a waveform.
  2. Week 2: demo the PDE solver bus-master output to audio bus_slave for a 10x10 (or greater) grid to produce a sound.
    Also demo the HPS producing a nonlinear drum sound through the FPGA audio codec.
  3. Week 3: demo full FPGA system and compare to HPS.

  1. Build a realtime drum solver which produces sound from the audio codec interface.
    Minimum FPGA grid size for full credit should be bigger than 24x24, the size that can be solved by the HPS in realtime.
    The grid should have no more than 4:1 aspect ratio.
    Grid expansion via symmetry does not count for size.
  2. The solver on the FPGA should solve the 2d wave equation to produce selectable effects.
    A minimum of three buttons on the DE1-SoC should produce different timbers.
    Timber can be set by boundary condition, eta, rho, tension modulation, or number of nodes.
    At least one timber must include audible nonlinear tension modulation effects.
  3. There should be exactly one computational update of all the drum nodes for each audio sample.
    Last year the highest number of nodes was around 300,000 on DE1-SoC.
    Each sample that you calculate must be output to the audio codec.
    Each node simulated will require around 10 additions/multiplications. You may be able to use clever shifting schemes to avoid multiplys. Thus the computation rate will be about
    10*(number of wave equation nodes)*(audio sample frequency) .
    For a minimal 24x24 grid you will need ~276x106 operations/sec. Clearly some parallel processing will be necessary as you go to higher numbers of nodes.
  4. You can use fine-grained parallelism or course-grained multiprocessors on the FPGA side.
  5. Plot grid size versus compute time for at least 3 different drum sizes for both FPGA and HPS solvers.
    At least one FPGA grid size should be bigger than the size that can be solved by the HPS in realtime.
  6. Record the audio output back into matlab to show that your simulation matches drum modes (under the correct boundary conditions, etc).

Be prepared to demo your design to your TA in lab.

Your written lab report should include the sections mentioned in the policy page, and :

Copyright Cornell University February 14, 2018