CS 179: GPU Computing

Recitation 1 - 4/1/16
Recap
● Device (GPU) runs CUDA kernel defined in .cu and .cuh files
- C++ code with a few extensions
- Compiled with proprietary NVCC compiler
- Kernel defines the behavior of each GPU thread
● Program control flow managed by host (CPU)
- Uses CUDA API calls to allocate GPU memory, and copy input data from
host RAM to device RAM
- In charge of calling kernel - (almost) like any other function
- Must also copy output data back from device to host
- Executable is ultimately C++ program compiled by G++
- Doesn’t treat object files (.o) produced by NVCC any differently
Recap
● GPU hardware abstraction consists of a grid of blocks of threads
- Grid and blocks can have up to three dimensions
- Each block assigned to an independent streaming multiprocessor (SM)
- SM divides blocks into warps of 32 threads
- All threads in a warp execute the same instruction concurrently
- Warp divergence occurs when threads must wait to execute different
instructions
- GPUs are slow - waiting adds up fast!
● Parallelizable problems can be broken into independent components
- Want to assign one thread per “thing that needs to get done”
- Even better if threads in a warp don’t diverge
Parallelizable Problems
● Most obvious example is adding two linear arrays
- CPU code:

- Need to allocate a, b, c and populate a, b beforehand

- (But you should know how to do that)
● Why is this parallelizable?
- For i ≠ j, operations for c[i] and c[j] don’t have any interdependence
- Could happen in any order
- Thus we can do them all at the same time (!) with the right hardware
Non-Parallelizable Problems
● Potentially harder to recognize
● Consider computing a moving average
- Input array x of n data points
- Output array y of n averages
- Two well-known options:
- Simple
- Exponential
● Simple method just weights all data points so far equally
- CPU code:
- Parallelizable? Yes!
- y[i] values separate
Non-Parallelizable Problems
● What about an exponential moving average?
- Uses a recurrence relation to decay point weight by a factor of 0 < 1 - c < 1
- Specifically, y[i] = c · x[i] + (1 - c) · y[i - 1]
- Thus y[n] = c · (x[n] + (1 - c) · x[n - 1] + … + (1 - c)n - 1 · x[1]) + (1 - c)n · x[0]
- CPU code:

- Parallelizable? Nope
- Need to know y[i] before calculating y[i + 1]
What Have We Learned?
● Not all problems are parallelizable
- Even similar-looking ones
● Harnessing the GPU’s power requires algorithms whose computations can be
done at the same time
- “Parallel execution”
- Opposite would be “serial execution,” CPU-style
● Output elements should probably not rely on one another
- Would require multiple kernel calls to compute otherwise
- Different blocks of threads can’t wait for each other, more on that
later in the course
- In addition to all the extra instructions, there’s a lot of overhead
Assignment 1: Small-Kernel Convolution
● First assignment involves manipulating an input signal
- In particular, a WAV (.wav) audio file
- We provide an example to test with
- Using audio will require libsndfile
- Installation instructions included in assignment
- Code also includes an option to use random data instead
● C++ and CUDA files provided, your job to fill in TODOs
- Code already includes CPU implementation of desired algorithm
- Your job is to write the equivalent CUDA kernel to parallelize it
- You’re also in charge of memory allocation and host-device data transfers
● Conceptually straightforward, goal is familiarity with integrating CUDA into C++
Some Background on Signals
● A system takes input signal(s), produces output signal(s)
● A signal can be a continuous or discrete stream of data
- Typically characterized by amplitude
- E.g. continuous acoustic input to a microphone
● A continuous signal can also be discretized
- Simply sample it at discrete intervals
- Ideally periodic in nature
- E.g. voltage waveform microphone output
● We will consider discrete signals
- Assignment uses two-channel audio
Linear Systems
● Suppose some system takes input signal xi[n] and produces output signal yi[n]
- We denote this as xi[n] → yi[n]
● If the system is linear, then for constants a, b we have:
- a · x1[n] + b · x2[n] → a · y1[n] + b · y2[n]
● Now suppose we want to pick out a single point in the signal
- We can do this with a delta function,
- If we treat it as a discrete signal, we can define it as:
- [n - k] = 1 if n = k, [n - k] = 0 if n ≠ k
- “Zero everywhere with a spike at k”
● This definition means that x[k] = x[n] · [n - k]
- Note: I was wrong about this in recitation. We use the delta function to
pick out the value of signal x[n] at constant point k.
Linear Systems
● Next we can define a system’s response to [n - k] as hk[n]
- I.e. [n - k] → hk[n]
● From linearity we then have x[n] · [n - k] → x[n] · hk[n]
- x[n] is the input signal, [n - k] is the delta function signal
- Note: I was wrong about this in recitation; see the previous slide for
details.
- Response at time k defined by response to delta function
Time-Invariance
● If a system is time-invariant, then it will satisfy:
- x[n] → y[n] ⇒ x[n + m] → y[n + m] for integer m
● Thus given [n - k] → hk[n] and [n - l] → hl[n], we can say that hk[n] and hl[n]
are “time-shifted” versions of each other
- Instead of a new response hk[n] for each k, we can define h[n] such that
[n] → h[n], and shift h with k such that [n - k] → h[n - k]
- By linearity, we then have x[n] · [n - k] → x[n] · hk[n]
● This lets us rewrite the system’s response x[n] → y[n]:
- x[n] = Σ x[k] · [n - k] → Σ x[k] · hk[n - k] = x[k] · hk[n] = y[n]
- Output must be equivalent to y[n] because x[n] → y[n]
- Note: sum is over all k.
What Have We Learned?
● Linear time-invariant systems have some very convenient properties
- Most importantly, they can be characterized entirely by h[n]
- This allows y[n] to be written entirely in terms of the input samples x[k]
and the delta function response h[n]
● Remember:
- y[n] = Σ x[k] · hk[n - k]
- x[n] is the input signal to our system
- y[n] is the output signal, or “impulse response” from our system
- [n] is the delta function signal
- h[n] is the impulse response from our system for [n]
Putting It All Together
● Assignment asks you to accelerate convolution of an input signal
- E.g. input x[0..99], system with h[0..3] delta function response
- For finite-duration h such as this, computable with y[n] = Σ x[k] · h[n - k]
- y[50] computation, for example, would be:
- y[50] = x[47] · h[3] + x[48] · h[2] + x[49] · h[1] + x[50] · h[0]
- All other h terms are 0
- Here y[50] etc. refer to the signal at that point
● This sum is parallelizable
- Pseudocode:
Assignment Details
● All you need to worry about is the kernel and memory operations
● We provide the skeleton and some useful tools
- CPU implementation - reference this for your GPU version
- Error checking code for your output
- Delta function response h[n] (default is Gaussian impulse response)
- Note: I was wrong in the recitation, saying that h[n] is the response to
any function we wish to convolve. Rather, the system is defined such
that its response to the delta function is the signal we to convolve.
- This derivation is a discrete-time version of https://en.wikipedia.
org/wiki/LTI_system_theory#Impulse_response_and_convolution.
Looking at this will help distinguish when we refer to the signal as a
function and when we refer to a specific point in it.
Assignment Details
● Code can be compiled in one of two modes
- Normal mode (AUDIO_ON defined to be 0)
- Generates random x[n]
- Can test performance on various input lengths
- Can run repeated trials by increasing number of channels
- Audio mode (AUDIO_ON defined to be 1)
- Reads x[n] from input WAV file
- Generates output WAV from y[n]
- Gaussian h[n] is an (imperfect) low-pass filter - high frequencies
should be attenuated
Debugging Tips
● printf() can be useful, but gets messy if all threads print
- Better to only print from certain threads, though your kernel will diverge
● If you want to check your kernel’s output, copy it back to the host
- More manageable than printing from the kernel and you can write normal
C++ to inspect the data
● Use the gpuErrchk() macro to check CUDA API calls for errors
- Example usage: gpuErrchk(cudaMalloc(&dev_in, length * sizeof (int)));
- Prints error info to stderr and exits
● Use small convolution test cases before trying large arrays or the test WAV
- E.g. 5-element x[n], 3-element h[n]
Any Questions?