libMHCUDA 2.3.0

Accelerated Weighted MinHash-ing on GPU

MinHashCuda

This project is the reimplementation of Weighted MinHash calculation from ekzhu/datasketch in NVIDIA CUDA and thus brings 600-1000x speedup over numpy with MKL (Titan X 2016 vs 12-core Xeon E5-1650). It supports running on multiple GPUs to be even faster, e.g., processing 10Mx12M matrix with sparsity 0.0014 takes 40 minutes using two Titan Xs. The produced results are bit-to-bit identical to the reference implementation. Read the article.

The input format is 32-bit float CSR matrix. The code is optimized for low memory consumption and speed.

What is Weighted MinHash

MinHash can be used to compress unweighted set or binary vector, and estimate unweighted Jaccard similarity. It is possible to modify MinHash for weighted Jaccard by expanding each item (or dimension) by its weight. However this approach does not support real number weights, and doing so can be very expensive if the weights are very large. Weighted MinHash is created by Sergey Ioffe, and its performance does not depend on the weights - as long as the universe of all possible items (or dimension for vectors) is known. This makes it unsuitable for stream processing, when the knowledge of unseen items cannot be assumed.

Building

cmake -DCMAKE_BUILD_TYPE=Release . && make

It requires cudart, curand >=8.0, OpenMP 4.0 compatible compiler (that is, not gcc <=4.8) and cmake >= 3.2. If numpy headers are not found, specify the includes path with defining NUMPY_INCLUDES. If you do not want to build the Python native module, add -D DISABLE_PYTHON=y. If CUDA is not automatically found, add -D CUDA_TOOLKIT_ROOT_DIR=/usr/local/cuda-8.0 (change the path to the actual one). If you are building in a Docker container you may encounter the following error: Could NOT find CUDA (missing: CUDA_TOOLKIT_ROOT_DIR CUDA_INCLUDE_DIRS CUDA_CUDART_LIBRARY) This means you need to install the rest of the CUDA toolkit, which can be installed like in the nvidia/cuda:8.0-devrel Dockerfile. If you still run into Could NOT find CUDA (missing: CUDA_INCLUDE_DIRS) then run: ln -s /usr/local/cuda/targets/x86_64-linux/include/* /usr/local/cuda/include/

Python users: if you are using Linux x86-64 and CUDA 8.0, then you can install this easily:

pip install libMHCUDA

Otherwise, you'll have to install it from source:

pip install git+https://github.com/src-d/minhashcuda.git

Building in Python virtual environments, e.g. pyenv or conda is officially not supported. You can still submit patches to fix the related problems.

Testing

test.py contains the unit tests based on unittest. They require datasketch and scipy.

Contributions

...are welcome! See CONTRIBUTING and code of conduct.

License

Apache 2.0

Python example

import libMHCUDA
import numpy
from scipy.sparse import csr_matrix

# Prepare the rows
numpy.random.seed(1)
data = numpy.random.randint(0, 100, (6400, 130))
mask = numpy.random.randint(0, 5, data.shape)
data *= (mask >= 4)
del mask
m = csr_matrix(data, dtype=numpy.float32)
del data

# We've got 80% sparse matrix 6400 x 130
# Initialize the hasher aka "generator" with 128 hash samples for every row
gen = libMHCUDA.minhash_cuda_init(m.shape[-1], 128, seed=1, verbosity=1)

# Calculate the hashes. Can be executed several times with different number of rows
hashes = libMHCUDA.minhash_cuda_calc(gen, m)

# Free the resources
libMHCUDA.minhash_cuda_fini(gen)

The functions can be easily wrapped into a class (not included).

Python API

Import "libMHCUDA".

def minhash_cuda_init(dim, samples, seed=time(), deferred=False, devices=0, verbosity=0)

Creates the hasher.

dim integer, the number of dimensions in the input. In other words, length of each weight vector. Must be less than 2³².

samples integer, the number of hash samples. The more the value, the more precise are the estimates, but the larger the hash size and the longer to calculate (linear). Must not be prime for performance considerations and less than 2¹⁶.

seed integer, the random generator seed for reproducible results.

deferred boolean, if True, disables the initialization of WMH parameters with random numbers. In that case, the user is expected to call minhash_cuda_assign_random_vars() afterwards.

devices integer, bitwise OR-ed CUDA device indices, e.g. 1 means first device, 2 means second device, 3 means using first and second device. Special value 0 enables all available devices. Default value is 0.

verbosity integer, 0 means complete silence, 1 means mere progress logging, 2 means lots of output.

return integer, pointer to generator struct (opaque).

def minhash_cuda_calc(gen, matrix, row_start=0, row_finish=0xffffffff)

Calculates Weighted MinHash-es. May reallocate memory on GPU but does it's best to reuse the buffers.

gen integer, pointer to generator struct obtained from init().

matrix scipy.sparse.csr_matrix instance, the number of columns must match dim. The number of rows must be less than 2³¹.

row_start integer, slice start offset (the index of the first row to process). Enables efficient zero-copy sparse matrix slicing.

row_finish integer, slice finish offset (the index of the row after the last one to process). The resulting matrix row slice is [row-start:row_finish].

return numpy.ndarray of shape (number of matrix rows, samples, 2) and dtype uint32.

def minhash_cuda_fini(gen)

Disposes any resources allocated by init() and subsequent calc()-s. Generator pointer is invalidated.

gen integer, pointer to generator struct obtained from init().

C API

Include "minhashcuda.h".

MinhashCudaGenerator* mhcuda_init(
    uint32_t dim, uint16_t samples, uint32_t seed, int deferred,
    uint32_t devices, int verbosity, MHCUDAResult *status)

Initializes the Weighted MinHash generator.

dim the number of dimensions in the input. In other words, length of each weight vector.

samples he number of hash samples. The more the value, the more precise are the estimates, but the larger the hash size and the longer to calculate (linear). Must not be prime for performance considerations.

seed the random generator seed for reproducible results.

deferred if set to anything except 0, disables the initialization of WMH parameters with random numbers. In that case, the user is expected to call mhcuda_assign_random_vars() afterwards.

devices bitwise OR-ed CUDA device indices, e.g. 1 means first device, 2 means second device, 3 means using first and second device. Special value 0 enables all available devices.

verbosity 0 means complete silence, 1 means mere progress logging, 2 means lots of output.

status pointer to the reported return code. May be nullptr. In case of any error, the returned result is nullptr and the code is stored into *status (with nullptr check).

return pointer to the allocated generator opaque struct.

MHCUDAResult mhcuda_calc(
    const MinhashCudaGenerator *gen, const float *weights,
    const uint32_t *cols, const uint32_t *rows, uint32_t length,
    uint32_t *output)

Calculates the Weighted MinHash-es for the specified CSR matrix.

gen pointer to the generator opaque struct obtained from mhcuda_init(). weights sparse matrix's values. cols sparse matrix's column indices, must be the same size as weights. rows sparse matrix's row indices. The first element is always 0, the last is effectively the size of weights and cols. length the number of rows. "rows" argument must have the size (rows + 1) because of the leading 0. output resulting hashes array of size rows x samples x 2.

return the status code.

MHCUDAResult mhcuda_fini(MinhashCudaGenerator *gen);

Frees any resources allocated by mhcuda_init() and mhcuda_calc(), including device buffers. Generator pointer is invalidated.

gen pointer to the generator opaque struct obtained from mhcuda_init().

return the status code.

README {#ignore_this_doxygen_anchor}