Category Archive

Introduction Zero Copy access has been a part of the CUDA Toolkit for a long time (~2009). However, there were very few applications using the capability because, with time, GPU memory has become reasonably large. The only applications using zero copy access were mainly ones with extremely large memory requirements, such as database processing. Zero Copy is a way to map host memory and access it directly over PCIe without doing an explicit memory transfer. It allows CUDA kernels to directly access host memory. Instead of reading data from global memory (limit ~200GB/s), data would be read over PCIe and be limited by PCIe bandwidth (upto 16GB/s). Hence there was no real performance advantage for most applications. However, with the …

Triangle counting is a building block for many social analytics, especially for the widely used clustering coefficient analytic. Clustering coefficients is used for finding key players in a network based on their local connectivity, which is expressed based on the number of triangles that they belong to. There are many ways for finding the triangles a vertex may belong to. One of the most popular approaches for finding the triangles is to do an intersection of two adjacency lists. If a pair of vertices (u,v) have any common neighbors, these will be found in the intersection process. While the above may sound simple, implementing an efficient intersection on the GPU is not trivial or straightforward, it requires smart partitioning of the work to make full use …

About a month ago, one of my colleagues did a post on how to author the most concise OpenCL program using the C++ API provided by Khronos. In today’s post, we shall further modify that example to achieve the following two goals. Enable the kernel to work with different integral data types out of the box Ensure that the kernels compile only once at run time per data type Let’s dive into the details now. We can template the OpenCL kernels by passing a build option -D T=”typename” to the kernel compilation step. To pass such options, we would need a construct that can give us a string literal that represents the corresponding integral type. Let us declare a struct …

In this blog we will review the benefit of using element-wise operations for your computations. Element-wise operations are operations that are applied to every element in an array and allow the user to avoid coding loops and nested loops for rudimentary operations. In a simple example of an element-wise operation, we use both the addition (+) and multiply (*)operations: array A = randu(1024, 1024), B = randu(1024, 1024); array C=A*A+B*B; An element-wise operator that is applied to a single element is unary operator. An operator that works on two elements is a binary operator. ArrayFire has implemented a large number element-wise operations that are applied to the elements of an array. These operators can help reduce the programming overhead for an application designer as: Performance – …