LSU EE 7700-2 - What's New

Final Exam Question Hint: The answer to Problem 2 applies to CC 1.X as well as CC 2.x GPUs. In other words, just think about CC 1.X GPUs when answering the question.

Take-Home Final assigned. Please start working on the take-home final. As mentioned in class it is based on Lee 2010, the "Debunking" paper. Please feel free to send me questions. Answers to some questions will be posted here so check back. A completed exam is due Saturday morning 14 May. Good luck! Note that I'm out of town now and I won't be back until Thursday.

Linked the radix sort and manycore notes to the lecture notes page.

The paper to read for the next exam, Debunking.., has been linked to the references page. The paper should be freely accessible within LSU, elsewhere use the user id "ee4720" and the password given in class. (A good warning to kids so they don't stick there hands in the cage near basketball stadium.)
In class a student asked about performance comparisons between Knights Ferry (called Larrabee by you-know-who) and NVIDIA GPUs. The answer was yes, an Intel Whitepaper does a detailed comparison of the performance of a radix sort on GPU and Knights Ferry. In fact, some of the authors also wrote the sorting paper assigned in class.

Updated the NVIDIA GPU Microarchitecture slides to include the additional material scheduling and Fermi (CC 2.X).

Updated the NVIDIA GPU Microarchitecture slides to include the material on branch divergence.

Here is a draft of the radix sort analysis shown in class today (with a few additions). A more complete version will be posted tomorrow or Friday.

Here is a reference for a basic parallel sum and prefix sum. The parallel sum is on page 1172 (3) and is headed "Sum of an Array of Numbers". The principal author of the paper, Daniel Hillis, is a co-founder of Thinking Machines, maker of the Connection Machine, an early parallel computer design. Remember that these basic parallel techniques are only the starting point for efficient GPU implementation.

Updated the solution to Homework 2, including the version linked here. Also posted the Homework 3 assignment but expect an E-mail message providing details.

Checked in a solution to Homework 2 Problem 3a to the svn repository. For your convenience, here is the kernel. Extensive comments have been added to the stencil_shared routine. The solution appears in stencil_shared_2, however it is not unrolled and it does not access memory efficiently. Unroll-friendly code will be put in another routine. For Homework 3 fix the global and shared memory problems with stencil_shared_2. I'll post more detailed instructions by Monday.

Linked NVIDIA GPU Microarchitecture notes used in todays class to the lecture notes page. They are under construction, please report errors or requested clarifications and additions.

Homework 2 deadline extended will be extended to Monday if at least two students ask non-trivial questions about the assignment in Friday's class or earlier.

Assigned Homework 2, due Friday 11 March 2011.

Linked Homework 1 solution to the previous work page. Also checked in solution code to the svn repository, do an "svn update" in your homework directory to get it. The solution is in new files with names starting "solution-".

Those getting errors about missing files on Homework 1 when issuing the "make" command should update their homework code:

 cd ~/hw1
 svn update

This will modifiy the makefile but won't change other files.

Homework 1 assigned, due Monday, 28 February 2011. ECE computer accounts are needed for this assignment they will be given out in class, or E-mail me for an account.

Updated the account setup instructions in the procedures page. I still haven't assigned anything but students are free to follow them. E-mail me if you need a class account.

Please look at the description of the NVIDIA 3-series GPUs in this paper. (Other GPU descriptions can be found in the GPU descriptions page.) Don't worry if the paper seems difficult, we'll go over the important points in class.