15-347/18-347 Spring '97
Notes on Assignments and Exams

Homework 4

• The xanim program is available on most of the campus Unix machines.
  • You'll want to work directly from a workstation rather than operate remotely with a server machine.
  • The SUN's in the Wean Hall cluster work well, but they seem to have their sound disabled. For the parts of the assignment where you need to check the sound quality, you'll have to move to a different machine.
  • Xanim runs very nicely on the Wean Hall SGI machines
• The SGI program movieplayer. This is a nice program, but it doesn't generate some of the quantitative information this assignment requires. You can get some useful information with the program mediaconverter.
• The Macintosh program Movie Player. Again, it gives a very good display, but not enough quantitative information.
• For those of you with a PowerPC Macintosh, try installing the program MovieInfo Pro that I downloaded from the Internet. You can do a frame-by-frame analysis of the video. The documentation is in Japanese, but the program interface is in English.
• There should be comparable programs on most other platforms.

Exam 2 (postscript) (solution)

Grade histogram for Exam 2

Homework 3

Lab 2

• Fix your bugs. Feel free to run the tests in HOME347/public/labs/lab2/microtests. You should not get any error messages, and the program should halt on "break 1".
• Turn in your code by copying the file to the directory: /afs/andrew.cmu.edu/scs/cs/15-347/lab2-new-handins
• Name the file as before: NAME-stages.c, where NAME is the same one you used on your first handin
• You have until 3:00pm Thursday, April 10 to do this.

• Rerun our tests
• Derive a "correctness measure" Cnew based on the same method as before. That is, for each of the two modes, there are up to 25 points of correctness, but you lose 1 point for each failing test. So, Cnew can range from 0 to 50.
• Suppose your old correctness score was Cold (also between 0 and 50).
• We will add to your original score 0.75 * (Cnew - Cold) points.
• For example, if you had Cold = 0 and Cnew = 50, you'll get 37.5 points.

Important Announcement: The due date for this assignment has been extended to Tuesday, March 11.

The documentation for the assignment is available in postscript. The files you need for Lab2 are available as HOME347//public/labs/lab2/files.tar
You can try out a working version of the simulator solve_tk in the directory HOME347public/sim/. Included in this directory are subdirectories containing tests cases:

• demos containing the demonstration examples from the Feb. 20 class.
• exc containing the demonstration examples for the Feb. 25 class on exception handing.
• tests containing the demonstration examples for the Feb. 25 class on exception handing.

Clarifications

• In the actual MIPS instruction set, the jalr has two register operands: one to specify the jump destination, and the other is the return pointer. In our simulator, we will assume the return pointer is always $31.

Some advice: This lab is going to take time. If you try to start the night before, you'll find yourselves in deep trouble. In particular, note that the quality of your testing procedure (task 4) is going to count a significant fraction (around 30%) of the total. If you're still trying to understand C code and add forwarding the night before the assignment is due, you're going to have to blow off this important part of the assignment.

Here's my advice on how to proceed:

• Try running the executable version on the class demos (demos/demoX.O) and the supplied test code (tests/XXX.O). Make sure you understand what is happening, especially how the various hazards are handled in both stall and forwarding mode.
• Study the code in stages.c carefully. It closely parallels the pipeline organization presented in Ch 6 of the text and in lecture.
• Implement the jump instructions. You'll need to think through how to compute the target address and how to deal with register 31 (for jal and jalr). Test them on code examples that don't have any hazard problems.
• Implement stall mode. It requires only adding some code to the procedure do_stall_check.
• Implement forwarding mode. This involves changes to do_ex_stage, do_mem_stage, and do_stall_check.
• See what kind of fault coverage you're getting from the supplied tests. For data hazards, you can be systematic, filling in the matrix shown in class (Lect. 10, Slide 24). It's hard to be systematic about control hazards. But think through different aspects of branches and jumps that need to be tested.
• Devise a comprehensive test set, based on your analysis from 6.

• It's OK to add more "wiring", multiplexors, and control logic.
• It's not OK to add more ALUs or adders.
• You must obey the register usage constraints: only one stage can set the next state of a pipe register, but many can read the current state.
• You must not introduce "hidden" state into your program in the form of static global or static variables.

Exam 1 (postscript) (solution)

Grade histogram for Exam 1

Lab 1

Disassembled code for #5

If your code did get cc to do software pipelining, and you've already written the C code to match it, don't do it again. However, be sure you indicate clearly that you used your disassembled code and not ours, so that we'll grade it properly.

Revised load/store count estimates

On a DECstation (I used ride.ww), you can eliminate...
problem #     outer loop       inner loop
  2            2 L, 0 S         3 L, 0 S
  3            0 L, 0 S         3 L, 2 S
  4            1 L, 0 S         0 L, 0 S


On an SGI (I used dakota.weh), you can eliminate...
problem #     outer loop       inner loop
  2            3 L, 0 S         3 L, 0 S
  3            0 L, 1 S         2 L, 1 S
  4            1 L, 0 S         0 L, 0 S

The abbreviation "x L, y S" means x loads, y stores.

BEQL and BNEL instructions

	BEQL (Branch On Equal Likely)
	format:  BEQL  rs, rt, target

	Compares the contents of registers rs and rt, and if these
	are equal, branches to the specified target, with a delay of
	one instruction.  If the conditional branch is not taken,
	the instruction in the delay slot is nullified.

There is a corresponding BNEL (Branch on Not Equal Likely) instruction, which treats the delay slot the same way as BEQL.

You might want to think about the reasons why instructions like these would have been added to the MIPS instruction set.

Ari

Homework 2

• integer registers: $0, $1, $2, ..., etc.
• floating point registers: $f0, $f1, $f2, ..., etc.
• floating point status register: (probably won't be genereated)

• integer registers: r1, r2, ..., etc.
• floating point registers: f1, f2, ..., etc.
• floating point status register: $31

Homework 1

• |t_{2n}^{'} - t_{n}^{'} | / t_{n}^{'} <= E

• |t_{2n}^{'} - 2t_{n}^{'} | / t_{2n}^{'} <= E, for t_{2n}^{'} > 0

Sorry about the error.
droh