Question

Assume we want to execute the DAXPY loop show on page 511 in MIPS assembly on the NVIDIA 8800 GTX GPU described in this chapter. In this problem, we will assume that all math operations are performed on single-precision floating point numbers (we will rename the loop SAXPY). Assume that instructions take the following number of cycles to execute.

Loads	Stores	Add.S	Mult.S
5	2	3	4

6.13.1Describe how you will constructs warps for the SAXPY loop to exploit the 8 cores provided in a single multiprocessor.

Short answer

6.13.1

The first step is to characterize the DAXPY loop for the core values that can be calculated with the scaling and adding the intrinsic functions for each statement.

Loop:

daxpy (int saxID, double *p,double *q,*r, double b)

saxSize=SIZE/N

saxS=chuck_id*saxSize

saxE=saxS+saxSize

for(i=saxS;i<sax;i++)

r[i]=b*p[i]+q[i]

Step-by-step solution

Determine DAXPY LOOP

When the MIPS instruction set architecture is extended with vector instructions and the vector registers, vector operations use same names as MIPS with “V” appended to it. The input for vector instructions is taken by a vector register of scalar register.

$Y=a\times X+Y$

In the above expression, X and Y are vectors of 64 double precision floating point numbers initially resident in memory and “a” is a scalar double precision variable. This is called DAXPY loop, stands for double precision$a\times XplusY$ .

Determine the SAXPY Loop code

6.13.1

The first step is to characterize the DAXPY loop for the core values that can be calculated with the scaling and adding the intrinsic functions for each statement. Then the damp loop for exploiting the subroutines which will fetch the default library for each individual processor and implements the large performance routines. The loop is given below.

Loop:

daxpy (int saxID, double *p,double *q,*r, double b)

saxSize=SIZE/N

saxS=chuck_id*saxSize

saxE=saxS+saxSize

for(i=saxS;i<sax;i++)

r[i]=b*p[i]+q[i]