FFT Accelerator Project Rohit Prakash Anand Silodia Date: June 7 th, 2007.

FFT Accelerator ProjectRohit PrakashAnand Silodia

Date: June 7th , 2007

Objectives

• Analysis using random input points

• %age improvement (from the previous implementations)

• Cache profiling

Improvements

• Calls to sine/cosine decreased• Separate arrays for power, some

other terms– Division decreased– Multiplications decreased

• Error in last time corrected (FFTW floating point)

System Configuration

• Intel Pentium 4 (HT) 3.0Ghz• RAM : 1GB• Cache : 1MB L2• O.S. : Fedora Core 3• Compiler icc• Flags used : -xW, -O3, -ipo-prec-

div, -static

User time : vs. FFTW (single precision)

Radix-4 works 1.5 times slower than fftw

Radix-8 works 1.6 times slower than fftw

User time : previous (double) vs. new (float)

Approximately 20% improvement

User time : previous (double) vs new (float)

Approximately 19% improvement

Cache Organization

Cache Level

Size Associativity

Line size

L2 1 MB 8-way 64

I1 16 KB 4-way 64

D1 16KB 4-way 64

Radix-4 L2 misses

Approximately 30% less L2 misses

Radix-4 D1 misses

Approximately 1.6% less D1 misses

Radix-8 L2 misses

Approximately 13.6% less L2 misses

Radix-8 D1 misses

Approximately .96% less D1 misses

Profiling results: using vtune

Profiling results: using gprof

Profiling results : using vtune

Profiling results: using gprof

Profiling results: using vtune

Profiling results: using gprof

Profiling results: using vtune

Profiling results: using gprof

Profiling results: using vtune

Profiling results: using vtune

Profiling results: using vtune

Profiling results: using vtune

Profiling results: using vtune

Profiling results: using vtune

Profiling results: using gprof

Profiling results: using gprof

Further Improvements : use sse instructions• Vectorize the loop

TA[r]Uw*A[r+p]Vw*w*A[r+2*p]Ww*w*w*A[r+3*p]----------------------------------Complex temp[4];For(i = 1; i<4;i++){

temp[i] = twiddle[i*p]*A[r+ i*l]

}

Thank You