CUDALucas (a.k.a. MaclucasFFTW/CUDA 2.3/CUFFTW)

[msft]

Hi,

New version on GTX260.

This version use Matrix transpose with Cuda(see NVIDIA_GPU_Computing_SDK/C/src/transpose).

This version support only 4096k.

MacLucasFFTW.cuda.t.tar.gz

$ time ./MacLucasFFTW 216091
Iteration 10000 M( 216091 )C, 0x30247786758b8792, n = 4194304, MacLucasFFTW v8.1 Ballester
...
Iteration 210000 M( 216091 )C, 0xcfe091c8f59f8a7b, n = 4194304, MacLucasFFTW v8.1 Ballester
M( 216091 )P, n = 4194304, MacLucasFFTW v8.1 Ballester

real 99m5.087s
user 42m45.096s
sys 0m0.272s

$ time ./MacLucasFFTW 63333333
Iteration 10000 M( 63333333 )C, 0xa5d7a917d728239a, n = 4194304, MacLucasFFTW v8.1 Ballester
M( 63333333 )C, 0xa5d7a917d728239a, n = 4194304, MacLucasFFTW v8.1 Ballester

real 4m20.656s
user 1m41.750s
sys 0m0.112s

4096k fft sec/iter = 0.026

Thank you,

[frmky]

I just retested the 4096K FFT for the recent versions on the C1060. Version k is the fastest.

Version k: 0.0264 sec/iter
Version o: 0.0278 sec/iter
Version t: 0.0324 sec/iter

[msft]

Hi, frmky

Originally Posted by frmky

I just retested the 4096K FFT for the recent versions on the C1060.
Version k is the fastest.

Version k: 0.0264 sec/iter
Version o: 0.0278 sec/iter
Version t: 0.0324 sec/iter

If you can execute version "t" with "CUDA Visual Profiler", tell me .csv result.

Arguments 63333333
Max.Execution Time:30

Thank you,

[frmky]

Attached is the csv. I'll upload the graph in the next post.

[frmky]

Here's the graph:

[msft]

Originally Posted by frmky

Here's the graph:

Thank you, frmky
transpose_kernel copt from NVIDIA_GPU_Computing_SDK/C/src/transpose/transpose_kernel.cu ,I don't touch.
mmm...

[frmky]

Not sure if it'll help, but...

[childers release]$ ./transpose 
Transposing a 256 by 4096 matrix of floats...
Naive transpose average time:     2.286 ms
Optimized transpose average time: 0.327 ms

Test PASSED

Press ENTER to exit...

[msft]

$ ./transpose
Transposing a 256 by 4096 matrix of floats...
Naive transpose average time:     1.281 ms
Optimized transpose average time: 0.184 ms

C1060 have 512 bank something wrong.

[frmky]

Try using transposeDiagonal from transposeNew.cu.

[TransposeNew]
> Device 0: "Tesla C1060"
> SM Capability 1.3 detected:
> CUDA device has 30 Multi-Processors
> SM performance scaling factor = 1.00

Matrix size: 2048x2048 (64x64 tiles), tile size: 32x32, block size: 32x8

Kernel                  Loop over kernel        Loop within kernel
------                  ----------------        ------------------
simple copy             73.02 GB/s              74.76 GB/s
shared memory copy      70.46 GB/s              71.84 GB/s
naive transpose          2.13 GB/s               2.14 GB/s
coalesced transpose     17.53 GB/s              18.25 GB/s
no bank conflict trans  17.70 GB/s              18.34 GB/s
coarse-grained          17.70 GB/s              18.33 GB/s
fine-grained            69.51 GB/s              72.12 GB/s
diagonal transpose      63.91 GB/s              69.27 GB/s

Test PASSED

Press ENTER to exit...

[msft]

Originally Posted by frmky

diagonal transpose      63.91 GB/s              69.27 GB/s

I use diagonal transpose code.

Only support 2048k,4096k.

$ time ./MacLucasFFTW 216091
Iteration 10000 M( 216091 )C, 0x30247786758b8792, n = 2097152, MacLucasFFTW v8.1 Ballester
...
Iteration 210000 M( 216091 )C, 0xcfe091c8f59f8a7b, n = 2097152, MacLucasFFTW v8.1 Ballester
M( 216091 )P, n = 2097152, MacLucasFFTW v8.1 Ballester

real 48m8.792s
user 0m22.289s
sys 0m9.177s

$ time ./MacLucasFFTW 33333333
Iteration 10000 M( 33333333 )C, 0xd717246f501c7d94, n = 2097152, MacLucasFFTW v8.1 Ballester
M( 33333333 )C, 0xd717246f501c7d94, n = 2097152, MacLucasFFTW v8.1 Ballester

real 2m9.478s
user 0m1.248s
sys 0m0.528s

2048k fft sec/iter = 0.0130

$ time ./MacLucasFFTW 63333333
Iteration 10000 M( 63333333 )C, 0xa5d7a917d728239a, n = 4194304, MacLucasFFTW v8.1 Ballester
M( 63333333 )C, 0xa5d7a917d728239a, n = 4194304, MacLucasFFTW v8.1 Ballester

real 4m6.204s
user 0m1.348s
sys 0m0.496s

4096k fft sec/iter = 0.0246

Thank you,

[frmky]

Version u is much better, but still about 5% slower than version k.

Version k: 0.0264 sec/iter
Version o: 0.0278 sec/iter
Version u: 0.0277 sec/iter

Here's the updated profiler graph. transpose no longer dominates the runtime.