MAGMA GEMM Sources for Fermi Released

[admin]

The MAGMA BLAS SGEMM and DGEMM sources for Fermi GPUs are now released.
These improved GEMMs, developed by Rajib Nath and Stan Tomov, will be
part of the up-coming MAGMA 0.3 library release and will be included in
CUBLAS 3.2 as well.

The basic algorithm is described in:
Nath, R., Tomov, S., Dongarra, J. "An Improved MAGMA GEMM for Fermi GPUs,"
University of Tennessee Computer Science Technical Report, UT-CS-10-655
(also LAPACK working note 227), July 29, 2010.
http://icl.cs.utk.edu/projectsfiles/mag ... i_gemm.pdf

On a C2050 GPU the new DGEMM gets up to 300 GFlop/s (58% of peak) and
the SGEMM up to 645 (63% of peak). On a GTX480 DGEMM gets up to 166 GFlop/s
and SGEMM up to 844 GFlop/s.

[mbibby]

When will we see the cgemm and zgemm equivalents?

Malcolm

[Stan Tomov]

I am not sure if we would personally write the equivalents. NVIDIA is preparing CUBLAS 3.2
that will have improved c/z gemms using ideas from the s/d gemms.
Stan

[Boxed Cylon]

I preface this post with the declaration that I know just about nothing about details of these routines...

I was looking through the fermi_sgemm.cu routine to get some sense of how the code was engineered. I noticed the __mul24 function, and wondered what it did. A google search turned up the Fermi Tuning Guide with:

Code: Select all

32-Bit Integer Multiplication
On devices of compute capability 1.x, 32-bit integer multiplication is implemented using multiple instructions as it is not natively supported. 24-bit integer multiplication is natively supported via the __[u]mul24 intrinsic.

On devices of compute capability 2.0, however, 32-bit integer multiplication is natively supported, but 24-bit integer multiplication is not. __[u]mul24 is therefore implemented using multiple instructions and should not be used (Section 5.4.1).

Should the fermi_sgemm.cu routine be using __mul24? (Or perhaps there are reasons 24-bit integers are employed?)

[Stan Tomov]

There is no reason to use __mul24. We will remove it. Thanks for pointing this out.

[Allan Menezes]

Dear Stan,
As this is just pointer arithmetic and used in only a few places it does not change the perfomance much at all as per my experiment below.
Just for fun I changed fermi_dgemm.cu and fermi_sgemm.cu with a single #define on top as #define __mul24(a,b) ((a)*(b)) and there was no significant difference in Gflops and err was still 0.00 on a GTX-480.
The device memory still on available fermi devices is < 4GB and is going to change in the future with the Tesla C2070 and CUDA 3.2 to 64 bit addresses.
Thank you,
Allan

[rramachand21]

Hello,

I am new to cuda and this api. Could I please get the source code for matrix vector multiplication (sgemv and dgemv) which is generic.

Thanks,
Ranjith