Ispc: A SPMD Compiler for High-Performance CPU Programming

http://ispc.github.io/documentation.html

by xstartup • 9 years ago

44 9 9 years ago

ispc.github.io

gonewest• 9 years ago

Used in the implementation of a production raytracer at Dreamworks Animation:

http://tabellion.org/et/paper17/index.html

meuk• 9 years ago

I think SIMD programming is a simple way to improve the performance of applications that use the CPU. The real solution for applications that are computation bound and not inherently sequential is to perform the computations on an architecture that allows massive parallelism (a manycore architecure like Xeon Phi, a GPU, or even an FPGA).

rurban• 9 years ago

Sure, but you still need to the software to manage this massive parallelism. It cannot just be OpenMP or OpenCL with C++. It needs to be a better and faster system, such as ispc or pony. ispc is faster, pony is safer, both are better and faster than C++.

MaxBarraclough3• 9 years ago

At a glance, ISPC looks a lot like OpenCL. Does it outperform Intel's OpenCL-on-CPU engine? If so, why?

dragontamer• 9 years ago

both are better and faster than C++

"Better" is a matter of opinion, so I'll stay away from that claim.

But if we include AMD's HCC or Microsoft's C++ AMP projects (which turn C++ code and directly run it on the GPU), then I think you are mistaken about "faster".

http://rocm-documentation.readthedocs.io/en/latest/ROCm_API_...

https://msdn.microsoft.com/en-us/library/hh265136.aspx

Full GPU acceleration from C++ code. The SIMT model is quite powerful for allowing massive parallelism. And the "tiling" feature indicates that C++ is powerful enough to represent the various memory structures inside of the GPU. (LDS vs Global memory)

------------

Transparency of the memory model is incredibly important. GPUs only get their massive TFlop numbers if they are calculating from their LDS cache / Shared Memory locations. Even the biggest, beefiest GPUs with HBM2 only have ~480GB/s of main-memory performance, which is insufficient to reach 10TFlops. (Remember: a single float is 4-bytes. So 480GB/s bandwidth only gives 120GFlops at best)

Careful programming of memory, and transparency of the "LDS" cache and awareness of the limited register space of the GPU is needed to even

Ispc: A SPMD Compiler for High-Performance CPU Programming

Related Stories