Sequoia

I am a grad student focusing on compiler and architecture.

I think the stream processor like Imagine and its compiler is very promising. The separation between computation and control makes sense for a lot of practical applications. Imagine and other streaming processors are great. But papers about them seem like not so active since 2007. At the meantime, commercial GPGPU and CELL BE populated. my dumb question is Can I regards them as evolve of streaming processor? i guess they are essentially bulk-style processors. Why you stop? do you think the pure streaming processor does not work out?

Sequoia drew my attention as applicable single-source compiler infrastructure, however, I realized that it's much general and flexible after reading the published articles about it [sc06],[ppopp07], [ppopp08]. all these 3 papers are excellent. The hierarchical approach to separate control and intensive computation is great idea. This way can solve scatter-gather deriving from streaming processor. [ppopp07] impressed me very much because it detailed a bunch of optimization for cell, which i considered.

What's confusing me it the development of Sequoia. Because many members are coming from Graphics lab, GPU should be one target. Only mentioned in [ppopp08], i can not find any farther information. any clue? In addition, I think the release package is a little out of date. I had gone through the source tar. the total extracted directory is about 60M, PPC compiler is amazingly 32M. counting x86 variety, 97% files are opaque for the innocent & motived reader. I carefully read the runtime directory. The interface of class runtime is a subtle inconsistent in the [ppopp08], and only virtual functions but implementation at there. The specialized code for cell is well-developed though, sqcc acts like a magic compiler can emit all the Makefile, entrypoints, taskfunctions, and even the implementation of Runtime...

How state-of-the-art compiler and runtime looks like? I believe Sequoia is great. due to the powerful xml description and IR, i strongly believe it is the perfect research platform for GPU and other streaming processor. actually, i am developing for cell using Sequoia now. i met some feature-requests and ideas to optimize of data-reuse. can i discuss them to you guys? btw, up-coming larrabbee is very close to cell in terms of physical layout. i believe the result on that platform are promising with tiny efforts.

Still, I am interested in the optimization stage. i want to read the concrete code about instantiate task and optimization describing in [ppopp07]. Can i check out the code. furthermore please disclose me the roadmap of development. ie. What features is going to appear next?

i feel attractive things for Sequoia as follows:
i) provide advisable values for tunable variable. bi-search at least.
ii) OO or struct support?
iii) TM-support for reduce functions.
iv) type-safe API for external functions. why does it use sqArray* instead of concrete types?
etc.

sqcc is a magic compiler. i think time will prove it is influential.