thechao

129 ms·

1621.
by thechao 11y ago
I think you're asking about the distinction between SIMT (supported by GPUs/compute coprocessors), and SIMD. In SIMT, you want to issue a "bunch of the same work on a crapload of pseudo-threads"—in the most hardware-ef
1622.
by thechao 11y ago
I remember reading, years ago, a paper which provided an algorithm for converting an IEEE 754 into a string using only a fixed number machine word length scalars—but I certainly can't find it now.
1623.
by thechao 12y ago
The swords illustrate the seven basic "strokes": six in a "flower", and one stab (the bottom one). Mastering all of the ways to get your weapon from point "A" (near you, in a proper guard), to point "B&quo
1624.
by thechao 12y ago
Concepts are significantly simpler to implement & check. I did a non-trivial amount of coding both with TMP-concepts, and with ConceptGCC. ConceptGCC produced better error messages earlier, was easier to symbolically debug, and produced
1625.
by thechao 12y ago
There were two fundamentally different proposals for concepts: Indiana— concepts are records of signatures; checking is done "by signature"; an arbitrary mapping (adapter) can be defined. Notable authors: Doug Gregor, Jeremy Siek,
1626.
by thechao 12y ago
Here's my implementation in c++, using alloca(): https://github.com/jaroslov/coro . Without debugger support, coroutines are pretty difficult to work with.
1627.
by thechao 12y ago
Finished this five years ago: "Stepanov-style generic programming is the bees knees; we don't understand why."
1628.
by thechao 12y ago
I enjoyed the heck out of developing drivers for that thing—it was a constant uphill battle to even get correctness, nonetheless, the pretty decent performance we got. In the end, my tablet's battery died, and it now sits in a drawer,
1629.
by thechao 12y ago
Rename tracking is what undoes a lot if the more naive variants.
1630.
by thechao 12y ago
> Some or all of this article's listed sources may not be reliable. Please help this article by looking for better, more reliable sources, or by checking whether the references meet the criteria for reliable sources. Unreliable cita
1631.
by thechao 12y ago
The last time I researched this, to sufficient depth, was about 3 years ago. LZ4 was, hands down, the winner---nothing was even close.
1632.
by thechao 12y ago
This popped up on reddit/r/programming yesterday, so I'll just copy my comment from over there: How well does density [perform] when there's competition for the I$ and L$? LZ4 in theory is slower than a number
1633.
by thechao 12y ago
Maybe... 'prefer'. I'm more confident that a really conservative Makefile will build my code correctly. My issue with -MD was not that it didn't provide precise (and correct!) dependencies; my issue was that the build sy
1634.
by thechao 12y ago
> Do you have some kind of way to verify that your makefile dependencies conform to your source dependencies? Nope. I explicitly use a conservative approximation—this guarantees correctness, over speed. Building everything every time wit
1635.
by thechao 12y ago
I've been burned by so many build tools over the years. I've finally settled (for C/++/asm) on the combination of Make + ccache: I build a _very_ paranoid Makefile that recompiles everything if it feels like anything cha
1636.
by thechao 12y ago
> Many years ago, I briefly worked at NVIDIA on the DirectX driver team (internship). I've been a full-time GPU engineer at a couple of major companies for about 6 years now. My colleagues at my current company have all been GPU eng
1637.
by thechao 12y ago
> but adding a map function on Vector would have not prevented this more sophisticated use. This was questioned, long ago, by Todd Veldhuizen in his Parsimony Principle paper[1]. The long-and-short of which is: do it! [1] http:/&#x
1638.
by thechao 12y ago
What about a contract like this: the copyright holder declares the value of the copyright (on the quarter; on the year; but the declared value may only increase); we then tax them on that value---say 1%. The flip side is that the copyright
1639.
by thechao 12y ago
There is a fairly straightforward generalization of the 'axis' concept to allow graph traversal: 'GPath'. (You have to add an annotation for a traversal kind , and an identifier for a property map.) XPath can be straigh
1640.
by thechao 12y ago
I used to write software rasterizers in a past life. AVX-512 is straight-up an observed 8x performance gain over SSE: a well-written software rasterizer is a very clever thread-scheduling algorithm wrapped around long lists of FMA-load&#x2F
1641.
by thechao 12y ago
Higher- or lower- level domains, I feel, would be more accurate; when I code my c-with-templates embedded code there's no STL, so I'm quite aware of intrusive containers, even though I rarely code straight-C. My favorite variant i
1642.
by thechao 12y ago
I used to be in the field of ultra-high-speed flow cytometry. We were working on developing 1-mega-event sorters around 10 years ago. http://onlinelibrary.wiley.com/doi/10.1002/cyto.a.20160/full James Leary w