xoranth

9 ms·

61.
by xoranth 3y ago
(Sorry, I was wrong. See here for the author's reply: https://news.ycombinator.com/item?id=36627385 )
62.
by xoranth 3y ago
> I guess the compiler's unrolling heuristics generally aren't as good as that blocking "mod then div" alternative to Duff's device? Obviously taking `s` out of the loop condition is part of the magic. The magic
63.
by xoranth 3y ago
Interesting. I think you can vectorize the prologue using movemask + popcnt instead of keeping a counter in the ymm registers (warning: untested code, still need to benchmark it): const __m256i zero = _mm256_setzero_si256(); const
64.
by xoranth 3y ago
I rerun the benchmark vs loop-5 and loop-7 from the second post. Runtime is basically the same on my machine. I would have expected yours to be faster given that it needs to execute fewer instructions per loop iteration. Though maybe the CP
65.
by xoranth 3y ago
I think I managed to improve on both this post, and its sequel, at the cost of specializing the function for the case of a string made only of 's' and 'p'. The benchmark only tests strings made of 's' and &#x27
66.
by xoranth 3y ago
Most blogs that have RSS also have a `<link rel="alternate" type="application/rss+xml">` tag that redirects you to the RSS feed. If you pass the link to the homepage to a feed reader[^0], it will follow the li
67.
by xoranth 3y ago
https://xoranth.net/ In-depth blogging about low level optimization. Two posts so far: - https://xoranth.net/memcmp-avx2/ A walkthrough of an highly optimized implementation of string comparisons. - h
68.
by xoranth 3y ago
Lower fees. Also, "exchange competition" is a thing in US stocks, but not, for example, in the futures' market. And there's HFTs in futures too, so having a single exchange wouldn't "remove" HFTs (not that
69.
by xoranth 3y ago
Your vanilla Debian box uses BMI2 whenever you do a string comparison, unless you are on a decade old CPU [^0]. The "strange" instructions are actually not that niche, it's just that usage tends to be "indirect" and
70.
by xoranth 3y ago
> Another simple trick is duplicating the buffer pointer field and buffer size field for consumer and producer Do you mean something like this? struct Inner { int\* data; size_t size; size_t cachedIdx; }
71.
by xoranth 3y ago
From the slides on `conflict`/`vpcompress`, it says. > In fact, we can re-use an existing structure with some minor changes What is the existing structure that they reused? The slides don't seem to specify.
72.
Optimizing the `pext` perfect hash function(xoranth.net)
3 points by xoranth 3y ago|0 comments
73.
by xoranth 3y ago
As a datapoint, numpy also chooses between radix and timsort based on data type. From the docs [^0]: > ‘stable’ automatically chooses the best stable sorting algorithm for the data type being sorted. It, along with ‘mergesort’ is current
74.
by xoranth 3y ago
They have sample searches for "best headphones", "steve jobs", and "python exceptions" on their main website [^0]. [0] https://kagi.com/
75.
by xoranth 3y ago
> For larger arrays, I believe the branchy search algorithms start to do better since half their branches get predicted correctly. (not sure if this is what you meant by "conflict misses") I believe the reason branchy binary se
76.
A look inside memcmp on Intel AVX2 hardware(xoranth.net)
3 points by xoranth 3y ago|0 comments
77.
by xoranth 3y ago
Thanks!
78.
Ask HN: Resources on Memory Prefetching
2 points by xoranth 3y ago|2 comments
79.
by xoranth 3y ago
Thank you! I'll take a look.
80.
by xoranth 3y ago
> 2. glibc strings functions are ok, but not particularly good, ime Do you mean for RISC-V, or in general? (and in particular, what about x64?) What problems do they have? [Also what are in your experience better implementations?]
81.
by xoranth 3y ago
Mmm, it looks like LLVM even has a pass to actively convert generated cmovs into branches if one of the operands touches memory [^1], and that only supported way to force things to be brancheless is inline assembly [^2]. [^1] https:/&
82.
by xoranth 3y ago
> However, I’ve found that some compilers, e.g. GCC on x86-64 will refuse to make this variant branchless. I hate how fickle compilers can be sometimes, and I wish compilers exposed not just the likely/unlikely attributes, but also
83.
by xoranth 8y ago
I cannot speak on Carreyrou, but parts of _When Genius Failed_ make me think that Lowenstein did not have a good grasp of the subject he was writing about. For example, when discussing LTCM volatility strategies: >The stock market, for i
84.
by xoranth 10y ago
Could you share more details about the robust regression you are using? All resources I could find online on robust regression would either point to Laplace distributed residuals, or some capped loss function.
85.
by xoranth 10y ago
The difference in the drop between UKX and IAS is due to fx. Namely, if you were a British investor that bought an european index ETF before the move, you would have lost on the drop, but would have made some of the money back on the fact t