<!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.0//EN" "http://www.w3.org/TR/REC-html40/strict.dtd">
<html><head><meta name="qrichtext" content="1" /><style type="text/css">
p, li { white-space: pre-wrap; }
</style></head><body style=" font-family:'monospace'; font-size:9pt; font-weight:400; font-style:normal;">
<p style=" margin-top:0px; margin-bottom:0px; margin-left:0px; margin-right:0px; -qt-block-indent:0; text-indent:0px; -qt-user-state:0;">> I assume the ">" operator will</p>
<p style=" margin-top:0px; margin-bottom:0px; margin-left:0px; margin-right:0px; -qt-block-indent:0; text-indent:0px; -qt-user-state:0;">> return an array of bools and then you'll have to arrange that different</p>
<p style=" margin-top:0px; margin-bottom:0px; margin-left:0px; margin-right:0px; -qt-block-indent:0; text-indent:0px; -qt-user-state:0;">> portions of your array are handled differently (like MATLAB/Octave?</p>
<p style=" margin-top:0px; margin-bottom:0px; margin-left:0px; margin-right:0px; -qt-block-indent:0; text-indent:0px; -qt-user-state:0;">> see also the old Cray vector merge intrinsics).</p>
<p style="-qt-paragraph-type:empty; margin-top:0px; margin-bottom:0px; margin-left:0px; margin-right:0px; -qt-block-indent:0; text-indent:0px; "> </p>
<p style=" margin-top:0px; margin-bottom:0px; margin-left:0px; margin-right:0px; -qt-block-indent:0; text-indent:0px; -qt-user-state:0;">In what I prototyped, you'd do something like</p>
<p style="-qt-paragraph-type:empty; margin-top:0px; margin-bottom:0px; margin-left:0px; margin-right:0px; -qt-block-indent:0; text-indent:0px; "> </p>
<p style=" margin-top:0px; margin-bottom:0px; margin-left:0px; margin-right:0px; -qt-block-indent:0; text-indent:0px; -qt-user-state:0;"> const auto mask = (fabs (Ce) <= 2.623395162778);</p>
<p style=" margin-top:0px; margin-bottom:0px; margin-left:0px; margin-right:0px; -qt-block-indent:0; text-indent:0px; -qt-user-state:0;"> const auto X = Q->Qn * Ce;</p>
<p style=" margin-top:0px; margin-bottom:0px; margin-left:0px; margin-right:0px; -qt-block-indent:0; text-indent:0px; -qt-user-state:0;"> const auto Y = Q->Qn * Cn + Q->Zb;</p>
<p style=" margin-top:0px; margin-bottom:0px; margin-left:0px; margin-right:0px; -qt-block-indent:0; text-indent:0px; -qt-user-state:0;"> return { ternary(mask, X, HUGE_VAL),</p>
<p style=" margin-top:0px; margin-bottom:0px; margin-left:0px; margin-right:0px; -qt-block-indent:0; text-indent:0px; -qt-user-state:0;"> ternary(mask, Y, HUGE_VAL) };</p>
<p style="-qt-paragraph-type:empty; margin-top:0px; margin-bottom:0px; margin-left:0px; margin-right:0px; -qt-block-indent:0; text-indent:0px; "> </p>
<p style=" margin-top:0px; margin-bottom:0px; margin-left:0px; margin-right:0px; -qt-block-indent:0; text-indent:0px; -qt-user-state:0;">With SSE/AVX "mask" will be just a SIMD register itself with bits set or unset depending on the rest of the comparison for each member of the vector, and ternary(mask, then, else) will expand to</p>
<p style=" margin-top:0px; margin-bottom:0px; margin-left:0px; margin-right:0px; -qt-block-indent:0; text-indent:0px; -qt-user-state:0;">or(and(mask, then), andnot(mask, else)). (In an ideal world, we would be able to overload the ? : operator, but that's not possible :-))</p>
<p style=" margin-top:0px; margin-bottom:0px; margin-left:0px; margin-right:0px; -qt-block-indent:0; text-indent:0px; -qt-user-state:0;">In cases where you really need a real branch, you can get an integer and then decide to go to a slower code path to deal with a non nominal situation. But in my idea, only a small portion of existing PROJ operations would be convered to a vector-enabled code: the ones that are particularly costly and have few branches to deal with the nominal input.</p>
<p style="-qt-paragraph-type:empty; margin-top:0px; margin-bottom:0px; margin-left:0px; margin-right:0px; -qt-block-indent:0; text-indent:0px; "> </p>
<p style=" margin-top:0px; margin-bottom:0px; margin-left:0px; margin-right:0px; -qt-block-indent:0; text-indent:0px; -qt-user-state:0;">> (And probably my assertion that the code can remain essentially</p>
<p style=" margin-top:0px; margin-bottom:0px; margin-left:0px; margin-right:0px; -qt-block-indent:0; text-indent:0px; -qt-user-state:0;">> the same, except for the declarations, is a pipe dream.)</p>
<p style="-qt-paragraph-type:empty; margin-top:0px; margin-bottom:0px; margin-left:0px; margin-right:0px; -qt-block-indent:0; text-indent:0px; "> </p>
<p style=" margin-top:0px; margin-bottom:0px; margin-left:0px; margin-right:0px; -qt-block-indent:0; text-indent:0px; -qt-user-state:0;">Not at all. My SIMD exact_e_fwd() is exactly the same as current one, with just "double" replaced by the vector type, and the final part with the if( fabs (Ce) <= 2.623395162778) branch replaced by the above snippet.</p>
<p style="-qt-paragraph-type:empty; margin-top:0px; margin-bottom:0px; margin-left:0px; margin-right:0px; -qt-block-indent:0; text-indent:0px; "> </p>
<p style=" margin-top:0px; margin-bottom:0px; margin-left:0px; margin-right:0px; -qt-block-indent:0; text-indent:0px; -qt-user-state:0;">> By the way, Eigen is an awesome package. It implements everything with</p>
<p style=" margin-top:0px; margin-bottom:0px; margin-left:0px; margin-right:0px; -qt-block-indent:0; text-indent:0px; -qt-user-state:0;">> header code and so the end result is often very efficient. The downside</p>
<p style=" margin-top:0px; margin-bottom:0px; margin-left:0px; margin-right:0px; -qt-block-indent:0; text-indent:0px; -qt-user-state:0;">> is that because of all the flexibility that Eigen provides with how</p>
<p style=" margin-top:0px; margin-bottom:0px; margin-left:0px; margin-right:0px; -qt-block-indent:0; text-indent:0px; -qt-user-state:0;">> arrays are laid out and allocated, you can sometimes end up in template</p>
<p style=" margin-top:0px; margin-bottom:0px; margin-left:0px; margin-right:0px; -qt-block-indent:0; text-indent:0px; -qt-user-state:0;">> hell.</p>
<p style="-qt-paragraph-type:empty; margin-top:0px; margin-bottom:0px; margin-left:0px; margin-right:0px; -qt-block-indent:0; text-indent:0px; "> </p>
<p style=" margin-top:0px; margin-bottom:0px; margin-left:0px; margin-right:0px; -qt-block-indent:0; text-indent:0px; -qt-user-state:0;">To be honest, I love my own header approach because I know how it works and can customize it quite easily, and it was fun to do. But its scope (base types & supported architectures) is more reduced than other mentioned alternatives.</p>
<p style="-qt-paragraph-type:empty; margin-top:0px; margin-bottom:0px; margin-left:0px; margin-right:0px; -qt-block-indent:0; text-indent:0px; "> </p>
<p style=" margin-top:0px; margin-bottom:0px; margin-left:0px; margin-right:0px; -qt-block-indent:0; text-indent:0px; -qt-user-state:0;">At the end of the day, everybody ends up more or less the same API, a templated type <base_type, number_of_elements>, so switching between implementations (with similar capabilities) could be done relatively easily. In a PROJ context, the availability and accuracy of transcendent functions is a key element, and I'm not aware of better alternatives than Sleef for that.</p>
<p style=" margin-top:0px; margin-bottom:0px; margin-left:0px; margin-right:0px; -qt-block-indent:0; text-indent:0px; -qt-user-state:0;">The fact that it was considered for clang and also in the paper Andrew quoted for a LLVM-based (proprietary) Arm Compiler for HPC makes it IMHO a serious candidate.</p>
<p style="-qt-paragraph-type:empty; margin-top:0px; margin-bottom:0px; margin-left:0px; margin-right:0px; -qt-block-indent:0; text-indent:0px; "> </p>
<p style=" margin-top:0px; margin-bottom:0px; margin-left:0px; margin-right:0px; -qt-block-indent:0; text-indent:0px; -qt-user-state:0;">I've also just given a quick try at https://github.com/xtensor-stack/xsimd suggested by Joris Van den Bossche, which is also another header based approach (the only one that can work...)</p>
<p style="-qt-paragraph-type:empty; margin-top:0px; margin-bottom:0px; margin-left:0px; margin-right:0px; -qt-block-indent:0; text-indent:0px; "> </p>
<p style=" margin-top:0px; margin-bottom:0px; margin-left:0px; margin-right:0px; -qt-block-indent:0; text-indent:0px; -qt-user-state:0;">One annoying limitation of it is that it refuses to instanciate for example a vector of 4-double on SSE2-only because it doesn't map to a single SIMD register, whereas Eigen or my approach would use 2 SSE registers for that, which is sometimes still valuable to make use of the fact that the CPU can sometimes execute 2 additions in parallel if they use separated registers. I could actually measure that on my prototype (even when going to a 8-double "vector" on SSE, which maps to 4 SSE registers), so this isn't just theoretical.</p>
<p style="-qt-paragraph-type:empty; margin-top:0px; margin-bottom:0px; margin-left:0px; margin-right:0px; -qt-block-indent:0; text-indent:0px; "> </p>
<p style=" margin-top:0px; margin-bottom:0px; margin-left:0px; margin-right:0px; -qt-block-indent:0; text-indent:0px; -qt-user-state:0;">I've not seen an explicit statement about the accuracy of transcendent functions in xsimd. When randomly trying sin(150), it returned a small difference compared to libc sin() or Sleef, but that's probably not in the range of values that PROJ would trigger. When trying between [-PI,PI], couldn't see a difference from a quick probing.</p>
<p style="-qt-paragraph-type:empty; margin-top:0px; margin-bottom:0px; margin-left:0px; margin-right:0px; -qt-block-indent:0; text-indent:0px; "> </p>
<p style=" margin-top:0px; margin-bottom:0px; margin-left:0px; margin-right:0px; -qt-block-indent:0; text-indent:0px; -qt-user-state:0;">One advantage with xsimd is that the transcendent functions are implemented in a header-only fashion too, whereas currently Sleef requires linking against a lib, but inlining of the code of the transcendent functions isn't really needed for performance, as the overhead of the call is quite neglectable compared to their cost.</p>
<p style="-qt-paragraph-type:empty; margin-top:0px; margin-bottom:0px; margin-left:0px; margin-right:0px; -qt-block-indent:0; text-indent:0px; "> </p>
<p style=" margin-top:0px; margin-bottom:0px; margin-left:0px; margin-right:0px; -qt-block-indent:0; text-indent:0px; -qt-user-state:0;">-- </p>
<p style=" margin-top:0px; margin-bottom:0px; margin-left:0px; margin-right:0px; -qt-block-indent:0; text-indent:0px; -qt-user-state:0;">Spatialys - Geospatial professional services</p>
<p style=" margin-top:0px; margin-bottom:0px; margin-left:0px; margin-right:0px; -qt-block-indent:0; text-indent:0px; -qt-user-state:0;">http://www.spatialys.com</p></body></html>