I knew counting bits individually was slow, but it never really mattered
until now. However, I didn't expect such a dramatic boost just by going
to mapping bytes to bit counts. 16-bit words would be faster still, but
the 64kB lookup table would probably start hurting cache performance,
and 32-bit words (4GB table) definitely would ruin the cache. The
universe isn't big enough for 64-bits :)
Having set_expand exposed is useful for loading data into a set.
However, it turns out there was a bug in its size calculation in that
when the requested set size was a multiple of SET_BITS (and greater than
the current set size), the new set size one be SET_BITS larger than
requested. There's now some tests for this :)
This reduces the overhead needed to manage the memory blocks as the
blocks are guaranteed to be page-aligned. Also, the superblock is now
alllocated from within one of the memory blocks it manages. While this
does slightly reduce the available cachelines within the first block (by
one or two depending on 32 vs 64 bit pointers), it removes the need for
an extra memory allocation (probably via malloc) for the superblock.
This failed with errors such as:
from ./include/QF/simd/vec4d.h:32,
from libs/util/simd.c:37:
./include/QF/simd/vec4d.h: In function ‘qmuld’:
/usr/lib/gcc/x86_64-pc-linux-gnu/10.3.0/include/avx2intrin.h:1049:1: error: inlining failed in call to ‘always_inline’ ‘_mm256_permute4x64_pd’: target specific option mismatch
1049 | _mm256_permute4x64_pd (__m256d __X, const int __M)
Fuzzy bsearch is useful for finding an entry in a prefix sum array
(value is >= ele[0], < ele[1]), and the reentrant version is good when
data needs to be passed to the compare function. Adapted from the code
used in pr_resolve.
A bit of a mess for optimized vs unoptimized, but the tests acknowledge
the differences in precision while checking that the code produces the
right results allowing for that precision.
It seems that i686 code generation is all over the place reguarding sse2
vs fp, with the resulting differences in carried precision. I'm not sure
I'm happy with the situation, but at least it's being tested to a
certain extent. Not sure if this broke basic (no sse) i686 tests.
GCC does a fairly nice job of producing code for vector types when the
hardware doesn't support SIMD, but it seems to break certain math
optimization rules due to excess precision (?). Still, it works well
enough for the core engine, but may not be well suited to the tools.
However, so far, only qfvis uses vector types (and it's not tested yet),
and tools should probably be used on suitable machines anyway (not
forces, of course).
I don't know that the cache line size is 64 bytes on 32 bit systems, but
it should be ok to assume that 64-byte alignment behaves well on systems
with smaller cache lines so long as they are powers of two. This does
mean there is some waste on 32-bit systems, but it should be fairly
minimal (32 bytes per memblock, which manages page sized regions).
This seems to be pretty close to as fast as it gets (might be able to do
better with some shuffles of the negation constants instead of loading
separate constants).
Care needs to be taken to ensure the right function is used with the
right arguments, but with these, the need to use qconj(d|f) for a
one-off inverse rotation is removed.
I think the sub-line allocator falling over is the final source of
qfvis's leaks. It certainly causes a mess of the sub-lines. But having
some tests to get working sure beats scratching my head over qfvis :)
The idea is to not search through blocks for an available allocation.
While the goal was to speed up allocation of cache lines of varying
cluster sizes, it's not enough due to fragmentation.
They take advantage of gcc's vector_size attribute and so only cross,
dot, qmul, qvmul and qrot (create rotation quaternion from two vectors)
are needed at this stage as basic (per-component) math is supported
natively by gcc.
The provided functions work on horizontal (array-of-structs) data, ie a
vec4d_t or vec4f_t represents a single vector, or traditional vector
layout. Vertical layout (struct-of-arrays) does not need any special
functions as the regular math can be used to operate on four vectors at
a time.
Functions are provided for loading a vec4 from a vec3 (4th element set
to 0) and storing a vec4 into a vec3 (discarding the 4th element).
With this, QF will require AVX2 support (needed for vec4d_t). Without
support for doubles, SSE is possible, but may not be worthwhile for
horizontal data.
Fused-multiply-add is NOT used because it alters the results between
unoptimized and optimized code, resulting in -mfma really meaning
-mfast-math-anyway. I really do not want to have to debug issues that
occur only in optimized code.
It is capable of parsing single expressions with fairly simple
operations. It current supports ints, enums, cvars and (external) data
structs. It is also thread-safe (in theory, needs proper testing) and
the memory it uses can be mass-freed.
This was inspired by
Hoard: A Scalable Memory Allocator
for Multithreaded Applications
Emery D. Berger, Kathryn S. McKinley, Robert D. Blumofe, Paul R.
Wilson,
It's not anywhere near the same implementation, but it did take a few
basic concepts. The idea is twofold:
1) A pool of memory from which blocks can be allocated and then freed
en-mass and is fairly efficient for small (4-16 byte) blocks
2) Tread safety for use with the Vulkan renderer (and any other
multi-threaded tasks).
However, based on the Hoard paper, small allocations are cache-line
aligned. On top of that, larger allocations are page aligned.
I suspect it would help qfvis somewhat if I ever get around to tweaking
qfvis to use cmem.
The calculation fails (produces NaN) if the vectors are anti-parallel,
but works for all other combinations. I came up with this implementation
when I discovered Unity's Quaternion.FromToRotation could did not work
with very small angles. This implementation will produce a usable
quaternion below 0.00255 degrees (though it will be slightly larger than
unit). Unity's failed such that I could see KSP's skybox snap while it
rotated around my test vessel.
There's still some cleanup to do, but everything seems to be working
nicely: `make -j` works, `make distcheck` passes. There is probably
plenty of bitrot in the package directories (RPM, debian), though.
The vc project files have been removed since those versions are way out
of date and quakeforge is pretty much dependent on gcc now anyway.
Most of the old Makefile.am files are now Makemodule.am. This should
allow for new Makefile.am files that allow local building (to be added
on an as-needed bases). The current remaining Makefile.am files are for
standalone sub-projects.a
The installable bins are currently built in the top-level build
directory. This may change if the clutter gets to be too much.
While this does make a noticeable difference in build times, the main
reason for the switch was to take care of the growing dependency issues:
now it's possible to build tools for code generation (eg, using qfcc and
ruamoko programs for code-gen).
When I ported SEB to python, I discovered that I apparently didn't
really understand the paper's description of the end condition and the
usage of the affine and convex sets for center testing. This cleans up
the test and makes SEB more correct for the cases that have less than 4
supporting points (especially when there are less than 4 points total).
The better accuracy is for specific cases (90 degree rotations around a
main axis: the matrix element for that axis is now 1 instead of
0.99999994). The speedup comes from doing fewer additions (multiply
seems to be faster than add for fp, at least in this situation).
After messing with SIMD stuff for a little, I think I now understand why
the industry went with xyzw instead of the mathematical wxyz. Anyway, this
will make for less pain in the future (assuming I got everything).
