esl_alloc.md

esl_alloc : portable aligned memory allocation

The alloc module provides for portable aligned allocation. This is generally only needed for SIMD vector code.

rationale

Yes, the C99 standard states that malloc() is suitably aligned so that it may be assigned to a pointer to any type of object. But SIMD vector types are not part of the C99 standard, vector types may be wider than any C99 object type, and vector memory should be aligned.

Most, if not all systems we work on will provide posix_memalign(). But I don't trust it to be there; plus we have a policy of making it as easy as possible to port to non-POSIX platforms when we can.

We use POSIX's posix_memalign(), C11's aligned_alloc, or Intel's _mm_malloc() (in that preference order), if available on the system. Easel configure.ac tests for them, and sets HAVE_POSIX_MEMALIGN, HAVE_ALIGNED_ALLOC, and/or HAVE__MM_MALLOC as appropriate. If none are available, we fall back to a handrolled implementation.

the quasi-portable fallback implementation

The fallback implementation does unspeakable things, things that are technically undefined-behaviour in C99, but which happen to work everywhere I know of. Specifically, given a pointer to a malloc() allocation, we cast that pointer to an integer (of type uintptr_t), mask off its low-order bits to achieve alignment, and store those low-order bits in a byte preceding our allocation:

  pp = pointer that malloc() gives us
  v
  ...X[------------------------]
     ^^
     | \
     \  p = aligned pointer we give to the caller
      \
        one byte storing r-1, where r=(p-pp), total alignment shift
        r is 1..256, so we can store r-1 in an unsigned byte.

When we free, we use the alignment shift to reconstruct what p was, so we can call free() with the pointer that malloc() originally gave us.

alignment is limited to <= 256 bytes

We only use one byte for the shift r, because we don't anticipate needing to align on more than a 256 byte boundary. Currently the largest vectors are AVX-512's 64-byte vectors, and Intel is projecting an AVX-1024 with 128-byte vectors. We can revisit if needed.

there is an overallocation cost

In the best case, malloc() gives us an allocation that's off by exactly one byte from a properly aligned location; r=1 and we store 0 in the byte. In the worst case, malloc() gives us a properly aligned allocation, in which case our extra byte looks pretty stupid, r=V, and we store V-1 in the byte.

Because the worst case behavior means we overallocate by V bytes, for a pointer that was already properly aligned, the fallback implementation is potentially wasteful, and to minimize the wastage, you should minimize allocation calls where possible. For example, it'd be better to do 2D arrays by setting pointers into a single allocation, for example.

It would be desirable to know when the system malloc() already returns a suitably aligned pointer, and we could then just call malloc() directly - but I don't know a reliable way to test for that.

may cause unnecessary unit test failure

Currently, the unit tests deliberately compile and test the fallback implementation, even if esl_alloc_aligned() is using a system call like posix_memalign(). Thus it may happen that esl_alloc_aligned() is working fine, but the unit test fails because esl_alloc_aligned() doesn't work on some system (perhaps because of the unspeakable things it does).

there is no realloc, by design

Aligned realloc() is a problem in general. There's no POSIX aligned realloc counterpart for posix_memalign(), nor for C11 aligned_alloc(), not for Intel _mm_malloc().

If we try to write our own realloc, we have a problem that the reallocated unaligned pointer could formally have a different offset $r$, so the system realloc() is not guaranteed to move our data correctly. To be sure, we would have to copy our data again in the correct alignment, and we would need to know the size of the data, not just the pointer to it.

Instead, at least for now, we will avoid reallocating aligned memory altogether; instead we will free() and do a fresh allocation. Thus we can only do _Reinit() style functions that do not guarantee preservation of data, not _Resize(), which assume that the data will be preserved.

---

benchmarking

Real time for -L 100, -N 10000: $10^6$ reallocations, so you can think of these as $u$sec per reallocation.

on Mac OS/X: timings are essentially the same w/ gcc vs. clang: [11 Feb 17 on wumpus. 2.5Ghz Core i7, Mac OS/X 10.10.5 Yosemite, gcc 4.9.3, gcc -O3]

	M=5000	M=500000	M=5000000
malloc/realloc	0.159	10.480	5.009
malloc/free/malloc	0.136	0.482	0.897
alloc_aligned_fallback	0.139	0.641	26.394
posix_memalign	0.189	0.481	0.908

on Linux: [11 Feb 17 on ody eddyfs01. icc -O3]

	M=5000	M=500000	M=5000000
malloc/realloc	0.115	0.662	1.094
malloc/free/malloc	0.100	0.252	1.868
alloc_aligned_fallback	0.106	0.249	1.877
posix_memalign	0.206	0.366	1.944

dependence on allocation size isn't obvious

Timings go up and down as max allocation size M changes. Maybe what's happening is that the system is treating different sizes with different strategies.

realloc copies data, so it can be slow

In general, if you don't need data to be preserved, allocating fresh memory (with free()/malloc()) may be faster than realloc(), because realloc() copies data if it has to move the allocation. However, note one example on Linux where realloc() is faster - perhaps because it's smart enough to recognize cases where it doesn't need to expand an allocation.

easel's aligned alloc can be slow on OS/X

I ran the -M5000000 case under Instruments. It is spending all its time in free(), in madvise(). Not sure why.

conclusion

• posix_memalign() is usually available and performs well.
• we'll design HMMER vector code to _Reinit() with fresh allocations, rather than using reallocation. This may even speed things up a small bit.
• the madvise() stall with the easel fallback code is puzzling and worrying, though it only happens on MacOS, not Linux.