Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Implement intersperse using SSE2 #310

Merged
merged 1 commit into from
Nov 20, 2020
Merged

Implement intersperse using SSE2 #310

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
4 changes: 4 additions & 0 deletions bench/BenchAll.hs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -361,6 +361,10 @@ main = do
, benchFE "floatHexFixed" $ fromIntegral >$< P.floatHexFixed
, benchFE "doubleHexFixed" $ fromIntegral >$< P.doubleHexFixed
]
, bgroup "intersperse"
[ bench "intersperse" $ whnf (S.intersperse 32) byteStringData
, bench "intersperse (unaligned)" $ whnf (S.intersperse 32) (S.drop 1 byteStringData)
]
, bgroup "partition"
[
bgroup "strict"
Expand Down
20 changes: 19 additions & 1 deletion cbits/fpstring.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -30,6 +30,10 @@
*/

#include "fpstring.h"
#if defined(__x86_64__)
#include <emmintrin.h>
#include <xmmintrin.h>
#endif

/* copy a string in reverse */
void fps_reverse(unsigned char *q, unsigned char *p, size_t n) {
Expand All @@ -44,7 +48,21 @@ void fps_intersperse(unsigned char *q,
unsigned char *p,
size_t n,
unsigned char c) {

#if defined(__x86_64__)
{
const __m128i separator = _mm_set1_epi8(c);
const unsigned char *const p_begin = p;
const unsigned char *const p_end = p_begin + n - 9;
while (p < p_end) {
const __m128i eight_src_bytes = _mm_loadl_epi64((__m128i *)p);
Copy link
Contributor

@vdukhovni vdukhovni Oct 24, 2020 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

What happens when p is not 64-bit aligned? Is there any performance degradation?
Are there OS platforms supported by GHC where __x86_64__ is defined, but the intrinsics are not available?
That is I guess, has this code been tested on FreeBSD, MacOS and Windows, and perhaps even less common platforms like NetBSD, that are not officially supported by GHC, but do maintain ports...

Lastly, I am curious about what sort of applications might care about the performance of intersperse? Is this a graphics thing? When would I want to bulk insert a fixed byte between every other byte in a large-enough buffer to want it done faster with intrinsics? (And yet be doing the task in Haskell...)

[ EDIT: FWIW, it compiles on a FreeBSD 12 system, and appears to work correctly in naïve tests... ]

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

What happens when p is not 64-bit aligned? Is there any performance degradation?

My understanding is that there is no performance degradation on CPUs released in the latest ~10 years. And for earlier ones I suspect it's still worth it.

Are there OS platforms supported by GHC where x86_64 is defined, but the intrinsics are not available? That is I guess, has this code been tested on FreeBSD, MacOS and Windows, and perhaps even less common platforms like NetBSD, that are not officially supported by GHC, but do maintain ports...

I think this code should work on all OSes, at least with gcc and clang. Does it ever happen that cbits of a Haskell package are compiled with MSVC? Intrinsics should still be the same but I'd need to verify that the header names are correct.

I can make a PR configuring GitHub Actions to build and test bytestring on Linux, Mac and Windows if you're OK with that. I'm not aware of any CI service that offers BSD machines.

Lastly, I am curious about what sort of applications might care about the performance of intersperse? Is this a graphics thing? When would I want to bulk insert a fixed byte between every other byte in a large-enough buffer to want it done faster with intrinsics? (And yet be doing the task in Haskell...)

I just want the concept of using SIMD in Haskell to be more mainstream. The reason I started with ByteString.intersperse is that it's very similar to ascii->utf16 conversion that I accelerated in text here: haskell/text#298. I'd like to continue with other functions like ByteString.reverse.

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

My understanding is that there is no performance degradation...

@ethercrow could you please add a benchmark for non-aligned bytestrings?

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@Bodigrim what's a good way to create one that's surely unaligned? Take a slice of global one with an odd offset like 7?

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yeah, Data.ByteString.drop 1 should be enough.

const __m128i sixteen_dst_bytes = _mm_unpacklo_epi8(eight_src_bytes, separator);
_mm_storeu_si128((__m128i *)q, sixteen_dst_bytes);
p += 8;
q += 16;
}
n -= p - p_begin;
}
#endif
while (n > 1) {
*q++ = *p++;
*q++ = c;
Expand Down