Re:HighwayHash: Strong hashing at 0.3 clock cycles per byte
by amock on Wednesday March 02, 2016 @07:33PM
(#11214240)
I'm guessing it was built and tested internally at Google and no one tested it with bazel. It also depends on smasher, which I believe should be declared in the WORKSPACE file. So unless you want to manually build it you'll have to wait for someone to fix the build.
[ Parent
| on HN ]
Re:HighwayHash: Strong hashing at 0.3 clock cycles per byte
by nkurz on Wednesday March 02, 2016 @08:48PM
(#11214514)
I went ahead and created a Makefile:
CXXFLAGS := -g -std=c++11 -Wall -O3 -march=native -Wno-sign-compare
OBJS := highway_tree_hash.o \
scalar_highway_tree_hash.o \
scalar_sip_tree_hash.o \
sip_hash.o \
sip_tree_hash.o
MAIN := sip_hash_main
all: $(OBJS)
$(CXX) $(CXXFLAGS) $(OBJS) $(MAIN).cc -o $(MAIN)
clean:
rm -f $(OBJS) $(MAIN)
Then I commented out two lines in "sip_hash_main.cc": // #include "smhasher.h"
// RunAllTests(argc, argv);
Then "make" and run: nate@skylake:~/git/highwayhash$ ./sip_hash_main
ScalarSipTreeHash ... GBps=2.44 c/b=1.44
ScalarHighwayTreeHash ... GBps=1.16 c/b=3.02
SipHash ... GBps=1.54 c/b=2.27
SipTreeHash ... GBps=4.56 c/b=0.77
HighwayTreeHash ... GBps=11.49 c/b=0.30
...
The inner loop for HighwayTreeHash looks like this: 0.89 │ 60:┌─→vpor %ymm5,%ymm0,%ymm0
4.47 │ │ vpaddq (%rcx),%ymm1,%ymm1
6.04 │ │ add $0x4,%rax
0.22 │ │ add $0x20,%rcx
2.24 │ │ cmp %rax,%rdi
3.13 │ │ vpsrlq $0x20,%ymm1,%ymm2
3.36 │ │ vpmulu %ymm1,%ymm0,%ymm4
13.20 │ │ vpshuf %ymm3,%ymm4,%ymm4
11.19 │ │ vpmulu %ymm2,%ymm0,%ymm2
│ │ vpaddq %ymm1,%ymm2,%ymm1
31.32 │ │ vpaddq %ymm4,%ymm0,%ymm0
7.61 │ └──ja 60
Re:HighwayHash: Strong hashing at 0.3 clock cycles per byte
by forgotpwtomain on Thursday March 03, 2016 @11:03AM
(#11217671)
Do you know if there is a reference for a comparative number of collisions anywhere e.g. (http://programmers.stackexchange.com/questions/49550/which-h...) ?
Re:HighwayHash: Strong hashing at 0.3 clock cycles per byte
by janwas on Thursday March 03, 2016 @12:55PM
(#11218538)
With a 64-bit hash, you shouldn't be seeing collisions at all on small datasets. A test such as smhasher can verify this.