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

Jump to bottom

CRC32 implementation with gcc intrinsics instead pure asm. #4

Merged
merged 16 commits into from
Jan 28, 2018
Merged

CRC32 implementation with gcc intrinsics instead pure asm. #4

merged 16 commits into from
Jan 28, 2018

Conversation

racardoso
Copy link

This commit implements CRC32 using power8 vector intrinsics
and gcc builtins instead pure assembly. The performance is
the same compared to .S version:

time ./vec_crc32_bench 32768 5000000
CRC: 165b4c91
real 0m2.799s
user 0m2.799s
sys 0m0.000s

time ./crc32_bench 32768 5000000
CRC: 165b4c91
real 0m2.803s
user 0m2.803s
sys 0m0.000s

Perf results:

perf stat -a ./vec_crc32_bench 32768 5000000
CRC: 165b4c91
Performance counter stats for 'system wide':
360774.660732 task-clock (msec) # 128.683 CPUs utilized
529 context-switches # 0.001 K/sec
8 cpu-migrations # 0.000 K/sec
208 page-faults # 0.001 K/sec
12,468,436,530 cycles # 0.035 GHz (66.62%)
18,068,249 stalled-cycles-frontend # 0.14% cycles idle
466,739,548 stalled-cycles-backend # 3.74% cycles idle
49,670,139,591 instructions # 3.98 insns per cycle
# 0.01 stalled cycles
per insn (66.82%)
1,370,729,619 branches # 3.799 M/sec (50.09%)
5,759,980 branch-misses # 0.42% of all branches

2.803581718 seconds time elapsed

perf stat -a ./crc32_bench 32768 5000000
CRC: 165b4c91
Performance counter stats for 'system wide':
360942.638504 task-clock (msec) # 128.498 CPUs utilized
535 context-switches # 0.001 K/sec
12 cpu-migrations # 0.000 K/sec
287 page-faults # 0.001 K/sec
12,476,309,108 cycles # 0.035 GHz (66.67%)
17,688,340 stalled-cycles-frontend # 0.14% cycles idle
477,872,611 stalled-cycles-backend # 3.83% cycles idle
48,459,294,347 instructions # 3.88 insns per cycle
# 0.01 stalled cycles
per insn (66.69%)
1,371,856,316 branches # 3.801 M/sec (50.01%)
5,771,271 branch-misses # 0.42% of all branches

2.808943029 seconds time elapsed

Tested on (tulibee): P8 / LE DD2.1 Murano 32G RAM, 16 Cores.
RHEL7.2 LE

@grooverdan
Copy link
Contributor

nice! Solved difficulties with clang missing ppc headers too.

@racardoso
Copy link
Author

Hi @antonblanchard @grooverdan . Any news about this PR?

@grooverdan grooverdan mentioned this pull request Aug 22, 2017
Merged
Rogerio Alves and others added 6 commits August 22, 2017 10:42
CRC32 implementation with gcc intrinsics instead pure asm.
76948c0 
This commit implements CRC32 using power8 vector intrinsics
and gcc builtins instead pure assembly. The performance is
the same compared to .S version:

time ./vec_crc32_bench 32768 5000000
CRC: 165b4c91
real  0m2.799s
user  0m2.799s
sys   0m0.000s

time ./crc32_bench 32768 5000000
CRC: 165b4c91
real  0m2.803s
user  0m2.803s
sys   0m0.000s

Perf results:

perf stat -a ./vec_crc32_bench 32768 5000000
CRC: 165b4c91
Performance counter stats for 'system wide':
360774.660732   task-clock (msec)   #   128.683 CPUs utilized
529             context-switches    #   0.001 K/sec
8               cpu-migrations      #   0.000 K/sec
208             page-faults         #   0.001 K/sec
12,468,436,530  cycles              #   0.035 GHz (66.62%)
18,068,249      stalled-cycles-frontend #   0.14% cycles idle
466,739,548     stalled-cycles-backend  #   3.74% cycles idle
49,670,139,591  instructions        #   3.98  insns per cycle
                                    #   0.01  stalled cycles
                                    per insn  (66.82%)
1,370,729,619  branches             #   3.799 M/sec (50.09%)
5,759,980      branch-misses        #   0.42% of all branches

2.803581718 seconds time elapsed

perf stat -a ./crc32_bench 32768 5000000
CRC: 165b4c91
Performance counter stats for 'system wide':
360942.638504   task-clock (msec)   #   128.498 CPUs utilized
535             context-switches    #   0.001 K/sec
12              cpu-migrations      #   0.000 K/sec
287             page-faults         #   0.001 K/sec
12,476,309,108  cycles              #   0.035 GHz (66.67%)
17,688,340      stalled-cycles-frontend #   0.14% cycles idle
477,872,611     stalled-cycles-backend  #   3.83% cycles idle
48,459,294,347  instructions        #   3.88  insns per cycle
                                    #   0.01  stalled cycles
                                        per insn  (66.69%)
1,371,856,316   branches            #   3.801 M/sec (50.01%)
5,771,271       branch-misses       #   0.42% of all branches

2.808943029 seconds time elapsed

Tested on (tulibee): P8 / LE DD2.1 Murano 32G RAM, 16 Cores.
RHEL7.2 LE

Signed-off-by: Rogerio Alves <rogealve@br.ibm.com>
Added a note about vec_crc32 use on README.
bc4b3bd 
Included quickstart instruction for vec_crc32.c on README.

Signed-off-by: Rogerio Alves <rogealve@br.ibm.com>
@grooverdan
Add generation command in crc32_constants header
737e20c 
Signed-off-by: Daniel Black <daniel.black@au.ibm.com>
@grooverdan
crc32_constants: reorder defines in generated crc32_constants.h
b3703d3 
This ensures that:

defining __ASSEMBLY__ (gcc builtin) isn't needed for C implementation.

MAX_SIZE is defined in both C and __ASSEMBLY__ generations

Signed-off-by: Daniel Black <daniel.black@au.ibm.com>
@grooverdan
include stdio.h for printf where needed
d214b66 
Signed-off-by: Daniel Black <daniel.black@au.ibm.com>
@grooverdan
Makefile: allow crc32_constants to be compiled on non-Power architect…
8d0e3f6 
…ures

Signed-off-by: Daniel Black <daniel.black@au.ibm.com>
@racardoso
Copy link
Author

Updated PR with @grooverdan review and contribution.

@grooverdan
Copy link
Contributor

https://gcc.gnu.org/onlinedocs/gcc-7.2.0/gcc/PowerPC-AltiVec_002fVSX-Built-in-Functions.html#PowerPC-AltiVec_002fVSX-Built-in-Functions indicates -mvsx needed for vector unsigned long (though in practice I never saw any warning). I'm not particular worried about it however.

Also lodged https://bugs.llvm.org/show_bug.cgi?id=34295 and https://bugs.llvm.org/show_bug.cgi?id=34296 as it currently doesn't compile with bleeding edge clang.

@grooverdan @racardoso
crc32_constants: generate cross-compilable header files
be0dd60 
Signed-off-by: Daniel Black <daniel.black@au.ibm.com>
@grooverdan @racardoso
Readme: acknowledge Rogerio Alves and Daniel Black
1da1dd4 
Signed-off-by: Daniel Black <daniel.black@au.ibm.com>
@grooverdan @racardoso
Fold vec_crc32_wrapper.c into vec_crc32.c
b64bd99 
Signed-off-by: Daniel Black <daniel.black@au.ibm.com>
@grooverdan @racardoso
Remove duplicate vec_crc32_{bench|test}.c code
8661035 
Signed-off-by: Daniel Black <daniel.black@au.ibm.com>
@grooverdan @racardoso
Use argv[0] in helper usage() functions and describe purpose
42a50fe 
Signed-off-by: Daniel Black <daniel.black@au.ibm.com>
@grooverdan @racardoso
crc32_constants: add C and/or Assembler options to generation
a477389 
Signed-off-by: Daniel Black <daniel.black@au.ibm.com>
@grooverdan @racardoso
slice_by_8_bench: use common crc32_bench code
aadb491 
Signed-off-by: Daniel Black <daniel.black@au.ibm.com>
@grooverdan @racardoso
Add configurability to crc32 function name and constants header
8d37f60 
Add example crc32_two_implementations on how to use this.

Signed-off-by: Daniel Black <daniel.black@au.ibm.com>
@grooverdan @racardoso
__ASSEMBLER__ is the buildin define for assembler builds
c812bcd 
Signed-off-by: Daniel Black <daniel.black@au.ibm.com>
@grooverdan grooverdan mentioned this pull request Sep 28, 2017
Closed
@grooverdan @racardoso
crc32_test: test C and ASM for various alignments
8220229 
This rolls vec_crc32_test into the crc32_test by comparing the
reference, ASM and C vpmsum CRC32 implementations.

Compares the result at up to 16 alignments as the codepath with change
for this.

A Makefile target of test is added to test the boundary conditions
of the implementations.

Signed-off-by: Daniel Black <daniel.black@au.ibm.com>
@grooverdan grooverdan mentioned this pull request Jan 16, 2018
Closed
@antonblanchard antonblanchard merged commit 90c45d7 into antonblanchard:master Jan 28, 2018
@antonblanchard
Copy link
Owner

Thanks Rogerio and Daniel!

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Reviewers
No reviews
Assignees
No one assigned
Labels
None yet
Projects
None yet
Milestone
No milestone
Development

Successfully merging this pull request may close these issues.
3 participants
@racardoso @grooverdan @antonblanchard