| Message ID | AM5PR0802MB261043A55245CE7BF460ACA9833F0@AM5PR0802MB2610.eurprd08.prod.outlook.com |
|---|---|
| State | New |
| Headers | show |
On Thursday 23 March 2017 04:45 PM, Wilco Dijkstra wrote: > The 2nd version of this patch removes the made-up distribution and one based on a real > trace instead: > > Add a new randomized memcpy test for copies up to 256 bytes. The distribution of size and > alignment is based on a trace of SPEC2006 (other traces could be added in the future). > Instead of repeating the same copy over and over again like the existing tests, it times several > thousand different copies to more accurately estimate the overhead of branch prediction due to > the different sizes and alignments. > > ChangeLog: > 2017-02-09 Wilco Dijkstra <wdijkstr@arm.com> > > * benchtests/Makefile (string-benchset): Add memcpy-random. > * benchtests/bench-memcpy-random.c: New file. Nice, this looks good to me. There may be additional cache effects in spec2006 that may make these results different from actual spec2006, but it is still a valuable data point, definitely more than the repeated copies. Siddhesh
On Thu, 2017-03-23 at 19:53 +0530, Siddhesh Poyarekar wrote: > > > > ChangeLog: > > 2017-02-09 Wilco Dijkstra <wdijkstr@arm.com> > > > > * benchtests/Makefile (string-benchset): Add memcpy-random. > > * benchtests/bench-memcpy-random.c: New file. > Nice, this looks good to me. There may be additional cache effects in > spec2006 that may make these results different from actual spec2006, but > it is still a valuable data point, definitely more than the repeated > copies. > > Siddhesh I think there is a Makefile problem in this patch. make -C ../src/glibc/benchtests objdir=`pwd` bench make[1]: Entering directory '/home/ubuntu/sellcey/glibc- ifunc/src/glibc/benchtests' make[1]: *** No rule to make target '/home/ubuntu/sellcey/glibc- ifunc/obj-glibc64/benchtests/bench-memcpy-random.o', needed by '/home/ubuntu/sellcey/glibc-ifunc/obj-glibc64/benchtests/bench-memcpy- random'. Stop. make[1]: Leaving directory '/home/ubuntu/sellcey/glibc- ifunc/src/glibc/benchtests' Makefile:16: recipe for target 'bench' failed make: *** [bench] Error 2
It looks like the Makefile change got pushed out but the new benchtests/bench-memcpy-random.c file did not. Steve Ellcey On Fri, 2017-03-24 at 14:49 -0700, Steve Ellcey wrote: > On Thu, 2017-03-23 at 19:53 +0530, Siddhesh Poyarekar wrote: > > > > ChangeLog: > > > 2017-02-09 Wilco Dijkstra <wdijkstr@arm.com> > > > > > > * benchtests/Makefile (string-benchset): Add memcpy- > > > random. > > > * benchtests/bench-memcpy-random.c: New file. > > Nice, this looks good to me. There may be additional cache effects in > > spec2006 that may make these results different from actual spec2006, but > > it is still a valuable data point, definitely more than the repeated > > copies. > > > > Siddhesh > I think there is a Makefile problem in this patch. > > > make -C ../src/glibc/benchtests objdir=`pwd` bench > make[1]: Entering directory '/home/ubuntu/sellcey/glibc- > ifunc/src/glibc/benchtests' > make[1]: *** No rule to make target '/home/ubuntu/sellcey/glibc- > ifunc/obj-glibc64/benchtests/bench-memcpy-random.o', needed by > '/home/ubuntu/sellcey/glibc-ifunc/obj-glibc64/benchtests/bench- > memcpy- > random'. Stop. > make[1]: Leaving directory '/home/ubuntu/sellcey/glibc- > ifunc/src/glibc/benchtests' > Makefile:16: recipe for target 'bench' failed > make: *** [bench] Error 2
Hi Steve, > It looks like the Makefile change got pushed out but the > new benchtests/bench-memcpy-random.c file did not. Yes it seems I forgot to do a git add of the actual file... Sorry about that. Since I'm on holiday with no access to my GLIBC tree, it's probably best to revert the Makefile change for now and I'll redo it when I get back. Cheers, Wilco
On Sunday 26 March 2017 12:54 AM, Wilco Dijkstra wrote: > Yes it seems I forgot to do a git add of the actual file... Sorry about that. > > Since I'm on holiday with no access to my GLIBC tree, it's probably best > to revert the Makefile change for now and I'll redo it when I get back. Don't worry about it, I'll push the file. Siddhesh
On Sun, 2017-03-26 at 18:12 +0530, Siddhesh Poyarekar wrote: > On Sunday 26 March 2017 12:54 AM, Wilco Dijkstra wrote: > > > > Yes it seems I forgot to do a git add of the actual file... Sorry > > about that. > > > > Since I'm on holiday with no access to my GLIBC tree, it's probably best > > to revert the Makefile change for now and I'll redo it when I get back. > Don't worry about it, I'll push the file. > > Siddhesh I see the bench-memcpy-random.c file now but I think there is still something wrong. I am using a glibc that has my aarch64 IFUNC version of memcpy and memmove implemented in it but I do not get measurements of the different implementations for the random memcpy test like I do for the other two memcpy tests. In bench-memcpy.out I have: builtin_memcpy simple_memcpy __memcpy_thunderx __memcpy_generic Length 1, alignment 0/ 0: 39.2188 19.375 22.03 12 23.5938 Length 1, alignment 0/ 0: 27.3438 16.25 22.187 5 22.1875 . . In bench-memcpy-large.out I have: __memcpy_thunderx __memcpy_generi c Length 65543, alignment 0/ 0: 8937.5 17000 Length 65551, alignment 0/ 3: 27275.6 35756.9 . . In bench-memcpy-random.out I have: memcpy Memory size 4096: 105751 Memory size 8192: 103559 Memory size 16384: 112045 Memory size 32768: 118391 Memory size 65536: 158526 I.e. there are no seperate timings for the generic and thunderx versions of memcpy like I have in bench-memcpy and bench-memcpy-large. There is code in bench-memcpy-random.c to do the different measurements, I see 'FOR_EACH_IMPL' but it doesn't seeem to be working. Has anyone else seen this? Steve Ellcey
diff --git a/benchtests/Makefile b/benchtests/Makefile index 81edf8a933ce7371ac60c118a4a46daee6391800..a96e9533b3b2d7223c2ce90a723b6a434ba1a1ea 100644 --- a/benchtests/Makefile +++ b/benchtests/Makefile @@ -37,7 +37,7 @@ string-benchset := bcopy bzero memccpy memchr memcmp memcpy memmem memmove \ strcat strchr strchrnul strcmp strcpy strcspn strlen \ strncasecmp strncat strncmp strncpy strnlen strpbrk strrchr \ strspn strstr strcpy_chk stpcpy_chk memrchr strsep strtok \ - strcoll memcpy-large memmove-large memset-large + strcoll memcpy-large memcpy-random memmove-large memset-large # Build and run locale-dependent benchmarks only if we're building natively. ifeq (no,$(cross-compiling)) diff --git a/benchtests/bench-memcpy-random.c b/benchtests/bench-memcpy-random.c new file mode 100644 index 0000000000000000000000000000000000000000..4cd620966d4e2b6790830ab005074c2b19ab922c --- /dev/null +++ b/benchtests/bench-memcpy-random.c @@ -0,0 +1,157 @@ +/* Measure memcpy performance. + Copyright (C) 2016-2017 Free Software Foundation, Inc. + This file is part of the GNU C Library. + + The GNU C Library is free software; you can redistribute it and/or + modify it under the terms of the GNU Lesser General Public + License as published by the Free Software Foundation; either + version 2.1 of the License, or (at your option) any later version. + + The GNU C Library is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + Lesser General Public License for more details. + + You should have received a copy of the GNU Lesser General Public + License along with the GNU C Library; if not, see + <http://www.gnu.org/licenses/>. */ + +#define MIN_PAGE_SIZE 131072 +#define TEST_MAIN +#define TEST_NAME "memcpy-random" +#include "bench-string.h" +#include <assert.h> + +IMPL (memcpy, 0) + +#define NUM_COPIES 4096 + +typedef struct { uint16_t size; uint16_t freq; } freq_data_t; +typedef struct { uint8_t align; uint8_t freq; } align_data_t; + +#define SIZE_NUM 1024 +#define SIZE_MASK (SIZE_NUM-1) +static uint8_t size_arr[SIZE_NUM]; + +/* Frequency data for memcpy of less than 256 bytes based on SPEC2006. */ +static freq_data_t size_freq[] = +{ + { 8, 576}, {104, 94}, { 24, 78}, { 48, 58}, { 32, 48}, { 16, 46}, + { 1, 30}, { 96, 12}, { 72, 11}, {216, 11}, {192, 8}, { 12, 7}, + {144, 5}, { 2, 4}, { 64, 4}, {120, 4}, { 4, 3}, { 40, 2}, + { 7, 2}, {168, 2}, {160, 2}, {128, 1}, { 3, 1}, { 9, 1}, + {176, 1}, {240, 1}, { 11, 1}, { 0, 1}, { 5, 1}, { 6, 1}, + { 80, 1}, { 52, 1}, {152, 1}, { 10, 1}, { 56, 1}, { 51, 1}, + { 14, 1}, {208, 1}, { 0, 0} +}; + +#define ALIGN_NUM 256 +#define ALIGN_MASK (ALIGN_NUM-1) +static uint8_t src_align_arr[ALIGN_NUM]; +static uint8_t dst_align_arr[ALIGN_NUM]; + +/* Source alignment frequency for memcpy based on SPEC2006. */ +static align_data_t src_align_freq[] = +{ + {16, 144}, {8, 86}, {3, 23}, {1, 3}, {0, 0} +}; + +/* Destination alignment frequency for memcpy based on SPEC2006. */ +static align_data_t dst_align_freq[] = +{ + {16, 197}, {8, 30}, {3, 23}, {1, 6}, {0, 0} +}; + +typedef struct +{ + uint16_t src; + uint16_t dst; + uint16_t len; +} copy_t; + +static copy_t copy[NUM_COPIES]; + +typedef char *(*proto_t) (char *, const char *, size_t); + +static void +init_copy_distribution (void) +{ + int i, j, freq, size, n; + + for (n = i = 0; (freq = size_freq[i].freq) != 0; i++) + for (j = 0, size = size_freq[i].size; j < freq; j++) + size_arr[n++] = size; + assert (n == SIZE_NUM); + + for (n = i = 0; (freq = src_align_freq[i].freq) != 0; i++) + for (j = 0, size = src_align_freq[i].align; j < freq; j++) + src_align_arr[n++] = size - 1; + assert (n == ALIGN_NUM); + + for (n = i = 0; (freq = dst_align_freq[i].freq) != 0; i++) + for (j = 0, size = dst_align_freq[i].align; j < freq; j++) + dst_align_arr[n++] = size - 1; + assert (n == ALIGN_NUM); +} + + +static void +do_one_test (impl_t *impl, char *dst, char *src, copy_t *copy, size_t n) +{ + timing_t start, stop, cur; + size_t iters = INNER_LOOP_ITERS * 20; + + TIMING_NOW (start); + for (int i = 0; i < iters; ++i) + for (int j = 0; j < n; j++) + CALL (impl, dst + copy[j].dst, src + copy[j].src, copy[j].len); + TIMING_NOW (stop); + + TIMING_DIFF (cur, start, stop); + + TIMING_PRINT_MEAN ((double) cur, (double) iters); +} + +static void +do_test (size_t max_size) +{ + for (int i = 0; i < max_size; i++) + buf1[i] = i * 3; + + /* Create a random set of copies with the given size and alignment + distributions. */ + for (int i = 0; i < NUM_COPIES; i++) + { + copy[i].dst = (rand () & (max_size - 1)) | 1; + copy[i].dst &= ~dst_align_arr[rand () & ALIGN_MASK]; + copy[i].src = (rand () & (max_size - 1)) | 3; + copy[i].src &= ~src_align_arr[rand () & ALIGN_MASK]; + copy[i].len = size_arr[rand () & SIZE_MASK]; + } + + printf ("Memory size %6zd:", max_size); + + FOR_EACH_IMPL (impl, 0) + do_one_test (impl, (char *) buf2, (char *) buf1, copy, NUM_COPIES); + + putchar ('\n'); +} + +int +test_main (void) +{ + test_init (); + init_copy_distribution (); + + printf ("%23s", ""); + FOR_EACH_IMPL (impl, 0) + printf ("\t%s", impl->name); + putchar ('\n'); + + for (int i = 4; i <= 64; i = i * 2) + do_test (i * 1024); + + return ret; +} + +#include <support/test-driver.c>