gem_fence_upload: Move to benchmarks
PATCH 1/2 gem_fence_upload: Move to benchmarks
From: Tvrtko Ursulin <[email protected]>
This is not a correctness test so move it to benchmarks and off the extended test list.
---
benchmarks/.gitignore | 1 +
benchmarks/Makefile.am | 3 +++
benchmarks/Makefile.sources | 1 +
{tests => benchmarks}/gem_fence_upload.c | 0
tests/Makefile.am | 2 --
tests/Makefile.sources | 1 -
tests/intel-ci/extended.testlist | 5 -----
7 files changed, 5 insertions(+), 8 deletions(-)
rename {tests => benchmarks}/gem_fence_upload.c (100%)benchmarks/.gitignore +1 -0
@@ -7,6 +7,7 @@ gem_exec_fault gem_exec_nop gem_exec_reloc gem_exec_trace +gem_fence_upload gem_latency gem_mmap gem_prw
benchmarks/Makefile.am +3 -0
@@ -15,6 +15,9 @@ benchmarks_LTLIBRARIES = gem_exec_tracer.la gem_exec_tracer_la_LDFLAGS = -module -avoid-version -no-undefined gem_exec_tracer_la_LIBADD = -ldl +gem_fence_upload_CFLAGS = $(AM_CFLAGS) $(THREAD_CFLAGS) +gem_fence_upload_LDADD = $(LDADD) -lpthread + gem_latency_CFLAGS = $(AM_CFLAGS) $(THREAD_CFLAGS) gem_latency_LDADD = $(LDADD) -lpthread gem_syslatency_CFLAGS = $(AM_CFLAGS) $(THREAD_CFLAGS)
benchmarks/Makefile.sources +1 -0
@@ -9,6 +9,7 @@ benchmarks_prog_list = \ gem_exec_nop \ gem_exec_reloc \ gem_exec_trace \ + gem_fence_upload \ gem_latency \ gem_mmap \ gem_prw \
benchmarks/gem_fence_upload.c renamed from tests/gem_fence_upload.c +0 -0
tests/Makefile.am +0 -2
@@ -89,8 +89,6 @@ gem_exec_parallel_CFLAGS = $(AM_CFLAGS) $(THREAD_CFLAGS) gem_exec_parallel_LDADD = $(LDADD) -lpthread gem_fence_thrash_CFLAGS = $(AM_CFLAGS) $(THREAD_CFLAGS) gem_fence_thrash_LDADD = $(LDADD) -lpthread -gem_fence_upload_CFLAGS = $(AM_CFLAGS) $(THREAD_CFLAGS) -gem_fence_upload_LDADD = $(LDADD) -lpthread gem_flink_race_CFLAGS = $(AM_CFLAGS) $(THREAD_CFLAGS) gem_flink_race_LDADD = $(LDADD) -lpthread gem_mmap_gtt_CFLAGS = $(AM_CFLAGS) $(THREAD_CFLAGS)
tests/Makefile.sources +0 -1
@@ -63,7 +63,6 @@ TESTS_progs_M = \ gem_exec_whisper \ gem_fenced_exec_thrash \ gem_fence_thrash \ - gem_fence_upload \ gem_flink_basic \ gem_flink_race \ gem_linear_blits \
tests/intel-ci/extended.testlist +0 -5
@@ -1446,11 +1446,6 @@ igt@gem_fence_thrash@bo-write-verify-threaded-x igt@gem_fence_thrash@bo-write-verify-threaded-y igt@gem_fence_thrash@bo-write-verify-x igt@gem_fence_thrash@bo-write-verify-y -igt@gem_fence_upload@performance -igt@gem_fence_upload@thread-contention -igt@gem_fence_upload@thread-performance-both -igt@gem_fence_upload@thread-performance-read -igt@gem_fence_upload@thread-performance-write igt@gem_fenced_exec_thrash@2-spare-fences igt@gem_fenced_exec_thrash@no-spare-fences igt@gem_fenced_exec_thrash@no-spare-fences-busy
Signature · 6 lines
2.9.4 _______________________________________________ Intel-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/intel-gfx
PATCH 2/2 igt/gem_fence_upload: Stabilise results
From: Tvrtko Ursulin <[email protected]>
Most of the subtest were failing on my SKL GT2 and on the various CI systems as well. Try to fix that by different tweaks per subtests:
performance:
We cannot say how big the performance drop will be once the fences are contented, just that there will be one, so modify the assert accordingly.
thread-performance-*:
Similar to "performance", but also on some machine the peak performance to compare against seems to be num_cpu dependant which the biggest aggregate throughput with num_cpu clients.
On other machines the results are more stable so this change should not affect those.
Also, on some machines writes are much faster than reads which was making the "-write" subtest a bit unstable. Increased the number of iterations to hopefully stabilise this.
thread-contention:
Same as "performance" - we don't know how big the drop to expect.
wc-contention:
This subtest should not expect a tiled performance drop since CPU WC maps are under test. Change a few things:
1.
Number of iterations was to small causing result instability on some machines.
2.
There seems to be a performance valley with num_threads == num_cpus so do a pre-warm step in those cases which seem to fix it.
3.
Only assert that the final step is slower than the reference, and for the reference take the single thred result which looks to be the fastest.
4.
Check that linear vs tiled results in each step are within 15% tolerance to verify the above attempts to stabilise the test.
--- benchmarks/gem_fence_upload.c | 51 ++++++++++++++++++++++++++++--------------- 1 file changed, 33 insertions(+), 18 deletions(-)
benchmarks/gem_fence_upload.c +33 -18
@@ -103,7 +103,7 @@ static void performance(void) errno =0 ;igt_assert (linear[1 ] >0.75 * linear[0 ]); -igt_assert (tiled[1 ] >0.75 * tiled[0 ]); +igt_assert (tiled[1 ] < tiled[0 ]); }struct thread_performance {
@@ -155,7 +155,8 @@ static const char *direction_string(unsigned mask) }static void thread_performance (unsigned mask) { -const int loops =4096 ; +const int ncpus =sysconf (_SC_NPROCESSORS_ONLN); +const int loops = (mask &READ ) ?4096 :32768 ;int n, count;int fd, num_fences;double linear[2 ], tiled[2 ];
@@ -164,9 +165,11 @@ static void thread_performance(unsigned mask) num_fences =gem_available_fences (fd);igt_require (num_fences >0 ); +igt_require (num_fences > ncpus);for (count =2 ; count <4 *num_fences; count *=2 ) {const int nthreads = (mask &READ ? count :0 ) + (mask &WRITE ? count :0 ); +const int idx = ncpus != count;struct timeval start, end;struct thread_performance readers[count];struct thread_performance writers[count];
@@ -209,8 +212,8 @@ static void thread_performance(unsigned mask) }gettimeofday (&end,NULL ); - linear[count !=2 ] = nthreads * loops /elapsed (&start, &end) / (OBJECT_SIZE /4096 ); -igt_info ("%s rate for %d linear surfaces, %d threads: %7.3fMiB/s\n" ,direction_string (mask), count, nthreads, linear[count !=2 ]); + linear[idx] = nthreads * loops /elapsed (&start, &end) / (OBJECT_SIZE /4096 ); +igt_info ("%s rate for %d linear surfaces, %d threads: %7.3fMiB/s\n" ,direction_string (mask), count, nthreads, linear[idx]);for (n =0 ; n < count; n++)gem_set_tiling (fd, handle[n],I915_TILING_X ,1024 );
@@ -230,8 +233,8 @@ static void thread_performance(unsigned mask) }gettimeofday (&end,NULL ); - tiled[count !=2 ] = nthreads * loops /elapsed (&start, &end) / (OBJECT_SIZE /4096 ); -igt_info ("%s rate for %d tiled surfaces, %d threads: %7.3fMiB/s\n" ,direction_string (mask), count, nthreads, tiled[count !=2 ]); + tiled[idx] = nthreads * loops /elapsed (&start, &end) / (OBJECT_SIZE /4096 ); +igt_info ("%s rate for %d tiled surfaces, %d threads: %7.3fMiB/s\n" ,direction_string (mask), count, nthreads, tiled[idx]);for (n =0 ; n < count; n++) {munmap (ptr[n],OBJECT_SIZE );
@@ -241,7 +244,7 @@ static void thread_performance(unsigned mask) errno =0 ;igt_assert (linear[1 ] >0.75 * linear[0 ]); -igt_assert (tiled[1 ] >0.75 * tiled[0 ]); +igt_assert (tiled[1 ] < tiled[0 ]); }struct thread_contention {
@@ -329,13 +332,13 @@ static void thread_contention(void) errno =0 ;igt_assert (linear[1 ] >0.75 * linear[0 ]); -igt_assert (tiled[1 ] >0.75 * tiled[0 ]); +igt_assert (tiled[1 ] < tiled[0 ]); }static void wc_contention (void ) { -const int loops =4096 ; -int n, count; +const int ncpus =sysconf (_SC_NPROCESSORS_ONLN); +int n, count, order;int fd, num_fences;double linear[2 ], tiled[2 ];
@@ -344,8 +347,13 @@ static void wc_contention(void) num_fences =gem_available_fences (fd);igt_require (num_fences >0 ); +igt_require (num_fences > ncpus); -for (count =1 ; count <4 *num_fences; count *=2 ) { +for (count =1 , order =log2 (num_fences *2 ) +1 ; + count <4 * num_fences; + count *=2 , order--) { +const int loops =4096 * order *2 ; +const int idx =1 != count;struct timeval start, end;struct thread_contention threads[count];
@@ -355,15 +363,20 @@ static void wc_contention(void) threads[n].fd = fd; } +if (count <= ncpus /2 ) { +for (n =0 ; n < count; n++) +pthread_create (&threads[n].thread,NULL , wc_mmap, &threads[n]); +for (n =0 ; n < count; n++) +pthread_join (threads[n].thread,NULL ); + }gettimeofday (&start,NULL );for (n =0 ; n < count; n++)pthread_create (&threads[n].thread,NULL , wc_mmap, &threads[n]);for (n =0 ; n < count; n++)pthread_join (threads[n].thread,NULL );gettimeofday (&end,NULL ); - - linear[count !=2 ] = count * loops /elapsed (&start, &end) / (OBJECT_SIZE /4096 ); -igt_info ("Contended upload rate for %d linear threads/wc: %7.3fMiB/s\n" , count, linear[count !=2 ]); + linear[idx] = count * loops /elapsed (&start, &end) / (OBJECT_SIZE /4096 ); +igt_info ("Contended upload rate for %d linear threads/wc: %7.3fMiB/s\n" , count, linear[idx]);for (n =0 ; n < count; n++)gem_set_tiling (fd, threads[n].handle,I915_TILING_X ,1024 );
@@ -375,17 +388,19 @@ static void wc_contention(void)pthread_join (threads[n].thread,NULL );gettimeofday (&end,NULL ); - tiled[count !=2 ] = count * loops /elapsed (&start, &end) / (OBJECT_SIZE /4096 ); -igt_info ("Contended upload rate for %d tiled threads/wc: %7.3fMiB/s\n" , count, tiled[count !=2 ]); + tiled[idx] = count * loops /elapsed (&start, &end) / (OBJECT_SIZE /4096 ); +igt_info ("Contended upload rate for %d tiled threads/wc: %7.3fMiB/s\n" , count, tiled[idx]);for (n =0 ; n < count; n++) {gem_close (fd, threads[n].handle); } + +igt_require (fabs ((linear[idx] - tiled[idx]) / linear[idx]) <0.15 ); } errno =0 ; -igt_assert (linear[1 ] >0.75 * linear[0 ]); -igt_assert (tiled[1 ] >0.75 * tiled[0 ]); +igt_assert (linear[1 ] < linear[0 ]); +igt_assert (tiled[1 ] < tiled[0 ]); } igt_main
Signature · 6 lines
2.9.4 _______________________________________________ Intel-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/intel-gfx