igt/gem_userptr_benchmark: Testing of overlapping synchronized objects
PATCH 1/1 igt/gem_userptr_benchmark: Testing of overlapping synchronized objects
Larger performance impact is expected with first draft of the kernel implementation for overlapping objects so this is just to confirm that.
Signed-off-by Tvrtko Ursulin <[email protected]>
Cc Chris Wilson <[email protected]>
--- benchmarks/gem_userptr_benchmark.c | 97 ++++++++++++++++++++++++++++---------- 1 file changed, 72 insertions(+), 25 deletions(-)
benchmarks/gem_userptr_benchmark.c +72 -25
@@ -213,7 +213,8 @@ static void exchange_ptr(void *array, unsigned i, unsigned j)static void test_malloc_free (int random) {unsigned long iter =0 ; -unsigned int i, tot =1000 ; +unsigned int i; +const unsigned int tot =1000 ;void *ptr[tot];start_test (test_duration_sec);
@@ -230,13 +231,14 @@ static void test_malloc_free(int random) iter++; } -printf ("%8lu iter/s\n" , iter / test_duration_sec); +printf ("%10lu op/s\n" ,2 * tot * iter / test_duration_sec); }static void test_malloc_realloc_free (int random) {unsigned long iter =0 ; -unsigned int i, tot =1000 ; +unsigned int i; +const unsigned int tot =1000 ;void *ptr[tot];start_test (test_duration_sec);
@@ -259,13 +261,14 @@ static void test_malloc_realloc_free(int random) iter++; } -printf ("%8lu iter/s\n" , iter / test_duration_sec); +printf ("%10lu op/s\n" ,3 * tot * iter / test_duration_sec); }static void test_mmap_unmap (int random) {unsigned long iter =0 ; -unsigned int i, tot =1000 ; +unsigned int i; +const unsigned int tot =1000 ;void *ptr[tot];start_test (test_duration_sec);
@@ -283,7 +286,7 @@ static void test_mmap_unmap(int random) iter++; } -printf ("%8lu iter/s\n" , iter / test_duration_sec); +printf ("%10lu op/s\n" ,2 * tot * iter / test_duration_sec); }static void test_ptr_read (void *ptr)
@@ -307,7 +310,7 @@ static void test_ptr_read(void *ptr) iter++; } -printf ("%8lu MB/s\n" , iter / test_duration_sec *BO_SIZE /1000000 ); +printf ("%10lu MB/s\n" , iter / test_duration_sec *BO_SIZE /1000000 ); }static void test_ptr_write (void *ptr)
@@ -331,7 +334,31 @@ static void test_ptr_write(void *ptr) iter++; } -printf ("%8lu MB/s\n" , iter / test_duration_sec *BO_SIZE /1000000 ); +printf ("%10lu MB/s\n" , iter / test_duration_sec *BO_SIZE /1000000 ); +} + +static void test_impact_common (void *ptr,unsigned int nr) +{ +printf ("ptr-read, %5u bos = " , nr); +test_ptr_read (ptr); + +printf ("ptr-write, %5u bos = " , nr); +test_ptr_write (ptr); + +printf ("malloc-free, %5u bos = " , nr); +test_malloc_free (0 ); +printf ("malloc-free-random, %5u bos = " , nr); +test_malloc_free (1 ); + +printf ("malloc-realloc-free, %5u bos = " , nr); +test_malloc_realloc_free (0 ); +printf ("malloc-realloc-free-random, %5u bos = " , nr); +test_malloc_realloc_free (1 ); + +printf ("mmap-unmap, %5u bos = " , nr); +test_mmap_unmap (0 ); +printf ("mmap-unmap-random, %5u bos = " , nr); +test_mmap_unmap (1 ); }static void test_impact (int fd)
@@ -351,29 +378,46 @@ static void test_impact(int fd)else ptr = buffer; -printf ("ptr-read, %5u bos = " , nr_bos[subtest]); -test_ptr_read (ptr); +test_impact_common (ptr, nr_bos[subtest]); -printf ("ptr-write %5u bos = " , nr_bos[subtest]); -test_ptr_write (ptr); +for (i =0 ; i < nr_bos[subtest]; i++) +free_userptr_bo (fd, handles[i]); + } +} -printf ("malloc-free, %5u bos = " , nr_bos[subtest]); -test_malloc_free (0 ); -printf ("malloc-free-random %5u bos = " , nr_bos[subtest]); -test_malloc_free (1 ); +static void test_impact_overlap (int fd) +{ +unsigned int total =sizeof (nr_bos) /sizeof (nr_bos[0 ]); +unsigned int subtest, i; +uint32_t handles[nr_bos[total-1 ]]; +void *ptr; +char buffer[BO_SIZE ]; +void *block; +uint32_t handle; +int ret; -printf ("malloc-realloc-free, %5u bos = " , nr_bos[subtest]); -test_malloc_realloc_free (0 ); -printf ("malloc-realloc-free-random, %5u bos = " , nr_bos[subtest]); -test_malloc_realloc_free (1 ); +for (subtest =0 ; subtest < total; subtest++) { +if (nr_bos[subtest] >0 ) { + ret =posix_memalign (&block,PAGE_SIZE ,BO_SIZE * nr_bos[subtest]); +igt_assert (ret ==0 ); +for (ptr = block, i =0 ; i < nr_bos[subtest]; i++) { + ret =gem_userptr (fd, ptr,BO_SIZE ,0 , &handle); +igt_assert (ret ==0 ); + handles[i] = handle; + ptr = (char *)ptr +BO_SIZE /2 ; + } + ptr = block; + }else { + ptr = buffer; + } -printf ("mmap-unmap, %5u bos = " , nr_bos[subtest]); -test_mmap_unmap (0 ); -printf ("mmap-unmap-random, %5u bos = " , nr_bos[subtest]); -test_mmap_unmap (1 ); +test_impact_common (ptr, nr_bos[subtest]);for (i =0 ; i < nr_bos[subtest]; i++) -free_userptr_bo (fd, handles[i]); +gem_close (fd, handles[i]); + +if (nr_bos[subtest]) +free (block); } }
@@ -494,6 +538,9 @@ int main(int argc, char **argv) igt_subtest("userptr-impact-sync")test_impact (fd); + igt_subtest("userptr-overlap-impact-sync") +test_impact_overlap (fd); +igt_exit ();return 0 ;
Signature · 1 line
1.9.3
Quoted 5 lines · Tvrtko Ursulin
Larger performance impact is expected with first draft of the kernel implementation for overlapping objects so this is just to confirm that.
Signed-off-by Tvrtko Ursulin <[email protected]>Cc Chris Wilson <[email protected]>
Looks fine. I can't comment on likely overlapping arrangements, so your overlap with adjacent neighbours seems a sensible first choice. -Chris
Signature · 1 line
Chris Wilson, Intel Open Source Technology Centre