When looking at the pwrite/pread/wc performance, it is useful to judge that against the performance of an ordinary CPU mmap.
Signed-off-by: Chris Wilson <[email protected]> --- lib/drmtest.c | 15 ++++++++++- lib/drmtest.h | 6 ++++- tests/gem_gtt_speed.c | 72 +++++++++++++++++++++++++++++++++++++++++++------ tests/gem_mmap.c | 22 +++++++-------- 4 files changed, 94 insertions(+), 21 deletions(-) diff --git a/lib/drmtest.c b/lib/drmtest.c index f9f10e1..cdf46aa 100644 --- a/lib/drmtest.c +++ b/lib/drmtest.c @@ -311,7 +311,7 @@ uint32_t gem_create(int fd, int size) return create.handle; } -void *gem_mmap(int fd, uint32_t handle, int size, int prot) +void *gem_mmap__gtt(int fd, uint32_t handle, int size, int prot) { struct drm_i915_gem_mmap_gtt mmap_arg; void *ptr; @@ -327,6 +327,19 @@ void *gem_mmap(int fd, uint32_t handle, int size, int prot) return ptr; } +void *gem_mmap__cpu(int fd, uint32_t handle, int size, int prot) +{ + struct drm_i915_gem_mmap mmap_arg; + + mmap_arg.handle = handle; + mmap_arg.offset = 0; + mmap_arg.size = size; + if (drmIoctl(fd, DRM_IOCTL_I915_GEM_MMAP, &mmap_arg)) + return NULL; + + return (void *)(uintptr_t)mmap_arg.addr_ptr; +} + uint64_t gem_aperture_size(int fd) { struct drm_i915_gem_get_aperture aperture; diff --git a/lib/drmtest.h b/lib/drmtest.h index 9f27e2c..4021104 100644 --- a/lib/drmtest.h +++ b/lib/drmtest.h @@ -52,7 +52,11 @@ void gem_set_domain(int fd, uint32_t handle, uint32_t read_domains, uint32_t write_domain); void gem_sync(int fd, uint32_t handle); uint32_t gem_create(int fd, int size); -void *gem_mmap(int fd, uint32_t handle, int size, int prot); + +void *gem_mmap__gtt(int fd, uint32_t handle, int size, int prot); +void *gem_mmap__cpu(int fd, uint32_t handle, int size, int prot); +#define gem_mmap gem_mmap__gtt + uint64_t gem_aperture_size(int fd); uint64_t gem_mappable_aperture_size(void); int gem_madvise(int fd, uint32_t handle, int state); diff --git a/tests/gem_gtt_speed.c b/tests/gem_gtt_speed.c index 69cd1ff..73a3c6d 100644 --- a/tests/gem_gtt_speed.c +++ b/tests/gem_gtt_speed.c @@ -83,6 +83,58 @@ int main(int argc, char **argv) } if (tiling == I915_TILING_NONE) { + gem_set_domain(fd, handle, + I915_GEM_DOMAIN_CPU, + I915_GEM_DOMAIN_CPU); + + { + uint32_t *base = gem_mmap__cpu(fd, handle, size, PROT_READ | PROT_WRITE); + volatile uint32_t *ptr = base; + int x = 0; + + for (i = 0; i < size/sizeof(*ptr); i++) + x += ptr[i]; + + /* force overtly clever gcc to actually compute x */ + ptr[0] = x; + + munmap(base, size); + + /* mmap read */ + gettimeofday(&start, NULL); + for (loop = 0; loop < 1000; loop++) { + base = gem_mmap__cpu(fd, handle, size, PROT_READ | PROT_WRITE); + ptr = base; + x = 0; + + for (i = 0; i < size/sizeof(*ptr); i++) + x += ptr[i]; + + /* force overtly clever gcc to actually compute x */ + ptr[0] = x; + + munmap(base, size); + } + gettimeofday(&end, NULL); + printf("Time to read %dk through a CPU map: %7.3fµs\n", + size/1024, elapsed(&start, &end, loop)); + + /* mmap write */ + gettimeofday(&start, NULL); + for (loop = 0; loop < 1000; loop++) { + base = gem_mmap__cpu(fd, handle, size, PROT_READ | PROT_WRITE); + ptr = base; + + for (i = 0; i < size/sizeof(*ptr); i++) + ptr[i] = i; + + munmap(base, size); + } + gettimeofday(&end, NULL); + printf("Time to write %dk through a CPU map: %7.3fµs\n", + size/1024, elapsed(&start, &end, loop)); + } + /* CPU pwrite */ gettimeofday(&start, NULL); for (loop = 0; loop < 1000; loop++) @@ -102,7 +154,8 @@ int main(int argc, char **argv) /* prefault into gtt */ { - uint32_t *ptr = gem_mmap(fd, handle, size, PROT_READ | PROT_WRITE); + uint32_t *base = gem_mmap(fd, handle, size, PROT_READ | PROT_WRITE); + volatile uint32_t *ptr = base; int x = 0; for (i = 0; i < size/sizeof(*ptr); i++) @@ -111,12 +164,13 @@ int main(int argc, char **argv) /* force overtly clever gcc to actually compute x */ ptr[0] = x; - munmap(ptr, size); + munmap(base, size); } /* mmap read */ gettimeofday(&start, NULL); for (loop = 0; loop < 1000; loop++) { - uint32_t *ptr = gem_mmap(fd, handle, size, PROT_READ | PROT_WRITE); + uint32_t *base = gem_mmap(fd, handle, size, PROT_READ | PROT_WRITE); + volatile uint32_t *ptr = base; int x = 0; for (i = 0; i < size/sizeof(*ptr); i++) @@ -125,7 +179,7 @@ int main(int argc, char **argv) /* force overtly clever gcc to actually compute x */ ptr[0] = x; - munmap(ptr, size); + munmap(base, size); } gettimeofday(&end, NULL); printf("Time to read %dk through a GTT map: %7.3fµs\n", @@ -134,12 +188,13 @@ int main(int argc, char **argv) /* mmap write */ gettimeofday(&start, NULL); for (loop = 0; loop < 1000; loop++) { - uint32_t *ptr = gem_mmap(fd, handle, size, PROT_READ | PROT_WRITE); + uint32_t *base = gem_mmap(fd, handle, size, PROT_READ | PROT_WRITE); + volatile uint32_t *ptr = base; for (i = 0; i < size/sizeof(*ptr); i++) ptr[i] = i; - munmap(ptr, size); + munmap(base, size); } gettimeofday(&end, NULL); printf("Time to write %dk through a GTT map: %7.3fµs\n", @@ -148,7 +203,8 @@ int main(int argc, char **argv) /* mmap read */ gettimeofday(&start, NULL); for (loop = 0; loop < 1000; loop++) { - uint32_t *ptr = gem_mmap(fd, handle, size, PROT_READ | PROT_WRITE); + uint32_t *base = gem_mmap(fd, handle, size, PROT_READ | PROT_WRITE); + volatile uint32_t *ptr = base; int x = 0; for (i = 0; i < size/sizeof(*ptr); i++) @@ -157,7 +213,7 @@ int main(int argc, char **argv) /* force overtly clever gcc to actually compute x */ ptr[0] = x; - munmap(ptr, size); + munmap(base, size); } gettimeofday(&end, NULL); printf("Time to read %dk (again) through a GTT map: %7.3fµs\n", diff --git a/tests/gem_mmap.c b/tests/gem_mmap.c index 6ba6e3b..f9369f4 100644 --- a/tests/gem_mmap.c +++ b/tests/gem_mmap.c @@ -45,7 +45,7 @@ int main(int argc, char **argv) { int fd; - struct drm_i915_gem_mmap gem_mmap; + struct drm_i915_gem_mmap arg; uint8_t expected[OBJECT_SIZE]; uint8_t buf[OBJECT_SIZE]; uint8_t *addr; @@ -54,23 +54,23 @@ int main(int argc, char **argv) fd = drm_open_any(); - memset(&gem_mmap, 0, sizeof(gem_mmap)); - gem_mmap.handle = 0x10101010; - gem_mmap.offset = 0; - gem_mmap.size = 4096; + memset(&arg, 0, sizeof(arg)); + arg.handle = 0x10101010; + arg.offset = 0; + arg.size = 4096; printf("Testing mmaping of bad object.\n"); - ret = ioctl(fd, DRM_IOCTL_I915_GEM_MMAP, &gem_mmap); + ret = ioctl(fd, DRM_IOCTL_I915_GEM_MMAP, &arg); assert(ret == -1 && errno == ENOENT); handle = gem_create(fd, OBJECT_SIZE); printf("Testing mmaping of newly created object.\n"); - gem_mmap.handle = handle; - gem_mmap.offset = 0; - gem_mmap.size = OBJECT_SIZE; - ret = ioctl(fd, DRM_IOCTL_I915_GEM_MMAP, &gem_mmap); + arg.handle = handle; + arg.offset = 0; + arg.size = OBJECT_SIZE; + ret = ioctl(fd, DRM_IOCTL_I915_GEM_MMAP, &arg); assert(ret == 0); - addr = (uint8_t *)(uintptr_t)gem_mmap.addr_ptr; + addr = (uint8_t *)(uintptr_t)arg.addr_ptr; printf("Testing contents of newly created object.\n"); memset(expected, 0, sizeof(expected)); -- 1.7.10 _______________________________________________ Intel-gfx mailing list [email protected] http://lists.freedesktop.org/mailman/listinfo/intel-gfx
