[RFC PATCH 0/2] Clear system pages with GPU
Nirmoy Das
nirmoy.das at intel.com
Thu Jun 20 13:46:50 UTC 2024
TTM does clear on free for pooled pages and clear on alloc for
non pooled pages using CPU this can have large latency for large
buffer objects. GPU can clear pages much faster but mostly for larger buffers as gpu
clearing requires a gpu job submission which can make latency worse. XE driver on device with flat CCS clears CCS meta data with a clear job submission for all buffers. This series extend that clear job to
also clear system pages using GPU to improve job submission latency. To test the series I created a small test that tries to submit a job after binding various sizes of buffer. Without the series: sudo ~/igt-gpu-tools/build/tests/xe_exec_store --run basic-store-benchmark
IGT-Version: 1.28-g2ed908c0b (x86_64) (Linux: 6.9.0-xe+ x86_64)
Using IGT_SRANDOM=1718889799 for randomisation
Opened device: /dev/dri/card0
Starting subtest: basic-store-benchmark
Starting dynamic subtest: WC
Dynamic subtest WC: SUCCESS (0.000s)
Time taken for size SZ_4K: 4882 us
Time taken for size SZ_2M: 3679 us
Time taken for size SZ_64M: 13367 us
Time taken for size SZ_128M: 21034 us
Time taken for size SZ_256M: 32940 us
Time taken for size SZ_1G: 116261 us
Starting dynamic subtest: WB
Dynamic subtest WB: SUCCESS (0.000s)
Time taken for size SZ_4K: 5417 us
Time taken for size SZ_2M: 5711 us
Time taken for size SZ_64M: 15718 us
Time taken for size SZ_128M: 26170 us
Time taken for size SZ_256M: 50529 us
Time taken for size SZ_1G: 177933 us
Subtest basic-store-benchmark: SUCCESS (0.504s)
With the series: ~/igt-gpu-tools/build/tests/xe_exec_store --run basic-store-benchmark IGT-Version: 1.28-g2ed908c0b (x86_64) (Linux: 6.9.0-xe+ x86_64) Using IGT_SRANDOM=1718889593 for randomisation Opened device: /dev/dri/card0 Starting subtest: basic-store-benchmark Starting dynamic subtest: WC Dynamic subtest WC: SUCCESS (0.000s) Time taken for size SZ_4K: 4479 us Time taken for size SZ_2M: 3291 us Time taken for size SZ_64M: 6595 us Time taken for size SZ_128M: 9069 us Time taken for size SZ_256M: 12681 us Time taken for size SZ_1G: 41806 us Starting dynamic subtest: WB Dynamic subtest WB: SUCCESS (0.000s) Time taken for size SZ_4K: 3317 us Time taken for size SZ_2M: 6458 us Time taken for size SZ_64M: 12802 us Time taken for size SZ_128M: 19579 us Time taken for size SZ_256M: 38768 us Time taken for size SZ_1G: 143250 us Subtest basic-store-benchmark: SUCCESS (0.328s)
Cc: Christian Koenig <christian.koenig at amd.com>
Cc: "Thomas Hellström" <thomas.hellstrom at linux.intel.com>
Cc: Matthew Auld <matthew.auld at intel.com>
Nirmoy Das (2):
drm/ttm/pool: Introduce a way to skip clear on free
drm/xe/lnl: Offload system clear page activity to GPU
drivers/gpu/drm/ttm/ttm_device.c | 42 +++++++++++++++++++++---
drivers/gpu/drm/ttm/ttm_pool.c | 49 +++++++++++++++++++++-------
drivers/gpu/drm/xe/xe_bo.c | 4 +++
drivers/gpu/drm/xe/xe_device.c | 38 ++++++++++++++++-----
drivers/gpu/drm/xe/xe_device_types.h | 2 ++
drivers/gpu/drm/xe/xe_migrate.c | 6 ++--
include/drm/ttm/ttm_device.h | 8 +++++
include/drm/ttm/ttm_pool.h | 11 +++++++
8 files changed, 133 insertions(+), 27 deletions(-)
--
2.42.0
More information about the dri-devel
mailing list