igt/gem_render_tiled_blits: Speed up by using the GPU to detile
Avoid accessing via the slow GTT to read back and compare the contents
of each bo against expected results. It is much faster, on llc at least,
to detile using the GPU and then copy to system memory for the compare.
Before:
IVB: time sudo ./gem_render_tiled_blits
IGT-Version: 1.6-ge46ff3f (x86_64) (Linux: 3.15.0-rc3+ x86_64)
Using 3072 1MiB buffers
Verifying initialisation...
Cyclic blits, forward...
Cyclic blits, backward...
Random blits...
real 6m26.005s
user 6m19.234s
sys 0m2.414s
PNV: time sudo ./gem_render_tiled_blits
IGT-Version: 1.6-g8556f8a (i686) (Linux: 3.15.0-rc2+ i686)
Using 768 1MiB buffers
Verifying initialisation...
Cyclic blits, forward...
Cyclic blits, backward...
Random blits...
real 1m45.431s
user 1m34.960s
sys 0m4.624s
Using pread:
IVB: time sudo ./gem_render_tiled_blits
IGT-Version: 1.6-ge46ff3f (x86_64) (Linux: 3.15.0-rc3+ x86_64)
Using 3072 1MiB buffers
Verifying initialisation...
Cyclic blits, forward...
Cyclic blits, backward...
Random blits...
real 0m14.717s
user 0m3.699s
sys 0m3.192s
Using snoop:
IVB: time sudo ./gem_render_tiled_blits
IGT-Version: 1.6-ge46ff3f (x86_64) (Linux: 3.15.0-rc3+ x86_64)
Using 3072 1MiB buffers
Using a snoop linear buffer for comparisons
Verifying initialisation...
Cyclic blits, forward...
Cyclic blits, backward...
Random blits...
real 0m13.774s
user 0m3.900s
sys 0m2.089s
PNV: time sudo ./gem_render_tiled_blits
IGT-Version: 1.6-g8556f8a (i686) (Linux: 3.15.0-rc2+ i686)
Using 768 1MiB buffers
Using a snoop linear buffer for comparisons
Verifying initialisation...
Cyclic blits, forward...
Cyclic blits, backward...
Random blits...
real 0m20.831s
user 0m4.384s
sys 0m5.032s
So roughly 10-30x faster depending on platform.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=78244
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>