L3 Cache Partitioning Coming For The Intel DRM Driver
Francisco Jerez has been tackling L3 cache partitioning for the Intel DRM driver, which will yield some interesting possibilities moving forward.
This L3 cache partitioning code will be useful for supporting compute shaders with shared variables, improved performance of scratch access, atomic counters, and images, and will be possible for other advantages down the road.
Those interested can see this Mesa mailing list post.
In the L3 atomics patch Francisco commented, "Improves performance of the arb_shader_image_load_store-atomicity piglit test by over 25x (which isn't a real benchmark it's just heavy on atomics -- the improvement in a microbenchmark I wrote a while ago seemed to be even greater). The drawback is one needs to be extra-careful not to hang the GPU (in fact the whole system)."
This L3 cache partitioning code will be useful for supporting compute shaders with shared variables, improved performance of scratch access, atomic counters, and images, and will be possible for other advantages down the road.
Those interested can see this Mesa mailing list post.
In the L3 atomics patch Francisco commented, "Improves performance of the arb_shader_image_load_store-atomicity piglit test by over 25x (which isn't a real benchmark it's just heavy on atomics -- the improvement in a microbenchmark I wrote a while ago seemed to be even greater). The drawback is one needs to be extra-careful not to hang the GPU (in fact the whole system)."
Add A Comment