anv: Disable snooping for allocator pools again
authorKristian Høgsberg Kristensen <kristian.h.kristensen@intel.com>
Fri, 12 Feb 2016 23:10:22 +0000 (15:10 -0800)
committerKristian Høgsberg Kristensen <kristian.h.kristensen@intel.com>
Fri, 12 Feb 2016 23:11:31 +0000 (15:11 -0800)
The race we were seeing on cherryview was caused by the multi-submit
problem with fences. We can now turn snooping off again an rely on
clflush and we intended.

src/vulkan/anv_allocator.c

index e935cd71df0d26b309d75cce41d3103b61868de2..4be149ea695aa783eb185a30a9d6aa5966d09006 100644 (file)
@@ -441,7 +441,6 @@ anv_block_pool_grow(struct anv_block_pool *pool, struct anv_block_state *state)
       goto fail;
    cleanup->gem_handle = gem_handle;
 
-#if 0
    /* Regular objects are created I915_CACHING_CACHED on LLC platforms and
     * I915_CACHING_NONE on non-LLC platforms. However, userptr objects are
     * always created as I915_CACHING_CACHED, which on non-LLC means
@@ -453,7 +452,6 @@ anv_block_pool_grow(struct anv_block_pool *pool, struct anv_block_state *state)
       anv_gem_set_domain(pool->device, gem_handle,
                          I915_GEM_DOMAIN_GTT, I915_GEM_DOMAIN_GTT);
    }
-#endif
 
    /* Now that we successfull allocated everything, we can write the new
     * values back into pool. */