summaryrefslogtreecommitdiff
path: root/src/intel
diff options
context:
space:
mode:
authorJason Ekstrand <jason.ekstrand@intel.com>2017-04-24 01:51:51 -0700
committerJason Ekstrand <jason.ekstrand@intel.com>2017-05-04 19:07:54 -0700
commitf82d3d38b62048246c4df999a1789b5cb60184c6 (patch)
treea2ebc05fe825f8149e87edcbca1ef10f9f927d7f /src/intel
parent8c079b566e59df2f6d0e0deb951078aba862991d (diff)
anv/allocator: Allow state pools to allocate large states
Previously, the maximum size of a state that could be allocated from a state pool was a block. However, this has caused us various issues particularly with shaders which are potentially very large. We've also hit issues with render passes with a large number of attachments when we go to allocate the block of surface state. This effectively removes the restriction on the maximum size of a single state. (There's still a limit of 1MB imposed by a fixed-length bucket array.) For states larger than the block size, we just grab a large block off of the block pool rather than sub-allocating. When we go to allocate some chunk of state and the current bucket does not have state, we try to pull a chunk from some larger bucket and split it up. This should improve memory usage if a client occasionally allocates a large block of state. This commit is inspired by some similar work done by Juan A. Suarez Romero <jasuarez@igalia.com>. Reviewed-by: Juan A. Suarez Romero <jasuarez@igalia.com>
Diffstat (limited to 'src/intel')
-rw-r--r--src/intel/vulkan/anv_allocator.c69
1 files changed, 69 insertions, 0 deletions
diff --git a/src/intel/vulkan/anv_allocator.c b/src/intel/vulkan/anv_allocator.c
index a74b8decd2..3988a1ab2b 100644
--- a/src/intel/vulkan/anv_allocator.c
+++ b/src/intel/vulkan/anv_allocator.c
@@ -663,6 +663,12 @@ anv_fixed_size_state_pool_alloc_new(struct anv_fixed_size_state_pool *pool,
struct anv_block_state block, old, new;
uint32_t offset;
+ /* If our state is large, we don't need any sub-allocation from a block.
+ * Instead, we just grab whole (potentially large) blocks.
+ */
+ if (state_size >= block_size)
+ return anv_block_pool_alloc(block_pool, state_size);
+
restart:
block.u64 = __sync_fetch_and_add(&pool->block.u64, state_size);
@@ -715,6 +721,69 @@ anv_state_pool_alloc_no_vg(struct anv_state_pool *pool,
goto done;
}
+ /* Try to grab a chunk from some larger bucket and split it up */
+ for (unsigned b = bucket + 1; b < ANV_STATE_BUCKETS; b++) {
+ int32_t chunk_offset;
+ if (anv_free_list_pop(&pool->buckets[b].free_list,
+ &pool->block_pool.map, &chunk_offset)) {
+ unsigned chunk_size = anv_state_pool_get_bucket_size(b);
+
+ /* We've found a chunk that's larger than the requested state size.
+ * There are a couple of options as to what we do with it:
+ *
+ * 1) We could fully split the chunk into state.alloc_size sized
+ * pieces. However, this would mean that allocating a 16B
+ * state could potentially split a 2MB chunk into 512K smaller
+ * chunks. This would lead to unnecessary fragmentation.
+ *
+ * 2) The classic "buddy allocator" method would have us split the
+ * chunk in half and return one half. Then we would split the
+ * remaining half in half and return one half, and repeat as
+ * needed until we get down to the size we want. However, if
+ * you are allocating a bunch of the same size state (which is
+ * the common case), this means that every other allocation has
+ * to go up a level and every fourth goes up two levels, etc.
+ * This is not nearly as efficient as it could be if we did a
+ * little more work up-front.
+ *
+ * 3) Split the difference between (1) and (2) by doing a
+ * two-level split. If it's bigger than some fixed block_size,
+ * we split it into block_size sized chunks and return all but
+ * one of them. Then we split what remains into
+ * state.alloc_size sized chunks and return all but one.
+ *
+ * We choose option (3).
+ */
+ if (chunk_size > pool->block_size &&
+ state.alloc_size < pool->block_size) {
+ assert(chunk_size % pool->block_size == 0);
+ /* We don't want to split giant chunks into tiny chunks. Instead,
+ * break anything bigger than a block into block-sized chunks and
+ * then break it down into bucket-sized chunks from there. Return
+ * all but the first block of the chunk to the block bucket.
+ */
+ const uint32_t block_bucket =
+ anv_state_pool_get_bucket(pool->block_size);
+ anv_free_list_push(&pool->buckets[block_bucket].free_list,
+ pool->block_pool.map,
+ chunk_offset + pool->block_size,
+ pool->block_size,
+ (chunk_size / pool->block_size) - 1);
+ chunk_size = pool->block_size;
+ }
+
+ assert(chunk_size % state.alloc_size == 0);
+ anv_free_list_push(&pool->buckets[bucket].free_list,
+ pool->block_pool.map,
+ chunk_offset + state.alloc_size,
+ state.alloc_size,
+ (chunk_size / state.alloc_size) - 1);
+
+ state.offset = chunk_offset;
+ goto done;
+ }
+ }
+
state.offset = anv_fixed_size_state_pool_alloc_new(&pool->buckets[bucket],
&pool->block_pool,
state.alloc_size,