From: Marek Olšák <marek.olsak at amd.com>
Sadly, the 3 games I tested (DeusEx:MD, DiRT Rally, DOTA 2) are unaffected
by the overallocation, because I guess their buffers don't fall into
the small range below a power-of-two size.
---
src/gallium/winsys/amdgpu/drm/amdgpu_bo.c | 24 +++++++++++++++++++++++
1 file changed, 24 insertions(+)
diff --git a/src/gallium/winsys/amdgpu/drm/amdgpu_bo.c b/src/gallium/winsys/amdgpu/drm/amdgpu_bo.c
index 3fc1da8b0b8..5139e765b72 100644
--- a/src/gallium/winsys/amdgpu/drm/amdgpu_bo.c
+++ b/src/gallium/winsys/amdgpu/drm/amdgpu_bo.c
@@ -403,20 +403,44 @@ static struct amdgpu_winsys_bo *amdgpu_create_bo(struct amdgpu_winsys *ws,
amdgpu_bo_handle buf_handle;
uint64_t va = 0;
struct amdgpu_winsys_bo *bo;
amdgpu_va_handle va_handle;
unsigned va_gap_size;
int r;
/* VRAM or GTT must be specified, but not both at the same time. */
assert(util_bitcount(initial_domain & RADEON_DOMAIN_VRAM_GTT) == 1);
+ /* Gfx9: Overallocate the size to the next power of two for faster address
+ * translation if we don't waste too much memory.
+ */
+ if (ws->info.chip_class >= GFX9) {
+ uint64_t next_pot_size = util_next_power_of_two64(size);
+
+ /* For slightly lower than 4 GB allocations, at most 32 MB are wasted.
+ * For slightly lower than 256 MB allocations, at most 2 MB are wasted.
+ * For slightly lower than 64 MB allocations, at most 512 KB are wasted.
+ *
+ * Waste at most 0.79% (1/127) of the size if we decide to overallocate.
+ */
+ uint64_t max_overalloc = next_pot_size >> 7;
+
+ /* If the next power-of-two size is <= the page size, waste up to
+ * 6.25% (1/16) of the size if we decide to overallocate.
+ */
+ if (next_pot_size <= ws->info.pte_fragment_size)
+ max_overalloc = next_pot_size >> 4;
+
+ if (size + max_overalloc >= next_pot_size)
+ size = next_pot_size;
+ }
+
bo = CALLOC_STRUCT(amdgpu_winsys_bo);
if (!bo) {
return NULL;
}
if (heap >= 0) {
pb_cache_init_entry(&ws->bo_cache, &bo->u.real.cache_entry, &bo->base,
heap);
}
request.alloc_size = size;
--
2.17.1