summaryrefslogtreecommitdiffstats
diff options
context:
space:
mode:
authorSamuel Pitoiset <[email protected]>2018-12-20 15:25:22 +0100
committerSamuel Pitoiset <[email protected]>2018-12-20 17:34:04 +0100
commit5b1ec10e4cf0b5ad3d21172333613f8f090a9a52 (patch)
tree240cb4b0bbfba8247fbe9fd97576af538d08b589
parent9f0bfbed11f5fc4f3b899b1eb90570dbeeef45c0 (diff)
radv: compute optimal VM alignment for imported buffers
This fixes GPU hangs on GFX9 with dEQP-VK.memory.external_memory_host.bind_image_memory_and_render.with_zero_offset.* Copied from RadeonSI. Signed-off-by: Samuel Pitoiset <[email protected]> Reviewed-by: Bas Nieuwenhuizen <[email protected]>
-rw-r--r--src/amd/vulkan/winsys/amdgpu/radv_amdgpu_bo.c31
1 files changed, 30 insertions, 1 deletions
diff --git a/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_bo.c b/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_bo.c
index ec126bfc7cb..a9bd55eac8f 100644
--- a/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_bo.c
+++ b/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_bo.c
@@ -410,6 +410,28 @@ radv_amdgpu_winsys_bo_unmap(struct radeon_winsys_bo *_bo)
amdgpu_bo_cpu_unmap(bo->bo);
}
+static uint64_t
+radv_amdgpu_get_optimal_vm_alignment(struct radv_amdgpu_winsys *ws,
+ uint64_t size, unsigned alignment)
+{
+ uint64_t vm_alignment = alignment;
+
+ /* Increase the VM alignment for faster address translation. */
+ if (size >= ws->info.pte_fragment_size)
+ vm_alignment = MAX2(vm_alignment, ws->info.pte_fragment_size);
+
+ /* Gfx9: Increase the VM alignment to the most significant bit set
+ * in the size for faster address translation.
+ */
+ if (ws->info.chip_class >= GFX9) {
+ unsigned msb = util_last_bit64(size); /* 0 = no bit is set */
+ uint64_t msb_alignment = msb ? 1ull << (msb - 1) : 0;
+
+ vm_alignment = MAX2(vm_alignment, msb_alignment);
+ }
+ return vm_alignment;
+}
+
static struct radeon_winsys_bo *
radv_amdgpu_winsys_bo_from_ptr(struct radeon_winsys *_ws,
void *pointer,
@@ -420,6 +442,7 @@ radv_amdgpu_winsys_bo_from_ptr(struct radeon_winsys *_ws,
struct radv_amdgpu_winsys_bo *bo;
uint64_t va;
amdgpu_va_handle va_handle;
+ uint64_t vm_alignment;
bo = CALLOC_STRUCT(radv_amdgpu_winsys_bo);
if (!bo)
@@ -428,8 +451,14 @@ radv_amdgpu_winsys_bo_from_ptr(struct radeon_winsys *_ws,
if (amdgpu_create_bo_from_user_mem(ws->dev, pointer, size, &buf_handle))
goto error;
+ /* Using the optimal VM alignment also fixes GPU hangs for buffers that
+ * are imported.
+ */
+ vm_alignment = radv_amdgpu_get_optimal_vm_alignment(ws, size,
+ ws->info.gart_page_size);
+
if (amdgpu_va_range_alloc(ws->dev, amdgpu_gpu_va_range_general,
- size, 1 << 12, 0, &va, &va_handle,
+ size, vm_alignment, 0, &va, &va_handle,
AMDGPU_VA_RANGE_HIGH))
goto error_va_alloc;