anv: add a new NO_LOCAL_MEM allocation flag

We found a perf regression with 9027c5df4c51b ("anv: remove the
LOCAL_MEM allocation bit") which seems to be that we over subscribe
local memory, leading i915 to swap things in/out too much.

This change avoid putting buffers in local memory if they are not
allocated from a DEVICE_LOCAL heap.

Maybe we can revisit this later if i915 is better able to deal with
more buffers in local memory.

v2: Remove implicit_css from anv_bo when not in lmem (Ivan)

Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Fixes: 9027c5df4c51b ("anv: remove the LOCAL_MEM allocation bit")
Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/7188
Reviewed-by: Ivan Briano <ivan.briano@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/18395>

diff --git a/src/intel/vulkan/anv_allocator.c b/src/intel/vulkan/anv_allocator.c
index 80f9603..4f637f4 100644 (file)
--- a/src/intel/vulkan/anv_allocator.c
+++ b/src/intel/vulkan/anv_allocator.c
@@ -1463,15 +1463,19 @@ anv_device_alloc_bo(struct anv_device *device,
       /* This always try to put the object in local memory. Here
        * vram_non_mappable & vram_mappable actually are the same region.
        */
-      regions[nregions++] = device->physical->vram_non_mappable.region;
+      if (alloc_flags & ANV_BO_ALLOC_NO_LOCAL_MEM)
+         regions[nregions++] = device->physical->sys.region;
+      else
+         regions[nregions++] = device->physical->vram_non_mappable.region;
 
       /* If the buffer is mapped on the host, add the system memory region.
        * This ensures that if the buffer cannot live in mappable local memory,
        * it can be spilled to system memory.
        */
       uint32_t flags = 0;
-      if ((alloc_flags & ANV_BO_ALLOC_MAPPED) ||
-          (alloc_flags & ANV_BO_ALLOC_LOCAL_MEM_CPU_VISIBLE)) {
+      if (!(alloc_flags & ANV_BO_ALLOC_NO_LOCAL_MEM) &&
+          ((alloc_flags & ANV_BO_ALLOC_MAPPED) ||
+           (alloc_flags & ANV_BO_ALLOC_LOCAL_MEM_CPU_VISIBLE))) {
          regions[nregions++] = device->physical->sys.region;
          if (device->physical->vram_non_mappable.size > 0)
             flags |= I915_GEM_CREATE_EXT_FLAG_NEEDS_CPU_ACCESS;
@@ -1497,7 +1501,8 @@ anv_device_alloc_bo(struct anv_device *device,
       .is_external = (alloc_flags & ANV_BO_ALLOC_EXTERNAL),
       .has_client_visible_address =
          (alloc_flags & ANV_BO_ALLOC_CLIENT_VISIBLE_ADDRESS) != 0,
-      .has_implicit_ccs = ccs_size > 0 || device->info->verx10 >= 125,
+      .has_implicit_ccs = ccs_size > 0 ||
+                          (device->info->verx10 >= 125 && !(alloc_flags & ANV_BO_ALLOC_NO_LOCAL_MEM)),
    };
 
    if (alloc_flags & ANV_BO_ALLOC_MAPPED) {

diff --git a/src/intel/vulkan/anv_device.c b/src/intel/vulkan/anv_device.c
index db84230..87760eb 100644 (file)
--- a/src/intel/vulkan/anv_device.c
+++ b/src/intel/vulkan/anv_device.c
@@ -3852,6 +3852,9 @@ VkResult anv_AllocateMemory(
        (mem_type->propertyFlags & VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT))
       alloc_flags |= ANV_BO_ALLOC_LOCAL_MEM_CPU_VISIBLE;
 
+   if (!(mem_type->propertyFlags & VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT))
+      alloc_flags |= ANV_BO_ALLOC_NO_LOCAL_MEM;
+
    if (vk_flags & VK_MEMORY_ALLOCATE_DEVICE_ADDRESS_BIT)
       alloc_flags |= ANV_BO_ALLOC_CLIENT_VISIBLE_ADDRESS;
 

diff --git a/src/intel/vulkan/anv_private.h b/src/intel/vulkan/anv_private.h
index 8e6eff5..deb0ce7 100644 (file)
--- a/src/intel/vulkan/anv_private.h
+++ b/src/intel/vulkan/anv_private.h
@@ -1237,6 +1237,9 @@ enum anv_bo_alloc_flags {
 
    /** This buffer is allocated from local memory and should be cpu visible */
    ANV_BO_ALLOC_LOCAL_MEM_CPU_VISIBLE = (1 << 10),
+
+   /** For non device local allocations */
+   ANV_BO_ALLOC_NO_LOCAL_MEM = (1 << 11),
 };
 
 VkResult anv_device_alloc_bo(struct anv_device *device,