radv: implement a new descriptor sets allocator

The previous implementation was horribly slow with a larger number
of descriptor sets.

The new approach uses util_vma_heap (like ANV) which is a perfect fit.

This fixes stuttering in Indiana Jones because that games seems to use
a huge number of descriptor sets which can also be freed.

Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/13901
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/37976>

src/amd/vulkan/radv_descriptor_pool.c

@@ -20,8 +20,8 @@ static void
 radv_destroy_descriptor_pool_entries(struct radv_device *device, struct radv_descriptor_pool *pool)
 {
    if (!pool->host_memory_base) {
-      for (uint32_t i = 0; i < pool->entry_count; ++i) {
+      list_for_each_entry_safe (struct radv_descriptor_set, set, &pool->sets, link) {
-         radv_descriptor_set_destroy(device, pool, pool->entries[i].set, false);
+         radv_descriptor_set_destroy(device, pool, set);
       }
    } else {
       list_for_each_entry_safe (struct radv_descriptor_set, set, &pool->sets, link) {
@@ -38,6 +38,8 @@ radv_destroy_descriptor_pool(struct radv_device *device, const VkAllocationCallb
 {
    radv_destroy_descriptor_pool_entries(device, pool);
 
+   if (!pool->host_memory_base && pool->size)
+      util_vma_heap_finish(&pool->bo_heap);
    if (pool->bo)
       radv_bo_destroy(device, &pool->base, pool->bo);
    if (pool->host_bo)
@@ -146,8 +148,6 @@ radv_create_descriptor_pool(struct radv_device *device, const VkDescriptorPoolCr
       size += pCreateInfo->maxSets * sizeof(struct radv_descriptor_set);
       size += sizeof(struct radeon_winsys_bo *) * bo_count;
       size += sizeof(struct radv_descriptor_range) * range_count;
-   } else {
-      size += sizeof(struct radv_descriptor_pool_entry) * pCreateInfo->maxSets;
    }
 
    pool = vk_zalloc2(&device->vk.alloc, pAllocator, size, 8, VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
@@ -162,6 +162,11 @@ radv_create_descriptor_pool(struct radv_device *device, const VkDescriptorPoolCr
       pool->host_memory_base = (uint8_t *)pool + sizeof(struct radv_descriptor_pool);
       pool->host_memory_ptr = pool->host_memory_base;
       pool->host_memory_end = (uint8_t *)pool + size;
+   } else {
+      if (bo_size) {
+         util_vma_heap_init(&pool->bo_heap, RADV_POOL_HEAP_OFFSET, bo_size + RADV_POOL_HEAP_OFFSET);
+         pool->bo_heap.alloc_high = false;
+      }
    }
 
    if (bo_size) {
@@ -231,6 +236,11 @@ radv_ResetDescriptorPool(VkDevice _device, VkDescriptorPool descriptorPool, VkDe
 
    radv_destroy_descriptor_pool_entries(device, pool);
 
+   if (!pool->host_memory_base && pool->size) {
+      util_vma_heap_finish(&pool->bo_heap);
+      util_vma_heap_init(&pool->bo_heap, RADV_POOL_HEAP_OFFSET, pool->size + RADV_POOL_HEAP_OFFSET);
+   }
+
    pool->entry_count = 0;
    pool->current_offset = 0;
    pool->host_memory_ptr = pool->host_memory_base;

src/amd/vulkan/radv_descriptor_pool.h

@@ -12,15 +12,15 @@
 #include <vulkan/vulkan.h>
 
 #include "util/list.h"
+#include "util/vma.h"
+
+/* The vma heap reserves 0 to mean NULL; we have to offset by some amount to ensure we can allocate
+ * the entire BO without hitting zero. The actual amount doesn't matter.
+ */
+#define RADV_POOL_HEAP_OFFSET 32
 
 struct radv_descriptor_set;
 
-struct radv_descriptor_pool_entry {
-   uint32_t offset;
-   uint32_t size;
-   struct radv_descriptor_set *set;
-};
-
 struct radv_descriptor_pool {
    struct vk_object_base base;
    struct radeon_winsys_bo *bo;
@@ -35,10 +35,10 @@ struct radv_descriptor_pool {
 
    struct list_head sets;
 
+   struct util_vma_heap bo_heap;
+
    uint32_t entry_count;
    uint32_t max_entry_count;
-
-   struct radv_descriptor_pool_entry entries[0];
 };
 
 VK_DEFINE_NONDISP_HANDLE_CASTS(radv_descriptor_pool, base, VkDescriptorPool, VK_OBJECT_TYPE_DESCRIPTOR_POOL)

src/amd/vulkan/radv_descriptor_set.c

@@ -377,47 +377,29 @@ radv_alloc_descriptor_pool_entry(struct radv_device *device, struct radv_descrip
                                  struct radv_descriptor_set *set)
 {
    uint64_t current_offset = pool->current_offset;
-   uint32_t entry_index = pool->entry_count;
 
    if (!pool->host_memory_base) {
-      /* Try to allocate linearly first, so that we don't spend time looking for gaps if the app
+      if (set->header.size) {
-       * only allocates & resets via the pool. */
+         uint64_t pool_vma_offset = util_vma_heap_alloc(&pool->bo_heap, set->header.size, 32);
-      if (current_offset + set->header.size <= pool->size) {
+         if (!pool_vma_offset)
-         pool->current_offset += set->header.size;
+            return VK_ERROR_FRAGMENTED_POOL;
-      } else {
-         current_offset = 0;
 
-         for (entry_index = 0; entry_index < pool->entry_count; ++entry_index) {
+         assert(pool_vma_offset >= RADV_POOL_HEAP_OFFSET && pool_vma_offset <= pool->size + RADV_POOL_HEAP_OFFSET);
-            if (pool->entries[entry_index].offset - current_offset >= set->header.size)
+         set->header.offset = pool_vma_offset - RADV_POOL_HEAP_OFFSET;
-               break;
+         current_offset = set->header.offset;
-            current_offset = pool->entries[entry_index].offset + pool->entries[entry_index].size;
-         }
-
-         if (pool->size - current_offset < set->header.size) {
-            vk_free2(&device->vk.alloc, NULL, set);
-            return VK_ERROR_OUT_OF_POOL_MEMORY;
-         }
-
-         memmove(&pool->entries[entry_index + 1], &pool->entries[entry_index],
-                 sizeof(pool->entries[0]) * (pool->entry_count - entry_index));
       }
-
-      pool->entries[entry_index].offset = current_offset;
-      pool->entries[entry_index].size = set->header.size;
-      pool->entries[entry_index].set = set;
    } else {
       if (current_offset + set->header.size > pool->size)
          return VK_ERROR_OUT_OF_POOL_MEMORY;
 
       pool->current_offset += set->header.size;
-
-      list_addtail(&set->link, &pool->sets);
    }
 
    set->header.bo = pool->bo;
    set->header.mapped_ptr = (uint32_t *)(pool->mapped_ptr + current_offset);
    set->header.va = pool->bo ? (radv_buffer_get_va(set->header.bo) + current_offset) : 0;
 
+   list_addtail(&set->link, &pool->sets);
    pool->entry_count++;
    return VK_SUCCESS;
 }
@@ -439,8 +421,8 @@ radv_descriptor_set_create(struct radv_device *device, struct radv_descriptor_po
       unsigned stride = radv_descriptor_type_buffer_count(layout->binding[layout->binding_count - 1].type);
       buffer_count = layout->binding[layout->binding_count - 1].buffer_offset + variable_count * stride;
    }
-   unsigned range_offset = sizeof(struct radv_descriptor_set_header) + sizeof(struct list_head) +
+   unsigned range_offset =
-                           sizeof(struct radeon_winsys_bo *) * buffer_count;
+      offsetof(struct radv_descriptor_set, descriptors) + sizeof(struct radeon_winsys_bo *) * buffer_count;
    const unsigned dynamic_offset_count = layout->dynamic_offset_count;
    unsigned mem_size = range_offset + sizeof(struct radv_descriptor_range) * dynamic_offset_count;
 
@@ -507,21 +489,17 @@ radv_descriptor_set_create(struct radv_device *device, struct radv_descriptor_po
 
 void
 radv_descriptor_set_destroy(struct radv_device *device, struct radv_descriptor_pool *pool,
-                            struct radv_descriptor_set *set, bool free_bo)
+                            struct radv_descriptor_set *set)
 {
    assert(!pool->host_memory_base);
 
+   list_del(&set->link);
    vk_descriptor_set_layout_unref(&device->vk, &set->header.layout->vk);
 
-   if (free_bo) {
+   if (set->header.size)
-      for (int i = 0; i < pool->entry_count; ++i) {
+      util_vma_heap_free(&pool->bo_heap, (uint64_t)set->header.offset + RADV_POOL_HEAP_OFFSET, set->header.size);
-         if (pool->entries[i].set == set) {
+   pool->entry_count--;
-            memmove(&pool->entries[i], &pool->entries[i + 1], sizeof(pool->entries[i]) * (pool->entry_count - i - 1));
+
-            --pool->entry_count;
-            break;
-         }
-      }
-   }
    vk_object_base_finish(&set->header.base);
    vk_free2(&device->vk.alloc, NULL, set);
 }
@@ -578,7 +556,7 @@ radv_FreeDescriptorSets(VkDevice _device, VkDescriptorPool descriptorPool, uint3
       VK_FROM_HANDLE(radv_descriptor_set, set, pDescriptorSets[i]);
 
       if (set && !pool->host_memory_base)
-         radv_descriptor_set_destroy(device, pool, set, true);
+         radv_descriptor_set_destroy(device, pool, set);
    }
    return VK_SUCCESS;
 }

src/amd/vulkan/radv_descriptor_set.h

@@ -90,6 +90,7 @@ struct radv_descriptor_range {
 struct radv_descriptor_set_header {
    struct vk_object_base base;
    struct radv_descriptor_set_layout *layout;
+   uint32_t offset;
    uint32_t size;
    uint32_t buffer_count;
 
@@ -136,6 +137,6 @@ void radv_cmd_update_descriptor_sets(struct radv_device *device, struct radv_cmd
                                      const VkCopyDescriptorSet *pDescriptorCopies);
 
 void radv_descriptor_set_destroy(struct radv_device *device, struct radv_descriptor_pool *pool,
-                                 struct radv_descriptor_set *set, bool free_bo);
+                                 struct radv_descriptor_set *set);
 
 #endif /* RADV_DESCRIPTOR_SET_H */