nvk: Flush descriptor tables and heap maps on submit

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/33959>

src/nouveau/vulkan/nvk_descriptor_table.c

@@ -121,6 +121,7 @@ nvk_descriptor_table_write_locked(struct nvk_descriptor_table *table,
 
    assert(desc_size == table->desc_size);
    memcpy(map, desc_data, table->desc_size);
+   nvk_mem_arena_set_map_dirty(&table->arena);
 }
 
 static void
@@ -130,6 +131,7 @@ nvk_descriptor_table_clear_locked(struct nvk_descriptor_table *table,
    void *map = nvk_descriptor_table_map_locked(table, index);
 
    memset(map, 0, table->desc_size);
+   nvk_mem_arena_set_map_dirty(&table->arena);
 }
 
 static VkResult

src/nouveau/vulkan/nvk_descriptor_table.h

@@ -55,6 +55,13 @@ void nvk_descriptor_table_remove(struct nvk_device *dev,
                                  struct nvk_descriptor_table *table,
                                  uint32_t index);
 
+static void
+nvk_descriptor_table_flush_map(struct nvk_device *dev,
+                               struct nvk_descriptor_table *table)
+{
+   nvk_mem_arena_flush_map(dev, &table->arena);
+}
+
 static inline uint64_t
 nvk_descriptor_table_base_address(struct nvk_descriptor_table *table)
 {

src/nouveau/vulkan/nvk_heap.h

@@ -35,6 +35,12 @@ VkResult nvk_heap_upload(struct nvk_device *dev, struct nvk_heap *heap,
 void nvk_heap_free(struct nvk_device *dev, struct nvk_heap *heap,
                    uint64_t addr, uint64_t size);
 
+static void
+nvk_heap_flush_maps(struct nvk_device *dev, struct nvk_heap *heap)
+{
+   nvk_mem_arena_flush_map(dev, &heap->arena);
+}
+
 static inline uint64_t
 nvk_heap_contiguous_base_address(struct nvk_heap *heap)
 {

src/nouveau/vulkan/nvk_mem_arena.c

@@ -158,8 +158,33 @@ nvk_mem_arena_map(const struct nvk_mem_arena *arena,
    return mem->mem->map + mem_offset_B;
 }
 
+static void
+nvk_mem_arena_flush_map_locked(struct nvk_device *dev,
+                               struct nvk_mem_arena *arena)
+{
+   if (p_atomic_xchg(&arena->map_dirty, 0) == 0)
+      return;
+
+   const uint32_t mem_count = nvk_mem_arena_mem_count(arena);
+
+   for (uint32_t mem_idx = 0; mem_idx < mem_count; mem_idx++) {
+      const struct nvk_arena_mem *mem = &arena->mem[mem_idx];
+      const uint64_t mem_size_B = nvk_mem_arena_mem_size_B(mem_idx);
+      nvkmd_mem_sync_map_to_gpu(mem->mem, 0, mem_size_B);
+   }
+}
+
 void
-nvk_mem_arena_copy_to_gpu(const struct nvk_mem_arena *arena,
+nvk_mem_arena_flush_map(struct nvk_device *dev,
+                        struct nvk_mem_arena *arena)
+{
+   simple_mtx_lock(&arena->mutex);
+   nvk_mem_arena_flush_map_locked(dev, arena);
+   simple_mtx_unlock(&arena->mutex);
+}
+
+void
+nvk_mem_arena_copy_to_gpu(struct nvk_mem_arena *arena,
                           uint64_t dst_addr, const void *src, size_t size_B)
 {
    assert(nvk_mem_arena_is_mapped(arena));
@@ -182,4 +207,6 @@ nvk_mem_arena_copy_to_gpu(const struct nvk_mem_arena *arena,
       src += copy_size_B;
       size_B -= copy_size_B;
    }
+
+   nvk_mem_arena_set_map_dirty(arena);
 }

src/nouveau/vulkan/nvk_mem_arena.h

@@ -72,6 +72,9 @@ struct nvk_mem_arena {
     */
    uint32_t mem_count;
 
+   /* Non-zero if any maps of this arena are dirty */
+   uint32_t map_dirty;
+
    struct nvk_arena_mem mem[NVK_MEM_ARENA_MAX_MEM_COUNT];
 };
 
@@ -188,7 +191,23 @@ nvk_contiguous_mem_arena_map_offset(const struct nvk_mem_arena *arena,
 void *nvk_mem_arena_map(const struct nvk_mem_arena *arena,
                         uint64_t addr, size_t map_range_B);
 
-void nvk_mem_arena_copy_to_gpu(const struct nvk_mem_arena *arena,
+/** Mark the arena map dirty
+ *
+ * This should be called after writing data into the arena via a pointer
+ * returned by nvk_mem_arena_map().  It must be called after the write, not
+ * before, to ensure that the next call to nvk_mem_arena_flush_map() will
+ * flush out the new writes.
+ */
+static inline void
+nvk_mem_arena_set_map_dirty(struct nvk_mem_arena *arena)
+{
+   return p_atomic_set(&arena->map_dirty, 1);
+}
+
+void nvk_mem_arena_flush_map(struct nvk_device *dev,
+                             struct nvk_mem_arena *arena);
+
+void nvk_mem_arena_copy_to_gpu(struct nvk_mem_arena *arena,
                                uint64_t dst_addr,
                                const void *src, size_t size_B);
 

src/nouveau/vulkan/nvk_queue.c

@@ -213,6 +213,11 @@ nvk_queue_submit_exec(struct nvk_queue *queue,
    VkResult result;
 
    if (submit->command_buffer_count > 0) {
+      nvk_descriptor_table_flush_map(dev, &dev->images);
+      nvk_descriptor_table_flush_map(dev, &dev->samplers);
+      nvk_heap_flush_maps(dev, &dev->shader_heap);
+      assert(dev->event_heap.arena.mem_flags & NVKMD_MEM_COHERENT);
+
       result = nvk_queue_state_update(queue, &queue->state);
       if (result != VK_SUCCESS)
          return result;