d3d12: Track up to 16 contexts worth of batch references locally in bos

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21909>

src/gallium/drivers/d3d12/d3d12_batch.cpp

@@ -72,6 +72,7 @@ d3d12_init_batch(struct d3d12_context *ctx, struct d3d12_batch *batch)
       return false;
 
    util_dynarray_init(&batch->zombie_samplers, NULL);
+   util_dynarray_init(&batch->local_bos, NULL);
 
    if (FAILED(screen->dev->CreateCommandAllocator(D3D12_COMMAND_LIST_TYPE_DIRECT,
                                                   IID_PPV_ARGS(&batch->cmdalloc))))
@@ -95,8 +96,13 @@ d3d12_init_batch(struct d3d12_context *ctx, struct d3d12_batch *batch)
    return true;
 }
 
+static inline void
+delete_bo(d3d12_bo *bo)
+{
+   d3d12_bo_unreference(bo);
+}
 static void
-delete_bo(hash_entry *entry)
+delete_bo_entry(hash_entry *entry)
 {
    struct d3d12_bo *bo = (struct d3d12_bo *)entry->key;
    d3d12_bo_unreference(bo);
@@ -143,12 +149,19 @@ d3d12_reset_batch(struct d3d12_context *ctx, struct d3d12_batch *batch, uint64_t
       d3d12_fence_reference(&batch->fence, NULL);
    }
 
-   _mesa_hash_table_clear(batch->bos, delete_bo);
+   _mesa_hash_table_clear(batch->bos, delete_bo_entry);
    _mesa_hash_table_clear(batch->sampler_tables, delete_sampler_view_table);
    _mesa_set_clear(batch->sampler_views, delete_sampler_view);
    _mesa_set_clear(batch->surfaces, delete_surface);
    _mesa_set_clear(batch->objects, delete_object);
 
+   
+   util_dynarray_foreach(&batch->local_bos, d3d12_bo*, bo) {
+      (*bo)->local_reference_mask[batch->ctx_id] &= ~(1 << batch->ctx_index);
+      delete_bo(*bo);
+   }
+   util_dynarray_clear(&batch->local_bos);
+
    util_dynarray_foreach(&batch->zombie_samplers, d3d12_descriptor_handle, handle)
       d3d12_descriptor_handle_free(handle);
    util_dynarray_clear(&batch->zombie_samplers);
@@ -178,6 +191,7 @@ d3d12_destroy_batch(struct d3d12_context *ctx, struct d3d12_batch *batch)
    _mesa_set_destroy(batch->surfaces, NULL);
    _mesa_set_destroy(batch->objects, NULL);
    util_dynarray_fini(&batch->zombie_samplers);
+   util_dynarray_fini(&batch->local_bos);
 }
 
 void
@@ -260,21 +274,60 @@ d3d12_end_batch(struct d3d12_context *ctx, struct d3d12_batch *batch)
    mtx_unlock(&screen->submit_mutex);
 }
 
-enum batch_bo_reference_state
+
+inline uint8_t*
+d3d12_batch_get_reference(struct d3d12_batch *batch,
+                          struct d3d12_bo *bo)
 {
-   batch_bo_reference_read = (1 << 0),
-   batch_bo_reference_written = (1 << 1),
-};
+   if (batch->ctx_id != D3D12_CONTEXT_NO_ID) {
+      if ((bo->local_reference_mask[batch->ctx_id] & (1 << batch->ctx_index)) != 0) {
+         return &bo->local_reference_state[(batch->ctx_id * 16) + batch->ctx_index];
+      }
+      else
+         return NULL;
+   }
+   else {
+      hash_entry* entry = _mesa_hash_table_search(batch->bos, bo);
+      if (entry == NULL)
+         return NULL;
+      else
+         return (uint8_t*)&entry->data;
+   }
+}
+
+inline uint8_t*
+d3d12_batch_acquire_reference(struct d3d12_batch *batch,
+                          struct d3d12_bo *bo)
+{
+   if (batch->ctx_id != D3D12_CONTEXT_NO_ID) {
+      if ((bo->local_reference_mask[batch->ctx_id] & (1 << batch->ctx_index)) == 0) {
+         d3d12_bo_reference(bo);
+         util_dynarray_append(&batch->local_bos, d3d12_bo*, bo);
+         bo->local_reference_mask[batch->ctx_id] |= (1 << batch->ctx_index);
+         bo->local_reference_state[(batch->ctx_id * 16) + batch->ctx_index] = batch_bo_reference_none;
+      }
+      return &bo->local_reference_state[(batch->ctx_id * 16) + batch->ctx_index];
+   }
+   else {
+      hash_entry* entry = _mesa_hash_table_search(batch->bos, bo);
+      if (entry == NULL) {
+         d3d12_bo_reference(bo);
+         entry = _mesa_hash_table_insert(batch->bos, bo, NULL);
+      }
+
+      return (uint8_t*)&entry->data;
+   }
+}
 
 bool
 d3d12_batch_has_references(struct d3d12_batch *batch,
                            struct d3d12_bo *bo,
                            bool want_to_write)
 {
-   hash_entry *entry = _mesa_hash_table_search(batch->bos, bo);
-   if (entry == NULL)
+   uint8_t*state = d3d12_batch_get_reference(batch, bo);
+   if (state == NULL)
       return false;
-   bool resource_was_written = ((batch_bo_reference_state)(size_t)entry->data & batch_bo_reference_written) != 0;
+   bool resource_was_written = ((batch_bo_reference_state)(size_t)*state & batch_bo_reference_written) != 0;
    return want_to_write || resource_was_written;
 }
 
@@ -283,14 +336,11 @@ d3d12_batch_reference_resource(struct d3d12_batch *batch,
                                struct d3d12_resource *res,
                                bool write)
 {
-   hash_entry *entry = _mesa_hash_table_search(batch->bos, res->bo);
-   if (entry == NULL) {
-      d3d12_bo_reference(res->bo);
-      entry = _mesa_hash_table_insert(batch->bos, res->bo, NULL);
-   }
-   size_t new_data = write ? batch_bo_reference_written : batch_bo_reference_read;
-   size_t old_data = (size_t)entry->data;
-   entry->data = (void*)(old_data | new_data);
+   uint8_t*state = d3d12_batch_acquire_reference(batch, res->bo);
+
+   uint8_t new_data = write ? batch_bo_reference_written : batch_bo_reference_read;
+   uint8_t old_data = (uint8_t)*state;
+   *state = (old_data | new_data);
 }
 
 void

src/gallium/drivers/d3d12/d3d12_batch.h

@@ -46,6 +46,7 @@ struct d3d12_batch {
    struct d3d12_fence *fence;
 
    struct hash_table *bos;
+   struct util_dynarray local_bos;
    struct hash_table *sampler_tables;
    struct set *sampler_views;
    struct set *surfaces;
@@ -60,6 +61,7 @@ struct d3d12_batch {
    bool pending_memory_barrier;
 
    uint64_t submit_id;
+   uint32_t ctx_id, ctx_index;
 };
 
 bool

src/gallium/drivers/d3d12/d3d12_bufmgr.h

@@ -41,6 +41,12 @@ enum d3d12_residency_status {
    d3d12_permanently_resident,
 };
 
+enum batch_bo_reference_state {
+   batch_bo_reference_none = 0,
+   batch_bo_reference_read = (1 << 0),
+   batch_bo_reference_written = (1 << 1),
+};
+
 struct d3d12_bo {
    struct pipe_reference reference;
    struct d3d12_screen *screen;
@@ -62,7 +68,10 @@ struct d3d12_bo {
    uint16_t local_needs_resolve_state;
 
    unsigned local_context_state_mask;
+   uint8_t local_reference_mask[16];
+
    d3d12_context_state_table_entry local_context_states[16];
+   uint8_t local_reference_state[128];
 };
 
 struct d3d12_buffer {

src/gallium/drivers/d3d12/d3d12_context.cpp

@@ -2610,6 +2610,11 @@ d3d12_context_create(struct pipe_screen *pscreen, void *priv, unsigned flags)
       ctx->id = D3D12_CONTEXT_NO_ID;
    mtx_unlock(&screen->submit_mutex);
 
+   for (unsigned i = 0; i < ARRAY_SIZE(ctx->batches); ++i) {
+      ctx->batches[i].ctx_id = ctx->id;
+      ctx->batches[i].ctx_index = i;
+   }
+
    if (flags & PIPE_CONTEXT_PREFER_THREADED)
       return threaded_context_create(&ctx->base,
          &screen->transfer_pool,

src/gallium/drivers/d3d12/d3d12_residency.cpp

@@ -126,22 +126,9 @@ get_eviction_grace_period(struct d3d12_memory_info *mem_info)
    return INT64_MAX;
 }
 
-void
-d3d12_process_batch_residency(struct d3d12_screen *screen, struct d3d12_batch *batch)
+static void 
+gather_base_bos(struct d3d12_screen *screen, set *base_bo_set, struct d3d12_bo *bo, uint64_t &size_to_make_resident, uint64_t pending_fence_value, int64_t current_time)
 {
-   d3d12_memory_info mem_info;
-   screen->get_memory_info(screen, &mem_info);
-
-   uint64_t completed_fence_value = screen->fence->GetCompletedValue();
-   uint64_t pending_fence_value = screen->fence_value + 1;
-   int64_t current_time = os_time_get();
-   int64_t grace_period = get_eviction_grace_period(&mem_info);
-
-   /* Gather base bos for the batch */
-   uint64_t size_to_make_resident = 0;
-   set *base_bo_set = _mesa_pointer_set_create(nullptr);
-   hash_table_foreach(batch->bos, entry) {
-      struct d3d12_bo *bo = (struct d3d12_bo *)entry->key;
    uint64_t offset;
    struct d3d12_bo *base_bo = d3d12_bo_get_base(bo, &offset);
 
@@ -162,7 +149,27 @@ d3d12_process_batch_residency(struct d3d12_screen *screen, struct d3d12_batch *b
 
    base_bo->last_used_fence = pending_fence_value;
    base_bo->last_used_timestamp = current_time;
-   }
+}
+
+void
+d3d12_process_batch_residency(struct d3d12_screen *screen, struct d3d12_batch *batch)
+{
+   d3d12_memory_info mem_info;
+   screen->get_memory_info(screen, &mem_info);
+
+   uint64_t completed_fence_value = screen->fence->GetCompletedValue();
+   uint64_t pending_fence_value = screen->fence_value + 1;
+   int64_t current_time = os_time_get();
+   int64_t grace_period = get_eviction_grace_period(&mem_info);
+
+   /* Gather base bos for the batch */
+   uint64_t size_to_make_resident = 0;
+   set *base_bo_set = _mesa_pointer_set_create(nullptr);
+
+   util_dynarray_foreach(&batch->local_bos, d3d12_bo*, bo)
+      gather_base_bos(screen, base_bo_set, *bo, size_to_make_resident, pending_fence_value, current_time);
+   hash_table_foreach(batch->bos, entry) 
+      gather_base_bos(screen, base_bo_set, (struct d3d12_bo *)entry->key, size_to_make_resident, pending_fence_value, current_time);
 
    /* Now that bos referenced by this batch are moved to the end of the LRU, trim it */
    evict_aged_allocations(screen, completed_fence_value, current_time, grace_period);

src/gallium/drivers/d3d12/d3d12_resource_state.cpp

@@ -339,6 +339,22 @@ resolve_global_state(struct d3d12_context *ctx, ID3D12Resource *res, d3d12_resou
    }
 }
 
+static void
+context_state_resolve_submission(struct d3d12_context *ctx, d3d12_bo *bo)
+{
+   d3d12_context_state_table_entry *bo_state = find_or_create_state_entry(ctx, bo);
+   if (!bo_state->batch_end.supports_simultaneous_access) {
+      assert(bo->res && bo->global_state.subresource_states);
+
+      resolve_global_state(ctx, bo->res, &bo_state->batch_begin, &bo->global_state);
+
+      copy_resource_state(&bo_state->batch_begin, &bo_state->batch_end);
+      copy_resource_state(&bo->global_state, &bo_state->batch_end);
+   } else {
+      reset_resource_state(&bo_state->batch_end);
+   }
+}
+
 bool
 d3d12_context_state_resolve_submission(struct d3d12_context *ctx, struct d3d12_batch *batch)
 {
@@ -351,20 +367,12 @@ d3d12_context_state_resolve_submission(struct d3d12_context *ctx, struct d3d12_b
 
    util_dynarray_clear(&ctx->recently_destroyed_bos);
 
-   hash_table_foreach(batch->bos, bo_entry) {
-      d3d12_bo *bo = (d3d12_bo *)bo_entry->key;
-      d3d12_context_state_table_entry *bo_state = find_or_create_state_entry(ctx, bo);
-      if (!bo_state->batch_end.supports_simultaneous_access) {
-         assert(bo->res && bo->global_state.subresource_states);
    
-         resolve_global_state(ctx, bo->res, &bo_state->batch_begin, &bo->global_state);
+   util_dynarray_foreach(&batch->local_bos, d3d12_bo*, bo)
+      context_state_resolve_submission(ctx, *bo);
+   hash_table_foreach(batch->bos, bo_entry) 
+      context_state_resolve_submission(ctx, (d3d12_bo *)bo_entry->key);
    
-         copy_resource_state(&bo_state->batch_begin, &bo_state->batch_end);
-         copy_resource_state(&bo->global_state, &bo_state->batch_end);
-      } else {
-         reset_resource_state(&bo_state->batch_end);
-      }
-   }
 
    bool needs_execute_fixup = false;
    if (ctx->barrier_scratch.size) {