radv: Hardcode root node id.

Optimizes code a tiny bit, and avoid the hack of encoding the root
node id in the low bits of the BLAS address in an instance node.

This is needed to adjust serialization/deserialization as the
instance address there has to be the base address, so this avoids
some wrapping/unwrapping.

Reviewed-by: Konstantin Seurer <konstantin.seurer@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/18692>

diff --git a/src/amd/vulkan/bvh/bvh.h b/src/amd/vulkan/bvh/bvh.h
index dec87c1..f7d2121 100644 (file)
--- a/src/amd/vulkan/bvh/bvh.h
+++ b/src/amd/vulkan/bvh/bvh.h
@@ -53,8 +53,8 @@ struct radv_accel_struct_geometry_info {
 };
 
 struct radv_accel_struct_header {
-   uint32_t root_node_offset;
    uint32_t reserved;
+   uint32_t reserved2;
    float aabb[2][3];
 
    /* Everything after this gets updated/copied from the CPU. */
@@ -119,4 +119,7 @@ struct radv_bvh_box32_node {
    uint32_t reserved[4];
 };
 
+/* 128 bytes of header & a box32 node */
+#define RADV_BVH_ROOT_NODE (0x10 + radv_bvh_node_internal)
+
 #endif

diff --git a/src/amd/vulkan/bvh/internal.comp b/src/amd/vulkan/bvh/internal.comp
index 9b26530..de4de86 100644 (file)
--- a/src/amd/vulkan/bvh/internal.comp
+++ b/src/amd/vulkan/bvh/internal.comp
@@ -95,7 +95,6 @@ main(void)
 
    if (fill_header) {
       REF(radv_accel_struct_header) header = REF(radv_accel_struct_header)(args.bvh);
-      DEREF(header).root_node_offset = node_id;
 
       DEREF(header).aabb[0][0] = total_bounds.min.x;
       DEREF(header).aabb[0][1] = total_bounds.min.y;

diff --git a/src/amd/vulkan/bvh/leaf.comp b/src/amd/vulkan/bvh/leaf.comp
index 4e98691..a0e86f6 100644 (file)
--- a/src/amd/vulkan/bvh/leaf.comp
+++ b/src/amd/vulkan/bvh/leaf.comp
@@ -211,8 +211,7 @@ build_instance(inout AABB bounds, VOID_REF src_ptr, VOID_REF dst_ptr, uint32_t g
 
    radv_accel_struct_header instance_header =
       DEREF(REF(radv_accel_struct_header)(instance.accelerationStructureReference));
-   DEREF(node).base_ptr =
-      instance.accelerationStructureReference | instance_header.root_node_offset;
+   DEREF(node).base_ptr = instance.accelerationStructureReference;
 
    for (uint32_t comp = 0; comp < 3; ++comp) {
       bounds.min[comp] = transform[3][comp];

diff --git a/src/amd/vulkan/radv_acceleration_structure.c b/src/amd/vulkan/radv_acceleration_structure.c
index d6f04ca..71b0cbc 100644 (file)
--- a/src/amd/vulkan/radv_acceleration_structure.c
+++ b/src/amd/vulkan/radv_acceleration_structure.c
@@ -843,9 +843,15 @@ radv_CmdBuildAccelerationStructuresKHR(
             (src_scratch_offset == buffer_1_offset) ? buffer_2_offset : buffer_1_offset;
 
          uint32_t dst_node_offset = bvh_states[i].node_offset;
-         if (final_iter)
+         if (final_iter) {
             dst_node_offset = ALIGN(sizeof(struct radv_accel_struct_header), 64);
 
+            /* Make sure we build the BVH so the hardcoded root node is valid. */
+            STATIC_ASSERT(RADV_BVH_ROOT_NODE ==
+                          DIV_ROUND_UP(sizeof(struct radv_accel_struct_header), 64) * 8 +
+                             radv_bvh_node_internal);
+         }
+
          const struct internal_args consts = {
             .bvh = accel_struct->va,
             .src_ids = pInfos[i].scratchData.deviceAddress + src_scratch_offset,

diff --git a/src/amd/vulkan/radv_nir_lower_ray_queries.c b/src/amd/vulkan/radv_nir_lower_ray_queries.c
index b25f0d8..7c78726 100644 (file)
--- a/src/amd/vulkan/radv_nir_lower_ray_queries.c
+++ b/src/amd/vulkan/radv_nir_lower_ray_queries.c
@@ -366,12 +366,9 @@ lower_rq_initialize(nir_builder *b, nir_ssa_def *index, nir_intrinsic_instr *ins
    {
       rq_store_var(b, index, vars->trav.bvh_base, build_addr_to_node(b, accel_struct), 1);
 
-      nir_ssa_def *bvh_root =
-         nir_build_load_global(b, 1, 32, accel_struct, .access = ACCESS_NON_WRITEABLE,
-                               .align_mul = 64, .align_offset = 0);
-
       rq_store_var(b, index, vars->trav.stack, nir_imm_int(b, 1), 0x1);
-      rq_store_array(b, index, vars->stack, nir_imm_int(b, 0), bvh_root, 0x1);
+      rq_store_array(b, index, vars->stack, nir_imm_int(b, 0), nir_imm_int(b, RADV_BVH_ROOT_NODE),
+                     0x1);
 
       rq_store_var(b, index, vars->trav.top_stack, nir_imm_int(b, 0), 1);
 
@@ -749,7 +746,7 @@ lower_rq_proceed(nir_builder *b, nir_ssa_def *index, struct ray_query_vars *vars
                                1);
 
                   rq_store_array(b, index, vars->stack, rq_load_var(b, index, vars->trav.stack),
-                                 nir_iand_imm(b, nir_channel(b, instance_data, 0), 63), 0x1);
+                                 nir_imm_int(b, RADV_BVH_ROOT_NODE), 0x1);
                   rq_store_var(b, index, vars->trav.stack,
                                nir_iadd_imm(b, rq_load_var(b, index, vars->trav.stack), 1), 1);
 

diff --git a/src/amd/vulkan/radv_pipeline_rt.c b/src/amd/vulkan/radv_pipeline_rt.c
index 0f9a8db..374ce4a 100644 (file)
--- a/src/amd/vulkan/radv_pipeline_rt.c
+++ b/src/amd/vulkan/radv_pipeline_rt.c
@@ -1422,9 +1422,6 @@ build_traversal_shader(struct radv_device *device,
    {
       nir_store_var(&b, trav_vars.bvh_base, build_addr_to_node(&b, accel_struct), 1);
 
-      nir_ssa_def *bvh_root = nir_build_load_global(
-         &b, 1, 32, accel_struct, .access = ACCESS_NON_WRITEABLE, .align_mul = 64);
-
       nir_ssa_def *desc = create_bvh_descriptor(&b);
       nir_ssa_def *vec3ones = nir_channels(&b, nir_imm_vec4(&b, 1.0, 1.0, 1.0, 1.0), 0x7);
 
@@ -1438,7 +1435,7 @@ build_traversal_shader(struct radv_device *device,
       nir_store_var(&b, trav_vars.lds_stack_base, stack_base, 1);
 
       nir_store_var(&b, trav_vars.top_stack, nir_imm_int(&b, 0), 1);
-      nir_store_var(&b, trav_vars.current_node, bvh_root, 0x1);
+      nir_store_var(&b, trav_vars.current_node, nir_imm_int(&b, RADV_BVH_ROOT_NODE), 0x1);
 
       nir_push_loop(&b);
 
@@ -1567,8 +1564,7 @@ build_traversal_shader(struct radv_device *device,
                              build_addr_to_node(
                                 &b, nir_pack_64_2x32(&b, nir_channels(&b, instance_data, 0x3))),
                              1);
-               nir_store_var(&b, trav_vars.current_node,
-                             nir_iand_imm(&b, nir_channel(&b, instance_data, 0), 63), 1);
+               nir_store_var(&b, trav_vars.current_node, nir_imm_int(&b, RADV_BVH_ROOT_NODE), 1);
 
                nir_store_var(
                   &b, trav_vars.origin,

diff --git a/src/amd/vulkan/radv_rra.c b/src/amd/vulkan/radv_rra.c
index 8eb1988..3c5ec15 100644 (file)
--- a/src/amd/vulkan/radv_rra.c
+++ b/src/amd/vulkan/radv_rra.c
@@ -373,7 +373,7 @@ rra_dump_blas_header(struct radv_accel_struct_header *header,
 }
 
 static void
-rra_dump_blas_geometry_infos(struct radv_accel_struct_geometry_info *geometry_infos, 
+rra_dump_blas_geometry_infos(struct radv_accel_struct_geometry_info *geometry_infos,
                              uint32_t geometry_count, FILE *output)
 {
    uint32_t accumulated_primitive_count = 0;
@@ -465,10 +465,9 @@ rra_validate_node(struct hash_table_u64 *accel_struct_vas, uint8_t *data,
             leaf_nodes_size, internal_nodes_size, parent_table_size, is_bottom_level);
       } else if (type == radv_bvh_node_instance) {
          struct radv_bvh_instance_node *src = (struct radv_bvh_instance_node *)(data + offset);
-         uint64_t blas_va = src->base_ptr & (~63UL);
-         if (!_mesa_hash_table_u64_search(accel_struct_vas, blas_va)) {
+         if (!_mesa_hash_table_u64_search(accel_struct_vas, src->base_ptr)) {
             rra_accel_struct_validation_fail(offset, "Invalid instance node pointer 0x%llx",
-                                             (unsigned long long)blas_va);
+                                             (unsigned long long)src->base_ptr);
             result = false;
          }
       }
@@ -503,14 +502,11 @@ rra_transcode_aabb_node(struct rra_aabb_node *dst, const struct radv_bvh_aabb_no
 static void
 rra_transcode_instance_node(struct rra_instance_node *dst, const struct radv_bvh_instance_node *src)
 {
-   /* Mask out root node offset from AS pointer to get the raw VA */
-   uint64_t blas_va = src->base_ptr & (~63UL);
-
    dst->custom_instance_id = src->custom_instance_and_mask & 0xffffff;
    dst->mask = src->custom_instance_and_mask >> 24;
    dst->sbt_offset = src->sbt_offset_and_flags & 0xffffff;
    dst->instance_flags = src->sbt_offset_and_flags >> 24;
-   dst->blas_va = (blas_va + sizeof(struct rra_accel_struct_metadata)) >> 3;
+   dst->blas_va = (src->base_ptr + sizeof(struct rra_accel_struct_metadata)) >> 3;
    dst->instance_id = src->instance_id;
    dst->blas_metadata_size = sizeof(struct rra_accel_struct_metadata);
 
@@ -640,7 +636,7 @@ rra_dump_acceleration_structure(struct rra_copied_accel_struct *copied_struct,
       ((dst_leaf_nodes_size + dst_internal_nodes_size) / 64) * sizeof(uint32_t);
 
    /* convert root node id to offset */
-   uint32_t src_root_offset = (header->root_node_offset & ~7) << 3;
+   uint32_t src_root_offset = (RADV_BVH_ROOT_NODE & ~7) << 3;
 
    if (should_validate)
       if (!rra_validate_node(accel_struct_vas, data,
@@ -656,7 +652,7 @@ rra_dump_acceleration_structure(struct rra_copied_accel_struct *copied_struct,
    }
 
    node_parent_table[rra_parent_table_index_from_offset(RRA_ROOT_NODE_OFFSET, node_parent_table_size)] = 0xffffffff;
-   
+
    uint32_t *leaf_node_ids = malloc(primitive_count * sizeof(uint32_t));
    if (!leaf_node_ids) {
       free(node_parent_table);
@@ -1005,7 +1001,7 @@ rra_copy_acceleration_structures(VkQueue vk_queue, struct rra_accel_struct_copy
 
       dst->copied_structures[i].handle = structure;
       dst->copied_structures[i].data = dst->map_data + dst_offset;
-      
+
       dst_offset += accel_struct->size;
 
       ++(*copied_structure_count);