path: root/35734.patch

From 3942a0b1af37ae3820a64afbd0438450e7120199 Mon Sep 17 00:00:00 2001
From: Konstantin Seurer <konstantin.seurer@gmail.com>
Date: Sat, 31 May 2025 09:54:41 +0200
Subject: [PATCH 1/8] radv: Add and use RADV_OFFSET_UNUSED

This deduplicates the logic to figure out what needs to be written.
---
 src/amd/vulkan/radv_acceleration_structure.c | 11 ++++++++---
 1 file changed, 8 insertions(+), 3 deletions(-)

diff --git a/src/amd/vulkan/radv_acceleration_structure.c b/src/amd/vulkan/radv_acceleration_structure.c
index 4a76a4a15e2f4..d4e5daf3f2cf4 100644
--- a/src/amd/vulkan/radv_acceleration_structure.c
+++ b/src/amd/vulkan/radv_acceleration_structure.c
@@ -48,6 +48,8 @@ static const uint32_t leaf_spv[] = {
 #include "bvh/radv_leaf.spv.h"
 };
 
+#define RADV_OFFSET_UNUSED 0xffffffff
+
 struct acceleration_structure_layout {
    uint32_t geometry_info_offset;
    uint32_t primitive_base_indices_offset;
@@ -124,11 +126,15 @@ radv_get_acceleration_structure_layout(struct radv_device *device,
    if (device->rra_trace.accel_structs) {
       accel_struct->geometry_info_offset = offset;
       offset += sizeof(struct radv_accel_struct_geometry_info) * state->build_info->geometryCount;
+   } else {
+      accel_struct->geometry_info_offset = RADV_OFFSET_UNUSED;
    }
 
    if (device->vk.enabled_features.rayTracingPositionFetch && geometry_type == VK_GEOMETRY_TYPE_TRIANGLES_KHR) {
       accel_struct->primitive_base_indices_offset = offset;
       offset += sizeof(uint32_t) * state->build_info->geometryCount;
+   } else {
+      accel_struct->primitive_base_indices_offset = RADV_OFFSET_UNUSED;
    }
 
    /* On GFX12, we need additional space for leaf node offsets since they do not have the same
@@ -670,7 +676,7 @@ radv_init_header(VkCommandBuffer commandBuffer, const struct vk_acceleration_str
    radv_update_memory_cp(cmd_buffer, vk_acceleration_structure_get_va(dst) + base, (const char *)&header + base,
                          sizeof(header) - base);
 
-   if (device->rra_trace.accel_structs) {
+   if (layout.geometry_info_offset != RADV_OFFSET_UNUSED) {
       uint64_t geometry_infos_size = state->build_info->geometryCount * sizeof(struct radv_accel_struct_geometry_info);
 
       struct radv_accel_struct_geometry_info *geometry_infos = malloc(geometry_infos_size);
@@ -691,8 +697,7 @@ radv_init_header(VkCommandBuffer commandBuffer, const struct vk_acceleration_str
       free(geometry_infos);
    }
 
-   VkGeometryTypeKHR geometry_type = vk_get_as_geometry_type(state->build_info);
-   if (device->vk.enabled_features.rayTracingPositionFetch && geometry_type == VK_GEOMETRY_TYPE_TRIANGLES_KHR) {
+   if (layout.primitive_base_indices_offset != RADV_OFFSET_UNUSED) {
       uint32_t base_indices_size = sizeof(uint32_t) * state->build_info->geometryCount;
       uint32_t *base_indices = malloc(base_indices_size);
       if (!base_indices) {
-- 
GitLab


From 5de247e1eaa04da7f17444c246cb31711e22eba4 Mon Sep 17 00:00:00 2001
From: Konstantin Seurer <konstantin.seurer@gmail.com>
Date: Sat, 31 May 2025 10:26:58 +0200
Subject: [PATCH 2/8] radv: Only write leaf node offsets when required

They are only used for serialization and position fetch which makes them
unnecessary most of the times.
---
 src/amd/vulkan/bvh/build_interface.h         |  9 ++++---
 src/amd/vulkan/bvh/encode_gfx12.comp         | 12 +++++----
 src/amd/vulkan/radv_acceleration_structure.c | 27 +++++++++++++++++---
 3 files changed, 35 insertions(+), 13 deletions(-)

diff --git a/src/amd/vulkan/bvh/build_interface.h b/src/amd/vulkan/bvh/build_interface.h
index 77149b1cd15a1..d5993b308a0a4 100644
--- a/src/amd/vulkan/bvh/build_interface.h
+++ b/src/amd/vulkan/bvh/build_interface.h
@@ -18,10 +18,11 @@
 #define VOID_REF  uint64_t
 #endif
 
-#define RADV_BUILD_FLAG_COMPACT         (1u << (VK_BUILD_FLAG_COUNT + 0))
-#define RADV_BUILD_FLAG_BVH8            (1u << (VK_BUILD_FLAG_COUNT + 1))
-#define RADV_BUILD_FLAG_UPDATE_IN_PLACE (1u << (VK_BUILD_FLAG_COUNT + 2))
-#define RADV_BUILD_FLAG_NO_INFS         (1u << (VK_BUILD_FLAG_COUNT + 3))
+#define RADV_BUILD_FLAG_COMPACT                 (1u << (VK_BUILD_FLAG_COUNT + 0))
+#define RADV_BUILD_FLAG_BVH8                    (1u << (VK_BUILD_FLAG_COUNT + 1))
+#define RADV_BUILD_FLAG_UPDATE_IN_PLACE         (1u << (VK_BUILD_FLAG_COUNT + 2))
+#define RADV_BUILD_FLAG_NO_INFS                 (1u << (VK_BUILD_FLAG_COUNT + 3))
+#define RADV_BUILD_FLAG_WRITE_LEAF_NODE_OFFSETS (1u << (VK_BUILD_FLAG_COUNT + 4))
 
 #define RADV_COPY_MODE_COPY        0
 #define RADV_COPY_MODE_SERIALIZE   1
diff --git a/src/amd/vulkan/bvh/encode_gfx12.comp b/src/amd/vulkan/bvh/encode_gfx12.comp
index 805b662e6fd18..eaf7ff081387f 100644
--- a/src/amd/vulkan/bvh/encode_gfx12.comp
+++ b/src/amd/vulkan/bvh/encode_gfx12.comp
@@ -193,11 +193,13 @@ encode_gfx12(uint32_t ir_leaf_node_size, REF(vk_ir_box_node) intermediate_intern
          if (type == vk_ir_node_internal) {
             encoded_type = 5;
          } else {
-            /* Write leaf node offset. */
-            uint32_t leaf_index = offset / ir_leaf_node_size;
-            REF(uint32_t) child_dst_offset = REF(uint32_t)(args.output_base + args.leaf_node_offsets_offset);
-            child_dst_offset = INDEX(uint32_t, child_dst_offset, leaf_index);
-            DEREF(child_dst_offset) = dst_offset;
+            if (VK_BUILD_FLAG(RADV_BUILD_FLAG_WRITE_LEAF_NODE_OFFSETS)) {
+               /* Write leaf node offset. */
+               uint32_t leaf_index = offset / ir_leaf_node_size;
+               REF(uint32_t) child_dst_offset = REF(uint32_t)(args.output_base + args.leaf_node_offsets_offset);
+               child_dst_offset = INDEX(uint32_t, child_dst_offset, leaf_index);
+               DEREF(child_dst_offset) = dst_offset;
+            }
 
             VOID_REF dst_leaf_addr = args.output_base + args.output_bvh_offset + dst_offset;
 
diff --git a/src/amd/vulkan/radv_acceleration_structure.c b/src/amd/vulkan/radv_acceleration_structure.c
index d4e5daf3f2cf4..8a0204bb33c59 100644
--- a/src/amd/vulkan/radv_acceleration_structure.c
+++ b/src/amd/vulkan/radv_acceleration_structure.c
@@ -68,7 +68,8 @@ struct update_scratch_layout {
 };
 
 enum radv_encode_key_bits {
-   RADV_ENCODE_KEY_COMPACT = 1,
+   RADV_ENCODE_KEY_COMPACT = (1 << 0),
+   RADV_ENCODE_KEY_WRITE_LEAF_NODE_OFFSETS = (1 << 1),
 };
 
 static void
@@ -409,9 +410,23 @@ radv_get_build_config(VkDevice _device, struct vk_acceleration_structure_build_s
    struct radv_physical_device *pdev = radv_device_physical(device);
 
    uint32_t encode_key = 0;
-   if (radv_use_bvh8(pdev))
+   if (radv_use_bvh8(pdev)) {
       encode_key |= RADV_ENCODE_KEY_COMPACT;
 
+      /*
+       * Leaf nodes are not written in the order provided by the application when BVH8 encoding is used.
+       * The proper order leaf nodes is used...
+       *    1. When fetching vertex positions for triangles.
+       *    2. When applying/writing BLAS pointers during TLAS deserialization/serialization. The type
+       *       compared to VK_ACCELERATION_STRUCTURE_TYPE_BOTTOM_LEVEL_KHR to handle the
+       *       VK_ACCELERATION_STRUCTURE_TYPE_GENERIC_KHR case when the application queries the
+       *       acceleration structure size.
+       */
+      if ((state->build_info->flags & VK_BUILD_ACCELERATION_STRUCTURE_ALLOW_DATA_ACCESS_KHR) ||
+          state->build_info->type != VK_ACCELERATION_STRUCTURE_TYPE_BOTTOM_LEVEL_KHR)
+         encode_key |= RADV_ENCODE_KEY_WRITE_LEAF_NODE_OFFSETS;
+   }
+
    if (state->build_info->flags & VK_BUILD_ACCELERATION_STRUCTURE_ALLOW_COMPACTION_BIT_KHR)
       encode_key |= RADV_ENCODE_KEY_COMPACT;
 
@@ -482,6 +497,8 @@ radv_build_flags(VkCommandBuffer commandBuffer, uint32_t key)
    /* gfx11 box intersection tests can return garbage with infs and non-standard box sorting */
    if (pdev->info.gfx_level == GFX11)
       flags |= RADV_BUILD_FLAG_NO_INFS;
+   if (key & RADV_ENCODE_KEY_WRITE_LEAF_NODE_OFFSETS)
+      flags |= RADV_BUILD_FLAG_WRITE_LEAF_NODE_OFFSETS;
 
    return flags;
 }
@@ -500,8 +517,10 @@ static VkResult
 radv_encode_bind_pipeline_gfx12(VkCommandBuffer commandBuffer,
                                 const struct vk_acceleration_structure_build_state *state)
 {
-   radv_bvh_build_bind_pipeline(commandBuffer, RADV_META_OBJECT_KEY_BVH_ENCODE, encode_gfx12_spv,
-                                sizeof(encode_gfx12_spv), sizeof(struct encode_gfx12_args), 0);
+   radv_bvh_build_bind_pipeline(
+      commandBuffer, RADV_META_OBJECT_KEY_BVH_ENCODE, encode_gfx12_spv, sizeof(encode_gfx12_spv),
+      sizeof(struct encode_gfx12_args),
+      radv_build_flags(commandBuffer, state->config->encode_key[0]) & RADV_BUILD_FLAG_WRITE_LEAF_NODE_OFFSETS);
 
    return VK_SUCCESS;
 }
-- 
GitLab


From 9a5d0d7357873ad5417b9d89d89f5f79fbe8161a Mon Sep 17 00:00:00 2001
From: Konstantin Seurer <konstantin.seurer@gmail.com>
Date: Thu, 5 Jun 2025 08:30:36 +0200
Subject: [PATCH 3/8] radv/bvh: Fix flush in bit_writer_skip_to

If temp is not cleared, the next flushed dword will contain data from
the previous one.
---
 src/amd/vulkan/bvh/encode.h | 1 +
 1 file changed, 1 insertion(+)

diff --git a/src/amd/vulkan/bvh/encode.h b/src/amd/vulkan/bvh/encode.h
index 25abeb2ac8d91..80e2f669eac44 100644
--- a/src/amd/vulkan/bvh/encode.h
+++ b/src/amd/vulkan/bvh/encode.h
@@ -103,6 +103,7 @@ bit_writer_skip_to(inout bit_writer writer, uint32_t target)
    if (writer.count > 0) {
       REF(uint32_t) dst = REF(uint32_t)(writer.addr + writer.offset);
       DEREF(dst) = writer.temp;
+      writer.temp = 0;
    }
 
    writer.count = target % 32;
-- 
GitLab


From b43a1748064a6831e0bc9c7833cbe71be82ce3ea Mon Sep 17 00:00:00 2001
From: Konstantin Seurer <konstantin.seurer@gmail.com>
Date: Thu, 5 Jun 2025 08:32:36 +0200
Subject: [PATCH 4/8] radv/bvh: Use a fixed indices midpoint on GFX12

This saves a vouple of loads inside the update shader.
---
 src/amd/vulkan/bvh/encode.h          | 12 ++++++++----
 src/amd/vulkan/bvh/update_gfx12.comp | 19 ++++---------------
 2 files changed, 12 insertions(+), 19 deletions(-)

diff --git a/src/amd/vulkan/bvh/encode.h b/src/amd/vulkan/bvh/encode.h
index 80e2f669eac44..625525ae07a52 100644
--- a/src/amd/vulkan/bvh/encode.h
+++ b/src/amd/vulkan/bvh/encode.h
@@ -125,6 +125,8 @@ bit_writer_finish(inout bit_writer writer)
    writer.total_count = 0;
 }
 
+#define RADV_GFX12_UPDATABLE_PRIMITIVE_NODE_INDICES_OFFSET (align(RADV_GFX12_PRIMITIVE_NODE_HEADER_SIZE, 32) / 8 + 9 * 4)
+
 void
 radv_encode_triangle_gfx12(VOID_REF dst, vk_ir_triangle_node src)
 {
@@ -141,8 +143,7 @@ radv_encode_triangle_gfx12(VOID_REF dst, vk_ir_triangle_node src)
    bit_writer_write(child_writer, 0, 1);  /* vertex_type */
    bit_writer_write(child_writer, 28, 5); /* primitive_index_base_bits */
    bit_writer_write(child_writer, 28, 5); /* primitive_index_bits */
-   /* header + 9 floats + geometry_id */
-   bit_writer_write(child_writer, RADV_GFX12_PRIMITIVE_NODE_HEADER_SIZE + 9 * 32 + 28, 10);
+   bit_writer_write(child_writer, RADV_GFX12_UPDATABLE_PRIMITIVE_NODE_INDICES_OFFSET * 8 + 32, 10);
 
    bit_writer_write(child_writer, floatBitsToUint(src.coords[0][0]), 32);
    bit_writer_write(child_writer, floatBitsToUint(src.coords[0][1]), 32);
@@ -154,6 +155,8 @@ radv_encode_triangle_gfx12(VOID_REF dst, vk_ir_triangle_node src)
    bit_writer_write(child_writer, floatBitsToUint(src.coords[2][1]), 32);
    bit_writer_write(child_writer, floatBitsToUint(src.coords[2][2]), 32);
 
+   bit_writer_write(child_writer, 0, 64 - RADV_GFX12_PRIMITIVE_NODE_HEADER_SIZE + 4);
+
    bit_writer_write(child_writer, src.geometry_id_and_flags & 0xfffffff, 28);
    bit_writer_write(child_writer, src.triangle_id, 28);
 
@@ -192,8 +195,7 @@ radv_encode_aabb_gfx12(VOID_REF dst, vk_ir_aabb_node src)
    bit_writer_write(child_writer, 0, 1);  /* vertex_type */
    bit_writer_write(child_writer, 28, 5); /* primitive_index_base_bits */
    bit_writer_write(child_writer, 28, 5); /* primitive_index_bits */
-   /* header + 6 floats + geometry_id */
-   bit_writer_write(child_writer, RADV_GFX12_PRIMITIVE_NODE_HEADER_SIZE + 6 * 32 + 28, 10);
+   bit_writer_write(child_writer, RADV_GFX12_UPDATABLE_PRIMITIVE_NODE_INDICES_OFFSET * 8 + 32, 10);
 
    bit_writer_write(child_writer, floatBitsToUint(src.base.aabb.min.x), 32);
    bit_writer_write(child_writer, floatBitsToUint(src.base.aabb.min.y), 32);
@@ -202,6 +204,8 @@ radv_encode_aabb_gfx12(VOID_REF dst, vk_ir_aabb_node src)
    bit_writer_write(child_writer, floatBitsToUint(src.base.aabb.max.y), 32);
    bit_writer_write(child_writer, floatBitsToUint(src.base.aabb.max.z), 32);
 
+   bit_writer_skip_to(child_writer, RADV_GFX12_UPDATABLE_PRIMITIVE_NODE_INDICES_OFFSET * 8 + 4);
+
    bit_writer_write(child_writer, src.geometry_id_and_flags & 0xfffffff, 28);
    bit_writer_write(child_writer, src.primitive_id, 28);
 
diff --git a/src/amd/vulkan/bvh/update_gfx12.comp b/src/amd/vulkan/bvh/update_gfx12.comp
index 8cbc930ab4352..32da58fca6890 100644
--- a/src/amd/vulkan/bvh/update_gfx12.comp
+++ b/src/amd/vulkan/bvh/update_gfx12.comp
@@ -33,18 +33,6 @@ layout(push_constant) uniform CONSTS
    update_gfx12_args args;
 };
 
-uint32_t
-read_bits(VOID_REF data, uint32_t start, uint32_t count)
-{
-   uint32_t shift = start % 32;
-   uint32_t lower = DEREF(INDEX(uint32_t, data, start / 32)) >> shift;
-   uint32_t upper = 0;
-   if (shift != 0 && shift + count > 32)
-      upper = DEREF(INDEX(uint32_t, data, start / 32 + 1)) << (32 - shift);
-   uint32_t total = lower | upper;
-   return count != 32 ? total & ((1u << count) - 1u) : total;
-}
-
 void
 main()
 {
@@ -109,9 +97,10 @@ main()
 
       if (is_leaf) {
          VOID_REF src_leaf_node = OFFSET(src_bvh, child_offset);
-         uint32_t indices_midpoint = read_bits(src_leaf_node, 42, 10);
-         uint32_t geometry_index = read_bits(src_leaf_node, indices_midpoint - 28, 28);
-         uint32_t primitive_index = read_bits(src_leaf_node, indices_midpoint, 28);
+         uint32_t geometry_index =
+            DEREF(REF(uint32_t)(src_leaf_node + RADV_GFX12_UPDATABLE_PRIMITIVE_NODE_INDICES_OFFSET)) >> 4;
+         uint32_t primitive_index =
+            DEREF(REF(uint32_t)(src_leaf_node + RADV_GFX12_UPDATABLE_PRIMITIVE_NODE_INDICES_OFFSET + 4)) & 0xfffffff;
 
          vk_bvh_geometry_data geom_data = DEREF(INDEX(vk_bvh_geometry_data, args.geom_data, geometry_index));
 
-- 
GitLab


From 86858fdd56bf4f2142dc5b49122002f68f177cf5 Mon Sep 17 00:00:00 2001
From: Konstantin Seurer <konstantin.seurer@gmail.com>
Date: Fri, 6 Jun 2025 08:31:38 +0200
Subject: [PATCH 5/8] radv/bvh: Update leaf nodes before refitting

This should reduce latency between refitting nodes and their parent
nodes.
---
 src/amd/vulkan/bvh/update_gfx12.comp | 70 +++++++++++++++-------------
 1 file changed, 38 insertions(+), 32 deletions(-)

diff --git a/src/amd/vulkan/bvh/update_gfx12.comp b/src/amd/vulkan/bvh/update_gfx12.comp
index 32da58fca6890..4c2890b65dcc9 100644
--- a/src/amd/vulkan/bvh/update_gfx12.comp
+++ b/src/amd/vulkan/bvh/update_gfx12.comp
@@ -81,6 +81,27 @@ main()
       child_offset += SIZEOF(radv_gfx12_box_node) * child_index;
    }
 
+   vk_aabb bounds;
+   bounds.min = vec3(INFINITY);
+   bounds.max = vec3(-INFINITY);
+   if (is_leaf) {
+      VOID_REF src_leaf_node = OFFSET(src_bvh, child_offset);
+      uint32_t geometry_index =
+         DEREF(REF(uint32_t)(src_leaf_node + RADV_GFX12_UPDATABLE_PRIMITIVE_NODE_INDICES_OFFSET)) >> 4;
+      uint32_t primitive_index =
+         DEREF(REF(uint32_t)(src_leaf_node + RADV_GFX12_UPDATABLE_PRIMITIVE_NODE_INDICES_OFFSET + 4)) & 0xfffffff;
+
+      vk_bvh_geometry_data geom_data = DEREF(INDEX(vk_bvh_geometry_data, args.geom_data, geometry_index));
+
+      VOID_REF dst_leaf_node = OFFSET(dst_bvh, child_offset);
+      if (geom_data.geometry_type == VK_GEOMETRY_TYPE_TRIANGLES_KHR) {
+         radv_build_triangle(bounds, dst_leaf_node, geom_data, primitive_index, true);
+      } else {
+         VOID_REF src_ptr = OFFSET(geom_data.data, primitive_index * geom_data.stride);
+         radv_build_aabb(bounds, src_ptr, dst_leaf_node, geometry_index, primitive_index, true);
+      }
+   }
+
    uint32_t child_index = (child_offset - internal_nodes_offset) / SIZEOF(radv_gfx12_box_node);
 
    bool is_ready = is_leaf_or_invalid;
@@ -91,27 +112,7 @@ main()
       if (radv_ballot(cluster, is_ready) != 0xff)
          continue;
 
-      vk_aabb bounds;
-      bounds.min = vec3(INFINITY);
-      bounds.max = vec3(-INFINITY);
-
-      if (is_leaf) {
-         VOID_REF src_leaf_node = OFFSET(src_bvh, child_offset);
-         uint32_t geometry_index =
-            DEREF(REF(uint32_t)(src_leaf_node + RADV_GFX12_UPDATABLE_PRIMITIVE_NODE_INDICES_OFFSET)) >> 4;
-         uint32_t primitive_index =
-            DEREF(REF(uint32_t)(src_leaf_node + RADV_GFX12_UPDATABLE_PRIMITIVE_NODE_INDICES_OFFSET + 4)) & 0xfffffff;
-
-         vk_bvh_geometry_data geom_data = DEREF(INDEX(vk_bvh_geometry_data, args.geom_data, geometry_index));
-
-         VOID_REF dst_leaf_node = OFFSET(dst_bvh, child_offset);
-         if (geom_data.geometry_type == VK_GEOMETRY_TYPE_TRIANGLES_KHR) {
-            radv_build_triangle(bounds, dst_leaf_node, geom_data, primitive_index, true);
-         } else {
-            VOID_REF src_ptr = OFFSET(geom_data.data, primitive_index * geom_data.stride);
-            radv_build_aabb(bounds, src_ptr, dst_leaf_node, geometry_index, primitive_index, true);
-         }
-      } else if (is_valid) {
+      if (!is_leaf_or_invalid) {
          bounds = DEREF(INDEX(vk_aabb, args.bounds, child_index));
       }
 
@@ -123,13 +124,20 @@ main()
       total_bounds.max.y = subgroupClusteredMax(bounds.max.y, 8);
       total_bounds.max.z = subgroupClusteredMax(bounds.max.z, 8);
 
-      if (!is_root_node) {
-         DEREF(INDEX(vk_aabb, args.bounds, node_index - 1)) = total_bounds;
+      if (cluster.invocation_index == 0) {
+         if (is_root_node) {
+            DEREF(args.dst).aabb = total_bounds;
+         } else {
+            DEREF(INDEX(vk_aabb, args.bounds, node_index - 1)) = total_bounds;
+
+            memoryBarrier(gl_ScopeDevice, gl_StorageSemanticsBuffer,
+                          gl_SemanticsAcquireRelease | gl_SemanticsMakeAvailable | gl_SemanticsMakeVisible);
 
-         memoryBarrier(gl_ScopeDevice, gl_StorageSemanticsBuffer,
-                       gl_SemanticsAcquireRelease | gl_SemanticsMakeAvailable | gl_SemanticsMakeVisible);
+            DEREF(INDEX(uint32_t, args.internal_ready_count, node_index - 1)) = 1;
 
-         DEREF(INDEX(uint32_t, args.internal_ready_count, node_index - 1)) = 1;
+            memoryBarrier(gl_ScopeDevice, gl_StorageSemanticsBuffer,
+                          gl_SemanticsAcquireRelease | gl_SemanticsMakeAvailable | gl_SemanticsMakeVisible);
+         }
       }
 
       vec3 origin = total_bounds.min;
@@ -139,6 +147,10 @@ main()
       uvec3 extent_exponents = floatBitsToUint(extent) >> 23;
 
       if (cluster.invocation_index == 0) {
+         if (!VK_BUILD_FLAG(RADV_BUILD_FLAG_UPDATE_IN_PLACE)) {
+            DEREF(dst_node).primitive_base_id = src_node.primitive_base_id;
+            DEREF(dst_node).internal_base_id = src_node.internal_base_id;
+         }
          DEREF(dst_node).origin = origin;
          DEREF(dst_node).child_count_exponents = extent_exponents.x | (extent_exponents.y << 8) |
                                                  (extent_exponents.z << 16) | (valid_child_count_minus_one << 28);
@@ -170,12 +182,6 @@ main()
          DEREF(dst_node).children[cluster.invocation_index] = null_child;
       }
 
-      if (is_root_node)
-         DEREF(args.dst).aabb = total_bounds;
-
-      /* Make changes to internal_ready_count available to the other invocations. */
-      memoryBarrier(gl_ScopeDevice, gl_StorageSemanticsBuffer,
-                    gl_SemanticsAcquireRelease | gl_SemanticsMakeAvailable | gl_SemanticsMakeVisible);
       break;
    }
 }
-- 
GitLab


From fe46f07812c0790754c995504d6fc2d9115c258a Mon Sep 17 00:00:00 2001
From: Konstantin Seurer <konstantin.seurer@gmail.com>
Date: Tue, 10 Jun 2025 14:05:22 +0200
Subject: [PATCH 6/8] radv/bvh: Specialize the update shader for
 geometryCount==1

The geometry data can be loaded from push constants in that case.
---
 src/amd/vulkan/bvh/build_interface.h         |  3 +++
 src/amd/vulkan/bvh/update_gfx12.comp         | 13 ++++++++++---
 src/amd/vulkan/radv_acceleration_structure.c | 17 +++++++++++++----
 3 files changed, 26 insertions(+), 7 deletions(-)

diff --git a/src/amd/vulkan/bvh/build_interface.h b/src/amd/vulkan/bvh/build_interface.h
index d5993b308a0a4..a49476d720392 100644
--- a/src/amd/vulkan/bvh/build_interface.h
+++ b/src/amd/vulkan/bvh/build_interface.h
@@ -23,6 +23,7 @@
 #define RADV_BUILD_FLAG_UPDATE_IN_PLACE         (1u << (VK_BUILD_FLAG_COUNT + 2))
 #define RADV_BUILD_FLAG_NO_INFS                 (1u << (VK_BUILD_FLAG_COUNT + 3))
 #define RADV_BUILD_FLAG_WRITE_LEAF_NODE_OFFSETS (1u << (VK_BUILD_FLAG_COUNT + 4))
+#define RADV_BUILD_FLAG_UPDATE_SINGLE_GEOMETRY  (1u << (VK_BUILD_FLAG_COUNT + 5))
 
 #define RADV_COPY_MODE_COPY        0
 #define RADV_COPY_MODE_SERIALIZE   1
@@ -78,6 +79,8 @@ struct update_gfx12_args {
    REF(vk_aabb) bounds;
    REF(uint32_t) internal_ready_count;
    uint32_t leaf_node_count;
+
+   vk_bvh_geometry_data geom_data0;
 };
 
 #endif /* BUILD_INTERFACE_H */
diff --git a/src/amd/vulkan/bvh/update_gfx12.comp b/src/amd/vulkan/bvh/update_gfx12.comp
index 4c2890b65dcc9..72f8358f115b1 100644
--- a/src/amd/vulkan/bvh/update_gfx12.comp
+++ b/src/amd/vulkan/bvh/update_gfx12.comp
@@ -86,12 +86,19 @@ main()
    bounds.max = vec3(-INFINITY);
    if (is_leaf) {
       VOID_REF src_leaf_node = OFFSET(src_bvh, child_offset);
-      uint32_t geometry_index =
-         DEREF(REF(uint32_t)(src_leaf_node + RADV_GFX12_UPDATABLE_PRIMITIVE_NODE_INDICES_OFFSET)) >> 4;
+
+      uint32_t geometry_index = 0;
+      if (!VK_BUILD_FLAG(RADV_BUILD_FLAG_UPDATE_SINGLE_GEOMETRY))
+         geometry_index = DEREF(REF(uint32_t)(src_leaf_node + RADV_GFX12_UPDATABLE_PRIMITIVE_NODE_INDICES_OFFSET)) >> 4;
+
       uint32_t primitive_index =
          DEREF(REF(uint32_t)(src_leaf_node + RADV_GFX12_UPDATABLE_PRIMITIVE_NODE_INDICES_OFFSET + 4)) & 0xfffffff;
 
-      vk_bvh_geometry_data geom_data = DEREF(INDEX(vk_bvh_geometry_data, args.geom_data, geometry_index));
+      vk_bvh_geometry_data geom_data;
+      if (VK_BUILD_FLAG(RADV_BUILD_FLAG_UPDATE_SINGLE_GEOMETRY))
+         geom_data = args.geom_data0;
+      else
+         geom_data = DEREF(INDEX(vk_bvh_geometry_data, args.geom_data, geometry_index));
 
       VOID_REF dst_leaf_node = OFFSET(dst_bvh, child_offset);
       if (geom_data.geometry_type == VK_GEOMETRY_TYPE_TRIANGLES_KHR) {
diff --git a/src/amd/vulkan/radv_acceleration_structure.c b/src/amd/vulkan/radv_acceleration_structure.c
index 8a0204bb33c59..5309d6388c533 100644
--- a/src/amd/vulkan/radv_acceleration_structure.c
+++ b/src/amd/vulkan/radv_acceleration_structure.c
@@ -437,6 +437,9 @@ radv_get_build_config(VkDevice _device, struct vk_acceleration_structure_build_s
    if (state->build_info->srcAccelerationStructure == state->build_info->dstAccelerationStructure)
       update_key |= RADV_BUILD_FLAG_UPDATE_IN_PLACE;
 
+   if (state->build_info->geometryCount == 1)
+      update_key |= RADV_BUILD_FLAG_UPDATE_SINGLE_GEOMETRY;
+
    state->config.update_key[0] = update_key;
 }
 
@@ -520,7 +523,7 @@ radv_encode_bind_pipeline_gfx12(VkCommandBuffer commandBuffer,
    radv_bvh_build_bind_pipeline(
       commandBuffer, RADV_META_OBJECT_KEY_BVH_ENCODE, encode_gfx12_spv, sizeof(encode_gfx12_spv),
       sizeof(struct encode_gfx12_args),
-      radv_build_flags(commandBuffer, state->config->encode_key[0]) & RADV_BUILD_FLAG_WRITE_LEAF_NODE_OFFSETS);
+      radv_build_flags(commandBuffer, state->config.encode_key[0]) & RADV_BUILD_FLAG_WRITE_LEAF_NODE_OFFSETS);
 
    return VK_SUCCESS;
 }
@@ -753,7 +756,8 @@ radv_init_update_scratch(VkCommandBuffer commandBuffer, const struct vk_accelera
    radv_fill_memory(cmd_buffer, scratch + layout.internal_ready_count_offset,
                     layout.size - layout.internal_ready_count_offset, 0x0, RADV_COPY_FLAGS_DEVICE_LOCAL);
 
-   if (radv_use_bvh8(pdev)) {
+   /* geometryCount == 1 passes the data as push constant. */
+   if (radv_use_bvh8(pdev) && !(state->config.update_key[0] & RADV_BUILD_FLAG_UPDATE_SINGLE_GEOMETRY)) {
       uint32_t data_size = sizeof(struct vk_bvh_geometry_data) * state->build_info->geometryCount;
       struct vk_bvh_geometry_data *data = malloc(data_size);
       if (!data) {
@@ -797,8 +801,7 @@ radv_update_bind_pipeline(VkCommandBuffer commandBuffer, const struct vk_acceler
    if (radv_device_physical(device)->info.cp_sdma_ge_use_system_memory_scope)
       cmd_buffer->state.flush_bits |= RADV_CMD_FLAG_INV_L2;
 
-   bool in_place = state->config.update_key[0] & RADV_BUILD_FLAG_UPDATE_IN_PLACE;
-   uint32_t flags = in_place ? RADV_BUILD_FLAG_UPDATE_IN_PLACE : 0;
+   uint32_t flags = state->config.update_key[0];
 
    if (radv_use_bvh8(pdev)) {
       radv_bvh_build_bind_pipeline(commandBuffer, RADV_META_OBJECT_KEY_BVH_UPDATE, update_gfx12_spv,
@@ -899,6 +902,12 @@ radv_update_as_gfx12(VkCommandBuffer commandBuffer, const struct vk_acceleration
       .leaf_node_count = state->leaf_node_count,
    };
 
+   if (state->config.update_key[0] & RADV_BUILD_FLAG_UPDATE_SINGLE_GEOMETRY) {
+      const VkAccelerationStructureGeometryKHR *geom =
+         state->build_info->pGeometries ? &state->build_info->pGeometries[0] : state->build_info->ppGeometries[0];
+      update_consts.geom_data0 = vk_fill_geometry_data(state->build_info->type, 0, 0, geom, state->build_range_infos);
+   }
+
    radv_bvh_build_set_args(commandBuffer, &update_consts, sizeof(update_consts));
 
    struct radv_dispatch_info dispatch = {
-- 
GitLab


From db8303d72894d567b268e4893f04a0ab3fd28182 Mon Sep 17 00:00:00 2001
From: Konstantin Seurer <konstantin.seurer@gmail.com>
Date: Fri, 2 May 2025 14:52:02 +0200
Subject: [PATCH 7/8] radv/rra/gfx12: Handle compressed primitive nodes

---
 src/amd/vulkan/radv_rra_gfx12.c | 47 ++++++++++++++++++++++++---------
 1 file changed, 34 insertions(+), 13 deletions(-)

diff --git a/src/amd/vulkan/radv_rra_gfx12.c b/src/amd/vulkan/radv_rra_gfx12.c
index 4cad241a38b89..a0114462881e9 100644
--- a/src/amd/vulkan/radv_rra_gfx12.c
+++ b/src/amd/vulkan/radv_rra_gfx12.c
@@ -37,19 +37,27 @@ rra_validate_node_gfx12(struct hash_table_u64 *accel_struct_vas, uint8_t *data,
 }
 
 static uint32_t
-get_geometry_id(const void *node, uint32_t node_type)
+get_geometry_id(const void *node, uint32_t triangle_index)
 {
-   if (node_type == radv_bvh_node_instance)
-      return 0;
+   uint32_t geometry_index_base_bits = BITSET_EXTRACT(node, 20, 4) * 2;
+   uint32_t geometry_index_bits = BITSET_EXTRACT(node, 24, 4) * 2;
 
    uint32_t indices_midpoint = BITSET_EXTRACT(node, 42, 10);
-   return BITSET_EXTRACT(node, indices_midpoint - 28, 28);
+   uint32_t geometry_id_base =
+      BITSET_EXTRACT(node, indices_midpoint - geometry_index_base_bits, geometry_index_base_bits);
+
+   if (triangle_index == 0)
+      return geometry_id_base;
+
+   return (geometry_id_base & ~BITFIELD64_MASK(geometry_index_bits)) |
+          BITSET_EXTRACT(node, indices_midpoint - geometry_index_base_bits - geometry_index_bits * triangle_index,
+                         geometry_index_bits);
 }
 
 void
 rra_gather_bvh_info_gfx12(const uint8_t *bvh, uint32_t node_id, struct rra_bvh_info *dst)
 {
-   uint32_t node_type = node_id & 7;
+   uint32_t node_type = node_id & 0xf;
 
    switch (node_type) {
    case radv_bvh_node_box32:
@@ -63,11 +71,12 @@ rra_gather_bvh_info_gfx12(const uint8_t *bvh, uint32_t node_id, struct rra_bvh_i
       dst->leaf_nodes_size += sizeof(struct radv_gfx12_primitive_node);
       break;
    default:
-      unreachable("Invalid node type");
+      if (node_type >= radv_bvh_node_triangle + 4 && !(node_type & 0x8))
+         unreachable("Invalid node type");
       break;
    }
 
-   const void *node = bvh + ((node_id & (~7u)) << 3);
+   const void *node = bvh + ((node_id & (~0xf)) << 3);
    if (node_type == radv_bvh_node_box32) {
       const struct radv_gfx12_box_node *src = node;
 
@@ -90,8 +99,16 @@ rra_gather_bvh_info_gfx12(const uint8_t *bvh, uint32_t node_id, struct rra_bvh_i
 
          rra_gather_bvh_info_gfx12(bvh, child_id, dst);
       }
+   } else if (node_type == radv_bvh_node_instance) {
+      dst->geometry_infos[0].primitive_count++;
    } else {
-      dst->geometry_infos[get_geometry_id(node, node_type)].primitive_count++;
+      uint32_t pair_index = (node_type & 0x3) | ((node_type & 0x8) >> 1);
+
+      if (BITSET_EXTRACT(node, 1024 - 29 * (pair_index + 1) + 17, 12))
+         dst->geometry_infos[get_geometry_id(node, pair_index * 2 + 0)].primitive_count++;
+
+      if (BITSET_EXTRACT(node, 1024 - 29 * (pair_index + 1) + 3, 12))
+         dst->geometry_infos[get_geometry_id(node, pair_index * 2 + 1)].primitive_count++;
    }
 }
 
@@ -114,7 +131,7 @@ rra_transcode_box8_node(struct rra_transcoding_context *ctx, const struct radv_g
       uint32_t child_type = (src->children[i].dword2 >> 24) & 0xf;
       if (child_type == radv_bvh_node_box32)
          internal_child_count++;
-      else
+      else if (child_type == radv_bvh_node_triangle || child_type == radv_bvh_node_instance)
          leaf_child_count++;
    }
 
@@ -141,19 +158,23 @@ rra_transcode_box8_node(struct rra_transcoding_context *ctx, const struct radv_g
          child_id = primitive_id | child_type;
          primitive_id += (child_size * RADV_GFX12_BVH_NODE_SIZE) >> 3;
          child_dst_offset = dst_leaf_offset;
-         dst_leaf_offset += RADV_GFX12_BVH_NODE_SIZE;
+         if (child_type == radv_bvh_node_triangle || child_type == radv_bvh_node_instance)
+            dst_leaf_offset += RADV_GFX12_BVH_NODE_SIZE;
       }
 
-      rra_transcode_node_gfx12(ctx, radv_bvh_node_box32 | (dst_offset >> 3), child_id, child_dst_offset);
+      if (child_type == radv_bvh_node_triangle || child_type == radv_bvh_node_instance ||
+          child_type == radv_bvh_node_box32)
+         rra_transcode_node_gfx12(ctx, radv_bvh_node_box32 | (dst_offset >> 3), child_id, child_dst_offset);
 
-      dst->children[i].dword2 = (dst->children[i].dword2 & 0x0fffffff) | (1 << 28);
+      if (child_type == radv_bvh_node_instance)
+         dst->children[i].dword2 = (dst->children[i].dword2 & 0x0fffffff) | (1 << 28);
    }
 }
 
 void
 rra_transcode_node_gfx12(struct rra_transcoding_context *ctx, uint32_t parent_id, uint32_t src_id, uint32_t dst_offset)
 {
-   uint32_t node_type = src_id & 7;
+   uint32_t node_type = src_id & 0xf;
    uint32_t src_offset = (src_id & (~7u)) << 3;
 
    const void *src_child_node = ctx->src + src_offset;
-- 
GitLab


From fc61f0304670f4a8a33153acbf5a655dc32f0b7b Mon Sep 17 00:00:00 2001
From: Konstantin Seurer <konstantin.seurer@gmail.com>
Date: Mon, 7 Apr 2025 22:19:42 +0200
Subject: [PATCH 8/8] radv: Emit compressed primitive nodes on GFX12

Emits two triangles per node whenever possible. The nir code will
revisit the triangle node to handle the second triangle only if both
triangles are interescted by the ray.
---
 src/amd/vulkan/bvh/build_interface.h          |  1 +
 src/amd/vulkan/bvh/encode.h                   | 64 +++++++++++++++++++
 src/amd/vulkan/bvh/encode_gfx12.comp          | 43 ++++++++++++-
 .../vulkan/nir/radv_nir_lower_ray_queries.c   |  4 ++
 src/amd/vulkan/nir/radv_nir_rt_common.c       | 38 ++++++++---
 src/amd/vulkan/nir/radv_nir_rt_common.h       |  3 +
 src/amd/vulkan/nir/radv_nir_rt_shader.c       |  5 ++
 src/amd/vulkan/radv_acceleration_structure.c  | 17 +++--
 8 files changed, 160 insertions(+), 15 deletions(-)

diff --git a/src/amd/vulkan/bvh/build_interface.h b/src/amd/vulkan/bvh/build_interface.h
index a49476d720392..22e76c3f1814b 100644
--- a/src/amd/vulkan/bvh/build_interface.h
+++ b/src/amd/vulkan/bvh/build_interface.h
@@ -24,6 +24,7 @@
 #define RADV_BUILD_FLAG_NO_INFS                 (1u << (VK_BUILD_FLAG_COUNT + 3))
 #define RADV_BUILD_FLAG_WRITE_LEAF_NODE_OFFSETS (1u << (VK_BUILD_FLAG_COUNT + 4))
 #define RADV_BUILD_FLAG_UPDATE_SINGLE_GEOMETRY  (1u << (VK_BUILD_FLAG_COUNT + 5))
+#define RADV_BUILD_FLAG_PAIR_COMPRESS_TRIANGLES (1u << (VK_BUILD_FLAG_COUNT + 6))
 
 #define RADV_COPY_MODE_COPY        0
 #define RADV_COPY_MODE_SERIALIZE   1
diff --git a/src/amd/vulkan/bvh/encode.h b/src/amd/vulkan/bvh/encode.h
index 625525ae07a52..eec16291c31b7 100644
--- a/src/amd/vulkan/bvh/encode.h
+++ b/src/amd/vulkan/bvh/encode.h
@@ -179,6 +179,70 @@ radv_encode_triangle_gfx12(VOID_REF dst, vk_ir_triangle_node src)
    bit_writer_finish(child_writer);
 }
 
+void
+radv_encode_triangle_gfx12(VOID_REF dst, vk_ir_triangle_node src0, vk_ir_triangle_node src1)
+{
+   bit_writer child_writer;
+   bit_writer_init(child_writer, dst);
+
+   bit_writer_write(child_writer, 31, 5); /* x_vertex_bits_minus_one */
+   bit_writer_write(child_writer, 31, 5); /* y_vertex_bits_minus_one */
+   bit_writer_write(child_writer, 31, 5); /* z_vertex_bits_minus_one */
+   bit_writer_write(child_writer, 0, 5);  /* trailing_zero_bits */
+   bit_writer_write(child_writer, 14, 4); /* geometry_index_base_bits_div_2 */
+   bit_writer_write(child_writer, 14, 4); /* geometry_index_bits_div_2 */
+   bit_writer_write(child_writer, 0, 3);  /* triangle_pair_count_minus_one */
+   bit_writer_write(child_writer, 0, 1);  /* vertex_type */
+   bit_writer_write(child_writer, 28, 5); /* primitive_index_base_bits */
+   bit_writer_write(child_writer, 28, 5); /* primitive_index_bits */
+   /* header + (9 floats + geometry_id) * 2 triangles */
+   bit_writer_write(child_writer, RADV_GFX12_PRIMITIVE_NODE_HEADER_SIZE + 2 * 9 * 32 + 2 * 28, 10);
+
+   bit_writer_write(child_writer, floatBitsToUint(src0.coords[0][0]), 32);
+   bit_writer_write(child_writer, floatBitsToUint(src0.coords[0][1]), 32);
+   bit_writer_write(child_writer, floatBitsToUint(src0.coords[0][2]), 32);
+   bit_writer_write(child_writer, floatBitsToUint(src0.coords[1][0]), 32);
+   bit_writer_write(child_writer, floatBitsToUint(src0.coords[1][1]), 32);
+   bit_writer_write(child_writer, floatBitsToUint(src0.coords[1][2]), 32);
+   bit_writer_write(child_writer, floatBitsToUint(src0.coords[2][0]), 32);
+   bit_writer_write(child_writer, floatBitsToUint(src0.coords[2][1]), 32);
+   bit_writer_write(child_writer, floatBitsToUint(src0.coords[2][2]), 32);
+
+   bit_writer_write(child_writer, floatBitsToUint(src1.coords[0][0]), 32);
+   bit_writer_write(child_writer, floatBitsToUint(src1.coords[0][1]), 32);
+   bit_writer_write(child_writer, floatBitsToUint(src1.coords[0][2]), 32);
+   bit_writer_write(child_writer, floatBitsToUint(src1.coords[1][0]), 32);
+   bit_writer_write(child_writer, floatBitsToUint(src1.coords[1][1]), 32);
+   bit_writer_write(child_writer, floatBitsToUint(src1.coords[1][2]), 32);
+   bit_writer_write(child_writer, floatBitsToUint(src1.coords[2][0]), 32);
+   bit_writer_write(child_writer, floatBitsToUint(src1.coords[2][1]), 32);
+   bit_writer_write(child_writer, floatBitsToUint(src1.coords[2][2]), 32);
+
+   bit_writer_write(child_writer, src1.geometry_id_and_flags & 0xfffffff, 28);
+   bit_writer_write(child_writer, src0.geometry_id_and_flags & 0xfffffff, 28);
+   bit_writer_write(child_writer, src0.triangle_id, 28);
+   bit_writer_write(child_writer, src1.triangle_id, 28);
+
+   bit_writer_skip_to(child_writer, 32 * 32 - RADV_GFX12_PRIMITIVE_NODE_PAIR_DESC_SIZE);
+
+   uint32_t opaque0 = (src0.geometry_id_and_flags & VK_GEOMETRY_OPAQUE) != 0 ? 1 : 0;
+   uint32_t opaque1 = (src1.geometry_id_and_flags & VK_GEOMETRY_OPAQUE) != 0 ? 1 : 0;
+
+   bit_writer_write(child_writer, 1, 1);       /* prim_range_stop */
+   bit_writer_write(child_writer, 0, 1);       /* tri1_double_sided */
+   bit_writer_write(child_writer, opaque1, 1); /* tri1_opaque */
+   bit_writer_write(child_writer, 3, 4);       /* tri1_v0_index */
+   bit_writer_write(child_writer, 4, 4);       /* tri1_v1_index */
+   bit_writer_write(child_writer, 5, 4);       /* tri1_v2_index */
+   bit_writer_write(child_writer, 0, 1);       /* tri0_double_sided */
+   bit_writer_write(child_writer, opaque0, 1); /* tri0_opaque */
+   bit_writer_write(child_writer, 0, 4);       /* tri0_v0_index */
+   bit_writer_write(child_writer, 1, 4);       /* tri0_v1_index */
+   bit_writer_write(child_writer, 2, 4);       /* tri0_v2_index */
+
+   bit_writer_finish(child_writer);
+}
+
 void
 radv_encode_aabb_gfx12(VOID_REF dst, vk_ir_aabb_node src)
 {
diff --git a/src/amd/vulkan/bvh/encode_gfx12.comp b/src/amd/vulkan/bvh/encode_gfx12.comp
index eaf7ff081387f..639587baefe63 100644
--- a/src/amd/vulkan/bvh/encode_gfx12.comp
+++ b/src/amd/vulkan/bvh/encode_gfx12.comp
@@ -81,6 +81,8 @@ encode_gfx12(uint32_t ir_leaf_node_size, REF(vk_ir_box_node) intermediate_intern
       if (cluster.invocation_index < 2)
          child = src.children[cluster.invocation_index];
 
+      uint32_t second_child = RADV_BVH_INVALID_NODE;
+
       while (true) {
          uint32_t valid_children = radv_ballot(cluster, child != RADV_BVH_INVALID_NODE);
          if ((valid_children & 0x80) != 0 || valid_children == 0)
@@ -110,12 +112,30 @@ encode_gfx12(uint32_t ir_leaf_node_size, REF(vk_ir_box_node) intermediate_intern
             if (left == RADV_BVH_INVALID_NODE) {
                left = right;
                right = RADV_BVH_INVALID_NODE;
+            } else if (right != RADV_BVH_INVALID_NODE && ir_id_to_type(left) == vk_ir_node_triangle &&
+                       ir_id_to_type(right) == vk_ir_node_triangle &&
+                       VK_BUILD_FLAG(RADV_BUILD_FLAG_PAIR_COMPRESS_TRIANGLES)) {
+               second_child = right;
+               right = RADV_BVH_INVALID_NODE;
             }
 
             child = left;
          }
          right = radv_read_invocation(cluster, collapse_index, right);
 
+         if (VK_BUILD_FLAG(RADV_BUILD_FLAG_PAIR_COMPRESS_TRIANGLES)) {
+            bool is_valid_triangle = child != RADV_BVH_INVALID_NODE && ir_id_to_type(child) == vk_ir_node_triangle;
+            uint32_t right_pair_mask =
+               radv_ballot(cluster, is_valid_triangle && second_child == RADV_BVH_INVALID_NODE &&
+                                       right != RADV_BVH_INVALID_NODE && ir_id_to_type(right) == vk_ir_node_triangle);
+
+            if (right_pair_mask != 0) {
+               if (cluster.invocation_index == findLSB(right_pair_mask))
+                  second_child = right;
+               continue;
+            }
+         }
+
          if (cluster.invocation_index == findMSB(valid_children) + 1)
             child = right;
       }
@@ -201,12 +221,25 @@ encode_gfx12(uint32_t ir_leaf_node_size, REF(vk_ir_box_node) intermediate_intern
                DEREF(child_dst_offset) = dst_offset;
             }
 
+            if (VK_BUILD_FLAG(RADV_BUILD_FLAG_WRITE_LEAF_NODE_OFFSETS) && second_child != RADV_BVH_INVALID_NODE) {
+               uint32_t leaf_index = ir_id_to_offset(second_child) / ir_leaf_node_size;
+               REF(uint32_t) child_dst_offset = REF(uint32_t)(args.output_base + args.leaf_node_offsets_offset);
+               child_dst_offset = INDEX(uint32_t, child_dst_offset, leaf_index);
+               DEREF(child_dst_offset) = dst_offset;
+            }
+
             VOID_REF dst_leaf_addr = args.output_base + args.output_bvh_offset + dst_offset;
 
             switch (args.geometry_type) {
             case VK_GEOMETRY_TYPE_TRIANGLES_KHR: {
-               vk_ir_triangle_node src_node = DEREF(REF(vk_ir_triangle_node)(OFFSET(args.intermediate_bvh, offset)));
-               radv_encode_triangle_gfx12(dst_leaf_addr, src_node);
+               vk_ir_triangle_node src_node0 = DEREF(REF(vk_ir_triangle_node)(OFFSET(args.intermediate_bvh, offset)));
+               if (VK_BUILD_FLAG(RADV_BUILD_FLAG_PAIR_COMPRESS_TRIANGLES) && second_child != RADV_BVH_INVALID_NODE) {
+                  vk_ir_triangle_node src_node1 =
+                     DEREF(REF(vk_ir_triangle_node)(OFFSET(args.intermediate_bvh, ir_id_to_offset(second_child))));
+                  radv_encode_triangle_gfx12(dst_leaf_addr, src_node0, src_node1);
+               } else {
+                  radv_encode_triangle_gfx12(dst_leaf_addr, src_node0);
+               }
                break;
             }
             case VK_GEOMETRY_TYPE_AABBS_KHR: {
@@ -229,6 +262,12 @@ encode_gfx12(uint32_t ir_leaf_node_size, REF(vk_ir_box_node) intermediate_intern
          }
 
          vk_aabb child_aabb = DEREF(REF(vk_ir_node) OFFSET(args.intermediate_bvh, offset)).aabb;
+         if (second_child != RADV_BVH_INVALID_NODE) {
+            vk_aabb second_child_aabb =
+               DEREF(REF(vk_ir_node) OFFSET(args.intermediate_bvh, ir_id_to_offset(second_child))).aabb;
+            child_aabb.min = min(child_aabb.min, second_child_aabb.min);
+            child_aabb.max = max(child_aabb.max, second_child_aabb.max);
+         }
 
          radv_gfx12_box_child box_child;
          /* TODO: subtree flags culling */
diff --git a/src/amd/vulkan/nir/radv_nir_lower_ray_queries.c b/src/amd/vulkan/nir/radv_nir_lower_ray_queries.c
index 2dc30e5a47425..c46cbdc3c678a 100644
--- a/src/amd/vulkan/nir/radv_nir_lower_ray_queries.c
+++ b/src/amd/vulkan/nir/radv_nir_lower_ray_queries.c
@@ -79,6 +79,7 @@ enum radv_ray_query_field {
    radv_ray_query_trav_previous_node,
    radv_ray_query_trav_instance_top_node,
    radv_ray_query_trav_instance_bottom_node,
+   radv_ray_query_trav_second_iteration,
    radv_ray_query_stack,
    radv_ray_query_field_count,
 };
@@ -115,6 +116,7 @@ radv_get_ray_query_type()
    FIELD(trav_previous_node, glsl_uint_type());
    FIELD(trav_instance_top_node, glsl_uint_type());
    FIELD(trav_instance_bottom_node, glsl_uint_type());
+   FIELD(trav_second_iteration, glsl_bool_type());
    FIELD(stack, glsl_array_type(glsl_uint_type(), MAX_SCRATCH_STACK_ENTRY_COUNT, 0));
 
 #undef FIELD
@@ -292,6 +294,7 @@ lower_rq_initialize(nir_builder *b, nir_intrinsic_instr *instr, struct ray_query
    rq_store(b, rq, trav_previous_node, nir_imm_int(b, RADV_BVH_INVALID_NODE));
    rq_store(b, rq, trav_instance_top_node, nir_imm_int(b, RADV_BVH_INVALID_NODE));
    rq_store(b, rq, trav_instance_bottom_node, nir_imm_int(b, RADV_BVH_NO_INSTANCE_ROOT));
+   rq_store(b, rq, trav_second_iteration, nir_imm_false(b));
 
    rq_store(b, rq, trav_top_stack, nir_imm_int(b, -1));
 
@@ -491,6 +494,7 @@ lower_rq_proceed(nir_builder *b, nir_intrinsic_instr *instr, struct ray_query_va
       .previous_node = rq_deref(b, rq, trav_previous_node),
       .instance_top_node = rq_deref(b, rq, trav_instance_top_node),
       .instance_bottom_node = rq_deref(b, rq, trav_instance_bottom_node),
+      .second_iteration = rq_deref(b, rq, trav_second_iteration),
       .instance_addr = isec_deref(b, candidate, instance_addr),
       .sbt_offset_and_flags = isec_deref(b, candidate, sbt_offset_and_flags),
    };
diff --git a/src/amd/vulkan/nir/radv_nir_rt_common.c b/src/amd/vulkan/nir/radv_nir_rt_common.c
index c41cb6258642c..f5bc2f35ed312 100644
--- a/src/amd/vulkan/nir/radv_nir_rt_common.c
+++ b/src/amd/vulkan/nir/radv_nir_rt_common.c
@@ -575,23 +575,43 @@ insert_traversal_triangle_case(struct radv_device *device, nir_builder *b, const
 static void
 insert_traversal_triangle_case_gfx12(struct radv_device *device, nir_builder *b,
                                      const struct radv_ray_traversal_args *args, const struct radv_ray_flags *ray_flags,
-                                     nir_def *result, nir_def *bvh_node)
+                                     nir_def *result, nir_def *global_bvh_node, nir_def *bvh_node)
 {
    if (!args->triangle_cb)
       return;
 
+   nir_def *t[2] = {
+      nir_channel(b, result, 0),
+      nir_channel(b, result, 4),
+   };
+   nir_def *triangle0_first = nir_flt(b, t[0], t[1]);
+
+   nir_def *second_t = nir_bcsel(b, triangle0_first, t[1], t[0]);
+   nir_def *second_iteration = nir_load_deref(b, args->vars.second_iteration);
+   nir_def *revisit = nir_iand(b, nir_inot(b, second_iteration), nir_flt(b, second_t, nir_load_deref(b, args->vars.tmax)));
+   nir_def *next_node = nir_bcsel(b, revisit, bvh_node, nir_imm_int(b, RADV_BVH_INVALID_NODE));
+   nir_store_deref(b, args->vars.current_node, next_node, 0x1);
+   nir_store_deref(b, args->vars.second_iteration, revisit, 0x1);
+
+   nir_def *triangle0 = nir_ixor(b, triangle0_first, second_iteration);
+
    struct radv_triangle_intersection intersection;
-   intersection.t = nir_channel(b, result, 0);
+   intersection.t = nir_bcsel(b, triangle0, t[0], t[1]);
 
    nir_push_if(b, nir_iand(b, nir_flt(b, intersection.t, nir_load_deref(b, args->vars.tmax)),
                            nir_flt(b, args->tmin, intersection.t)));
    {
-      intersection.frontface = nir_inot(b, nir_test_mask(b, nir_channel(b, result, 3), 1));
-      intersection.base.node_addr = build_node_to_addr(device, b, bvh_node, false);
-      intersection.base.primitive_id = nir_ishr_imm(b, nir_channel(b, result, 3), 1);
-      intersection.base.geometry_id_and_flags = nir_ishr_imm(b, nir_channel(b, result, 8), 2);
-      intersection.base.opaque = nir_inot(b, nir_test_mask(b, nir_channel(b, result, 2), 1u << 31));
-      intersection.barycentrics = nir_fabs(b, nir_channels(b, result, 0x3 << 1));
+      nir_def *dword1 = nir_bcsel(b, triangle0, nir_channel(b, result, 1), nir_channel(b, result, 5));
+      nir_def *dword2 = nir_bcsel(b, triangle0, nir_channel(b, result, 2), nir_channel(b, result, 6));
+      nir_def *dword3 = nir_bcsel(b, triangle0, nir_channel(b, result, 3), nir_channel(b, result, 7));
+
+      intersection.frontface = nir_inot(b, nir_test_mask(b, dword3, 1));
+      intersection.base.node_addr = build_node_to_addr(device, b, global_bvh_node, false);
+      intersection.base.primitive_id = nir_ishr_imm(b, dword3, 1);
+      intersection.base.geometry_id_and_flags =
+         nir_ishr_imm(b, nir_bcsel(b, triangle0, nir_channel(b, result, 8), nir_channel(b, result, 9)), 2);
+      intersection.base.opaque = nir_inot(b, nir_test_mask(b, dword2, 1u << 31));
+      intersection.barycentrics = nir_fabs(b, nir_vec2(b, dword1, dword2));
 
       nir_push_if(b, nir_bcsel(b, intersection.base.opaque, ray_flags->no_cull_opaque, ray_flags->no_cull_no_opaque));
       {
@@ -1086,7 +1106,7 @@ radv_build_ray_traversal_gfx12(struct radv_device *device, nir_builder *b, const
          nir_push_else(b, NULL);
          {
             nir_push_if(b, ray_flags.no_skip_triangles);
-            insert_traversal_triangle_case_gfx12(device, b, args, &ray_flags, result, global_bvh_node);
+            insert_traversal_triangle_case_gfx12(device, b, args, &ray_flags, result, global_bvh_node, bvh_node);
             nir_pop_if(b, NULL);
          }
          nir_pop_if(b, NULL);
diff --git a/src/amd/vulkan/nir/radv_nir_rt_common.h b/src/amd/vulkan/nir/radv_nir_rt_common.h
index aaef6ca72cd15..e3efb46bc4115 100644
--- a/src/amd/vulkan/nir/radv_nir_rt_common.h
+++ b/src/amd/vulkan/nir/radv_nir_rt_common.h
@@ -103,6 +103,9 @@ struct radv_ray_traversal_vars {
    nir_deref_instr *instance_top_node;
    nir_deref_instr *instance_bottom_node;
 
+   /* Whether the current iteration revisits the last triangle node to handle the second triangle. */
+   nir_deref_instr *second_iteration;
+
    /* Information about the current instance used for culling. */
    nir_deref_instr *instance_addr;
    nir_deref_instr *sbt_offset_and_flags;
diff --git a/src/amd/vulkan/nir/radv_nir_rt_shader.c b/src/amd/vulkan/nir/radv_nir_rt_shader.c
index 086726052ecc2..7fa8e183213f2 100644
--- a/src/amd/vulkan/nir/radv_nir_rt_shader.c
+++ b/src/amd/vulkan/nir/radv_nir_rt_shader.c
@@ -1223,6 +1223,7 @@ struct rt_traversal_vars {
    nir_variable *previous_node;
    nir_variable *instance_top_node;
    nir_variable *instance_bottom_node;
+   nir_variable *second_iteration;
 };
 
 static struct rt_traversal_vars
@@ -1248,6 +1249,8 @@ init_traversal_vars(nir_builder *b)
    ret.instance_top_node = nir_variable_create(b->shader, nir_var_shader_temp, glsl_uint_type(), "instance_top_node");
    ret.instance_bottom_node =
       nir_variable_create(b->shader, nir_var_shader_temp, glsl_uint_type(), "instance_bottom_node");
+   ret.second_iteration =
+      nir_variable_create(b->shader, nir_var_shader_temp, glsl_bool_type(), "second_iteration");
    return ret;
 }
 
@@ -1542,6 +1545,7 @@ radv_build_traversal(struct radv_device *device, struct radv_ray_tracing_pipelin
    nir_store_var(b, trav_vars.previous_node, nir_imm_int(b, RADV_BVH_INVALID_NODE), 0x1);
    nir_store_var(b, trav_vars.instance_top_node, nir_imm_int(b, RADV_BVH_INVALID_NODE), 0x1);
    nir_store_var(b, trav_vars.instance_bottom_node, nir_imm_int(b, RADV_BVH_NO_INSTANCE_ROOT), 0x1);
+   nir_store_var(b, trav_vars.second_iteration, nir_imm_false(b), 0x1);
 
    nir_store_var(b, trav_vars.top_stack, nir_imm_int(b, -1), 1);
 
@@ -1558,6 +1562,7 @@ radv_build_traversal(struct radv_device *device, struct radv_ray_tracing_pipelin
       .previous_node = nir_build_deref_var(b, trav_vars.previous_node),
       .instance_top_node = nir_build_deref_var(b, trav_vars.instance_top_node),
       .instance_bottom_node = nir_build_deref_var(b, trav_vars.instance_bottom_node),
+      .second_iteration = nir_build_deref_var(b, trav_vars.second_iteration),
       .instance_addr = nir_build_deref_var(b, trav_vars.instance_addr),
       .sbt_offset_and_flags = nir_build_deref_var(b, trav_vars.sbt_offset_and_flags),
    };
diff --git a/src/amd/vulkan/radv_acceleration_structure.c b/src/amd/vulkan/radv_acceleration_structure.c
index 5309d6388c533..1ae3a419af30f 100644
--- a/src/amd/vulkan/radv_acceleration_structure.c
+++ b/src/amd/vulkan/radv_acceleration_structure.c
@@ -70,6 +70,7 @@ struct update_scratch_layout {
 enum radv_encode_key_bits {
    RADV_ENCODE_KEY_COMPACT = (1 << 0),
    RADV_ENCODE_KEY_WRITE_LEAF_NODE_OFFSETS = (1 << 1),
+   RADV_ENCODE_KEY_PAIR_COMPRESS_GFX12 = (1 << 2),
 };
 
 static void
@@ -425,6 +426,12 @@ radv_get_build_config(VkDevice _device, struct vk_acceleration_structure_build_s
       if ((state->build_info->flags & VK_BUILD_ACCELERATION_STRUCTURE_ALLOW_DATA_ACCESS_KHR) ||
           state->build_info->type != VK_ACCELERATION_STRUCTURE_TYPE_BOTTOM_LEVEL_KHR)
          encode_key |= RADV_ENCODE_KEY_WRITE_LEAF_NODE_OFFSETS;
+
+      VkGeometryTypeKHR geometry_type = vk_get_as_geometry_type(state->build_info);
+      if (!(state->build_info->flags & (VK_BUILD_ACCELERATION_STRUCTURE_ALLOW_UPDATE_BIT_KHR |
+                                        VK_BUILD_ACCELERATION_STRUCTURE_ALLOW_DATA_ACCESS_KHR)) &&
+          geometry_type == VK_GEOMETRY_TYPE_TRIANGLES_KHR)
+         encode_key |= RADV_ENCODE_KEY_PAIR_COMPRESS_GFX12;
    }
 
    if (state->build_info->flags & VK_BUILD_ACCELERATION_STRUCTURE_ALLOW_COMPACTION_BIT_KHR)
@@ -502,6 +509,8 @@ radv_build_flags(VkCommandBuffer commandBuffer, uint32_t key)
       flags |= RADV_BUILD_FLAG_NO_INFS;
    if (key & RADV_ENCODE_KEY_WRITE_LEAF_NODE_OFFSETS)
       flags |= RADV_BUILD_FLAG_WRITE_LEAF_NODE_OFFSETS;
+   if (key & RADV_ENCODE_KEY_PAIR_COMPRESS_GFX12)
+      flags |= RADV_BUILD_FLAG_PAIR_COMPRESS_TRIANGLES;
 
    return flags;
 }
@@ -520,10 +529,10 @@ static VkResult
 radv_encode_bind_pipeline_gfx12(VkCommandBuffer commandBuffer,
                                 const struct vk_acceleration_structure_build_state *state)
 {
-   radv_bvh_build_bind_pipeline(
-      commandBuffer, RADV_META_OBJECT_KEY_BVH_ENCODE, encode_gfx12_spv, sizeof(encode_gfx12_spv),
-      sizeof(struct encode_gfx12_args),
-      radv_build_flags(commandBuffer, state->config.encode_key[0]) & RADV_BUILD_FLAG_WRITE_LEAF_NODE_OFFSETS);
+   radv_bvh_build_bind_pipeline(commandBuffer, RADV_META_OBJECT_KEY_BVH_ENCODE, encode_gfx12_spv,
+                                sizeof(encode_gfx12_spv), sizeof(struct encode_gfx12_args),
+                                radv_build_flags(commandBuffer, state->config.encode_key[0]) &
+                                   (RADV_BUILD_FLAG_WRITE_LEAF_NODE_OFFSETS | RADV_BUILD_FLAG_PAIR_COMPRESS_TRIANGLES));
 
    return VK_SUCCESS;
 }
-- 
GitLab