Use start index of the instanced transform instead of the buffer devi…

…ce address, and additionally pass the transform buffer as storage buffer.
stripe2933 · Jan 27, 2025 · 3332184 · 3332184
1 parent 00e2736
commit 3332184
Show file tree

Hide file tree

Showing 15 changed files with 65 additions and 48 deletions.
diff --git a/impl/MainApp.cpp b/impl/MainApp.cpp
@@ -852,9 +852,10 @@ void vk_gltf_viewer::MainApp::loadGltf(const std::filesystem::path &path) {
     }));
     gpu.device.updateDescriptorSets({
         sharedData.assetDescriptorSet.getWriteOne<0>({ gltf->assetGpuBuffers.getPrimitiveBuffer(), 0, vk::WholeSize }),
-        sharedData.assetDescriptorSet.getWrite<1>(gltf->nodeBuffer.descriptorInfo),
-        sharedData.assetDescriptorSet.getWriteOne<2>({ gltf->assetGpuBuffers.materialBuffer, 0, vk::WholeSize }),
-        sharedData.assetDescriptorSet.getWrite<3>(imageInfos),
+        sharedData.assetDescriptorSet.getWrite<1>(gltf->nodeBuffer.getDescriptorInfo()),
+        sharedData.assetDescriptorSet.getWrite<2>(gltf->instancedNodeWorldTransformBuffer.getDescriptorInfo()),
+        sharedData.assetDescriptorSet.getWriteOne<3>({ gltf->assetGpuBuffers.materialBuffer, 0, vk::WholeSize }),
+        sharedData.assetDescriptorSet.getWrite<4>(imageInfos),
     }, {});
 
     // TODO: due to the ImGui's gamma correction issue, base color/emissive texture is rendered darker than it should be.

diff --git a/interface/vulkan/buffer/InstancedNodeWorldTransforms.cppm b/interface/vulkan/buffer/InstancedNodeWorldTransforms.cppm
@@ -41,10 +41,14 @@ namespace vk_gltf_viewer::vulkan::buffer {
         ) : asset { asset },
             instanceOffsets { createInstanceOffsets() },
             buffer { createBuffer(gpu.allocator, adapter) },
-            bufferAddress { gpu.device.getBufferAddress({ buffer }) } { }
+            descriptorInfo { buffer, 0, vk::WholeSize } { }
 
-        [[nodiscard]] vk::DeviceAddress getTransformStartAddress(std::size_t nodeIndex) const noexcept {
-            return bufferAddress + sizeof(fastgltf::math::fmat4x4) * instanceOffsets[nodeIndex];
+        [[nodiscard]] std::uint32_t getStartIndex(std::size_t nodeIndex) const noexcept {
+            return instanceOffsets[nodeIndex];
+        }
+
+        [[nodiscard]] const vk::DescriptorBufferInfo &getDescriptorInfo() const noexcept {
+            return descriptorInfo;
         }
 
         /**
@@ -121,7 +125,7 @@ namespace vk_gltf_viewer::vulkan::buffer {
          * Be careful that there is no transform matrix related about node C, because it is meshless.
          */
         vku::MappedBuffer buffer;
-        vk::DeviceAddress bufferAddress;
+        vk::DescriptorBufferInfo descriptorInfo;
 
         [[nodiscard]] std::vector<std::uint32_t> createInstanceOffsets() const {
             std::vector<std::uint32_t> result;
@@ -154,7 +158,7 @@ namespace vk_gltf_viewer::vulkan::buffer {
             vku::MappedBuffer result { allocator, vk::BufferCreateInfo {
                 {},
                 sizeof(fastgltf::math::fmat4x4) * instanceOffsets.back(),
-                vk::BufferUsageFlagBits::eStorageBuffer | vk::BufferUsageFlagBits::eShaderDeviceAddress,
+                vk::BufferUsageFlagBits::eStorageBuffer,
             } };
 
             const std::span data = result.asRange<fastgltf::math::fmat4x4>();

diff --git a/interface/vulkan/buffer/Nodes.cppm b/interface/vulkan/buffer/Nodes.cppm
@@ -7,22 +7,25 @@ export import :vulkan.buffer.InstancedNodeWorldTransforms;
 
 namespace vk_gltf_viewer::vulkan::buffer {
     export class Nodes {
-        /**
-         * @brief Buffer with the start address of the instanced node world transform buffer.
-         */
-        vku::AllocatedBuffer buffer;
-
     public:
-        vk::DescriptorBufferInfo descriptorInfo;
-
         Nodes(
             const fastgltf::Asset &asset,
             const InstancedNodeWorldTransforms &instancedNodeWorldTransformBuffer,
             const Gpu &gpu [[clang::lifetimebound]]
         ) : buffer { createBuffer(asset, instancedNodeWorldTransformBuffer, gpu) },
             descriptorInfo { buffer, 0, vk::WholeSize } { }
 
+        [[nodiscard]] const vk::DescriptorBufferInfo &getDescriptorInfo() const noexcept {
+            return descriptorInfo;
+        }
+
     private:
+        /**
+         * @brief Buffer with the start index of the instanced node world transform buffer.
+         */
+        vku::AllocatedBuffer buffer;
+        vk::DescriptorBufferInfo descriptorInfo;
+
         [[nodiscard]] vku::AllocatedBuffer createBuffer(
             const fastgltf::Asset &asset,
             const InstancedNodeWorldTransforms &instancedNodeWorldTransformBuffer,
@@ -31,7 +34,7 @@ namespace vk_gltf_viewer::vulkan::buffer {
             vku::AllocatedBuffer buffer = vku::MappedBuffer {
                 gpu.allocator,
                 std::from_range, ranges::views::upto(asset.nodes.size()) | std::views::transform([&](std::size_t nodeIndex) {
-                    return instancedNodeWorldTransformBuffer.getTransformStartAddress(nodeIndex);
+                    return instancedNodeWorldTransformBuffer.getStartIndex(nodeIndex);
                 }),
                 gpu.isUmaDevice ? vk::BufferUsageFlagBits::eStorageBuffer : vk::BufferUsageFlagBits::eTransferSrc,
             }.unmap();

diff --git a/interface/vulkan/descriptor_set_layout/Asset.cppm b/interface/vulkan/descriptor_set_layout/Asset.cppm
@@ -4,12 +4,13 @@ import std;
 export import vku;
 
 namespace vk_gltf_viewer::vulkan::dsl {
-    export struct Asset : vku::DescriptorSetLayout<vk::DescriptorType::eStorageBuffer, vk::DescriptorType::eStorageBuffer, vk::DescriptorType::eStorageBuffer, vk::DescriptorType::eCombinedImageSampler> {
+    export struct Asset : vku::DescriptorSetLayout<vk::DescriptorType::eStorageBuffer, vk::DescriptorType::eStorageBuffer, vk::DescriptorType::eStorageBuffer, vk::DescriptorType::eStorageBuffer, vk::DescriptorType::eCombinedImageSampler> {
         Asset(const vk::raii::Device &device [[clang::lifetimebound]], std::uint32_t textureCount)
             : DescriptorSetLayout { device, vk::StructureChain {
                 vk::DescriptorSetLayoutCreateInfo {
                     vk::DescriptorSetLayoutCreateFlagBits::eUpdateAfterBindPool,
                     vku::unsafeProxy(getBindings(
+                        { 1, vk::ShaderStageFlagBits::eVertex },
                         { 1, vk::ShaderStageFlagBits::eVertex },
                         { 1, vk::ShaderStageFlagBits::eVertex },
                         { 1, vk::ShaderStageFlagBits::eVertex | vk::ShaderStageFlagBits::eFragment },
@@ -20,6 +21,7 @@ namespace vk_gltf_viewer::vulkan::dsl {
                         {},
                         {},
                         {},
+                        {},
                         vk::DescriptorBindingFlagBits::eUpdateAfterBind,
                     }),
                 },

diff --git a/shaders/depth.vert b/shaders/depth.vert
@@ -12,15 +12,16 @@
 #include "indexing.glsl"
 #include "types.glsl"
 
-layout (std430, buffer_reference, buffer_reference_align = 64) readonly buffer Node { mat4 transforms[]; };
-
 layout (location = 0) flat out uint outNodeIndex;
 
 layout (set = 0, binding = 0) readonly buffer PrimitiveBuffer {
     Primitive primitives[];
 };
 layout (set = 0, binding = 1, std430) readonly buffer NodeBuffer {
-    Node nodes[];
+    uint instancedTransformStartIndices[];
+};
+layout (set = 0, binding = 2) readonly buffer InstancedTransformBuffer {
+    mat4 instancedTransforms[];
 };
 
 layout (push_constant) uniform PushConstant {

diff --git a/shaders/indexing.glsl b/shaders/indexing.glsl
@@ -7,7 +7,8 @@
 #define PRIMITIVE_INDEX gl_BaseInstance & 0xFFFFU
 #define PRIMITIVE primitives[PRIMITIVE_INDEX]
 #define NODE_INDEX gl_BaseInstance >> 16U
-#define TRANSFORM Node(nodes[NODE_INDEX]).transforms[gl_InstanceIndex - gl_BaseInstance]
+#define INSTANCE_INDEX gl_InstanceIndex - gl_BaseInstance
+#define TRANSFORM instancedTransforms[instancedTransformStartIndices[NODE_INDEX] + INSTANCE_INDEX]
 #define MATERIAL_INDEX PRIMITIVE.materialIndex
 #define MATERIAL materials[MATERIAL_INDEX]
 

diff --git a/shaders/jump_flood_seed.vert b/shaders/jump_flood_seed.vert
@@ -12,13 +12,14 @@
 #include "indexing.glsl"
 #include "types.glsl"
 
-layout (std430, buffer_reference, buffer_reference_align = 64) readonly buffer Node { mat4 transforms[]; };
-
 layout (set = 0, binding = 0) readonly buffer PrimitiveBuffer {
     Primitive primitives[];
 };
 layout (set = 0, binding = 1, std430) readonly buffer NodeBuffer {
-    Node nodes[];
+    uint instancedTransformStartIndices[];
+};
+layout (set = 0, binding = 2) readonly buffer InstancedTransformBuffer {
+    mat4 instancedTransforms[];
 };
 
 layout (push_constant) uniform PushConstant {

diff --git a/shaders/mask_depth.frag b/shaders/mask_depth.frag
@@ -27,10 +27,10 @@ layout (location = 2) in FRAG_VARIADIC_IN {
 
 layout (location = 0) out uint outNodeIndex;
 
-layout (set = 0, binding = 2, std430) readonly buffer MaterialBuffer {
+layout (set = 0, binding = 3, std430) readonly buffer MaterialBuffer {
     Material materials[];
 };
-layout (set = 0, binding = 3) uniform sampler2D textures[];
+layout (set = 0, binding = 4) uniform sampler2D textures[];
 
 void main(){
     float baseColorAlpha = MATERIAL.baseColorFactor.a;

diff --git a/shaders/mask_depth.vert b/shaders/mask_depth.vert
@@ -17,8 +17,6 @@
 layout (constant_id = 0) const uint TEXCOORD_COMPONENT_TYPE = 5126; // FLOAT
 layout (constant_id = 1) const uint COLOR_COMPONENT_TYPE = 5126; // FLOAT
 
-layout (std430, buffer_reference, buffer_reference_align = 64) readonly buffer Node { mat4 transforms[]; };
-
 layout (location = 0) flat out uint outNodeIndex;
 layout (location = 1) flat out uint outMaterialIndex;
 #if HAS_VARIADIC_OUT
@@ -36,9 +34,12 @@ layout (set = 0, binding = 0) readonly buffer PrimitiveBuffer {
     Primitive primitives[];
 };
 layout (set = 0, binding = 1, std430) readonly buffer NodeBuffer {
-    Node nodes[];
+    uint instancedTransformStartIndices[];
+};
+layout (set = 0, binding = 2) readonly buffer InstancedTransformBuffer {
+    mat4 instancedTransforms[];
 };
-layout (set = 0, binding = 2, std430) readonly buffer MaterialBuffer {
+layout (set = 0, binding = 3, std430) readonly buffer MaterialBuffer {
     Material materials[];
 };
 

diff --git a/shaders/mask_jump_flood_seed.frag b/shaders/mask_jump_flood_seed.frag
@@ -26,10 +26,10 @@ layout (location = 1) in FRAG_VARIDIC_IN {
 
 layout (location = 0) out uvec2 outCoordinate;
 
-layout (set = 0, binding = 2, std430) readonly buffer MaterialBuffer {
+layout (set = 0, binding = 3, std430) readonly buffer MaterialBuffer {
     Material materials[];
 };
-layout (set = 0, binding = 3) uniform sampler2D textures[];
+layout (set = 0, binding = 4) uniform sampler2D textures[];
 
 void main(){
     float baseColorAlpha = MATERIAL.baseColorFactor.a;

diff --git a/shaders/mask_jump_flood_seed.vert b/shaders/mask_jump_flood_seed.vert
@@ -17,8 +17,6 @@
 layout (constant_id = 0) const uint TEXCOORD_COMPONENT_TYPE = 5126; // FLOAT
 layout (constant_id = 1) const uint COLOR_COMPONENT_TYPE = 5126; // FLOAT
 
-layout (std430, buffer_reference, buffer_reference_align = 64) readonly buffer Node { mat4 transforms[]; };
-
 layout (location = 0) flat out uint outMaterialIndex;
 #if HAS_VARIADIC_OUT
 layout (location = 1) out VS_VARIADIC_OUT {
@@ -35,9 +33,12 @@ layout (set = 0, binding = 0) readonly buffer PrimitiveBuffer {
     Primitive primitives[];
 };
 layout (set = 0, binding = 1, std430) readonly buffer NodeBuffer {
-    Node nodes[];
+    uint instancedTransformStartIndices[];
+};
+layout (set = 0, binding = 2) readonly buffer InstancedTransformBuffer {
+    mat4 instancedTransforms[];
 };
-layout (set = 0, binding = 2, std430) readonly buffer MaterialBuffer {
+layout (set = 0, binding = 3, std430) readonly buffer MaterialBuffer {
     Material materials[];
 };
 

diff --git a/shaders/primitive.frag b/shaders/primitive.frag
@@ -54,10 +54,10 @@ layout (set = 0, binding = 0, scalar) uniform SphericalHarmonicsBuffer {
 layout (set = 0, binding = 1) uniform samplerCube prefilteredmap;
 layout (set = 0, binding = 2) uniform sampler2D brdfmap;
 
-layout (set = 1, binding = 2, std430) readonly buffer MaterialBuffer {
+layout (set = 1, binding = 3, std430) readonly buffer MaterialBuffer {
     Material materials[];
 };
-layout (set = 1, binding = 3) uniform sampler2D textures[];
+layout (set = 1, binding = 4) uniform sampler2D textures[];
 
 layout (push_constant, std430) uniform PushConstant {
     mat4 projectionView;

diff --git a/shaders/primitive.vert b/shaders/primitive.vert
@@ -18,8 +18,6 @@ layout (constant_id = 0) const uint PACKED_TEXCOORD_COMPONENT_TYPES = 0x06060606
 layout (constant_id = 1) const uint COLOR_COMPONENT_COUNT = 0;
 layout (constant_id = 2) const uint COLOR_COMPONENT_TYPE = 5126; // FLOAT
 
-layout (std430, buffer_reference, buffer_reference_align = 64) readonly buffer Node { mat4 transforms[]; };
-
 layout (location = 0) out vec3 outPosition;
 layout (location = 1) flat out uint outMaterialIndex;
 #if HAS_VARIADIC_OUT
@@ -50,9 +48,12 @@ layout (set = 1, binding = 0) readonly buffer PrimitiveBuffer {
     Primitive primitives[];
 };
 layout (set = 1, binding = 1, std430) readonly buffer NodeBuffer {
-    Node nodes[];
+    uint instancedTransformStartIndices[];
+};
+layout (set = 1, binding = 2) readonly buffer InstancedTransformBuffer {
+    mat4 instancedTransforms[];
 };
-layout (set = 1, binding = 2, std430) readonly buffer MaterialBuffer {
+layout (set = 1, binding = 3, std430) readonly buffer MaterialBuffer {
     Material materials[];
 };
 

diff --git a/shaders/unlit_primitive.frag b/shaders/unlit_primitive.frag
@@ -30,10 +30,10 @@ layout (location = 0) out vec4 outColor;
 layout (location = 1) out float outRevealage;
 #endif
 
-layout (set = 1, binding = 2, std430) readonly buffer MaterialBuffer {
+layout (set = 1, binding = 3, std430) readonly buffer MaterialBuffer {
     Material materials[];
 };
-layout (set = 1, binding = 3) uniform sampler2D textures[];
+layout (set = 1, binding = 4) uniform sampler2D textures[];
 
 #if ALPHA_MODE == 0 || ALPHA_MODE == 2
 layout (early_fragment_tests) in;

diff --git a/shaders/unlit_primitive.vert b/shaders/unlit_primitive.vert
@@ -18,8 +18,6 @@ layout (constant_id = 0) const uint TEXCOORD_COMPONENT_TYPE = 5126; // FLOAT
 layout (constant_id = 1) const uint COLOR_COMPONENT_COUNT = 0;
 layout (constant_id = 2) const uint COLOR_COMPONENT_TYPE = 5126; // FLOAT
 
-layout (std430, buffer_reference, buffer_reference_align = 64) readonly buffer Node { mat4 transforms[]; };
-
 layout (location = 0) flat out uint outMaterialIndex;
 #if HAS_VARIADIC_OUT
 layout (location = 1) out VS_VARIADIC_OUT {
@@ -37,9 +35,12 @@ layout (set = 1, binding = 0) readonly buffer PrimitiveBuffer {
     Primitive primitives[];
 };
 layout (set = 1, binding = 1, std430) readonly buffer NodeBuffer {
-    Node nodes[];
+    uint instancedTransformStartIndices[];
+};
+layout (set = 1, binding = 2) readonly buffer InstancedTransformBuffer {
+    mat4 instancedTransforms[];
 };
-layout (set = 1, binding = 2, std430) readonly buffer MaterialBuffer {
+layout (set = 1, binding = 3, std430) readonly buffer MaterialBuffer {
     Material materials[];
 };