Deploying to master from @ gfx-rs/wgpu@6f5014f 🚀

Author: nical

doc/search-index.js

@@ -1293,14 +1293,17 @@
 <a href="#1292" id="1292">1292</a>
 <a href="#1293" id="1293">1293</a>
 <a href="#1294" id="1294">1294</a>
+<a href="#1295" id="1295">1295</a>
+<a href="#1296" id="1296">1296</a>
+<a href="#1297" id="1297">1297</a>
 </pre></div><pre class="rust"><code><span class="kw">use crate</span>::{
     device::{queue::TempResource, Device},
     global::Global,
     hub::Hub,
     id::CommandEncoderId,
     init_tracker::MemoryInitKind,
     ray_tracing::{
-        tlas_instance_into_bytes, BlasAction, BlasBuildEntry, BlasGeometries, BlasTriangleGeometry,
+        BlasAction, BlasBuildEntry, BlasGeometries, BlasTriangleGeometry,
         BuildAccelerationStructureError, TlasAction, TlasBuildEntry, TlasInstance, TlasPackage,
         TraceBlasBuildEntry, TraceBlasGeometries, TraceBlasTriangleGeometry, TraceTlasInstance,
         TraceTlasPackage, ValidateBlasActionsError, ValidateTlasActionsError,
@@ -1355,9 +1358,6 @@
     entry: TlasBuildEntry,
 }
 
-<span class="comment">// TODO: Get this from the device (e.g. VkPhysicalDeviceAccelerationStructurePropertiesKHR.minAccelerationStructureScratchOffsetAlignment) this is currently the largest possible some devices have 0, 64, 128 (lower limits) so this could create excess allocation (Note: dx12 has 256).
-</span><span class="kw">const </span>SCRATCH_BUFFER_ALIGNMENT: u32 = <span class="number">256</span>;
-
 <span class="kw">impl </span>Global {
     <span class="comment">// Currently this function is very similar to its safe counterpart, however certain parts of it are very different,
     // making for the two to be implemented differently, the main difference is this function has separate buffers for each
@@ -1488,6 +1488,7 @@
             <span class="kw-2">&amp;mut </span>scratch_buffer_blas_size,
             <span class="kw-2">&amp;mut </span>blas_storage,
             hub,
+            device.alignments.ray_tracing_scratch_buffer_alignment,
         )<span class="question-mark">?</span>;
 
         <span class="kw">let </span><span class="kw-2">mut </span>scratch_buffer_tlas_size = <span class="number">0</span>;
@@ -1555,7 +1556,7 @@
             <span class="kw">let </span>scratch_buffer_offset = scratch_buffer_tlas_size;
             scratch_buffer_tlas_size += align_to(
                 tlas.size_info.build_scratch_size <span class="kw">as </span>u32,
-                SCRATCH_BUFFER_ALIGNMENT,
+                device.alignments.ray_tracing_scratch_buffer_alignment,
             ) <span class="kw">as </span>u64;
 
             tlas_storage.push(UnsafeTlasStore {
@@ -1803,6 +1804,7 @@
             <span class="kw-2">&amp;mut </span>scratch_buffer_blas_size,
             <span class="kw-2">&amp;mut </span>blas_storage,
             hub,
+            device.alignments.ray_tracing_scratch_buffer_alignment,
         )<span class="question-mark">?</span>;
         <span class="kw">let </span><span class="kw-2">mut </span>tlas_lock_store = Vec::&lt;(<span class="prelude-ty">Option</span>&lt;TlasPackage&gt;, Arc&lt;Tlas&gt;)&gt;::new();
 
@@ -1830,7 +1832,7 @@
             <span class="kw">let </span>scratch_buffer_offset = scratch_buffer_tlas_size;
             scratch_buffer_tlas_size += align_to(
                 tlas.size_info.build_scratch_size <span class="kw">as </span>u32,
-                SCRATCH_BUFFER_ALIGNMENT,
+                device.alignments.ray_tracing_scratch_buffer_alignment,
             ) <span class="kw">as </span>u64;
 
             <span class="kw">let </span>first_byte_index = instance_buffer_staging_source.len();
@@ -1853,10 +1855,13 @@
 
                 cmd_buf_data.trackers.blas_s.set_single(blas.clone());
 
-                instance_buffer_staging_source.extend(tlas_instance_into_bytes(
-                    <span class="kw-2">&amp;</span>instance,
-                    blas.handle,
-                    device.backend(),
+                instance_buffer_staging_source.extend(device.raw().tlas_instance_to_bytes(
+                    hal::TlasInstance {
+                        transform: <span class="kw-2">*</span>instance.transform,
+                        custom_index: instance.custom_index,
+                        mask: instance.mask,
+                        blas_address: blas.handle,
+                    },
                 ));
 
                 instance_count += <span class="number">1</span>;
@@ -2308,6 +2313,7 @@
     scratch_buffer_blas_size: <span class="kw-2">&amp;mut </span>u64,
     blas_storage: <span class="kw-2">&amp;mut </span>Vec&lt;BlasStore&lt;<span class="lifetime">'a</span>&gt;&gt;,
     hub: <span class="kw-2">&amp;</span>Hub,
+    ray_tracing_scratch_buffer_alignment: u32,
 ) -&gt; <span class="prelude-ty">Result</span>&lt;(), BuildAccelerationStructureError&gt; {
     <span class="kw">let </span><span class="kw-2">mut </span>triangle_entries =
         Vec::&lt;hal::AccelerationStructureTriangles&lt;<span class="kw">dyn </span>hal::DynBuffer&gt;&gt;::new();
@@ -2487,7 +2493,7 @@
             <span class="kw">let </span>scratch_buffer_offset = <span class="kw-2">*</span>scratch_buffer_blas_size;
             <span class="kw-2">*</span>scratch_buffer_blas_size += align_to(
                 blas.size_info.build_scratch_size <span class="kw">as </span>u32,
-                SCRATCH_BUFFER_ALIGNMENT,
+                ray_tracing_scratch_buffer_alignment,
             ) <span class="kw">as </span>u64;
 
             blas_storage.push(BlasStore {

doc/src/wgpu_core/command/ray_tracing.rs.html

@@ -368,7 +368,7 @@
     global::Global,
     id::{<span class="self">self</span>, BlasId, TlasId},
     lock::RwLock,
-    ray_tracing::{get_raw_tlas_instance_size, CreateBlasError, CreateTlasError},
+    ray_tracing::{CreateBlasError, CreateTlasError},
     resource, LabelHelpers,
 };
 <span class="kw">use </span>hal::AccelerationStructureTriangleIndices;
@@ -491,7 +491,7 @@
         .map_err(DeviceError::from_hal)<span class="question-mark">?</span>;
 
         <span class="kw">let </span>instance_buffer_size =
-            get_raw_tlas_instance_size(<span class="self">self</span>.backend()) * desc.max_instances.max(<span class="number">1</span>) <span class="kw">as </span>usize;
+            <span class="self">self</span>.alignments.raw_tlas_instance_size * desc.max_instances.max(<span class="number">1</span>) <span class="kw">as </span>usize;
         <span class="kw">let </span>instance_buffer = <span class="kw">unsafe </span>{
             <span class="self">self</span>.raw().create_buffer(<span class="kw-2">&amp;</span>hal::BufferDescriptor {
                 label: <span class="prelude-val">Some</span>(<span class="string">"(wgpu-core) instances_buffer"</span>),

doc/src/wgpu_core/device/ray_tracing.rs.html

@@ -277,51 +277,6 @@
 <a href="#276" id="276">276</a>
 <a href="#277" id="277">277</a>
 <a href="#278" id="278">278</a>
-<a href="#279" id="279">279</a>
-<a href="#280" id="280">280</a>
-<a href="#281" id="281">281</a>
-<a href="#282" id="282">282</a>
-<a href="#283" id="283">283</a>
-<a href="#284" id="284">284</a>
-<a href="#285" id="285">285</a>
-<a href="#286" id="286">286</a>
-<a href="#287" id="287">287</a>
-<a href="#288" id="288">288</a>
-<a href="#289" id="289">289</a>
-<a href="#290" id="290">290</a>
-<a href="#291" id="291">291</a>
-<a href="#292" id="292">292</a>
-<a href="#293" id="293">293</a>
-<a href="#294" id="294">294</a>
-<a href="#295" id="295">295</a>
-<a href="#296" id="296">296</a>
-<a href="#297" id="297">297</a>
-<a href="#298" id="298">298</a>
-<a href="#299" id="299">299</a>
-<a href="#300" id="300">300</a>
-<a href="#301" id="301">301</a>
-<a href="#302" id="302">302</a>
-<a href="#303" id="303">303</a>
-<a href="#304" id="304">304</a>
-<a href="#305" id="305">305</a>
-<a href="#306" id="306">306</a>
-<a href="#307" id="307">307</a>
-<a href="#308" id="308">308</a>
-<a href="#309" id="309">309</a>
-<a href="#310" id="310">310</a>
-<a href="#311" id="311">311</a>
-<a href="#312" id="312">312</a>
-<a href="#313" id="313">313</a>
-<a href="#314" id="314">314</a>
-<a href="#315" id="315">315</a>
-<a href="#316" id="316">316</a>
-<a href="#317" id="317">317</a>
-<a href="#318" id="318">318</a>
-<a href="#319" id="319">319</a>
-<a href="#320" id="320">320</a>
-<a href="#321" id="321">321</a>
-<a href="#322" id="322">322</a>
-<a href="#323" id="323">323</a>
 </pre></div><pre class="rust"><code><span class="comment">// Ray tracing
 // Major missing optimizations (no api surface changes needed):
 // - use custom tracker to track build state
@@ -337,8 +292,8 @@
     id::{BlasId, BufferId, TlasId},
     resource::CreateBufferError,
 };
-<span class="kw">use </span>std::{mem::size_of, sync::Arc};
-<span class="kw">use </span>std::{num::NonZeroU64, slice};
+<span class="kw">use </span>std::num::NonZeroU64;
+<span class="kw">use </span>std::sync::Arc;
 
 <span class="kw">use </span><span class="kw">crate</span>::resource::{Blas, ResourceErrorIdent, Tlas};
 <span class="kw">use </span>thiserror::Error;
@@ -600,49 +555,4 @@
     <span class="kw">pub </span>instances: Vec&lt;<span class="prelude-ty">Option</span>&lt;TraceTlasInstance&gt;&gt;,
     <span class="kw">pub </span>lowest_unmodified: u32,
 }
-
-<span class="kw">pub</span>(<span class="kw">crate</span>) <span class="kw">fn </span>get_raw_tlas_instance_size(backend: wgt::Backend) -&gt; usize {
-    <span class="comment">// TODO: this should be provided by the backend
-    </span><span class="kw">match </span>backend {
-        wgt::Backend::Empty =&gt; <span class="number">0</span>,
-        wgt::Backend::Vulkan =&gt; <span class="number">64</span>,
-        <span class="kw">_ </span>=&gt; <span class="macro">unimplemented!</span>(),
-    }
-}
-
-<span class="attr">#[derive(Clone)]
-#[repr(C)]
-</span><span class="kw">struct </span>RawTlasInstance {
-    transform: [f32; <span class="number">12</span>],
-    custom_index_and_mask: u32,
-    shader_binding_table_record_offset_and_flags: u32,
-    acceleration_structure_reference: u64,
-}
-
-<span class="kw">pub</span>(<span class="kw">crate</span>) <span class="kw">fn </span>tlas_instance_into_bytes(
-    instance: <span class="kw-2">&amp;</span>TlasInstance,
-    blas_address: u64,
-    backend: wgt::Backend,
-) -&gt; Vec&lt;u8&gt; {
-    <span class="comment">// TODO: get the device to do this
-    </span><span class="kw">match </span>backend {
-        wgt::Backend::Empty =&gt; <span class="macro">vec!</span>[],
-        wgt::Backend::Vulkan =&gt; {
-            <span class="kw">const </span>MAX_U24: u32 = (<span class="number">1u32 </span>&lt;&lt; <span class="number">24u32</span>) - <span class="number">1u32</span>;
-            <span class="kw">let </span>temp = RawTlasInstance {
-                transform: <span class="kw-2">*</span>instance.transform,
-                custom_index_and_mask: (instance.custom_index &amp; MAX_U24)
-                    | (u32::from(instance.mask) &lt;&lt; <span class="number">24</span>),
-                shader_binding_table_record_offset_and_flags: <span class="number">0</span>,
-                acceleration_structure_reference: blas_address,
-            };
-            <span class="kw">let </span>temp: <span class="kw-2">*const </span><span class="kw">_ </span>= <span class="kw-2">&amp;</span>temp;
-            <span class="kw">unsafe </span>{
-                slice::from_raw_parts::&lt;u8&gt;(temp.cast::&lt;u8&gt;(), size_of::&lt;RawTlasInstance&gt;())
-                    .to_vec()
-            }
-        }
-        <span class="kw">_ </span>=&gt; <span class="macro">unimplemented!</span>(),
-    }
-}
 </code></pre></div></section></main></body></html>

doc/src/wgpu_core/ray_tracing.rs.html

@@ -529,14 +529,19 @@
 <a href="#528" id="528">528</a>
 <a href="#529" id="529">529</a>
 <a href="#530" id="530">530</a>
+<a href="#531" id="531">531</a>
+<a href="#532" id="532">532</a>
+<a href="#533" id="533">533</a>
+<a href="#534" id="534">534</a>
+<a href="#535" id="535">535</a>
 </pre></div><pre class="rust"><code><span class="kw">use crate</span>::{
     AccelerationStructureBuildSizes, AccelerationStructureDescriptor, Api, BindGroupDescriptor,
     BindGroupLayoutDescriptor, BufferDescriptor, BufferMapping, CommandEncoderDescriptor,
     ComputePipelineDescriptor, Device, DeviceError, FenceValue,
     GetAccelerationStructureBuildSizesDescriptor, Label, MemoryRange, PipelineCacheDescriptor,
     PipelineCacheError, PipelineError, PipelineLayoutDescriptor, RenderPipelineDescriptor,
     SamplerDescriptor, ShaderError, ShaderInput, ShaderModuleDescriptor, TextureDescriptor,
-    TextureViewDescriptor,
+    TextureViewDescriptor, TlasInstance,
 };
 
 <span class="kw">use super</span>::{
@@ -689,6 +694,7 @@
         <span class="kw-2">&amp;</span><span class="self">self</span>,
         acceleration_structure: Box&lt;<span class="kw">dyn </span>DynAccelerationStructure&gt;,
     );
+    <span class="kw">fn </span>tlas_instance_to_bytes(<span class="kw-2">&amp;</span><span class="self">self</span>, instance: TlasInstance) -&gt; Vec&lt;u8&gt;;
 
     <span class="kw">fn </span>get_internal_counters(<span class="kw-2">&amp;</span><span class="self">self</span>) -&gt; wgt::HalCounters;
     <span class="kw">fn </span>generate_allocator_report(<span class="kw-2">&amp;</span><span class="self">self</span>) -&gt; <span class="prelude-ty">Option</span>&lt;wgt::AllocatorReport&gt;;
@@ -1051,6 +1057,10 @@
         <span class="kw">unsafe </span>{ D::destroy_acceleration_structure(<span class="self">self</span>, acceleration_structure.unbox()) }
     }
 
+    <span class="kw">fn </span>tlas_instance_to_bytes(<span class="kw-2">&amp;</span><span class="self">self</span>, instance: TlasInstance) -&gt; Vec&lt;u8&gt; {
+        D::tlas_instance_to_bytes(<span class="self">self</span>, instance)
+    }
+
     <span class="kw">fn </span>get_internal_counters(<span class="kw-2">&amp;</span><span class="self">self</span>) -&gt; wgt::HalCounters {
         D::get_internal_counters(<span class="self">self</span>)
     }

doc/src/wgpu_hal/dynamic/device.rs.html

@@ -508,9 +508,15 @@
 <a href="#507" id="507">507</a>
 <a href="#508" id="508">508</a>
 <a href="#509" id="509">509</a>
+<a href="#510" id="510">510</a>
+<a href="#511" id="511">511</a>
+<a href="#512" id="512">512</a>
+<a href="#513" id="513">513</a>
+<a href="#514" id="514">514</a>
 </pre></div><pre class="rust"><code><span class="attr">#![allow(unused_variables)]
 
-</span><span class="kw">use </span>std::ops::Range;
+</span><span class="kw">use </span><span class="kw">crate</span>::TlasInstance;
+<span class="kw">use </span>std::ops::Range;
 
 <span class="attr">#[derive(Clone, Debug)]
 </span><span class="kw">pub struct </span>Api;
@@ -816,6 +822,10 @@
     }
     <span class="kw">unsafe fn </span>destroy_acceleration_structure(<span class="kw-2">&amp;</span><span class="self">self</span>, _acceleration_structure: Resource) {}
 
+    <span class="kw">fn </span>tlas_instance_to_bytes(<span class="kw-2">&amp;</span><span class="self">self</span>, instance: TlasInstance) -&gt; Vec&lt;u8&gt; {
+        <span class="macro">vec!</span>[]
+    }
+
     <span class="kw">fn </span>get_internal_counters(<span class="kw-2">&amp;</span><span class="self">self</span>) -&gt; wgt::HalCounters {
         Default::default()
     }

doc/src/wgpu_hal/empty.rs.html

@@ -1279,6 +1279,8 @@
 <a href="#1278" id="1278">1278</a>
 <a href="#1279" id="1279">1279</a>
 <a href="#1280" id="1280">1280</a>
+<a href="#1281" id="1281">1281</a>
+<a href="#1282" id="1282">1282</a>
 </pre></div><pre class="rust"><code><span class="kw">use </span>glow::HasContext;
 <span class="kw">use </span>parking_lot::Mutex;
 <span class="kw">use </span>std::sync::{atomic::AtomicU8, Arc};
@@ -2132,6 +2134,8 @@
                     // being, provide 1 as the value here, to cause as little
                     // trouble as possible.
                     </span>uniform_bounds_check_alignment: wgt::BufferSize::new(<span class="number">1</span>).unwrap(),
+                    raw_tlas_instance_size: <span class="number">0</span>,
+                    ray_tracing_scratch_buffer_alignment: <span class="number">0</span>,
                 },
             },
         })

doc/src/wgpu_hal/gles/adapter.rs.html

@@ -1643,6 +1643,10 @@
 <a href="#1642" id="1642">1642</a>
 <a href="#1643" id="1643">1643</a>
 <a href="#1644" id="1644">1644</a>
+<a href="#1645" id="1645">1645</a>
+<a href="#1646" id="1646">1646</a>
+<a href="#1647" id="1647">1647</a>
+<a href="#1648" id="1648">1648</a>
 </pre></div><pre class="rust"><code><span class="kw">use super</span>::{conv, PrivateCapabilities};
 <span class="kw">use </span><span class="kw">crate</span>::auxil::map_naga_stage;
 <span class="kw">use </span>glow::HasContext;
@@ -1653,7 +1657,7 @@
     sync::{Arc, Mutex},
 };
 
-<span class="kw">use </span><span class="kw">crate</span>::AtomicFenceValue;
+<span class="kw">use crate</span>::{AtomicFenceValue, TlasInstance};
 <span class="kw">use </span>arrayvec::ArrayVec;
 <span class="kw">use </span>std::sync::atomic::Ordering;
 
@@ -3278,6 +3282,10 @@
     ) {
     }
 
+    <span class="kw">fn </span>tlas_instance_to_bytes(<span class="kw-2">&amp;</span><span class="self">self</span>, _instance: TlasInstance) -&gt; Vec&lt;u8&gt; {
+        <span class="macro">unimplemented!</span>()
+    }
+
     <span class="kw">fn </span>get_internal_counters(<span class="kw-2">&amp;</span><span class="self">self</span>) -&gt; wgt::HalCounters {
         <span class="self">self</span>.counters.clone()
     }