WIP: fix(dx12): align tex. <-> buf. copies via intermediate buffer if !UnrestrictedBufferTextureCopyPitchSupported

TODO: resolve `TODO` comments

Author: ErichDonGubler

CHANGELOG.md

@@ -59,6 +59,12 @@ By @Vecvec in [#7913](https://github.com/gfx-rs/wgpu/pull/7913).
 
 Naga now requires that no type be larger than 1 GB. This limit may be lowered in the future; feedback on an appropriate value for the limit is welcome. By @andyleiserson in [#7950](https://github.com/gfx-rs/wgpu/pull/7950).
 
+### Bug Fixes
+
+#### DX12
+
+- Align copies b/w textures and buffers when `D3D12_FEATURE_DATA_D3D12_OPTIONS13.UnrestrictedBufferTextureCopyPitchSupported` is `false`. By @ErichDonGubler in [#7721](https://github.com/gfx-rs/wgpu/pull/7721).
+
 ## v26.0.1 (2025-07-10)
 
 ### Bug Fixes

tests/tests/wgpu-gpu/regression/issue_6827.rs

@@ -19,17 +19,7 @@ static TEST_SCATTER: GpuTestConfiguration = GpuTestConfiguration::new()
             .expect_fail(FailureCase::backend_adapter(
                 wgpu::Backends::METAL,
                 "Apple Paravirtual device", // CI on M1
-            ))
-            .expect_fail(
-                // Unfortunately this depends on if `D3D12_FEATURE_DATA_D3D12_OPTIONS13.UnrestrictedBufferTextureCopyPitchSupported`
-                // is true, which we have no way to encode. This reproduces in CI though, so not too worried about it.
-                FailureCase::backend(wgpu::Backends::DX12)
-                    .flaky()
-                    .validation_error(
-                        "D3D12_PLACED_SUBRESOURCE_FOOTPRINT::Offset must be a multiple of 512",
-                    )
-                    .panic("GraphicsCommandList::close failed: The parameter is incorrect"),
-            ),
+            )),
     )
     .run_async(|ctx| async move { run_test(ctx, true).await });
 

wgpu-hal/src/dx12/adapter.rs

@@ -317,8 +317,7 @@ impl super::Adapter {
             suballocation_supported: !info.name.contains("Iris(R) Xe"),
             shader_model,
             max_sampler_descriptor_heap_size,
-            _unrestricted_buffer_texture_copy_pitch_supported:
-                unrestricted_buffer_texture_copy_pitch_supported,
+            unrestricted_buffer_texture_copy_pitch_supported,
         };
 
         // Theoretically vram limited, but in practice 2^20 is the limit
@@ -713,6 +712,7 @@ impl crate::Adapter for super::Adapter {
             queue: super::Queue {
                 raw: queue,
                 temp_lists: Mutex::new(Vec::new()),
+                intermediate_copy_bufs: Mutex::new(Vec::new()),
             },
         })
     }

wgpu-hal/src/dx12/command.rs

@@ -14,7 +14,7 @@ use crate::{
         dxgi::{name::ObjectExt, result::HResult as _},
     },
     dx12::borrow_interface_temporarily,
-    AccelerationStructureEntries,
+    AccelerationStructureEntries, CommandEncoder as _,
 };
 
 fn make_box(origin: &wgt::Origin3d, size: &crate::CopyExtent) -> Direct3D12::D3D12_BOX {
@@ -312,6 +312,78 @@ impl super::CommandEncoder {
             }
         }
     }
+
+    unsafe fn buf_tex_intermediate<T>(
+        &mut self,
+        region: crate::BufferTextureCopy,
+        tex_fmt: wgt::TextureFormat,
+        copy_op: impl FnOnce(&mut Self, &super::Buffer, wgt::BufferSize, crate::BufferTextureCopy) -> T,
+    ) -> Option<(T, super::Buffer)> {
+        let size = {
+            let copy_info = region.buffer_layout.get_buffer_texture_copy_info(
+                tex_fmt,
+                region.texture_base.aspect.map(),
+                &region.size.into(),
+            );
+            copy_info.unwrap().bytes_in_copy
+        };
+
+        let size = wgt::BufferSize::new(size)?;
+
+        let buffer = {
+            let (resource, allocation) =
+                super::suballocation::DeviceAllocationContext::from(&*self)
+                    .create_buffer(&crate::BufferDescriptor {
+                        label: None,
+                        size: size.get(),
+                        usage: wgt::BufferUses::COPY_SRC | wgt::BufferUses::COPY_DST,
+                        memory_flags: crate::MemoryFlags::empty(),
+                    })
+                    .expect(concat!(
+                        "internal error: ",
+                        "failed to allocate intermediate buffer ",
+                        "for offset alignment"
+                    ));
+            super::Buffer {
+                resource,
+                size: size.get(),
+                allocation,
+            }
+        };
+
+        let mut region = region;
+        region.buffer_layout.offset = 0;
+
+        unsafe {
+            self.transition_buffers(
+                [crate::BufferBarrier {
+                    buffer: &buffer,
+                    usage: crate::StateTransition {
+                        from: wgt::BufferUses::empty(),
+                        to: wgt::BufferUses::COPY_DST,
+                    },
+                }]
+                .into_iter(),
+            )
+        };
+
+        let t = copy_op(self, &buffer, size, region);
+
+        unsafe {
+            self.transition_buffers(
+                [crate::BufferBarrier {
+                    buffer: &buffer,
+                    usage: crate::StateTransition {
+                        from: wgt::BufferUses::COPY_DST,
+                        to: wgt::BufferUses::COPY_SRC,
+                    },
+                }]
+                .into_iter(),
+            )
+        };
+
+        Some((t, buffer))
+    }
 }
 
 impl crate::CommandEncoder for super::CommandEncoder {
@@ -363,7 +435,10 @@ impl crate::CommandEncoder for super::CommandEncoder {
     unsafe fn end_encoding(&mut self) -> Result<super::CommandBuffer, crate::DeviceError> {
         let raw = self.list.take().unwrap();
         unsafe { raw.Close() }.into_device_result("GraphicsCommandList::close")?;
-        Ok(super::CommandBuffer { raw })
+        Ok(super::CommandBuffer {
+            raw,
+            intermediate_copy_bufs: mem::take(&mut self.intermediate_copy_bufs).into(),
+        })
     }
     unsafe fn reset_all<I: Iterator<Item = super::CommandBuffer>>(&mut self, command_buffers: I) {
         for cmd_buf in command_buffers {
@@ -612,7 +687,32 @@ impl crate::CommandEncoder for super::CommandEncoder {
     ) where
         T: Iterator<Item = crate::BufferTextureCopy>,
     {
+        let offset_alignment = self.shared.private_caps.texture_data_placement_alignment();
+
         for r in regions {
+            let is_offset_aligned = r.buffer_layout.offset % offset_alignment == 0;
+            let (r, src) = if is_offset_aligned {
+                (r, src)
+            } else {
+                let Some((r, src)) = (unsafe {
+                    let src_offset = r.buffer_layout.offset;
+                    self.buf_tex_intermediate(r, dst.format, |this, buf, size, r| {
+                        let layout = crate::BufferCopy {
+                            src_offset,
+                            dst_offset: 0,
+                            size,
+                        };
+                        this.copy_buffer_to_buffer(src, buf, [layout].into_iter());
+                        r
+                    })
+                }) else {
+                    continue;
+                };
+                self.intermediate_copy_bufs.push(src);
+                let src = self.intermediate_copy_bufs.last().unwrap();
+                (r, src)
+            };
+
             let list = self.list.as_ref().unwrap();
 
             let src_location = Direct3D12::D3D12_TEXTURE_COPY_LOCATION {
@@ -680,8 +780,33 @@ impl crate::CommandEncoder for super::CommandEncoder {
             };
         };
 
+        let offset_alignment = self.shared.private_caps.texture_data_placement_alignment();
+
         for r in regions {
-            copy_aligned(this, src, dst, r);
+            let is_offset_aligned = r.buffer_layout.offset % offset_alignment == 0;
+            if is_offset_aligned {
+                copy_aligned(self, src, dst, r)
+            } else {
+                let orig_offset = r.buffer_layout.offset;
+                let Some((r, src)) = (unsafe {
+                    self.buf_tex_intermediate(r, src.format, |this, buf, size, r| {
+                        copy_aligned(this, src, buf, r);
+                        crate::BufferCopy {
+                            src_offset: orig_offset,
+                            dst_offset: 0,
+                            size,
+                        }
+                    })
+                }) else {
+                    continue;
+                };
+
+                unsafe {
+                    self.copy_buffer_to_buffer(&src, dst, [r].into_iter());
+                }
+
+                self.intermediate_copy_bufs.push(src);
+            };
         }
     }
 

wgpu-hal/src/dx12/device.rs

@@ -778,6 +778,7 @@ impl crate::Device for super::Device {
             mem_allocator: self.mem_allocator.clone(),
             rtv_pool: Arc::clone(&self.rtv_pool),
             temp_rtv_handles: Vec::new(),
+            intermediate_copy_bufs: Vec::new(),
             null_rtv_handle: self.null_rtv_handle,
             list: None,
             free_lists: Vec::new(),

wgpu-hal/src/dx12/mod.rs

@@ -95,7 +95,11 @@ use windows::{
     core::{Free, Interface},
     Win32::{
         Foundation,
-        Graphics::{Direct3D, Direct3D12, DirectComposition, Dxgi},
+        Graphics::{
+            Direct3D,
+            Direct3D12::{self, D3D12_TEXTURE_DATA_PLACEMENT_ALIGNMENT},
+            DirectComposition, Dxgi,
+        },
         System::Threading,
     },
 };
@@ -581,7 +585,17 @@ struct PrivateCapabilities {
     suballocation_supported: bool,
     shader_model: naga::back::hlsl::ShaderModel,
     max_sampler_descriptor_heap_size: u32,
-    _unrestricted_buffer_texture_copy_pitch_supported: bool,
+    unrestricted_buffer_texture_copy_pitch_supported: bool,
+}
+
+impl PrivateCapabilities {
+    fn texture_data_placement_alignment(&self) -> u64 {
+        if self.unrestricted_buffer_texture_copy_pitch_supported {
+            D3D12_TEXTURE_DATA_PLACEMENT_ALIGNMENT.into()
+        } else {
+            4
+        }
+    }
 }
 
 #[derive(Default)]
@@ -695,6 +709,7 @@ unsafe impl Sync for Device {}
 pub struct Queue {
     raw: Direct3D12::ID3D12CommandQueue,
     temp_lists: Mutex<Vec<Option<Direct3D12::ID3D12CommandList>>>,
+    intermediate_copy_bufs: Mutex<Vec<Arc<Vec<Buffer>>>>,
 }
 
 impl Queue {
@@ -817,6 +832,8 @@ pub struct CommandEncoder {
     rtv_pool: Arc<Mutex<descriptor::CpuPool>>,
     temp_rtv_handles: Vec<descriptor::Handle>,
 
+    intermediate_copy_bufs: Vec<Buffer>,
+
     null_rtv_handle: descriptor::Handle,
     list: Option<Direct3D12::ID3D12GraphicsCommandList>,
     free_lists: Vec<Direct3D12::ID3D12GraphicsCommandList>,
@@ -845,6 +862,7 @@ impl fmt::Debug for CommandEncoder {
 #[derive(Debug)]
 pub struct CommandBuffer {
     raw: Direct3D12::ID3D12GraphicsCommandList,
+    intermediate_copy_bufs: Arc<Vec<Buffer>>,
 }
 
 impl crate::DynCommandBuffer for CommandBuffer {}
@@ -1454,8 +1472,11 @@ impl crate::Queue for Queue {
     ) -> Result<(), crate::DeviceError> {
         let mut temp_lists = self.temp_lists.lock();
         temp_lists.clear();
+        let mut intermediate_copy_bufs = self.intermediate_copy_bufs.lock();
         for cmd_buf in command_buffers {
             temp_lists.push(Some(cmd_buf.raw.clone().into()));
+            intermediate_copy_bufs.push(Arc::clone(&cmd_buf.intermediate_copy_bufs));
+            // TODO: When to clean accumulated copy buffers up?
         }
 
         {