hal: switch texture copies to expect physical sizes

wgpu-core/src/command/transfer.rs

@@ -339,21 +339,12 @@ pub(crate) fn validate_texture_copy_range(
         wgt::TextureDimension::D1 | wgt::TextureDimension::D2 => {
             (1, copy_size.depth_or_array_layers)
         }
-        wgt::TextureDimension::D3 => (
-            copy_size
-                .depth_or_array_layers
-                .min(extent_virtual.depth_or_array_layers),
-            1,
-        ),
+        wgt::TextureDimension::D3 => (copy_size.depth_or_array_layers, 1),
     };
 
-    // WebGPU uses the physical size of the texture for copies whereas vulkan uses
-    // the virtual size. We have passed validation, so it's safe to use the
-    // image extent data directly. We want the provided copy size to be no larger than
-    // the virtual size.
     let copy_extent = hal::CopyExtent {
-        width: copy_size.width.min(extent_virtual.width),
-        height: copy_size.height.min(extent_virtual.height),
+        width: copy_size.width,
+        height: copy_size.height,
         depth,
     };
     Ok((copy_extent, array_layer_count))

wgpu-hal/src/auxil/mod.rs

@@ -19,3 +19,54 @@ pub fn map_naga_stage(stage: naga::ShaderStage) -> wgt::ShaderStages {
         naga::ShaderStage::Compute => wgt::ShaderStages::COMPUTE,
     }
 }
+
+impl crate::CopyExtent {
+    pub fn min(&self, other: &Self) -> Self {
+        Self {
+            width: self.width.min(other.width),
+            height: self.height.min(other.height),
+            depth: self.depth.min(other.depth),
+        }
+    }
+
+    // Get the copy size at a specific mipmap level. This doesn't make most sense,
+    // since the copy extents are provided *for* a mipmap level to start with.
+    // But backends use `CopyExtent` more sparingly, and this piece is shared.
+    pub fn at_mip_level(&self, level: u32) -> Self {
+        Self {
+            width: (self.width >> level).max(1),
+            height: (self.height >> level).max(1),
+            depth: (self.depth >> level).max(1),
+        }
+    }
+}
+
+impl crate::TextureCopyBase {
+    pub fn max_copy_size(&self, full_size: &crate::CopyExtent) -> crate::CopyExtent {
+        let mip = full_size.at_mip_level(self.mip_level);
+        crate::CopyExtent {
+            width: mip.width - self.origin.x,
+            height: mip.height - self.origin.y,
+            depth: mip.depth - self.origin.z,
+        }
+    }
+}
+
+impl crate::BufferTextureCopy {
+    pub fn clamp_size_to_virtual(&mut self, full_size: &crate::CopyExtent) {
+        let max_size = self.texture_base.max_copy_size(full_size);
+        self.size = self.size.min(&max_size);
+    }
+}
+
+impl crate::TextureCopy {
+    pub fn clamp_size_to_virtual(
+        &mut self,
+        full_src_size: &crate::CopyExtent,
+        full_dst_size: &crate::CopyExtent,
+    ) {
+        let max_src_size = self.src_base.max_copy_size(full_src_size);
+        let max_dst_size = self.dst_base.max_copy_size(full_dst_size);
+        self.size = self.size.min(&max_src_size).min(&max_dst_size);
+    }
+}

wgpu-hal/src/dx12/command.rs

@@ -13,18 +13,6 @@ fn make_box(origin: &wgt::Origin3d, size: &crate::CopyExtent) -> d3d12::D3D12_BO
     }
 }
 
-fn upround_extent(size: crate::CopyExtent, block_size: u32) -> crate::CopyExtent {
-    debug_assert!(block_size.is_power_of_two());
-
-    let block_mask = block_size - 1;
-
-    crate::CopyExtent {
-        width: (size.width + block_mask) & !(block_mask),
-        height: (size.height + block_mask) & !(block_mask),
-        depth: size.depth,
-    }
-}
-
 impl super::Temp {
     fn prepare_marker(&mut self, marker: &str) -> (&[u16], u32) {
         self.marker.clear();
@@ -514,11 +502,8 @@ impl crate::CommandEncoder<super::Api> for super::CommandEncoder {
             u: mem::zeroed(),
         };
 
-        let block_size = src.format.describe().block_dimensions.0 as u32;
         for r in regions {
-            let uprounded_size = upround_extent(r.size, block_size);
-
-            let src_box = make_box(&r.src_base.origin, &uprounded_size);
+            let src_box = make_box(&r.src_base.origin, &r.size);
             *src_location.u.SubresourceIndex_mut() = src.calc_subresource_for_copy(&r.src_base);
             *dst_location.u.SubresourceIndex_mut() = dst.calc_subresource_for_copy(&r.dst_base);
 
@@ -556,19 +541,17 @@ impl crate::CommandEncoder<super::Api> for super::CommandEncoder {
 
         let block_size = dst.format.describe().block_dimensions.0 as u32;
         for r in regions {
-            let uprounded_size = upround_extent(r.size, block_size);
-
-            let src_box = make_box(&wgt::Origin3d::ZERO, &uprounded_size);
+            let src_box = make_box(&wgt::Origin3d::ZERO, &r.size);
             *src_location.u.PlacedFootprint_mut() = d3d12::D3D12_PLACED_SUBRESOURCE_FOOTPRINT {
                 Offset: r.buffer_layout.offset,
                 Footprint: d3d12::D3D12_SUBRESOURCE_FOOTPRINT {
                     Format: raw_format,
-                    Width: uprounded_size.width,
+                    Width: r.size.width,
                     Height: r
                         .buffer_layout
                         .rows_per_image
-                        .map_or(uprounded_size.height, |count| count.get() * block_size),
-                    Depth: uprounded_size.depth,
+                        .map_or(r.size.height, |count| count.get() * block_size),
+                    Depth: r.size.depth,
                     RowPitch: r.buffer_layout.bytes_per_row.map_or(0, |count| {
                         count.get().max(d3d12::D3D12_TEXTURE_DATA_PITCH_ALIGNMENT)
                     }),
@@ -611,20 +594,18 @@ impl crate::CommandEncoder<super::Api> for super::CommandEncoder {
 
         let block_size = src.format.describe().block_dimensions.0 as u32;
         for r in regions {
-            let uprounded_size = upround_extent(r.size, block_size);
-
-            let src_box = make_box(&r.texture_base.origin, &uprounded_size);
+            let src_box = make_box(&r.texture_base.origin, &r.size);
             *src_location.u.SubresourceIndex_mut() = src.calc_subresource_for_copy(&r.texture_base);
             *dst_location.u.PlacedFootprint_mut() = d3d12::D3D12_PLACED_SUBRESOURCE_FOOTPRINT {
                 Offset: r.buffer_layout.offset,
                 Footprint: d3d12::D3D12_SUBRESOURCE_FOOTPRINT {
                     Format: raw_format,
-                    Width: uprounded_size.width,
+                    Width: r.size.width,
                     Height: r
                         .buffer_layout
                         .rows_per_image
-                        .map_or(uprounded_size.height, |count| count.get() * block_size),
-                    Depth: uprounded_size.depth,
+                        .map_or(r.size.height, |count| count.get() * block_size),
+                    Depth: r.size.depth,
                     RowPitch: r.buffer_layout.bytes_per_row.map_or(0, |count| count.get()),
                 },
             };

wgpu-hal/src/gles/command.rs

@@ -290,7 +290,8 @@ impl crate::CommandEncoder<super::Api> for super::CommandEncoder {
     {
         let (src_raw, src_target) = src.inner.as_native();
         let (dst_raw, dst_target) = dst.inner.as_native();
-        for copy in regions {
+        for mut copy in regions {
+            copy.clamp_size_to_virtual(&src.copy_size, &dst.copy_size);
             self.cmd_buffer.commands.push(C::CopyTextureToTexture {
                 src: src_raw,
                 src_target,
@@ -310,7 +311,8 @@ impl crate::CommandEncoder<super::Api> for super::CommandEncoder {
         T: Iterator<Item = crate::BufferTextureCopy>,
     {
         let (dst_raw, dst_target) = dst.inner.as_native();
-        for copy in regions {
+        for mut copy in regions {
+            copy.clamp_size_to_virtual(&dst.copy_size);
             self.cmd_buffer.commands.push(C::CopyBufferToTexture {
                 src: src.raw,
                 src_target: src.target,
@@ -332,7 +334,8 @@ impl crate::CommandEncoder<super::Api> for super::CommandEncoder {
         T: Iterator<Item = crate::BufferTextureCopy>,
     {
         let (src_raw, src_target) = src.inner.as_native();
-        for copy in regions {
+        for mut copy in regions {
+            copy.clamp_size_to_virtual(&src.copy_size);
             self.cmd_buffer.commands.push(C::CopyTextureToBuffer {
                 src: src_raw,
                 src_target,

wgpu-hal/src/gles/device.rs

@@ -429,6 +429,12 @@ impl crate::Device<super::Api> for super::Device {
             | crate::TextureUses::DEPTH_STENCIL_READ;
         let format_desc = self.shared.describe_texture_format(desc.format);
 
+        let mut copy_size = crate::CopyExtent {
+            width: desc.size.width,
+            height: desc.size.height,
+            depth: 1,
+        };
+
         let inner = if render_usage.contains(desc.usage)
             && desc.dimension == wgt::TextureDimension::D2
             && desc.size.depth_or_array_layers == 1
@@ -516,6 +522,7 @@ impl crate::Device<super::Api> for super::Device {
                     }
                 }
                 wgt::TextureDimension::D3 => {
+                    copy_size.depth = desc.size.depth_or_array_layers;
                     let target = glow::TEXTURE_3D;
                     gl.bind_texture(target, Some(raw));
                     gl.tex_storage_3d(
@@ -562,6 +569,7 @@ impl crate::Device<super::Api> for super::Device {
             },
             format: desc.format,
             format_desc,
+            copy_size,
         })
     }
     unsafe fn destroy_texture(&self, texture: super::Texture) {

wgpu-hal/src/gles/egl.rs

@@ -984,6 +984,11 @@ impl crate::Surface<super::Api> for Surface {
             mip_level_count: 1,
             format: sc.format,
             format_desc: sc.format_desc.clone(),
+            copy_size: crate::CopyExtent {
+                width: sc.extent.width,
+                height: sc.extent.height,
+                depth: 1,
+            },
         };
         Ok(Some(crate::AcquiredSurfaceTexture {
             texture,

wgpu-hal/src/gles/mod.rs

@@ -230,6 +230,7 @@ pub struct Texture {
     array_layer_count: u32,
     format: wgt::TextureFormat,
     format_desc: TextureFormatDesc,
+    copy_size: crate::CopyExtent,
 }
 
 #[derive(Clone, Debug)]

wgpu-hal/src/lib.rs

@@ -360,6 +360,7 @@ pub trait CommandEncoder<A: Api>: Send + Sync {
     /// Copy from one texture to another.
     /// Works with a single array layer.
     /// Note: `dst` current usage has to be `TextureUses::COPY_DST`.
+    /// Note: the copy extent is in physical size (rounded to the block size)
     unsafe fn copy_texture_to_texture<T>(
         &mut self,
         src: &A::Texture,
@@ -372,12 +373,14 @@ pub trait CommandEncoder<A: Api>: Send + Sync {
     /// Copy from buffer to texture.
     /// Works with a single array layer.
     /// Note: `dst` current usage has to be `TextureUses::COPY_DST`.
+    /// Note: the copy extent is in physical size (rounded to the block size)
     unsafe fn copy_buffer_to_texture<T>(&mut self, src: &A::Buffer, dst: &A::Texture, regions: T)
     where
         T: Iterator<Item = BufferTextureCopy>;
 
     /// Copy from texture to buffer.
     /// Works with a single array layer.
+    /// Note: the copy extent is in physical size (rounded to the block size)
     unsafe fn copy_texture_to_buffer<T>(
         &mut self,
         src: &A::Texture,

wgpu-hal/src/metal/command.rs

@@ -153,26 +153,19 @@ impl crate::CommandEncoder<super::Api> for super::CommandEncoder {
 
         for mip_level in mip_range {
             // Note that Metal requires this only to be a multiple of the pixel size, not some other constant like in other APIs.
-            let mip_size = texture
-                .size
-                .mip_level_size(mip_level, texture.raw_type == mtl::MTLTextureType::D3);
-            let depth = if texture.raw_type == mtl::MTLTextureType::D3 {
-                mip_size.depth_or_array_layers as u64
-            } else {
-                1
-            };
+            let mip_size = texture.copy_size.at_mip_level(mip_level);
             let bytes_per_row = mip_size.width as u64 / format_desc.block_dimensions.0 as u64
                 * format_desc.block_size as u64;
             let max_rows_per_copy = super::ZERO_BUFFER_SIZE / bytes_per_row;
             // round down to a multiple of rows needed by the texture format
             let max_rows_per_copy = max_rows_per_copy / format_desc.block_dimensions.1 as u64
                 * format_desc.block_dimensions.1 as u64;
             assert!(max_rows_per_copy > 0, "Zero buffer size is too small to fill a single row of a texture of type {:?}, size {:?} and format {:?}",
-                        texture.raw_type, texture.size, texture.format);
+                        texture.raw_type, texture.copy_size, texture.format);
 
             for array_layer in array_range.clone() {
                 // 3D textures are quickly massive in memory size, so we don't bother trying to do more than one layer at once.
-                for z in 0..depth {
+                for z in 0..mip_size.depth as u64 {
                     // May need multiple copies for each subresource! We assume that we never need to split a row.
                     let mut num_rows_left = mip_size.height as u64;
                     while num_rows_left > 0 {
@@ -239,6 +232,7 @@ impl crate::CommandEncoder<super::Api> for super::CommandEncoder {
         for copy in regions {
             let src_origin = conv::map_origin(&copy.src_base.origin);
             let dst_origin = conv::map_origin(&copy.dst_base.origin);
+            // no clamping is done: Metal expects physical sizes here
             let extent = conv::map_copy_extent(&copy.size);
             encoder.copy_from_texture(
                 &src.raw,
@@ -265,7 +259,11 @@ impl crate::CommandEncoder<super::Api> for super::CommandEncoder {
         let encoder = self.enter_blit();
         for copy in regions {
             let dst_origin = conv::map_origin(&copy.texture_base.origin);
-            let extent = conv::map_copy_extent(&copy.size);
+            // Metal expects buffer-texture copies in virtual sizes
+            let extent = copy
+                .texture_base
+                .max_copy_size(&dst.copy_size)
+                .min(&copy.size);
             let bytes_per_row = copy
                 .buffer_layout
                 .bytes_per_row
@@ -279,7 +277,7 @@ impl crate::CommandEncoder<super::Api> for super::CommandEncoder {
                 copy.buffer_layout.offset,
                 bytes_per_row,
                 bytes_per_image,
-                extent,
+                conv::map_copy_extent(&extent),
                 &dst.raw,
                 copy.texture_base.array_layer as u64,
                 copy.texture_base.mip_level as u64,
@@ -301,7 +299,11 @@ impl crate::CommandEncoder<super::Api> for super::CommandEncoder {
         let encoder = self.enter_blit();
         for copy in regions {
             let src_origin = conv::map_origin(&copy.texture_base.origin);
-            let extent = conv::map_copy_extent(&copy.size);
+            // Metal expects texture-buffer copies in virtual sizes
+            let extent = copy
+                .texture_base
+                .max_copy_size(&src.copy_size)
+                .min(&copy.size);
             let bytes_per_row = copy
                 .buffer_layout
                 .bytes_per_row
@@ -315,7 +317,7 @@ impl crate::CommandEncoder<super::Api> for super::CommandEncoder {
                 copy.texture_base.array_layer as u64,
                 copy.texture_base.mip_level as u64,
                 src_origin,
-                extent,
+                conv::map_copy_extent(&extent),
                 &dst.raw,
                 copy.buffer_layout.offset,
                 bytes_per_row,

wgpu-hal/src/metal/device.rs

@@ -236,6 +236,11 @@ impl crate::Device<super::Api> for super::Device {
 
         let descriptor = mtl::TextureDescriptor::new();
         let mut array_layers = desc.size.depth_or_array_layers;
+        let mut copy_size = crate::CopyExtent {
+            width: desc.size.width,
+            height: desc.size.height,
+            depth: 1,
+        };
         let mtl_type = match desc.dimension {
             wgt::TextureDimension::D1 => {
                 if desc.size.depth_or_array_layers > 1 {
@@ -259,6 +264,7 @@ impl crate::Device<super::Api> for super::Device {
             wgt::TextureDimension::D3 => {
                 descriptor.set_depth(desc.size.depth_or_array_layers as u64);
                 array_layers = 1;
+                copy_size.depth = desc.size.depth_or_array_layers;
                 mtl::MTLTextureType::D3
             }
         };
@@ -283,7 +289,7 @@ impl crate::Device<super::Api> for super::Device {
             raw_type: mtl_type,
             mip_levels: desc.mip_level_count,
             array_layers,
-            size: desc.size,
+            copy_size,
         })
     }
 

wgpu-hal/src/metal/mod.rs

@@ -297,6 +297,7 @@ pub struct Surface {
     render_layer: Mutex<mtl::MetalLayer>,
     swapchain_format: wgt::TextureFormat,
     raw_swapchain_format: mtl::MTLPixelFormat,
+    extent: wgt::Extent3d,
     main_thread_id: thread::ThreadId,
     // Useful for UI-intensive applications that are sensitive to
     // window resizing.
@@ -425,7 +426,7 @@ pub struct Texture {
     raw_type: mtl::MTLTextureType,
     array_layers: u32,
     mip_levels: u32,
-    size: wgt::Extent3d,
+    copy_size: crate::CopyExtent,
 }
 
 unsafe impl Send for Texture {}