wgpu_hal/auxil/
mod.rs

1// Mostly DX12-only, but also compiled for Vulkan-on-Windows, which reuses
2// `dxgi::hdr` to query display HDR info. The DX12-only submodules stay gated
3// behind `dx12` in `dxgi/mod.rs`.
4#[cfg(any(dx12, all(vulkan, windows)))]
5pub(super) mod dxgi;
6
7#[cfg(all(native, feature = "renderdoc"))]
8pub(super) mod renderdoc;
9
10pub mod db {
11    pub mod amd {
12        /// cbindgen:ignore
13        pub const VENDOR: u32 = 0x1002;
14    }
15    pub mod apple {
16        /// cbindgen:ignore
17        pub const VENDOR: u32 = 0x106B;
18    }
19    pub mod arm {
20        /// cbindgen:ignore
21        pub const VENDOR: u32 = 0x13B5;
22    }
23    pub mod broadcom {
24        /// cbindgen:ignore
25        pub const VENDOR: u32 = 0x14E4;
26    }
27    pub mod imgtec {
28        /// cbindgen:ignore
29        pub const VENDOR: u32 = 0x1010;
30    }
31    pub mod intel {
32        /// cbindgen:ignore
33        pub const VENDOR: u32 = 0x8086;
34        pub const DEVICE_KABY_LAKE_MASK: u32 = 0x5900;
35        pub const DEVICE_SKY_LAKE_MASK: u32 = 0x1900;
36    }
37    pub mod mesa {
38        // Mesa does not actually have a PCI vendor id.
39        //
40        // To match Vulkan, we use the VkVendorId for Mesa in the gles backend so that lavapipe (Vulkan) and
41        // llvmpipe (OpenGL) have the same vendor id.
42        /// cbindgen:ignore
43        pub const VENDOR: u32 = 0x10005;
44    }
45    pub mod nvidia {
46        /// cbindgen:ignore
47        pub const VENDOR: u32 = 0x10DE;
48    }
49    pub mod qualcomm {
50        /// cbindgen:ignore
51        pub const VENDOR: u32 = 0x5143;
52    }
53}
54
55/// Maximum binding size for the shaders that only support `i32` indexing.
56/// Interestingly, the index itself can't reach that high, because the minimum
57/// element size is 4 bytes, but the compiler toolchain still computes the
58/// offset at some intermediate point, internally, as i32.
59pub const MAX_I32_BINDING_SIZE: u32 = (1 << 31) - 1;
60
61pub use wgpu_naga_bridge::map_naga_stage;
62
63impl crate::CopyExtent {
64    pub fn map_extent_to_copy_size(extent: &wgt::Extent3d, dim: wgt::TextureDimension) -> Self {
65        Self {
66            width: extent.width,
67            height: extent.height,
68            depth: match dim {
69                wgt::TextureDimension::D1 | wgt::TextureDimension::D2 => 1,
70                wgt::TextureDimension::D3 => extent.depth_or_array_layers,
71            },
72        }
73    }
74
75    pub fn min(&self, other: &Self) -> Self {
76        Self {
77            width: self.width.min(other.width),
78            height: self.height.min(other.height),
79            depth: self.depth.min(other.depth),
80        }
81    }
82
83    // Get the copy size at a specific mipmap level. This doesn't make most sense,
84    // since the copy extents are provided *for* a mipmap level to start with.
85    // But backends use `CopyExtent` more sparingly, and this piece is shared.
86    pub fn at_mip_level(&self, level: u32) -> Self {
87        Self {
88            width: (self.width >> level).max(1),
89            height: (self.height >> level).max(1),
90            depth: (self.depth >> level).max(1),
91        }
92    }
93}
94
95impl crate::TextureCopyBase {
96    pub fn max_copy_size(&self, full_size: &crate::CopyExtent) -> crate::CopyExtent {
97        let mip = full_size.at_mip_level(self.mip_level);
98        crate::CopyExtent {
99            width: mip.width - self.origin.x,
100            height: mip.height - self.origin.y,
101            depth: mip.depth - self.origin.z,
102        }
103    }
104}
105
106impl crate::BufferTextureCopy {
107    pub fn clamp_size_to_virtual(&mut self, full_size: &crate::CopyExtent) {
108        let max_size = self.texture_base.max_copy_size(full_size);
109        self.size = self.size.min(&max_size);
110    }
111}
112
113impl crate::TextureCopy {
114    pub fn clamp_size_to_virtual(
115        &mut self,
116        full_src_size: &crate::CopyExtent,
117        full_dst_size: &crate::CopyExtent,
118    ) {
119        let max_src_size = self.src_base.max_copy_size(full_src_size);
120        let max_dst_size = self.dst_base.max_copy_size(full_dst_size);
121        self.size = self.size.min(&max_src_size).min(&max_dst_size);
122    }
123}
124
125/// Adjust `limits` to honor HAL-imposed maximums and comply with WebGPU's
126/// adapter capability guarantees.
127#[cfg_attr(not(any_backend), allow(dead_code))]
128pub(crate) fn adjust_raw_limits(mut limits: wgt::Limits) -> wgt::Limits {
129    // Apply hal limits.
130    limits.max_bind_groups = limits.max_bind_groups.min(crate::MAX_BIND_GROUPS as u32);
131    limits.max_vertex_buffers = limits
132        .max_vertex_buffers
133        .min(crate::MAX_VERTEX_BUFFERS as u32);
134    // Once we allow the 2 limits above to be higher than 24 we should use
135    // `cap_limits_to_be_under_the_sum_limit` to cap them under
136    // `max_bind_groups_plus_vertex_buffers`.
137    const { assert!(crate::MAX_BIND_GROUPS + crate::MAX_VERTEX_BUFFERS == 24) };
138    limits.max_bind_groups_plus_vertex_buffers = limits.max_bind_groups_plus_vertex_buffers.min(24);
139    limits.max_color_attachments = limits
140        .max_color_attachments
141        .min(crate::MAX_COLOR_ATTACHMENTS as u32);
142
143    // Adjust limits according to WebGPU adapter capability guarantees.
144    // See <https://gpuweb.github.io/gpuweb/#adapter-capability-guarantees>.
145
146    // WebGPU requires maxBindingsPerBindGroup to be at least the sum of all
147    // per-stage limits multiplied with the maximum shader stages per pipeline.
148    //
149    // Since backends already report their maximum maxBindingsPerBindGroup,
150    // we need to lower all per-stage limits to satisfy this guarantee.
151    const MAX_SHADER_STAGES_PER_PIPELINE: u32 = 2;
152    let max_per_stage_resources =
153        limits.max_bindings_per_bind_group / MAX_SHADER_STAGES_PER_PIPELINE;
154
155    cap_limits_to_be_under_the_sum_limit(
156        [
157            &mut limits.max_sampled_textures_per_shader_stage,
158            &mut limits.max_uniform_buffers_per_shader_stage,
159            &mut limits.max_storage_textures_per_shader_stage,
160            &mut limits.max_storage_buffers_per_shader_stage,
161            &mut limits.max_samplers_per_shader_stage,
162            &mut limits.max_acceleration_structures_per_shader_stage,
163        ],
164        max_per_stage_resources,
165    );
166
167    // Not required by the spec but dynamic buffers count
168    // towards non-dynamic buffer limits as well.
169    limits.max_dynamic_uniform_buffers_per_pipeline_layout = limits
170        .max_dynamic_uniform_buffers_per_pipeline_layout
171        .min(limits.max_uniform_buffers_per_shader_stage);
172    limits.max_dynamic_storage_buffers_per_pipeline_layout = limits
173        .max_dynamic_storage_buffers_per_pipeline_layout
174        .min(limits.max_storage_buffers_per_shader_stage);
175
176    limits.min_uniform_buffer_offset_alignment = limits.min_uniform_buffer_offset_alignment.max(32);
177    limits.min_storage_buffer_offset_alignment = limits.min_storage_buffer_offset_alignment.max(32);
178
179    limits.max_uniform_buffer_binding_size = limits
180        .max_uniform_buffer_binding_size
181        .min(limits.max_buffer_size);
182    limits.max_storage_buffer_binding_size = limits
183        .max_storage_buffer_binding_size
184        .min(limits.max_buffer_size);
185
186    limits.max_storage_buffer_binding_size &= !(u64::from(wgt::STORAGE_BINDING_SIZE_ALIGNMENT) - 1);
187    limits.max_vertex_buffer_array_stride &= !(wgt::VERTEX_ALIGNMENT as u32 - 1);
188
189    let x = limits.max_compute_workgroup_size_x;
190    let y = limits.max_compute_workgroup_size_y;
191    let z = limits.max_compute_workgroup_size_z;
192    let m = limits.max_compute_invocations_per_workgroup;
193    limits.max_compute_workgroup_size_x = x.min(m);
194    limits.max_compute_workgroup_size_y = y.min(m);
195    limits.max_compute_workgroup_size_z = z.min(m);
196    limits.max_compute_invocations_per_workgroup = m.min(x.saturating_mul(y).saturating_mul(z));
197
198    limits.max_immediate_size = limits.max_immediate_size.min(256);
199
200    limits
201}
202
203/// Evenly allocates space to each limit,
204/// capping them only if strictly necessary.
205pub fn cap_limits_to_be_under_the_sum_limit<const N: usize>(
206    mut limits: [&mut u32; N],
207    sum_limit: u32,
208) {
209    limits.sort();
210
211    let mut rem_limit = sum_limit;
212    let mut divisor = limits.len() as u32;
213    for limit_to_adjust in limits {
214        let limit = rem_limit / divisor;
215        *limit_to_adjust = (*limit_to_adjust).min(limit);
216        rem_limit -= *limit_to_adjust;
217        divisor -= 1;
218    }
219}
220
221#[cfg(test)]
222mod tests {
223    use super::*;
224
225    #[test]
226    fn test_cap_limits_to_be_under_the_sum_limit() {
227        test([3, 3, 3], 3, [1, 1, 1]);
228        test([3, 2, 1], 3, [1, 1, 1]);
229        test([1, 2, 3], 6, [1, 2, 3]);
230        test([1, 2, 3], 3, [1, 1, 1]);
231        test([1, 8, 100], 6, [1, 2, 3]);
232        test([2, 80, 80], 6, [2, 2, 2]);
233        test([2, 80, 80], 12, [2, 5, 5]);
234
235        #[track_caller]
236        fn test<const N: usize>(mut input: [u32; N], limit: u32, output: [u32; N]) {
237            cap_limits_to_be_under_the_sum_limit(input.each_mut(), limit);
238            assert_eq!(input, output);
239        }
240    }
241}
wgpu_hal/auxil/mod.rs

wgpu_hal/auxil/
mod.rs
