Allowlist mali drivers for gpu preprocessing support. (#18769)

Fixes #17591

Looking at the arm downloads page, "r48p0" is a version number that
increments, where rXX is the major version and pX seems to be a patch
version. Take the conservative approach here that we know gpu
preprocessing is working on at least version 48 and presumably higher.
The assumption here is that the driver_info string will be reported
similarly on non-pixel devices.

crates/bevy_core_pipeline/src/experimental/mip_generation/mod.rs

@@ -7,6 +7,10 @@
 
 use core::array;
 
+use crate::core_3d::{
+    graph::{Core3d, Node3d},
+    prepare_core_3d_depth_textures,
+};
 use bevy_app::{App, Plugin};
 use bevy_asset::{load_internal_asset, weak_handle, Handle};
 use bevy_derive::{Deref, DerefMut};
@@ -21,6 +25,7 @@ use bevy_ecs::{
     world::{FromWorld, World},
 };
 use bevy_math::{uvec2, UVec2, Vec4Swizzles as _};
+use bevy_render::batching::gpu_preprocessing::GpuPreprocessingSupport;
 use bevy_render::{
     experimental::occlusion_culling::{
         OcclusionCulling, OcclusionCullingSubview, OcclusionCullingSubviewEntities,
@@ -30,23 +35,19 @@ use bevy_render::{
         binding_types::{sampler, texture_2d, texture_2d_multisampled, texture_storage_2d},
         BindGroup, BindGroupEntries, BindGroupLayout, BindGroupLayoutEntries,
         CachedComputePipelineId, ComputePassDescriptor, ComputePipeline, ComputePipelineDescriptor,
-        DownlevelFlags, Extent3d, IntoBinding, PipelineCache, PushConstantRange, Sampler,
-        SamplerBindingType, SamplerDescriptor, Shader, ShaderStages, SpecializedComputePipeline,
+        Extent3d, IntoBinding, PipelineCache, PushConstantRange, Sampler, SamplerBindingType,
+        SamplerDescriptor, Shader, ShaderStages, SpecializedComputePipeline,
         SpecializedComputePipelines, StorageTextureAccess, TextureAspect, TextureDescriptor,
         TextureDimension, TextureFormat, TextureSampleType, TextureUsages, TextureView,
         TextureViewDescriptor, TextureViewDimension,
     },
-    renderer::{RenderAdapter, RenderContext, RenderDevice},
+    renderer::{RenderContext, RenderDevice},
     texture::TextureCache,
     view::{ExtractedView, NoIndirectDrawing, ViewDepthTexture},
     Render, RenderApp, RenderSet,
 };
 use bitflags::bitflags;
-
-use crate::core_3d::{
-    graph::{Core3d, Node3d},
-    prepare_core_3d_depth_textures,
-};
+use tracing::debug;
 
 /// Identifies the `downsample_depth.wgsl` shader.
 pub const DOWNSAMPLE_DEPTH_SHADER_HANDLE: Handle<Shader> =
@@ -325,26 +326,14 @@ pub struct DownsampleDepthPipelines {
     sampler: Sampler,
 }
 
-fn supports_compute_shaders(device: &RenderDevice, adapter: &RenderAdapter) -> bool {
-    adapter
-        .get_downlevel_capabilities()
-        .flags
-        .contains(DownlevelFlags::COMPUTE_SHADERS)
-    // Even if the adapter supports compute, we might be simulating a lack of
-    // compute via device limits (see `WgpuSettingsPriority::WebGL2` and
-    // `wgpu::Limits::downlevel_webgl2_defaults()`). This will have set all the
-    // `max_compute_*` limits to zero, so we arbitrarily pick one as a canary.
-    && (device.limits().max_compute_workgroup_storage_size != 0)
-}
-
 /// Creates the [`DownsampleDepthPipelines`] if downsampling is supported on the
 /// current platform.
 fn create_downsample_depth_pipelines(
     mut commands: Commands,
     render_device: Res<RenderDevice>,
-    render_adapter: Res<RenderAdapter>,
     pipeline_cache: Res<PipelineCache>,
     mut specialized_compute_pipelines: ResMut<SpecializedComputePipelines<DownsampleDepthPipeline>>,
+    gpu_preprocessing_support: Res<GpuPreprocessingSupport>,
     mut has_run: Local<bool>,
 ) {
     // Only run once.
@@ -356,9 +345,8 @@ fn create_downsample_depth_pipelines(
     }
     *has_run = true;
 
-    // If we don't have compute shaders, we can't invoke the downsample depth
-    // compute shader.
-    if !supports_compute_shaders(&render_device, &render_adapter) {
+    if !gpu_preprocessing_support.is_culling_supported() {
+        debug!("Downsample depth is not supported on this platform.");
         return;
     }
 

crates/bevy_render/src/batching/gpu_preprocessing.rs

@@ -20,7 +20,7 @@ use bytemuck::{Pod, Zeroable};
 use encase::{internal::WriteInto, ShaderSize};
 use indexmap::IndexMap;
 use nonmax::NonMaxU32;
-use tracing::error;
+use tracing::{error, info};
 use wgpu::{BindingResource, BufferUsages, DownlevelFlags, Features};
 
 use crate::{
@@ -1093,12 +1093,13 @@ impl FromWorld for GpuPreprocessingSupport {
         let adapter = world.resource::<RenderAdapter>();
         let device = world.resource::<RenderDevice>();
 
-        // Filter some Qualcomm devices on Android as they crash when using GPU
-        // preprocessing.
-        // We filter out Adreno 730 and earlier GPUs (except 720, as it's newer
+        // Filter Android drivers that are incompatible with GPU preprocessing:
+        // - We filter out Adreno 730 and earlier GPUs (except 720, as it's newer
         //   than 730).
+        // - We filter out Mali GPUs with driver versions lower than 48.
         fn is_non_supported_android_device(adapter: &RenderAdapter) -> bool {
             crate::get_adreno_model(adapter).is_some_and(|model| model != 720 && model <= 730)
+                || crate::get_mali_driver_version(adapter).is_some_and(|version| version < 48)
         }
 
         let culling_feature_support = device.features().contains(
@@ -1107,18 +1108,31 @@ impl FromWorld for GpuPreprocessingSupport {
                 | Features::PUSH_CONSTANTS,
         );
         // Depth downsampling for occlusion culling requires 12 textures
-        let limit_support = device.limits().max_storage_textures_per_shader_stage >= 12;
+        let limit_support = device.limits().max_storage_textures_per_shader_stage >= 12 &&
+            // Even if the adapter supports compute, we might be simulating a lack of
+            // compute via device limits (see `WgpuSettingsPriority::WebGL2` and
+            // `wgpu::Limits::downlevel_webgl2_defaults()`). This will have set all the
+            // `max_compute_*` limits to zero, so we arbitrarily pick one as a canary.
+            device.limits().max_compute_workgroup_storage_size != 0;
+
         let downlevel_support = adapter.get_downlevel_capabilities().flags.contains(
+            DownlevelFlags::COMPUTE_SHADERS |
             DownlevelFlags::VERTEX_AND_INSTANCE_INDEX_RESPECTS_RESPECTIVE_FIRST_VALUE_IN_INDIRECT_DRAW
         );
 
         let max_supported_mode = if device.limits().max_compute_workgroup_size_x == 0
             || is_non_supported_android_device(adapter)
         {
+            info!(
+                "GPU preprocessing is not supported on this device. \
+                Falling back to CPU preprocessing.",
+            );
             GpuPreprocessingMode::None
         } else if !(culling_feature_support && limit_support && downlevel_support) {
+            info!("Some GPU preprocessing are limited on this device.");
             GpuPreprocessingMode::PreprocessingOnly
         } else {
+            info!("GPU preprocessing is fully supported on this device.");
             GpuPreprocessingMode::Culling
         };
 

crates/bevy_render/src/lib.rs

@@ -584,3 +584,26 @@ pub fn get_adreno_model(adapter: &RenderAdapter) -> Option<u32> {
             .fold(0, |acc, digit| acc * 10 + digit),
     )
 }
+
+/// Get the Mali driver version if the adapter is a Mali GPU.
+pub fn get_mali_driver_version(adapter: &RenderAdapter) -> Option<u32> {
+    if !cfg!(target_os = "android") {
+        return None;
+    }
+
+    let driver_name = adapter.get_info().name;
+    if !driver_name.contains("Mali") {
+        return None;
+    }
+    let driver_info = adapter.get_info().driver_info;
+    if let Some(start_pos) = driver_info.find("v1.r") {
+        if let Some(end_pos) = driver_info[start_pos..].find('p') {
+            let start_idx = start_pos + 4; // Skip "v1.r"
+            let end_idx = start_pos + end_pos;
+
+            return driver_info[start_idx..end_idx].parse::<u32>().ok();
+        }
+    }
+
+    None
+}