RUST.md: iris's bindless texture array does not survive real Android GPUs

Iris asked whether the 'unknown number of images' approach even works on
mobile. It does not, measured with a new rig (rigs/gpu-probe, no APK
needed) and sourced rather than recalled: the emulator's software Vulkan
refuses iris's descriptor-indexing request outright, and on real hardware
the current Android Vulkan Profile baseline (80.1% of active devices)
does not require VK_EXT_descriptor_indexing either -- Arm's own docs say
only Valhall/5th-Gen Mali (2019+) support it.

iris already solved the identical problem for text in I1 (the glyph
atlas). The recommendation is to generalize it to images rather than
widen the binding array further; not yet implemented, since it changes
iris's render core.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

rigs/gpu-probe/.gitignore

@@ -0,0 +1 @@
+target

rigs/gpu-probe/Cargo.toml

@@ -0,0 +1,18 @@
+[package]
+name = "gpu-probe"
+version = "0.1.0"
+edition = "2024"
+
+# Deliberately its own crate rather than a member of iris's workspace: the
+# vendored `iris/` tree is meant to stay reconcilable with the iris/iris
+# repository, and this is a rig belonging to ai-app.
+
+[dependencies]
+# Pinned to what iris asks for, so the answer is about iris rather than
+# about a different wgpu.
+wgpu = "28.0.0"
+pollster = "0.4.0"
+# Queried directly, because wgpu and `cmd gpu vkjson` disagreed about
+# descriptor indexing in the emulator and only the raw call says which is
+# right.
+ash = "0.38"

rigs/gpu-probe/src/main.rs

@@ -0,0 +1,154 @@
+//! Ask a device whether it can give iris the GPU it asks for.
+//!
+//! iris's renderer binds every texture it has drawn as one binding array and
+//! indexes it non-uniformly from the shader, which needs descriptor indexing
+//! and a very large per-stage binding-array limit (101,000 elements: 100,000
+//! textures and 1,000 samplers, `UiLimits::default`). Those are ordinary on a
+//! desktop and not obviously available on a phone, so this reports what the
+//! adapter offers before anything is built on the assumption.
+//!
+//! It runs as a plain executable with no window and no APK, because
+//! `request_adapter` needs no surface -- so it can be pushed to a device with
+//! `adb push` and run from `/data/local/tmp`, which is far cheaper than an
+//! app. What it therefore cannot answer is anything about presenting to a
+//! surface; that is the Android backend's own problem.
+
+mod vk;
+
+use wgpu::*;
+
+/// What `iris/src/default/render.rs` asks `request_device` for.
+fn iris_features() -> Features {
+    Features::TEXTURE_BINDING_ARRAY
+        | Features::PARTIALLY_BOUND_BINDING_ARRAY
+        | Features::SAMPLED_TEXTURE_AND_STORAGE_BUFFER_ARRAY_NON_UNIFORM_INDEXING
+}
+
+/// `UiLimits::default()`: 100,000 textures + 1,000 samplers.
+const IRIS_MAX_BINDING_ARRAY: u32 = 101_000;
+const IRIS_MAX_BINDING_ARRAY_SAMPLERS: u32 = 1_000;
+
+fn main() {
+    vk::report();
+
+    let instance = Instance::new(&InstanceDescriptor {
+        backends: Backends::from_env().unwrap_or(Backends::PRIMARY),
+        ..Default::default()
+    });
+
+    let adapters = pollster::block_on(instance.enumerate_adapters(Backends::all()));
+    println!("adapters: {}", adapters.len());
+    for adapter in &adapters {
+        let info = adapter.get_info();
+        println!(
+            "  {:?} {} ({:?})",
+            info.backend, info.name, info.device_type
+        );
+    }
+
+    let Some(adapter) = pollster::block_on(instance.request_adapter(&RequestAdapterOptions {
+        power_preference: PowerPreference::default(),
+        compatible_surface: None,
+        force_fallback_adapter: false,
+    }))
+    .ok() else {
+        println!("\nNO ADAPTER");
+        std::process::exit(1);
+    };
+
+    let info = adapter.get_info();
+    println!(
+        "\nchosen: {:?} {} ({:?})",
+        info.backend, info.name, info.device_type
+    );
+    println!("driver: {} {}", info.driver, info.driver_info);
+
+    let have = adapter.features();
+    println!("\nfeatures iris requires:");
+    let mut missing = Features::empty();
+    for f in iris_features().iter() {
+        let ok = have.contains(f);
+        println!(
+            "  {:60} {}",
+            format!("{f:?}"),
+            if ok { "yes" } else { "NO" }
+        );
+        if !ok {
+            missing |= f;
+        }
+    }
+
+    let limits = adapter.limits();
+    println!("\nlimits iris requires:");
+    for (name, want, got) in [
+        (
+            "max_binding_array_elements_per_shader_stage",
+            IRIS_MAX_BINDING_ARRAY,
+            limits.max_binding_array_elements_per_shader_stage,
+        ),
+        (
+            "max_binding_array_sampler_elements_per_shader_stage",
+            IRIS_MAX_BINDING_ARRAY_SAMPLERS,
+            limits.max_binding_array_sampler_elements_per_shader_stage,
+        ),
+    ] {
+        println!(
+            "  {name:52} want {want:>7}  have {got:>7}  {}",
+            if got >= want { "ok" } else { "TOO SMALL" }
+        );
+    }
+    println!(
+        "  {:52} want {:>7}  have {:>7}",
+        "max_buffer_size (iris asks 1<<30)",
+        1u64 << 30,
+        limits.max_buffer_size
+    );
+
+    // The question that actually matters: does the device iris builds come
+    // back, or does wgpu refuse it?
+    let mut wanted = Limits {
+        max_binding_array_elements_per_shader_stage: IRIS_MAX_BINDING_ARRAY,
+        max_binding_array_sampler_elements_per_shader_stage: IRIS_MAX_BINDING_ARRAY_SAMPLERS,
+        max_buffer_size: 1 << 30,
+        ..Default::default()
+    };
+    match pollster::block_on(adapter.request_device(&DeviceDescriptor {
+        required_features: iris_features(),
+        required_limits: wanted.clone(),
+        ..Default::default()
+    })) {
+        Ok(_) => println!("\nIRIS DEVICE: ok"),
+        Err(e) => println!("\nIRIS DEVICE: FAILED -- {e}"),
+    }
+
+    // If it failed, say how far down it has to be turned before it works, so
+    // the report names a number to design against rather than just "no".
+    if missing.is_empty() {
+        for cap in [
+            limits.max_binding_array_elements_per_shader_stage,
+            1024,
+            128,
+            16,
+        ] {
+            if cap >= IRIS_MAX_BINDING_ARRAY {
+                continue;
+            }
+            wanted.max_binding_array_elements_per_shader_stage = cap;
+            wanted.max_binding_array_sampler_elements_per_shader_stage =
+                cap.min(IRIS_MAX_BINDING_ARRAY_SAMPLERS);
+            let ok = pollster::block_on(adapter.request_device(&DeviceDescriptor {
+                required_features: iris_features(),
+                required_limits: wanted.clone(),
+                ..Default::default()
+            }))
+            .is_ok();
+            println!(
+                "  binding array capped at {cap:>7}: {}",
+                if ok { "ok" } else { "no" }
+            );
+            if ok {
+                break;
+            }
+        }
+    }
+}

rigs/gpu-probe/src/vk.rs

@@ -0,0 +1,108 @@
+//! The raw Vulkan half of the probe.
+//!
+//! wgpu reports a feature only after a chain of its own decisions -- which
+//! physical device, which API version, which extension list -- so "wgpu says
+//! no" and "the driver says no" are different claims. This asks
+//! `vkGetPhysicalDeviceFeatures2` itself and prints the inputs to that chain,
+//! so a disagreement can be attributed rather than guessed at.
+
+use ash::{Entry, vk};
+use std::ffi::CStr;
+
+pub fn report() {
+    let entry = match unsafe { Entry::load() } {
+        Ok(e) => e,
+        Err(e) => {
+            println!("\nraw vulkan: cannot load loader -- {e}");
+            return;
+        }
+    };
+
+    let instance_version = match unsafe { entry.try_enumerate_instance_version() } {
+        Ok(Some(v)) => v,
+        Ok(None) => vk::API_VERSION_1_0,
+        Err(e) => {
+            println!("\nraw vulkan: enumerate_instance_version failed -- {e}");
+            return;
+        }
+    };
+    println!("\nraw vulkan:");
+    println!("  loader instance version: {}", ver(instance_version));
+
+    // Ask for the highest instance version the loader admits to: wgpu clamps
+    // the device version by the instance's, so an instance created at 1.0
+    // makes a 1.3 device look like 1.0.
+    let app_info = vk::ApplicationInfo::default().api_version(instance_version);
+    let create = vk::InstanceCreateInfo::default().application_info(&app_info);
+    let instance = match unsafe { entry.create_instance(&create, None) } {
+        Ok(i) => i,
+        Err(e) => {
+            println!("  create_instance failed -- {e}");
+            return;
+        }
+    };
+
+    let devices = unsafe { instance.enumerate_physical_devices() }.unwrap_or_default();
+    for phd in devices {
+        let props = unsafe { instance.get_physical_device_properties(phd) };
+        let name = unsafe { CStr::from_ptr(props.device_name.as_ptr()) };
+        println!("  device: {}", name.to_string_lossy());
+        println!("    device api version: {}", ver(props.api_version));
+
+        let exts =
+            unsafe { instance.enumerate_device_extension_properties(phd) }.unwrap_or_default();
+        let has_ext = exts.iter().any(|e| {
+            (unsafe { CStr::from_ptr(e.extension_name.as_ptr()) }) == c"VK_EXT_descriptor_indexing"
+        });
+        println!("    VK_EXT_descriptor_indexing advertised: {has_ext}");
+        println!("    device extensions: {}", exts.len());
+
+        let mut indexing = vk::PhysicalDeviceDescriptorIndexingFeatures::default();
+        let mut features2 = vk::PhysicalDeviceFeatures2::default().push_next(&mut indexing);
+        unsafe { instance.get_physical_device_features2(phd, &mut features2) };
+
+        for (name, v) in [
+            (
+                "shaderSampledImageArrayNonUniformIndexing",
+                indexing.shader_sampled_image_array_non_uniform_indexing,
+            ),
+            (
+                "descriptorBindingSampledImageUpdateAfterBind",
+                indexing.descriptor_binding_sampled_image_update_after_bind,
+            ),
+            (
+                "shaderStorageImageArrayNonUniformIndexing",
+                indexing.shader_storage_image_array_non_uniform_indexing,
+            ),
+            (
+                "descriptorBindingStorageImageUpdateAfterBind",
+                indexing.descriptor_binding_storage_image_update_after_bind,
+            ),
+            (
+                "shaderStorageBufferArrayNonUniformIndexing",
+                indexing.shader_storage_buffer_array_non_uniform_indexing,
+            ),
+            (
+                "descriptorBindingStorageBufferUpdateAfterBind",
+                indexing.descriptor_binding_storage_buffer_update_after_bind,
+            ),
+            (
+                "descriptorBindingPartiallyBound",
+                indexing.descriptor_binding_partially_bound,
+            ),
+        ] {
+            println!("    {name:48} {}", if v != 0 { "yes" } else { "NO" });
+        }
+    }
+
+    unsafe { instance.destroy_instance(None) };
+}
+
+fn ver(v: u32) -> String {
+    format!(
+        "{}.{}.{}",
+        vk::api_version_major(v),
+        vk::api_version_minor(v),
+        vk::api_version_patch(v)
+    )
+}