edgefirst_image/gl/

mod.rs

// SPDX-FileCopyrightText: Copyright 2025 Au-Zone Technologies
// SPDX-License-Identifier: Apache-2.0

#![cfg(any(
    target_os = "linux",
    target_os = "macos",
    target_os = "ios",
    target_os = "android",
    target_os = "windows"
))]
#![cfg(feature = "opengl")]
// Several types defined at the `gl` module root (EglDisplayKind,
// TransferBackend, RegionOfInterest, etc.) are consumed only by the
// Linux-only inner modules (`context`, `dma_import`, `fourcc`). The
// macOS/iOS and Android paths import platform buffers through their own
// modules (`iosurface_import`, `ahardwarebuffer_import`) and do not
// touch every shared type, so some appear unused off Linux.
// Rather than fragmenting the type definitions per platform, suppress
// the dead-code/unused-import lints on non-Linux targets.
#![cfg_attr(not(target_os = "linux"), allow(dead_code, unused_imports))]

// Module layout:
//   - `platform/` — cross-platform display/EGL-loader seam (all OSes)
//   - Linux-only:   `context`, `dma_import`, `fourcc`
//   - macOS/iOS:    `iosurface_import`
//   - Android-only: `ahardwarebuffer_import`
//   - Windows:      `platform/windows.rs` (ANGLE/D3D11, PBO transfers only)
// The engine (`processor`, `threaded`, `cache`, `resources`, `shaders`)
// is portable and reaches platform buffers only through the `GlPlatform`
// trait — see `crates/image/ARCHITECTURE.md`.

macro_rules! function {
    () => {{
        fn f() {}
        fn type_name_of<T>(_: T) -> &'static str {
            std::any::type_name::<T>()
        }
        let name = type_name_of(f);

        // Find and cut the rest of the path
        match &name[..name.len() - 3].rfind(':') {
            Some(pos) => &name[pos + 1..name.len() - 3],
            None => &name[..name.len() - 3],
        }
    }};
}

mod cache;
#[cfg(target_os = "linux")]
mod context;
#[cfg(target_os = "linux")]
mod dma_import;
// Cfg-agnostic: the float render-path classifier is the single source of
// truth for the "(PixelFormat, DType, TensorMemory) → float path" decision and
// is compiled on every platform so both the Linux and macOS backends share one
// definition (see `crates/image/ARCHITECTURE.md`, review item #4).
// Portable renderer helpers (crop uniforms, …) shared by both platform
// backends. No gbm/IOSurface types; compiled on both.
mod core;
mod float_dispatch;
// Pure decision table for the proto-segmentation render path (upload
// strategy × program × count uniform). No GL types; host-tested
// exhaustively. See `proto_dispatch.rs`.
mod proto_dispatch;
// Portable GL render lowering (y-flip viewport, source UV, batch chunk planner).
// No platform types; compiled on both platforms, consumed by the converged
// tile/batch renderer. See `render.rs`.
mod render;
// PixelFormat -> DRM FourCC mapping via the portable `drm_fourcc` crate (NOT
// `gbm`). DRM FourCC is a Linux/DMA-BUF concept, so this lives with the other
// Linux graphics modules; the point is that it carries no `gbm` coupling, so
// `shaders.rs` and the format code no longer pull in `gbm`.
#[cfg(target_os = "android")]
mod ahardwarebuffer_import;
// CUDA test policy: turns a vacuous CUDA skip into a required-coverage
// failure under HAL_TEST_REQUIRE_CUDA=1. Test-only -- every item in it is
// consumed by `tests.rs` under #[cfg(test)] and by its own unit tests, so an
// ungated module would be empty (and dead-code-flagged) in a release build.
#[cfg(test)]
mod cuda_policy;
// `EGL_ANGLE_image_d3d11_texture` attribute assembly (pure); consumed by the
// Windows leaf (`platform/windows.rs`) to build the EGLImage import target.
#[cfg(target_os = "windows")]
pub(super) mod d3d11_import;
#[cfg(target_os = "linux")]
mod fourcc;
#[cfg(any(target_os = "macos", target_os = "ios"))]
mod iosurface_import;
#[cfg(target_os = "android")]
mod native_fence;
mod platform;
mod processor;
mod resources;
mod shaders;
mod shaders_common;
// Engine GL tests, in three tiers (see crates/image/TESTING.md):
//   portable                       — run on Linux AND macOS/ANGLE
//   cfg(target_os = "linux")       — display probing, PBO/CUDA paths
//   cfg(all(linux, dma_test_formats)) — DMA-BUF import/pool specifics
// Per-item cfg gates inside the module select the tier; the mount
// itself is unconditional so the macOS lane runs the portable tier.
mod tests;
mod threaded;

#[cfg(target_os = "linux")]
pub use context::probe_egl_displays;
// These are accessed by sibling sub-modules via `super::context::` directly.
// No re-export needed at the mod.rs level.
pub use cache::{CacheStats, ConvertStats, GlCacheStats};
pub use threaded::GLProcessorThreaded;

/// Dynamically-loaded EGL 1.4 instance. The lifetime parameter is
/// `'static` because the underlying `libloading::Library` is intentionally
/// leaked at first load (see `EGL_LIB` in `context.rs` and the equivalent
/// on macOS — drivers may retain internal state past explicit cleanup, so
/// dlclose can SIGBUS on process exit).
///
/// Defined here at the `gl` module root so the `platform/` trait and both
/// platform implementations can name it without dragging in a cross-cfg
/// re-export. The Linux `context.rs` and the macOS `platform/macos.rs`
/// both use this same alias.
pub(super) type Egl = edgefirst_egl::Instance<
    edgefirst_egl::Dynamic<&'static libloading::Library, edgefirst_egl::EGL1_4>,
>;

/// Handle returned by `GlPlatform::export_completion_fence` (and surfaced
/// by `convert_with_fence`): a sync_file fd on Unix
/// (`EGL_ANDROID_native_fence_sync`), an event NT handle on Windows, set
/// when the D3D11 device fence reaches the value the convert queued.
/// Keeps the engine free of `cfg` branches (precedent: `BlobFd` in
/// `crates/tensor/src/blob.rs`).
#[cfg(unix)]
pub type CompletionFence = std::os::fd::OwnedFd;
#[cfg(windows)]
pub type CompletionFence = std::os::windows::io::OwnedHandle;

/// Identifies the type of EGL display used for headless OpenGL ES rendering.
///
/// The HAL creates a surfaceless GLES 3.0 context
/// (`EGL_KHR_surfaceless_context` + `EGL_KHR_no_config_context`) and
/// renders exclusively through FBOs backed by EGLImages imported from
/// DMA-buf file descriptors. No window or PBuffer surface is created.
///
/// Displays are probed in priority order: PlatformDevice first (zero
/// external dependencies), then GBM, then Default. Use
/// [`probe_egl_displays`] to discover which are available and
/// [`ImageProcessorConfig::egl_display`](crate::ImageProcessorConfig::egl_display)
/// to override the auto-detection.
///
/// # Display Types
///
/// - **`PlatformDevice`** — Uses `EGL_EXT_device_enumeration` to query
///   available EGL devices via `eglQueryDevicesEXT`, then selects the first
///   device with `eglGetPlatformDisplay(EGL_EXT_platform_device, ...)`.
///   Headless and compositor-free with zero external library dependencies.
///   Works on NVIDIA GPUs and newer Vivante drivers. Skipped on Qualcomm
///   Adreno, whose device display crashes in `eglQueryString`; GBM is used
///   there instead.
///
/// - **`Gbm`** — Opens a DRM render node (e.g. `/dev/dri/renderD128`) and
///   creates a GBM (Generic Buffer Manager) device, then calls
///   `eglGetPlatformDisplay(EGL_PLATFORM_GBM_KHR, gbm_device)`. Requires
///   `libgbm` and a DRM render node. Needed on ARM Mali (i.MX95) and older
///   Vivante drivers that do not expose `EGL_EXT_platform_device`.
///
/// - **`Default`** — Calls `eglGetDisplay(EGL_DEFAULT_DISPLAY)`, letting the
///   EGL implementation choose the display. On Wayland systems this connects
///   to the compositor; on X11 it connects to the X server. May block on
///   headless systems where a compositor is expected but not running.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
pub enum EglDisplayKind {
    Gbm,
    PlatformDevice,
    Default,
}

impl std::fmt::Display for EglDisplayKind {
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        match self {
            EglDisplayKind::Gbm => write!(f, "GBM"),
            EglDisplayKind::PlatformDevice => write!(f, "PlatformDevice"),
            EglDisplayKind::Default => write!(f, "Default"),
        }
    }
}

/// A validated, available EGL display discovered by [`probe_egl_displays`].
#[derive(Debug, Clone)]
pub struct EglDisplayInfo {
    /// The type of EGL display.
    pub kind: EglDisplayKind,
    /// Human-readable description for logging/diagnostics
    /// (e.g. "GBM via /dev/dri/renderD128").
    pub description: String,
}

/// Tracks which data-transfer method is active for moving pixels
/// between CPU memory and GPU textures/framebuffers.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub(crate) enum TransferBackend {
    /// Zero-copy via EGLImage imported from DMA-buf file descriptors.
    /// Available on i.MX8 (Vivante), i.MX95 (Mali), Jetson, and any
    /// platform where `EGL_EXT_image_dma_buf_import` is present AND
    /// the GPU can actually render through DMA-buf-backed textures.
    DmaBuf,

    /// Zero-copy via `EGL_ANGLE_iosurface_client_buffer` (macOS/iOS).
    /// Available when ANGLE's Metal backend is loaded and the EGL
    /// extension is advertised. The IOSurface is wrapped as an EGL
    /// pbuffer and bound to a 2D texture via `eglBindTexImage`.
    #[cfg(any(target_os = "macos", target_os = "ios"))]
    IOSurface,

    /// Zero-copy via `EGL_ANDROID_image_native_buffer` (Android). The
    /// AHardwareBuffer is wrapped as an EGLImage via
    /// `eglGetNativeClientBufferANDROID` → `eglCreateImageKHR` and bound
    /// to a 2D texture via `glEGLImageTargetTexture2DOES` — persistent
    /// bindings, like Linux DMA-BUF (unlike the per-pass IOSurface
    /// pbuffer binds).
    #[cfg(target_os = "android")]
    AHardwareBuffer,

    /// Zero-copy via `EGL_ANGLE_image_d3d11_texture` (Windows): the tensor's
    /// `ID3D11Texture2D` wrapped as an EGLImage and bound with
    /// `glEGLImageTargetTexture2DOES`. Unlike Linux and Android the binding
    /// is re-issued per use, because ANGLE does not observe writes made to
    /// the texture outside GL (see `AngleD3d11::PERSISTENT_TEX_BINDINGS`).
    #[cfg(target_os = "windows")]
    D3d11Texture,

    /// GPU buffer via Pixel Buffer Object. Used when DMA-buf is unavailable
    /// but OpenGL is present. Data stays in GPU-accessible memory. Windows
    /// took this path until the D3D11 texture import landed and now reports
    /// [`TransferBackend::D3d11Texture`]; PBO destinations still exist there
    /// as `TensorMemory::Pbo` tensors.
    Pbo,

    /// Synchronous `glTexSubImage2D` upload + `glReadPixels` readback.
    /// Used when DMA-buf is unavailable or when the DMA-buf verification
    /// probe fails (e.g. NVIDIA discrete GPUs where EGLImage creation
    /// succeeds but rendered data is all zeros).
    Sync,
}

impl TransferBackend {
    /// Returns `true` if DMA-buf zero-copy is available.
    pub(crate) fn is_dma(self) -> bool {
        self == TransferBackend::DmaBuf
    }

    /// Returns `true` if the platform can import `TensorMemory::DmaBuf`
    /// tensors zero-copy: DMA-BUF EGLImages on Linux, IOSurface pbuffers
    /// on macOS. Path-selection sites use this; probes that are
    /// specifically about DMA-BUF semantics (e.g. the render-roundtrip
    /// verification) keep `is_dma`.
    pub(crate) fn is_zero_copy(self) -> bool {
        #[cfg(any(target_os = "macos", target_os = "ios"))]
        if self == TransferBackend::IOSurface {
            return true;
        }
        #[cfg(target_os = "android")]
        if self == TransferBackend::AHardwareBuffer {
            return true;
        }
        #[cfg(target_os = "windows")]
        if self == TransferBackend::D3d11Texture {
            return true;
        }
        self == TransferBackend::DmaBuf
    }
}

/// Interpolation mode for int8 proto textures (GL_R8I cannot use GL_LINEAR).
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum Int8InterpolationMode {
    /// texelFetch at nearest texel — simplest, fastest GPU execution.
    Nearest,
    /// texelFetch × 4 neighbors with shader-computed bilinear weights (default).
    Bilinear,
    /// Two-pass: dequant int8→f16 FBO, then existing f16 shader with GL_LINEAR.
    TwoPass,
}

/// A rectangular region of interest expressed as normalised [0, 1] coordinates.
#[derive(Debug, Clone, Copy)]
pub(super) struct RegionOfInterest {
    pub(super) left: f32,
    pub(super) top: f32,
    pub(super) right: f32,
    pub(super) bottom: f32,
}

impl RegionOfInterest {
    /// Build a source ROI spanning a pixel-space crop rectangle's true edges,
    /// clamped to [0, 1]. The `LINEAR` paths keep their kernel inside the crop
    /// by clamping each sample to
    /// [`render::sample_clamp_rect_within`](super::render::sample_clamp_rect_within)
    /// of [`Self::uv_bounds`], not by insetting these coordinates: an inset
    /// maps the crop's edge pixels to the edge texels' centres, which shifts
    /// every sample of a 1:1 crop by up to half a texel and blurs it.
    ///
    /// `crop`: pixel-space rectangle (left, top, width, height).
    /// `tex_w`, `tex_h`: full texture dimensions in pixels.
    pub(super) fn from_crop(crop: &crate::Rect, tex_w: usize, tex_h: usize) -> Self {
        RegionOfInterest {
            left: (crop.left as f32 / tex_w as f32).clamp(0.0, 1.0),
            top: ((crop.top + crop.height) as f32 / tex_h as f32).clamp(0.0, 1.0),
            right: ((crop.left + crop.width) as f32 / tex_w as f32).clamp(0.0, 1.0),
            bottom: (crop.top as f32 / tex_h as f32).clamp(0.0, 1.0),
        }
    }

    /// `[u_min, v_min, u_max, v_max]` of this ROI, whichever way it is flipped.
    pub(super) fn uv_bounds(&self) -> [f32; 4] {
        [
            self.left.min(self.right),
            self.top.min(self.bottom),
            self.left.max(self.right),
            self.top.max(self.bottom),
        ]
    }
}

#[cfg(all(test, target_os = "windows"))]
mod d3d11_backend_tests {
    use super::TransferBackend;

    #[test]
    fn d3d11_texture_is_zero_copy_and_not_dma() {
        assert!(TransferBackend::D3d11Texture.is_zero_copy());
        assert!(!TransferBackend::D3d11Texture.is_dma());
    }
}