gfx-rs/hal跨平台图形抽象库使用介绍

文档列表见：Rust 移动端跨平台复杂图形渲染项目开发系列总结（目录）

gfx-rs/gfx是一个Rust编写的底层、跨平台图形抽象库，包含如下层或组件：

• gfx-HAL
• gfx-backend-
  • Metal
  • Vulkan
  • OpenGL，开发中，由于GL与下一代接口Vulkan差异过大，这个模块可能做不完
  • OpenGL ES，开发中，由于GL与下一代接口Vulkan差异过大，这个模块可能做不完
  • DirectX 11
  • DirectX 12
  • WebGL，开发中，由于GL与下一代接口Vulkan差异过大，这个模块可能做不完
• gfx-warden

本文档只考虑master分支，对应HAL新接口，忽略pre-II老接口。 另外，只考虑用gfx-hal实现离线渲染和计算着色器功能，即，渲染到纹理和GPGPU。渲染到窗口及鼠标、键盘事件处理可参考gfx自带DEMO。

初始化具体图形库后端

gfx-hal接口近乎1:1仿造Vulkan接口，可以参考Vulkan各种教程，Vulkan的操作从Instance创建开始。

#[cfg(any(feature = "vulkan", feature = "dx12", feature = "metal"))]
let instance = backend::Instance::create("name", 1 /* version */);
复制代码

创建不同的设备和队列需要适配器满足不同的能力要求，下面逐一描述。

创建不同功能的设备

整体流程为backend::Instance::create() -> enumerate_adapters() -> open_with()。

• enumerate_adapters()是目前可用的Adapter列表，在这一步我们先选择支持指定队列功能要求的适配器。在下一步打开具体逻辑Device时直接返回true。
• open_with()第一个参数count表示要打开的QueueFamily数量，目前移动设备通常只有一个GPU，此时传递1即可，我目前主观认为打开多个QueueFamily并不能提高整个App的图形性能。

配置完在macOS上运行可得到类似如下信息，配置细节参考后面内容。

AdapterInfo {
  name: "Intel Iris Pro Graphics", 
  vendor: 0, 
  device: 0, 
  device_type: IntegratedGpu }
Limits { 
  max_texture_size: 4096, 
  max_texel_elements: 16777216, 
  ... }
Memory types: [
  MemoryType { properties: DEVICE_LOCAL, heap_index: 0 }, 
  MemoryType { properties: COHERENT | CPU_VISIBLE, heap_index: 1 },
  ...]
复制代码

只支持渲染的设备

渲染是图形设备存在的意义，故简单粗暴地取出第1个适配器进行后面操作。

let mut adapter = instance.enumerate_adapters().remove(0);
let (mut device, mut queue_group) = adapter
  .open_with::<_, Graphics>(1, |_family| true
  .unwrap();
复制代码

只支持计算的设备

考虑到低版本的OpenGL不支持Compute Shader，此时需要过滤，如果只编译Metal/Vulkan，和上面一样enumerate_adapters().remove(0)即可。相应地，open_with()作了调整。

let mut adapter = instance
    .enumerate_adapters()
    .into_iter()
    .find(|a| {
        a.queue_families
            .iter()
            .any(|family| family.supports_compute())
    })
    .expect("Failed to find a GPU with compute support!");
let (mut device, mut queue_group) = adapter
    .open_with::<_, Compute>(1, |_family| true)
    .unwrap();    
复制代码

同时支持渲染+计算的设备

同上，适配器过滤条件和open_with()都调整成同时满足渲染与计算要求。

let mut adapter = instance
    .enumerate_adapters()
    .into_iter()
    .find(|a| {
        a.queue_families
            .iter()
            .any(|family| family.supports_graphics() && family.supports_compute())
    }).expect("Failed to find a GPU with graphics and compute support!");
let (mut device, mut queue_group) = adapter
    .open_with::<_, General>(1, |_family| true)
    .unwrap();    
复制代码

有了Device和QueueGroup可开始创建Image（可看作Vulkan版Texture）、Pipeline等资源。

创建资源

Buffer

创建Buffer

Buffer和Image本身并不存储数据，它们表达了存储数据要满足的条件，这些条件用于创建Memory。

let stride = size_of::<T>() as u64;
let upload_size = data_source.len() as u64 * stride;
let usage = buffer::Usage::TRANSFER_SRC | buffer::Usage::TRANSFER_DST;
let unbound = device.create_buffer(upload_size, usage).unwrap();
复制代码

• data_source为需要分配的存储空间大小，在此只作为一个标志，实际创建存储空间由后面介绍的Memory实现。
• usage根据Buffer的实际用途进行配置，不同组合对性能影响较大，需注意。

销毁Buffer

device.destroy_buffer(buffer);
复制代码

连接Buffer与Memory

buffer = device.bind_buffer_memory(&memory, 0, unbound).unwrap();
复制代码

bind_buffer_memory后Buffer对象才拥有实际的存储空间，但是，数据还是存在Memory对象中。

连接Buffer和Memory后，在macOS上通常输出如下信息，其中length = 256的256与前面输出的Limits某一项相关，具体内容视显卡而定：

buffer = Buffer { raw: <MTLIGAccelBuffer: 0x7fa9a472d470>
    label = <none> 
    length = 256 
    cpuCacheMode = MTLCPUCacheModeDefaultCache 
    storageMode = MTLStorageModeShared 
    resourceOptions = MTLResourceCPUCacheModeDefaultCache MTLResourceStorageModeShared  
    purgeableState = MTLPurgeableStateNonVolatile, range: 0..6, options: CPUCacheModeDefaultCache | StorageModeShared }
memory = Memory { heap: Public(MemoryTypeId(1), <MTLIGAccelBuffer: 0x7fa9a472d470>
    label = <none> 
    length = 256 
    cpuCacheMode = MTLCPUCacheModeDefaultCache 
    storageMode = MTLStorageModeShared 
    resourceOptions = MTLResourceCPUCacheModeDefaultCache MTLResourceStorageModeShared  
    purgeableState = MTLPurgeableStateNonVolatile), size: 256 }
复制代码

创建BufferView

let format = Some(format::Format::Rg4Unorm);
let size = data_source.len();
let buffer_view = device.create_buffer_view(buffer, format, 0..size);
复制代码

销毁BufferView

device.destroy_buffer_view(buffer_view);
复制代码

Memory

Memory分配用于存储Buffer和Image所需数据的内存空间。

创建Memory

// A note about performance: Using CPU_VISIBLE memory is convenient because it can be
// directly memory mapped and easily updated by the CPU, but it is very slow and so should
// only be used for small pieces of data that need to be updated very frequently. For something like
// a vertex buffer that may be much larger and should not change frequently, you should instead
// use a DEVICE_LOCAL buffer that gets filled by copying data from a CPU_VISIBLE staging buffer.
let upload_type = memory_types
    .iter()
    .enumerate()
    .position(|(id, mem_type)| {
        mem_req.type_mask & (1 << id) != 0 && mem_type.properties.contains(memory::Properties::CPU_VISIBLE)
    })
    .unwrap()
    .into();
let mem_req = device.get_buffer_requirements(&unbound);    
let memory = device.allocate_memory(upload_type, mem_req.size).unwrap();
复制代码

Memory写入

Memory的读写都要映射相应的Write/Reader，为了线程安全，需要手工加上合适的Fence。

let mut data_target = device.acquire_mapping_writer::<T>(&memory, 0..size).unwrap();
data_target[0..data_source.len()].copy_from_slice(data_source);
device.release_mapping_writer(data_target);
复制代码

Memory读取

let reader = device.acquire_mapping_reader::<u32>(&staging_memory, 0..staging_size).unwrap();
println!("Times: {:?}", reader[0..numbers.len()].into_iter().map(|n| *n).collect::<Vec<u32>>());
device.release_mapping_reader(reader);
复制代码

Image

创建Image

类似Buffer对象，Image对象本身也不存储实际的纹理数据。

let kind = image::Kind::D2(dims.width as image::Size, dims.height as image::Size,
                           1/* Layer */, 1/* NumSamples */);
let unbound = device
    .create_image(
        kind,
        1,
        ColorFormat::SELF,
        image::Tiling::Optimal,
        image::Usage::TRANSFER_DST | image::Usage::SAMPLED,
        image::StorageFlags::empty(),
    )
    .unwrap();
复制代码

同样，创建Image时指定的Usage也要根据Image的实际用途来组合，不合理的组合会降低性能。

销毁Image

device.destroy_image_view(image_view);
复制代码

连接Image到Memory

let image = device.bind_image_memory(&memory, 0, unbound).unwrap();
复制代码

创建ImageView

创建Sampler

组织绘制命令

创建Submission

hal-buffer创建、读写

buffer::Usage::TRANSFER_SRC | buffer::Usage::TRANSFER_DST,
复制代码

功能、区别

let (staging_memory, staging_buffer, staging_size) = create_buffer::<back::Backend>(
        &mut device,
        &memory_properties.memory_types,
        memory::Properties::CPU_VISIBLE | memory::Properties::COHERENT,
        buffer::Usage::TRANSFER_SRC | buffer::Usage::TRANSFER_DST,
        stride,
        numbers.len() as u64,
    );
复制代码

    let (device_memory, device_buffer, _device_buffer_size) = create_buffer::<back::Backend>(
        &mut device,
        &memory_properties.memory_types,
        memory::Properties::DEVICE_LOCAL,
        buffer::Usage::TRANSFER_SRC | buffer::Usage::TRANSFER_DST | buffer::Usage::STORAGE,
        stride,
        numbers.len() as u64,
    );
复制代码

    {
        let mut writer = device.acquire_mapping_writer::<u32>(&staging_memory, 0..staging_size).unwrap();
        writer[0..numbers.len()].copy_from_slice(&numbers);
        device.release_mapping_writer(writer);
    }
复制代码

Metal模块

fn create_buffer(
    &self, size: u64, usage: buffer::Usage
) -> Result<n::UnboundBuffer, buffer::CreationError> {
    debug!("create_buffer of size {} and usage {:?}", size, usage);
    Ok(n::UnboundBuffer {
        size,
        usage,
    })
}
复制代码

fn get_buffer_requirements(&self, buffer: &n::UnboundBuffer) -> memory::Requirements {
    let mut max_size = buffer.size;
    let mut max_alignment = self.private_caps.buffer_alignment;

    if self.private_caps.resource_heaps {
        // We don't know what memory type the user will try to allocate the buffer with, 
        // so we test them all get the most stringent ones.
        for (i, _mt) in self.memory_types.iter().enumerate() {
            let (storage, cache) = MemoryTypes::describe(i);
            let options = conv::resource_options_from_storage_and_cache(storage, cache);
            let requirements = self.shared.device.lock()
                .heap_buffer_size_and_align(buffer.size, options);
            max_size = cmp::max(max_size, requirements.size);
            max_alignment = cmp::max(max_alignment, requirements.align);
        }
    }

    // based on Metal validation error for view creation:
    // failed assertion `BytesPerRow of a buffer-backed texture with pixelFormat(XXX) must be aligned to 256 bytes
    const SIZE_MASK: u64 = 0xFF;
    let supports_texel_view = buffer.usage.intersects(
        buffer::Usage::UNIFORM_TEXEL | buffer::Usage::STORAGE_TEXEL
    );

    memory::Requirements {
        size: (max_size + SIZE_MASK) & !SIZE_MASK,
        alignment: max_alignment,
        type_mask: if !supports_texel_view || self.private_caps.shared_textures {
            MemoryTypes::all().bits()
        } else {
            (MemoryTypes::all() ^ MemoryTypes::SHARED).bits()
        },
    }
}
复制代码

fn allocate_memory(&self, memory_type: hal::MemoryTypeId, size: u64) -> Result<n::Memory, AllocationError> {
    let (storage, cache) = MemoryTypes::describe(memory_type.0);
    let device = self.shared.device.lock();
    debug!("allocate_memory type {:?} of size {}", memory_type, size);

    // Heaps cannot be used for CPU coherent resources
    //TEMP: MacOS supports Private only, iOS and tvOS can do private/shared
    let heap = if self.private_caps.resource_heaps && storage != MTLStorageMode::Shared && false {
        let descriptor = metal::HeapDescriptor::new();
        descriptor.set_storage_mode(storage);
        descriptor.set_cpu_cache_mode(cache);
        descriptor.set_size(size);
        let heap_raw = device.new_heap(&descriptor);
        n::MemoryHeap::Native(heap_raw)
    } else if storage == MTLStorageMode::Private {
        n::MemoryHeap::Private
    } else {
        let options = conv::resource_options_from_storage_and_cache(storage, cache);
        let cpu_buffer = device.new_buffer(size, options);
        debug!("\tbacked by cpu buffer {:?}", cpu_buffer.as_ptr());
        n::MemoryHeap::Public(memory_type, cpu_buffer)
    };

    Ok(n::Memory::new(heap, size))
}
复制代码

fn bind_buffer_memory(
    &self, memory: &n::Memory, offset: u64, buffer: n::UnboundBuffer
) -> Result<n::Buffer, BindError> {
    debug!("bind_buffer_memory of size {} at offset {}", buffer.size, offset);
    let (raw, options, range) = match memory.heap {
        n::MemoryHeap::Native(ref heap) => {
            let resource_options = conv::resource_options_from_storage_and_cache(
                heap.storage_mode(),
                heap.cpu_cache_mode(),
            );
            let raw = heap.new_buffer(buffer.size, resource_options)
                .unwrap_or_else(|| {
                    // TODO: disable hazard tracking?
                    self.shared.device
                        .lock()
                        .new_buffer(buffer.size, resource_options)
                });
            (raw, resource_options, 0 .. buffer.size) //TODO?
        }
        n::MemoryHeap::Public(mt, ref cpu_buffer) => {
            debug!("\tmapped to public heap with address {:?}", cpu_buffer.as_ptr());
            let (storage, cache) = MemoryTypes::describe(mt.0);
            let options = conv::resource_options_from_storage_and_cache(storage, cache);
            (cpu_buffer.clone(), options, offset .. offset + buffer.size)
        }
        n::MemoryHeap::Private => {
            //TODO: check for aliasing
            let options = MTLResourceOptions::StorageModePrivate |
                MTLResourceOptions::CPUCacheModeDefaultCache;
            let raw = self.shared.device
                .lock()
                .new_buffer(buffer.size, options);
            (raw, options, 0 .. buffer.size)
        }
    };

    Ok(n::Buffer {
        raw,
        range,
        options,
    })
}
复制代码

let size = data_source.len() as u64;
let mut data_target = device.acquire_mapping_writer::<T>(&memory, 0..size).unwrap();
data_target[0..data_source.len()].copy_from_slice(data_source);
let _ = device.release_mapping_writer(data_target);
复制代码

/// Acquire a mapping Writer.
///
/// The accessible slice will correspond to the specified range (in bytes).
fn acquire_mapping_writer<'a, T>(
    &self,
    memory: &'a B::Memory,
    range: Range<u64>,
) -> Result<mapping::Writer<'a, B, T>, mapping::Error>
    where
        T: Copy,
{
    let count = (range.end - range.start) as usize / mem::size_of::<T>();
    self.map_memory(memory, range.clone()).map(|ptr| unsafe {
        let start_ptr = ptr as *mut _;
        mapping::Writer {
            slice: slice::from_raw_parts_mut(start_ptr, count),
            memory,
            range,
            released: false,
        }
    })
}
复制代码

fn map_memory<R: RangeArg<u64>>(
    &self, memory: &n::Memory, generic_range: R
) -> Result<*mut u8, mapping::Error> {
    let range = memory.resolve(&generic_range);
    debug!("map_memory of size {} at {:?}", memory.size, range);

    let base_ptr = match memory.heap {
        n::MemoryHeap::Public(_, ref cpu_buffer) => cpu_buffer.contents() as *mut u8,
        n::MemoryHeap::Native(_) |
        n::MemoryHeap::Private => panic!("Unable to map memory!"),
    };
    Ok(unsafe { base_ptr.offset(range.start as _) })
}
复制代码