GpuBuffer read no longer allocate an staging buffer itself. Instead need to manually do it and use new copy function to move data from a buffer to a mapable one

Now GpuArena doesn't do everything.

- Created GpuDeviceAllocator that allocate to the device
- GpuArena become GpuArenaAllocator and use a child_allocator like
std.heap.ArenaAllocator

README.md

@@ -1,26 +1,28 @@
-# Minimal Zig WebGPU Compute Library
+# Minimal Zig WebGPU Compute & Render Library
 
-This is a minimal, self-contained Zig library designed to simplify running compute shaders using WebGPU.
-It abstracts away much of the boilerplate required for GPU device initialization, memory management, and pipeline execution.
+This is a minimal, self-contained Zig library designed to simplify running compute shaders and rendering pipelines using WebGPU. It abstracts away much of the boilerplate required for GPU device initialization, memory management, bind groups, and pipeline execution.
 
 ## Core Modules
 
-The library exports five primary components:
+The library exports the following primary components:
 
-* **`GpuDevice`**: Initializes the WebGPU instance, adapter, device, and queue. It is configured to prioritize high performance and automatically requests the `ShaderF16` feature if the adapter supports it. By default, it enforces a 2 GB VRAM limit.
-* **`GpuArena` / `GpuAllocator`**: A memory management layer that tracks allocated VRAM bytes to prevent exceeding the device budget. The arena automatically destroys and releases all tracked WebGPU buffers when deinitialized.
-* **`GpuBuffer`**: Wraps native WebGPU buffers. It automatically aligns buffer sizes forward to a multiple of 4 bytes. It provides a `.load()` method for CPU-to-GPU data transfers (handling both aligned and unaligned lengths smoothly) and a `.read()` method that utilizes a staging buffer to map GPU data back to the CPU.
-* **`GpuCompute`**: Compiles WGSL source code into a compute pipeline. When running, it automatically splits the work into manageable chunks (up to 1 GB at a time) and dispatches workgroups of size 256.
+* **`GpuDevice`**: Initializes the WebGPU instance, adapter, device, and queue. It is configured to prioritize high performance and automatically requests the `ShaderF16` feature if the adapter supports it. It provides the base `GpuAllocator` for raw VRAM allocations.
+* **`GpuArenaAllocator`**: A memory management layer that wraps a base allocator to track and automatically destroy all allocated WebGPU buffers, textures, views, and pipelines when deinitialized.
+* **`GpuBuffer`**: Wraps native WebGPU buffers. It provides a `.load()` method for CPU-to-GPU data transfers and a `.read()` method that utilizes a staging buffer to map GPU data back to the CPU.
+* **`GpuCompute`**: Compiles WGSL source code into a compute pipeline and dispatches compute workgroups.
+* **`GpuRender` / `GpuTexture` / `GpuTextureView`**: Components used to initialize render pipelines, set up render attachments (textures), and bind render targets for offscreen drawing.
 
-## Quick Start Example
+---
 
-Below is a complete, self-contained example demonstrating how to initialize the GPU, load data, run a compute shader, and read the results back to the CPU:
+## Example 1: Compute Pipeline
+
+Below is a complete example demonstrating how to initialize the GPU via the device allocator, manage VRAM using a GPU Arena, run a compute shader, and read the results back to the CPU:
 
 ```zig
 const std = @import("std");
 const gpu = @import("gpu");
 const GpuDevice = gpu.GpuDevice;
-const GpuArena = gpu.GpuArena;
+const GpuArenaAllocator = gpu.GpuArenaAllocator;
 const GpuBuffer = gpu.GpuBuffer;
 const GpuCompute = gpu.GpuCompute;
 
@@ -32,21 +34,23 @@ pub fn main(init: std.process.Init) !void {
     defer device.deinit();
 
     // 2. Create a GPU Arena to manage VRAM
-    var grena = GpuArena.init(allocator, device);
+    var grena = GpuArenaAllocator.init(allocator, device.gpuAllocator());
     defer grena.deinit();
     const gloc = grena.gpuAllocator();
 
     // 3. Load the WGSL compute pipeline
     const add_cp = try GpuCompute.init(
-        device,
+        gloc,
         @embedFile("shaders/add.wgsl"),
-        .{ .bindings = &.{
-            .{ .element_size = @sizeOf(f16) },
-            .{ .element_size = @sizeOf(f16) },
-            .{ .element_size = @sizeOf(f16) },
-        } },
+        .{
+            .label = "add",
+            .bindings = &.{
+                .{ .element_size = @sizeOf(f16) },
+                .{ .element_size = @sizeOf(f16) },
+                .{ .element_size = @sizeOf(f16) },
+            },
+        },
     );
-    defer add_cp.deinit();
 
     // 4. Setup CPU data
     const len: usize = 16;
@@ -62,12 +66,13 @@ pub fn main(init: std.process.Init) !void {
 
     // 5. Initialize raw GPU Buffers
     const byte_size = len * @sizeOf(f16);
-    const buf_a = try GpuBuffer.init(gloc, byte_size, .initMany(&.{ .Storage, .CopyDst, .CopySrc }));
-    const buf_b = try GpuBuffer.init(gloc, byte_size, .initMany(&.{ .Storage, .CopyDst, .CopySrc }));
-    const buf_out = try GpuBuffer.init(gloc, byte_size, .initMany(&.{ .Storage, .CopyDst, .CopySrc }));
+    const buf_a = try GpuBuffer.init(gloc, .{ .label = "a", .size = byte_size, .usage = .initMany(&.{ .Storage, .CopyDst, .CopySrc }) });
+    const buf_b = try GpuBuffer.init(gloc, .{ .label = "b", .size = byte_size, .usage = .initMany(&.{ .Storage, .CopyDst, .CopySrc }) });
+    const buf_out = try GpuBuffer.init(gloc, .{ .label = "out", .size = byte_size, .usage = .initMany(&.{ .Storage, .CopyDst, .CopySrc }) });
 
-    // Note: The buffers are safely tied to the GpuArena which will automatically
+    // Note: Buffers are safely tied to the GpuArenaAllocator which will automatically
     // release them at the end. You can also manually call buf_x.deinit() if desired.
+    // This will also release pipelines, textures, ect. Everything using a GpuAllocator to init.
 
     // 6. Transfer data from CPU slices to GPU Buffers
     try buf_a.load(f16, data_a);
@@ -84,53 +89,145 @@ pub fn main(init: std.process.Init) !void {
 }
 ```
 
+---
+
+## Example 2: Rendering Pipeline (Offscreen to PPM Image)
+
+This example demonstrates how to initialize a rendering pipeline, allocate an output texture target, draw primitives via WebGPU,
+and pull the frame pixels back to the CPU to write a standard image file:
+
+```zig
+const std = @import("std");
+const gpu = @import("gpu");
+const GpuDevice = gpu.GpuDevice;
+const GpuArenaAllocator = gpu.GpuArenaAllocator;
+const GpuBuffer = gpu.GpuBuffer;
+const GpuRender = gpu.GpuRender;
+const GpuTexture = gpu.GpuTexture;
+const GpuTextureView = gpu.GpuTextureView;
+
+const width: u32 = 512;
+const height: u32 = 512;
+
+pub fn main(init: std.process.Init) !void {
+    const allocator = init.gpa;
+
+    // 1. Open GPU Device
+    const device = try GpuDevice.init(.{});
+    defer device.deinit();
+
+    // 2. Init VRAM Arena
+    var grena = GpuArenaAllocator.init(allocator, device.gpuAllocator());
+    defer grena.deinit();
+    const gloc = grena.gpuAllocator();
+
+    // 3. Load Render Pipeline
+    const circle_rp = try GpuRender.init(
+        gloc,
+        @embedFile("shaders/circle.wgsl"),
+        .{ .bindings = &.{}, .texture_format = .RGBA8Unorm, .topology = .TriangleStrip },
+    );
+    defer circle_rp.deinit();
+
+    // 4. Create VRAM texture to render into
+    const texture = try GpuTexture.init(gloc, .{
+        .format = .RGBA8Unorm,
+        .size = .{ .width = width, .height = height, .depthOrArrayLayers = 1 },
+        .usage = .initMany(&.{ .RenderAttachment, .CopySrc }),
+    });
+    defer texture.deinit();
+
+    // 5. Create a view from texture
+    const view = try GpuTextureView.init(gloc, texture, .{});
+    defer view.deinit();
+
+    // 6. Run the rendering pipeline
+    try circle_rp.draw(gloc, view, 4, .{});
+
+    // 7. Load Texture into GpuBuffer
+    const cpu_staging_cpu = try texture.buffCopy(gloc);
+    defer cpu_staging_cpu.deinit();
+
+    // 8. Read GpuBuffer to CPU
+    // This need to be free manually because CPU memory
+    const pixels = try cpu_staging_cpu.read(allocator, u8);
+    defer allocator.free(pixels);
+
+    // 9. Write a simple ppm image
+    try savePpm(init.io, "circle.ppm", width, height, pixels);
+}
+
+fn savePpm(io: std.Io, filename: []const u8, w: u32, h: u32, rgba_pixels: []const u8) !void {
+    const file = try std.Io.Dir.cwd().createFile(io, filename, .{});
+    defer file.close(io);
+
+    var buf: [255]u8 = undefined;
+    var writer = file.writer(io, &buf);
+
+    // PPM Header: P6 format means raw RGB bytes
+    try writer.interface.print("P6\n{d} {d}\n255\n", .{ w, h });
+
+    // Strip Alpha channel when writing out to standard RGB PPM format
+    var i: usize = 0;
+    while (i < rgba_pixels.len) : (i += 4) {
+        try writer.interface.writeAll(rgba_pixels[i .. i + 3]);
+    }
+}
+```
+
+---
+
+## Running Examples Locally
+
+If you have cloned the repository, you can run the included examples directly using the Zig build system:
+
+```bash
+# Run the rendering example (generates circle.ppm)
+zig build circle
+
+# Run the compute example
+zig build compute
+
+# Run the compute benchmark
+zig build bench_cp
+```
+
+---
+
 ## Dependencies
 
-* **`wgpu.h`**: The library relies on the WebGPU C API headers to bind to the native system graphics.
+* **`wgpu.h`**: The library relies on WebGPU C API headers to bind to the native system graphics.
 
 ## System Requirements
 
-Because this library binds to native system graphics APIs via `wgpu-native`,
-you must ensure the appropriate development headers and libraries are available on your system before compiling.
+Because this library binds to native system graphics APIs via `wgpu-native`, you must ensure the appropriate development headers and libraries are available on your system before compiling.
 
 ### Linux (Vulkan)
-You need the Vulkan development headers to compile the project, and a Vulkan-compatible driver to run it. 
 
-Depending on your distribution, install the following packages:
-* **Ubuntu / Debian:**
-  ```bash
-  sudo apt update
-  sudo apt install libvulkan-dev mesa-vulkan-drivers
-  ```
-* **Fedora / RHEL:**
-  ```bash
-  sudo dnf install vulkan-devel mesa-vulkan-drivers
-  ```
-* **Arch Linux:**
-  ```bash
-  sudo pacman -S vulkan-headers vulkan-icd-loader
-  ```
+* **Ubuntu / Debian:** `sudo apt update && sudo apt install libvulkan-dev mesa-vulkan-drivers`
+* **Fedora / RHEL:** `sudo dnf install vulkan-devel mesa-vulkan-drivers`
+* **Arch Linux:** `sudo pacman -S vulkan-headers vulkan-icd-loader`
 
 ### macOS (Metal)
-No extra installation is required.
-The build script automatically links against the standard Apple frameworks provided by the Xcode Command Line Tools
-(`Metal`, `QuartzCore`, `Foundation`, `CoreGraphics`).
+
+No extra installation required. Automatically links against `Metal`, `QuartzCore`, `Foundation`, and `CoreGraphics`.
 
 ### Windows (DirectX 12)
-No extra installation is required.
-The build script automatically links against the standard Windows SDK libraries (`d3d12`, `dxgi`, `user32`).
-Ensure you have the MSVC build tools installed.
+
+No extra installation required. Automatically links against `d3d12`, `dxgi`, and `user32`. Ensure you have MSVC build tools installed.
 
 ---
 
 ## Adding to your project
 
-To use this module in your own Zig project, add it to your `build.zig.zon`:
+Add it to your `build.zig.zon`:
+
 ```bash
-zig fetch --save git+https://git.bouvais.lu/adrien/zig-wgpu
+zig fetch --save git+[https://git.bouvais.lu/adrien/zig-wgpu#ref=0.2.0](https://git.bouvais.lu/adrien/zig-wgpu)
 ```
 
-Then, in your `build.zig`, import and add the module to your executable:
+Then, expose it in your `build.zig`:
+
 ```zig
 const zig_wgpu = b.dependency("zig-wgpu", .{
     .target = target,

build.zig

@@ -29,66 +29,33 @@ pub fn build(b: *std.Build) !void {
         mod.linkSystemLibrary("gcc_s", .{});
     }
 
-    var threaded: std.Io.Threaded = .init_single_threaded;
-    const io = threaded.io();
+    if (b.pkg_hash.len == 0) {
+        var threaded: std.Io.Threaded = .init_single_threaded;
+        const io = threaded.io();
 
-    var buf: [1024]u8 = undefined;
-    const exemples = try std.Io.Dir.cwd().openDir(io, "examples", .{ .access_sub_paths = false, .iterate = true });
-    var iter = exemples.iterate();
-    while (try iter.next(io)) |entry| {
-        if (entry.kind != .file) continue;
-        if (!std.mem.eql(u8, entry.name[entry.name.len - 4 ..], ".zig")) continue;
+        var buf: [1024]u8 = undefined;
+        const exemples = try std.Io.Dir.cwd().openDir(io, "examples", .{ .access_sub_paths = false, .iterate = true });
+        var iter = exemples.iterate();
+        while (try iter.next(io)) |entry| {
+            if (entry.kind != .file) continue;
+            if (!std.mem.eql(u8, entry.name[entry.name.len - 4 ..], ".zig")) continue;
 
-        const exe = b.addExecutable(.{
-            .name = entry.name[0 .. entry.name.len - 4],
-            .root_module = b.createModule(.{
-                .root_source_file = b.path(try std.fmt.bufPrint(&buf, "examples/{s}", .{entry.name})),
-                .target = target,
-                .optimize = optimize,
-                .imports = &.{},
-            }),
-        });
-        exe.root_module.addImport("gpu", mod);
+            const exe = b.addExecutable(.{
+                .name = entry.name[0 .. entry.name.len - 4],
+                .root_module = b.createModule(.{
+                    .root_source_file = b.path(try std.fmt.bufPrint(&buf, "examples/{s}", .{entry.name})),
+                    .target = target,
+                    .optimize = optimize,
+                    .imports = &.{},
+                }),
+            });
+            exe.root_module.addImport("gpu", mod);
 
-        b.installArtifact(exe);
+            b.installArtifact(exe);
 
-        const run_step = b.step(entry.name[0 .. entry.name.len - 4], try std.fmt.bufPrint(&buf, "Run {s} demo", .{entry.name}));
-        const run_cmd = b.addRunArtifact(exe);
-        run_step.dependOn(&run_cmd.step);
+            const run_step = b.step(entry.name[0 .. entry.name.len - 4], try std.fmt.bufPrint(&buf, "Run {s} demo", .{entry.name}));
+            const run_cmd = b.addRunArtifact(exe);
+            run_step.dependOn(&run_cmd.step);
+        }
     }
-
-    const exe = b.addExecutable(.{
-        .name = "circle",
-        .root_module = b.createModule(.{
-            .root_source_file = b.path("src/circle.zig"),
-            .target = target,
-            .optimize = optimize,
-            .imports = &.{},
-        }),
-    });
-
-    exe.root_module.addIncludePath(b.path("libs/wgpu-native/include"));
-    exe.root_module.addLibraryPath(b.path("libs/wgpu-native/lib"));
-    exe.root_module.addObjectFile(b.path("libs/wgpu-native/lib/libwgpu_native.a"));
-
-    // Platform-specific system frameworks needed by wgpu-native
-    if (t.os.tag == .macos) {
-        exe.root_module.linkFramework("Metal", .{});
-        exe.root_module.linkFramework("QuartzCore", .{});
-        exe.root_module.linkFramework("Foundation", .{});
-        exe.root_module.linkFramework("CoreGraphics", .{});
-    } else if (t.os.tag == .windows) {
-        exe.root_module.linkSystemLibrary("d3d12", .{});
-        exe.root_module.linkSystemLibrary("dxgi", .{});
-        exe.root_module.linkSystemLibrary("user32", .{});
-    } else {
-        exe.root_module.linkSystemLibrary("vulkan", .{});
-        exe.root_module.linkSystemLibrary("gcc_s", .{});
-    }
-
-    b.installArtifact(exe);
-
-    const run_step = b.step("circle", "Run circle");
-    const run_cmd = b.addRunArtifact(exe);
-    run_step.dependOn(&run_cmd.step);
 }

build.zig.zon

@@ -1,6 +1,6 @@
 .{
     .name = .zig_wgpu,
-    .version = "0.1.0",
+    .version = "0.2.0",
     .fingerprint = 0x5d0e853acbc0c2c6,
     .minimum_zig_version = "0.16.0",
     .dependencies = .{},

examples/bench_cp.zig

@@ -1,11 +1,13 @@
 const std = @import("std");
 const gpu = @import("gpu");
 const GpuDevice = gpu.GpuDevice;
-const GpuArena = gpu.GpuArena;
+const GpuArenaAllocator = gpu.GpuArenaAllocator;
 const GpuAllocator = gpu.GpuAllocator;
 const GpuBuffer = gpu.GpuBuffer;
 const GpuCompute = gpu.GpuCompute;
 
+pub const std_options = std.Options{ .log_level = .info };
+
 /// Minimal implementation of a f16 Vector
 const Vec = struct {
     buf: GpuBuffer,
@@ -14,11 +16,10 @@ const Vec = struct {
     // Changed: gloc is passed by value (const)
     pub fn initZero(gloc: GpuAllocator, len: usize) !Vec {
         return .{
-            .buf = try GpuBuffer.init(
-                gloc,
-                len * @sizeOf(f16),
-                .initMany(&.{ .Storage, .CopyDst, .CopySrc }),
-            ),
+            .buf = try GpuBuffer.init(gloc, .{
+                .size = len * @sizeOf(f16),
+                .usage = .initMany(&.{ .Storage, .CopyDst, .CopySrc }),
+            }),
             .len = len,
         };
     }
@@ -60,12 +61,11 @@ pub fn main(init: std.process.Init) !void {
     const device = try GpuDevice.init(.{ .vram_bytes_limit = 4 * 1024 * 1024 * 1024 });
     defer device.deinit();
 
-    var grena = GpuArena.init(init.gpa, device);
+    var grena = GpuArenaAllocator.init(init.gpa, device.gpuAllocator());
     defer grena.deinit();
-
     const gloc = grena.gpuAllocator();
 
-    const add_pip = try GpuCompute.init(device, @embedFile("shaders/add.wgsl"), .{ .bindings = &.{
+    const add_pip = try GpuCompute.init(gloc, @embedFile("shaders/add.wgsl"), .{ .bindings = &.{
         .{ .element_size = @sizeOf(f16) },
         .{ .element_size = @sizeOf(f16) },
         .{ .element_size = @sizeOf(f16) },

examples/circle.zig

@@ -0,0 +1,76 @@
+const std = @import("std");
+const gpu = @import("gpu");
+const GpuDevice = gpu.GpuDevice;
+const GpuArenaAllocator = gpu.GpuArenaAllocator;
+const GpuBuffer = gpu.GpuBuffer;
+const GpuRender = gpu.GpuRender;
+const GpuTexture = gpu.GpuTexture;
+const GpuTextureView = gpu.GpuTextureView;
+
+const width: u32 = 512;
+const height: u32 = 512;
+
+pub fn main(init: std.process.Init) !void {
+    const allocator = init.gpa;
+
+    // 1. Open GPU Device
+    const device = try GpuDevice.init(.{});
+    defer device.deinit();
+
+    // 2. Init VRAM Arena
+    var grena = GpuArenaAllocator.init(allocator, device.gpuAllocator());
+    defer grena.deinit();
+    const gloc = grena.gpuAllocator();
+
+    // 3. Load Render Pipeline
+    const circle_rp = try GpuRender.init(
+        gloc,
+        @embedFile("shaders/circle.wgsl"),
+        .{ .bindings = &.{}, .texture_format = .RGBA8Unorm, .topology = .TriangleStrip },
+    );
+    defer circle_rp.deinit();
+
+    // 4. Create VRAM texture to render into
+    const texture = try GpuTexture.init(gloc, .{
+        .format = .RGBA8Unorm,
+        .size = .{ .width = width, .height = height, .depthOrArrayLayers = 1 },
+        .usage = .initMany(&.{ .RenderAttachment, .CopySrc }),
+    });
+    defer texture.deinit();
+
+    // 5. Create a view from texture
+    const view = try GpuTextureView.init(gloc, texture, .{});
+    defer view.deinit();
+
+    // 6. Run the rendering pipeline
+    try circle_rp.draw(gloc, view, 4, .{});
+
+    // 7. Load Texture into GpuBuffer
+    const cpu_staging_cpu = try texture.buffCopy(gloc);
+    defer cpu_staging_cpu.deinit();
+
+    // 8. Read GpuBuffer to CPU
+    // This need to be free manually because CPU memory
+    const pixels = try cpu_staging_cpu.read(allocator, u8);
+    defer allocator.free(pixels);
+
+    // 9. Write a simple ppm image
+    try savePpm(init.io, "circle.ppm", width, height, pixels);
+}
+
+fn savePpm(io: std.Io, filename: []const u8, w: u32, h: u32, rgba_pixels: []const u8) !void {
+    const file = try std.Io.Dir.cwd().createFile(io, filename, .{});
+    defer file.close(io);
+
+    var buf: [255]u8 = undefined;
+    var writer = file.writer(io, &buf);
+
+    // PPM Header: P6 format means raw RGB bytes
+    try writer.interface.print("P6\n{d} {d}\n255\n", .{ w, h });
+
+    // Strip Alpha channel when writing out to standard RGB PPM format
+    var i: usize = 0;
+    while (i < rgba_pixels.len) : (i += 4) {
+        try writer.interface.writeAll(rgba_pixels[i .. i + 3]);
+    }
+}

examples/compute.zig

@@ -1,7 +1,7 @@
 const std = @import("std");
 const gpu = @import("gpu");
 const GpuDevice = gpu.GpuDevice;
-const GpuArena = gpu.GpuArena;
+const GpuArenaAllocator = gpu.GpuArenaAllocator;
 const GpuBuffer = gpu.GpuBuffer;
 const GpuCompute = gpu.GpuCompute;
 
@@ -13,24 +13,26 @@ pub fn main(init: std.process.Init) !void {
     defer device.deinit();
 
     // 2. Create a GPU Arena to manage VRAM
-    var grena = GpuArena.init(allocator, device);
+    var grena = GpuArenaAllocator.init(allocator, device.gpuAllocator());
     defer grena.deinit();
     const gloc = grena.gpuAllocator();
 
     // 3. Load the WGSL compute pipeline
     const add_cp = try GpuCompute.init(
-        device,
+        gloc,
         @embedFile("shaders/add.wgsl"),
-        .{ .bindings = &.{
-            .{ .element_size = @sizeOf(f16) },
-            .{ .element_size = @sizeOf(f16) },
-            .{ .element_size = @sizeOf(f16) },
-        } },
+        .{
+            .label = "add",
+            .bindings = &.{
+                .{ .element_size = @sizeOf(f16) },
+                .{ .element_size = @sizeOf(f16) },
+                .{ .element_size = @sizeOf(f16) },
+            },
+        },
     );
-    defer add_cp.deinit();
 
     // 4. Setup CPU data
-    const len: usize = 16;
+    const len: usize = 1024;
     const data_a = try allocator.alloc(f16, len);
     defer allocator.free(data_a);
     const data_b = try allocator.alloc(f16, len);
@@ -43,12 +45,13 @@ pub fn main(init: std.process.Init) !void {
 
     // 5. Initialize raw GPU Buffers
     const byte_size = len * @sizeOf(f16);
-    const buf_a = try GpuBuffer.init(gloc, byte_size, .initMany(&.{ .Storage, .CopyDst, .CopySrc }));
-    const buf_b = try GpuBuffer.init(gloc, byte_size, .initMany(&.{ .Storage, .CopyDst, .CopySrc }));
-    const buf_out = try GpuBuffer.init(gloc, byte_size, .initMany(&.{ .Storage, .CopyDst, .CopySrc }));
+    const buf_a = try GpuBuffer.init(gloc, .{ .label = "a", .size = byte_size, .usage = .initMany(&.{ .Storage, .CopyDst, .CopySrc }) });
+    const buf_b = try GpuBuffer.init(gloc, .{ .label = "b", .size = byte_size, .usage = .initMany(&.{ .Storage, .CopyDst, .CopySrc }) });
+    const buf_out = try GpuBuffer.init(gloc, .{ .label = "out", .size = byte_size, .usage = .initMany(&.{ .Storage, .CopyDst, .CopySrc }) });
 
-    // Note: The buffers are safely tied to the GpuArena which will automatically
+    // Note: Buffers are safely tied to the GpuArenaAllocator which will automatically
     // release them at the end. You can also manually call buf_x.deinit() if desired.
+    // This will also release pipelines, textures, ect. Everything using a GpuAllocator to init.
 
     // 6. Transfer data from CPU slices to GPU Buffers
     try buf_a.load(f16, data_a);
@@ -58,8 +61,15 @@ pub fn main(init: std.process.Init) !void {
     try add_cp.run(gloc, .{ buf_a, buf_b, buf_out });
 
     // 8. Map and copy the resulting buffer back to the CPU
-    const out = try buf_out.read(allocator, f16);
+    const staging = try GpuBuffer.init(gloc, .{
+        .size = byte_size,
+        .usage = .initMany(&.{ .MapRead, .CopyDst }),
+    });
+    defer staging.deinit();
+
+    try buf_out.copy(staging);
+    const out = try staging.read(allocator, f16);
     defer allocator.free(out);
 
-    std.debug.print("Result: {any}\n", .{out});
+    std.debug.print("Result: {any}\n", .{out[0..@min(6, len)]});
 }

examples/digit.zig

@@ -1,77 +0,0 @@
-// I am using this mnist reduced dataset https://www.kaggle.com/datasets/mohamedgamal07/reduced-mnist
-
-const std = @import("std");
-const gpu = @import("gpu");
-const GpuDevice = gpu.GpuDevice;
-const GpuArena = gpu.GpuArena;
-const GpuBuffer = gpu.GpuBuffer;
-const GpuProcess = gpu.GpuProcess;
-
-const BATCHSIZE = 10;
-const EPOCH = 10;
-
-pub fn main(init: std.process.Init) !void {
-    const allocator = init.gpa;
-    const io = init.io;
-
-    // 1. Open GPU Device
-    const device = try GpuDevice.init(.{});
-    defer device.deinit();
-
-    // 2. Create a GPU Arena to manage VRAM
-    var grena = GpuArena.init(allocator, device);
-    defer grena.deinit();
-    const gloc = grena.gpuAllocator();
-
-    // 3. Load the WGSL compute pipeline
-    const add_process = try GpuProcess.init(device, @embedFile("shaders/add.wgsl"));
-    defer add_process.deinit();
-
-    var train_dir = try std.Io.Dir.cwd().openDir(io, "mnist/train", .{});
-
-    var images: [BATCHSIZE * 28 * 28]f16 = undefined;
-    for (EPOCH) |epoch| {
-        // Load random images from train dir
-        train_dir.openDir(io, "0", .{});
-        for (BATCHSIZE) |i| {
-            const file = try train_dir.openFile(io, "0.jpg", .{});
-            images[28 * 28 * i .. 28 * 28 * (i + 1)] = file.read
-        }
-    }
-
-    // 4. Setup CPU data
-    const len: usize = 16;
-    const data_a = try allocator.alloc(f16, len);
-    defer allocator.free(data_a);
-    const data_b = try allocator.alloc(f16, len);
-    defer allocator.free(data_b);
-
-    for (0..len) |i| {
-        data_a[i] = @floatFromInt(i);
-        data_b[i] = @floatFromInt(len - 1 - i);
-    }
-
-    // 5. Initialize raw GPU Buffers
-    // We pass the EnumSet inline using `.initMany` since the Enum itself isn't exported
-    const byte_size = len * @sizeOf(f16);
-    const buf_a = try GpuBuffer.init(gloc, byte_size, .initMany(&.{ .Storage, .CopyDst, .CopySrc }));
-    const buf_b = try GpuBuffer.init(gloc, byte_size, .initMany(&.{ .Storage, .CopyDst, .CopySrc }));
-    const buf_out = try GpuBuffer.init(gloc, byte_size, .initMany(&.{ .Storage, .CopyDst, .CopySrc }));
-
-    // Note: The buffers are safely tied to the GpuArena which will automatically
-    // release them at the end. You can also manually call buf_x.deinit() if desired.
-
-    // 6. Transfer data from CPU slices to GPU Buffers
-    try buf_a.load(f16, data_a);
-    try buf_b.load(f16, data_b);
-
-    // 7. Dispatch the Compute Process
-    // We pass the data type (f16) to allow GpuProcess to calculate chunks correctly
-    try add_process.run(gloc, f16, buf_a, buf_b, buf_out);
-
-    // 8. Map and copy the resulting buffer back to the CPU
-    const out = try buf_out.read(allocator, f16);
-    defer allocator.free(out);
-
-    std.debug.print("Result: {any}\n", .{out});
-}

src/GpuAllocator.zig

@@ -6,6 +6,12 @@ pub const VTable = struct {
     freeBuffer: *const fn (ctx: *anyopaque, buf_raw: c.WGPUBuffer) void,
     allocTexture: *const fn (ctx: *anyopaque, desc: c.WGPUTextureDescriptor) anyerror!c.WGPUTexture,
     freeTexture: *const fn (ctx: *anyopaque, buf_raw: c.WGPUTexture) void,
+    allocTextureView: *const fn (ctx: *anyopaque, texture: c.WGPUTexture, desc: c.WGPUTextureViewDescriptor) anyerror!c.WGPUTextureView,
+    freeTextureView: *const fn (ctx: *anyopaque, buf_raw: c.WGPUTextureView) void,
+    allocRenderPipeline: *const fn (ctx: *anyopaque, desc: c.WGPURenderPipelineDescriptor) anyerror!c.WGPURenderPipeline,
+    freeRenderPipeline: *const fn (ctx: *anyopaque, buf_raw: c.WGPURenderPipeline) void,
+    allocComputePipeline: *const fn (ctx: *anyopaque, desc: c.WGPUComputePipelineDescriptor) anyerror!c.WGPUComputePipeline,
+    freeComputePipeline: *const fn (ctx: *anyopaque, buf_raw: c.WGPUComputePipeline) void,
 };
 
 device: GpuDevice,
@@ -16,14 +22,38 @@ pub fn allocBuffer(self: @This(), desc: c.WGPUBufferDescriptor) !c.WGPUBuffer {
     return self.vtable.allocBuffer(self.ptr, desc);
 }
 
-pub fn freeBuffer(self: @This(), buf_raw: c.WGPUBuffer) void {
-    self.vtable.freeBuffer(self.ptr, buf_raw);
+pub fn freeBuffer(self: @This(), raw: c.WGPUBuffer) void {
+    self.vtable.freeBuffer(self.ptr, raw);
 }
 
 pub fn allocTexture(self: @This(), desc: c.WGPUTextureDescriptor) !c.WGPUTexture {
     return self.vtable.allocTexture(self.ptr, desc);
 }
 
-pub fn freeTexture(self: @This(), buf_raw: c.WGPUTexture) void {
-    self.vtable.freeTexture(self.ptr, buf_raw);
+pub fn freeTexture(self: @This(), raw: c.WGPUTexture) void {
+    self.vtable.freeTexture(self.ptr, raw);
+}
+
+pub fn allocTextureView(self: @This(), texture: c.WGPUTexture, desc: c.WGPUTextureViewDescriptor) !c.WGPUTextureView {
+    return self.vtable.allocTextureView(self.ptr, texture, desc);
+}
+
+pub fn freeTextureView(self: @This(), raw: c.WGPUTextureView) void {
+    self.vtable.freeTextureView(self.ptr, raw);
+}
+
+pub fn allocRenderPipeline(self: @This(), desc: c.WGPURenderPipelineDescriptor) !c.WGPURenderPipeline {
+    return self.vtable.allocRenderPipeline(self.ptr, desc);
+}
+
+pub fn freeRenderPipeline(self: @This(), raw: c.WGPURenderPipeline) void {
+    self.vtable.freeRenderPipeline(self.ptr, raw);
+}
+
+pub fn allocComputePipeline(self: @This(), desc: c.WGPUComputePipelineDescriptor) !c.WGPUComputePipeline {
+    return self.vtable.allocComputePipeline(self.ptr, desc);
+}
+
+pub fn freeComputePipeline(self: @This(), raw: c.WGPUComputePipeline) void {
+    self.vtable.freeComputePipeline(self.ptr, raw);
 }

src/GpuArena.zig

@@ -1,107 +0,0 @@
-const std = @import("std");
-const GpuDevice = @import("GpuDevice.zig");
-const GpuAllocator = @import("GpuAllocator.zig");
-const GpuTextureFormat = @import("lib.zig").GpuTextureFormat;
-const c = @import("utils.zig").c;
-
-device: GpuDevice,
-tracked_buffers: std.AutoHashMap(c.WGPUBuffer, c.WGPUBufferDescriptor),
-tracked_textures: std.AutoHashMap(c.WGPUTexture, c.WGPUTextureDescriptor),
-allocated_vram_bytes: u64 = 0,
-
-pub fn init(cpu_allocator: std.mem.Allocator, device: GpuDevice) @This() {
-    return .{
-        .device = device,
-        .tracked_buffers = .init(cpu_allocator),
-        .tracked_textures = .init(cpu_allocator),
-    };
-}
-
-pub fn deinit(self: *@This()) void {
-    var it_buffer = self.tracked_buffers.keyIterator();
-    while (it_buffer.next()) |buf_ptr| {
-        c.wgpuBufferDestroy(buf_ptr.*);
-        c.wgpuBufferRelease(buf_ptr.*);
-    }
-    self.tracked_buffers.deinit();
-
-    var it_texture = self.tracked_textures.keyIterator();
-    while (it_texture.next()) |tex_ptr|
-        c.wgpuTextureRelease(tex_ptr.*);
-    self.tracked_textures.deinit();
-}
-
-/// Returns the type-erased immutable interface wrapper
-pub fn gpuAllocator(self: *@This()) GpuAllocator {
-    return .{
-        .device = self.device,
-        .ptr = self,
-        .vtable = &.{
-            .allocBuffer = allocBuffer,
-            .freeBuffer = freeBuffer,
-            .allocTexture = allocTexture,
-            .freeTexture = freeTexture,
-        },
-    };
-}
-
-fn allocBuffer(ctx: *anyopaque, desc: c.WGPUBufferDescriptor) anyerror!c.WGPUBuffer {
-    const self: *@This() = @ptrCast(@alignCast(ctx));
-
-    if (desc.size > self.device.limits.maxBufferSize)
-        return error.SingleBufferExceedsLimit;
-
-    if (desc.size + self.allocated_vram_bytes > self.device.config.vram_bytes_limit)
-        return error.ExceedsVramBudget;
-
-    const buf = c.wgpuDeviceCreateBuffer(self.device.device, &desc) orelse return error.BufferAlloc;
-    errdefer {
-        c.wgpuBufferDestroy(buf);
-        c.wgpuBufferRelease(buf);
-    }
-
-    try self.tracked_buffers.put(buf, desc);
-    self.allocated_vram_bytes += desc.size;
-    return buf;
-}
-
-fn freeBuffer(ctx: *anyopaque, buf_raw: c.WGPUBuffer) void {
-    const self: *@This() = @ptrCast(@alignCast(ctx));
-
-    if (self.tracked_buffers.fetchRemove(buf_raw)) |kv| {
-        c.wgpuBufferDestroy(buf_raw);
-        c.wgpuBufferRelease(buf_raw);
-        self.allocated_vram_bytes -= kv.value.size;
-    }
-}
-
-fn allocTexture(ctx: *anyopaque, desc: c.WGPUTextureDescriptor) anyerror!c.WGPUTexture {
-    const self: *@This() = @ptrCast(@alignCast(ctx));
-
-    const format: GpuTextureFormat = @enumFromInt(desc.format);
-    const bytes_size = desc.size.width * desc.size.height * format.bytesPerPixel();
-    if (bytes_size > self.device.limits.maxBufferSize)
-        return error.SingleBufferExceedsLimit;
-
-    if (bytes_size + self.allocated_vram_bytes > self.device.config.vram_bytes_limit)
-        return error.ExceedsVramBudget;
-
-    const texture = c.wgpuDeviceCreateTexture(self.device.device, &desc) orelse return error.Texture;
-
-    try self.tracked_textures.put(texture, desc);
-    self.allocated_vram_bytes += bytes_size;
-    return texture;
-}
-
-fn freeTexture(ctx: *anyopaque, texture_raw: c.WGPUTexture) void {
-    const self: *@This() = @ptrCast(@alignCast(ctx));
-
-    if (self.tracked_textures.fetchRemove(texture_raw)) |kv| {
-        c.wgpuTextureRelease(texture_raw);
-
-        const desc = kv.value;
-        const format: GpuTextureFormat = @enumFromInt(desc.format);
-        const bytes_size = desc.size.width * desc.size.height * format.bytesPerPixel();
-        self.allocated_vram_bytes -= bytes_size;
-    }
-}

src/GpuArenaAllocator.zig

@@ -0,0 +1,222 @@
+const std = @import("std");
+const GpuDevice = @import("GpuDevice.zig");
+const GpuAllocator = @import("GpuAllocator.zig");
+const GpuTextureFormat = @import("lib.zig").GpuTextureFormat;
+const c = @import("utils.zig").c;
+const viewStr = @import("utils.zig").viewStr;
+
+child_allocator: GpuAllocator, // I use Zig naming child_allocator, but that should be a parent for me. Likely something idk
+tracked_buffers: std.AutoHashMap(c.WGPUBuffer, c.WGPUBufferDescriptor),
+tracked_textures: std.AutoHashMap(c.WGPUTexture, c.WGPUTextureDescriptor),
+tracked_views: std.AutoHashMap(c.WGPUTextureView, c.WGPUTextureViewDescriptor),
+tracked_renders: std.AutoHashMap(c.WGPURenderPipeline, c.WGPURenderPipelineDescriptor),
+tracked_computes: std.AutoHashMap(c.WGPUComputePipeline, c.WGPUComputePipelineDescriptor),
+allocated_vram_bytes: u64 = 0,
+
+pub fn init(cpu_allocator: std.mem.Allocator, child_allocator: GpuAllocator) @This() {
+    return .{
+        .child_allocator = child_allocator,
+        .tracked_buffers = .init(cpu_allocator),
+        .tracked_textures = .init(cpu_allocator),
+        .tracked_views = .init(cpu_allocator),
+        .tracked_computes = .init(cpu_allocator),
+        .tracked_renders = .init(cpu_allocator),
+    };
+}
+
+pub fn deinit(self: *@This()) void {
+    std.log.debug("Freeing GpuArenaAllocator (Used VRAM: {}/{} MB)", .{
+        self.allocated_vram_bytes / 1024 / 1024,
+        self.child_allocator.device.def.vram_bytes_limit / 1024 / 1024,
+    });
+
+    var it_buffer = self.tracked_buffers.keyIterator();
+    while (it_buffer.next()) |buf_ptr|
+        freeBuffer(self, buf_ptr.*);
+    self.tracked_buffers.deinit();
+
+    var it_tex = self.tracked_textures.keyIterator();
+    while (it_tex.next()) |buf_ptr|
+        freeTexture(self, buf_ptr.*);
+    self.tracked_textures.deinit();
+
+    var it_view = self.tracked_views.keyIterator();
+    while (it_view.next()) |buf_ptr|
+        freeTextureView(self, buf_ptr.*);
+    self.tracked_views.deinit();
+
+    var it_render = self.tracked_renders.keyIterator();
+    while (it_render.next()) |buf_ptr|
+        freeRenderPipeline(self, buf_ptr.*);
+    self.tracked_renders.deinit();
+
+    var it_compute = self.tracked_computes.keyIterator();
+    while (it_compute.next()) |buf_ptr|
+        freeComputePipeline(self, buf_ptr.*);
+    self.tracked_computes.deinit();
+
+    std.log.debug("Freed GpuArenaAllocator (Used VRAM: {}/{} MB)", .{
+        self.allocated_vram_bytes / 1024 / 1024,
+        self.child_allocator.device.def.vram_bytes_limit / 1024 / 1024,
+    });
+}
+
+/// Returns the type-erased immutable interface wrapper
+pub fn gpuAllocator(self: *@This()) GpuAllocator {
+    return .{
+        .device = self.child_allocator.device,
+        .ptr = self,
+        .vtable = &.{
+            .allocBuffer = allocBuffer,
+            .freeBuffer = freeBuffer,
+            .allocTexture = allocTexture,
+            .freeTexture = freeTexture,
+            .allocTextureView = allocTextureView,
+            .freeTextureView = freeTextureView,
+            .allocRenderPipeline = allocRenderPipeline,
+            .freeRenderPipeline = freeRenderPipeline,
+            .allocComputePipeline = allocComputePipeline,
+            .freeComputePipeline = freeComputePipeline,
+        },
+    };
+}
+
+// NOTE: I use ensureTotalCapacity so I know that try self.tracked_x.put will not fail!
+// Like that I dont have to use errdefer to release what I just allocated in VRAM
+
+fn allocBuffer(ctx: *anyopaque, desc: c.WGPUBufferDescriptor) anyerror!c.WGPUBuffer {
+    const self: *@This() = @ptrCast(@alignCast(ctx));
+    try self.tracked_buffers.ensureTotalCapacity(self.tracked_buffers.count() + 1);
+
+    const bytes = desc.size;
+    if (bytes > self.child_allocator.device.limits.maxBufferSize)
+        return error.SingleBufferExceedsLimit;
+
+    if (bytes + self.allocated_vram_bytes > self.child_allocator.device.def.vram_bytes_limit)
+        return error.ExceedsVramBudget;
+
+    const raw = try self.child_allocator.allocBuffer(desc);
+    self.tracked_buffers.putAssumeCapacity(raw, desc);
+    self.allocated_vram_bytes += desc.size;
+
+    std.log.debug("Allocated Buffer '{s}': {d} B (Total VRAM: {}/{} MB)", .{
+        viewStr(desc.label),
+        desc.size,
+        self.allocated_vram_bytes / 1024 / 1024,
+        self.child_allocator.device.def.vram_bytes_limit / 1024 / 1024,
+    });
+
+    return raw;
+}
+
+fn freeBuffer(ctx: *anyopaque, raw: c.WGPUBuffer) void {
+    const self: *@This() = @ptrCast(@alignCast(ctx));
+    if (self.tracked_buffers.fetchRemove(raw)) |kv| {
+        self.child_allocator.freeBuffer(raw);
+        self.allocated_vram_bytes -= kv.value.size;
+
+        std.log.debug("Freed Buffer '{s}' (Total VRAM: {}/{} MB)", .{
+            viewStr(kv.value.label),
+            self.allocated_vram_bytes / 1024 / 1024,
+            self.child_allocator.device.def.vram_bytes_limit / 1024 / 1024,
+        });
+    }
+}
+
+fn allocTexture(ctx: *anyopaque, desc: c.WGPUTextureDescriptor) anyerror!c.WGPUTexture {
+    const self: *@This() = @ptrCast(@alignCast(ctx));
+    try self.tracked_textures.ensureTotalCapacity(self.tracked_textures.count() + 1);
+
+    const format: GpuTextureFormat = @enumFromInt(desc.format);
+    const bytes_size = desc.size.width * desc.size.height * format.bytesPerPixel();
+
+    if (bytes_size + self.allocated_vram_bytes > self.child_allocator.device.def.vram_bytes_limit)
+        return error.ExceedsVramBudget;
+
+    const raw = try self.child_allocator.allocTexture(desc);
+
+    self.tracked_textures.putAssumeCapacity(raw, desc);
+    self.allocated_vram_bytes += bytes_size;
+
+    std.log.debug("Allocated Texture '{s}': {d} B (Total VRAM: {}/{} MB)", .{
+        viewStr(desc.label),
+        bytes_size,
+        self.allocated_vram_bytes / 1024 / 1024,
+        self.child_allocator.device.def.vram_bytes_limit / 1024 / 1024,
+    });
+
+    return raw;
+}
+
+fn freeTexture(ctx: *anyopaque, raw: c.WGPUTexture) void {
+    const self: *@This() = @ptrCast(@alignCast(ctx));
+
+    if (self.tracked_textures.fetchRemove(raw)) |kv| {
+        self.child_allocator.freeTexture(raw);
+
+        const desc = kv.value;
+        const format: GpuTextureFormat = @enumFromInt(desc.format);
+        const bytes_size = desc.size.width * desc.size.height * format.bytesPerPixel();
+        self.allocated_vram_bytes -= bytes_size;
+
+        std.log.debug("Freed Texture '{s}' (Total VRAM: {}/{} MB)", .{
+            viewStr(desc.label),
+            self.allocated_vram_bytes / 1024 / 1024,
+            self.child_allocator.device.def.vram_bytes_limit / 1024 / 1024,
+        });
+    }
+}
+
+fn allocTextureView(ctx: *anyopaque, texture: c.WGPUTexture, desc: c.WGPUTextureViewDescriptor) anyerror!c.WGPUTextureView {
+    const self: *@This() = @ptrCast(@alignCast(ctx));
+    try self.tracked_views.ensureTotalCapacity(self.tracked_views.count() + 1);
+    const raw = try self.child_allocator.allocTextureView(texture, desc);
+    self.tracked_views.putAssumeCapacity(raw, desc);
+    std.log.debug("Allocated Texture View '{s}'", .{viewStr(desc.label)});
+    return raw;
+}
+
+fn freeTextureView(ctx: *anyopaque, raw: c.WGPUTextureView) void {
+    const self: *@This() = @ptrCast(@alignCast(ctx));
+    if (self.tracked_views.fetchRemove(raw)) |kv| {
+        self.child_allocator.freeTextureView(raw);
+        const desc = kv.value;
+        std.log.debug("Freed Texture View '{s}'", .{viewStr(desc.label)});
+    }
+}
+
+fn allocRenderPipeline(ctx: *anyopaque, desc: c.WGPURenderPipelineDescriptor) anyerror!c.WGPURenderPipeline {
+    const self: *@This() = @ptrCast(@alignCast(ctx));
+    try self.tracked_renders.ensureTotalCapacity(self.tracked_renders.count() + 1);
+    const raw = try self.child_allocator.allocRenderPipeline(desc);
+    self.tracked_renders.putAssumeCapacity(raw, desc);
+    std.log.debug("Allocated Render Pipeline '{s}'", .{viewStr(desc.label)});
+    return raw;
+}
+
+fn freeRenderPipeline(ctx: *anyopaque, raw: c.WGPURenderPipeline) void {
+    const self: *@This() = @ptrCast(@alignCast(ctx));
+    if (self.tracked_renders.fetchRemove(raw)) |kv| {
+        self.child_allocator.freeRenderPipeline(raw);
+        const desc = kv.value;
+        std.log.debug("Freed Render Pipeline '{s}'", .{viewStr(desc.label)});
+    }
+}
+
+fn allocComputePipeline(ctx: *anyopaque, desc: c.WGPUComputePipelineDescriptor) anyerror!c.WGPUComputePipeline {
+    const self: *@This() = @ptrCast(@alignCast(ctx));
+    try self.tracked_computes.ensureTotalCapacity(self.tracked_computes.count() + 1);
+    const raw = try self.child_allocator.allocComputePipeline(desc);
+    self.tracked_computes.putAssumeCapacity(raw, desc);
+    std.log.debug("Allocated Compute Pipeline '{s}'", .{viewStr(desc.label)});
+    return raw;
+}
+
+fn freeComputePipeline(ctx: *anyopaque, raw: c.WGPUComputePipeline) void {
+    const self: *@This() = @ptrCast(@alignCast(ctx));
+    if (self.tracked_computes.fetchRemove(raw)) |kv| {
+        self.child_allocator.freeComputePipeline(raw);
+        const desc = kv.value;
+        std.log.debug("Freed Compute Pipeline '{s}'", .{viewStr(desc.label)});
+    }
+}

src/GpuBuffer.zig

@@ -1,13 +1,13 @@
 const std = @import("std");
 const c = @import("utils.zig").c;
 const GpuAllocator = @import("GpuAllocator.zig");
+const svOpt = @import("utils.zig").svOpt;
 
 raw: c.WGPUBuffer,
-size: u64,
-usage: c.WGPUBufferUsage,
 gloc: GpuAllocator,
+def: GpuBufferDef,
 
-const BufferUsage = enum(u64) {
+pub const GpuBufferUsage = enum(u64) {
     None = 0x0000000000000000,
     MapRead = 0x0000000000000001,
     MapWrite = 0x0000000000000002,
@@ -19,37 +19,46 @@ const BufferUsage = enum(u64) {
     Storage = 0x0000000000000080,
     Indirect = 0x0000000000000100,
     QueryResolve = 0x0000000000000200,
+
+    fn enumSetToWGPUBufferUsage(set: std.EnumSet(GpuBufferUsage)) c.WGPUBufferUsage {
+        var use: u64 = 0;
+        var iter = set.iterator();
+        while (iter.next()) |flag| use |= @intFromEnum(flag);
+        return use;
+    }
 };
 
-/// Allocates the underlying WebGPU handle and registers it to the parent GpuAllocator
-pub fn init(gloc: GpuAllocator, size: u64, usage: std.EnumSet(BufferUsage)) !@This() {
-    var use: u64 = 0;
-    var iter = usage.iterator();
-    while (iter.next()) |flag| use |= @intFromEnum(flag);
+pub const GpuBufferDef = struct {
+    label: ?[]const u8 = null,
+    size: u64,
+    usage: std.EnumSet(GpuBufferUsage),
+};
+
+pub fn init(gloc: GpuAllocator, def: GpuBufferDef) !@This() {
 
     // Automatically align the buffer size forward to a multiple of 4 bytes under the hood
-    const aligned_size = std.mem.alignForward(u64, size, 4);
+    const aligned_size = std.mem.alignForward(u64, def.size, 4);
 
-    const raw_handle = try gloc.allocBuffer(.{ .size = aligned_size, .usage = use });
+    const raw_handle = try gloc.allocBuffer(.{
+        .size = aligned_size,
+        .usage = GpuBufferUsage.enumSetToWGPUBufferUsage(def.usage),
+        .label = svOpt(def.label),
+    });
     return .{
         .raw = raw_handle,
-        .size = aligned_size,
-        .usage = use,
+        .def = def,
         .gloc = gloc,
     };
 }
 
-/// Unregisters from the parent GpuAllocator and cleanly destroys GPU resources
 pub fn deinit(self: @This()) void {
     self.gloc.freeBuffer(self.raw);
 }
 
-/// Native getConstMappedRange wrapper
 pub fn getConstMappedRange(self: @This(), offset: u64, size: u64) ?*const anyopaque {
     return c.wgpuBufferGetConstMappedRange(self.raw, offset, size);
 }
 
-/// Native mapAsync wrapper
 pub fn mapAsync(
     self: @This(),
     mode: c.WGPUMapMode,
@@ -60,12 +69,11 @@ pub fn mapAsync(
     _ = c.wgpuBufferMapAsync(self.raw, mode, offset, size, callback_info);
 }
 
-/// Native unmap wrapper
 pub fn unmap(self: @This()) void {
     c.wgpuBufferUnmap(self.raw);
 }
 
-/// CPU to GPU.
+/// CPU to GPU
 pub fn load(
     self: @This(),
     T: type,
@@ -73,9 +81,9 @@ pub fn load(
 ) !void {
     const bytes = data.len * @sizeOf(T);
 
-    if (bytes == self.size) {
+    if (bytes == self.def.size) {
         // Aligned path: direct download
-        c.wgpuQueueWriteBuffer(self.gloc.device.queue, self.raw, 0, data.ptr, self.size);
+        c.wgpuQueueWriteBuffer(self.gloc.device.queue, self.raw, 0, data.ptr, self.def.size);
     } else {
         // Unaligned path: Split the write into an aligned chunk and a padded remainder
         // to support arbitrary lengths without any allocations or large stack arrays.
@@ -92,37 +100,27 @@ pub fn load(
     }
 }
 
+/// GPU to CPU
+/// Buffer must have MapRead usage or returns error.BufferNotMappable.
 pub fn read(self: @This(), alloc: std.mem.Allocator, T: type) ![]T {
-    const out = try alloc.alloc(T, @divExact(self.size, @sizeOf(T)));
+    if (!self.def.usage.contains(.MapRead)) return error.BufferNotMappable;
 
-    const staging = try init(
-        self.gloc,
-        self.size,
-        .initMany(&.{ .MapRead, .CopyDst }),
-    );
-    defer staging.deinit();
-
-    const enc = c.wgpuDeviceCreateCommandEncoder(self.gloc.device.device, null) orelse return error.Encoder;
-    c.wgpuCommandEncoderCopyBufferToBuffer(enc, self.raw, 0, staging.raw, 0, self.size);
-    const cmd = c.wgpuCommandEncoderFinish(enc, null);
-    defer c.wgpuCommandEncoderRelease(enc);
-    defer c.wgpuCommandBufferRelease(cmd);
-    c.wgpuQueueSubmit(self.gloc.device.queue, 1, &cmd);
+    const out = try alloc.alloc(T, @divExact(self.def.size, @sizeOf(T)));
 
     var mapped = false;
-    staging.mapAsync(
+    self.mapAsync(
         c.WGPUMapMode_Read,
         0,
-        self.size,
+        self.def.size,
         .{ .callback = onMapped, .userdata1 = &mapped },
     );
     while (!mapped) self.gloc.device.poll();
 
     const ptr: [*]const T = @ptrCast(@alignCast(
-        staging.getConstMappedRange(0, self.size),
+        self.getConstMappedRange(0, self.def.size),
     ));
     @memcpy(out[0..out.len], ptr[0..out.len]);
-    staging.unmap();
+    self.unmap();
 
     return out;
 }
@@ -136,3 +134,20 @@ fn onMapped(
     const flag: *bool = @ptrCast(@alignCast(userdata1.?));
     flag.* = (status == c.WGPUMapAsyncStatus_Success);
 }
+
+/// GPU to GPU. Both buffers must be same size, src needs CopySrc, dst needs CopyDst.
+pub fn copy(src: @This(), dst: @This()) !void {
+    if (src.def.size != dst.def.size) return error.SizeMismatch;
+
+    const copy_src: u64 = @intFromEnum(GpuBufferUsage.CopySrc);
+    const copy_dst: u64 = @intFromEnum(GpuBufferUsage.CopyDst);
+    if (@as(u64, GpuBufferUsage.enumSetToWGPUBufferUsage(src.def.usage)) & copy_src == 0) return error.SrcNotCopyable;
+    if (@as(u64, GpuBufferUsage.enumSetToWGPUBufferUsage(dst.def.usage)) & copy_dst == 0) return error.DstNotWritable;
+
+    const enc = c.wgpuDeviceCreateCommandEncoder(src.gloc.device.device, null) orelse return error.Encoder;
+    c.wgpuCommandEncoderCopyBufferToBuffer(enc, src.raw, 0, dst.raw, 0, src.def.size);
+    const cmd = c.wgpuCommandEncoderFinish(enc, null);
+    defer c.wgpuCommandEncoderRelease(enc);
+    defer c.wgpuCommandBufferRelease(cmd);
+    c.wgpuQueueSubmit(src.gloc.device.queue, 1, &cmd);
+}

src/GpuCompute.zig

@@ -1,5 +1,6 @@
 const c = @import("utils.zig").c;
 const sv = @import("utils.zig").sv;
+const svOpt = @import("utils.zig").svOpt;
 const GpuAllocator = @import("GpuAllocator.zig");
 const GpuBuffer = @import("GpuBuffer.zig");
 const GpuDevice = @import("GpuDevice.zig");
@@ -11,6 +12,7 @@ pub const Binding = struct {
 };
 
 pub const ComputeDef = struct {
+    label: ?[]const u8 = null,
     bindings: []const Binding,
     workgroup_size: u32 = 256,
     max_workgroups: u32 = 65535,
@@ -20,30 +22,33 @@ pub const ComputeDef = struct {
 };
 
 pip: c.WGPUComputePipeline,
+gloc: GpuAllocator,
 def: ComputeDef,
 
-pub fn init(device: GpuDevice, wgsl: []const u8, def: ComputeDef) !@This() {
+pub fn init(gloc: GpuAllocator, wgsl: []const u8, def: ComputeDef) !@This() {
     var wgsl_src = c.WGPUShaderSourceWGSL{
         .chain = .{ .sType = c.WGPUSType_ShaderSourceWGSL },
         .code = sv(wgsl),
     };
-    const shader = c.wgpuDeviceCreateShaderModule(device.device, &.{
+    const shader = c.wgpuDeviceCreateShaderModule(gloc.device.device, &.{
         .nextInChain = @ptrCast(&wgsl_src),
     }) orelse return error.Shader;
     defer c.wgpuShaderModuleRelease(shader);
 
-    const pip = c.wgpuDeviceCreateComputePipeline(device.device, &.{
+    const pip = try gloc.allocComputePipeline(.{
+        .label = svOpt(def.label),
         .compute = .{ .module = shader, .entryPoint = sv("main") },
-    }) orelse return error.Pipeline;
+    });
 
     return .{
+        .gloc = gloc,
         .pip = pip,
         .def = def,
     };
 }
 
 pub fn deinit(self: @This()) void {
-    c.wgpuComputePipelineRelease(self.pip);
+    self.gloc.freeComputePipeline(self.pip);
 }
 
 /// Execute the compute pass with arbitrary buffer bindings via a tuple.
@@ -69,7 +74,7 @@ pub fn run(
             const buf = @field(args, field.name);
             const el_size = self.def.bindings[i].element_size;
             if (el_size > 0) {
-                elements_count = @intCast(buf.size / el_size);
+                elements_count = @intCast(buf.def.size / el_size);
             }
         }
     }
@@ -80,7 +85,7 @@ pub fn run(
         const el_size = self.def.bindings[i].element_size;
         if (el_size > 0) {
             const expected_min_bytes = @as(u64, elements_count) * el_size;
-            if (buf.size < expected_min_bytes)
+            if (buf.def.size < expected_min_bytes)
                 return error.BufferTooSmall;
         }
     }
@@ -98,7 +103,7 @@ pub fn run(
             .binding = @intCast(i),
             .buffer = buf.raw,
             .offset = 0,
-            .size = buf.size, // Size exposes the fully allocated length
+            .size = buf.def.size, // Size exposes the fully allocated length
         };
         entry_count += 1;
     }
@@ -108,11 +113,11 @@ pub fn run(
     defer if (info_buf) |b| b.deinit();
 
     if (self.def.append_info_buffer) {
-        info_buf = try GpuBuffer.init(
-            gloc,
-            @sizeOf(u32),
-            .initMany(&.{ .Uniform, .CopyDst }),
-        );
+        info_buf = try GpuBuffer.init(gloc, .{
+            .size = @sizeOf(u32),
+            .usage = .initMany(&.{ .Uniform, .CopyDst }),
+            .label = "compute_info_buffer",
+        });
         c.wgpuQueueWriteBuffer(gloc.device.queue, info_buf.?.raw, 0, &elements_count, @sizeOf(u32));
 
         entries_buf[entry_count] = .{

src/GpuDevice.zig

@@ -1,13 +1,17 @@
 const std = @import("std");
 const c = @import("utils.zig").c;
 const sv = @import("utils.zig").sv;
+const svOpt = @import("utils.zig").svOpt;
+const GpuAllocator = @import("GpuAllocator.zig");
+const GpuTextureFormat = @import("lib.zig").GpuTextureFormat;
 
 const Ctx = struct {
     adapter: c.WGPUAdapter = null,
     device: c.WGPUDevice = null,
 };
 
-const GpuDeviceConfig = struct {
+const GpuDeviceDef = struct {
+    label: ?[]const u8 = null,
     /// VRAM limit. Default 2 GB
     vram_bytes_limit: u64 = 2 * 1024 * 1024 * 1024,
     power_preference: enum(c_uint) {
@@ -24,9 +28,9 @@ device: c.WGPUDevice,
 queue: c.WGPUQueue,
 limits: c.WGPULimits,
 
-config: GpuDeviceConfig,
+def: GpuDeviceDef,
 
-pub fn init(config: GpuDeviceConfig) !@This() {
+pub fn init(def: GpuDeviceDef) !@This() {
     const instance = c.wgpuCreateInstance(
         &std.mem.zeroes(c.WGPUInstanceDescriptor),
     ) orelse return error.NoInstance;
@@ -35,13 +39,25 @@ pub fn init(config: GpuDeviceConfig) !@This() {
     var ctx = Ctx{};
     _ = c.wgpuInstanceRequestAdapter(
         instance,
-        &.{ .powerPreference = @intFromEnum(config.power_preference) },
+        &.{ .powerPreference = @intFromEnum(def.power_preference) },
         .{ .callback = onAdapter, .userdata1 = &ctx },
     );
     c.wgpuInstanceProcessEvents(instance);
     const adapter = ctx.adapter orelse return error.NoAdapter;
     errdefer c.wgpuAdapterRelease(adapter);
 
+    var adapter_info = std.mem.zeroes(c.WGPUAdapterInfo);
+    _ = c.wgpuAdapterGetInfo(adapter, &adapter_info);
+
+    std.log.info("=== WebGPU Device Initialized ===", .{});
+    if (adapter_info.device.length > 0 and adapter_info.device.data != null) {
+        std.log.info("  Device Name  : {s}", .{adapter_info.device.data[0..adapter_info.device.length]});
+    }
+    if (adapter_info.architecture.length > 0 and adapter_info.architecture.data != null) {
+        std.log.info("  Architecture : {s}", .{adapter_info.architecture.data[0..adapter_info.architecture.length]});
+    }
+    std.log.info("  Backend Type : {d}", .{adapter_info.backendType});
+
     var supported_features = std.mem.zeroes(c.WGPUSupportedFeatures);
     c.wgpuAdapterGetFeatures(adapter, &supported_features);
 
@@ -49,6 +65,11 @@ pub fn init(config: GpuDeviceConfig) !@This() {
     supported_limits.nextInChain = null;
     if (c.wgpuAdapterGetLimits(adapter, &supported_limits) != 1) return error.FailedToGetAdapterLimits;
 
+    std.log.info("  Max Buf Size : {d} MB", .{supported_limits.maxBufferSize / 1024 / 1024});
+    std.log.info("  Max Storage  : {d} MB", .{supported_limits.maxStorageBufferBindingSize / 1024 / 1024});
+    std.log.info("  Max Workgroup: X: {d}, Y: {d}, Z: {d}", .{ supported_limits.maxComputeWorkgroupSizeX, supported_limits.maxComputeWorkgroupSizeY, supported_limits.maxComputeWorkgroupSizeZ });
+    std.log.info("  VRAM Budget  : {d} MB", .{def.vram_bytes_limit / 1024 / 1024});
+
     var has_f16 = false;
     for (0..supported_features.featureCount) |i| {
         if (supported_features.features[i] == c.WGPUFeatureName_ShaderF16) {
@@ -56,6 +77,8 @@ pub fn init(config: GpuDeviceConfig) !@This() {
             break;
         }
     }
+    std.log.info("  Shader F16   : {}", .{has_f16});
+    std.log.info("=================================", .{});
 
     var feature_buf = [_]c.WGPUFeatureName{c.WGPUFeatureName_ShaderF16};
     const required_features: []const c.WGPUFeatureName =
@@ -63,7 +86,7 @@ pub fn init(config: GpuDeviceConfig) !@This() {
 
     const device_descriptor = c.WGPUDeviceDescriptor{
         .nextInChain = null,
-        .label = sv("TensorCompilerDevice"),
+        .label = svOpt(def.label),
         .requiredFeatureCount = required_features.len,
         .requiredFeatures = if (required_features.len > 0) required_features.ptr else null,
         .requiredLimits = &supported_limits,
@@ -82,7 +105,7 @@ pub fn init(config: GpuDeviceConfig) !@This() {
         .device = device,
         .queue = c.wgpuDeviceGetQueue(device),
         .limits = supported_limits,
-        .config = config,
+        .def = def,
     };
 }
 
@@ -127,3 +150,75 @@ fn onDevice(
     const ctx: *Ctx = @ptrCast(@alignCast(userdata1.?));
     ctx.device = device;
 }
+
+// Allocation stuff
+
+/// Returns the type-erased immutable interface wrapper
+pub fn gpuAllocator(self: *const @This()) GpuAllocator {
+    return .{
+        .device = self.*,
+        .ptr = @ptrCast(@constCast(self)),
+        .vtable = &.{
+            .allocBuffer = allocBuffer,
+            .freeBuffer = freeBuffer,
+            .allocTexture = allocTexture,
+            .freeTexture = freeTexture,
+            .allocTextureView = allocTextureView,
+            .freeTextureView = freeTextureView,
+            .allocRenderPipeline = allocRenderPipeline,
+            .freeRenderPipeline = freeRenderPipeline,
+            .allocComputePipeline = allocComputePipeline,
+            .freeComputePipeline = freeComputePipeline,
+        },
+    };
+}
+
+fn allocBuffer(ctx: *anyopaque, desc: c.WGPUBufferDescriptor) anyerror!c.WGPUBuffer {
+    const self: *@This() = @ptrCast(@alignCast(ctx));
+    if (desc.size > self.limits.maxBufferSize)
+        return error.SingleBufferExceedsLimit;
+    return c.wgpuDeviceCreateBuffer(self.device, &desc) orelse return error.BufferAlloc;
+}
+
+fn freeBuffer(_: *anyopaque, raw: c.WGPUBuffer) void {
+    c.wgpuBufferDestroy(raw);
+    c.wgpuBufferRelease(raw);
+}
+
+fn allocTexture(ctx: *anyopaque, desc: c.WGPUTextureDescriptor) anyerror!c.WGPUTexture {
+    const self: *@This() = @ptrCast(@alignCast(ctx));
+    const format: GpuTextureFormat = @enumFromInt(desc.format);
+    if (desc.size.width * desc.size.height * format.bytesPerPixel() > self.limits.maxBufferSize)
+        return error.SingleBufferExceedsLimit;
+    return c.wgpuDeviceCreateTexture(self.device, &desc) orelse return error.Texture;
+}
+
+fn freeTexture(_: *anyopaque, raw: c.WGPUTexture) void {
+    c.wgpuTextureRelease(raw);
+}
+
+fn allocTextureView(_: *anyopaque, texture: c.WGPUTexture, desc: c.WGPUTextureViewDescriptor) anyerror!c.WGPUTextureView {
+    return c.wgpuTextureCreateView(texture, &desc) orelse return error.View;
+}
+
+fn freeTextureView(_: *anyopaque, raw: c.WGPUTextureView) void {
+    c.wgpuTextureViewRelease(raw);
+}
+
+fn allocRenderPipeline(ctx: *anyopaque, desc: c.WGPURenderPipelineDescriptor) anyerror!c.WGPURenderPipeline {
+    const self: *@This() = @ptrCast(@alignCast(ctx));
+    return c.wgpuDeviceCreateRenderPipeline(self.device, &desc) orelse return error.Pipeline;
+}
+
+fn freeRenderPipeline(_: *anyopaque, raw: c.WGPURenderPipeline) void {
+    c.wgpuRenderPipelineRelease(raw);
+}
+
+fn allocComputePipeline(ctx: *anyopaque, desc: c.WGPUComputePipelineDescriptor) anyerror!c.WGPUComputePipeline {
+    const self: *@This() = @ptrCast(@alignCast(ctx));
+    return c.wgpuDeviceCreateComputePipeline(self.device, &desc) orelse return error.Pipeline;
+}
+
+fn freeComputePipeline(_: *anyopaque, raw: c.WGPUComputePipeline) void {
+    c.wgpuComputePipelineRelease(raw);
+}

src/GpuRender.zig

@@ -1,16 +1,19 @@
 const std = @import("std");
 const c = @import("utils.zig").c;
 const sv = @import("utils.zig").sv;
+const svOpt = @import("utils.zig").svOpt;
 const GpuAllocator = @import("GpuAllocator.zig");
 const GpuBuffer = @import("GpuBuffer.zig");
 const GpuDevice = @import("GpuDevice.zig");
+const GpuTextureView = @import("GpuTextureView.zig");
 const GpuTextureFormat = @import("lib.zig").GpuTextureFormat;
 
 pub const Binding = struct {
     element_size: u32 = 0,
 };
 
-pub const RenderDef = struct {
+pub const GpuRenderDef = struct {
+    label: ?[]const u8 = null,
     bindings: []const Binding = &.{},
     /// The surface texture format we are rendering to (e.g., BGRA8Unorm)
     texture_format: GpuTextureFormat,
@@ -31,15 +34,16 @@ const GpuPrimitiveTopology = enum(c_uint) {
     Force32 = 0x7FFFFFFF,
 };
 
+gloc: GpuAllocator,
 pip: c.WGPURenderPipeline,
-def: RenderDef,
+def: GpuRenderDef,
 
-pub fn init(device: GpuDevice, wgsl: []const u8, def: RenderDef) !@This() {
+pub fn init(gloc: GpuAllocator, wgsl: []const u8, def: GpuRenderDef) !@This() {
     var wgsl_src = c.WGPUShaderSourceWGSL{
         .chain = .{ .sType = c.WGPUSType_ShaderSourceWGSL },
         .code = sv(wgsl),
     };
-    const shader = c.wgpuDeviceCreateShaderModule(device.device, &.{
+    const shader = c.wgpuDeviceCreateShaderModule(gloc.device.device, &.{
         .nextInChain = @ptrCast(&wgsl_src),
     }) orelse return error.Shader;
     defer c.wgpuShaderModuleRelease(shader);
@@ -65,7 +69,8 @@ pub fn init(device: GpuDevice, wgsl: []const u8, def: RenderDef) !@This() {
     };
 
     // 3. Compile the Complete Render Pipeline
-    const pip = c.wgpuDeviceCreateRenderPipeline(device.device, &.{
+    const pip = try gloc.allocRenderPipeline(.{
+        .label = svOpt(def.label),
         .vertex = .{
             .module = shader,
             .entryPoint = sv(def.vertex_entry),
@@ -82,16 +87,17 @@ pub fn init(device: GpuDevice, wgsl: []const u8, def: RenderDef) !@This() {
             .alphaToCoverageEnabled = 0,
         },
         .fragment = &fragment_state,
-    }) orelse return error.Pipeline;
+    });
 
     return .{
+        .gloc = gloc,
         .pip = pip,
         .def = def,
     };
 }
 
 pub fn deinit(self: @This()) void {
-    c.wgpuRenderPipelineRelease(self.pip);
+    self.gloc.freeRenderPipeline(self.pip);
 }
 
 /// Execute the render pass targeting a specific frame texture view.
@@ -99,7 +105,7 @@ pub fn deinit(self: @This()) void {
 pub fn draw(
     self: @This(),
     gloc: GpuAllocator,
-    target_view: c.WGPUTextureView,
+    target_view: GpuTextureView,
     vertex_count: u32,
     args: anytype,
 ) !void {
@@ -144,7 +150,7 @@ pub fn draw(
     defer c.wgpuCommandEncoderRelease(enc);
 
     const color_attachment = c.WGPURenderPassColorAttachment{
-        .view = target_view,
+        .view = target_view.raw,
         .resolveTarget = null,
         .loadOp = c.WGPULoadOp_Clear,
         .storeOp = c.WGPUStoreOp_Store,

src/GpuTexture.zig

@@ -1,54 +1,92 @@
 const std = @import("std");
 const c = @import("utils.zig").c;
+const svOpt = @import("utils.zig").svOpt;
 const GpuAllocator = @import("GpuAllocator.zig");
+const GpuBuffer = @import("GpuBuffer.zig");
 const GpuTextureFormat = @import("lib.zig").GpuTextureFormat;
+const GpuTextureUsage = @import("lib.zig").GpuTextureUsage;
 
-const TextureUsage = enum(u64) {
-    None = 0x0000000000000000,
-    CopySrc = 0x0000000000000001,
-    CopyDst = 0x0000000000000002,
-    TextureBinding = 0x0000000000000004,
-    StorageBinding = 0x0000000000000008,
-    RenderAttachment = 0x0000000000000010,
-    TransientAttachment = 0x0000000000000020,
+pub const GpuTextureDef = struct {
+    label: ?[]const u8 = null,
+    size: c.WGPUExtent3D,
+    usage: std.EnumSet(GpuTextureUsage),
+    format: GpuTextureFormat,
 };
 
 raw: c.WGPUTexture,
-size: c.WGPUExtent3D,
-usage: c.WGPUTextureUsage,
-format: GpuTextureFormat,
 gloc: GpuAllocator,
+def: GpuTextureDef,
 
-/// Allocates the underlying WebGPU handle and registers it to the parent GpuAllocator
-pub fn init(gloc: GpuAllocator, format: GpuTextureFormat, size: c.WGPUExtent3D, usage: std.EnumSet(TextureUsage)) !@This() {
+pub fn init(gloc: GpuAllocator, def: GpuTextureDef) !@This() {
     var use: u64 = 0;
-    var iter = usage.iterator();
+    var iter = def.usage.iterator();
     while (iter.next()) |flag| use |= @intFromEnum(flag);
 
     const desc = c.WGPUTextureDescriptor{
+        .label = svOpt(def.label),
         .usage = use,
         .dimension = c.WGPUTextureDimension_2D,
-        .size = size,
-        .format = @intFromEnum(format),
+        .size = def.size,
+        .format = @intFromEnum(def.format),
         .mipLevelCount = 1,
         .sampleCount = 1,
     };
     const raw = try gloc.allocTexture(desc);
 
-    return .{ .gloc = gloc, .raw = raw, .size = size, .format = format, .usage = use };
+    return .{ .gloc = gloc, .raw = raw, .def = def };
 }
 
-/// Unregisters from the parent GpuAllocator and cleanly destroys GPU resources
 pub fn deinit(self: @This()) void {
     self.gloc.freeTexture(self.raw);
 }
 
-/// Native getConstMappedRange wrapper
 pub fn getConstMappedRange(self: @This(), offset: u64, size: u64) ?*const anyopaque {
     return c.wgpuBufferGetConstMappedRange(self.raw, offset, size);
 }
 
-/// Native mapAsync wrapper
+pub fn bytesSize(self: @This()) u32 {
+    return self.bytesSizeRow() * self.def.size.height;
+}
+
+pub fn bytesSizeRow(self: @This()) u32 {
+    return self.def.size.width * self.def.format.bytesPerPixel();
+}
+
+/// Return a GpuBuffer containing a copy of the texture.
+pub fn buffCopy(self: @This(), gloc: GpuAllocator) !GpuBuffer {
+    const buf = try GpuBuffer.init(gloc, .{
+        .size = self.bytesSize(),
+        .usage = .initMany(&.{ .CopyDst, .CopySrc }),
+        .label = "texture_copy_buffer",
+    });
+
+    const enc = c.wgpuDeviceCreateCommandEncoder(gloc.device.device, null) orelse return error.Encoder;
+    defer c.wgpuCommandEncoderRelease(enc);
+
+    const src_copy = c.WGPUTexelCopyTextureInfo{
+        .texture = self.raw,
+        .mipLevel = 0,
+        .origin = .{ .x = 0, .y = 0, .z = 0 },
+        .aspect = c.WGPUTextureAspect_All,
+    };
+    const dst_copy = c.WGPUTexelCopyBufferInfo{
+        .buffer = buf.raw,
+        .layout = .{
+            .offset = 0,
+            .bytesPerRow = self.bytesSizeRow(),
+            .rowsPerImage = self.def.size.height,
+        },
+    };
+
+    c.wgpuCommandEncoderCopyTextureToBuffer(enc, &src_copy, &dst_copy, &self.def.size);
+
+    const cmd = c.wgpuCommandEncoderFinish(enc, null);
+    defer c.wgpuCommandBufferRelease(cmd);
+    c.wgpuQueueSubmit(gloc.device.queue, 1, &cmd);
+
+    return buf;
+}
+
 pub fn mapAsync(
     self: @This(),
     mode: c.WGPUMapMode,
@@ -59,12 +97,11 @@ pub fn mapAsync(
     _ = c.wgpuBufferMapAsync(self.raw, mode, offset, size, callback_info);
 }
 
-/// Native unmap wrapper
 pub fn unmap(self: @This()) void {
     c.wgpuBufferUnmap(self.raw);
 }
 
-/// CPU to GPU.
+/// CPU to GPU
 pub fn load(
     self: @This(),
     T: type,
@@ -72,37 +109,52 @@ pub fn load(
 ) !void {
     const bytes = data.len * @sizeOf(T);
 
-    if (bytes == self.size) {
-        // Aligned path: direct download
-        c.wgpuQueueWriteBuffer(self.gloc.device.queue, self.raw, 0, data.ptr, self.size);
-    } else {
-        // Unaligned path: Split the write into an aligned chunk and a padded remainder
-        // to support arbitrary lengths without any allocations or large stack arrays.
-        const aligned_part = (bytes / 4) * 4;
-        if (aligned_part > 0) {
-            c.wgpuQueueWriteBuffer(self.gloc.device.queue, self.raw, 0, data.ptr, aligned_part);
-        }
-
-        var remainder_buf: [4]u8 = .{ 0, 0, 0, 0 };
-        const data_bytes = std.mem.sliceAsBytes(data);
-        @memcpy(remainder_buf[0 .. bytes - aligned_part], data_bytes[aligned_part..bytes]);
-
-        c.wgpuQueueWriteBuffer(self.gloc.device.queue, self.raw, aligned_part, &remainder_buf, 4);
-    }
+    c.wgpuQueueWriteTexture(
+        self.gloc.device.queue,
+        &.{
+            .texture = self.raw,
+            .mipLevel = 0,
+            .origin = .{ .x = 0, .y = 0, .z = 0 },
+            .aspect = c.WGPUTextureAspect_All,
+        },
+        data.ptr,
+        bytes,
+        &.{
+            .offset = 0,
+            .bytesPerRow = self.bytesSizeRow(),
+            .rowsPerImage = self.def.size.height,
+        },
+        &self.def.size,
+    );
 }
 
+// GPU to CPU
 pub fn read(self: @This(), alloc: std.mem.Allocator, T: type) ![]T {
     const out = try alloc.alloc(T, @divExact(self.size, @sizeOf(T)));
 
-    const staging = try init(
-        self.gloc,
-        self.size,
-        .initMany(&.{ .MapRead, .CopyDst }),
-    );
+    const staging = try init(self.gloc, .{
+        .size = self.size,
+        .usage = .initMany(&.{ .MapRead, .CopyDst }),
+        .label = "texture_read_staging",
+    });
     defer staging.deinit();
 
     const enc = c.wgpuDeviceCreateCommandEncoder(self.gloc.device.device, null) orelse return error.Encoder;
-    c.wgpuCommandEncoderCopyBufferToBuffer(enc, self.raw, 0, staging.raw, 0, self.size);
+    const src_copy = c.WGPUTexelCopyTextureInfo{
+        .texture = self.raw,
+        .mipLevel = 0,
+        .origin = .{ .x = 0, .y = 0, .z = 0 },
+        .aspect = c.WGPUTextureAspect_All,
+    };
+    const dst_copy = c.WGPUTexelCopyBufferInfo{
+        .buffer = staging.raw,
+        .layout = .{
+            .offset = 0,
+            .bytesPerRow = self.bytesSizeRow(),
+            .rowsPerImage = self.def.size.height,
+        },
+    };
+    c.wgpuCommandEncoderCopyTextureToBuffer(enc, &src_copy, &dst_copy, &self.def.size);
     const cmd = c.wgpuCommandEncoderFinish(enc, null);
     defer c.wgpuCommandEncoderRelease(enc);
     defer c.wgpuCommandBufferRelease(cmd);

src/GpuTextureView.zig

@@ -0,0 +1,35 @@
+const std = @import("std");
+const c = @import("utils.zig").c;
+const svOpt = @import("utils.zig").svOpt;
+const GpuAllocator = @import("GpuAllocator.zig");
+const GpuTexture = @import("lib.zig").GpuTexture;
+const GpuTextureFormat = @import("lib.zig").GpuTextureFormat;
+const GpuTextureUsage = @import("lib.zig").GpuTextureUsage;
+
+pub const GpuViewDef = struct {
+    label: ?[]const u8 = null,
+    usage: std.EnumSet(GpuTextureUsage) = .empty,
+    format: GpuTextureFormat = .Undefined,
+};
+
+raw: c.WGPUTextureView,
+gloc: GpuAllocator,
+
+pub fn init(gloc: GpuAllocator, texture: GpuTexture, def: GpuViewDef) !@This() {
+    var use: u64 = 0;
+    var iter = def.usage.iterator();
+    while (iter.next()) |flag| use |= @intFromEnum(flag);
+
+    const raw = try gloc.allocTextureView(texture.raw, .{
+        .label = svOpt(def.label),
+        .format = @intFromEnum(def.format),
+        .usage = use,
+        .mipLevelCount = 1,
+        .arrayLayerCount = 1,
+    });
+    return .{ .gloc = gloc, .raw = raw };
+}
+
+pub fn deinit(self: @This()) void {
+    self.gloc.freeTextureView(self.raw);
+}

src/circle.zig

@@ -1,112 +0,0 @@
-const std = @import("std");
-const gpu = @import("lib.zig");
-const c = @import("utils.zig").c;
-const sv = @import("utils.zig").sv;
-const GpuDevice = gpu.GpuDevice;
-const GpuArena = gpu.GpuArena;
-const GpuBuffer = gpu.GpuBuffer;
-const GpuRender = gpu.GpuRender;
-const GpuTexture = gpu.GpuTexture;
-
-pub fn main(init: std.process.Init) !void {
-    const allocator = init.gpa;
-
-    // 1. Open the raw headless GPU Device you shared
-    const device = try GpuDevice.init(.{});
-    defer device.deinit();
-
-    var grena = GpuArena.init(allocator, device);
-    defer grena.deinit();
-    const gloc = grena.gpuAllocator();
-
-    const width: u32 = 512;
-    const height: u32 = 512;
-
-    // 2. Load our Render Pipeline (Procedural Triangle Strip)
-    const circle_rp = try GpuRender.init(
-        device,
-        @embedFile("shaders/circle.wgsl"),
-        .{
-            .bindings = &.{},
-            .texture_format = .RGBA8Unorm,
-            .topology = .TriangleStrip,
-        },
-    );
-    defer circle_rp.deinit();
-
-    // 3. Create the offscreen VRAM texture to render into
-    const texture = try GpuTexture.init(
-        gloc,
-        .RGBA8Unorm,
-        .{ .width = width, .height = height, .depthOrArrayLayers = 1 },
-        .initMany(&.{ .RenderAttachment, .CopySrc }),
-    );
-    defer texture.deinit();
-
-    const target_view = c.wgpuTextureCreateView(texture.raw, null) orelse return error.View;
-    defer c.wgpuTextureViewRelease(target_view);
-
-    // 4. Create a staging buffer to pull pixels from VRAM to CPU
-    // 4 bytes per pixel (RGBA8)
-    const row_bytes = width * 4;
-    const buffer_bytes = row_bytes * height;
-
-    // Create a regular GpuBuffer set up to receive texture copy transfers
-    const cpu_staging_buf = try GpuBuffer.init(gloc, buffer_bytes, .initMany(&.{ .CopyDst, .CopySrc }));
-
-    // 5. Draw the Circle Frame into the texture view!
-    try circle_rp.draw(gloc, target_view, 4, .{});
-
-    // 6. Copy the texture data into our CPU staging buffer
-    const enc = c.wgpuDeviceCreateCommandEncoder(device.device, null) orelse return error.Encoder;
-    defer c.wgpuCommandEncoderRelease(enc);
-
-    const src_copy = c.WGPUTexelCopyTextureInfo{
-        .texture = texture.raw,
-        .mipLevel = 0,
-        .origin = .{ .x = 0, .y = 0, .z = 0 },
-        .aspect = c.WGPUTextureAspect_All,
-    };
-    const dst_copy = c.WGPUTexelCopyBufferInfo{
-        .buffer = cpu_staging_buf.raw,
-        .layout = .{
-            .offset = 0,
-            .bytesPerRow = row_bytes,
-            .rowsPerImage = height,
-        },
-    };
-    const copy_size = c.WGPUExtent3D{ .width = width, .height = height, .depthOrArrayLayers = 1 };
-
-    c.wgpuCommandEncoderCopyTextureToBuffer(enc, &src_copy, &dst_copy, &copy_size);
-
-    const cmd = c.wgpuCommandEncoderFinish(enc, null);
-    defer c.wgpuCommandBufferRelease(cmd);
-    c.wgpuQueueSubmit(device.queue, 1, &cmd);
-
-    // 7. Map and read the raw image bytes back to CPU
-    // (This uses whatever slice-reading helpers your `GpuBuffer` wrapper provides)
-    const pixels = try cpu_staging_buf.read(allocator, u8);
-    defer allocator.free(pixels);
-
-    // Now you have the raw binary image data! Let's output a simple Netpbm PPM image file
-    // so you can actually open and look at your rendered circle.
-    try savePpm(init.io, "circle.ppm", width, height, pixels);
-    std.debug.print("Successfully rendered circle to circle.ppm!\n", .{});
-}
-
-fn savePpm(io: std.Io, filename: []const u8, w: u32, h: u32, rgba_pixels: []const u8) !void {
-    const file = try std.Io.Dir.cwd().createFile(io, filename, .{});
-    defer file.close(io);
-
-    var buf: [255]u8 = undefined;
-    var writer = file.writer(io, &buf);
-
-    // PPM Header: P6 format means raw RGB bytes
-    try writer.interface.print("P6\n{d} {d}\n255\n", .{ w, h });
-
-    // Strip Alpha channel when writing out to standard RGB PPM format
-    var i: usize = 0;
-    while (i < rgba_pixels.len) : (i += 4) {
-        try writer.interface.writeAll(rgba_pixels[i .. i + 3]);
-    }
-}

src/lib.zig

@@ -1,10 +1,11 @@
 pub const GpuAllocator = @import("GpuAllocator.zig");
-pub const GpuArena = @import("GpuArena.zig");
+pub const GpuArenaAllocator = @import("GpuArenaAllocator.zig");
 pub const GpuBuffer = @import("GpuBuffer.zig");
 pub const GpuDevice = @import("GpuDevice.zig");
 pub const GpuCompute = @import("GpuCompute.zig");
 pub const GpuRender = @import("GpuRender.zig");
 pub const GpuTexture = @import("GpuTexture.zig");
+pub const GpuTextureView = @import("GpuTextureView.zig");
 
 pub const GpuTextureFormat = enum(c_uint) {
     Undefined = 0,
@@ -174,3 +175,13 @@ pub const GpuTextureFormat = enum(c_uint) {
         };
     }
 };
+
+pub const GpuTextureUsage = enum(u64) {
+    None = 0x0000000000000000,
+    CopySrc = 0x0000000000000001,
+    CopyDst = 0x0000000000000002,
+    TextureBinding = 0x0000000000000004,
+    StorageBinding = 0x0000000000000008,
+    RenderAttachment = 0x0000000000000010,
+    TransientAttachment = 0x0000000000000020,
+};

src/shaders/add.wgsl

@@ -1,24 +0,0 @@
-enable f16;
-
-@group(0) @binding(0) var<storage, read> A: array<f16>;
-@group(0) @binding(1) var<storage, read> B: array<f16>;
-@group(0) @binding(2) var<storage, read_write> C: array<f16>;
-@group(0) @binding(3) var<uniform> size: u32; 
-
-@compute @workgroup_size(256)
-fn main(
-    @builtin(global_invocation_id) global_id : vec3<u32>,
-    @builtin(num_workgroups) num_workgroups: vec3<u32>
-) {
-    // 1. Calculate the total number of threads across the entire grid
-    let total_threads = num_workgroups.x * 256u; 
-    
-    // 2. Start at this thread's unique global ID
-    var index = global_id.x;
-    
-    // 3. Stride through the tensor elements
-    while (index < size) {
-        C[index] = A[index] + B[index];
-        index += total_threads; // Jump forward by the total thread count
-    }
-}

src/utils.zig

@@ -3,3 +3,17 @@ pub const c = @cImport(@cInclude("wgpu.h"));
 pub fn sv(s: []const u8) c.WGPUStringView {
     return .{ .data = s.ptr, .length = s.len };
 }
+
+/// Allows safely passing an optional Zig string to a WebGPU string view.
+pub fn svOpt(s: ?[]const u8) c.WGPUStringView {
+    if (s) |str| return sv(str);
+    return .{ .data = null, .length = 0 };
+}
+
+/// Helper to print a WGPUStringView in your logs.
+pub fn viewStr(view: c.WGPUStringView) []const u8 {
+    if (view.data != null and view.length > 0) {
+        return view.data[0..view.length];
+    }
+    return "unnamed";
+}