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

README.md

@@ -42,11 +42,14 @@ pub fn main(init: std.process.Init) !void {
     const add_cp = try GpuCompute.init(
         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) },
+            },
+        },
     );
 
     // 4. Setup CPU data
@@ -63,12 +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: Buffers, pipelines, and other objects initialized with 'gloc' 
-    // are safely tied to the GpuArenaAllocator and will automatically release.
+    // 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);
@@ -96,6 +100,7 @@ and pull the frame pixels back to the CPU to write a standard image file:
 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;
@@ -111,8 +116,10 @@ pub fn main(init: std.process.Init) !void {
     const device = try GpuDevice.init(.{});
     defer device.deinit();
 
-    // 2. Get base device GPU Allocator
-    const gloc = device.gpuAllocator();
+    // 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(
@@ -137,15 +144,16 @@ pub fn main(init: std.process.Init) !void {
     // 6. Run the rendering pipeline
     try circle_rp.draw(gloc, view, 4, .{});
 
-    // 7. Copy Texture into a readable GPU staging buffer
-    const cpu_staging_buf = try texture.buffCopy(gloc);
-    defer cpu_staging_buf.deinit();
+    // 7. Load Texture into GpuBuffer
+    const cpu_staging_cpu = try texture.buffCopy(gloc);
+    defer cpu_staging_cpu.deinit();
 
-    // 8. Read GpuBuffer to CPU memory
-    const pixels = try cpu_staging_buf.read(allocator, u8);
+    // 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 out to a simple PPM image
+    // 9. Write a simple ppm image
     try savePpm(init.io, "circle.ppm", width, height, pixels);
 }
 

examples/compute.zig

@@ -32,7 +32,7 @@ pub fn main(init: std.process.Init) !void {
     );
 
     // 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);
@@ -61,8 +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)]});
 }

src/GpuArenaAllocator.zig

@@ -87,6 +87,14 @@ pub fn gpuAllocator(self: *@This()) GpuAllocator {
 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;

src/GpuBuffer.zig

@@ -4,11 +4,10 @@ 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,
@@ -20,31 +19,34 @@ 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;
+    }
 };
 
-const GpuBufferDef = struct {
+pub const GpuBufferDef = struct {
     label: ?[]const u8 = null,
     size: u64,
-    usage: std.EnumSet(BufferUsage),
+    usage: std.EnumSet(GpuBufferUsage),
 };
 
 pub fn init(gloc: GpuAllocator, def: GpuBufferDef) !@This() {
-    var use: u64 = 0;
-    var iter = def.usage.iterator();
-    while (iter.next()) |flag| use |= @intFromEnum(flag);
 
     // Automatically align the buffer size forward to a multiple of 4 bytes under the hood
     const aligned_size = std.mem.alignForward(u64, def.size, 4);
 
     const raw_handle = try gloc.allocBuffer(.{
         .size = aligned_size,
-        .usage = use,
+        .usage = GpuBufferUsage.enumSetToWGPUBufferUsage(def.usage),
         .label = svOpt(def.label),
     });
     return .{
         .raw = raw_handle,
-        .size = aligned_size,
-        .usage = use,
+        .def = def,
         .gloc = gloc,
     };
 }
@@ -79,9 +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.
@@ -99,37 +101,26 @@ 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, .{
-        .size = self.size,
-        .usage = .initMany(&.{ .MapRead, .CopyDst }),
-        .label = "staging_read_buffer",
-    });
-    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;
 }
@@ -143,3 +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

@@ -74,7 +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);
             }
         }
     }
@@ -85,7 +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;
         }
     }
@@ -103,7 +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;
     }