adrien/zig-wgpu
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

GpuBuffer align with size 4 as required by WebGPU

Browse Source 
Otherwise a GpuBuffer of a single f16 is too small (only 2) and a eneven
f16 wouldn't work either
This commit is contained in:
adrien 2026-05-18 16:24:43 +02:00
parent d9cb4c4672
commit c7afe28cc6
2 changed files with 24 additions and 22 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

26
src/GpuBuffer.zig
View File

@ -3,7 +3,7 @@ const c = @import("utils.zig").c;
const GpuAllocator = @import("GpuAllocator.zig");
raw: c.WGPUBuffer,
size: u64,
size: u64, // Now tracks the 4-byte aligned size directly
usage: c.WGPUBufferUsage,
gloc: GpuAllocator,
@ -27,11 +27,13 @@ pub fn init(gloc: GpuAllocator, size: u64, usage: std.EnumSet(BufferUsage)) !@Th
var iter = usage.iterator();
while (iter.next()) |flag| use |= @intFromEnum(flag);
const raw_handle = try gloc.allocBuffer(size, use);
// 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 raw_handle = try gloc.allocBuffer(aligned_size, use);
return .{
.raw = raw_handle,
.size = size,
.size = aligned_size, // Expose the aligned size to the rest of the application
.usage = use,
.gloc = gloc,
};
@ -69,7 +71,25 @@ pub fn load(
T: type,
data: []const T,
) !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);
}
}
pub fn read(self: @This(), alloc: std.mem.Allocator, T: type) ![]T {

18
src/bench.zig
View File

@ -40,10 +40,6 @@ const Vec = struct {
try self.buf.load(f16, data);
}
pub fn byteSize(self: Vec) u64 {
return @as(u64, self.len) * @sizeOf(f16);
}
// Changed: gloc is passed by value instead of *GpuAllocator
pub fn run(self: Vec, gloc: GpuAllocator, other: Vec, process: GpuProcess) !Vec {
std.debug.assert(self.len == other.len);
@ -75,20 +71,6 @@ pub fn main(init: std.process.Init) !void {
const allocator = init.gpa;
// --- WARM-UP PHASE ---
{
var warmup_a = [_]f16{1.0};
var warmup_b = [_]f16{1.0};
const wa = try Vec.initLoad(gloc, &warmup_a);
defer wa.deinit();
const wb = try Vec.initLoad(gloc, &warmup_b);
defer wb.deinit();
const wsum = try wa.run(gloc, wb, add_pip);
defer wsum.deinit();
const wout = try wsum.read(allocator);
defer allocator.free(wout);
}
const sizes = [_]usize{
1,
256,