diff --git a/src/GpuAllocator.zig b/src/GpuAllocator.zig
index 9600710..efc005d 100644
--- a/src/GpuAllocator.zig
+++ b/src/GpuAllocator.zig
@@ -1,5 +1,4 @@
 const std = @import("std");
-const sh = @import("shaders.zig");
 const GpuDevice = @import("GpuDevice.zig");
 const c = @import("c.zig").c;
 
@@ -8,25 +7,16 @@ const GpuAllocator = @This();
 device: GpuDevice,
 cpu_allocator: std.mem.Allocator,
 tracked_buffers: std.AutoHashMap(c.WGPUBuffer, void),
-pipelines: struct {
-    add: c.WGPUComputePipeline,
-},
 
 pub fn init(cpu_allocator: std.mem.Allocator, device: GpuDevice) !GpuAllocator {
     return .{
         .device = device,
         .cpu_allocator = cpu_allocator,
         .tracked_buffers = .init(cpu_allocator),
-        .pipelines = .{
-            .add = try buildPipeline(device.device, sh.SHADER_ADD),
-        },
     };
 }
 
 pub fn deinit(self: *GpuAllocator) void {
-    inline for (@typeInfo(@TypeOf(self.pipelines)).@"struct".fields) |field|
-        c.wgpuComputePipelineRelease(@field(self.pipelines, field.name));
-
     var it = self.tracked_buffers.keyIterator();
     while (it.next()) |buf_ptr| {
         const buf = buf_ptr.*;
@@ -56,22 +46,3 @@ pub fn unregisterAndDestroyBuffer(self: *GpuAllocator, buf: c.WGPUBuffer) void {
         c.wgpuBufferRelease(buf);
     }
 }
-
-fn buildPipeline(device: c.WGPUDevice, wgsl: []const u8) !c.WGPUComputePipeline {
-    var wgsl_src = c.WGPUShaderSourceWGSL{
-        .chain = .{ .sType = c.WGPUSType_ShaderSourceWGSL },
-        .code = sv(wgsl),
-    };
-    const shader = c.wgpuDeviceCreateShaderModule(device, &.{
-        .nextInChain = @ptrCast(&wgsl_src),
-    }) orelse return error.Shader;
-    defer c.wgpuShaderModuleRelease(shader);
-
-    return c.wgpuDeviceCreateComputePipeline(device, &.{
-        .compute = .{ .module = shader, .entryPoint = sv("main") },
-    }) orelse error.Pipeline;
-}
-
-fn sv(s: []const u8) c.WGPUStringView {
-    return .{ .data = s.ptr, .length = s.len };
-}
diff --git a/src/GpuDevice.zig b/src/GpuDevice.zig
index be4026c..601cd02 100644
--- a/src/GpuDevice.zig
+++ b/src/GpuDevice.zig
@@ -1,5 +1,4 @@
 const std = @import("std");
-const sh = @import("shaders.zig");
 const c = @import("c.zig").c;
 
 const Ctx = struct {
diff --git a/src/GpuPipeline.zig b/src/GpuPipeline.zig
new file mode 100644
index 0000000..68b3540
--- /dev/null
+++ b/src/GpuPipeline.zig
@@ -0,0 +1,28 @@
+const std = @import("std");
+const GpuDevice = @import("GpuDevice.zig");
+const c = @import("c.zig").c;
+
+raw: c.WGPUComputePipeline,
+
+pub fn init(device: GpuDevice, wgsl: []const u8) !@This() {
+    var wgsl_src = c.WGPUShaderSourceWGSL{
+        .chain = .{ .sType = c.WGPUSType_ShaderSourceWGSL },
+        .code = sv(wgsl),
+    };
+    const shader = c.wgpuDeviceCreateShaderModule(device.device, &.{
+        .nextInChain = @ptrCast(&wgsl_src),
+    }) orelse return error.Shader;
+    defer c.wgpuShaderModuleRelease(shader);
+
+    return .{ .raw = c.wgpuDeviceCreateComputePipeline(device.device, &.{
+        .compute = .{ .module = shader, .entryPoint = sv("main") },
+    }) orelse return error.Pipeline };
+}
+
+pub fn deinit(self: @This()) void {
+    c.wgpuComputePipelineRelease(self.raw);
+}
+
+fn sv(s: []const u8) c.WGPUStringView {
+    return .{ .data = s.ptr, .length = s.len };
+}
diff --git a/src/Mat.zig b/src/Mat.zig
index 144132a..06fcfbc 100644
--- a/src/Mat.zig
+++ b/src/Mat.zig
@@ -3,6 +3,7 @@ const std = @import("std");
 const c = @import("c.zig").c;
 const GpuAllocator = @import("GpuAllocator.zig");
 const GpuBuffer = @import("GpuBuffer.zig");
+const GpuPipeline = @import("GpuPipeline.zig");
 
 const Mat = @This();
 
@@ -52,13 +53,13 @@ pub fn byteSize(self: Mat) u64 {
     return @as(u64, self.len()) * @sizeOf(f32);
 }
 
-pub fn add(self: Mat, gloc: *GpuAllocator, other: Mat) !Mat {
+pub fn run(self: Mat, gloc: *GpuAllocator, other: Mat, pip: GpuPipeline) !Mat {
     std.debug.assert(self.rows == other.rows and self.cols == other.cols);
 
     const result = try Mat.zeros(gloc, self.rows, self.cols);
     errdefer result.deinit();
 
-    try dispatch2in1out(gloc, gloc.pipelines.add, self.buf, other.buf, result.buf, self.byteSize());
+    try dispatch2in1out(gloc, pip.raw, self.buf, other.buf, result.buf, self.byteSize());
 
     return result;
 }
diff --git a/src/main.zig b/src/main.zig
index 8f7748a..673ce9c 100644
--- a/src/main.zig
+++ b/src/main.zig
@@ -1,6 +1,7 @@
 const std = @import("std");
 const GpuDevice = @import("GpuDevice.zig");
 const GpuAllocator = @import("GpuAllocator.zig");
+const GpuPipeline = @import("GpuPipeline.zig");
 const Mat = @import("Mat.zig");
 
 pub fn main(init: std.process.Init) !void {
@@ -10,6 +11,9 @@ pub fn main(init: std.process.Init) !void {
     var gloc = try GpuAllocator.init(init.gpa, device);
     defer gloc.deinit();
 
+    const add_pip = try GpuPipeline.init(device, @embedFile("shaders/add.wgsl"));
+    defer add_pip.deinit();
+
     // Define the sizes you want to benchmark
     const sizes = [_]usize{
         1,
@@ -54,7 +58,7 @@ pub fn main(init: std.process.Init) !void {
         defer b.deinit();
 
         // a + b
-        const sum = try a.add(&gloc, b);
+        const sum = try a.run(&gloc, b, add_pip);
         defer sum.deinit();
 
         const out = try sum.read(&gloc, allocator);
diff --git a/src/shaders.zig b/src/shaders.zig
deleted file mode 100644
index 26f2330..0000000
--- a/src/shaders.zig
+++ /dev/null
@@ -1,3 +0,0 @@
-const std = @import("std");
-
-pub const SHADER_ADD = @embedFile("shaders/add.wgsl");