bench(vulkan): add Vulkan-specific benchmarks for Steam Deck optimization

GPU GEMV vs CPU matmul at Qwen3-0.6B model sizes, dispatch overhead,
elementwise ops, and full MLP pass benchmarks.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: evilsocket

cake-core/benches/bench_vulkan.rs

@@ -0,0 +1,157 @@
+/// Vulkan backend benchmarks — GPU GEMV vs CPU matmul at model-realistic sizes,
+/// dispatch overhead, upload/download costs, and elementwise ops.
+///
+/// Run on Steam Deck: `cargo bench -p cake-core --features vulkan -- vulkan`
+
+use cake_core::backends::{ComputeBackend, VulkanBackend};
+use candle_core::{DType, Device, Tensor};
+
+fn vk() -> VulkanBackend {
+    VulkanBackend::new().expect("Vulkan backend required for these benchmarks")
+}
+
+fn cpu_tensor(shape: &[usize], seed: u64) -> Tensor {
+    super::bench_helpers::make_tensor(shape, seed)
+}
+
+// ── Dispatch overhead ────────────────────────────────────────────────
+
+#[divan::bench]
+fn vulkan_dispatch_overhead(bencher: divan::Bencher) {
+    let backend = vk();
+    // Tiny tensor to measure fixed dispatch cost (not data transfer)
+    let a = cpu_tensor(&[1, 1, 16], 1000);
+    let b = cpu_tensor(&[1, 1, 16], 1001);
+    bencher.bench_local(|| backend.silu_mul(&a, &b).unwrap());
+}
+
+// ── GPU GEMV vs CPU matmul at model sizes ────────────────────────────
+// Qwen3-0.6B: hidden=1024, intermediate=3072, head_dim=128
+// QKV: (1,1024) × (1024,4096), O: (1,1024) × (1024,1024)
+// gate_up: (1,1024) × (1024,6144), down: (1,3072) × (3072,1024)
+
+#[divan::bench(args = [1024, 4096, 6144])]
+fn vulkan_gemv_1024xN(bencher: divan::Bencher, n: usize) {
+    let backend = vk();
+    let a = cpu_tensor(&[1, 1024], 1100);
+    let b = cpu_tensor(&[1024, n], 1101);
+    bencher.bench_local(|| backend.matmul(&a, &b).unwrap());
+}
+
+#[divan::bench]
+fn vulkan_gemv_3072x1024(bencher: divan::Bencher) {
+    let backend = vk();
+    let a = cpu_tensor(&[1, 3072], 1200);
+    let b = cpu_tensor(&[3072, 1024], 1201);
+    bencher.bench_local(|| backend.matmul(&a, &b).unwrap());
+}
+
+#[divan::bench(args = [1024, 4096, 6144])]
+fn cpu_gemv_1024xN(bencher: divan::Bencher, n: usize) {
+    let a = cpu_tensor(&[1, 1024], 1100);
+    let b = cpu_tensor(&[1024, n], 1101);
+    bencher.bench_local(|| a.matmul(&b).unwrap());
+}
+
+#[divan::bench]
+fn cpu_gemv_3072x1024(bencher: divan::Bencher) {
+    let a = cpu_tensor(&[1, 3072], 1200);
+    let b = cpu_tensor(&[3072, 1024], 1201);
+    bencher.bench_local(|| a.matmul(&b).unwrap());
+}
+
+// ── GPU GEMM (prefill) at model sizes ────────────────────────────────
+
+#[divan::bench(args = [8, 32, 64])]
+fn vulkan_gemm_Mx1024x4096(bencher: divan::Bencher, m: usize) {
+    let backend = vk();
+    let a = cpu_tensor(&[m, 1024], 1300);
+    let b = cpu_tensor(&[1024, 4096], 1301);
+    bencher.bench_local(|| backend.matmul(&a, &b).unwrap());
+}
+
+#[divan::bench(args = [8, 32, 64])]
+fn cpu_gemm_Mx1024x4096(bencher: divan::Bencher, m: usize) {
+    let a = cpu_tensor(&[m, 1024], 1300);
+    let b = cpu_tensor(&[1024, 4096], 1301);
+    bencher.bench_local(|| a.matmul(&b).unwrap());
+}
+
+// ── Elementwise ops at model sizes ───────────────────────────────────
+
+#[divan::bench(args = [1024, 3072])]
+fn vulkan_silu_mul(bencher: divan::Bencher, size: usize) {
+    let backend = vk();
+    let gate = cpu_tensor(&[1, 1, size], 1400);
+    let up = cpu_tensor(&[1, 1, size], 1401);
+    bencher.bench_local(|| backend.silu_mul(&gate, &up).unwrap());
+}
+
+#[divan::bench(args = [1024, 3072])]
+fn cpu_silu_mul(bencher: divan::Bencher, size: usize) {
+    let gate = cpu_tensor(&[1, 1, size], 1400);
+    let up = cpu_tensor(&[1, 1, size], 1401);
+    bencher.bench_local(|| {
+        (candle_nn::ops::silu(&gate).unwrap() * &up).unwrap()
+    });
+}
+
+#[divan::bench(args = [1024, 3072])]
+fn vulkan_add3(bencher: divan::Bencher, size: usize) {
+    let backend = vk();
+    let a = cpu_tensor(&[1, 1, size], 1500);
+    let b = cpu_tensor(&[1, 1, size], 1501);
+    let c = cpu_tensor(&[1, 1, size], 1502);
+    bencher.bench_local(|| backend.add3(&a, &b, &c).unwrap());
+}
+
+// ── RMS norm (CPU-only in current backend) ───────────────────────────
+
+#[divan::bench(args = [1024, 3072])]
+fn vulkan_rms_norm_gated(bencher: divan::Bencher, size: usize) {
+    let backend = vk();
+    let x = cpu_tensor(&[1, 1, size], 1600);
+    let z = cpu_tensor(&[1, 1, size], 1601);
+    let w = Tensor::ones(size, DType::F32, &Device::Cpu).unwrap();
+    bencher.bench_local(|| backend.rms_norm_gated(&x, &z, &w, 1e-6).unwrap());
+}
+
+#[divan::bench(args = [1024, 3072])]
+fn vulkan_add_rms_norm(bencher: divan::Bencher, size: usize) {
+    let backend = vk();
+    let a = cpu_tensor(&[1, 1, size], 1700);
+    let b = cpu_tensor(&[1, 1, size], 1701);
+    let w = Tensor::ones(size, DType::F32, &Device::Cpu).unwrap();
+    bencher.bench_local(|| backend.add_rms_norm(&a, &b, &w, 1e-6).unwrap());
+}
+
+// ── Full MLP pass (gate_up + silu_mul + down) ────────────────────────
+
+#[divan::bench]
+fn vulkan_mlp_full(bencher: divan::Bencher) {
+    let backend = vk();
+    let x = cpu_tensor(&[1, 1024], 1800);
+    let gate_up_w = cpu_tensor(&[6144, 1024], 1801);
+    let down_w = cpu_tensor(&[1024, 3072], 1802);
+    bencher.bench_local(|| {
+        let fused = backend.matmul(&x, &gate_up_w.t().unwrap()).unwrap();
+        let gate = fused.narrow(1, 0, 3072).unwrap().contiguous().unwrap();
+        let up = fused.narrow(1, 3072, 3072).unwrap().contiguous().unwrap();
+        let act = backend.silu_mul(&gate, &up).unwrap();
+        backend.matmul(&act, &down_w.t().unwrap()).unwrap()
+    });
+}
+
+#[divan::bench]
+fn cpu_mlp_full(bencher: divan::Bencher) {
+    let x = cpu_tensor(&[1, 1024], 1800);
+    let gate_up_w = cpu_tensor(&[6144, 1024], 1801);
+    let down_w = cpu_tensor(&[1024, 3072], 1802);
+    bencher.bench_local(|| {
+        let fused = x.matmul(&gate_up_w.t().unwrap()).unwrap();
+        let gate = fused.narrow(1, 0, 3072).unwrap();
+        let up = fused.narrow(1, 3072, 3072).unwrap();
+        let act = (candle_nn::ops::silu(&gate).unwrap() * &up).unwrap();
+        act.matmul(&down_w.t().unwrap()).unwrap()
+    });
+}

cake-core/benches/benchmarks.rs

@@ -21,3 +21,5 @@ mod bench_flux;
 mod bench_vibevoice;
 #[cfg(feature = "luxtts")]
 mod bench_luxtts;
+#[cfg(feature = "vulkan")]
+mod bench_vulkan;