small_ints.um
Buffers of 8- and 16-bit integers (i8 / u8 / i16 / u16), and using the same weights on the CPU and the GPU.
Source: examples/small_ints.um
// 8 / 16 ビットの整数の Buffer(i8 / u8 / i16 / u16)と、CPU と GPU で同じ重みを使う書き方。//// - Buffer<i8> などは、CPU では普通の配列と同じに読み書きでき、kernel にも渡せる(GPU でも同じ値になる)。// - kernel の中の 8 / 16 ビットの値は、読み書き・比較・`as` での変換だけ。計算は `as i32` で広げてから。// - gpu.submit には、kernel が読むだけの普通の Buffer も渡せる。重みを1つの Buffer に置けば、// CPU の評価と GPU の評価が同じ領域を読む(写しを2つ持たなくてよい)。//// expect: -540// expect: -540// expect: -98// expect: 300
kernel function affine(i: u32, out: mut RacyBuffer<i32>, w: Buffer<i8>, x: Buffer<u8>, b: Buffer<i16>, n: u32): void { let acc: i32 = b[i] as i32; for (let k: u32 = 0; k < n; k++) { acc = acc + (w[i * n + k] as i32) * (x[k] as i32); } out[i] = acc;}
// CPU 版: 同じ Buffer をそのまま読む。function affineCpu(w: Buffer<i8>, x: Buffer<u8>, b: Buffer<i16>, row: u32, n: u32): i32 { let acc: i32 = b[row] as i32; for (let k: u32 = 0; k < n; k++) { acc = acc + (w[row * n + k] as i32) * (x[k] as i32); } return acc;}
export function main(): i32 { const rows: u32 = 4; const n: u32 = 8; let w = new Buffer<i8>(rows * n); let x = new Buffer<u8>(n); let b = new Buffer<i16>(rows); for (let i = 0; i < w.length; i++) { w[i] = ((i as i32) * 7 % 11 - 5) as i8; } for (let i = 0; i < x.length; i++) { x[i] = (i * 30 + 20) as u8; } for (let i = 0; i < b.length; i++) { b[i] = ((i as i32) * 100 - 150) as i16; }
let out = new RacyBuffer<i32>(rows); const t = gpu.submit(affine, rows, mut out, w, x, b, n); const mine = affineCpu(w, x, b, 2, n); // GPU が計算しているあいだに、CPU も同じ重みを読む gpu.wait(t); print(out[2]); print(mine);
// 切り詰め: i32 から i8 へは下位 8 ビット(ラップ)。u8 から i32 へは 0 で広げる。 const big: i32 = 1950; print(big as i8); // 1950 = 7 * 256 + 158 → -98 print((200 as u8) as i32 + 100); return 0;}Expected output
Section titled “Expected output”From the // expect: lines. The test runner checks this output on wasm32 and native.
-540-540-98300