demo · v135

Image tensor downsample — fp32 vs fp16

The TC39 proposal calls out machine-learning and graphics pipelines as the primary motivation for Float16Array. Here a 256×256 image is loaded as a normalized tensor (RGB / 255). The middle panel keeps it as Float32Array; the right panel rounds through Float16Array. Compare bytes used and reconstructed pixels.

Float16Array: ?

source (uint8 rgb)

tensor → Float32 → pixels

tensor → Float16 → pixels

fp32 tensor

—

bytes (rgb only)

fp16 tensor

—

bytes (rgb only)

peak abs error

—

per-channel (0..1)

the code

// Real ML preprocessing: HWC uint8 → normalized fp tensor.
const px = ctx.getImageData(0, 0, W, H).data;
const n = W * H * 3;

const fp32 = new Float32Array(n);
const fp16 = new Float16Array(n);   // Chrome 135+

for (let i = 0, j = 0; i < px.length; i += 4, j += 3) {
  fp32[j    ] = px[i    ] / 255;
  fp32[j + 1] = px[i + 1] / 255;
  fp32[j + 2] = px[i + 2] / 255;
}
fp16.set(fp32);                      // rounds to binary16 on write

console.log(fp32.byteLength, fp16.byteLength); // 49152 vs 24576