code/bench_decode.py

"""Throughput of on-GPU weight reconstruction (codebook gather) and of the
online activation-side Hadamard transforms, which is what an inference engine
must sustain to keep the storage stream busy.
"""
import json, os, sys, time
import torch

sys.path.insert(0, os.path.dirname(__file__))
import codec

DEV = "cuda"
RES = os.path.join(os.path.dirname(__file__), "..", "results")


def bench_decode(stages, nweights=1 << 24, reps=20):
    C = torch.randn(stages, 256, codec.D_SUB, device=DEV, dtype=torch.float16)
    n = nweights // codec.D_SUB
    idx = torch.randint(0, 256, (stages, n), device=DEV, dtype=torch.uint8)
    scale = torch.randn(nweights // 2048, 1, device=DEV, dtype=torch.float16)

    def run():
        acc = C[0][idx[0].long()]
        for s in range(1, stages):
            acc = acc + C[s][idx[s].long()]
        return (acc.view(-1, 2048) * scale).view(-1)

    for _ in range(3):
        run()
    torch.cuda.synchronize()
    t0 = time.perf_counter()
    for _ in range(reps):
        run()
    torch.cuda.synchronize()
    dt = (time.perf_counter() - t0) / reps
    packed = nweights * stages * codec.CB_BITS / codec.D_SUB / 8
    return dict(stages=stages, bits=stages * codec.BITS_PER_STAGE,
                weights_per_s=nweights / dt,
                fp16_equiv_GBs=nweights * 2 / dt / 1e9,
                packed_GBs=packed / dt / 1e9)


def bench_hadamard(dim=8192, batch=1, reps=200):
    x = torch.randn(batch, dim, device=DEV)
    for _ in range(3):
        codec._fwht(x)
    torch.cuda.synchronize()
    t0 = time.perf_counter()
    for _ in range(reps):
        codec._fwht(x)
    torch.cuda.synchronize()
    return (time.perf_counter() - t0) / reps * 1e6      # microseconds


def bench_matmul():
    out = {}
    for m, k, n in [(1, 8192, 2048), (32, 8192, 2048), (2048, 8192, 2048)]:
        a = torch.randn(m, k, device=DEV, dtype=torch.float16)
        b = torch.randn(k, n, device=DEV, dtype=torch.float16)
        for _ in range(3):
            a @ b
        torch.cuda.synchronize()
        reps = 50
        t0 = time.perf_counter()
        for _ in range(reps):
            a @ b
        torch.cuda.synchronize()
        dt = (time.perf_counter() - t0) / reps
        out[f"{m}x{k}x{n}"] = dict(tflops=2 * m * k * n / dt / 1e12, ms=dt * 1e3)
    return out


if __name__ == "__main__":
    res = {"decode": [bench_decode(s) for s in [2, 3, 4]],
           "hadamard_us": {str(d): bench_hadamard(d) for d in [2048, 4096, 8192]},
           "matmul": bench_matmul(),
           "gpu": torch.cuda.get_device_name(0)}
    for d in res["decode"]:
        print(f"decode {d['bits']:.1f} bit: {d['weights_per_s']/1e9:.2f} Gweight/s "
              f"= {d['fp16_equiv_GBs']:.1f} GB/s fp16-equivalent, "
              f"{d['packed_GBs']:.2f} GB/s of packed bytes consumed")
    print("hadamard (us):", {k: round(v, 1) for k, v in res["hadamard_us"].items()})
    for k, v in res["matmul"].items():
        print(f"matmul {k}: {v['tflops']:.2f} TFLOP/s")
    json.dump(res, open(os.path.join(RES, "decode_bench.json"), "w"), indent=2)