File size: 2,199 Bytes
80692f2 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 | name: 05_topk_bitonic
display_name: "TopK via Bitonic Sort"
precision: fp32
regime: memory
# Top-k is dominated by the input read (small output, no reduction over k).
# Comparator-network FLOPs are not the bottleneck on real hardware, so we
# track them but score on bandwidth.
flops_formula: "batch * n * 4" # ~O(n log^2 n) compares total, but the
# bitonic network is so cheap relative
# to memory that a coarse 4*n estimate
# is fine for telemetry only.
bytes_formula: "batch * n * 4 + batch * k * (4 + 8)" # fp32 input read + (fp32 value + int64 idx) output
hardware: [RTX_PRO_6000]
peak_tflops_key: fp32
peak_bandwidth_key: dram
# Top-k correctness:
# - VALUES must match within fp32 atol/rtol (the kth largest value is
# well-defined modulo float-equal ties, so we use a loose-ish tol).
# - INDICES are checked leniently: for each row, the multiset of returned
# indices must select values that match ref values within tol. Direct
# index equality is NOT required (ties in x can yield different valid
# index sets).
tolerance:
float32: 1.0e-4
# Forbidden ops — using any of these in solution.py fails correctness post-hoc.
# This problem is about IMPLEMENTING the selection, not dispatching to PyTorch's
# tuned top-k. torch.sort is also banned because torch.topk falls back to it.
forbidden:
- "torch.topk"
- "torch.kthvalue"
- "torch.sort"
- "torch.argsort"
- "Tensor.topk"
- "Tensor.kthvalue"
- "Tensor.sort"
- "Tensor.argsort"
- "torch.ops.aten.topk"
- "torch.ops.aten.sort"
- "torch.ops.aten.kthvalue"
sota:
name: "torch.topk (cuTOPK / CUB internals)"
url: "https://github.com/pytorch/pytorch/blob/main/aten/src/ATen/native/cuda/TensorTopK.cu"
function: "torch.topk"
deps: []
# Informational: torch.topk dispatches to a radix-select kernel for moderate
# k and to a bitonic sort kernel for small n. Beating it on the (1, 131072,
# 64) decoder shape requires saturating DRAM bandwidth on the input read.
reference_throughput_gbps_h100: 2400
num_correct_trials: 3
num_perf_trials: 50
|