c0bfc8c93d
Translated from CCCL cub/device/dispatch/tuning/tuning_segmented_scan.cuh. Uses hardware_capability dispatch instead of compute_capability.
46 lines
1.4 KiB
Plaintext
46 lines
1.4 KiB
Plaintext
// muh/include/muh/tuning/tuning_segmented_scan.cuh — BI-V100
|
|
//
|
|
// Mirrors: cccl_upstream/cub/cub/device/dispatch/tuning/tuning_segmented_scan.cuh
|
|
// CCCL: threads=128, items=9 (nominal), scale_mem_bound for SMEM
|
|
//
|
|
// vllm relevance: per-sequence softmax denominator accumulation
|
|
// SMEM risk: scale_mem_bound handles it. tuple<AccumT, bool> → effective size = accum_size + 1 (padded to alignment)
|
|
|
|
#pragma once
|
|
|
|
#include "muh/hardware.cuh"
|
|
#include "muh/tuning/common.cuh"
|
|
|
|
namespace muh::tuning::segmented_scan {
|
|
|
|
struct SegmentedScanPolicy {
|
|
int threads_per_block;
|
|
int items_per_thread;
|
|
int max_segments_per_block;
|
|
BlockLoadAlgorithm load_algorithm;
|
|
BlockStoreAlgorithm store_algorithm;
|
|
BlockScanAlgorithm scan_algorithm;
|
|
};
|
|
|
|
struct policy_selector {
|
|
int accum_size;
|
|
int accum_align;
|
|
|
|
constexpr SegmentedScanPolicy operator()(const hardware_capability& /*hw*/) const {
|
|
constexpr int nominal_threads = 128;
|
|
constexpr int nominal_items = 9;
|
|
constexpr int max_segments = 512;
|
|
|
|
// Multi-segment agent uses tuple<AccumT, bool>: size = round_up(accum_size + 1, accum_align)
|
|
int tuple_size = accum_size + accum_align; // conservative estimate
|
|
auto [items, threads] = scale_mem_bound(nominal_threads, nominal_items, tuple_size);
|
|
|
|
return {threads, items, max_segments,
|
|
BLOCK_LOAD_WARP_TRANSPOSE,
|
|
BLOCK_STORE_WARP_TRANSPOSE,
|
|
BLOCK_SCAN_WARP_SCANS};
|
|
}
|
|
};
|
|
|
|
} // namespace muh::tuning::segmented_scan
|