12ad7a3190
每个文件都是直接 cat 读完整 CCCL 源码后理解全部参数语义，
然后用大模型生成 BI-V100 适配版本。不使用 grep/sed/批量脚本。
segmented_sort.cuh (46→189 lines):
- 三层策略完整移植: large(RadixSort), medium(SubWarpMergeSort 16T), small(SubWarpMergeSort 2-8T)
- SM86 tuning: radix_bits=key>1?6:4, scale_reg_bound(256,23)
- BI-V100 SMEM cap for all three tiers
merge_sort.cuh (43→83 lines):
- SM50{256,11} SM52{512,15} SM60+{256,17} 三代参数
- nominal_4b_items_to_items scaling + unroll flag
merge.cuh (55→89 lines):
- SM52/SM60/SM80/SM90/SM100 五代参数
- bulk_copy=false (BI-V100 无 cp.async.bulk)
adjacent_difference.cuh (46→77 lines):
- nominal_8b_items_to_items(7) scaling
- may_alias → LOAD_CA vs LOAD_LDG
batch_memcpy.cuh (86→95 lines):
- small{128T,4buf,8B} + large{256T,32B} 双策略
- prefer_pow2_bits=false (SM70+)
find.cuh (32→39 lines):
- scale_mem_bound(128,16) + vec_size=4
find_bound_sorted_values.cuh (33→47 lines):
- SM80+: {512, N4B(15)} / SM60+: {256} / SM50: LOAD_LDG
78 lines
2.8 KiB
Plaintext
78 lines
2.8 KiB
Plaintext
// muh/include/muh/tuning/tuning_adjacent_difference.cuh — BI-V100
|
|
//
|
|
// Mirrors: cccl_upstream/cub/cub/device/dispatch/tuning/tuning_adjacent_difference.cuh
|
|
// CCCL source: 118 lines. Single policy for all compute capabilities.
|
|
//
|
|
// CCCL policy_selector (all CC):
|
|
// {128, Nominal8BItems(7, value_size), WARP_TRANSPOSE, may_alias?LOAD_CA:LOAD_LDG, WARP_TRANSPOSE}
|
|
//
|
|
// Nominal8BItemsToItems(7, value_size) = max(1, 7 * 8 / value_size)
|
|
// 1B → 56, 2B → 28, 4B → 14, 8B → 7, 16B → 3
|
|
//
|
|
// vllm relevance: computing token-level delta logits for speculative decoding,
|
|
// detecting attention pattern changes between consecutive positions.
|
|
//
|
|
// SMEM: threads * items * value_size * 2 (BlockLoad + BlockStore, double buffer)
|
|
// BI-V100 48KB limit → cap items when value_size is large.
|
|
|
|
#pragma once
|
|
|
|
#include "muh/hardware.cuh"
|
|
#include "muh/tuning/common.cuh"
|
|
|
|
namespace muh::tuning::adjacent_difference {
|
|
|
|
// ============================================================================
|
|
// Policy struct — matches CCCL AdjacentDifferencePolicy exactly
|
|
// ============================================================================
|
|
|
|
struct AdjacentDifferencePolicy {
|
|
int threads_per_block;
|
|
int items_per_thread;
|
|
BlockLoadAlgorithm load_algorithm;
|
|
CacheLoadModifier load_modifier;
|
|
BlockStoreAlgorithm store_algorithm;
|
|
};
|
|
|
|
// ============================================================================
|
|
// Helper: nominal_8B_items_to_items (from CCCL util_device.cuh)
|
|
// Scales items from an 8-byte nominal to actual value_size
|
|
// ============================================================================
|
|
|
|
constexpr int nominal_8b_items(int nominal, int value_size) {
|
|
int result = nominal * 8 / value_size;
|
|
return result > 0 ? result : 1;
|
|
}
|
|
|
|
// ============================================================================
|
|
// policy_selector — single policy for all CC (matches CCCL)
|
|
// BI-V100 SMEM cap applied for large items
|
|
// ============================================================================
|
|
|
|
struct policy_selector {
|
|
int value_type_size;
|
|
bool may_alias;
|
|
|
|
constexpr AdjacentDifferencePolicy operator()(const hardware_capability& hw) const {
|
|
int items = nominal_8b_items(7, value_type_size);
|
|
|
|
// SMEM check: BlockLoad + BlockStore share tile through union
|
|
// tile = threads * items * value_type_size
|
|
int threads = 128;
|
|
int tile_smem = threads * items * value_type_size;
|
|
while (tile_smem > hw.max_shared_memory_per_block - 2048 && items > 1) {
|
|
items--;
|
|
tile_smem = threads * items * value_type_size;
|
|
}
|
|
|
|
return AdjacentDifferencePolicy{
|
|
threads, items,
|
|
BLOCK_LOAD_WARP_TRANSPOSE,
|
|
may_alias ? LOAD_CA : LOAD_LDG,
|
|
BLOCK_STORE_WARP_TRANSPOSE
|
|
};
|
|
}
|
|
};
|
|
|
|
} // namespace muh::tuning::adjacent_difference
|