af8fc0caeb
Translated from CCCL cub/device/dispatch/tuning/tuning_merge.cuh. Uses hardware_capability dispatch instead of compute_capability.
56 lines
1.7 KiB
Plaintext
56 lines
1.7 KiB
Plaintext
// muh/include/muh/tuning/tuning_merge.cuh — BI-V100
|
|
//
|
|
// Mirrors: cccl_upstream/cub/cub/device/dispatch/tuning/tuning_merge.cuh
|
|
// CCCL: threads=256/512, items=nominal_4B scaled, depends on bulk copy support
|
|
//
|
|
// vllm relevance: beam search candidate merging
|
|
// SMEM risk: tile = threads * items * (key_size + value_size), can overflow for large types
|
|
|
|
#pragma once
|
|
|
|
#include "muh/hardware.cuh"
|
|
#include "muh/tuning/common.cuh"
|
|
|
|
namespace muh::tuning::merge {
|
|
|
|
struct MergePolicy {
|
|
int threads_per_block;
|
|
int items_per_thread;
|
|
CacheLoadModifier load_modifier;
|
|
BlockStoreAlgorithm store_algorithm;
|
|
bool use_bulk_copy;
|
|
};
|
|
|
|
// CCCL's nominal_4B_items_to_items
|
|
constexpr int nominal_4B_items(int nominal, int type_size) {
|
|
int items = nominal * 4 / type_size;
|
|
return items < 1 ? 1 : items;
|
|
}
|
|
|
|
struct policy_selector {
|
|
int key_size;
|
|
int value_size; // 0 for keys-only
|
|
bool can_bulk_copy;
|
|
|
|
constexpr MergePolicy operator()(const hardware_capability& hw) const {
|
|
int tune_size = key_size + value_size;
|
|
|
|
if (hw.at_least(hardware_capability::vendor_t::iluvatar, 100) && can_bulk_copy) {
|
|
// SM100-like: 512 threads, bulk copy
|
|
int items = nominal_4B_items(11, tune_size);
|
|
// SMEM check: 512 * items * tune_size <= 49152
|
|
while (512 * items * tune_size > hw.max_shared_memory_per_block && items > 1)
|
|
items--;
|
|
return {512, items, LOAD_DEFAULT, BLOCK_STORE_WARP_TRANSPOSE, true};
|
|
}
|
|
|
|
// Default: 256 threads, no bulk
|
|
int items = nominal_4B_items(15, tune_size);
|
|
while (256 * items * tune_size > hw.max_shared_memory_per_block && items > 1)
|
|
items--;
|
|
return {256, items, LOAD_DEFAULT, BLOCK_STORE_WARP_TRANSPOSE, false};
|
|
}
|
|
};
|
|
|
|
} // namespace muh::tuning::merge
|