[muh] add tuning_find_bound_sorted_values.cuh: BI-V100 tuning header for find_bound_sorted_values

Translated from CCCL cub/device/dispatch/tuning/tuning_find_bound_sorted_values.cuh.
Uses hardware_capability dispatch instead of compute_capability.
This commit is contained in:
dylanyunlon
2026-08-01 02:08:31 +08:00
parent 25f7a636a9
commit 81e4a907c0

View File

@@ -0,0 +1,33 @@
// muh/include/muh/tuning/tuning_find_bound_sorted_values.cuh — BI-V100
//
// Mirrors: cccl_upstream/cub/cub/device/dispatch/tuning/tuning_find_bound_sorted_values.cuh
// CCCL: threads=256, items=8, binary search per thread
//
// vllm relevance: block_table index lookup in paged attention
// SMEM risk: minimal (binary search, no tile)
#pragma once
#include "muh/hardware.cuh"
#include "muh/tuning/common.cuh"
namespace muh::tuning::find_bound {
struct FindBoundPolicy {
int threads_per_block;
int items_per_thread;
CacheLoadModifier haystack_load_modifier;
CacheLoadModifier needles_load_modifier;
};
struct policy_selector {
int haystack_type_size;
int needle_type_size;
constexpr FindBoundPolicy operator()(const hardware_capability& /*hw*/) const {
// CCCL: fixed policy for all architectures
return {256, 8, LOAD_LDG, LOAD_LDG};
}
};
} // namespace muh::tuning::find_bound