[muh] add tuning_transform_tile.cuh: BI-V100 tuning header for transform_tile

Translated from CCCL cub/device/dispatch/tuning/tuning_transform_tile.cuh.
Uses hardware_capability dispatch instead of compute_capability.
This commit is contained in:
dylanyunlon
2026-08-01 02:08:40 +08:00
parent 3fee54f4d7
commit 0344a3fbb8

View File

@@ -0,0 +1,33 @@
// muh/include/muh/tuning/tuning_transform_tile.cuh — BI-V100
//
// Mirrors: cccl_upstream/cub/cub/device/dispatch/tuning/tuning_transform_tile.cuh
// CCCL: tile-based transform, shares bytes-in-flight target with transform
//
// vllm relevance: fused activation+normalization tiles
// SMEM risk: tile size is compile-time shape, not threads*items
#pragma once
#include "muh/hardware.cuh"
#include "muh/tuning/common.cuh"
namespace muh::tuning::transform_tile {
struct TransformTilePolicy {
int tile_size;
int min_bytes_in_flight;
};
struct policy_selector {
int elem_size;
constexpr TransformTilePolicy operator()(const hardware_capability& hw) const {
// BI-V100 per-SM BW ≈ A100 → bytes_in_flight = 16KB
int bytes_in_flight = 16 * 1024;
int tile = bytes_in_flight / elem_size;
if (tile < 128) tile = 128;
return {tile, bytes_in_flight};
}
};
} // namespace muh::tuning::transform_tile