ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-04-19 22:39:03 +00:00
Code Issues Packages Projects Releases Wiki Activity

[CK_TILE] Refine Generic2dBlockShape to fix ck_tile example 2,10,11,14 on rdna3 and 4 (#2795)

Browse Source 
BlockWarps, WarpTile in Generic2dBlockShape are wave size dependent, it causes mangled name mismatch between host and device side.

Solution: Replace them with ThreadPerBlock and move BlockWarps, WarpTile calculation into Generic2dBlockShape
This commit is contained in:
linqunAMD
2025-09-10 08:29:20 +08:00
committed by GitHub
parent df4ee556d6
commit c254f3d7b4
14 changed files with 103 additions and 453 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
example/ck_tile/12_smoothquant/example_smoothquant.cpp
View File

@@ -94,12 +94,11 @@ bool run(const ck_tile::ArgParser& arg_parser)
constexpr bool kTwoPass = true;
using BlockWarps = ck_tile::sequence<2, 2>;
using BlockTile = ck_tile::sequence<2, 128>;
using WarpTile = ck_tile::sequence<1, 64>;
using Vector = ck_tile::sequence<1, 1>;
using BlockTile = ck_tile::sequence<2, 128>;
using Vector = ck_tile::sequence<1, 1>;
using ThreadPerBlock = ck_tile::sequence<2, 128>;
using Shape = ck_tile::Generic2dBlockShape<BlockTile, BlockWarps, WarpTile, Vector>;
using Shape = ck_tile::Generic2dBlockShape<BlockTile, ThreadPerBlock, Vector>;
using Problem = ck_tile::SmoothquantPipelineProblem<XDataType,
SmoothScaleDataType,
ComputeDataType,

46
example/ck_tile/12_smoothquant/smoothquant.hpp
View File

@@ -49,54 +49,16 @@ struct smoothquant_traits_
{
using DataType = ck_tile::remove_cvref_t<DataType_>;
static constexpr bool is_warp_per_row = ThreadPerBlock_N_ <= ck_tile::get_warp_size();
static_assert((ThreadPerBlock_M_ * ThreadPerBlock_N_) % ck_tile::get_warp_size() == 0);
static constexpr ck_tile::index_t total_warps =
(ThreadPerBlock_M_ * ThreadPerBlock_N_) / ck_tile::get_warp_size();
// num of warps along m
static constexpr ck_tile::index_t BlockWarps_M = []() {
if constexpr(is_warp_per_row)
{
static_assert(ck_tile::get_warp_size() % ThreadPerBlock_N_ == 0);
return total_warps * (ck_tile::get_warp_size() / ThreadPerBlock_N_);
}
else
{
// static_assert(ck_tile::get_warp_size() % ThreadPerBlock_M_ == 0);
return total_warps / (ThreadPerBlock_N_ / ck_tile::get_warp_size());
}
}();
// num of warps along n
static constexpr ck_tile::index_t BlockWarps_N = []() {
if constexpr(is_warp_per_row)
{
static_assert(ck_tile::get_warp_size() % ThreadPerBlock_N_ == 0);
return 1;
}
else
{
static_assert(ThreadPerBlock_N_ % ck_tile::get_warp_size() == 0);
return ThreadPerBlock_N_ / ck_tile::get_warp_size();
}
}();
static constexpr ck_tile::index_t Repeat_M = Repeat_M_;
static constexpr ck_tile::index_t Repeat_N = Repeat_N_;
static constexpr ck_tile::index_t Block_M = Repeat_M_ * ThreadPerBlock_M_;
static constexpr ck_tile::index_t Block_N = Repeat_N_ * ThreadPerBlock_N_ * Vector_N_;
static constexpr ck_tile::index_t Warp_M = ThreadPerBlock_M_ / BlockWarps_M;
static constexpr ck_tile::index_t Warp_N = ThreadPerBlock_N_ / BlockWarps_N * Vector_N_;
using BlockTile = ck_tile::sequence<Block_M, Block_N>;
using BlockWarps = ck_tile::sequence<BlockWarps_M, BlockWarps_N>;
using WarpTile = ck_tile::sequence<Warp_M, Warp_N>;
using Vector = ck_tile::sequence<1, Vector_N_>;
using Shape = ck_tile::Generic2dBlockShape<BlockTile, BlockWarps, WarpTile, Vector>;
using BlockTile = ck_tile::sequence<Block_M, Block_N>;
using Vector = ck_tile::sequence<1, Vector_N_>;
using ThreadPerBlock = ck_tile::sequence<ThreadPerBlock_M_, ThreadPerBlock_N_>;
using Shape = ck_tile::Generic2dBlockShape<BlockTile, ThreadPerBlock, Vector>;
static constexpr bool kPadN = kPadN_;
static constexpr bool kTwoPass = kTwoPass_;