Files
composable_kernel/include/ck_tile/ops/reduce/block/block_reduce2d.hpp
Yashvardhan Agarwal 191c62967b Fixes to "General 2D Reduction Kernel" (#2535) (#2656)
* fix reduce2d

- revret the combine_partial_results() chnages
- remove auto from function def

* clang-format
2025-08-11 15:01:33 +02:00

445 lines
18 KiB
C++

// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck_tile/core.hpp"
namespace ck_tile {
// BlockReduce2d implements a hierarchical 2D reduction operator that reduces data along the second
// dimension using a user-specified reduction function.
//
// The reduction is performed in a three-stage hierarchical approach:
//
// STAGE 1: Thread-level reduction (BlockReduce2d)
// ===============================================
// - Each thread processes multiple elements from the input tensor within its assigned data
// partition
// - Reduction is performed locally within each thread by iterating over assigned elements
// - ReducePacksPerXDim controls how many elements sweep_tile processes in one iteration per
// dimension
// (e.g., {1,1} = 1 element at a time from each dimension, {2,4} = 2 from dim0, 4 from dim1)
// - Results are accumulated into a thread-local output tensor stored in registers
// - The output tensor distribution is derived from the input tensor's distribution using
// make_reduce_tile_distribution_encoding() to handle dimension reduction
//
// STAGE 2: Warp-level reduction (BlockReduce2dSync)
// ================================================
// - Performs inter-thread reduction within each warp
// - Uses warp shuffle operations to exchange data between threads in the same warp
// - Implements a tree-reduction pattern with power-of-2 stages
// - Only reduces along dimensions that map to lane IDs within the warp
//
// STAGE 3: Cross-warp reduction (BlockReduce2dCrossWarpSync)
// ========================================================
// - Performs reduction across multiple warps within the same thread block
// - Uses shared memory (LDS) to facilitate data exchange between warps
// - Each warp's lane-0 thread stores its partial results to shared memory
// - All threads participate in loading and reducing data from shared memory
// - Implements block-level synchronization to ensure memory consistency
// BlockReduce2d: Thread-level reduction (Stage 1)
template <typename Problem_, typename Policy_ = void>
struct BlockReduce2d
{
// Thread-level reduction implementation
using Problem = remove_cvref_t<Problem_>;
using XDataType = typename Problem::XDataType;
using ComputeDataType = typename Problem::ComputeDataType;
CK_TILE_DEVICE constexpr BlockReduce2d() {}
template <
typename XDistributedTensor_,
typename YDistributedTensor_,
typename ReduceFunc,
typename ReducePacksPerXDim =
uniform_sequence_gen_t<2, 1>> // {1,1} = process 1 element at a time from each dimension
CK_TILE_DEVICE void operator()(const XDistributedTensor_& x_tensor,
YDistributedTensor_& y_tensor,
const ReduceFunc& reduce_func,
ReducePacksPerXDim = {})
{
sweep_tile<XDistributedTensor_>(
[&](auto... idx_) {
constexpr auto idx_0 = make_tuple(make_tuple(idx_[number<0>{}]...)[number<0>{}]);
y_tensor(idx_0) = reduce_func(
y_tensor(idx_0), ck_tile::type_convert<ComputeDataType>(x_tensor[idx_])...);
},
ReducePacksPerXDim{});
#if 0
constexpr auto I0 = number<0>{};
constexpr auto I1 = number<1>{};
constexpr auto spans = XDistributedTensor_::get_distributed_spans();
// FIXME: hard coded to reduce 2nd axis
sweep_tile_span(spans[I0], [&](auto dstr_idx_i0) {
constexpr auto y_dstr_idx = make_tuple(dstr_idx_i0);
auto y = y_tensor[y_dstr_idx];
sweep_tile_span(spans[I1], [&](auto dstr_idx_i1) {
constexpr auto in_dstr_idx = make_tuple(dstr_idx_i0, dstr_idx_i1);
const auto x = ck_tile::type_convert<ComputeDataType>(x_tensor[in_dstr_idx]);
y = reduce_func(y, x);
});
y_tensor(y_dstr_idx) = y;
});
#endif
}
template <typename XDistributedTensor_>
CK_TILE_DEVICE static auto MakeYBlockTile()
{
static_assert(std::is_same_v<XDataType, typename XDistributedTensor_::DataType>, "wrong!");
// FIXME: hard coded to reduce 2nd axis
constexpr auto reduce_dims = sequence<1>{};
constexpr auto dstr =
make_static_tile_distribution(detail::make_reduce_tile_distribution_encoding(
XDistributedTensor_::get_tile_distribution()
.get_static_tile_distribution_encoding(),
reduce_dims));
auto tensor = make_static_distributed_tensor<ComputeDataType>(dstr);
return tensor;
}
// uniform_sequence_gen_t<NSize, Value> generates sequence of NSize elements filled with Value
// e.g., uniform_sequence_gen_t<2, 1> → {1, 1} and uniform_sequence_gen_t<3, 4> → {4, 4, 4}
template <typename XDistributedTensor_,
typename ReduceFunc,
typename ReducePacksPerXDim = uniform_sequence_gen_t<2, 1>>
CK_TILE_DEVICE auto operator()(const XDistributedTensor_& x_tensor,
const ComputeDataType& reduce_init,
const ReduceFunc& reduce_func,
ReducePacksPerXDim = {})
{
auto y_tensor = MakeYBlockTile<XDistributedTensor_>();
set_tile(y_tensor, reduce_init);
(*this)(x_tensor, y_tensor, reduce_func, ReducePacksPerXDim{});
return y_tensor;
}
};
// BlockReduce2dSync: Warp-level reduction (Stage 2)
template <typename Problem_, typename Policy_ = void>
struct BlockReduce2dSync
{
using Problem = remove_cvref_t<Problem_>;
template <typename YDistributedTensor_, typename ReduceFunc>
CK_TILE_DEVICE void operator()(YDistributedTensor_& y_tensor, const ReduceFunc& reduce_func)
{
using Dstr = typename YDistributedTensor_::StaticTileDistribution;
using DstrEncode = typename Dstr::DstrEncode;
using DstrEncodeDetail = typename DstrEncode::detail;
constexpr index_t NDimP = Dstr::get_num_of_dimension_p();
constexpr index_t NDimR = Dstr::get_num_of_dimension_r();
constexpr index_t idim_p_lane = NDimP - 1;
// const auto ps_idx = make_array<index_t>(get_warp_id(), get_lane_id());
// const auto rs_idx =
// y_tensor.get_tile_distribution().calculate_rs_index_from_ps_index(ps_idx);
constexpr index_t thread_buf_size = YDistributedTensor_::get_thread_buffer_size();
// loop over thread data
static_for<0, thread_buf_size, 1>{}([&](auto i) {
auto v_local = y_tensor.get_thread_buffer()[i];
// cross-lane reduce for replication
// only reduce on R dimension correspond to lane
// (lane id maps to this R dimension)
static_for<0, NDimR, 1>{}([&](auto idim_r) {
// FIXME: nasty to use does_p_own_r_
if constexpr(DstrEncodeDetail::does_p_own_r_[idim_p_lane][idim_r])
{
constexpr index_t r_length = DstrEncode::rs_lengths_[idim_r];
constexpr index_t lid_over_rid_derivative =
DstrEncodeDetail::ps_over_rs_derivative_[idim_p_lane][idim_r];
static_assert(is_power_of_two_integer(r_length),
"wrong! only support power of 2 reduction");
constexpr index_t nstage = integer_log2_floor(r_length);
// reduction sweep forward
static_for<0, nstage, 1>{}([&](auto istage) {
// xor
index_t src_lane =
(__lane_id()) ^
(number<lid_over_rid_derivative << istage.value>{}.value);
// pull data from remote lane
const auto v_remote = warp_shuffle(v_local, src_lane);
v_local = reduce_func(v_local, v_remote);
});
}
});
// TODO - Do we need to broadcast to other lane?
y_tensor.get_thread_buffer()(i) = v_local;
});
}
};
// BlockReduce2dCrossWarpSync: Cross-warp reduction (Stage 3)
template <typename Problem_, typename Policy_ = void>
struct BlockReduce2dCrossWarpSync
{
using Problem = remove_cvref_t<Problem_>;
using BlockShape = typename Problem::BlockShape;
template <typename YDistributedTensor_>
CK_TILE_DEVICE static constexpr index_t GetReduceWarps()
{
constexpr index_t num_reduce_warps = [&]() {
using Dstr = typename YDistributedTensor_::StaticTileDistribution;
using DstrEncode = typename Dstr::DstrEncode;
using DstrEncodeDetail = typename DstrEncode::detail;
constexpr index_t NDimR = Dstr::get_num_of_dimension_r();
constexpr index_t idim_p_warp = 0;
index_t len_ = 1;
static_for<0, NDimR, 1>{}([&](auto idim_r) {
if constexpr(DstrEncodeDetail::does_p_own_r_[idim_p_warp][idim_r])
{
constexpr index_t r_length = DstrEncode::rs_lengths_[idim_r];
len_ *= r_length;
}
});
return len_;
}();
return num_reduce_warps;
}
// return in byte
template <typename YDistributedTensor_>
CK_TILE_HOST_DEVICE static constexpr index_t GetSmemSize()
{
using DataType = typename YDistributedTensor_::DataType;
// constexpr auto num_reduce_warps = GetReduceWarps<YDistributedTensor_>();
constexpr index_t thread_buf_size = YDistributedTensor_::get_thread_buffer_size();
// we need to store all data from every wave into smem
// e.g. 2x2 reduce along N
// -------------> reduce N
// | w0 | w1 | ___> | w01 |
// | w2 | w3 | | w23 |
//
// -> store data from every wave into LDS
//
//
// -------------> reduce N
// | w0 | w1 | w2 | w3 | -----> | w0123 |
//
// -> also store data from every wave into LDS
constexpr index_t num_warps = BlockShape::BlockSize / get_warp_size();
return num_warps * thread_buf_size * sizeof(DataType);
}
template <typename YDistributedTensor_, typename ReduceFunc>
CK_TILE_DEVICE void
operator()(YDistributedTensor_& y_tensor, void* smem, const ReduceFunc& reduce_func)
{
using DataType = typename YDistributedTensor_::DataType;
constexpr index_t thread_buf_size = YDistributedTensor_::get_thread_buffer_size();
DataType* smem_ptr = reinterpret_cast<DataType*>(smem);
const index_t lane_id = get_lane_id();
const index_t warp_id = get_warp_id();
constexpr auto num_reduce_warps = GetReduceWarps<YDistributedTensor_>();
constexpr index_t num_warps = BlockShape::BlockSize / get_warp_size();
const index_t smem_offset = warp_id;
// skip if nonthing to do
if constexpr(num_reduce_warps == 1)
return;
// store into smem only for lane-0 within one warp
if(lane_id == 0)
{
static_for<0, thread_buf_size, 1>{}([&](auto i) {
smem_ptr[smem_offset + i * num_warps] = y_tensor.get_thread_buffer()[i];
});
}
block_sync_lds();
// load from smem. here we let everythread to do compute :)
index_t local_warp_id = warp_id / num_reduce_warps;
index_t local_smem_os = local_warp_id * num_reduce_warps;
DataType all_scratch[thread_buf_size * num_reduce_warps];
static_for<0, thread_buf_size, 1>{}([&](auto i_0) {
static_for<0, num_reduce_warps, 1>{}([&](auto i_1) {
all_scratch[i_0 * num_reduce_warps + i_1] =
smem_ptr[i_0 * num_warps + local_smem_os + i_1];
});
});
block_sync_lds(); // TODO: we don't need sync here
static_for<0, thread_buf_size, 1>{}([&](auto i_0) {
// TODO: use descriptor for this
auto v_local = all_scratch[i_0 * num_reduce_warps];
// further reduce mean/var
static_for<0, num_reduce_warps - 1, 1>{}([&](auto i_1_n1) {
constexpr auto i_1 = number<i_1_n1 + 1>{};
const DataType v_remote = all_scratch[i_0 * num_reduce_warps + i_1];
v_local = reduce_func(v_local, v_remote);
});
y_tensor.get_thread_buffer()(i_0) = v_local;
});
}
};
template <typename Problem_, typename Policy_ = void>
struct BlockReduce2dTreeCrossWarpSync
{
using Problem = remove_cvref_t<Problem_>;
using BlockShape = typename Problem::BlockShape;
template <typename YDistributedTensor_>
CK_TILE_DEVICE static constexpr index_t GetReduceWarps()
{
constexpr index_t num_reduce_warps = [&]() {
using Dstr = typename YDistributedTensor_::StaticTileDistribution;
using DstrEncode = typename Dstr::DstrEncode;
using DstrEncodeDetail = typename DstrEncode::detail;
constexpr index_t NDimR = Dstr::get_num_of_dimension_r();
constexpr index_t idim_p_warp = 0;
index_t len_ = 1;
static_for<0, NDimR, 1>{}([&](auto idim_r) {
if constexpr(DstrEncodeDetail::does_p_own_r_[idim_p_warp][idim_r])
{
constexpr index_t r_length = DstrEncode::rs_lengths_[idim_r];
len_ *= r_length;
}
});
return len_;
}();
return num_reduce_warps;
}
// return in byte
template <typename YDistributedTensor_>
CK_TILE_HOST_DEVICE static constexpr index_t GetSmemSize()
{
using DataType = typename YDistributedTensor_::DataType;
constexpr index_t thread_buf_size = YDistributedTensor_::get_thread_buffer_size();
// we need to store all data from every wave into smem
// e.g. 2x2 reduce along N
// -------------> reduce N
// | w0 | w1 | ___> | w01 |
// | w2 | w3 | | w23 |
//
// -> store data from every wave into LDS
//
//
// -------------> reduce N
// | w0 | w1 | w2 | w3 | -----> | w0123 |
//
// -> also store data from every wave into LDS
constexpr index_t num_warps = BlockShape::BlockSize / warpSize;
return num_warps * thread_buf_size * sizeof(DataType);
}
template <typename YDistributedTensor_, typename ReduceFunc>
CK_TILE_DEVICE void
operator()(YDistributedTensor_& y_tensor, void* smem, const ReduceFunc& reduce_func)
{
using Dstr = typename YDistributedTensor_::StaticTileDistribution;
using DstrEncode = typename Dstr::DstrEncode;
using DstrEncodeDetail = typename DstrEncode::detail;
using DataType = typename YDistributedTensor_::DataType;
constexpr index_t NDimP = Dstr::get_num_of_dimension_p();
constexpr index_t NDimR = Dstr::get_num_of_dimension_r();
constexpr index_t idim_p_lane = NDimP - 1;
constexpr index_t thread_buf_size = YDistributedTensor_::get_thread_buffer_size();
DataType* smem_ptr = reinterpret_cast<DataType*>(smem);
const index_t lane_id = get_lane_id();
const index_t warp_id = get_warp_id();
constexpr index_t num_warps = BlockShape::BlockSize / get_warp_size();
constexpr index_t num_reduce_warps = GetReduceWarps<YDistributedTensor_>();
if constexpr(num_reduce_warps == 1)
return;
// Each warp's lane 0 writes its partial results to shared memory
const index_t smem_offset = warp_id;
if(lane_id == 0)
{
static_for<0, thread_buf_size, 1>{}([&](auto i) {
// Store the i-th element of this warp's thread_buffer into SMEM
smem_ptr[smem_offset + i * num_warps] = y_tensor.get_thread_buffer()[i];
});
}
block_sync_lds();
// We let each warp holds a duplication to do reduction.
static_for<0, thread_buf_size, 1>{}([&](auto i) {
DataType v = 0;
if(lane_id < num_reduce_warps)
{
v = smem_ptr[lane_id + i * num_warps];
}
// cross-lane reduce for replication
// only reduce on R dimension correspond to lane
// (lane id maps to this R dimension)
static_for<0, NDimR, 1>{}([&](auto idim_r) {
// FIXME: nasty to use does_p_own_r_
if constexpr(DstrEncodeDetail::does_p_own_r_[idim_p_lane][idim_r])
{
constexpr index_t r_length = DstrEncode::rs_lengths_[idim_r];
constexpr index_t lid_over_rid_derivative =
DstrEncodeDetail::ps_over_rs_derivative_[idim_p_lane][idim_r];
static_assert(is_power_of_two_integer(r_length),
"wrong! only support power of 2 reduction");
constexpr index_t nstage = integer_log2_floor(r_length);
// reduction sweep forward
static_for<0, nstage, 1>{}([&](auto istage) {
// pull data from remote lane
const auto o =
__shfl_xor(v, number<lid_over_rid_derivative << istage.value>{}.value);
// reduce
v = reduce_func(v, o);
});
}
});
y_tensor.get_thread_buffer()(i) = v;
});
}
};
} // namespace ck_tile