ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-11 17:00:18 +00:00
Code Issues Packages Projects Releases Wiki Activity

enable prefill overload operator().

Browse Source
This commit is contained in:
aska-0096
2025-07-30 03:51:06 +00:00
parent 13bcc913de
commit 8dacc35c4c
4 changed files with 701 additions and 159 deletions
Download Patch File Download Diff File Expand all files Collapse all files

65
include/ck_tile/ops/fmha/kernel/fmha_fwd_decode_kernel.hpp
View File

@@ -582,9 +582,6 @@ struct FmhaFwdDecodeKernel
CK_TILE_DEVICE void operator()(Kargs kargs) const
{
// allocate LDS
__shared__ char smem_ptr[GetSmemSize()];
// divide problem
const auto [i_tile_m, i_tile_n, i_split, i_nhead, i_batch] = GetTileIndex(kargs);
@@ -948,23 +945,51 @@ struct FmhaFwdDecodeKernel
}();
auto o_acc_tile = [&, i_split_ = i_split]() {
return FmhaPipeline{}(q_dram_window,
k_dram_window,
// k_page_block_navigator, // Remove it
v_dram_window,
// v_page_block_navigator, // Remove it
bias_dram_window,
lse_acc_dram_window,
kargs.num_splits, // Remove it
i_split_, // Remove it
mask,
position_encoding,
kargs.scale_s,
// variant, // Remove it
// variant_params, // Remove it
// block_indices, // Remove it
// kv_l2p_offset, // Remove it
smem_ptr);
if constexpr(FmhaPipeline::kM0 == 128)
{
// allocate double lds
// add __restrict__ here to avoid aliasing
__shared__ char
smem_ptrk0[FmhaPipeline::Policy::
template GetSmemSizeK<typename FmhaPipeline::Problem, true>()];
__shared__ char
smem_ptrk1[FmhaPipeline::Policy::
template GetSmemSizeK<typename FmhaPipeline::Problem, true>()];
__shared__ char smem_ptrv0
[FmhaPipeline::Policy::template GetSmemSizeV<typename FmhaPipeline::Problem>()];
__shared__ char smem_ptrv1
[FmhaPipeline::Policy::template GetSmemSizeV<typename FmhaPipeline::Problem>()];
return FmhaPipeline{}(q_dram_window,
k_dram_window,
v_dram_window,
bias_dram_window,
lse_acc_dram_window,
kargs.num_splits, // Remove it
i_split_, // Remove it
mask,
position_encoding,
kargs.scale_s,
smem_ptrk0,
smem_ptrk1,
smem_ptrv0,
smem_ptrv1);
}
else
{
__shared__ char smem_ptr[GetSmemSize()];
return FmhaPipeline{}(q_dram_window,
k_dram_window,
v_dram_window,
bias_dram_window,
lse_acc_dram_window,
kargs.num_splits, // Remove it
i_split_, // Remove it
mask,
position_encoding,
kargs.scale_s,
smem_ptr);
}
}();
// Oacc DRAM and Oacc DRAM window

683
include/ck_tile/ops/fmha/pipeline/block_fmha_fwd_decode_pipeline_qr_ks_vs.hpp
View File

@@ -14,6 +14,9 @@ namespace ck_tile {
template <typename Problem_, typename Policy_ = BlockFmhaFwdDecodePipelineQRKSVSDefaultPolicy>
struct BlockFmhaFwdDecodePipelineQRKSVS
{
static constexpr auto I0 = number<0>{};
static constexpr auto I1 = number<1>{};
using Problem = remove_cvref_t<Problem_>;
using Policy = remove_cvref_t<Policy_>;
using QDataType = remove_cvref_t<typename Problem::QDataType>;
@@ -43,8 +46,8 @@ struct BlockFmhaFwdDecodePipelineQRKSVS
static constexpr index_t kK1 = BlockFmhaShape::kK1;
static constexpr index_t kQKHeaddim = BlockFmhaShape::kQKHeaddim;
static constexpr index_t kSubQKHeaddim = BlockFmhaShape::kSubQKHeaddim;
static constexpr index_t kNWarp = BlockFmhaShape::Gemm0BlockWarps::at(number<1>{});
static constexpr index_t kNXdl = BlockFmhaShape::Gemm0WarpTile::at(number<1>{});
static constexpr index_t kNWarp = BlockFmhaShape::Gemm0BlockWarps::at(I1);
static constexpr index_t kNXdl = BlockFmhaShape::Gemm0WarpTile::at(I1);
static_assert(kSubQKHeaddim <= 256, "hdim bigger than 256 is not suitable for this pipeline!");
@@ -124,6 +127,7 @@ struct BlockFmhaFwdDecodePipelineQRKSVS
return Policy::template GetSmemSize<Problem>();
}
// Decode
template <typename QDramBlockWindowTmp,
typename KDramBlockWindowTmp,
typename VDramBlockWindowTmp,
@@ -149,15 +153,14 @@ struct BlockFmhaFwdDecodePipelineQRKSVS
std::is_same_v<VDataType, remove_cvref_t<typename VDramBlockWindowTmp::DataType>>,
"wrong!");
static_assert(kM0 == QDramBlockWindowTmp{}.get_window_lengths()[number<0>{}] &&
kSubQKHeaddim ==
QDramBlockWindowTmp{}.get_window_lengths()[number<1>{}] &&
kN0 == KDramBlockWindowTmp{}.get_window_lengths()[number<0>{}] &&
kK0 == KDramBlockWindowTmp{}.get_window_lengths()[number<1>{}] &&
kN1 == VDramBlockWindowTmp{}.get_window_lengths()[number<0>{}] &&
kK1 == VDramBlockWindowTmp{}.get_window_lengths()[number<1>{}] &&
kM0 == BiasDramBlockWindowTmp{}.get_window_lengths()[number<0>{}] &&
kN0 == BiasDramBlockWindowTmp{}.get_window_lengths()[number<1>{}],
static_assert(kM0 == QDramBlockWindowTmp{}.get_window_lengths()[I0] &&
kSubQKHeaddim == QDramBlockWindowTmp{}.get_window_lengths()[I1] &&
kN0 == KDramBlockWindowTmp{}.get_window_lengths()[I0] &&
kK0 == KDramBlockWindowTmp{}.get_window_lengths()[I1] &&
kN1 == VDramBlockWindowTmp{}.get_window_lengths()[I0] &&
kK1 == VDramBlockWindowTmp{}.get_window_lengths()[I1] &&
kM0 == BiasDramBlockWindowTmp{}.get_window_lengths()[I0] &&
kN0 == BiasDramBlockWindowTmp{}.get_window_lengths()[I1],
"wrong!");
ignore = bias_dram_block_window_tmp;
ignore = position_encoding;
@@ -192,7 +195,7 @@ struct BlockFmhaFwdDecodePipelineQRKSVS
const auto q_origin = q_dram_block_window_tmp.get_window_origin();
const auto [logical_seqlen_k_start, logical_seqlen_k_end] = mask.GetTileRangeAlongX(
q_origin.at(number<0>{}), number<kM0>{}, number<kN0>{}, num_splits, i_split);
q_origin.at(I0), number<kM0>{}, number<kN0>{}, num_splits, i_split);
// check early exit if no work to do
if constexpr(FmhaMask::IsMasking || kPadSeqLenK || kHasUnevenSplits)
@@ -287,20 +290,10 @@ struct BlockFmhaFwdDecodePipelineQRKSVS
auto v_lds_read_window =
make_tile_window(v_lds,
Policy::template MakeVLdsBlockDescriptor<Problem>().get_lengths(),
make_tuple(number<kK1>{}, number<kN1>{}),
{0, 0},
Policy::template MakeVRegTileDistribution<Problem>());
// Bias tile in LDS
// const auto bias_origin = bias_dram_block_window_tmp.get_window_origin();
// auto bias_dram_window =
// make_tile_window(bias_dram_block_window_tmp.get_bottom_tensor_view(),
// bias_dram_block_window_tmp.get_window_lengths(),
// {bias_origin.at(number<0>{}),
// logical_seqlen_k_start - (physical_seqlen_k_start -
// aligned_physical_seqlen_k_start)}, // M/N
// Policy::template MakeBiasDramTileDistribution<decltype(gemm_0)>());
block_sync_lds_direct_load<0>();
auto q_tile = load_tile(q_lds_read_window);
@@ -312,7 +305,7 @@ struct BlockFmhaFwdDecodePipelineQRKSVS
constexpr index_t k1_loops = kN0 / kK1;
static_assert(1 <= k0_loops);
static_assert(1 == k1_loops);
static_assert(1 <= k1_loops);
block_sync_lds();
async_load_tile(k_lds_write_window, k_dram_window);
@@ -326,40 +319,31 @@ struct BlockFmhaFwdDecodePipelineQRKSVS
async_load_tile(v_lds_write_window, v_dram_window); // prefetch load v tile
// move V tile windows
move_tile_window(v_dram_window, {kK1, 0});
move_tile_window(v_dram_window, {kN0, 0});
// STAGE 1, QK gemm
clear_tile(s_acc); // initialize C
// if constexpr(BiasEnum == BlockAttentionBiasEnum::ELEMENTWISE_BIAS)
// {
// __builtin_amdgcn_sched_barrier(
// 0); // prevent from messing up the order of global loads
// }
// const auto bias_tile = load_tile(bias_dram_window); // load bias tile
// if constexpr(BiasEnum == BlockAttentionBiasEnum::ELEMENTWISE_BIAS)
// {
// __builtin_amdgcn_sched_barrier(
// 0); // prevent from messing up the order of global loads
// }
if constexpr(1 < k0_loops)
{
static_for<0, k0_loops - 1, 1>{}([&](auto i_k0) {
if constexpr(i_k0 == 0){
if constexpr(i_k0 == 0)
{
block_sync_lds_direct_load<v_vmem_insts>();
}
else{
else
{
block_sync_lds_direct_load<0>();
}
auto k_tile = load_tile(k_lds_read_window);
gemm_0(s_acc,
get_slice_tile(q_tile,
sequence<0, i_k0 * kK0>{},
sequence<kM0, (i_k0 + 1) * kK0>{}),
k_tile);
// loop over along the [K]ey head dimension
move_tile_window(k_dram_window, {0, kK0});
block_sync_lds();
@@ -369,56 +353,23 @@ struct BlockFmhaFwdDecodePipelineQRKSVS
move_tile_window(k_dram_window, {0, -kK0 * (k0_loops - 1)});
}
if constexpr(k0_loops==1){
if constexpr(k0_loops == 1)
{
block_sync_lds_direct_load<v_vmem_insts>();
}
else{
else
{
block_sync_lds_direct_load<0>();
}
auto k_tile = load_tile(k_lds_read_window);
gemm_0(s_acc,
get_slice_tile(q_tile,
sequence<0, (k0_loops - 1) * kK0>{},
sequence<kM0, k0_loops * kK0>{}),
k_tile);
// // STAGE 2, scale_s, add bias, mask, softmax
// if constexpr(BiasEnum == BlockAttentionBiasEnum::ELEMENTWISE_BIAS)
// {
// tile_elementwise_inout([&scale_s](auto& x) { x = x * scale_s; }, s_acc);
// tile_elementwise_inout(
// [&](auto& x, const auto& y) {
// x += log2e_v<SaccDataType> *
// type_convert<SaccDataType>(bias_element_func(y));
// },
// s_acc,
// bias_tile);
// }
// else if constexpr(BiasEnum == BlockAttentionBiasEnum::ALIBI)
// {
// const auto k_origin = k_page_block_navigator.to_global_window_origin(
// i_page_block_k, k_dram_block_window.get_window_origin());
// constexpr auto s_spans = decltype(s_acc)::get_distributed_spans();
// sweep_tile_span(s_spans[number<0>{}], [&](auto idx0) {
// sweep_tile_span(s_spans[number<1>{}], [&](auto idx1) {
// const auto tile_idx = get_x_indices_from_distributed_indices(
// s_acc.get_tile_distribution(), make_tuple(idx0, idx1));
// const auto row = q_origin.at(number<0>{}) + tile_idx.at(number<0>{});
// const auto col = k_origin.at(number<0>{}) + tile_idx.at(number<1>{});
// constexpr auto i_j_idx = make_tuple(idx0, idx1);
// s_acc(i_j_idx) *= scale_s;
// // position_encoding accept only logical coordinates, do conversion here
// position_encoding.update(s_acc(i_j_idx), row, col - kv_l2p_offset);
// });
// });
// }
// move_tile_window(bias_dram_window, {0, kN0});
/// TODO: only check in first/last iteration without increasing code size
if constexpr(kHasUnevenSplits)
{
if(i_total_loops == (num_total_loop - 1))
@@ -429,8 +380,7 @@ struct BlockFmhaFwdDecodePipelineQRKSVS
[&,
physical_seqlen_k_start_ = physical_seqlen_k_start,
physical_seqlen_k_end_ = physical_seqlen_k_end](auto tile_idx) {
const auto col =
k_origin.at(number<0>{}) + tile_idx.at(number<1>{});
const auto col = k_origin.at(I0) + tile_idx.at(I1);
if constexpr(kIsPagedKV)
{
return col < physical_seqlen_k_start_ ||
@@ -447,19 +397,15 @@ struct BlockFmhaFwdDecodePipelineQRKSVS
if constexpr(kPadSeqLenK || FmhaMask::IsMasking)
{
const auto k_origin = make_tuple(kN0 * i_total_loops, 0);
// const auto k_origin = k_page_block_navigator.to_global_window_origin(
// i_page_block_k, k_dram_block_window.get_window_origin());
// mask accept only logical coordinates, do conversion here
bool need_perpixel_check = mask.IsEdgeTile(q_origin.at(number<0>{}),
k_origin.at(number<0>{}),
number<kM0>{},
number<kN0>{});
bool need_perpixel_check =
mask.IsEdgeTile(q_origin.at(I0), k_origin.at(I0), number<kM0>{}, number<kN0>{});
if(need_perpixel_check)
{
set_tile_if(
s_acc, -numeric<SMPLComputeDataType>::infinity(), [&](auto tile_idx) {
const auto row = q_origin.at(number<0>{}) + tile_idx.at(number<0>{});
const auto col = k_origin.at(number<0>{}) + tile_idx.at(number<1>{});
const auto row = q_origin.at(I0) + tile_idx.at(I0);
const auto col = k_origin.at(I0) + tile_idx.at(I1);
return mask.IsOutOfBound(row, col);
});
}
@@ -519,10 +465,10 @@ struct BlockFmhaFwdDecodePipelineQRKSVS
};
constexpr auto p_spans = decltype(p_compute)::get_distributed_spans();
sweep_tile_span(p_spans[number<0>{}], [&](auto idx0) {
sweep_tile_span(p_spans[I0], [&](auto idx0) {
constexpr auto i_idx = make_tuple(idx0);
auto row_max = scale_s * get_validated_m(m[i_idx]);
sweep_tile_span(p_spans[number<1>{}], [&](auto idx1) {
sweep_tile_span(p_spans[I1], [&](auto idx1) {
constexpr auto i_j_idx = make_tuple(idx0, idx1);
if constexpr(BiasEnum == BlockAttentionBiasEnum::ELEMENTWISE_BIAS ||
BiasEnum == BlockAttentionBiasEnum::ALIBI)
@@ -554,7 +500,7 @@ struct BlockFmhaFwdDecodePipelineQRKSVS
// l{j}, Oacc{j}
constexpr auto o_spans = decltype(o_acc)::get_distributed_spans();
sweep_tile_span(o_spans[number<0>{}], [&](auto idx0) {
sweep_tile_span(o_spans[I0], [&](auto idx0) {
constexpr auto i_idx = make_tuple(idx0);
const auto tmp = [&]() {
if constexpr(BiasEnum == BlockAttentionBiasEnum::ELEMENTWISE_BIAS ||
@@ -576,24 +522,39 @@ struct BlockFmhaFwdDecodePipelineQRKSVS
}
}();
l(i_idx) = tmp * l[i_idx] + rowsum_p[i_idx];
sweep_tile_span(o_spans[number<1>{}], [&](auto idx1) {
sweep_tile_span(o_spans[I1], [&](auto idx1) {
constexpr auto i_j_idx = make_tuple(idx0, idx1);
// FIXME: this use different equation from FA v2 paper,
// but produce correc result.
// Is the equation wrong?
o_acc(i_j_idx) *= tmp;
});
});
block_sync_lds_direct_load<k_vmem_insts>();
auto v_tile = load_tile_transpose(v_lds_read_window);
gemm_1(
o_acc,
get_slice_tile(
p_tile, sequence<0, (k1_loops - 1) * kK1>{}, sequence<kM0, k1_loops * kK1>{}),
get_slice_tile(
v_tile, sequence<0, (k1_loops - 1) * kK1>{}, sequence<kN1, k1_loops * kK1>{}));
if constexpr(1 < k1_loops)
{
static_for<0, k1_loops - 1, 1>{}([&](auto i_k1) {
gemm_1(o_acc,
get_slice_tile(p_tile,
sequence<0, i_k1 * kK1>{},
sequence<kM0, (i_k1 + 1) * kK1>{}),
v_tile);
// loop over along the [V]alue Sequence length
move_tile_window(v_lds_read_window, {kK1, 0});
v_tile = load_tile_transpose(v_lds_read_window);
});
// move back to the origin
move_tile_window(v_lds_read_window, {-kK1 * (k1_loops - 1), 0});
}
gemm_1(o_acc,
get_slice_tile(p_tile,
sequence<0, (k1_loops - 1) * kK1>{},
sequence<kM0, k1_loops * kK1>{}),
v_tile);
} while(++i_total_loops < num_total_loop);
@@ -603,7 +564,7 @@ struct BlockFmhaFwdDecodePipelineQRKSVS
auto lse_acc = make_static_distributed_tensor<LSEDataType>(m.get_tile_distribution());
constexpr auto lse_acc_spans = decltype(lse_acc)::get_distributed_spans();
sweep_tile_span(lse_acc_spans[number<0>{}], [&, m_ = m, l_ = l](auto idx0) {
sweep_tile_span(lse_acc_spans[I0], [&, m_ = m, l_ = l](auto idx0) {
constexpr auto i_idx = make_tuple(idx0);
if constexpr(BiasEnum == BlockAttentionBiasEnum::ELEMENTWISE_BIAS ||
BiasEnum == BlockAttentionBiasEnum::ALIBI)
@@ -632,7 +593,7 @@ struct BlockFmhaFwdDecodePipelineQRKSVS
// finally, O
constexpr auto o_spans = decltype(o_acc)::get_distributed_spans();
sweep_tile_span(o_spans[number<0>{}], [&](auto idx0) {
sweep_tile_span(o_spans[I0], [&](auto idx0) {
constexpr auto i_idx = make_tuple(idx0);
const auto tmp = [&]() {
if constexpr(BiasEnum == BlockAttentionBiasEnum::ELEMENTWISE_BIAS ||
@@ -643,7 +604,519 @@ struct BlockFmhaFwdDecodePipelineQRKSVS
else
return 1 / l[i_idx];
}();
sweep_tile_span(o_spans[number<1>{}], [&](auto idx1) {
sweep_tile_span(o_spans[I1], [&](auto idx1) {
constexpr auto i_j_idx = make_tuple(idx0, idx1);
o_acc(i_j_idx) *= tmp;
});
});
return o_acc;
}
// Prefill, double lds
template <typename QDramBlockWindowTmp,
typename KDramBlockWindowTmp,
typename VDramBlockWindowTmp,
typename BiasDramBlockWindowTmp,
typename LSEaccDramBlockWindowTmp,
typename PositionEncoding>
CK_TILE_HOST_DEVICE auto
operator()(const QDramBlockWindowTmp& q_dram_block_window_tmp, // M0*K0 tile
const KDramBlockWindowTmp& k_dram_block_window_tmp, // N0*K0 tile
const VDramBlockWindowTmp& v_dram_block_window_tmp, // N1*K1 tile
const BiasDramBlockWindowTmp& bias_dram_block_window_tmp, // M0*N0 tile
LSEaccDramBlockWindowTmp& lse_acc_dram_window_tmp, // M0*1 tile
index_t num_splits,
index_t i_split,
FmhaMask mask,
PositionEncoding position_encoding,
float scale_s,
void* __restrict__ smem_ptrk0,
void* __restrict__ smem_ptrk1,
void* __restrict__ smem_ptrv0,
void* __restrict__ smem_ptrv1) const
{
static_assert(
std::is_same_v<QDataType, remove_cvref_t<typename QDramBlockWindowTmp::DataType>> &&
std::is_same_v<KDataType, remove_cvref_t<typename KDramBlockWindowTmp::DataType>> &&
std::is_same_v<VDataType, remove_cvref_t<typename VDramBlockWindowTmp::DataType>>,
"wrong!");
static_assert(kM0 == QDramBlockWindowTmp{}.get_window_lengths()[I0] &&
kSubQKHeaddim == QDramBlockWindowTmp{}.get_window_lengths()[I1] &&
kN0 == KDramBlockWindowTmp{}.get_window_lengths()[I0] &&
kK0 == KDramBlockWindowTmp{}.get_window_lengths()[I1] &&
kN1 == VDramBlockWindowTmp{}.get_window_lengths()[I0] &&
kK1 == VDramBlockWindowTmp{}.get_window_lengths()[I1] &&
kM0 == BiasDramBlockWindowTmp{}.get_window_lengths()[I0] &&
kN0 == BiasDramBlockWindowTmp{}.get_window_lengths()[I1],
"wrong!");
ignore = bias_dram_block_window_tmp;
ignore = position_encoding;
// Block GEMM
constexpr auto gemm_0 = Policy::template GetQKBlockGemm<Problem>();
constexpr auto gemm_1 = Policy::template GetPVBlockGemm<Problem>();
using SaccBlockTileType = decltype(gemm_0.MakeCBlockTile());
auto s_acc = SaccBlockTileType{};
// reduction function for softmax
const auto f_max = [](auto e0, auto e1) { return max(e0, e1); };
const auto f_sum = [](auto e0, auto e1) { return e0 + e1; };
using OaccBlockTileType = decltype(gemm_1.MakeCBlockTile());
auto o_acc = OaccBlockTileType{};
// infer Sacc, S, P, M, L, Oacc type
using SBlockTileType = decltype(cast_tile<SMPLComputeDataType>(o_acc));
using MLBlockTileType = decltype(block_tile_reduce<SMPLComputeDataType>(
SBlockTileType{}, sequence<1>{}, f_max, SMPLComputeDataType{0}));
// init M, L
auto m = MLBlockTileType{};
auto l = MLBlockTileType{};
clear_tile(o_acc);
set_tile(m, -numeric<SMPLComputeDataType>::infinity());
clear_tile(l);
const auto q_origin = q_dram_block_window_tmp.get_window_origin();
const auto [logical_seqlen_k_start, logical_seqlen_k_end] = mask.GetTileRangeAlongX(
q_origin.at(I0), number<kM0>{}, number<kN0>{}, num_splits, i_split);
// check early exit if no work to do
if constexpr(FmhaMask::IsMasking || kPadSeqLenK || kHasUnevenSplits)
{
const index_t logical_num_total_loop =
integer_divide_ceil(logical_seqlen_k_end - logical_seqlen_k_start, kN0);
if(logical_num_total_loop <= 0)
{
if constexpr(kStoreLSE)
{
auto lse_acc =
make_static_distributed_tensor<LSEDataType>(m.get_tile_distribution());
set_tile(lse_acc, -numeric<SMPLComputeDataType>::infinity());
if(get_thread_local_1d_id() < kM0)
{
store_tile(lse_acc_dram_window_tmp, lse_acc);
}
}
// Note: here occ are all cleard, return it
// Note: q loaded but no fence, ignore it.
return o_acc;
}
}
// Q tile in LDS
auto q_dram_window = make_tile_window(
q_dram_block_window_tmp, Policy::template MakeQDramTileDistribution<Problem, true>());
// auto q_lds = make_tensor_view<address_space_enum::lds>(
// static_cast<QDataType*>(smem_ptrk0),
// Policy::template MakeQLdsBlockDescriptor<Problem>());
// auto q_lds_store_window = make_tile_window(
// q_lds, Policy::template MakeQLdsBlockDescriptor<Problem>().get_lengths(), {0, 0});
// auto q_lds_read_window =
// make_tile_window(q_lds,
// Policy::template MakeQLdsBlockDescriptor<Problem>().get_lengths(),
// {0, 0},
// Policy::template MakeQRegTileDistribution<Problem>());
// async_load_tile(q_lds_store_window, q_dram_window);
auto q_tile = load_tile(q_dram_window);
// K tile in LDS
const index_t physical_seqlen_k_start = logical_seqlen_k_start;
const index_t physical_seqlen_k_end = logical_seqlen_k_end;
// make sure the first tile is completely located in page-block (page-block size should be
// divisible by kN0)
// relationship between each *_start variables: aligned_physical_seqlen_k_start <=
// physical_seqlen_k_start, logical_seqlen_k_start <= physical_seqlen_k_start
const index_t aligned_physical_seqlen_k_start = physical_seqlen_k_start;
auto k_dram_window = make_tile_window(
k_dram_block_window_tmp, Policy::template MakeKDramTileDistribution<Problem, true>());
auto k_lds = make_tuple(make_tensor_view<address_space_enum::lds>(
static_cast<KDataType* __restrict__>(smem_ptrk0),
Policy::template MakeKLdsBlockDescriptor<Problem, true>()),
make_tensor_view<address_space_enum::lds>(
static_cast<KDataType* __restrict__>(smem_ptrk1),
Policy::template MakeKLdsBlockDescriptor<Problem, true>()));
auto k_lds_write_windows =
make_tuple(make_tile_window(
k_lds.at(I0),
Policy::template MakeKLdsBlockDescriptor<Problem, true>().get_lengths(),
{0, 0}),
make_tile_window(
k_lds.at(I1),
Policy::template MakeKLdsBlockDescriptor<Problem, true>().get_lengths(),
{0, 0}));
auto k_lds_read_windows =
make_tuple(make_tile_window(k_lds.at(I0),
make_tuple(number<kN0>{}, number<kK0>{}),
{0, 0},
Policy::template MakeKRegTileDistribution<Problem>()),
make_tile_window(k_lds.at(I1),
make_tuple(number<kN0>{}, number<kK0>{}),
{0, 0},
Policy::template MakeKRegTileDistribution<Problem>()));
// S tile in LDS
auto s_lds = make_tensor_view<address_space_enum::lds>(
reinterpret_cast<SaccDataType*>(reinterpret_cast<char*>(smem_ptrk0) +
Policy::template GetSmemSizeK<Problem>()),
Policy::template MakeSLdsBlockDescriptor<Problem>());
auto s_write_lds_window = make_tile_window(
s_lds, Policy::template MakeSLdsBlockDescriptor<Problem>().get_lengths(), {0, 0});
auto s_read_lds_window =
make_tile_window(s_lds,
Policy::template MakeSLdsBlockDescriptor<Problem>().get_lengths(),
{0, 0},
Policy::template MakeSRegTileDistribution<Problem>());
// V tile in LDS
auto v_dram_window = make_tile_window(
v_dram_block_window_tmp, Policy::template MakeVDramTileDistribution<Problem>());
auto v_lds = make_tuple(
make_tensor_view<address_space_enum::lds>(
reinterpret_cast<VDataType* __restrict__>(static_cast<char*>(smem_ptrv0)),
Policy::template MakeVLdsBlockDescriptor<Problem>()),
make_tensor_view<address_space_enum::lds>(
reinterpret_cast<VDataType* __restrict__>(static_cast<char*>(smem_ptrv1)),
Policy::template MakeVLdsBlockDescriptor<Problem>()));
auto v_lds_write_windows = make_tuple(
make_tile_window(v_lds.at(I0),
Policy::template MakeVLdsBlockDescriptor<Problem>().get_lengths(),
{0, 0}),
make_tile_window(v_lds.at(I1),
Policy::template MakeVLdsBlockDescriptor<Problem>().get_lengths(),
{0, 0}));
auto v_lds_read_windows =
make_tuple(make_tile_window(v_lds.at(I0),
make_tuple(number<kK1>{}, number<kN1>{}),
{0, 0},
Policy::template MakeVRegTileDistribution<Problem>()),
make_tile_window(v_lds.at(I1),
make_tuple(number<kK1>{}, number<kN1>{}),
{0, 0},
Policy::template MakeVRegTileDistribution<Problem>()));
// block_sync_lds_direct_load<0>();
// auto q_tile = load_tile(q_lds_read_window);
const index_t num_total_loop =
integer_divide_ceil(physical_seqlen_k_end - aligned_physical_seqlen_k_start, kN0);
index_t i_total_loops = 0;
constexpr index_t k0_loops = kQKHeaddim / kK0;
constexpr index_t k1_loops = kN0 / kK1;
static_assert(1 <= k0_loops);
static_assert(1 <= k1_loops);
// block_sync_lds();
async_load_tile(k_lds_write_windows.at(I0), k_dram_window);
async_load_tile(v_lds_write_windows.at(I0), v_dram_window);
constexpr index_t k_vmem_insts = k_dram_window.get_num_of_access();
constexpr index_t v_vmem_insts = v_dram_window.get_num_of_access();
auto mainloop = [&](auto lds_write_buf, auto lds_read_buf) {
auto k_lds_write_window = k_lds_write_windows.at(lds_write_buf);
auto k_lds_read_window = k_lds_read_windows.at(lds_read_buf);
auto v_lds_write_window = v_lds_write_windows.at(lds_write_buf);
auto v_lds_read_window = v_lds_read_windows.at(lds_read_buf);
block_sync_lds();
// move K tile windows
move_tile_window(k_dram_window, {kN0, 0});
async_load_tile(k_lds_write_window, k_dram_window);
// STAGE 1, QK gemm
clear_tile(s_acc); // initialize C
block_sync_lds_direct_load<k_vmem_insts + v_vmem_insts>();
auto k_tile = load_tile(k_lds_read_window);
if constexpr(1 < k0_loops)
{
static_for<0, k0_loops - 1, 1>{}([&](auto i_k0) {
gemm_0(s_acc,
get_slice_tile(q_tile,
sequence<0, i_k0 * kK0>{},
sequence<kM0, (i_k0 + 1) * kK0>{}),
k_tile);
// loop over along the [K]ey head dimension
move_tile_window(k_lds_read_window, {0, kK0});
k_tile = load_tile(k_lds_read_window);
});
// move back to the origin
move_tile_window(k_lds_read_window, {0, -kK0 * (k0_loops - 1)});
}
gemm_0(s_acc,
get_slice_tile(q_tile,
sequence<0, (k0_loops - 1) * kK0>{},
sequence<kM0, k0_loops * kK0>{}),
k_tile);
if constexpr(kHasUnevenSplits)
{
if(i_total_loops == (num_total_loop - 1))
{
const auto k_origin = make_tuple(kN0 * i_total_loops, 0);
set_tile_if(s_acc,
-numeric<SMPLComputeDataType>::infinity(),
[&,
physical_seqlen_k_start_ = physical_seqlen_k_start,
physical_seqlen_k_end_ = physical_seqlen_k_end](auto tile_idx) {
const auto col = k_origin.at(I0) + tile_idx.at(I1);
if constexpr(kIsPagedKV)
{
return col < physical_seqlen_k_start_ ||
physical_seqlen_k_end_ <= col;
}
else
{
return physical_seqlen_k_end_ <= col;
}
});
}
}
if constexpr(kPadSeqLenK || FmhaMask::IsMasking)
{
const auto k_origin = make_tuple(kN0 * i_total_loops, 0);
bool need_perpixel_check =
mask.IsEdgeTile(q_origin.at(I0), k_origin.at(I0), number<kM0>{}, number<kN0>{});
if(need_perpixel_check)
{
set_tile_if(
s_acc, -numeric<SMPLComputeDataType>::infinity(), [&](auto tile_idx) {
const auto row = q_origin.at(I0) + tile_idx.at(I0);
const auto col = k_origin.at(I0) + tile_idx.at(I1);
return mask.IsOutOfBound(row, col);
});
}
}
// Gemm1
auto s_new = [&]() {
if constexpr(kNWarp > 1)
{
auto s = cast_tile<SMPLComputeDataType>(s_acc); // S{j}
store_tile(s_write_lds_window, s);
block_sync_lds();
return load_tile(s_read_lds_window);
}
else
{
return cast_tile<SMPLComputeDataType>(s_acc); // S{j}
}
}();
auto m_local = block_tile_reduce<SMPLComputeDataType>(
s_new,
sequence<1>{},
f_max,
-numeric<SMPLComputeDataType>::infinity()); // m_local = rowmax(S{j})
block_tile_reduce_sync(m_local, f_max, bool_constant<false>{});
const auto m_old = m; // m{j-1}
tile_elementwise_inout(
[](auto& e0, auto e1, auto e2) { e0 = max(e1, e2); }, m, m_old, m_local); // m{j}
auto p_compute = make_static_distributed_tensor<SMPLComputeDataType>(
s_new.get_tile_distribution()); // Pcompute{j}
static const auto get_validated_m = [](SMPLComputeDataType raw_m) {
/// NOTICE: bias might be materialized mask including -inf values, need
/// consideration
if constexpr(BiasEnum == BlockAttentionBiasEnum::ELEMENTWISE_BIAS ||
FmhaMask::IsMasking)
{
return raw_m == -numeric<SMPLComputeDataType>::infinity()
? type_convert<SMPLComputeDataType>(0.f)
: raw_m;
}
else
{
return raw_m;
}
};
constexpr auto p_spans = decltype(p_compute)::get_distributed_spans();
sweep_tile_span(p_spans[I0], [&](auto idx0) {
constexpr auto i_idx = make_tuple(idx0);
auto row_max = scale_s * get_validated_m(m[i_idx]);
sweep_tile_span(p_spans[I1], [&](auto idx1) {
constexpr auto i_j_idx = make_tuple(idx0, idx1);
if constexpr(BiasEnum == BlockAttentionBiasEnum::ELEMENTWISE_BIAS ||
BiasEnum == BlockAttentionBiasEnum::ALIBI)
{
p_compute(i_j_idx) = exp2(s_new[i_j_idx] - get_validated_m(m[i_idx]));
}
else
{
if constexpr(kHasLogitsSoftCap)
{
p_compute(i_j_idx) = exp2(s_new[i_j_idx] - get_validated_m(m[i_idx]));
}
else
{
p_compute(i_j_idx) = exp2(scale_s * s_new[i_j_idx] - row_max);
}
}
});
});
auto rowsum_p = block_tile_reduce<SMPLComputeDataType>(
p_compute, sequence<1>{}, f_sum, SMPLComputeDataType{0}); // rowsum(Pcompute{j})
block_tile_reduce_sync(rowsum_p, f_sum, bool_constant<false>{});
auto p_tile = make_static_distributed_tensor<PDataType>(
Policy::template MakePRegTileDistribution<Problem>());
p_tile.get_thread_buffer() = cast_tile<PDataType>(p_compute).get_thread_buffer();
// l{j}, Oacc{j}
constexpr auto o_spans = decltype(o_acc)::get_distributed_spans();
sweep_tile_span(o_spans[I0], [&](auto idx0) {
constexpr auto i_idx = make_tuple(idx0);
const auto tmp = [&]() {
if constexpr(BiasEnum == BlockAttentionBiasEnum::ELEMENTWISE_BIAS ||
BiasEnum == BlockAttentionBiasEnum::ALIBI)
{
return exp2(m_old[i_idx] - get_validated_m(m[i_idx]));
}
else
{
if constexpr(kHasLogitsSoftCap)
{
return exp2(m_old[i_idx] - get_validated_m(m[i_idx]));
}
else
{
auto row_max = scale_s * get_validated_m(m[i_idx]);
return exp2(scale_s * m_old[i_idx] - row_max);
}
}
}();
l(i_idx) = tmp * l[i_idx] + rowsum_p[i_idx];
sweep_tile_span(o_spans[I1], [&](auto idx1) {
constexpr auto i_j_idx = make_tuple(idx0, idx1);
o_acc(i_j_idx) *= tmp;
});
});
move_tile_window(v_dram_window, {kN0, 0});
async_load_tile(v_lds_write_window, v_dram_window);
block_sync_lds_direct_load<k_vmem_insts + v_vmem_insts>();
// Will insert unexpected vmcnt(0) here, probably the aliasing issue.
auto v_tile = load_tile_transpose(v_lds_read_window);
if constexpr(1 < k1_loops)
{
static_for<0, k1_loops - 1, 1>{}([&](auto i_k1) {
gemm_1(o_acc,
get_slice_tile(p_tile,
sequence<0, i_k1 * kK1>{},
sequence<kM0, (i_k1 + 1) * kK1>{}),
v_tile);
// loop over along the [V]alue Sequence length
move_tile_window(v_lds_read_window, {kK1, 0});
v_tile = load_tile_transpose(v_lds_read_window);
});
// move back to the origin
move_tile_window(v_lds_read_window, {-kK1 * (k1_loops - 1), 0});
}
gemm_1(o_acc,
get_slice_tile(p_tile,
sequence<0, (k1_loops - 1) * kK1>{},
sequence<kM0, k1_loops * kK1>{}),
v_tile);
};
do
{
mainloop(I1, I0);
i_total_loops++;
if(i_total_loops == (num_total_loop))
{
continue;
}
mainloop(I0, I1);
i_total_loops++;
} while(i_total_loops < num_total_loop);
if constexpr(kStoreLSE)
{
// store lse acc
auto lse_acc = make_static_distributed_tensor<LSEDataType>(m.get_tile_distribution());
constexpr auto lse_acc_spans = decltype(lse_acc)::get_distributed_spans();
sweep_tile_span(lse_acc_spans[I0], [&, m_ = m, l_ = l](auto idx0) {
constexpr auto i_idx = make_tuple(idx0);
if constexpr(BiasEnum == BlockAttentionBiasEnum::ELEMENTWISE_BIAS ||
BiasEnum == BlockAttentionBiasEnum::ALIBI)
{
lse_acc(i_idx) = m_[i_idx] / C_LOG2E + log(l_[i_idx]);
}
else
{
if constexpr(kHasLogitsSoftCap)
{
lse_acc(i_idx) = m_[i_idx] / C_LOG2E + log(l_[i_idx]);
}
else
{
lse_acc(i_idx) = m_[i_idx] * scale_s / C_LOG2E + log(l_[i_idx]);
}
}
});
if(get_thread_local_1d_id() < kM0)
{
store_tile(lse_acc_dram_window_tmp, lse_acc);
}
}
// finally, O
constexpr auto o_spans = decltype(o_acc)::get_distributed_spans();
sweep_tile_span(o_spans[I0], [&](auto idx0) {
constexpr auto i_idx = make_tuple(idx0);
const auto tmp = [&]() {
if constexpr(BiasEnum == BlockAttentionBiasEnum::ELEMENTWISE_BIAS ||
FmhaMask::IsMasking)
{
return l[i_idx] == 0.f ? 0.f : 1 / l[i_idx];
}
else
return 1 / l[i_idx];
}();
sweep_tile_span(o_spans[I1], [&](auto idx1) {
constexpr auto i_j_idx = make_tuple(idx0, idx1);
o_acc(i_j_idx) *= tmp;
});

97
include/ck_tile/ops/fmha/pipeline/block_fmha_fwd_decode_pipeline_qr_ks_vs_policy.hpp
View File

@@ -68,7 +68,7 @@ struct BlockFmhaFwdDecodePipelineQRKSVSDefaultPolicy
{
constexpr index_t kBlockSize = Problem::kBlockSize;
constexpr index_t kNPerBlock = Problem::BlockFmhaShape::kN1;
constexpr index_t kKPerBlock = Problem::BlockFmhaShape::kK1;
constexpr index_t kKPerBlock = Problem::BlockFmhaShape::kN0;
constexpr index_t MaxVectorSize = 16 / sizeof(typename Problem::VDataType);
@@ -77,43 +77,77 @@ struct BlockFmhaFwdDecodePipelineQRKSVSDefaultPolicy
return min(ElemPerThread, MaxVectorSize);
}
template <typename Problem>
template <typename Problem, bool BypassLDS = false>
CK_TILE_HOST_DEVICE static constexpr auto MakeQDramTileDistribution()
{
constexpr index_t kBlockSize = Problem::kBlockSize;
constexpr index_t kMPerBlock = Problem::BlockFmhaShape::kM0;
constexpr index_t kKPerBlock = Problem::BlockFmhaShape::kSubQKHeaddim;
if constexpr(!BypassLDS)
{
constexpr index_t kBlockSize = Problem::kBlockSize;
constexpr index_t kMPerBlock = Problem::BlockFmhaShape::kM0;
constexpr index_t kKPerBlock = Problem::BlockFmhaShape::kSubQKHeaddim;
constexpr index_t MaxVectorSize = 16 / sizeof(typename Problem::QDataType);
constexpr index_t MaxVectorSize = 16 / sizeof(typename Problem::QDataType);
constexpr index_t ElemPerThread = (kMPerBlock * kKPerBlock) / kBlockSize;
static_assert(0 < ElemPerThread);
constexpr index_t kMaxVecLoad = min(ElemPerThread, MaxVectorSize);
constexpr index_t ElemPerThread = (kMPerBlock * kKPerBlock) / kBlockSize;
static_assert(0 < ElemPerThread);
constexpr index_t kMaxVecLoad = min(ElemPerThread, MaxVectorSize);
constexpr index_t KPerThread = kMaxVecLoad;
constexpr index_t KThreads = kKPerBlock / KPerThread;
constexpr index_t MThreadPerWarp = get_warp_size() / KThreads;
constexpr index_t NumWarps = kBlockSize / get_warp_size();
constexpr index_t MPerThread = kMPerBlock / (MThreadPerWarp * NumWarps);
constexpr index_t KPerThread = kMaxVecLoad;
constexpr index_t KThreads = kKPerBlock / KPerThread;
constexpr index_t MThreadPerWarp = get_warp_size() / KThreads;
constexpr index_t NumWarps = kBlockSize / get_warp_size();
constexpr index_t MPerThread = kMPerBlock / (MThreadPerWarp * NumWarps);
return make_static_tile_distribution(
tile_distribution_encoding<sequence<1>,
tuple<sequence<MPerThread, NumWarps, MThreadPerWarp>,
sequence<KThreads, KPerThread>>,
tuple<sequence<1>, sequence<1, 2>>,
tuple<sequence<1>, sequence<2, 0>>,
sequence<1, 2>,
sequence<0, 1>>{});
return make_static_tile_distribution(
tile_distribution_encoding<sequence<1>,
tuple<sequence<MPerThread, NumWarps, MThreadPerWarp>,
sequence<KThreads, KPerThread>>,
tuple<sequence<1>, sequence<1, 2>>,
tuple<sequence<1>, sequence<2, 0>>,
sequence<1, 2>,
sequence<0, 1>>{});
}
else
{
using BlockGemm = remove_cvref_t<decltype(GetQKBlockGemm<Problem>())>;
constexpr auto config = BlockGemm::Policy::template GetWarpGemmMWarpNWarp<Problem>();
using WarpGemm = remove_cvref_t<decltype(config.template at<0>())>;
constexpr index_t MWarp = Problem::BlockFmhaShape::Gemm0BlockWarps::at(number<0>{});
constexpr index_t NWarp = Problem::BlockFmhaShape::Gemm0BlockWarps::at(number<1>{});
constexpr index_t kMPerBlock = Problem::BlockFmhaShape::kM0;
constexpr index_t kKPerBlock = Problem::BlockFmhaShape::kSubQKHeaddim;
constexpr index_t MIterPerWarp = kMPerBlock / (MWarp * WarpGemm::kM);
constexpr index_t KIterPerWarp = kKPerBlock / WarpGemm::kK;
constexpr auto q_block_outer_dstr_encoding = tile_distribution_encoding<
sequence<NWarp>,
tuple<sequence<MIterPerWarp, MWarp>, sequence<KIterPerWarp>>,
tuple<sequence<1, 0>>,
tuple<sequence<1, 0>>,
sequence<1, 2>,
sequence<0, 0>>{};
constexpr auto q_block_dstr_encode = detail::make_embed_tile_distribution_encoding(
q_block_outer_dstr_encoding, typename WarpGemm::AWarpDstrEncoding{});
constexpr auto q_block_dstr = make_static_tile_distribution(q_block_dstr_encode);
return q_block_dstr;
}
}
template <typename Problem>
template <typename Problem, bool LoadOnce = false>
CK_TILE_HOST_DEVICE static constexpr auto MakeKDramTileDistribution()
{
using KDataType = remove_cvref_t<typename Problem::KDataType>;
constexpr index_t kBlockSize = Problem::kBlockSize;
constexpr index_t kNPerBlock = Problem::BlockFmhaShape::kN0;
constexpr index_t kKPerBlock = Problem::BlockFmhaShape::kK0;
constexpr index_t kKPerBlock =
LoadOnce ? Problem::BlockFmhaShape::kSubQKHeaddim : Problem::BlockFmhaShape::kK0;
constexpr index_t MaxVectorSize = 16 / sizeof(KDataType);
constexpr index_t ElemPerThread = (kNPerBlock * kKPerBlock) / kBlockSize;
@@ -190,11 +224,12 @@ struct BlockFmhaFwdDecodePipelineQRKSVSDefaultPolicy
return q_lds_block_desc;
}
template <typename Problem>
template <typename Problem, bool LoadOnce = false>
CK_TILE_HOST_DEVICE static constexpr auto MakeKLdsBlockDescriptor()
{
constexpr index_t kNPerBlock = Problem::BlockFmhaShape::kN0;
constexpr index_t kKPerBlock = Problem::BlockFmhaShape::kK0;
constexpr index_t kKPerBlock =
LoadOnce ? Problem::BlockFmhaShape::kSubQKHeaddim : Problem::BlockFmhaShape::kK0;
constexpr index_t kKPack = GetSmemKPackK<Problem>();
@@ -211,7 +246,7 @@ struct BlockFmhaFwdDecodePipelineQRKSVSDefaultPolicy
CK_TILE_HOST_DEVICE static constexpr auto MakeVLdsBlockDescriptor()
{
constexpr index_t kNPerBlock = Problem::BlockFmhaShape::kN1;
constexpr index_t kKPerBlock = Problem::BlockFmhaShape::kK1;
constexpr index_t kKPerBlock = Problem::BlockFmhaShape::kN0;
constexpr index_t kKPack = GetSmemKPackV<Problem>();
@@ -335,7 +370,7 @@ struct BlockFmhaFwdDecodePipelineQRKSVSDefaultPolicy
{
constexpr index_t kBlockSize = Problem::kBlockSize;
constexpr index_t kNPerBlock = Problem::BlockFmhaShape::kN1;
constexpr index_t kKPerBlock = Problem::BlockFmhaShape::kK1;
constexpr index_t kKPerBlock = Problem::BlockFmhaShape::kN0;
constexpr index_t MaxVectorSize = 16 / sizeof(typename Problem::VDataType);
@@ -370,7 +405,7 @@ struct BlockFmhaFwdDecodePipelineQRKSVSDefaultPolicy
constexpr index_t NWarp = Problem::BlockFmhaShape::Gemm1BlockWarps::at(number<1>{});
constexpr index_t kMPerBlock = Problem::BlockFmhaShape::kM0;
constexpr index_t kKPerBlock = Problem::BlockFmhaShape::kK1;
constexpr index_t kKPerBlock = Problem::BlockFmhaShape::kN0;
constexpr index_t MIterPerWarp = kMPerBlock / (MWarp * WarpGemm::kM);
constexpr index_t KIterPerWarp = kKPerBlock / WarpGemm::kK;
@@ -502,10 +537,10 @@ struct BlockFmhaFwdDecodePipelineQRKSVSDefaultPolicy
sizeof(typename Problem::QDataType);
}
template <typename Problem>
template <typename Problem, bool LoadOnce = false>
CK_TILE_HOST_DEVICE static constexpr ck_tile::index_t GetSmemSizeK()
{
return MakeKLdsBlockDescriptor<Problem>().get_element_space_size() *
return MakeKLdsBlockDescriptor<Problem, LoadOnce>().get_element_space_size() *
sizeof(typename Problem::KDataType);
}