ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-03 21:21:22 +00:00
Code Issues Packages Projects Releases Wiki Activity

Add attention sink support for FMHA FWD (#3368)

Browse Source 
* Revert "Revert "Add attn sink (#2892)" (#3250)"

This reverts commit 5adaa201ed.

* fix conflict

Signed-off-by: Linjun-AMD <Jun.Lin@amd.com>

* Add F_sink parameter to FmhaFwdPipeline

* Update tile_fmha_traits.hpp

* Refactor pipeline creation in fmha_fwd.py

Updated the pipeline creation logic to include 'sink' parameter in product combinations and adjusted the FmhaFwdPipeline calls accordingly.

* Update fmha_fwd.py

* Update fmha_fwd.py

* Update example/ck_tile/01_fmha/script/correct_test_fwd_sink.sh

Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>

* update CHANGELOG.md

Signed-off-by: Linjun-AMD <Jun.Lin@amd.com>

* Update CHANGELOG with new features and support

* Update fmha_fwd.hpp

* Update CHANGELOG.md

* Update smoke_test_fwd_sink.sh

* Update correct_test_fwd_sink.sh

* Update smoke_test_fwd_sink.sh

---------

Signed-off-by: Linjun-AMD <Jun.Lin@amd.com>
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
This commit is contained in:
Linjun-AMD
2025-12-15 12:21:59 +08:00
committed by GitHub
parent 22b945e06e
commit f5573f56d9
26 changed files with 948 additions and 196 deletions
Download Patch File Download Diff File Expand all files Collapse all files

94
include/ck_tile/ops/fmha/pipeline/block_fmha_fwd_pagedkv_pipeline_qr_ks_vs.hpp
View File

@@ -57,6 +57,7 @@ struct BlockFmhaFwdPagedKVPipelineQRKSVS
static constexpr auto BiasEnum = Problem::BiasEnum;
static constexpr bool kStoreLSE = Problem::kStoreLSE;
static constexpr bool kIsPagedKV = Problem::kIsPagedKV;
static constexpr bool kHasSink = Problem::kHasSink;
static_assert((CK_TILE_FMHA_FWD_FAST_EXP2 &&
(kHasLogitsSoftCap && Problem::BiasEnum == BlockAttentionBiasEnum::NO_BIAS ||
@@ -228,10 +229,22 @@ struct BlockFmhaFwdPagedKVPipelineQRKSVS
clear_tile(o_acc);
set_tile(m, -numeric<SMPLComputeDataType>::infinity());
clear_tile(l);
const auto q_origin = q_dram_window.get_window_origin();
const auto [logical_seqlen_k_start, logical_seqlen_k_end] =
mask.GetTileRangeAlongX(q_origin.at(number<0>{}), number<kM0>{}, number<kN0>{});
const auto q_origin = q_dram_window.get_window_origin();
const auto tile_range_result = [&mask, &q_origin]() {
if constexpr(kHasSink)
return mask.GetSinkTileRangeAlongX(
q_origin.at(number<0>{}), number<kM0>{}, number<kN0>{});
else
{
auto [start, end] =
mask.GetTileRangeAlongX(q_origin.at(number<0>{}), number<kM0>{}, number<kN0>{});
return ck_tile::make_tuple(0, start, end);
}
}();
const auto sink_seq_end = tile_range_result.get(ck_tile::number<0>{});
const auto logical_seqlen_k_start = tile_range_result.get(ck_tile::number<1>{});
const auto logical_seqlen_k_end = tile_range_result.get(ck_tile::number<2>{});
const auto num_sink_loop = integer_divide_ceil(sink_seq_end, kN0);
// check early exit if no work to do
if constexpr(FmhaMask::IsMasking || kPadSeqLenK)
@@ -255,7 +268,6 @@ struct BlockFmhaFwdPagedKVPipelineQRKSVS
return o_acc;
}
}
// k_dram_block_window
const index_t physical_seqlen_k_start = logical_seqlen_k_start + kv_l2p_offset;
const index_t physical_seqlen_k_end = logical_seqlen_k_end + kv_l2p_offset;
@@ -274,27 +286,36 @@ struct BlockFmhaFwdPagedKVPipelineQRKSVS
return physical_seqlen_k_start_;
}
}();
const auto kv_load_start = (sink_seq_end == 0 && aligned_physical_seqlen_k_start > 0)
? aligned_physical_seqlen_k_start
: 0;
const index_t num_total_loop =
integer_divide_ceil(physical_seqlen_k_end - aligned_physical_seqlen_k_start, kN0);
integer_divide_ceil(physical_seqlen_k_end - aligned_physical_seqlen_k_start, kN0) +
num_sink_loop;
auto [i_page_block_k, k_dram_block_window] = k_page_block_navigator.make_tile_window(
k_dram_block_window_lengths, {aligned_physical_seqlen_k_start, 0});
k_dram_block_window_lengths, {kv_load_start, 0});
const auto bias_origin = bias_dram_block_window_tmp.get_window_origin();
const index_t bias_n_offset = [&]() {
if constexpr(kHasSink)
return kv_load_start;
else
return logical_seqlen_k_start -
(physical_seqlen_k_start - aligned_physical_seqlen_k_start);
}();
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
{bias_origin.at(number<0>{}), bias_n_offset},
Policy::template MakeBiasDramTileDistribution<decltype(gemm_0)>());
// v_dram_window
auto [i_page_block_v, v_dram_window] = v_page_block_navigator.make_tile_window(
v_dram_block_window_lengths,
{0, aligned_physical_seqlen_k_start}, // TODO: hdim split?
{0, kv_load_start}, // TODO: hdim split?
Policy::template MakeVDramTileDistribution<Problem>());
auto q_tile = tile_elementwise_in(q_element_func, q);
// prefetch K tile
@@ -321,9 +342,16 @@ struct BlockFmhaFwdPagedKVPipelineQRKSVS
store_tile(k_lds_window, tile_elementwise_in(k_element_func, k_block_tile));
k_block_tile = load_tile(k_dram_window);
}
const bool is_sink_tile = ((num_sink_loop - 1) == i_total_loops);
const auto k_move_offset = [&]() {
if constexpr(kHasSink)
return is_sink_tile ? logical_seqlen_k_start - sink_seq_end + kN0 : kN0;
else
return kN0;
}();
auto physical_next_block_id_k =
amd_wave_read_first_lane(k_page_block_navigator.prefetch_table_id(
i_page_block_k, k_dram_block_window, {kN0, 0}));
i_page_block_k, k_dram_block_window, {k_move_offset, 0}));
auto physical_next_block_id_v = amd_wave_read_first_lane(
v_page_block_navigator.prefetch_table_id(i_page_block_v, v_dram_window, {0, kK1}));
@@ -442,7 +470,7 @@ struct BlockFmhaFwdPagedKVPipelineQRKSVS
#endif
}
}
move_tile_window(bias_dram_window, {0, kN0});
move_tile_window(bias_dram_window, {0, k_move_offset});
{
const auto k_origin = k_page_block_navigator.to_global_window_origin(
@@ -474,14 +502,29 @@ struct BlockFmhaFwdPagedKVPipelineQRKSVS
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>{});
return mask.IsOutOfBound(row, col - kv_l2p_offset);
auto apply_mask = [&](auto&& mask_func) {
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>{});
return mask_func(row, col - kv_l2p_offset);
});
};
if constexpr(kHasSink)
{
apply_mask([&](auto row, auto col) {
return mask.IsOutOfSinkBound(row, col);
});
}
else
{
apply_mask(
[&](auto row, auto col) { return mask.IsOutOfBound(row, col); });
}
}
}
}
@@ -647,7 +690,12 @@ struct BlockFmhaFwdPagedKVPipelineQRKSVS
}
// move K tile windows
i_page_block_k = k_page_block_navigator.move_tile_window(
i_page_block_k, k_dram_block_window, {kN0, 0}, physical_next_block_id_k);
i_page_block_k, k_dram_block_window, {k_move_offset, 0}, physical_next_block_id_k);
physical_next_block_id_v =
amd_wave_read_first_lane(v_page_block_navigator.prefetch_table_id(
i_page_block_v, v_dram_window, {0, k_move_offset - kN0}));
i_page_block_v = v_page_block_navigator.move_tile_window(
i_page_block_v, v_dram_window, {0, k_move_offset - kN0}, physical_next_block_id_v);
// tail
{
block_sync_lds();

86
include/ck_tile/ops/fmha/pipeline/block_fmha_fwd_splitkv_pipeline_nwarp_sshuffle_qr_ks_vs.hpp
View File

@@ -57,6 +57,7 @@ struct BlockFmhaFwdSplitKVPipelineNWarpSShuffleQRKSVS
static constexpr bool kStoreLSE = Problem::kStoreLSE;
static constexpr bool kIsPagedKV = Problem::kIsPagedKV;
static constexpr bool kHasUnevenSplits = Problem::kHasUnevenSplits;
static constexpr bool kHasSink = Problem::kHasSink;
static_assert((CK_TILE_FMHA_FWD_FAST_EXP2 &&
(kHasLogitsSoftCap && Problem::BiasEnum == BlockAttentionBiasEnum::NO_BIAS ||
@@ -256,11 +257,23 @@ struct BlockFmhaFwdSplitKVPipelineNWarpSShuffleQRKSVS
set_tile(m, -numeric<SMPLComputeDataType>::infinity());
clear_tile(l);
const auto q_origin = q_dram_window.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);
const auto q_origin = q_dram_window.get_window_origin();
const auto tile_range_result = [&mask, &q_origin, num_splits, i_split]() {
if constexpr(kHasSink)
return mask.GetSinkTileRangeAlongX(
q_origin.at(number<0>{}), number<kM0>{}, number<kN0>{}, num_splits, i_split);
else
{
auto [start, end] = mask.GetTileRangeAlongX(
q_origin.at(number<0>{}), number<kM0>{}, number<kN0>{}, num_splits, i_split);
return ck_tile::make_tuple(0, start, end);
}
}();
const auto sink_seq_end = tile_range_result.get(ck_tile::number<0>{});
const auto logical_seqlen_k_start = tile_range_result.get(ck_tile::number<1>{});
const auto logical_seqlen_k_end = tile_range_result.get(ck_tile::number<2>{});
// check early exit if no work to do
const auto num_sink_loop = integer_divide_ceil(sink_seq_end, kN0);
if constexpr(FmhaMask::IsMasking || kPadSeqLenK || kHasUnevenSplits)
{
const index_t logical_num_total_loop =
@@ -304,24 +317,33 @@ struct BlockFmhaFwdSplitKVPipelineNWarpSShuffleQRKSVS
return physical_seqlen_k_start_;
}
}();
const auto kv_load_start = (sink_seq_end == 0 && aligned_physical_seqlen_k_start > 0)
? aligned_physical_seqlen_k_start
: 0;
const index_t num_total_loop =
integer_divide_ceil(physical_seqlen_k_end - aligned_physical_seqlen_k_start, kN0);
integer_divide_ceil(physical_seqlen_k_end - aligned_physical_seqlen_k_start, kN0) +
num_sink_loop;
auto [i_page_block_k, k_dram_block_window] = k_page_block_navigator.make_tile_window(
k_dram_block_window_lengths, {aligned_physical_seqlen_k_start, 0});
k_dram_block_window_lengths, {kv_load_start, 0});
const auto bias_origin = bias_dram_block_window_tmp.get_window_origin();
const auto bias_origin = bias_dram_block_window_tmp.get_window_origin();
const index_t bias_n_offset = [&]() {
if constexpr(kHasSink)
return kv_load_start;
else
return logical_seqlen_k_start -
(physical_seqlen_k_start - aligned_physical_seqlen_k_start);
}();
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
{bias_origin.at(number<0>{}), bias_n_offset},
Policy::template MakeBiasDramTileDistribution<decltype(gemm_0)>());
auto [i_page_block_v, v_dram_window] = v_page_block_navigator.make_tile_window(
v_dram_block_window_lengths,
{0, aligned_physical_seqlen_k_start}, // TODO: hdim split?
{0, kv_load_start}, // TODO: hdim split?
Policy::template MakeVDramTileDistribution<Problem>());
// store Q into LDS
@@ -369,7 +391,13 @@ struct BlockFmhaFwdSplitKVPipelineNWarpSShuffleQRKSVS
{
// STAGE 1, QK gemm
clear_tile(s_acc); // initialize C
const bool is_sink_tile = ((num_sink_loop - 1) == i_total_loops);
const auto k_move_offset = [&]() {
if constexpr(kHasSink)
return is_sink_tile ? logical_seqlen_k_start - sink_seq_end + kN0 : kN0;
else
return kN0;
}();
// load the second tile of the first iteration
k_block_tile = load_tile(k_dram_window);
@@ -494,7 +522,7 @@ struct BlockFmhaFwdSplitKVPipelineNWarpSShuffleQRKSVS
#endif
}
}
move_tile_window(bias_dram_window, {0, kN0});
move_tile_window(bias_dram_window, {0, k_move_offset});
/// TODO: only check in first/last iteration without increasing code size
if constexpr(kHasUnevenSplits)
@@ -505,7 +533,7 @@ struct BlockFmhaFwdSplitKVPipelineNWarpSShuffleQRKSVS
s_acc,
-numeric<SMPLComputeDataType>::infinity(),
[&,
physical_seqlen_k_start_ = physical_seqlen_k_start,
physical_seqlen_k_start_ = is_sink_tile ? 0 : 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>{});
if constexpr(kIsPagedKV)
@@ -530,12 +558,26 @@ struct BlockFmhaFwdSplitKVPipelineNWarpSShuffleQRKSVS
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>{});
return mask.IsOutOfBound(row, col - kv_l2p_offset);
});
auto apply_mask = [&](auto&& mask_func) {
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>{});
return mask_func(row, col - kv_l2p_offset);
});
};
if constexpr(kHasSink)
{
apply_mask(
[&](auto row, auto col) { return mask.IsOutOfSinkBound(row, col); });
}
else
{
apply_mask([&](auto row, auto col) { return mask.IsOutOfBound(row, col); });
}
}
}
@@ -546,7 +588,7 @@ struct BlockFmhaFwdSplitKVPipelineNWarpSShuffleQRKSVS
{
// move K tile windows
i_page_block_k = k_page_block_navigator.move_tile_window(
i_page_block_k, k_dram_block_window, {kN0, 0});
i_page_block_k, k_dram_block_window, {k_move_offset, 0});
k_dram_window = make_tile_window(
k_dram_block_window,
@@ -742,6 +784,8 @@ struct BlockFmhaFwdSplitKVPipelineNWarpSShuffleQRKSVS
// moving k_dram_window is an in-page-block operation, so there is
// no need to invoke k_page_block_navigator.move_tile_window() here.
move_tile_window(k_dram_window, {0, kK0});
i_page_block_v = v_page_block_navigator.move_tile_window(
i_page_block_v, v_dram_window, {0, k_move_offset - kN0});
store_tile(k_lds_window, tile_elementwise_in(k_element_func, k_block_tile));
}
} while(++i_total_loops < num_total_loop);

92
include/ck_tile/ops/fmha/pipeline/block_fmha_fwd_splitkv_pipeline_qr_ks_vs.hpp
View File

@@ -56,6 +56,7 @@ struct BlockFmhaFwdSplitKVPipelineQRKSVS
static constexpr bool kStoreLSE = Problem::kStoreLSE;
static constexpr bool kIsPagedKV = Problem::kIsPagedKV;
static constexpr bool kHasUnevenSplits = Problem::kHasUnevenSplits;
static constexpr bool kHasSink = Problem::kHasSink;
static_assert((CK_TILE_FMHA_FWD_FAST_EXP2 &&
(kHasLogitsSoftCap && Problem::BiasEnum == BlockAttentionBiasEnum::NO_BIAS ||
@@ -229,9 +230,23 @@ struct BlockFmhaFwdSplitKVPipelineQRKSVS
set_tile(m, -numeric<SMPLComputeDataType>::infinity());
clear_tile(l);
const auto q_origin = q_dram_window.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);
const auto q_origin = q_dram_window.get_window_origin();
const auto tile_range_result = [&mask, &q_origin, num_splits, i_split]() {
if constexpr(kHasSink)
return mask.GetSinkTileRangeAlongX(
q_origin.at(number<0>{}), number<kM0>{}, number<kN0>{}, num_splits, i_split);
else
{
auto [start, end] = mask.GetTileRangeAlongX(
q_origin.at(number<0>{}), number<kM0>{}, number<kN0>{}, num_splits, i_split);
return ck_tile::make_tuple(0, start, end);
}
}();
const auto sink_seq_end = tile_range_result.get(ck_tile::number<0>{});
const auto logical_seqlen_k_start = tile_range_result.get(ck_tile::number<1>{});
const auto logical_seqlen_k_end = tile_range_result.get(ck_tile::number<2>{});
const auto num_sink_loop = integer_divide_ceil(sink_seq_end, kN0);
// check early exit if no work to do
if constexpr(FmhaMask::IsMasking || kPadSeqLenK || kHasUnevenSplits)
@@ -274,24 +289,35 @@ struct BlockFmhaFwdSplitKVPipelineQRKSVS
return physical_seqlen_k_start_;
}
}();
const auto kv_load_start = (sink_seq_end == 0 && aligned_physical_seqlen_k_start > 0)
? aligned_physical_seqlen_k_start
: 0;
const index_t num_total_loop =
integer_divide_ceil(physical_seqlen_k_end - aligned_physical_seqlen_k_start, kN0);
integer_divide_ceil(physical_seqlen_k_end - aligned_physical_seqlen_k_start, kN0) +
num_sink_loop;
auto [i_page_block_k, k_dram_block_window] = k_page_block_navigator.make_tile_window(
k_dram_block_window_lengths, {aligned_physical_seqlen_k_start, 0});
k_dram_block_window_lengths, {kv_load_start, 0});
const auto bias_origin = bias_dram_block_window_tmp.get_window_origin();
const index_t bias_n_offset = [&]() {
if constexpr(kHasSink)
return kv_load_start;
else
return logical_seqlen_k_start -
(physical_seqlen_k_start - aligned_physical_seqlen_k_start);
}();
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
{bias_origin.at(number<0>{}), bias_n_offset},
Policy::template MakeBiasDramTileDistribution<decltype(gemm_0)>());
auto [i_page_block_v, v_dram_window] = v_page_block_navigator.make_tile_window(
v_dram_block_window_lengths,
{0, aligned_physical_seqlen_k_start}, // TODO: hdim split?
{0, kv_load_start}, // TODO: hdim split?
Policy::template MakeVDramTileDistribution<Problem>());
auto q_tile = tile_elementwise_in(q_element_func, q);
@@ -320,9 +346,18 @@ struct BlockFmhaFwdSplitKVPipelineQRKSVS
store_tile(k_lds_window, tile_elementwise_in(k_element_func, k_block_tile));
k_block_tile = load_tile(k_dram_window);
}
const bool is_sink_tile = ((num_sink_loop - 1) == i_total_loops);
const auto k_move_offset = [&]() {
if constexpr(kHasSink)
return is_sink_tile ? logical_seqlen_k_start - sink_seq_end + kN0 : kN0;
else
return kN0;
}();
auto physical_next_block_id_k =
amd_wave_read_first_lane(k_page_block_navigator.prefetch_table_id(
i_page_block_k, k_dram_block_window, {kN0, 0}));
i_page_block_k, k_dram_block_window, {k_move_offset, 0}));
auto physical_next_block_id_v = amd_wave_read_first_lane(
v_page_block_navigator.prefetch_table_id(i_page_block_v, v_dram_window, {0, kK1}));
@@ -441,7 +476,7 @@ struct BlockFmhaFwdSplitKVPipelineQRKSVS
#endif
}
}
move_tile_window(bias_dram_window, {0, kN0});
move_tile_window(bias_dram_window, {0, k_move_offset});
/// TODO: only check in first/last iteration without increasing code size
if constexpr(kHasUnevenSplits)
@@ -452,7 +487,7 @@ struct BlockFmhaFwdSplitKVPipelineQRKSVS
s_acc,
-numeric<SMPLComputeDataType>::infinity(),
[&,
physical_seqlen_k_start_ = physical_seqlen_k_start,
physical_seqlen_k_start_ = is_sink_tile ? 0 : 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>{});
if constexpr(kIsPagedKV)
@@ -477,12 +512,26 @@ struct BlockFmhaFwdSplitKVPipelineQRKSVS
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>{});
return mask.IsOutOfBound(row, col - kv_l2p_offset);
});
auto apply_mask = [&](auto&& mask_func) {
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>{});
return mask_func(row, col - kv_l2p_offset);
});
};
if constexpr(kHasSink)
{
apply_mask(
[&](auto row, auto col) { return mask.IsOutOfSinkBound(row, col); });
}
else
{
apply_mask([&](auto row, auto col) { return mask.IsOutOfBound(row, col); });
}
}
}
@@ -647,7 +696,12 @@ struct BlockFmhaFwdSplitKVPipelineQRKSVS
}
// move K tile windows
i_page_block_k = k_page_block_navigator.move_tile_window(
i_page_block_k, k_dram_block_window, {kN0, 0}, physical_next_block_id_k);
i_page_block_k, k_dram_block_window, {k_move_offset, 0}, physical_next_block_id_k);
physical_next_block_id_v =
amd_wave_read_first_lane(v_page_block_navigator.prefetch_table_id(
i_page_block_v, v_dram_window, {0, k_move_offset - kN0}));
i_page_block_v = v_page_block_navigator.move_tile_window(
i_page_block_v, v_dram_window, {0, k_move_offset - kN0}, physical_next_block_id_v);
// tail
{
block_sync_lds();

3
include/ck_tile/ops/fmha/pipeline/block_fmha_pipeline_problem.hpp
View File

@@ -62,6 +62,7 @@ struct BlockFmhaPipelineProblem
static constexpr bool kHasDropout = Traits::kHasDropout;
static constexpr auto QScaleEnum = Traits::QScaleEnum;
static constexpr index_t kBlockPerCu = Traits::kBlockPerCu;
static constexpr bool kHasSink = Traits::kHasSink;
};
template <typename QDataType_,
@@ -114,6 +115,7 @@ struct BlockFmhaFwdPagedKVPipelineProblem
static constexpr bool kDoFp8StaticQuant = Traits::kDoFp8StaticQuant;
static constexpr bool kIsPagedKV = Traits::kIsPagedKV;
static constexpr index_t kBlockPerCu = Traits::kBlockPerCu;
static constexpr bool kHasSink = Traits::kHasSink;
};
template <typename QDataType_,
@@ -167,6 +169,7 @@ struct BlockFmhaFwdSplitKVPipelineProblem
static constexpr bool kHasUnevenSplits = kIsGroupMode || Traits::kHasUnevenSplits;
static constexpr bool kMergeNumHeadGroupsSeqLenQ = Traits::kMergeNumHeadGroupsSeqLenQ;
static constexpr index_t kBlockPerCu = Traits::kBlockPerCu;
static constexpr bool kHasSink = Traits::kHasSink;
};
// extract tile size attributes to remove dependency on traits

99
include/ck_tile/ops/fmha/pipeline/block_fmha_pipeline_qr_ks_vs.hpp
View File

@@ -57,6 +57,7 @@ struct BlockFmhaPipelineQRKSVS
static constexpr auto BiasEnum = Problem::BiasEnum;
static constexpr bool kStoreLSE = Problem::kStoreLSE;
static constexpr bool kHasDropout = Problem::kHasDropout;
static constexpr bool kHasSink = Problem::kHasSink;
static constexpr uint32_t DS_READ = 0x100; // Barrier for DS (data share) read
static constexpr uint32_t MFMA = 0x008; // Barrier for MFMA (matrix multiply-accumulate)
@@ -233,10 +234,26 @@ struct BlockFmhaPipelineQRKSVS
clear_tile(l);
const auto q_origin = q_dram_window.get_window_origin();
const auto [seqlen_k_start, seqlen_k_end] =
mask.GetTileRangeAlongX(q_origin.at(number<0>{}), number<kM0>{}, number<kN0>{});
const auto num_total_loop = integer_divide_ceil(seqlen_k_end - seqlen_k_start, kN0);
const auto tile_range_result = [&mask, &q_origin]() {
if constexpr(kHasSink)
return mask.GetSinkTileRangeAlongX(
q_origin.at(number<0>{}), number<kM0>{}, number<kN0>{});
else
{
auto [start, end] =
mask.GetTileRangeAlongX(q_origin.at(number<0>{}), number<kM0>{}, number<kN0>{});
return ck_tile::make_tuple(0, start, end);
}
}();
const auto sink_seq_end = tile_range_result.get(ck_tile::number<0>{});
const auto seqlen_k_start = tile_range_result.get(ck_tile::number<1>{});
const auto seqlen_k_end = tile_range_result.get(ck_tile::number<2>{});
const auto kv_load_start = (sink_seq_end == 0 && seqlen_k_start > 0) ? seqlen_k_start : 0;
const auto num_sink_loop = integer_divide_ceil(sink_seq_end, kN0);
const auto num_total_loop =
integer_divide_ceil(seqlen_k_end - seqlen_k_start, kN0) + num_sink_loop;
// check early exit if no work to do
if constexpr(FmhaMask::IsMasking || kPadSeqLenK)
@@ -262,22 +279,22 @@ struct BlockFmhaPipelineQRKSVS
auto k_dram_block_window =
make_tile_window(k_dram_block_window_tmp.get_bottom_tensor_view(),
k_dram_block_window_tmp.get_window_lengths(),
{seqlen_k_start, 0});
{kv_load_start, 0});
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>{}), seqlen_k_start}, // M/N
{bias_origin.at(number<0>{}), kv_load_start}, // M/N
Policy::template MakeBiasDramTileDistribution<decltype(gemm_0)>());
auto randval_dram_window = dropout.template MakeRandvalDramWindow<decltype(gemm_0)>(
randval_dram_block_window_tmp, seqlen_k_start);
randval_dram_block_window_tmp, kv_load_start);
auto v_dram_window =
make_tile_window(v_dram_block_window_tmp.get_bottom_tensor_view(),
v_dram_block_window_tmp.get_window_lengths(),
{0, seqlen_k_start}, // TODO: hdim split?
{0, kv_load_start}, // TODO: hdim split?
Policy::template MakeVDramTileDistribution<Problem>());
auto q_tile = tile_elementwise_in(q_element_func, q);
@@ -450,6 +467,11 @@ struct BlockFmhaPipelineQRKSVS
#endif
}
}
if constexpr(kHasSink)
{
if(i_total_loops == 0)
move_tile_window(bias_dram_window, {0, seqlen_k_start - sink_seq_end});
}
move_tile_window(bias_dram_window, {0, kN0});
if constexpr(kPadSeqLenK || FmhaMask::IsMasking)
{
@@ -460,17 +482,34 @@ struct BlockFmhaPipelineQRKSVS
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>{});
return !variant.LogitsMask(variant_params,
block_indices.batch_idx,
row,
col,
block_indices.qo_head_idx,
block_indices.kv_head_idx);
auto apply_mask = [&](auto&& mask_func) {
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>{});
return !mask_func(variant_params,
block_indices.batch_idx,
row,
col,
block_indices.qo_head_idx,
block_indices.kv_head_idx);
});
};
if constexpr(kHasSink)
{
apply_mask([&](auto&&... args) {
return variant.LogitsSinkMask(std::forward<decltype(args)>(args)...);
});
}
else
{
apply_mask([&](auto&&... args) {
return variant.LogitsMask(std::forward<decltype(args)>(args)...);
});
}
}
}
@@ -580,11 +619,23 @@ struct BlockFmhaPipelineQRKSVS
if constexpr(kHasDropout)
{
// K and dropout use the same address in LDS, finish loading from k_lds_window by
// gemm_0 to reuse LDS.
block_sync_lds();
auto randval_ptr = reinterpret_cast<char*>(smem_ptr);
index_t seq_offset = [&]() {
if constexpr(!kHasSink)
return seqlen_k_start + i_total_loops * kN0;
const bool in_sink_phase = (num_sink_loop > i_total_loops);
if(i_total_loops == num_sink_loop)
move_tile_window(randval_dram_window, {0, seqlen_k_start - sink_seq_end});
return in_sink_phase ? (kv_load_start + i_total_loops * kN0)
: (seqlen_k_start + (i_total_loops - num_sink_loop) * kN0);
}();
dropout.template Run<decltype(gemm_0), SMPLComputeDataType, RandValOutputDataType>(
smem_ptr, seqlen_k_start + i_total_loops * kN0, p_compute, randval_dram_window);
randval_ptr, seq_offset, p_compute, randval_dram_window);
}
block_sync_lds();
@@ -636,6 +687,14 @@ struct BlockFmhaPipelineQRKSVS
});
}
// move K tile windows
if constexpr(kHasSink)
{
if(i_total_loops == 0)
{
move_tile_window(k_dram_block_window, {seqlen_k_start - sink_seq_end, 0});
move_tile_window(v_dram_window, {0, seqlen_k_start - sink_seq_end});
}
}
move_tile_window(k_dram_block_window, {kN0, 0});
// tail
{

102
include/ck_tile/ops/fmha/pipeline/block_fmha_pipeline_qr_ks_vs_async.hpp
View File

@@ -62,6 +62,7 @@ struct BlockFmhaPipelineQRKSVSAsync
static constexpr auto BiasEnum = Problem::BiasEnum;
static constexpr bool kStoreLSE = Problem::kStoreLSE;
static constexpr bool kHasDropout = Problem::kHasDropout;
static constexpr bool kHasSink = Problem::kHasSink;
static_assert((CK_TILE_FMHA_FWD_FAST_EXP2 &&
(kHasLogitsSoftCap && Problem::BiasEnum == BlockAttentionBiasEnum::NO_BIAS ||
@@ -277,11 +278,26 @@ struct BlockFmhaPipelineQRKSVSAsync
clear_tile(l);
__builtin_amdgcn_sched_barrier(0);
const auto q_origin = q_dram_window.get_window_origin();
const auto [seqlen_k_start, seqlen_k_end] =
mask.GetTileRangeAlongX(q_origin.at(number<0>{}), number<kM0>{}, number<kN0>{});
const auto q_origin = q_dram_window.get_window_origin();
const auto tile_range_result = [&mask, &q_origin]() {
if constexpr(kHasSink)
return mask.GetSinkTileRangeAlongX(
q_origin.at(number<0>{}), number<kM0>{}, number<kN0>{});
else
{
auto [start, end] =
mask.GetTileRangeAlongX(q_origin.at(number<0>{}), number<kM0>{}, number<kN0>{});
return ck_tile::make_tuple(0, start, end);
}
}();
const auto sink_seq_end = tile_range_result.get(ck_tile::number<0>{});
const auto seqlen_k_start = tile_range_result.get(ck_tile::number<1>{});
const auto seqlen_k_end = tile_range_result.get(ck_tile::number<2>{});
const auto num_total_loop = integer_divide_ceil(seqlen_k_end - seqlen_k_start, kN0);
const auto kv_load_start = (sink_seq_end == 0 && seqlen_k_start > 0) ? seqlen_k_start : 0;
const auto num_sink_loop = integer_divide_ceil(sink_seq_end, kN0);
const auto num_total_loop =
integer_divide_ceil(seqlen_k_end - seqlen_k_start, kN0) + num_sink_loop;
// check early exit if no work to do
if constexpr(FmhaMask::IsMasking || kPadSeqLenK)
@@ -309,7 +325,7 @@ struct BlockFmhaPipelineQRKSVSAsync
auto k_dram_block_window =
make_tile_window(k_dram_block_window_tmp.get_bottom_tensor_view(),
k_dram_block_window_tmp.get_window_lengths(),
{seqlen_k_start, 0});
{kv_load_start, 0});
auto k_dram_window = make_tile_window(
k_dram_block_window.get_bottom_tensor_view(),
@@ -332,16 +348,16 @@ struct BlockFmhaPipelineQRKSVSAsync
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>{}), seqlen_k_start}, // M/N
{bias_origin.at(number<0>{}), kv_load_start}, // M/N
Policy::template MakeBiasDramTileDistribution<decltype(gemm_0)>());
auto randval_dram_window = dropout.template MakeRandvalDramWindow<decltype(gemm_0)>(
randval_dram_block_window_tmp, seqlen_k_start);
randval_dram_block_window_tmp, kv_load_start);
auto v_dram_window =
make_tile_window(v_dram_block_window_tmp.get_bottom_tensor_view(),
v_dram_block_window_tmp.get_window_lengths(),
{0, seqlen_k_start}, // TODO: hdim split?
{0, kv_load_start}, // TODO: hdim split?
Policy::template MakeVDramTileDistribution<Problem>());
// prefetch K tile
@@ -478,6 +494,11 @@ struct BlockFmhaPipelineQRKSVSAsync
#endif
}
}
if constexpr(kHasSink)
{
if(i_total_loops == 0)
move_tile_window(bias_dram_window, {0, seqlen_k_start - sink_seq_end});
}
move_tile_window(bias_dram_window, {0, kN0});
if constexpr(kPadSeqLenK || FmhaMask::IsMasking)
{
@@ -489,17 +510,34 @@ struct BlockFmhaPipelineQRKSVSAsync
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>{});
return !variant.LogitsMask(variant_params,
block_indices.batch_idx,
row,
col,
block_indices.qo_head_idx,
block_indices.kv_head_idx);
auto apply_mask = [&](auto&& mask_func) {
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>{});
return !mask_func(variant_params,
block_indices.batch_idx,
row,
col,
block_indices.qo_head_idx,
block_indices.kv_head_idx);
});
};
if constexpr(kHasSink)
{
apply_mask([&](auto&&... args) {
return variant.LogitsSinkMask(std::forward<decltype(args)>(args)...);
});
}
else
{
apply_mask([&](auto&&... args) {
return variant.LogitsMask(std::forward<decltype(args)>(args)...);
});
}
}
}
@@ -647,11 +685,21 @@ struct BlockFmhaPipelineQRKSVSAsync
{
auto randval_ptr =
reinterpret_cast<char*>(smem_ptr) + Policy::template GetSmemSizeKV<Problem>();
index_t seq_offset = [&]() {
if constexpr(!kHasSink)
return seqlen_k_start + i_total_loops * kN0;
const bool in_sink_phase = (num_sink_loop > i_total_loops);
if(i_total_loops == num_sink_loop)
move_tile_window(randval_dram_window, {0, seqlen_k_start - sink_seq_end});
return in_sink_phase ? (kv_load_start + i_total_loops * kN0)
: (seqlen_k_start + (i_total_loops - num_sink_loop) * kN0);
}();
dropout.template Run<decltype(gemm_0), SMPLComputeDataType, RandValOutputDataType>(
randval_ptr,
seqlen_k_start + i_total_loops * kN0,
p_compute,
randval_dram_window);
randval_ptr, seq_offset, p_compute, randval_dram_window);
}
const auto p = [&]() {
@@ -717,8 +765,16 @@ struct BlockFmhaPipelineQRKSVSAsync
i_total_loops++;
if(i_total_loops < num_total_loop)
{
// move K tile windows
if constexpr(kHasSink)
{
if(i_total_loops == 0)
{
move_tile_window(k_dram_block_window, {seqlen_k_start - sink_seq_end, 0});
move_tile_window(v_dram_window, {0, seqlen_k_start - sink_seq_end});
}
}
move_tile_window(k_dram_block_window, {kN0, 0});
k_dram_window.set_window_origin(k_dram_block_window.get_window_origin());
if constexpr(k1_loops >= 2 &&

1
include/ck_tile/ops/fmha/pipeline/block_fmha_pipeline_qr_ks_vs_async_trload.hpp
View File

@@ -69,6 +69,7 @@ struct BlockFmhaPipelineQRKSVSAsyncTrload
static constexpr auto BiasEnum = Problem::BiasEnum;
static constexpr bool kStoreLSE = Problem::kStoreLSE;
static constexpr bool kHasUnevenSplits = true;
static constexpr bool kHasSink = Problem::kHasSink;
static_assert((CK_TILE_FMHA_FWD_FAST_EXP2 &&
(kHasLogitsSoftCap && Problem::BiasEnum == BlockAttentionBiasEnum::NO_BIAS ||

16
include/ck_tile/ops/fmha/pipeline/tile_fmha_traits.hpp
View File

@@ -20,8 +20,9 @@ template <bool kPadSeqLenQ_ /* padding for seqlen_q */,
bool kStoreLSE_,
bool kHasDropout_,
BlockAttentionQuantScaleEnum QScaleEnum_,
index_t kBlockPerCu_ = -1, /* overwrite occupancy if not -1 */
bool kSkipMinSeqlenQ_ = false /* skip min seqlen q while chunked prefill */>
index_t kBlockPerCu_ = -1, /* overwrite occupancy if not -1 */
bool kSkipMinSeqlenQ_ = false, /* skip min seqlen q while chunked prefill */
bool kHasSink_ = false>
struct TileFmhaTraits
{
static constexpr bool kPadSeqLenQ = kPadSeqLenQ_;
@@ -36,6 +37,7 @@ struct TileFmhaTraits
static constexpr auto QScaleEnum = QScaleEnum_;
static constexpr index_t kBlockPerCu = kBlockPerCu_;
static constexpr bool kSkipMinSeqlenQ = kSkipMinSeqlenQ_;
static constexpr bool kHasSink = kHasSink_;
};
template <index_t kPadHeadDimQ_ /* paddding for hdim_q */,
@@ -65,8 +67,9 @@ template <bool kPadSeqLenQ_ /* padding for seqlen_q */,
bool kStoreLSE_, /* set to true if either num_splits > 1 or fwd training is running */
bool kIsPagedKV_,
bool kDoFp8StaticQuant_,
index_t kBlockPerCu_ = -1, /* overwrite occupancy if not -1 */
bool kSkipMinSeqlenQ_ = false /* skip min seqlen q while chunked prefill */>
index_t kBlockPerCu_ = -1, /* overwrite occupancy if not -1 */
bool kSkipMinSeqlenQ_ = false, /* skip min seqlen q while chunked prefill */
bool kHasSink_ = false>
struct TileFmhaFwdPagedKVTraits
{
static constexpr bool kPadSeqLenQ = kPadSeqLenQ_;
@@ -81,6 +84,7 @@ struct TileFmhaFwdPagedKVTraits
static constexpr bool kDoFp8StaticQuant = kDoFp8StaticQuant_;
static constexpr index_t kBlockPerCu = kBlockPerCu_;
static constexpr bool kSkipMinSeqlenQ = kSkipMinSeqlenQ_;
static constexpr bool kHasSink = kHasSink_;
};
template <bool kPadSeqLenQ_ /* padding for seqlen_q */,
@@ -95,7 +99,8 @@ template <bool kPadSeqLenQ_ /* padding for seqlen_q */,
bool kIsPagedKV_,
bool kHasUnevenSplits_,
bool kMergeNumHeadGroupsSeqLenQ_ = false,
index_t kBlockPerCu_ = -1 /* overwrite occupancy if not -1 */>
index_t kBlockPerCu_ = -1, /* overwrite occupancy if not -1 */
bool kHasSink_ = false>
struct TileFmhaFwdSplitKVTraits
{
static constexpr bool kPadSeqLenQ = kPadSeqLenQ_;
@@ -112,6 +117,7 @@ struct TileFmhaFwdSplitKVTraits
static constexpr bool kHasUnevenSplits = kHasUnevenSplits_;
static constexpr bool kMergeNumHeadGroupsSeqLenQ = kMergeNumHeadGroupsSeqLenQ_;
static constexpr index_t kBlockPerCu = kBlockPerCu_;
static constexpr bool kHasSink = kHasSink_;
};
template <bool kPadSeqLenQ_ /* padding for seqlen_q */,