Skip to content

GitLab

Commit 4d914af3 authored by Jun Liu's avatar Jun Liu
Browse files

Merge branch 'amd-develop' into amd-master

parents 223a2abe 4b798833
Hide whitespace changes
Inline Side-by-side
Showing with 453 additions and 243 deletions
include/ck_tile/ops/epilogue/default_2d_epilogue.hpp
View file @ 4d914af3
......@@ -9,23 +9,29 @@ namespace ck_tile {
// this epilogue just store out a M*N matrix, row major
template <typename AccDataType_, typename ODataType_, bool kPadM_, bool kPadN_>
template <typename AccDataType_,
typename ODataType_,
bool kPadM_,
bool kPadN_,
bool UseRawStore_ = true>
struct Default2DEpilogueProblem
{
using AccDataType = remove_cvref_t<AccDataType_>;
using ODataType = remove_cvref_t<ODataType_>;
static constexpr bool kPadM = kPadM_;
static constexpr bool kPadN = kPadN_;
using AccDataType = remove_cvref_t<AccDataType_>;
using ODataType = remove_cvref_t<ODataType_>;
static constexpr bool kPadM = kPadM_;
static constexpr bool kPadN = kPadN_;
static constexpr bool UseRawStore = UseRawStore_;
};
template <typename Problem_, typename Policy_ = void>
struct Default2DEpilogue
{
using Problem = remove_cvref_t<Problem_>;
using AccDataType = remove_cvref_t<typename Problem::AccDataType>;
using ODataType = remove_cvref_t<typename Problem::ODataType>;
static constexpr bool kPadM = Problem::kPadM;
static constexpr bool kPadN = Problem::kPadN;
using Problem = remove_cvref_t<Problem_>;
using AccDataType = remove_cvref_t<typename Problem::AccDataType>;
using ODataType = remove_cvref_t<typename Problem::ODataType>;
static constexpr bool kPadM = Problem::kPadM;
static constexpr bool kPadN = Problem::kPadN;
static constexpr bool UseRawStore = Problem::UseRawStore;
CK_TILE_HOST_DEVICE static constexpr index_t GetSmemSize() { return 0; }
......@@ -36,7 +42,7 @@ struct Default2DEpilogue
{
// TODO: this is ugly
if constexpr(kPadM || kPadN)
if constexpr(UseRawStore && (kPadM || kPadN))
{
store_tile_raw(o_dram_window_tmp, cast_tile<ODataType>(o_acc_tile));
buffer_store_fence();
......
include/ck_tile/ops/epilogue/dynamic_quant_epilogue.hpp 0 → 100644
View file @ 4d914af3
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck_tile/core.hpp"
#include "ck_tile/ops/reduce.hpp"
namespace ck_tile {
template <bool kPadM_, bool kPadN_, bool UseRawStore_ = true, bool UseMax3_ = false>
struct DynamicQuantEpilogueTraits
{
static constexpr bool kPadM = kPadM_;
static constexpr bool kPadN = kPadN_;
static constexpr bool UseRawStore = UseRawStore_;
static constexpr bool UseMax3 = UseMax3_;
};
// this epilogue just store out a M*N matrix, row major
template <typename AccDataType_,
typename YScaleDataType_,
typename ODataType_,
typename BlockShape_,
typename Traits_>
struct DynamicQuantEpilogueProblem
{
using AccDataType = remove_cvref_t<AccDataType_>;
using YScaleDataType = remove_cvref_t<YScaleDataType_>;
using ODataType = remove_cvref_t<ODataType_>;
using BlockShape = remove_cvref_t<BlockShape_>; // can consum generic 2d shape
using Traits = remove_cvref_t<Traits_>;
};
template <typename Problem_, typename Policy_ = void>
struct DynamicQuantEpilogue
{
using Problem = remove_cvref_t<Problem_>;
using AccDataType = remove_cvref_t<typename Problem::AccDataType>;
using YScaleDataType = remove_cvref_t<typename Problem::YScaleDataType>;
using ODataType = remove_cvref_t<typename Problem::ODataType>;
using BlockShape = remove_cvref_t<typename Problem::BlockShape>;
static constexpr bool kPadM = Problem::Traits::kPadM;
static constexpr bool kPadN = Problem::Traits::kPadN;
static constexpr bool UseRawStore = Problem::Traits::UseRawStore;
static constexpr bool UseMax3 = Problem::Traits::UseMax3;
CK_TILE_HOST_DEVICE static constexpr auto GetBlockReduce2d()
{
using P_ = BlockReduce2dProblem<AccDataType, AccDataType, BlockShape>;
return BlockReduce2d<P_>{};
}
CK_TILE_HOST_DEVICE static constexpr auto GetBlockReduce2dSync()
{
using P_ = BlockReduce2dProblem<AccDataType, AccDataType, BlockShape>;
return BlockReduce2dSync<P_>{};
}
CK_TILE_HOST_DEVICE static constexpr auto GetBlockReduce2dCrossWarpSync()
{
using P_ = BlockReduce2dProblem<AccDataType, AccDataType, BlockShape>;
return BlockReduce2dCrossWarpSync<P_>{};
}
CK_TILE_HOST_DEVICE static constexpr index_t GetSmemSize()
{
auto reduce_crosswarp_sync = GetBlockReduce2dCrossWarpSync();
return reduce_crosswarp_sync.GetSmemSize();
}
// TODO: this function assume store out vector size is the same as OAccTile last dimension size
// how do we fix this ?
template <typename ODramWindowTmp, typename YScaleWindow, typename OAccTile>
CK_TILE_DEVICE auto operator()(ODramWindowTmp& o_dram_window_tmp,
YScaleWindow& y_scale_window,
const OAccTile& o_acc_tile,
void* smem)
{
auto reduce = GetBlockReduce2d();
auto reduce_sync = GetBlockReduce2dSync();
auto reduce_crosswarp_sync = GetBlockReduce2dCrossWarpSync();
const auto f_absmax = [](auto acc_, auto v_0_) { return max(acc_, abs(v_0_)); };
auto row_absmax = [&]() {
constexpr auto y_size_per_row =
OAccTile{}.get_tile_distribution().get_ys_to_d_descriptor().get_lengths().at(
number<1>{});
// constexpr auto y_size_per_row = OAccTile::get_lengths()[number<1>{}];
if constexpr(UseMax3 && std::is_same_v<AccDataType, float> && y_size_per_row % 2 == 0)
{
// fast max3 implementation
const auto f_max3 = [](auto acc_, auto v_0_, auto v_1_) {
float rtn;
asm volatile("v_max3_f32 %0, %1, abs(%2), abs(%3)"
: "=v"(rtn)
: "v"(acc_), "v"(v_0_), "v"(v_1_));
return rtn;
};
return reduce(o_acc_tile, type_convert<AccDataType>(0), f_max3, sequence<1, 2>{});
}
else
{
return reduce(o_acc_tile, type_convert<AccDataType>(0), f_absmax);
}
}();
reduce_sync(row_absmax, f_absmax);
reduce_crosswarp_sync(row_absmax, smem, f_absmax);
// here y_scale is Acc TYpe, need convert to YScale type later
auto y_scale = tile_elementwise_in(
[&](const auto& v_) {
return v_ / type_convert<AccDataType>(numeric<ODataType>::max());
},
row_absmax);
store_tile(y_scale_window, cast_tile<YScaleDataType>(y_scale));
auto o_acc_scaled_tile =
make_static_distributed_tensor<AccDataType>(o_acc_tile.get_tile_distribution());
sweep_tile(o_acc_tile, [&](auto idx) {
constexpr auto row_id = make_tuple(idx[number<0>{}]);
o_acc_scaled_tile(idx) = o_acc_tile[idx] / y_scale(row_id);
});
// TODO: this is ugly
if constexpr(UseRawStore && (kPadM || kPadN))
{
store_tile_raw(o_dram_window_tmp, cast_tile<ODataType>(o_acc_scaled_tile));
buffer_store_fence();
}
else
{
store_tile(o_dram_window_tmp, cast_tile<ODataType>(o_acc_scaled_tile));
}
}
};
} // namespace ck_tile
include/ck_tile/ops/fmha.hpp
View file @ 4d914af3
......@@ -43,4 +43,5 @@
#include "ck_tile/ops/fmha/pipeline/block_fmha_pipeline_qx_ks_vs_custom_policy.hpp"
#include "ck_tile/ops/fmha/pipeline/tile_fmha_shape.hpp"
#include "ck_tile/ops/fmha/pipeline/tile_fmha_traits.hpp"
#include "ck_tile/ops/common/generic_2d_block_shape.hpp"
#include "ck_tile/ops/common/tensor_layout.hpp"
include/ck_tile/ops/fmha/kernel/fmha_fwd_kernel.hpp
View file @ 4d914af3
......@@ -69,7 +69,8 @@ struct FmhaFwdKernel
// sync with generate.py
// clang-format off
using bfs = typename FmhaPipeline::BlockFmhaShape;
using gbr = typename bfs::Gemm0BlockWarps;
using g0br = typename bfs::Gemm0BlockWarps;
using g1br = typename bfs::Gemm1BlockWarps;
using gwt = typename bfs::Gemm0WarpTile;
#define _SS_ std::string
#define _TS_ std::to_string
......@@ -81,11 +82,12 @@ struct FmhaFwdKernel
if (kPadHeadDimV) n += "dv";
return n.empty() ? n : std::string("p") + n; }();
return
_SS_("fmha_fwd_d") + _TS_(bfs::kK0BlockLength) + "_" + _SS_(t2s<QDataType>::name) +
_SS_("fmha_fwd_d") + _TS_(bfs::kQKHeaddim) + "_" + _SS_(t2s<QDataType>::name) +
"_" + (kIsGroupMode ? "group" : "batch") + "_" + _SS_(TilePartitioner::name) + "_"
"b" + _TS_(bfs::kM0) + "x" + _TS_(bfs::kN0) + "x" + _TS_(bfs::kK0) + "x" +
_TS_(bfs::kN1) + "x" + _TS_(bfs::kK1) + "x" + _TS_(bfs::kK0BlockLength) + "_" +
"r" + _TS_(gbr::at(ck_tile::number<0>{})) + "x" + _TS_(gbr::at(ck_tile::number<1>{})) + "x" + _TS_(gbr::at(ck_tile::number<2>{})) + "_" +
_TS_(bfs::kN1) + "x" + _TS_(bfs::kK1) + "x" + _TS_(bfs::kQKHeaddim) + "_" +
"r" + _TS_(g0br::at(ck_tile::number<0>{})) + "x" + _TS_(g0br::at(ck_tile::number<1>{})) + "x" + _TS_(g0br::at(ck_tile::number<2>{})) + "_" +
"r" + _TS_(g1br::at(ck_tile::number<0>{})) + "x" + _TS_(g1br::at(ck_tile::number<1>{})) + "x" + _TS_(g1br::at(ck_tile::number<2>{})) + "_" +
"w" + _TS_(gwt::at(ck_tile::number<0>{})) + "x" + _TS_(gwt::at(ck_tile::number<1>{})) + "x" + _TS_(gwt::at(ck_tile::number<2>{})) + "_" +
(kBlockPerCuInput == -1 ? "" : ("o" + _TS_(kBlockPerCu) + "_")) + _SS_(FmhaPipeline::name) + "_" +
"v" + (std::is_same_v<VLayout, ck_tile::tensor_layout::gemm::RowMajor> ? "r" : "c") + (pn.empty() ? "" : "_" + pn) +
......@@ -655,7 +657,7 @@ struct FmhaFwdKernel
{
return pad_tensor_view(
q_dram_naive,
make_tuple(number<FmhaPipeline::kM0>{}, number<FmhaPipeline::kK0BlockLength>{}),
make_tuple(number<FmhaPipeline::kM0>{}, number<FmhaPipeline::kSubQKHeaddim>{}),
sequence<kPadSeqLenQ, kPadHeadDimQ>{});
}
else
......@@ -722,7 +724,7 @@ struct FmhaFwdKernel
[&]() {
if constexpr(FmhaPipeline::kQLoadOnce)
return make_tuple(number<FmhaPipeline::kM0>{},
number<FmhaPipeline::kK0BlockLength>{});
number<FmhaPipeline::kSubQKHeaddim>{});
else
return make_tuple(number<FmhaPipeline::kM0>{}, number<FmhaPipeline::kK0>{});
}(),
......
include/ck_tile/ops/fmha/kernel/fmha_fwd_splitkv_kernel.hpp
View file @ 4d914af3
......@@ -65,7 +65,8 @@ struct FmhaFwdSplitKVKernel
// sync with generate.py
// clang-format off
using bfs = typename FmhaPipeline::BlockFmhaShape;
using gbr = typename bfs::Gemm0BlockWarps;
using g0br = typename bfs::Gemm0BlockWarps;
using g1br = typename bfs::Gemm1BlockWarps;
using gwt = typename bfs::Gemm0WarpTile;
#define _SS_ std::string
#define _TS_ std::to_string
......@@ -77,11 +78,12 @@ struct FmhaFwdSplitKVKernel
if (kPadHeadDimV) n += "dv";
return n.empty() ? n : std::string("p") + n; }();
return
_SS_("fmha_fwd_splitkv_d") + _TS_(bfs::kK0BlockLength) + "_" + _SS_(t2s<QDataType>::name) +
_SS_("fmha_fwd_splitkv_d") + _TS_(bfs::kQKHeaddim) + "_" + _SS_(t2s<QDataType>::name) +
"_" + (kIsGroupMode ? "group" : "batch") + "_"
"b" + _TS_(bfs::kM0) + "x" + _TS_(bfs::kN0) + "x" + _TS_(bfs::kK0) + "x" +
_TS_(bfs::kN1) + "x" + _TS_(bfs::kK1) + "x" + _TS_(bfs::kK0BlockLength) + "_" +
"r" + _TS_(gbr::at(ck_tile::number<0>{})) + "x" + _TS_(gbr::at(ck_tile::number<1>{})) + "x" + _TS_(gbr::at(ck_tile::number<2>{})) + "_" +
_TS_(bfs::kN1) + "x" + _TS_(bfs::kK1) + "x" + _TS_(bfs::kQKHeaddim) + "_" +
"r" + _TS_(g0br::at(ck_tile::number<0>{})) + "x" + _TS_(g0br::at(ck_tile::number<1>{})) + "x" + _TS_(g0br::at(ck_tile::number<2>{})) + "_" +
"r" + _TS_(g1br::at(ck_tile::number<0>{})) + "x" + _TS_(g1br::at(ck_tile::number<1>{})) + "x" + _TS_(g1br::at(ck_tile::number<2>{})) + "_" +
"w" + _TS_(gwt::at(ck_tile::number<0>{})) + "x" + _TS_(gwt::at(ck_tile::number<1>{})) + "x" + _TS_(gwt::at(ck_tile::number<2>{})) + "_" +
(kBlockPerCuInput == -1 ? "" : ("o" + _TS_(kBlockPerCu) + "_")) + _SS_(FmhaPipeline::name) + "_" +
"v" + (std::is_same_v<VLayout, ck_tile::tensor_layout::gemm::RowMajor> ? "r" : "c") + (pn.empty() ? "" : "_" + pn) +
......@@ -584,7 +586,7 @@ struct FmhaFwdSplitKVKernel
{
return pad_tensor_view(
q_dram_naive,
make_tuple(number<FmhaPipeline::kM0>{}, number<FmhaPipeline::kK0BlockLength>{}),
make_tuple(number<FmhaPipeline::kM0>{}, number<FmhaPipeline::kSubQKHeaddim>{}),
sequence<kPadSeqLenQ, kPadHeadDimQ>{});
}
else
......@@ -733,7 +735,7 @@ struct FmhaFwdSplitKVKernel
[&]() {
if constexpr(FmhaPipeline::kQLoadOnce)
return make_tuple(number<FmhaPipeline::kM0>{},
number<FmhaPipeline::kK0BlockLength>{});
number<FmhaPipeline::kSubQKHeaddim>{});
else
return make_tuple(number<FmhaPipeline::kM0>{}, number<FmhaPipeline::kK0>{});
}(),
......@@ -894,7 +896,7 @@ struct FmhaFwdSplitKVKernel
o_acc_ptr,
make_tuple(kargs.seqlen_q, kargs.hdim_v),
make_tuple(kargs.stride_o_acc, 1),
number<1>{},
number<FmhaPipeline::kAlignmentOacc>{},
number<1>{});
return pad_tensor_view(
......
include/ck_tile/ops/fmha/kernel/fmha_fwd_splitkv_tile_partitioner.hpp
View file @ 4d914af3
......@@ -26,8 +26,8 @@ struct FmhaFwdSplitKVTilePartitioner
{
// TODO: this may need tuning
return dim3(ck_tile::integer_divide_ceil(max_seqlen_q, kM0) *
ck_tile::integer_divide_ceil(hdim_v, kN1),
nhead * num_splits,
ck_tile::integer_divide_ceil(hdim_v, kN1) * num_splits,
nhead,
batch_size);
}
......@@ -42,8 +42,9 @@ struct FmhaFwdSplitKVTilePartitioner
return ck_tile::make_tuple(quotient, modulus);
};
const auto [i_tile_m, i_tile_n] = f(blockIdx.x, num_tile_n1);
const auto [i_nhead, i_split] = f(blockIdx.y, num_splits);
const auto [mn, i_split] = f(blockIdx.x, num_splits);
const auto [i_tile_m, i_tile_n] = f(mn, num_tile_n1);
const index_t i_nhead = blockIdx.y;
const index_t i_batch = blockIdx.z;
return ck_tile::make_tuple(i_tile_m, i_tile_n, i_split, i_nhead, i_batch);
......
include/ck_tile/ops/fmha/pipeline/block_fmha_fwd_splitkv_combine_pipeline.hpp
View file @ 4d914af3
......@@ -12,6 +12,16 @@ namespace detail {
template <index_t N>
struct log2;
template <>
struct log2<4> : std::integral_constant<index_t, 2>
{
};
template <>
struct log2<8> : std::integral_constant<index_t, 3>
{
};
template <>
struct log2<16> : std::integral_constant<index_t, 4>
{
......@@ -72,18 +82,18 @@ struct BlockFmhaFwdSplitKVCombinePipeline
{
if constexpr(kHeadDimV <= 32)
{
constexpr std::array<int, 4> occupancy{3, 3, 3, 1};
return occupancy[detail::log2<kMaxSplits>::value - 4];
constexpr std::array occupancy{3, 3, 3, 3, 3, 1};
return occupancy[detail::log2<kMaxSplits>::value - 2];
}
else if constexpr(kHeadDimV <= 128)
{
constexpr std::array<int, 4> occupancy{3, 3, 2, 1};
return occupancy[detail::log2<kMaxSplits>::value - 4];
constexpr std::array occupancy{3, 3, 3, 3, 2, 1};
return occupancy[detail::log2<kMaxSplits>::value - 2];
}
else if constexpr(kHeadDimV <= 256)
{
constexpr std::array<int, 4> occupancy{2, 2, 2, 1};
return occupancy[detail::log2<kMaxSplits>::value - 4];
constexpr std::array occupancy{2, 2, 2, 2, 2, 1};
return occupancy[detail::log2<kMaxSplits>::value - 2];
}
}
}();
......@@ -138,9 +148,8 @@ struct BlockFmhaFwdSplitKVCombinePipeline
auto lse_accum = make_static_distributed_tensor<LSEDataType>(
Policy::template MakeLSEaccRegTileDistribution<Problem>());
// copy LDS (shape=[kM0, kMaxSplits]) to lse_accum (shape=[kM0, max(kMaxSplits, warp_size)])
// this will extend the distributed tensor width so that each thread in wave have data to
// reduce.
// copy LDS (shape=[kM0, kMaxSplits]) to lse_accum (shape=[kM0, kMaxSplits])
// and fill up -INF values outside the [kM0, num_splits] region.
{
constexpr auto spans = decltype(lse_accum)::get_distributed_spans();
sweep_tile_span(spans[number<0>{}], [&](auto idx0) {
......
include/ck_tile/ops/fmha/pipeline/block_fmha_fwd_splitkv_combine_pipeline_default_policy.hpp
View file @ 4d914af3
......@@ -10,11 +10,26 @@ namespace ck_tile {
struct BlockFmhaFwdSplitKVCombinePipelineDefaultPolicy
{
template <index_t BlockSize, index_t M, index_t N, typename DataType>
CK_TILE_HOST_DEVICE static constexpr auto GetVectorSizeForTile()
{
constexpr index_t PixelsPerThread = (M * N) / BlockSize;
static_assert(0 < PixelsPerThread);
constexpr index_t MaxNPerThread = 16 / sizeof(DataType);
constexpr index_t NPerThread = min(MaxNPerThread, PixelsPerThread);
return NPerThread;
}
// alignment for dram lse tile (shape=[kMaxSplits, kM0])
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto GetAlignmentLSE()
{
using LSEDataType = remove_cvref_t<typename Problem::LSEDataType>;
return 16 / sizeof(LSEDataType);
return GetVectorSizeForTile<Problem::kBlockSize,
Problem::kMaxSplits,
Problem::kM0,
typename Problem::LSEDataType>();
}
template <typename Problem>
......@@ -47,29 +62,31 @@ struct BlockFmhaFwdSplitKVCombinePipelineDefaultPolicy
MakeLSEaccLdsBlockDescriptor<Problem>().get_element_space_size();
}
// shape=[kMaxSplits, kM0]
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto MakeLSEaccDramTileDistribution()
{
using LSEDataType = remove_cvref_t<typename Problem::LSEDataType>;
constexpr index_t kBlockSize = Problem::kBlockSize;
constexpr index_t kNumWarps = Problem::kNumWarps;
constexpr index_t kNPerBlock = Problem::kM0;
constexpr index_t kMPerBlock = Problem::kMaxSplits;
constexpr index_t NPerThread = 16 / sizeof(LSEDataType);
constexpr index_t NThreads = kNPerBlock / NPerThread;
constexpr index_t NPerThread =
GetVectorSizeForTile<kBlockSize, kMPerBlock, kNPerBlock, LSEDataType>();
constexpr index_t NThreads = kNPerBlock / NPerThread;
constexpr index_t MThreadsPerWarp = get_warp_size() / NThreads;
constexpr index_t TotalWarps = kBlockSize / get_warp_size();
constexpr index_t MPerThread = kMPerBlock / (TotalWarps * MThreadsPerWarp);
constexpr index_t MPerThread = kMPerBlock / (kNumWarps * MThreadsPerWarp);
static_assert(NThreads * NPerThread == kNPerBlock);
static_assert(MPerThread * TotalWarps * MThreadsPerWarp == kMPerBlock);
static_assert(MPerThread * kNumWarps * MThreadsPerWarp == kMPerBlock);
return make_static_tile_distribution(
tile_distribution_encoding<sequence<1>,
tuple<sequence<MPerThread, TotalWarps, MThreadsPerWarp>,
tuple<sequence<MPerThread, kNumWarps, MThreadsPerWarp>,
sequence<NThreads, NPerThread>>,
tuple<sequence<1>, sequence<1, 2>>,
tuple<sequence<1>, sequence<2, 0>>,
......@@ -77,15 +94,18 @@ struct BlockFmhaFwdSplitKVCombinePipelineDefaultPolicy
sequence<0, 1>>{});
}
// 3d + padding, [kMaxSplits, kM0]
// 3d + padding, shape=[kMaxSplits, kM0]
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto MakeLSEaccLdsStoreBlockDescriptor()
{
using LSEDataType = remove_cvref_t<typename Problem::LSEDataType>;
constexpr index_t kBlockSize = Problem::kBlockSize;
constexpr index_t kMPerBlock = Problem::kMaxSplits;
constexpr index_t kNPerBlock = Problem::kM0;
constexpr index_t NPack = 16 / sizeof(LSEDataType);
constexpr index_t NPack =
GetVectorSizeForTile<kBlockSize, kMPerBlock, kNPerBlock, LSEDataType>();
constexpr auto lse_acc_lds_block_desc_0 = make_naive_tensor_descriptor(
make_tuple(number<kNPerBlock / NPack>{}, number<kMPerBlock>{}, number<NPack>{}),
......@@ -103,15 +123,18 @@ struct BlockFmhaFwdSplitKVCombinePipelineDefaultPolicy
return lse_acc_lds_block_desc;
}
// 3d + padding, [kM0, kMaxSplits]
// 3d + padding, shape=[kM0, kMaxSplits]
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto MakeLSEaccLdsBlockDescriptor()
{
using LSEDataType = remove_cvref_t<typename Problem::LSEDataType>;
constexpr index_t kBlockSize = Problem::kBlockSize;
constexpr index_t kMPerBlock = Problem::kMaxSplits;
constexpr index_t kNPerBlock = Problem::kM0;
constexpr index_t NPack = 16 / sizeof(LSEDataType);
constexpr index_t NPack =
GetVectorSizeForTile<kBlockSize, kMPerBlock, kNPerBlock, LSEDataType>();
constexpr auto lse_acc_lds_block_desc_0 = make_naive_tensor_descriptor(
make_tuple(number<kNPerBlock / NPack>{}, number<kMPerBlock>{}, number<NPack>{}),
......@@ -134,26 +157,28 @@ struct BlockFmhaFwdSplitKVCombinePipelineDefaultPolicy
{
constexpr index_t kBlockSize = Problem::kBlockSize;
constexpr index_t kNPerBlock = max(Problem::kMaxSplits, get_warp_size());
constexpr index_t kNPerBlock = Problem::kMaxSplits;
constexpr index_t kMPerBlock = Problem::kM0;
constexpr index_t NThreads = get_warp_size();
constexpr index_t NThreads = 4;
constexpr index_t NPerThread = kNPerBlock / NThreads;
constexpr index_t MThreads = kBlockSize / NThreads;
constexpr index_t MPerThread = kMPerBlock / MThreads;
constexpr index_t MThreads = kBlockSize / NThreads;
constexpr index_t MPerThread = kMPerBlock / MThreads;
constexpr index_t MWarps = kBlockSize / get_warp_size();
constexpr index_t MThreadPerWarp = get_warp_size() / NThreads;
static_assert(NThreads * NPerThread == kNPerBlock);
static_assert(MThreads * MPerThread == kMPerBlock);
static_assert(MWarps * MThreadPerWarp * MPerThread == kMPerBlock);
return make_static_tile_distribution(
tile_distribution_encoding<
sequence<1>,
tuple<sequence<MThreads, MPerThread>, sequence<NThreads, NPerThread>>,
tuple<sequence<1>, sequence<2>>,
tuple<sequence<0>, sequence<0>>,
tuple<sequence<MWarps, MThreadPerWarp, MPerThread>, sequence<NThreads, NPerThread>>,
tuple<sequence<1>, sequence<2, 1>>,
tuple<sequence<0>, sequence<0, 1>>,
sequence<1, 2>,
sequence<1, 1>>{});
sequence<2, 1>>{});
}
template <typename Problem>
......
include/ck_tile/ops/fmha/pipeline/block_fmha_fwd_splitkv_pipeline_qr_ks_vs.hpp
View file @ 4d914af3
......@@ -34,12 +34,13 @@ struct BlockFmhaFwdSplitKVPipelineQRKSVS
static constexpr index_t kBlockSize = Problem::kBlockSize;
static constexpr index_t kM0 = BlockFmhaShape::kM0;
static constexpr index_t kN0 = BlockFmhaShape::kN0;
static constexpr index_t kK0 = BlockFmhaShape::kK0;
static constexpr index_t kN1 = BlockFmhaShape::kN1;
static constexpr index_t kK1 = BlockFmhaShape::kK1;
static constexpr index_t kK0BlockLength = BlockFmhaShape::kK0BlockLength;
static constexpr index_t kM0 = BlockFmhaShape::kM0;
static constexpr index_t kN0 = BlockFmhaShape::kN0;
static constexpr index_t kK0 = BlockFmhaShape::kK0;
static constexpr index_t kN1 = BlockFmhaShape::kN1;
static constexpr index_t kK1 = BlockFmhaShape::kK1;
static constexpr index_t kQKHeaddim = BlockFmhaShape::kQKHeaddim;
static constexpr index_t kSubQKHeaddim = BlockFmhaShape::kSubQKHeaddim;
static constexpr bool kIsGroupMode = Problem::kIsGroupMode;
static constexpr bool kPadSeqLenQ = Problem::kPadSeqLenQ;
......@@ -64,6 +65,9 @@ struct BlockFmhaFwdSplitKVPipelineQRKSVS
return kPadSeqLenK ? 1 : Policy::template GetAlignmentV<Problem>();
}();
static constexpr index_t kAlignmentOacc =
kPadHeadDimV ? 1 : Policy::template GetAlignmentOacc<Problem>();
static constexpr index_t kAlignmentBias =
kPadSeqLenK ? 1 : Policy::template GetAlignmentBias<Problem>();
......@@ -72,22 +76,22 @@ struct BlockFmhaFwdSplitKVPipelineQRKSVS
return Problem::kBlockPerCu;
else
{
if constexpr(kK0BlockLength <= 32)
if constexpr(kQKHeaddim <= 32)
{
return 2;
}
else if constexpr(kK0BlockLength <= 64)
else if constexpr(kQKHeaddim <= 64)
{
return 3;
}
else if constexpr(kK0BlockLength <= 128)
else if constexpr(kQKHeaddim <= 128)
{
if constexpr(BiasEnum == BlockAttentionBiasEnum::ELEMENTWISE_BIAS)
return 1;
else
return 2;
}
else if constexpr(kK0BlockLength <= 256)
else if constexpr(kQKHeaddim <= 256)
{
return 1;
}
......@@ -252,11 +256,11 @@ struct BlockFmhaFwdSplitKVPipelineQRKSVS
k_dram_block_window_lengths, {adjusted_seqlen_k_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>{}), adjusted_seqlen_k_start}, // M/N
Policy::template MakeBiasDramTileDistribution<Problem, decltype(gemm_0)>());
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>{}), adjusted_seqlen_k_start}, // M/N
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,
......@@ -267,7 +271,7 @@ struct BlockFmhaFwdSplitKVPipelineQRKSVS
// prefetch K tile
index_t i_total_loops = 0;
constexpr index_t k0_loops = kK0BlockLength / kK0;
constexpr index_t k0_loops = kQKHeaddim / kK0;
constexpr index_t k1_loops = kN0 / kK1;
static_assert(2 <= k0_loops);
......
include/ck_tile/ops/fmha/pipeline/block_fmha_fwd_splitkv_pipeline_qr_ks_vs_default_policy.hpp
View file @ 4d914af3
......@@ -9,11 +9,20 @@
namespace ck_tile {
// This pipeline is qkv all located in LDS
using BlockFmhaFwdSplitKVPipelineQRKSVSDefaultPolicy =
BlockFmhaPipelineQXKSVSCustomPolicy</* QLoadOnce = */ true,
/* AsyncCopyK = */ false,
/* AsyncCopyV = */ false,
/* NumPrefetchK = */ 1,
/* NumPrefetchV = */ 1>;
struct BlockFmhaFwdSplitKVPipelineQRKSVSDefaultPolicy
: BlockFmhaPipelineQXKSVSCustomPolicy</* QLoadOnce = */ true,
/* AsyncCopyK = */ false,
/* AsyncCopyV = */ false,
/* NumPrefetchK = */ 1,
/* NumPrefetchV = */ 1>
{
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto GetAlignmentOacc()
{
using OaccDataType = remove_cvref_t<typename Problem::OaccDataType>;
return static_cast<index_t>(16 / sizeof(OaccDataType));
}
};
} // namespace ck_tile
include/ck_tile/ops/fmha/pipeline/block_fmha_pipeline_problem.hpp
View file @ 4d914af3
......@@ -39,8 +39,11 @@ struct BlockFmhaPipelineProblem
using FmhaMask = remove_cvref_t<FmhaMask_>;
using Traits = remove_cvref_t<Traits_>;
static constexpr index_t kBlockSize = BlockFmhaShape::NumWarps * get_warp_size();
static constexpr bool kIsGroupMode = kIsGroupMode_;
static constexpr index_t kNumGemm0Warps = BlockFmhaShape::NumGemm0Warps;
static constexpr index_t kNumGemm1Warps = BlockFmhaShape::NumGemm1Warps;
static constexpr index_t kBlockSize = BlockFmhaShape::NumWarps * get_warp_size();
static constexpr bool kIsGroupMode = kIsGroupMode_;
// attributes from traits
static constexpr bool kPadSeqLenQ = Traits::kPadSeqLenQ;
......@@ -84,8 +87,11 @@ struct BlockFmhaFwdSplitKVPipelineProblem
using FmhaMask = remove_cvref_t<FmhaMask_>;
using Traits = remove_cvref_t<Traits_>;
static constexpr index_t kBlockSize = BlockFmhaShape::NumWarps * get_warp_size();
static constexpr bool kIsGroupMode = kIsGroupMode_;
static constexpr index_t kNumGemm0Warps = BlockFmhaShape::NumGemm0Warps;
static constexpr index_t kNumGemm1Warps = BlockFmhaShape::NumGemm1Warps;
static constexpr index_t kBlockSize = BlockFmhaShape::NumWarps * get_warp_size();
static constexpr bool kIsGroupMode = kIsGroupMode_;
// attributes from traits
static constexpr bool kPadSeqLenQ = Traits::kPadSeqLenQ;
......@@ -115,7 +121,8 @@ struct BlockFmhaSplitKVCombinePipelineProblem
using ODataType = remove_cvref_t<ODataType_>;
using Traits = remove_cvref_t<Traits_>;
static constexpr index_t kBlockSize = 256;
static constexpr index_t kNumWarps = kM0_ / (get_warp_size() / 4);
static constexpr index_t kBlockSize = kNumWarps * get_warp_size();
static constexpr bool kIsGroupMode = kIsGroupMode_;
static constexpr index_t kHeadDimV = HeadDimV_;
......
include/ck_tile/ops/fmha/pipeline/block_fmha_pipeline_qr_ks_vs.hpp
View file @ 4d914af3
......@@ -37,12 +37,13 @@ struct BlockFmhaPipelineQRKSVS
static constexpr index_t kBlockSize = Problem::kBlockSize;
static constexpr index_t kM0 = BlockFmhaShape::kM0;
static constexpr index_t kN0 = BlockFmhaShape::kN0;
static constexpr index_t kK0 = BlockFmhaShape::kK0;
static constexpr index_t kN1 = BlockFmhaShape::kN1;
static constexpr index_t kK1 = BlockFmhaShape::kK1;
static constexpr index_t kK0BlockLength = BlockFmhaShape::kK0BlockLength;
static constexpr index_t kM0 = BlockFmhaShape::kM0;
static constexpr index_t kN0 = BlockFmhaShape::kN0;
static constexpr index_t kK0 = BlockFmhaShape::kK0;
static constexpr index_t kN1 = BlockFmhaShape::kN1;
static constexpr index_t kK1 = BlockFmhaShape::kK1;
static constexpr index_t kQKHeaddim = BlockFmhaShape::kQKHeaddim;
static constexpr index_t kSubQKHeaddim = BlockFmhaShape::kSubQKHeaddim;
static constexpr bool kIsGroupMode = Problem::kIsGroupMode;
static constexpr bool kPadSeqLenQ = Problem::kPadSeqLenQ;
......@@ -76,22 +77,22 @@ struct BlockFmhaPipelineQRKSVS
return Problem::kBlockPerCu;
else
{
if constexpr(kK0BlockLength <= 32)
if constexpr(kQKHeaddim <= 32)
{
return 2;
}
else if constexpr(kK0BlockLength <= 64)
else if constexpr(kQKHeaddim <= 64)
{
return 3;
}
else if constexpr(kK0BlockLength <= 128)
else if constexpr(kQKHeaddim <= 128)
{
if constexpr(BiasEnum == BlockAttentionBiasEnum::ELEMENTWISE_BIAS)
return 1;
else
return 2;
}
else if constexpr(kK0BlockLength <= 256)
else if constexpr(kQKHeaddim <= 256)
{
return 1;
}
......@@ -242,11 +243,11 @@ struct BlockFmhaPipelineQRKSVS
{seqlen_k_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
Policy::template MakeBiasDramTileDistribution<Problem, decltype(gemm_0)>());
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
Policy::template MakeBiasDramTileDistribution<decltype(gemm_0)>());
auto randval_dram_window = dropout.template MakeRandvalDramWindow<decltype(gemm_0)>(
randval_dram_block_window_tmp, seqlen_k_start);
......@@ -261,7 +262,7 @@ struct BlockFmhaPipelineQRKSVS
// prefetch K tile
index_t i_total_loops = 0;
constexpr index_t k0_loops = kK0BlockLength / kK0;
constexpr index_t k0_loops = kQKHeaddim / kK0;
constexpr index_t k1_loops = kN0 / kK1;
static_assert(2 <= k0_loops);
......
include/ck_tile/ops/fmha/pipeline/block_fmha_pipeline_qr_ks_vs_async.hpp
View file @ 4d914af3
......@@ -38,12 +38,13 @@ struct BlockFmhaPipelineQRKSVSAsync
static constexpr index_t kBlockSize = Problem::kBlockSize;
static constexpr index_t kM0 = BlockFmhaShape::kM0;
static constexpr index_t kN0 = BlockFmhaShape::kN0;
static constexpr index_t kK0 = BlockFmhaShape::kK0;
static constexpr index_t kN1 = BlockFmhaShape::kN1;
static constexpr index_t kK1 = BlockFmhaShape::kK1;
static constexpr index_t kK0BlockLength = BlockFmhaShape::kK0BlockLength;
static constexpr index_t kM0 = BlockFmhaShape::kM0;
static constexpr index_t kN0 = BlockFmhaShape::kN0;
static constexpr index_t kK0 = BlockFmhaShape::kK0;
static constexpr index_t kN1 = BlockFmhaShape::kN1;
static constexpr index_t kK1 = BlockFmhaShape::kK1;
static constexpr index_t kQKHeaddim = BlockFmhaShape::kQKHeaddim;
static constexpr index_t kSubQKHeaddim = BlockFmhaShape::kSubQKHeaddim;
static constexpr bool kIsGroupMode = Problem::kIsGroupMode;
// TODO: seq_q always support padding, hdim_q/v support multiple of vector(like 8x)
......@@ -87,7 +88,7 @@ struct BlockFmhaPipelineQRKSVSAsync
return 1;
}
if constexpr(kK0BlockLength <= 32)
if constexpr(kQKHeaddim <= 32)
{
if constexpr(kPadSeqLenK && BiasEnum == BlockAttentionBiasEnum::ELEMENTWISE_BIAS &&
FmhaMask::IsMasking)
......@@ -95,21 +96,21 @@ struct BlockFmhaPipelineQRKSVSAsync
else
return 2;
}
else if constexpr(kK0BlockLength <= 64)
else if constexpr(kQKHeaddim <= 64)
{
if constexpr(kPadSeqLenK && BiasEnum == BlockAttentionBiasEnum::ELEMENTWISE_BIAS)
return 2;
else
return 3;
}
else if constexpr(kK0BlockLength <= 128)
else if constexpr(kQKHeaddim <= 128)
{
if constexpr(kPadSeqLenK && BiasEnum == BlockAttentionBiasEnum::ELEMENTWISE_BIAS)
return 1;
else
return 2;
}
else if constexpr(kK0BlockLength <= 256)
else if constexpr(kQKHeaddim <= 256)
{
return 1;
}
......@@ -314,11 +315,11 @@ struct BlockFmhaPipelineQRKSVSAsync
}();
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
Policy::template MakeBiasDramTileDistribution<Problem, decltype(gemm_0)>());
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
Policy::template MakeBiasDramTileDistribution<decltype(gemm_0)>());
auto randval_dram_window = dropout.template MakeRandvalDramWindow<decltype(gemm_0)>(
randval_dram_block_window_tmp, seqlen_k_start);
......@@ -334,12 +335,12 @@ struct BlockFmhaPipelineQRKSVSAsync
move_tile_window(k_dram_window, {0, kK0});
__builtin_amdgcn_sched_barrier(0);
buffer_load_fence(k_dram_window.get_num_access(), q.get_thread_buffer());
buffer_load_fence(k_dram_window.get_num_of_access(), q.get_thread_buffer());
(void)q_element_func; // ??? rocm-6.x if use q element func will have scratch on hdim=64/32
// auto q_tile = q; // tile_elementwise_in(q_element_func, q);
index_t i_total_loops = 0;
constexpr index_t k0_loops = kK0BlockLength / kK0;
constexpr index_t k0_loops = kQKHeaddim / kK0;
constexpr index_t k1_loops = kN0 / kK1;
static_assert(1 <= k0_loops);
......@@ -359,7 +360,7 @@ struct BlockFmhaPipelineQRKSVSAsync
if constexpr(i_k0 < k0_loops - 1)
move_tile_window(k_dram_window, {0, kK0});
async_load_fence(k_dram_window.get_num_access());
async_load_fence(k_dram_window.get_num_of_access());
__builtin_amdgcn_s_barrier();
__builtin_amdgcn_sched_barrier(0);
gemm_0(s_acc,
......
include/ck_tile/ops/fmha/pipeline/block_fmha_pipeline_qr_ks_vs_fp8.hpp
View file @ 4d914af3
......@@ -36,12 +36,12 @@ struct [[deprecated]] BlockFmhaPipelineQRKSVSFp8
static constexpr index_t kBlockSize = Problem::kBlockSize;
static constexpr index_t kM0 = BlockFmhaShape::kM0;
static constexpr index_t kN0 = BlockFmhaShape::kN0;
static constexpr index_t kK0 = BlockFmhaShape::kK0;
static constexpr index_t kN1 = BlockFmhaShape::kN1;
static constexpr index_t kK1 = BlockFmhaShape::kK1;
static constexpr index_t kK0BlockLength = BlockFmhaShape::kK0BlockLength;
static constexpr index_t kM0 = BlockFmhaShape::kM0;
static constexpr index_t kN0 = BlockFmhaShape::kN0;
static constexpr index_t kK0 = BlockFmhaShape::kK0;
static constexpr index_t kN1 = BlockFmhaShape::kN1;
static constexpr index_t kK1 = BlockFmhaShape::kK1;
static constexpr index_t kQKHeaddim = BlockFmhaShape::kQKHeaddim;
static constexpr bool kIsGroupMode = Problem::kIsGroupMode;
static constexpr bool kPadSeqLenQ = Problem::kPadSeqLenQ;
......@@ -75,22 +75,22 @@ struct [[deprecated]] BlockFmhaPipelineQRKSVSFp8
return Problem::kBlockPerCu;
else
{
if constexpr(kK0BlockLength <= 32)
if constexpr(kQKHeaddim <= 32)
{
return 2;
}
else if constexpr(kK0BlockLength <= 64)
else if constexpr(kQKHeaddim <= 64)
{
return 3;
}
else if constexpr(kK0BlockLength <= 128)
else if constexpr(kQKHeaddim <= 128)
{
if constexpr(BiasEnum == BlockAttentionBiasEnum::ELEMENTWISE_BIAS)
return 1;
else
return 2;
}
else if constexpr(kK0BlockLength <= 256)
else if constexpr(kQKHeaddim <= 256)
{
return 1;
}
......@@ -232,7 +232,7 @@ struct [[deprecated]] BlockFmhaPipelineQRKSVSFp8
// prefetch K tile
index_t i_total_loops = 0;
constexpr index_t k0_loops = kK0BlockLength / kK0;
constexpr index_t k0_loops = kQKHeaddim / kK0;
constexpr index_t k1_loops = kN0 / kK1;
static_assert(2 <= k0_loops);
......
include/ck_tile/ops/fmha/pipeline/block_fmha_pipeline_qs_ks_vs.hpp
View file @ 4d914af3
......@@ -9,9 +9,10 @@
namespace ck_tile {
/// NOTICE: we no-longer use this pipeline.
// This pipeline is qkv all located in LDS
template <typename Problem_, typename Policy_ = BlockFmhaPipelineQSKSVSDefaultPolicy>
struct BlockFmhaPipelineQSKSVS
struct [[deprecated]] BlockFmhaPipelineQSKSVS
{
using Problem = remove_cvref_t<Problem_>;
using Policy = remove_cvref_t<Policy_>;
......@@ -35,12 +36,13 @@ struct BlockFmhaPipelineQSKSVS
static constexpr index_t kBlockSize = Problem::kBlockSize;
static constexpr index_t kM0 = BlockFmhaShape::kM0;
static constexpr index_t kN0 = BlockFmhaShape::kN0;
static constexpr index_t kK0 = BlockFmhaShape::kK0;
static constexpr index_t kN1 = BlockFmhaShape::kN1;
static constexpr index_t kK1 = BlockFmhaShape::kK1;
static constexpr index_t kK0BlockLength = BlockFmhaShape::kK0BlockLength;
static constexpr index_t kM0 = BlockFmhaShape::kM0;
static constexpr index_t kN0 = BlockFmhaShape::kN0;
static constexpr index_t kK0 = BlockFmhaShape::kK0;
static constexpr index_t kN1 = BlockFmhaShape::kN1;
static constexpr index_t kK1 = BlockFmhaShape::kK1;
static constexpr index_t kQKHeaddim = BlockFmhaShape::kQKHeaddim;
static constexpr index_t kSubQKHeaddim = BlockFmhaShape::kSubQKHeaddim;
static constexpr bool kIsGroupMode = Problem::kIsGroupMode;
static constexpr bool kPadSeqLenQ = Problem::kPadSeqLenQ;
......@@ -55,22 +57,22 @@ struct BlockFmhaPipelineQSKSVS
return Problem::kBlockPerCu;
else
{
if constexpr(kK0BlockLength <= 32)
if constexpr(kQKHeaddim <= 32)
{
return 2;
}
else if constexpr(kK0BlockLength <= 64)
else if constexpr(kQKHeaddim <= 64)
{
return 3;
}
else if constexpr(kK0BlockLength <= 128)
else if constexpr(kQKHeaddim <= 128)
{
if constexpr(BiasEnum == BlockAttentionBiasEnum::ELEMENTWISE_BIAS)
return 1;
else
return 2;
}
else if constexpr(kK0BlockLength <= 256)
else if constexpr(kQKHeaddim <= 256)
{
return 1;
}
......@@ -234,7 +236,7 @@ struct BlockFmhaPipelineQSKSVS
// prefetch K tile
index_t i_total_loops = 0;
constexpr index_t k0_loops = kK0BlockLength / kK0;
constexpr index_t k0_loops = kQKHeaddim / kK0;
constexpr index_t k1_loops = kN0 / kK1;
static_assert(2 <= k0_loops);
......
include/ck_tile/ops/fmha/pipeline/block_fmha_pipeline_qx_ks_vs_custom_policy.hpp
View file @ 4d914af3
......@@ -15,6 +15,7 @@
#include "ck_tile/ops/gemm/block/block_gemm_areg_bsmem_creg_v1_custom_policy.hpp"
#include "ck_tile/ops/gemm/block/block_gemm_areg_bsmem_creg_v2_custom_policy.hpp"
#include "ck_tile/ops/gemm/block/block_gemm_areg_bsmem_creg_v2.hpp"
#include "ck_tile/ops/gemm/block/block_gemm_areg_bsmem_creg_one_warp_v1.hpp"
// TODO: remove this
#define K_LDS_LOAD_USE_OFFSET_TRANSFORM 0
......@@ -54,7 +55,7 @@ struct BlockFmhaPipelineQXCustomPolicy</* QLoadOnce = */ true>
constexpr index_t MWarp = config.template at<1>();
constexpr index_t kMPerBlock = Problem::BlockFmhaShape::kM0;
constexpr index_t kKPerBlock = Problem::BlockFmhaShape::kK0BlockLength;
constexpr index_t kKPerBlock = Problem::BlockFmhaShape::kSubQKHeaddim;
constexpr index_t K2 = WG::kK / WG::WarpGemmAttribute::Impl::kABKLane;
constexpr index_t K1 = WG::WarpGemmAttribute::Impl::kABKLane;
......@@ -64,13 +65,28 @@ struct BlockFmhaPipelineQXCustomPolicy</* QLoadOnce = */ true>
constexpr index_t M1 = MWarp;
constexpr index_t M0 = kMPerBlock / (M2 * M1);
return make_static_tile_distribution(
tile_distribution_encoding<sequence<1>,
tuple<sequence<M0, M1, M2>, sequence<K0, K1, K2>>,
tuple<sequence<1>, sequence<2, 1>>,
tuple<sequence<1>, sequence<1, 2>>,
sequence<1, 2, 2>,
sequence<0, 0, 2>>{});
if constexpr(1 < Problem::kNumGemm0Warps)
{
return make_static_tile_distribution(
tile_distribution_encoding<sequence<1>,
tuple<sequence<M0, M1, M2>, sequence<K0, K1, K2>>,
tuple<sequence<1>, sequence<2, 1>>,
tuple<sequence<1>, sequence<1, 2>>,
sequence<1, 2, 2>,
sequence<0, 0, 2>>{});
}
else
{
static_assert(MWarp == 1);
return make_static_tile_distribution(
tile_distribution_encoding<sequence<1>,
tuple<sequence<M0, M1, M2>, sequence<K0, K1, K2>>,
tuple<sequence<2, 1>>,
tuple<sequence<1, 2>>,
sequence<1, 2, 2>,
sequence<0, 0, 2>>{});
}
}
template <typename Problem>
......@@ -80,7 +96,7 @@ struct BlockFmhaPipelineQXCustomPolicy</* QLoadOnce = */ true>
BlockGemmProblem<typename Problem::QDataType,
typename Problem::KDataType,
typename Problem::SaccDataType,
Problem::kBlockSize,
Problem::kNumGemm0Warps * get_warp_size(),
TileGemmShape<sequence<Problem::BlockFmhaShape::kM0,
Problem::BlockFmhaShape::kN0,
Problem::BlockFmhaShape::kK0>,
......@@ -88,22 +104,33 @@ struct BlockFmhaPipelineQXCustomPolicy</* QLoadOnce = */ true>
typename Problem::BlockFmhaShape::Gemm0WarpTile>>;
constexpr auto warp_gemm = []() {
constexpr index_t WarpGemmM = Problem::BlockFmhaShape::Gemm0WarpTile::at(number<0>{});
static_assert(WarpGemmM == 16 || WarpGemmM == 32);
if constexpr(std::is_same_v<typename Problem::QDataType, half_t> &&
std::is_same_v<typename Problem::KDataType, half_t> &&
std::is_same_v<typename Problem::SaccDataType, float>)
{
return WarpGemmMfmaF16F16F32M32N32K16SwizzleBTransposedCDistribution{};
if constexpr(WarpGemmM == 32)
return WarpGemmMfmaF16F16F32M32N32K16SwizzleBTransposedCDistribution{};
else // WarpGemmM == 16
return WarpGemmMfmaF16F16F32M16N16K16TransposedCDistribution{};
}
else if constexpr(std::is_same_v<typename Problem::QDataType, bf16_t> &&
std::is_same_v<typename Problem::KDataType, bf16_t> &&
std::is_same_v<typename Problem::SaccDataType, float>)
{
return WarpGemmMfmaBf16Bf16F32M32N32K16SwizzleBTransposedCDistribution{};
if constexpr(WarpGemmM == 32)
return WarpGemmMfmaBf16Bf16F32M32N32K16SwizzleBTransposedCDistribution{};
else // WarpGemmM == 16
return WarpGemmMfmaBf16Bf16F32M16N16K16TransposedCDistribution{};
}
else if constexpr(std::is_same_v<typename Problem::QDataType, fp8_t> &&
std::is_same_v<typename Problem::KDataType, fp8_t> &&
std::is_same_v<typename Problem::SaccDataType, float>)
{
static_assert(WarpGemmM == 32);
// TODO: hard coded here. Otherwise, it may incorrect result
constexpr index_t swizzle_factor = 4;
return WarpGemmMfmaFp8Fp8F32M32N32K16SwizzleBTransposedCDistribution<
......@@ -118,12 +145,16 @@ struct BlockFmhaPipelineQXCustomPolicy</* QLoadOnce = */ true>
typename Problem::BlockFmhaShape::Gemm0BlockWarps,
decltype(warp_gemm)>;
return BlockGemmARegBSmemCRegV2<GemmProblem, BlockGemmPolicy>{};
if constexpr(1 < Problem::kNumGemm0Warps)
return BlockGemmARegBSmemCRegV2<GemmProblem, BlockGemmPolicy>{};
else
return BlockGemmARegBSmemCRegOneWarpV1<GemmProblem, BlockGemmPolicy>{};
}
};
/// NOTICE: we no-longer use this policy.
template <>
struct BlockFmhaPipelineQXCustomPolicy</* QLoadOnce = */ false>
struct [[deprecated]] BlockFmhaPipelineQXCustomPolicy</* QLoadOnce = */ false>
{
static constexpr bool QLoadOnce = false;
......@@ -292,6 +323,9 @@ struct BlockFmhaPipelineQXKSVSCustomPolicy : BlockFmhaPipelineQXCustomPolicy<QLo
template<> struct
LdsBufferSequence<3, 3, 3, 3> { using type = sequence<1, 2, 0, 1, 2, 0>; };
template<> struct
LdsBufferSequence<3, 3, 3, 4> { using type = sequence<1, 2, 0, 0, 1, 2, 0>; };
template<> struct
LdsBufferSequence<3, 3, 2, 2> { using type = sequence<1, 2, 1, 0>;};
// clang-format on
......@@ -301,12 +335,12 @@ struct BlockFmhaPipelineQXKSVSCustomPolicy : BlockFmhaPipelineQXCustomPolicy<QLo
{
using BlockFmhaShape = remove_cvref_t<typename Problem::BlockFmhaShape>;
constexpr index_t kN0 = BlockFmhaShape::kN0;
constexpr index_t kK0 = BlockFmhaShape::kK0;
constexpr index_t kK1 = BlockFmhaShape::kK1;
constexpr index_t kK0BlockLength = BlockFmhaShape::kK0BlockLength;
constexpr index_t kN0 = BlockFmhaShape::kN0;
constexpr index_t kK0 = BlockFmhaShape::kK0;
constexpr index_t kK1 = BlockFmhaShape::kK1;
constexpr index_t kQKHeaddim = BlockFmhaShape::kQKHeaddim;
constexpr index_t k0_loops = kK0BlockLength / kK0;
constexpr index_t k0_loops = kQKHeaddim / kK0;
constexpr index_t k1_loops = kN0 / kK1;
return typename LdsBufferSequence<NumPrefetchK, NumPrefetchV, k0_loops, k1_loops>::type{};
......@@ -353,12 +387,15 @@ struct BlockFmhaPipelineQXKSVSCustomPolicy : BlockFmhaPipelineQXCustomPolicy<QLo
constexpr index_t kNPerBlock = Problem::BlockFmhaShape::kN1;
constexpr index_t kKPerBlock = Problem::BlockFmhaShape::kK1;
constexpr index_t total_pixels = kNPerBlock * kKPerBlock / kBlockSize;
constexpr index_t kMaxVecLoad =
min(total_pixels, static_cast<index_t>(16 / sizeof(VDataType)));
constexpr index_t kMinVecLoad = 4 / sizeof(VDataType);
// TODO: not correct!
if constexpr(total_pixels > 4)
return 4;
else
return 2;
constexpr index_t kVecLoad = ((total_pixels / kMaxVecLoad) >= kMinVecLoad)
? kMaxVecLoad
: (total_pixels / kMinVecLoad);
return kVecLoad;
}
else
{
......@@ -372,10 +409,8 @@ struct BlockFmhaPipelineQXKSVSCustomPolicy : BlockFmhaPipelineQXCustomPolicy<QLo
using BlockGemm = remove_cvref_t<decltype(QXPolicy::template GetQKBlockGemm<Problem>())>;
constexpr auto config = BlockGemm::Policy::template GetWarpGemmMWarpNWarp<Problem>();
using WG = remove_cvref_t<decltype(config.template at<0>())>;
using CWarpDstr = typename WG::CWarpDstr;
constexpr auto vec =
CWarpDstr{}.get_ys_to_d_descriptor().get_lengths().at(number<CWarpDstr::NDimY - 1>{});
return vec;
return WG::WarpGemmAttribute::Impl::kCM1PerLane;
}
template <typename Problem>
......@@ -384,10 +419,8 @@ struct BlockFmhaPipelineQXKSVSCustomPolicy : BlockFmhaPipelineQXCustomPolicy<QLo
using BlockGemm = remove_cvref_t<decltype(GetKVBlockGemm<Problem>())>;
constexpr auto config = BlockGemm::Policy::template GetWarpGemmMWarpNWarp<Problem>();
using WG = remove_cvref_t<decltype(config.template at<0>())>;
using CWarpDstr = typename WG::CWarpDstr;
constexpr auto vec =
CWarpDstr{}.get_ys_to_d_descriptor().get_lengths().at(number<CWarpDstr::NDimY - 1>{});
return vec;
return WG::WarpGemmAttribute::Impl::kCM1PerLane;
}
template <typename Problem>
......@@ -438,44 +471,12 @@ struct BlockFmhaPipelineQXKSVSCustomPolicy : BlockFmhaPipelineQXCustomPolicy<QLo
return max(SingleKSize, SingleVSize);
}
template <typename Problem, typename BlockGemm>
CK_TILE_HOST_DEVICE static constexpr auto MakeQRegBlockDescriptor()
{
constexpr index_t kMPerBlock = Problem::BlockFmhaShape::kM0;
constexpr index_t kKPerBlock = Problem::BlockFmhaShape::kK0BlockLength;
constexpr auto config = BlockGemm::Policy::template GetWarpGemmMWarpNWarp<Problem>();
using WG = remove_cvref_t<decltype(config.template at<0>())>;
constexpr index_t MWarp = config.template at<1>();
constexpr index_t NWarp = config.template at<2>();
constexpr index_t MIterPerWarp = kMPerBlock / (MWarp * WG::kM);
constexpr index_t KIterPerWarp = kKPerBlock / WG::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 WG::AWarpDstrEncoding{});
constexpr auto q_block_dstr = make_static_tile_distribution(q_block_dstr_encode);
return q_block_dstr;
}
// TODO: this is used for non async copy desc. unify in the future
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto MakeKLdsBlockDescriptor()
{
constexpr index_t kNPerBlock = Problem::BlockFmhaShape::kN0;
constexpr index_t kKPerBlock = Problem::BlockFmhaShape::kK1;
constexpr index_t kKPerBlock = Problem::BlockFmhaShape::kK0;
constexpr index_t kKPack = GetSmemKPackK<Problem>();
constexpr auto k_lds_block_desc_0 = make_naive_tensor_descriptor(
......@@ -875,36 +876,10 @@ struct BlockFmhaPipelineQXKSVSCustomPolicy : BlockFmhaPipelineQXCustomPolicy<QLo
}
}
template <typename Problem, typename BlockGemm>
template <typename BlockGemm>
CK_TILE_HOST_DEVICE static constexpr auto MakeBiasDramTileDistribution()
{
constexpr index_t MPerBlock = Problem::BlockFmhaShape::kM0;
constexpr index_t NPerBlock = Problem::BlockFmhaShape::kN0;
constexpr auto config = BlockGemm::Policy::template GetWarpGemmMWarpNWarp<Problem>();
using WG = remove_cvref_t<decltype(config.template at<0>())>;
constexpr index_t MWarp = config.template at<1>();
constexpr index_t NWarp = config.template at<2>();
constexpr index_t MIterPerWarp = MPerBlock / (MWarp * WG::kM);
constexpr index_t NIterPerWarp = NPerBlock / (NWarp * WG::kN);
// Construct C-Block-HostTensor
constexpr auto c_block_outer_dstr_encoding = tile_distribution_encoding<
sequence<>,
tuple<sequence<MIterPerWarp, MWarp>, sequence<NIterPerWarp, NWarp>>,
tuple<sequence<1, 2>>,
tuple<sequence<1, 1>>,
sequence<1, 2>,
sequence<0, 0>>{};
constexpr auto c_block_dstr_encode = detail::make_embed_tile_distribution_encoding(
c_block_outer_dstr_encoding, typename WG::CWarpDstrEncoding{});
constexpr auto c_block_dstr = make_static_tile_distribution(c_block_dstr_encode);
return c_block_dstr;
return BlockGemm::MakeCBlockTile().get_tile_distribution();
}
template <typename Problem>
......@@ -961,7 +936,7 @@ struct BlockFmhaPipelineQXKSVSCustomPolicy : BlockFmhaPipelineQXCustomPolicy<QLo
BlockGemmProblem<typename Problem::PDataType,
typename Problem::VDataType,
typename Problem::OaccDataType,
Problem::kBlockSize,
Problem::kNumGemm1Warps * get_warp_size(),
TileGemmShape<sequence<Problem::BlockFmhaShape::kM0,
Problem::BlockFmhaShape::kN1,
Problem::BlockFmhaShape::kK1>,
......
include/ck_tile/ops/fmha/pipeline/tile_fmha_shape.hpp
View file @ 4d914af3
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
......@@ -7,6 +7,20 @@
namespace ck_tile {
static CK_TILE_HOST_DEVICE constexpr index_t ceil_to_qualified_tile_length(index_t len)
{
if(len == 96)
return 128;
if(len == 160)
return 256;
// only length of 96, 160 and power-of-two is supported
if(!(len & (len - 1)))
return len;
return 0;
};
template <typename BlockTile_, // sequence<...
typename Gemm0BlockWarps_,
typename Gemm0WarpTile_,
......@@ -21,20 +35,27 @@ struct TileFmhaShape
using Gemm1BlockWarps = remove_cvref_t<Gemm1BlockWarps_>;
using Gemm1WarpTile = remove_cvref_t<Gemm1WarpTile_>;
static constexpr index_t NumWarps =
static constexpr index_t NumGemm0Warps =
reduce_on_sequence(Gemm0BlockWarps{}, multiplies{}, number<1>{});
static constexpr index_t NumGemm1Warps =
reduce_on_sequence(Gemm1BlockWarps{}, multiplies{}, number<1>{});
static_assert(NumGemm1Warps % NumGemm0Warps == 0);
static constexpr index_t NumWarps = max(NumGemm0Warps, NumGemm1Warps);
static_assert(NumWarps == reduce_on_sequence(Gemm1BlockWarps{}, multiplies{}, number<1>{}));
static_assert(std::is_same_v<Gemm0WarpTile, Gemm1WarpTile>);
static constexpr index_t kM0 = BlockTile::at(number<0>{}); // tile size along q seqlen
static constexpr index_t kN0 = BlockTile::at(number<1>{}); // tile size along k seqlen
static constexpr index_t kK0 = BlockTile::at(number<2>{}); // tile size along qk gemm unroll
static constexpr index_t kN1 = BlockTile::at(number<3>{}); // tile size along v head_dim
static constexpr index_t kK1 = BlockTile::at(number<4>{}); // tile size along kv gemm unroll
static constexpr index_t kK0BlockLength =
static constexpr index_t kQKHeaddim =
BlockTile::at(number<5>{}); // total length of K0, used for pipeline that need load Q at
// once (or repeately load Q as a whole tile)
static_assert(kK0BlockLength % kK0 == 0, "kK0BlockLength should be divisible by kK0");
static_assert(kQKHeaddim % kK0 == 0, "kQKHeaddim should be divisible by kK0");
static constexpr index_t kSubQKHeaddim = ceil_to_qualified_tile_length(kQKHeaddim);
// v, rowmajor : seqlen*hdim, colmajor : hdim*seqlen
static constexpr bool IsVLayoutRowMajor = IsVLayoutRowMajor_;
......
include/ck_tile/ops/gemm.hpp
View file @ 4d914af3
......@@ -8,6 +8,7 @@
#include "ck_tile/ops/gemm/block/block_gemm_areg_breg_creg_v1.hpp"
#include "ck_tile/ops/gemm/block/block_gemm_areg_breg_creg_v1_custom_policy.hpp"
#include "ck_tile/ops/gemm/block/block_gemm_areg_breg_creg_v1_default_policy.hpp"
#include "ck_tile/ops/gemm/block/block_gemm_areg_bsmem_creg_one_warp_v1.hpp"
#include "ck_tile/ops/gemm/block/block_gemm_areg_bsmem_creg_v1.hpp"
#include "ck_tile/ops/gemm/block/block_gemm_areg_bsmem_creg_v1_custom_policy.hpp"
#include "ck_tile/ops/gemm/block/block_gemm_areg_bsmem_creg_v1_default_policy.hpp"
......@@ -23,12 +24,14 @@
#include "ck_tile/ops/gemm/block/block_gemm_problem.hpp"
#include "ck_tile/ops/gemm/kernel/gemm_kernel.hpp"
#include "ck_tile/ops/gemm/kernel/gemm_tile_partitioner.hpp"
#include "ck_tile/ops/gemm/pipeline/gemm_universal_pipeline_ag_bg_cr_policy.hpp"
#include "ck_tile/ops/gemm/pipeline/gemm_pipeline_ag_bg_cr_mem.hpp"
#include "ck_tile/ops/gemm/pipeline/gemm_pipeline_ag_bg_cr_scheduler.hpp"
#include "ck_tile/ops/gemm/pipeline/gemm_pipeline_agmem_bgmem_creg_v1.hpp"
#include "ck_tile/ops/gemm/pipeline/gemm_pipeline_agmem_bgmem_creg_v1_default_policy.hpp"
#include "ck_tile/ops/gemm/pipeline/gemm_pipeline_agmem_bgmem_creg_v2.hpp"
#include "ck_tile/ops/gemm/pipeline/gemm_pipeline_agmem_bgmem_creg_v2_default_policy.hpp"
#include "ck_tile/ops/gemm/pipeline/gemm_pipeline_problem.hpp"
#include "ck_tile/ops/gemm/pipeline/gemm_universal_pipeline_ag_bg_cr_policy.hpp"
#include "ck_tile/ops/gemm/pipeline/tile_gemm_shape.hpp"
#include "ck_tile/ops/gemm/pipeline/tile_gemm_traits.hpp"
#include "ck_tile/ops/gemm/warp/warp_gemm.hpp"
......@@ -36,4 +39,5 @@
#include "ck_tile/ops/gemm/warp/warp_gemm_attribute_mfma_impl.hpp"
#include "ck_tile/ops/gemm/warp/warp_gemm_dispatcher.hpp"
#include "ck_tile/ops/gemm/warp/warp_gemm_impl.hpp"
#include "ck_tile/ops/common/generic_2d_block_shape.hpp"
#include "ck_tile/ops/common/tensor_layout.hpp"
include/ck_tile/ops/gemm/block/block_gemm_areg_bgmem_creg_v1.hpp
View file @ 4d914af3
......@@ -32,7 +32,7 @@ struct BlockGemmARegBGmemCRegV1
BlockGemmProblem<ADataType, BDataType, CDataType, kBlockSize, BlockGemmShape>,
BlockGemmARegBGmemCRegV1DefaultPolicy>;
CK_TILE_HOST_DEVICE static constexpr ck_tile::index_t GetStaticLdsSize()
CK_TILE_HOST_DEVICE static constexpr index_t GetStaticLdsSize()
{
return sizeof(BDataType) *
Policy::template MakeBSmemBlockDescriptor<Problem>().get_element_space_size();
......
include/ck_tile/ops/gemm/block/block_gemm_areg_breg_creg_v1.hpp
View file @ 4d914af3
......@@ -157,7 +157,7 @@ struct BlockGemmARegBRegCRegV1
});
}
CK_TILE_DEVICE constexpr auto MakeCBlockTile() const
CK_TILE_DEVICE static constexpr auto MakeCBlockTile()
{
constexpr index_t MPerBlock = BlockGemmShape::kM;
constexpr index_t NPerBlock = BlockGemmShape::kN;
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment