ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-06-06 07:51:52 +00:00
Code Issues Packages Projects Releases Wiki Activity

Add SplitKV kernel codegen logics

Browse Source
This commit is contained in:
PoYen, Chen
2024-06-02 17:41:00 +00:00
parent 78b64d11c4
commit cacce74f2c
13 changed files with 2598 additions and 25 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
example/ck_tile/01_fmha/fmha_fwd.cpp
View File

@@ -597,6 +597,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
shape_seqlen_q,
shape_seqlen_k,
batch,
nhead,
max_seqlen_q,
hdim_q,
hdim_v,

33
example/ck_tile/01_fmha/fmha_fwd.hpp
View File

@@ -104,6 +104,7 @@ struct fmha_fwd_args
ck_tile::index_t seqlen_q;
ck_tile::index_t seqlen_k;
ck_tile::index_t batch;
ck_tile::index_t nhead;
ck_tile::index_t max_seqlen_q;
ck_tile::index_t hdim_q;
ck_tile::index_t hdim_v;
@@ -238,7 +239,6 @@ auto fmha_fwd_create_kargs_and_grids(fmha_fwd_args args)
return ck_tile::make_tuple(kargs, grids);
}
#if 0
template <typename FmhaFwdSplitKVKernel>
auto fmha_fwd_splitkv_create_kargs_and_grids(fmha_fwd_args args)
{
@@ -251,6 +251,7 @@ auto fmha_fwd_splitkv_create_kargs_and_grids(fmha_fwd_args args)
args.k_ptr,
args.v_ptr,
args.bias_ptr,
args.rand_val_ptr,
args.lse_acc_ptr,
args.o_acc_ptr,
args.batch,
@@ -261,6 +262,7 @@ auto fmha_fwd_splitkv_create_kargs_and_grids(fmha_fwd_args args)
args.seqlen_k_ptr,
args.hdim_q,
args.hdim_v,
args.nhead_q,
args.nhead_q / args.nhead_k,
args.num_splits,
args.scale_s,
@@ -269,13 +271,18 @@ auto fmha_fwd_splitkv_create_kargs_and_grids(fmha_fwd_args args)
args.stride_k,
args.stride_v,
args.stride_bias,
args.stride_randval,
args.nhead_stride_q,
args.nhead_stride_k,
args.nhead_stride_v,
args.nhead_stride_bias,
args.nhead_stride_randval,
args.window_size_left,
args.window_size_right,
args.mask_type);
args.mask_type,
args.p_drop,
args.s_randval,
args.drop_seed_offset);
}
else
{ // create batch mode kernel arguments
@@ -283,6 +290,7 @@ auto fmha_fwd_splitkv_create_kargs_and_grids(fmha_fwd_args args)
args.k_ptr,
args.v_ptr,
args.bias_ptr,
args.rand_val_ptr,
args.lse_acc_ptr,
args.o_acc_ptr,
args.batch,
@@ -292,6 +300,7 @@ auto fmha_fwd_splitkv_create_kargs_and_grids(fmha_fwd_args args)
args.seqlen_k,
args.hdim_q,
args.hdim_v,
args.nhead_q,
args.nhead_q / args.nhead_k,
args.num_splits,
args.scale_s,
@@ -300,17 +309,23 @@ auto fmha_fwd_splitkv_create_kargs_and_grids(fmha_fwd_args args)
args.stride_k,
args.stride_v,
args.stride_bias,
args.stride_randval,
args.nhead_stride_q,
args.nhead_stride_k,
args.nhead_stride_v,
args.nhead_stride_bias,
args.nhead_stride_randval,
args.batch_stride_q,
args.batch_stride_k,
args.batch_stride_v,
args.batch_stride_bias,
args.batch_stride_randval,
args.window_size_left,
args.window_size_right,
args.mask_type);
args.mask_type,
args.p_drop,
args.s_randval,
args.drop_seed_offset);
}
}();
@@ -319,6 +334,7 @@ auto fmha_fwd_splitkv_create_kargs_and_grids(fmha_fwd_args args)
return ck_tile::make_tuple(kargs, grids);
}
#if 0
template <typename FmhaFwdSplitKVCombineKernel>
auto fmha_fwd_splitkv_combine_create_kargs_and_grids(fmha_fwd_args args)
{
@@ -430,7 +446,10 @@ template <typename Traits_>
float fmha_fwd_(const ck_tile::stream_config&, fmha_fwd_args);
template <typename Traits_>
float fmha_fwd_splitkv_(const ck_tile::stream_config&, fmha_fwd_args);
float fmha_fwd_splitkv_oneshot_(const ck_tile::stream_config&, fmha_fwd_args);
template <typename Traits_>
std::string fmha_fwd_splitkv_get_name_();
// This is the public API, will be generated by script
struct fmha_fwd_traits
@@ -448,8 +467,4 @@ struct fmha_fwd_traits
// TODO: padding check is inside this api
};
float fmha_fwd(fmha_fwd_traits, fmha_fwd_args, const ck_tile::stream_config&);
inline float fmha_fwd_splitkv(fmha_fwd_traits, fmha_fwd_args, const ck_tile::stream_config&)
{
std::cout << __PRETTY_FUNCTION__ << std::endl;
return 0;
}
float fmha_fwd_splitkv(fmha_fwd_traits, fmha_fwd_args, const ck_tile::stream_config&);

142
example/ck_tile/01_fmha/generate.py
View File

@@ -165,6 +165,90 @@ float fmha_fwd_<trait_{F_idx}>(const ck_tile::stream_config& s, fmha_fwd_args a)
}}
"""
FMHA_FWD_SPLITKV_PIPELINE_MAP = {
"qr" : "ck_tile::BlockFmhaFwdSplitKVPipelineQRKSVS",
"qr_async" : "ck_tile::BlockFmhaFwdSplitKVPipelineQRKSVSAsync",
}
FMHA_FWD_SPLITKV_KERNEL_BODY="""
using fmha_dtype_{F_idx} = {F_dtype};
using fmha_block_tile_{F_idx} = ck_tile::sequence<{F_bm0}, {F_bn0}, {F_bk0}, {F_bn1}, {F_bk1}, {F_bk0blen}>;
using fmha_block_warps_{F_idx} = ck_tile::sequence<{F_rm}, {F_rn}, {F_rk}>;
using fmha_warp_tile_{F_idx} = ck_tile::sequence<{F_wm}, {F_wn}, {F_wk}>;
using fmha_shape_{F_idx} = ck_tile::TileFmhaShape<fmha_block_tile_{F_idx},
fmha_block_warps_{F_idx},
fmha_warp_tile_{F_idx},
fmha_block_warps_{F_idx},
fmha_warp_tile_{F_idx},
{F_vlayout}>;
using fmha_trait_{F_idx} = ck_tile::TileFmhaTraits<{F_spad},
{F_skpad},
{F_dpad},
{F_dvpad},
{F_bias},
false,
{F_lse},
{F_dropout},
{F_squant},
{F_occupancy}>;
using fmha_mask_{F_idx} = {F_mask};
using fmha_pipeline_problem_{F_idx} = ck_tile::BlockFmhaPipelineProblem<
typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::QDataType,
typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::KDataType,
typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::VDataType,
typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::SaccDataType,
typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::SMPLComputeDataType,
typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::BiasDataType,
typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::RandValOutputDataType,
typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::LSEDataType,
typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::PDataType,
typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::OaccDataType,
typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::OaccDataType,
fmha_shape_{F_idx},
{F_mode},
fmha_mask_{F_idx},
fmha_trait_{F_idx}>;
using fmha_pipeline_{F_idx} = {F_pipeline}<
fmha_pipeline_problem_{F_idx}>;
using fmha_epilogue_{F_idx} =
ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<{F_dtype}>::OaccDataType,
typename FmhaFwdTypeConfig<{F_dtype}>::OaccDataType,
{F_spad}, {F_dvpad}>>;
using fmha_kernel_{F_idx} =
ck_tile::FmhaFwdSplitKVKernel<ck_tile::FmhaFwdSplitKVTilePartitioner<fmha_shape_{F_idx}>,
fmha_pipeline_{F_idx},
fmha_epilogue_{F_idx}>;
using trait_{F_idx} = fmha_fwd_traits_<{F_hdim}, {F_dtype}, {F_mode},{F_bm0}, {F_bn0}, {F_bk0}, {F_bn1}, {F_bk1}, {F_bk0blen}, {F_vlayout},
{F_pipeline_enum}, fmha_mask_{F_idx}, {F_bias}, {F_lse}, {F_dropout}, {F_squant}, {F_spad}, {F_skpad}, {F_dpad}, {F_dvpad}>;
#include <iostream>
template<>
float fmha_fwd_splitkv_oneshot_<trait_{F_idx}>(const ck_tile::stream_config& s, fmha_fwd_args a)
{{
using k_ = fmha_kernel_{F_idx};
auto [kargs, grids] = fmha_fwd_splitkv_create_kargs_and_grids<k_>(a);
constexpr dim3 blocks = k_::BlockSize();
constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
return ck_tile::launch_kernel(s, ck_tile::make_kernel<blocks.x, kBlockPerCu>(k_{{}}, grids, blocks, 0, kargs));
}}
template<>
std::string fmha_fwd_splitkv_get_name_<trait_{F_idx}>()
{{
using k_ = fmha_kernel_{F_idx};
return k_::GetName();
}}
"""
FMHA_FWD_API_FILENAME="fmha_fwd_api.cpp"
FMHA_FWD_API="""
float fmha_fwd(fmha_fwd_traits t, fmha_fwd_args a, const ck_tile::stream_config& s){{
@@ -196,11 +280,43 @@ MASK_SIMPLIFIED_CHECK_MAP = {
FMHA_FWD_API_INNER_DISPATCH=""" {F_if}((t.is_group_mode == {F_mode}) && (t.is_v_rowmajor == {F_vlayout}) && ({F_mask_check}) && (t.bias_type == {F_bias_check}) && (t.has_lse == {F_lse}) && (t.has_dropout == {F_dropout}) && (t.do_fp8_static_quant == {F_squant}) &&
({F_scheck}) && ({F_skcheck}) && ({F_dcheck}) && ({F_dvcheck})) {{
using trait_ = fmha_fwd_traits_<{F_hdim}, {F_dtype}, {F_mode}, {F_bm0}, {F_bn0}, {F_bk0}, {F_bn1}, {F_bk1}, {F_bk0blen}, {F_vlayout}, {F_pipeline_enum}, {F_mask}, {F_bias}, {F_lse}, {F_dropout}, {F_squant}, {F_spad}, {F_skpad}, {F_dpad}, {F_dvpad}>;
return {F_callee}<trait_>(s, a);
return fmha_fwd_<trait_>(s, a);
}}
"""
FMHA_FWD_SPLITKV_API_FILENAME="fmha_fwd_splitkv_api.cpp"
FMHA_FWD_SPLITKV_API="""
#include <iostream>
template<typename fmha_fwd_splitkv_trait_
// , typename fmha_fwd_splitkv_combine_trait_
>
float fmha_fwd_splitkv_(const ck_tile::stream_config& s, fmha_fwd_args a)
{{
if(s.log_level_ > 0)
std::cout << ", " << fmha_fwd_splitkv_get_name_<fmha_fwd_splitkv_trait_>()
// << ", " << fmha_fwd_splitkv_combine_get_name_<fmha_fwd_splitkv_combine_trait_>()
<< std::flush;
return ck_tile::launch_kernel(s,
[=](const ck_tile::stream_config& s_){{ fmha_fwd_splitkv_oneshot_<fmha_fwd_splitkv_trait_>(s_, a); }}
// , [=](const ck_tile::stream_config& s_){{ fmha_fwd_splitkv_combine_oneshot_<dq_dk_dv_trait_>(s_, a); }}
);
}}
float fmha_fwd_splitkv(fmha_fwd_traits t, fmha_fwd_args a, const ck_tile::stream_config& s){{
float r = -1;
{F_dispatch}
return r;
}}
"""
FMHA_FWD_SPLITKV_API_INNER_DISPATCH=""" {F_if}((t.is_group_mode == {F_mode}) && (t.is_v_rowmajor == {F_vlayout}) && ({F_mask_check}) && (t.bias_type == {F_bias_check}) && (t.has_lse == {F_lse}) && (t.has_dropout == {F_dropout}) && (t.do_fp8_static_quant == {F_squant}) &&
({F_scheck}) && ({F_skcheck}) && ({F_dcheck}) && ({F_dvcheck})) {{
using trait_ = fmha_fwd_traits_<{F_hdim}, {F_dtype}, {F_mode}, {F_bm0}, {F_bn0}, {F_bk0}, {F_bn1}, {F_bk1}, {F_bk0blen}, {F_vlayout}, {F_pipeline_enum}, {F_mask}, {F_bias}, {F_lse}, {F_dropout}, {F_squant}, {F_spad}, {F_skpad}, {F_dpad}, {F_dvpad}>;
return fmha_fwd_splitkv_<trait_>(s, a);
}}
"""
def get_mask_map(mask : str):
if mask == "generic":
@@ -386,7 +502,9 @@ class FmhaFwdApiPool:
@property
def api(self) -> str:
callee = "fmha_fwd_splitkv_" if self.is_splitkv else "fmha_fwd_"
inner_dispatch = FMHA_FWD_SPLITKV_API_INNER_DISPATCH if self.is_splitkv else FMHA_FWD_API_INNER_DISPATCH
fwd_api = FMHA_FWD_SPLITKV_API if self.is_splitkv else FMHA_FWD_API
per_dtypes=str()
for i, dtype in enumerate(self.pool.keys()):
per_hdim_case=str()
@@ -395,20 +513,19 @@ class FmhaFwdApiPool:
inners=str()
for k, trait in enumerate(traits):
if_k = 'if' if k == 0 else 'else if'
inners = inners + FMHA_FWD_API_INNER_DISPATCH.format(F_if=if_k, F_mode=MODE_MAP[trait.mode], F_vlayout=LAYOUT_MAP[trait.vlayout],
inners = inners + inner_dispatch.format(F_if=if_k, F_mode=MODE_MAP[trait.mode], F_vlayout=LAYOUT_MAP[trait.vlayout],
F_pipeline_enum=PIPELINE_ENUM_MAP[trait.pipeline_tag], F_mask=get_mask_map(self.mask_impl)[trait.mask],
F_mask_check=get_mask_check_map(self.mask_impl)[trait.mask], F_bias_check=BIAS_CHECK_MAP[trait.bias], F_bias=BIAS_MAP[trait.bias],
F_lse=BOOL_MAP[trait.lse], F_dropout=BOOL_MAP[trait.dropout] ,
F_squant=BOOL_MAP[trait.squant], F_scheck=trait.scheck, F_skcheck=trait.skcheck, F_dcheck=trait.dcheck, F_dvcheck=trait.dvcheck,
F_spad=BOOL_MAP[trait.spad], F_skpad=BOOL_MAP[trait.skpad], F_dpad=BOOL_MAP[trait.dpad], F_dvpad=BOOL_MAP[trait.dvpad],
F_bm0=trait.bm0, F_bn0=trait.bn0, F_bk0=trait.bk0, F_bn1=trait.bn1, F_bk1=trait.bk1, F_bk0blen=trait.bk0blen,
F_hdim=hdim, F_dtype=DTYPE_MAP[dtype],
F_callee=callee)
F_hdim=hdim, F_dtype=DTYPE_MAP[dtype])
if_j = 'if' if j == 0 else 'else if'
per_hdim_case = per_hdim_case + FMHA_FWD_API_PER_HDIM_CASE.format(F_if=if_j, F_hdim=hdim, F_inner_dispatch=inners)
if_i = 'if' if i == 0 else 'else if'
per_dtypes = per_dtypes + FMHA_FWD_API_PER_DTYPE.format(F_if=if_i, F_dtype=dtype, F_hdim_case=per_hdim_case)
return FMHA_FWD_KERNEL_HEADER + FMHA_FWD_API.format(F_dispatch = per_dtypes)
return FMHA_FWD_KERNEL_HEADER + fwd_api.format(F_dispatch = per_dtypes)
@dataclass
class FmhaFwdTileSize:
@@ -528,17 +645,11 @@ class FmhaFwdSplitKVKernel:
F_pipeline : FmhaFwdPipeline
mask_impl : str
def get_tp(self) -> str:
if self.F_mode == 'group':
return 'hbs'
else:
return 'shb'
@property
def template(self) -> str:
kernel_body = str()
return FMHA_FWD_KERNEL_HEADER + \
FMHA_FWD_KERNEL_BODY.format(
FMHA_FWD_SPLITKV_KERNEL_BODY.format(
F_idx = self.F_idx,
F_hdim = self.F_hdim,
F_dtype = DTYPE_MAP[self.F_dtype],
@@ -567,13 +678,12 @@ class FmhaFwdSplitKVKernel:
F_pipeline_enum = PIPELINE_ENUM_MAP[self.F_pipeline.tag],
F_mask = get_mask_map(self.mask_impl)[self.F_pipeline.F_mask],
F_mode = MODE_MAP[self.F_mode],
F_pipeline = PIPELINE_MAP[self.F_pipeline.tag],
F_tile_partitioner = TILE_PARTITIONER_MAP[self.get_tp()])
F_pipeline = FMHA_FWD_SPLITKV_PIPELINE_MAP[self.F_pipeline.tag])
@property
def name(self) -> str:
# TODO: we don't encode idx here
return f"fmha_{self.direction}_splitkv_d{self.F_hdim}_{self.F_dtype}_{self.F_mode}_{self.get_tp()}_" + \
return f"fmha_{self.direction}_splitkv_d{self.F_hdim}_{self.F_dtype}_{self.F_mode}_" + \
self.F_tile.name + '_' + self.F_pipeline.name
@property

9
include/ck_tile/ops/fmha.hpp
View File

@@ -10,6 +10,8 @@
#include "ck_tile/ops/fmha/kernel/fmha_bwd_kernel.hpp"
#include "ck_tile/ops/fmha/kernel/fmha_bwd_tile_partitioner.hpp"
#include "ck_tile/ops/fmha/kernel/fmha_fwd_kernel.hpp"
#include "ck_tile/ops/fmha/kernel/fmha_fwd_splitkv_kernel.hpp"
#include "ck_tile/ops/fmha/kernel/fmha_fwd_splitkv_tile_partitioner.hpp"
#include "ck_tile/ops/fmha/kernel/fmha_fwd_tile_partitioner.hpp"
#include "ck_tile/ops/fmha/pipeline/block_fmha_bwd_dot_do_o.hpp"
#include "ck_tile/ops/fmha/pipeline/block_fmha_bwd_dot_do_o_default_policy.hpp"
@@ -22,6 +24,13 @@
#include "ck_tile/ops/fmha/pipeline/block_fmha_bwd_pipeline_default_policy.hpp"
#include "ck_tile/ops/fmha/pipeline/block_fmha_bwd_pipeline_enum.hpp"
#include "ck_tile/ops/fmha/pipeline/block_fmha_bwd_pipeline_problem.hpp"
#include "ck_tile/ops/fmha/pipeline/block_fmha_fwd_splitkv_combine_pipeline.hpp"
#include "ck_tile/ops/fmha/pipeline/block_fmha_fwd_splitkv_combine_pipeline_default_policy.hpp"
#include "ck_tile/ops/fmha/pipeline/block_fmha_fwd_splitkv_pipeline_default_policy.hpp"
#include "ck_tile/ops/fmha/pipeline/block_fmha_fwd_splitkv_pipeline_qr_ks_vs.hpp"
#include "ck_tile/ops/fmha/pipeline/block_fmha_fwd_splitkv_pipeline_qr_ks_vs_async.hpp"
#include "ck_tile/ops/fmha/pipeline/block_fmha_fwd_splitkv_pipeline_qr_ks_vs_async_default_policy.hpp"
#include "ck_tile/ops/fmha/pipeline/block_fmha_fwd_splitkv_pipeline_qr_ks_vs_default_policy.hpp"
#include "ck_tile/ops/fmha/pipeline/block_fmha_pipeline_enum.hpp"
#include "ck_tile/ops/fmha/pipeline/block_fmha_pipeline_problem.hpp"
#include "ck_tile/ops/fmha/pipeline/block_fmha_pipeline_qr_ks_vs.hpp"

17
include/ck_tile/ops/fmha/block/block_masking.hpp
View File

@@ -299,6 +299,23 @@ struct SimplifiedGenericAttentionMask
}
}
template <index_t TileHeight, index_t TileWidth>
CK_TILE_HOST_DEVICE constexpr auto GetTileRangeAlongX(index_t i_y,
number<TileHeight> height,
number<TileWidth> width,
index_t i_split,
index_t num_splits) const
{
auto [origin_start, origin_end] = GetTileRangeAlongX(i_y, height, width);
const index_t x_per_split = x_total / num_splits;
index_t split_start = x_per_split * i_split;
index_t split_end = (i_split == num_splits - 1 ? x_total : split_start + x_per_split);
return ck_tile::make_tuple(ck_tile::max(origin_start, split_start),
ck_tile::min(origin_end, split_end));
}
// to get the loop length along Y axis, return index:[start, end), end-start=length
// use this if need loop over Y axis tile by tile (like q-seqlen loopover)
// TODO: y_end still could be negative, so end-start could be negative(need check)

887
include/ck_tile/ops/fmha/kernel/fmha_fwd_splitkv_kernel.hpp Normal file
View File

@@ -0,0 +1,887 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck_tile/core.hpp"
#include "ck_tile/ops/common.hpp"
#include "ck_tile/ops/fmha/block/block_attention_bias_enum.hpp"
#include <string>
#include <type_traits>
// S[seqlen_q, seqlen_k] = Q[seqlen_q, hdim_q] @ K[seqlen_k, hdim_q]
// S'[seqlen_q, seqlen_k] = S[seqlen_q, seqlen_k] * Scale[1]
// S''[seqlen_q, seqlen_k] = S'[seqlen_q, seqlen_k] + Bias[seqlen_q, seqlen_k]
// P[seqlen_q, seqlen_k] = Softmax(S''[seqlen_q, seqlen_k])
// O[seqlen_q, hdim_v] = P[seqlen_q, seqlen_k] @ V^T[hdim_v, seqlen_k]
namespace ck_tile {
template <typename TilePartitioner_, typename FmhaPipeline_, typename EpiloguePipeline_>
struct FmhaFwdSplitKVKernel
{
using TilePartitioner = ck_tile::remove_cvref_t<TilePartitioner_>;
using FmhaPipeline = ck_tile::remove_cvref_t<FmhaPipeline_>;
using EpiloguePipeline = ck_tile::remove_cvref_t<EpiloguePipeline_>;
static constexpr ck_tile::index_t kBlockSize = FmhaPipeline::kBlockSize;
static constexpr ck_tile::index_t kBlockPerCu = FmhaPipeline::kBlockPerCu;
static_assert(kBlockPerCu > 0);
static constexpr ck_tile::index_t kBlockPerCuInput = FmhaPipeline::Problem::kBlockPerCu;
using QDataType = ck_tile::remove_cvref_t<typename FmhaPipeline::QDataType>;
using KDataType = ck_tile::remove_cvref_t<typename FmhaPipeline::KDataType>;
using VDataType = ck_tile::remove_cvref_t<typename FmhaPipeline::VDataType>;
using BiasDataType = ck_tile::remove_cvref_t<typename FmhaPipeline::BiasDataType>;
using RandValOutputDataType =
ck_tile::remove_cvref_t<typename FmhaPipeline::RandValOutputDataType>;
using LSEDataType = ck_tile::remove_cvref_t<typename FmhaPipeline::LSEDataType>;
using SaccDataType = ck_tile::remove_cvref_t<typename FmhaPipeline::SaccDataType>;
using OaccDataType = remove_cvref_t<typename FmhaPipeline::OaccDataType>;
using VLayout = ck_tile::remove_cvref_t<typename FmhaPipeline::VLayout>;
static constexpr bool kIsGroupMode = FmhaPipeline::kIsGroupMode;
static constexpr bool kPadSeqLenQ = FmhaPipeline::kPadSeqLenQ;
static constexpr bool kPadSeqLenK = FmhaPipeline::kPadSeqLenK;
static constexpr bool kPadHeadDimQ = FmhaPipeline::kPadHeadDimQ;
static constexpr bool kPadHeadDimV = FmhaPipeline::kPadHeadDimV;
static constexpr auto BiasEnum = FmhaPipeline::BiasEnum;
static constexpr bool kHasDropout = FmhaPipeline::kHasDropout;
static constexpr bool kDoFp8StaticQuant = FmhaPipeline::Problem::kDoFp8StaticQuant;
using FmhaMask = ck_tile::remove_cvref_t<typename FmhaPipeline::FmhaMask>;
static constexpr bool kHasMask = FmhaMask::IsMasking;
// clang-format off
template <typename T> struct t2s;
template <> struct t2s<float> { static constexpr const char * name = "fp32"; };
template <> struct t2s<ck_tile::fp16_t> { static constexpr const char * name = "fp16"; };
template <> struct t2s<ck_tile::bf16_t> { static constexpr const char * name = "bf16"; };
template <> struct t2s<ck_tile::fp8_t> { static constexpr const char * name = "fp8"; };
template <> struct t2s<ck_tile::bf8_t> { static constexpr const char * name = "bf8"; };
// clang-format on
__host__ static std::string GetName()
{
// sync with generate.py
// clang-format off
using bfs = typename FmhaPipeline::BlockFmhaShape;
using gbr = typename bfs::Gemm0BlockWarps;
using gwt = typename bfs::Gemm0WarpTile;
#define _SS_ std::string
#define _TS_ std::to_string
auto pn = [&] () {
std::string n;
if (kPadSeqLenQ) n += "s";
if (kPadSeqLenK) n += "sk";
if (kPadHeadDimQ) n += "d";
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) +
"_" + (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>{})) + "_" +
"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) +
(BiasEnum == BlockAttentionBiasEnum::NO_BIAS ? _SS_("") : (_SS_("_") + BlockAttentionBiasEnumToStr<BiasEnum>::name)) +
(kHasMask ? "_" + _SS_(FmhaMask::name) : "") + (kHasDropout ? "_dropout" : "" ) + (kDoFp8StaticQuant ? "_squant" : "" );
#undef _SS_
#undef _TS_
// clang-format on
}
template <ck_tile::index_t I> // to avoid duplicated base class prblem, introduce an template
// arg
struct EmptyKargs
{
};
// kargs use aggregate initializer, so no constructor will provided
// use inheritance to minimize karg size
// user need to use MakeKargs() function to create kargs.
struct CommonKargs
{
const void* q_ptr;
const void* k_ptr;
const void* v_ptr;
void* lse_acc_ptr;
void* o_acc_ptr;
ck_tile::index_t batch;
ck_tile::index_t nhead;
ck_tile::index_t max_seqlen_q;
ck_tile::index_t seqlen_q;
ck_tile::index_t seqlen_k;
ck_tile::index_t hdim_q;
ck_tile::index_t hdim_v;
ck_tile::index_t num_head_q;
// for MQA/GQA, nhead could be different. This parameter is nhead_q / nhead_k
// if this param is larger than 1, indicate MQA/GQA case
ck_tile::index_t nhead_ratio_qk;
ck_tile::index_t num_splits;
float scale_s;
ck_tile::index_t stride_q;
ck_tile::index_t stride_k;
ck_tile::index_t stride_v;
ck_tile::index_t nhead_stride_q;
ck_tile::index_t nhead_stride_k;
ck_tile::index_t nhead_stride_v;
};
struct CommonBiasKargs
{
const void* bias_ptr = nullptr;
ck_tile::index_t stride_bias = 0;
ck_tile::index_t nhead_stride_bias = 0;
};
struct BatchModeBiasKargs : CommonBiasKargs
{
ck_tile::index_t batch_stride_bias = 0;
};
struct AlibiKargs
{
// alibi is batch*nhead*1, no matter in batch/group mode, they are the same
const void* alibi_slope_ptr;
ck_tile::index_t alibi_slope_stride; // stride in batch, or 0 for all batch share same slope
};
struct MaskKargs
{
// ck_tile::index_t window_size_left, window_size_right;
ck_tile::index_t window_size_left, window_size_right;
ck_tile::GenericAttentionMaskEnum mask_type;
};
struct Fp8StaticQuantKargs
{
float scale_p;
};
struct CommonDropoutKargs
{
void init_dropout(const float p_drop,
const std::tuple<uint64_t, uint64_t>& drop_seed_offset)
{
float p_undrop = 1.0 - p_drop;
p_undrop_in_uint8_t =
uint8_t(std::floor(p_undrop * std::numeric_limits<uint8_t>::max()));
rp_undrop = 1.0 / p_undrop;
drop_seed = std::get<0>(drop_seed_offset);
drop_offset = std::get<1>(drop_seed_offset);
}
float rp_undrop = 1;
uint8_t p_undrop_in_uint8_t = std::numeric_limits<uint8_t>::max();
bool is_store_randval = false;
uint64_t drop_seed = 1;
uint64_t drop_offset = 0;
void* rand_val_ptr = nullptr;
ck_tile::index_t stride_randval = 0;
ck_tile::index_t nhead_stride_randval = 0;
};
struct BatchModeDropoutKargs : CommonDropoutKargs
{
ck_tile::index_t batch_stride_randval = 0;
};
struct BatchModeKargs
: CommonKargs,
std::conditional_t<BiasEnum == BlockAttentionBiasEnum::ELEMENTWISE_BIAS,
BatchModeBiasKargs,
std::conditional_t<BiasEnum == BlockAttentionBiasEnum::ALIBI,
AlibiKargs,
EmptyKargs<0>>>,
std::conditional_t<kHasMask, MaskKargs, EmptyKargs<1>>,
std::conditional_t<kDoFp8StaticQuant, Fp8StaticQuantKargs, EmptyKargs<2>>,
std::conditional_t<kHasDropout, BatchModeDropoutKargs, EmptyKargs<3>>
{
ck_tile::index_t batch_stride_q;
ck_tile::index_t batch_stride_k;
ck_tile::index_t batch_stride_v;
};
struct GroupModeKargs
: CommonKargs,
std::conditional_t<BiasEnum == BlockAttentionBiasEnum::ELEMENTWISE_BIAS,
CommonBiasKargs,
std::conditional_t<BiasEnum == BlockAttentionBiasEnum::ALIBI,
AlibiKargs,
EmptyKargs<0>>>,
std::conditional_t<kHasMask, MaskKargs, EmptyKargs<1>>,
std::conditional_t<kDoFp8StaticQuant, Fp8StaticQuantKargs, EmptyKargs<2>>,
std::conditional_t<kHasDropout, CommonDropoutKargs, EmptyKargs<3>>
{
const int32_t* seqstart_q_ptr;
const int32_t* seqstart_k_ptr;
const int32_t* seqlen_k_ptr;
};
using Kargs = std::conditional_t<kIsGroupMode, GroupModeKargs, BatchModeKargs>;
template <bool Cond = !kIsGroupMode>
__host__ static constexpr std::enable_if_t<Cond, Kargs>
MakeKargs(const void* q_ptr,
const void* k_ptr,
const void* v_ptr,
const void* bias_ptr,
void* rand_val_ptr,
void* lse_acc_ptr,
void* o_acc_ptr,
ck_tile::index_t batch,
ck_tile::index_t nhead,
ck_tile::index_t max_seqlen_q,
ck_tile::index_t seqlen_q,
ck_tile::index_t seqlen_k,
ck_tile::index_t hdim_q,
ck_tile::index_t hdim_v,
ck_tile::index_t num_head_q,
ck_tile::index_t nhead_ratio_qk,
ck_tile::index_t num_splits,
float scale_s,
float scale_p,
ck_tile::index_t stride_q,
ck_tile::index_t stride_k,
ck_tile::index_t stride_v,
ck_tile::index_t stride_bias,
ck_tile::index_t stride_randval,
ck_tile::index_t nhead_stride_q,
ck_tile::index_t nhead_stride_k,
ck_tile::index_t nhead_stride_v,
ck_tile::index_t nhead_stride_bias,
ck_tile::index_t nhead_stride_randval,
ck_tile::index_t batch_stride_q,
ck_tile::index_t batch_stride_k,
ck_tile::index_t batch_stride_v,
ck_tile::index_t batch_stride_bias,
ck_tile::index_t batch_stride_randval,
ck_tile::index_t window_size_left,
ck_tile::index_t window_size_right,
ck_tile::index_t mask_type,
float p_drop,
bool s_randval,
const std::tuple<uint64_t, uint64_t>& drop_seed_offset)
{
Kargs kargs{{q_ptr,
k_ptr,
v_ptr,
lse_acc_ptr,
o_acc_ptr,
batch,
nhead,
max_seqlen_q,
seqlen_q,
seqlen_k,
hdim_q,
hdim_v,
num_head_q,
nhead_ratio_qk,
num_splits,
#if CK_TILE_FMHA_FWD_FAST_EXP2
static_cast<float>(scale_s * ck_tile::log2e_v<>),
#else
scale_s,
#endif
stride_q,
stride_k,
stride_v,
nhead_stride_q,
nhead_stride_k,
nhead_stride_v}, // args for common karg
{}, // placeholder for bias
{}, // placeholder for mask
{}, // placeholder for fp8_static_quant args
{}, // placeholder for dropout
batch_stride_q,
batch_stride_k,
batch_stride_v};
if constexpr(BiasEnum == BlockAttentionBiasEnum::ELEMENTWISE_BIAS)
{
kargs.bias_ptr = bias_ptr;
kargs.stride_bias = stride_bias;
kargs.nhead_stride_bias = nhead_stride_bias;
kargs.batch_stride_bias = batch_stride_bias;
}
else if constexpr(BiasEnum == BlockAttentionBiasEnum::ALIBI)
{
kargs.alibi_slope_ptr = bias_ptr;
kargs.alibi_slope_stride = stride_bias;
}
if constexpr(kHasMask)
{
kargs.window_size_left = window_size_left;
kargs.window_size_right = window_size_right;
kargs.mask_type = static_cast<ck_tile::GenericAttentionMaskEnum>(mask_type);
}
if constexpr(kDoFp8StaticQuant)
{
kargs.scale_p = scale_p;
}
if constexpr(kHasDropout)
{
kargs.init_dropout(p_drop, drop_seed_offset);
kargs.rand_val_ptr = rand_val_ptr;
kargs.stride_randval = stride_randval;
kargs.nhead_stride_randval = nhead_stride_randval;
kargs.batch_stride_randval = batch_stride_randval;
kargs.is_store_randval = s_randval;
}
return kargs;
}
template <bool Cond = kIsGroupMode>
__host__ static constexpr std::enable_if_t<Cond, Kargs>
MakeKargs(const void* q_ptr,
const void* k_ptr,
const void* v_ptr,
const void* bias_ptr,
void* rand_val_ptr,
void* lse_acc_ptr,
void* o_acc_ptr,
ck_tile::index_t batch,
ck_tile::index_t nhead,
ck_tile::index_t max_seqlen_q,
const void* seqstart_q_ptr,
const void* seqstart_k_ptr,
const void* seqlen_k_ptr,
ck_tile::index_t hdim_q,
ck_tile::index_t hdim_v,
ck_tile::index_t num_head_q,
ck_tile::index_t nhead_ratio_qk,
ck_tile::index_t num_splits,
float scale_s,
float scale_p,
ck_tile::index_t stride_q,
ck_tile::index_t stride_k,
ck_tile::index_t stride_v,
ck_tile::index_t stride_bias,
ck_tile::index_t stride_randval,
ck_tile::index_t nhead_stride_q,
ck_tile::index_t nhead_stride_k,
ck_tile::index_t nhead_stride_v,
ck_tile::index_t nhead_stride_bias,
ck_tile::index_t nhead_stride_randval,
ck_tile::index_t window_size_left,
ck_tile::index_t window_size_right,
ck_tile::index_t mask_type,
float p_drop,
bool s_randval,
const std::tuple<uint64_t, uint64_t>& drop_seed_offset)
{
Kargs kargs{{q_ptr,
k_ptr,
v_ptr,
lse_acc_ptr,
o_acc_ptr,
batch,
nhead,
max_seqlen_q,
-1, // seqlen will be updated by another pointer
-1, //
hdim_q,
hdim_v,
num_head_q,
nhead_ratio_qk,
num_splits,
#if CK_TILE_FMHA_FWD_FAST_EXP2
static_cast<float>(scale_s * ck_tile::log2e_v<>),
#else
scale_s,
#endif
stride_q,
stride_k,
stride_v,
nhead_stride_q,
nhead_stride_k,
nhead_stride_v}, // args for common karg
{}, // placeholder for bias
{}, // placeholder for mask
{}, // placeholder for fp8_static_quant args
{}, // placeholder for dropout
reinterpret_cast<const int32_t*>(seqstart_q_ptr),
reinterpret_cast<const int32_t*>(seqstart_k_ptr),
reinterpret_cast<const int32_t*>(seqlen_k_ptr)};
if constexpr(BiasEnum == BlockAttentionBiasEnum::ELEMENTWISE_BIAS)
{
kargs.bias_ptr = bias_ptr;
kargs.stride_bias = stride_bias;
kargs.nhead_stride_bias = nhead_stride_bias;
}
else if constexpr(BiasEnum == BlockAttentionBiasEnum::ALIBI)
{
kargs.alibi_slope_ptr = bias_ptr;
kargs.alibi_slope_stride = stride_bias;
}
if constexpr(kHasMask)
{
kargs.window_size_left = window_size_left;
kargs.window_size_right = window_size_right;
kargs.mask_type = static_cast<ck_tile::GenericAttentionMaskEnum>(mask_type);
}
if constexpr(kDoFp8StaticQuant)
{
kargs.scale_p = scale_p;
}
if constexpr(kHasDropout)
{
kargs.init_dropout(p_drop, drop_seed_offset);
kargs.rand_val_ptr = rand_val_ptr;
kargs.stride_randval = stride_randval;
kargs.nhead_stride_randval = nhead_stride_randval;
kargs.is_store_randval = s_randval;
}
return kargs;
}
__host__ static constexpr auto GridSize(ck_tile::index_t batch_size,
ck_tile::index_t nhead,
ck_tile::index_t seqlen_q,
ck_tile::index_t hdim_v,
ck_tile::index_t num_splits)
{
return TilePartitioner::GridSize(batch_size, nhead, seqlen_q, hdim_v, num_splits);
}
__host__ static constexpr auto BlockSize() { return dim3(kBlockSize); }
CK_TILE_HOST_DEVICE static constexpr ck_tile::index_t GetSmemSize()
{
return ck_tile::max(FmhaPipeline::GetSmemSize(), EpiloguePipeline::GetSmemSize());
}
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] =
TilePartitioner{}(kargs.seqlen_q, kargs.hdim_v, kargs.num_splits);
const index_t i_m0 = __builtin_amdgcn_readfirstlane(i_tile_m * FmhaPipeline::kM0);
const index_t i_n1 = __builtin_amdgcn_readfirstlane(i_tile_n * FmhaPipeline::kN1);
long_index_t batch_offset_q = 0;
long_index_t batch_offset_k = 0;
long_index_t batch_offset_v = 0;
long_index_t batch_offset_bias = 0;
long_index_t batch_offset_randval = 0;
long_index_t batch_offset_lse_acc = 0;
long_index_t batch_offset_o_acc = 0;
if constexpr(kIsGroupMode)
{
// get starting offset for each batch
const long_index_t query_start = kargs.seqstart_q_ptr[i_batch];
const long_index_t key_start = kargs.seqstart_k_ptr[i_batch];
batch_offset_q = query_start * kargs.stride_q;
batch_offset_k = key_start * kargs.stride_k;
if constexpr(std::is_same_v<VLayout, ck_tile::tensor_layout::gemm::RowMajor>)
{
batch_offset_v = key_start * kargs.stride_v;
}
else
{
batch_offset_v = key_start;
}
if constexpr(BiasEnum == BlockAttentionBiasEnum::ELEMENTWISE_BIAS)
{
batch_offset_bias = query_start * kargs.stride_bias + key_start;
}
if constexpr(kHasDropout)
{
batch_offset_randval = query_start * kargs.stride_randval;
}
batch_offset_lse_acc = query_start;
batch_offset_o_acc = query_start * kargs.hdim_v;
// get real # queries & # keys under group mode
const auto adjusted_seqstart_q_ptr = kargs.seqstart_q_ptr + i_batch;
kargs.seqlen_q = adjusted_seqstart_q_ptr[1] - adjusted_seqstart_q_ptr[0];
// # of required blocks is different in each groups, terminate unnecessary blocks
// earlier
if(kargs.seqlen_q <= i_m0)
{
return;
}
if(kargs.seqlen_k_ptr != nullptr)
{
kargs.seqlen_k = kargs.seqlen_k_ptr[i_batch];
}
else
{
const auto adjusted_seqstart_k_ptr = kargs.seqstart_k_ptr + i_batch;
kargs.seqlen_k = adjusted_seqstart_k_ptr[1] - adjusted_seqstart_k_ptr[0];
}
}
else
{
batch_offset_q = static_cast<long_index_t>(i_batch) * kargs.batch_stride_q;
batch_offset_k = static_cast<long_index_t>(i_batch) * kargs.batch_stride_k;
batch_offset_v = static_cast<long_index_t>(i_batch) * kargs.batch_stride_v;
if constexpr(BiasEnum == BlockAttentionBiasEnum::ELEMENTWISE_BIAS)
{
batch_offset_bias = static_cast<long_index_t>(i_batch) * kargs.batch_stride_bias;
}
if constexpr(kHasDropout)
{
batch_offset_randval =
static_cast<long_index_t>(i_batch) * kargs.batch_stride_randval;
}
batch_offset_lse_acc =
static_cast<long_index_t>(i_batch) * (kargs.nhead * kargs.max_seqlen_q);
batch_offset_o_acc = static_cast<long_index_t>(i_batch) *
(kargs.nhead * kargs.max_seqlen_q * kargs.hdim_v);
}
// for simplicity, batch stride we just modify the pointer
const QDataType* q_ptr = reinterpret_cast<const QDataType*>(kargs.q_ptr) +
static_cast<long_index_t>(i_nhead) * kargs.nhead_stride_q +
batch_offset_q;
const KDataType* k_ptr =
reinterpret_cast<const KDataType*>(kargs.k_ptr) +
static_cast<long_index_t>(i_nhead / kargs.nhead_ratio_qk) * kargs.nhead_stride_k +
batch_offset_k;
const VDataType* v_ptr =
reinterpret_cast<const VDataType*>(kargs.v_ptr) +
static_cast<long_index_t>(i_nhead / kargs.nhead_ratio_qk) * kargs.nhead_stride_v +
batch_offset_v;
OaccDataType* o_acc_ptr =
reinterpret_cast<OaccDataType*>(kargs.o_acc_ptr) +
static_cast<long_index_t>(i_nhead) * (kargs.max_seqlen_q * kargs.hdim_v) +
batch_offset_o_acc +
i_split * (kargs.batch * kargs.nhead * kargs.max_seqlen_q * kargs.hdim_v);
// Q/K/V DRAM and DRAM window
const auto q_dram = [&]() {
const auto q_dram_naive = make_naive_tensor_view<address_space_enum::global>(
q_ptr,
make_tuple(kargs.seqlen_q, kargs.hdim_q),
make_tuple(kargs.stride_q, 1),
number<FmhaPipeline::kAlignmentQ>{},
number<1>{});
if constexpr(FmhaPipeline::kQLoadOnce)
{
return pad_tensor_view(
q_dram_naive,
make_tuple(number<FmhaPipeline::kM0>{}, number<FmhaPipeline::kK0BlockLength>{}),
sequence<kPadSeqLenQ, kPadHeadDimQ>{});
}
else
{
return pad_tensor_view(
q_dram_naive,
make_tuple(number<FmhaPipeline::kM0>{}, number<FmhaPipeline::kK0>{}),
sequence<kPadSeqLenQ, kPadHeadDimQ>{});
}
}();
const auto k_dram = [&]() {
const auto k_dram_naive = make_naive_tensor_view<address_space_enum::global>(
k_ptr,
make_tuple(kargs.seqlen_k, kargs.hdim_q),
make_tuple(kargs.stride_k, 1),
number<FmhaPipeline::kAlignmentK>{},
number<1>{});
return pad_tensor_view(
k_dram_naive,
make_tuple(number<FmhaPipeline::kN0>{}, number<FmhaPipeline::kK0>{}),
sequence<kPadSeqLenK, kPadHeadDimQ>{});
}();
const auto v_dram = [&]() {
if constexpr(std::is_same_v<VLayout, ck_tile::tensor_layout::gemm::RowMajor>)
{
const auto v_dram_naive = make_naive_tensor_view<address_space_enum::global>(
v_ptr,
make_tuple(kargs.seqlen_k, kargs.hdim_v),
make_tuple(kargs.stride_v, 1),
number<FmhaPipeline::kAlignmentV>{},
number<1>{});
const auto v_dram_transposed =
transform_tensor_view(v_dram_naive,
make_tuple(make_pass_through_transform(kargs.hdim_v),
make_pass_through_transform(kargs.seqlen_k)),
make_tuple(sequence<1>{}, sequence<0>{}),
make_tuple(sequence<0>{}, sequence<1>{}));
return pad_tensor_view(
v_dram_transposed,
make_tuple(number<FmhaPipeline::kN1>{}, number<FmhaPipeline::kK1>{}),
sequence<kPadHeadDimV, kPadSeqLenK>{});
}
else
{
const auto v_dram_naive = make_naive_tensor_view<address_space_enum::global>(
v_ptr,
make_tuple(kargs.hdim_v, kargs.seqlen_k),
make_tuple(kargs.stride_v, 1),
number<FmhaPipeline::kAlignmentV>{},
number<1>{});
return pad_tensor_view(
v_dram_naive,
make_tuple(number<FmhaPipeline::kN1>{}, number<FmhaPipeline::kK1>{}),
sequence<kPadHeadDimV, kPadSeqLenK>{});
}
}();
auto q_dram_window = make_tile_window(
q_dram,
[&]() {
if constexpr(FmhaPipeline::kQLoadOnce)
return make_tuple(number<FmhaPipeline::kM0>{},
number<FmhaPipeline::kK0BlockLength>{});
else
return make_tuple(number<FmhaPipeline::kM0>{}, number<FmhaPipeline::kK0>{});
}(),
{i_m0, 0});
auto k_dram_window = make_tile_window(
k_dram, make_tuple(number<FmhaPipeline::kN0>{}, number<FmhaPipeline::kK0>{}), {0, 0});
auto v_dram_window =
make_tile_window(v_dram,
make_tuple(number<FmhaPipeline::kN1>{}, number<FmhaPipeline::kK1>{}),
{i_n1, 0});
/// FIXME: Before C++20, capturing structured binding variables are not supported. Remove
/// following copy capture of the 'i_nhead' if in C++20
const auto bias_dram_window = [&, i_nhead_ = i_nhead]() {
constexpr auto bias_dram_window_lengths =
make_tuple(number<FmhaPipeline::kM0>{}, number<FmhaPipeline::kN0>{});
if constexpr(BiasEnum == BlockAttentionBiasEnum::ELEMENTWISE_BIAS)
{
const BiasDataType* bias_ptr =
reinterpret_cast<const BiasDataType*>(kargs.bias_ptr) +
static_cast<long_index_t>(i_nhead_) * kargs.nhead_stride_bias +
batch_offset_bias;
const auto bias_dram = [&]() {
const auto bias_dram_naive = make_naive_tensor_view<address_space_enum::global>(
bias_ptr,
make_tuple(kargs.seqlen_q, kargs.seqlen_k),
make_tuple(kargs.stride_bias, 1),
number<FmhaPipeline::kAlignmentBias>{},
number<1>{});
return pad_tensor_view(bias_dram_naive,
bias_dram_window_lengths,
sequence<kPadSeqLenQ, kPadSeqLenK>{});
}();
return make_tile_window(bias_dram, bias_dram_window_lengths, {i_m0, 0});
}
else
{
return make_null_tile_window(bias_dram_window_lengths);
}
}();
// lse acc
auto lse_acc_dram_window = [&, i_nhead_ = i_nhead, i_split_ = i_split]() {
constexpr auto lse_acc_dram_window_lengths = make_tuple(number<FmhaPipeline::kM0>{});
LSEDataType* lse_acc_ptr = reinterpret_cast<LSEDataType*>(kargs.lse_acc_ptr) +
static_cast<long_index_t>(i_nhead_) * (kargs.max_seqlen_q) +
batch_offset_lse_acc +
i_split_ * (kargs.batch * kargs.nhead * kargs.max_seqlen_q);
const auto lse_acc_dram = [&]() {
const auto lse_acc_dram_naive =
make_naive_tensor_view<address_space_enum::global>(lse_acc_ptr,
make_tuple(kargs.seqlen_q),
make_tuple(1),
number<1>{},
number<1>{});
return pad_tensor_view(
lse_acc_dram_naive, lse_acc_dram_window_lengths, sequence<kPadSeqLenQ>{});
}();
return make_tile_window(lse_acc_dram, lse_acc_dram_window_lengths, {i_m0});
}();
// dropout
float rp_undrop = 1;
uint8_t p_undrop_in_uint8_t = std::numeric_limits<uint8_t>::max();
uint64_t drop_seed = 0;
uint64_t drop_offset = 0;
bool is_store_randval = false;
if constexpr(kHasDropout)
{
rp_undrop = kargs.rp_undrop;
p_undrop_in_uint8_t = kargs.p_undrop_in_uint8_t;
drop_seed = kargs.drop_seed;
drop_offset = kargs.drop_offset;
is_store_randval = kargs.is_store_randval;
}
BlockDropout dropout(i_batch,
i_nhead,
kargs.num_head_q,
drop_seed,
drop_offset,
rp_undrop,
p_undrop_in_uint8_t,
is_store_randval);
auto randval_dram_window = [&, i_nhead_ = i_nhead]() {
constexpr auto randval_dram_window_lengths =
make_tuple(number<FmhaPipeline::kM0>{}, number<FmhaPipeline::kN0>{});
if constexpr(kHasDropout)
{
RandValOutputDataType* rand_val_ptr =
reinterpret_cast<RandValOutputDataType*>(kargs.rand_val_ptr) +
static_cast<long_index_t>(i_nhead_) * kargs.nhead_stride_randval +
batch_offset_randval;
const auto randval_dram = [&]() {
const auto randval_dram_naive =
make_naive_tensor_view<address_space_enum::global>(
rand_val_ptr,
make_tuple(kargs.seqlen_q, kargs.seqlen_k),
make_tuple(kargs.stride_randval, 1),
number<1>{},
number<1>{});
return pad_tensor_view(randval_dram_naive,
randval_dram_window_lengths,
sequence<kPadSeqLenQ, kPadSeqLenK>{});
}();
return make_tile_window(randval_dram, randval_dram_window_lengths, {i_m0, 0});
}
else
{
return make_null_tile_window(randval_dram_window_lengths);
}
}();
FmhaMask mask = [&]() {
if constexpr(kHasMask)
return ck_tile::make_generic_attention_mask_from_lr_window<FmhaMask>(
kargs.window_size_left,
kargs.window_size_right,
kargs.seqlen_q,
kargs.seqlen_k,
kargs.mask_type == GenericAttentionMaskEnum::MASK_FROM_TOP_LEFT);
else
return FmhaMask{kargs.seqlen_q, kargs.seqlen_k};
}();
// WA i_batch capture structure binding before c++20
auto position_encoding = [&, i_batch_ = i_batch, i_nhead_ = i_nhead]() {
if constexpr(BiasEnum == BlockAttentionBiasEnum::ALIBI)
{
// data loading, shared by entire wg
// TODO: how to use s_read?
SaccDataType slope =
*(reinterpret_cast<const SaccDataType*>(kargs.alibi_slope_ptr) +
i_batch_ * kargs.alibi_slope_stride + i_nhead_);
#if CK_TILE_FMHA_FWD_FAST_EXP2
slope *= ck_tile::log2e_v<>;
#endif
if constexpr(kHasMask)
{
return make_alibi_from_lr_mask<SaccDataType, true>(slope,
kargs.window_size_left,
kargs.window_size_right,
kargs.seqlen_q,
kargs.seqlen_k,
kargs.mask_type);
}
else
{
return Alibi<SaccDataType, true>{
slope, kargs.seqlen_q, kargs.seqlen_k, AlibiMode::FROM_BOTTOM_RIGHT};
}
}
else
{
return EmptyPositionEncoding<SaccDataType>{};
}
}();
auto o_acc_tile = [&, i_split_ = i_split]() {
if constexpr(kDoFp8StaticQuant)
{
return FmhaPipeline{}(q_dram_window,
identity{}, // q_element_func
k_dram_window,
identity{}, // k_element_func
v_dram_window,
identity{}, // v_element_func
bias_dram_window,
identity{}, // bias_element_func
randval_dram_window,
lse_acc_dram_window,
identity{}, // lse_element_func
identity{}, // s_acc_element_func
scales{kargs.scale_p}, // p_compute_element_func
identity{}, // o_acc_element_func
mask,
position_encoding,
kargs.scale_s,
smem_ptr,
dropout,
i_split_,
kargs.num_splits);
}
else
{
return FmhaPipeline{}(q_dram_window,
k_dram_window,
v_dram_window,
bias_dram_window,
randval_dram_window,
lse_acc_dram_window,
mask,
position_encoding,
kargs.scale_s,
smem_ptr,
dropout,
i_split_,
kargs.num_splits);
}
}();
// Oacc DRAM and Oacc DRAM window
auto o_acc_dram = [&]() {
const auto o_acc_dram_naive = make_naive_tensor_view<address_space_enum::global>(
o_acc_ptr,
make_tuple(kargs.seqlen_q, kargs.hdim_v),
make_tuple(kargs.hdim_v, 1),
number<FmhaPipeline::kAlignmentO>{},
number<1>{});
return pad_tensor_view(
o_acc_dram_naive,
make_tuple(number<FmhaPipeline::kM0>{}, number<FmhaPipeline::kN1>{}),
sequence<kPadSeqLenQ, kPadHeadDimV>{});
}();
auto o_acc_dram_window =
make_tile_window(o_acc_dram,
make_tuple(number<FmhaPipeline::kM0>{}, number<FmhaPipeline::kN1>{}),
{i_m0, i_n1});
EpiloguePipeline{}(o_acc_dram_window, o_acc_tile);
}
};
} // namespace ck_tile

57
include/ck_tile/ops/fmha/kernel/fmha_fwd_splitkv_tile_partitioner.hpp Normal file
View File

@@ -0,0 +1,57 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck_tile/core.hpp"
namespace ck_tile {
template <typename BlockFmhaShape_>
struct FmhaFwdSplitKVTilePartitioner
{
using BlockFmhaShape = ck_tile::remove_cvref_t<BlockFmhaShape_>;
static constexpr ck_tile::index_t kM0 = BlockFmhaShape::kM0;
static constexpr ck_tile::index_t kN0 = BlockFmhaShape::kN0;
static constexpr ck_tile::index_t kK0 = BlockFmhaShape::kK0;
static constexpr ck_tile::index_t kN1 = BlockFmhaShape::kN1;
static constexpr ck_tile::index_t kK1 = BlockFmhaShape::kK1;
__host__ static constexpr auto GridSize(ck_tile::index_t batch_size_,
ck_tile::index_t nhead_,
ck_tile::index_t seqlen_q_,
ck_tile::index_t hdim_v_,
ck_tile::index_t num_splits)
{
// TODO: this may need tuning
return dim3(ck_tile::integer_divide_ceil(seqlen_q_, kM0) *
ck_tile::integer_divide_ceil(hdim_v_, kN1),
nhead_ * num_splits,
batch_size_);
}
CK_TILE_DEVICE auto
operator()(ck_tile::index_t /*seqlen_q*/, ck_tile::index_t hdim_v, ck_tile::index_t num_splits)
{
// const index_t num_tile_m0 = seqlen_q / kM0;
const index_t num_tile_n1 = ck_tile::integer_divide_ceil(hdim_v, kN1);
const index_t i_block = blockIdx.x;
const index_t i_nhead = blockIdx.y / num_splits;
const index_t i_split = blockIdx.y - (i_nhead * num_splits);
const index_t i_batch = blockIdx.z;
const auto f = [](index_t dividend, index_t divisor) {
index_t quotient = dividend / divisor;
index_t modulus = dividend - quotient * divisor;
return ck_tile::make_tuple(quotient, modulus);
};
const auto [i_tile_m, i_tile_n] = f(i_block, num_tile_n1);
return ck_tile::make_tuple(i_tile_m, i_tile_n, i_split, i_nhead, i_batch);
}
};
} // namespace ck_tile

0
include/ck_tile/ops/fmha/pipeline/block_fmha_fwd_splitkv_pipeline_default_policy.hpp Normal file
View File

664
include/ck_tile/ops/fmha/pipeline/block_fmha_fwd_splitkv_pipeline_qr_ks_vs.hpp Normal file
View File

@@ -0,0 +1,664 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck_tile/core.hpp"
#include "ck_tile/ops/fmha/block/block_attention_bias_enum.hpp"
#include "ck_tile/ops/fmha/pipeline/block_fmha_fwd_splitkv_pipeline_qr_ks_vs_default_policy.hpp"
#include "ck_tile/ops/fmha/block/block_dropout.hpp"
#include "ck_tile/ops/reduce/block/block_reduce.hpp"
namespace ck_tile {
// This pipeline is qkv all located in LDS
template <typename Problem_, typename Policy_ = BlockFmhaFwdSplitKVPipelineQRKSVSDefaultPolicy>
struct BlockFmhaFwdSplitKVPipelineQRKSVS
{
using Problem = remove_cvref_t<Problem_>;
using Policy = remove_cvref_t<Policy_>;
using QDataType = remove_cvref_t<typename Problem::QDataType>;
using KDataType = remove_cvref_t<typename Problem::KDataType>;
using VDataType = remove_cvref_t<typename Problem::VDataType>;
using SaccDataType = remove_cvref_t<typename Problem::SaccDataType>;
using SMPLComputeDataType = remove_cvref_t<typename Problem::SMPLComputeDataType>;
using BiasDataType = remove_cvref_t<typename Problem::BiasDataType>;
using RandValOutputDataType = remove_cvref_t<typename Problem::RandValOutputDataType>;
using LSEDataType = remove_cvref_t<typename Problem::LSEDataType>;
using PDataType = remove_cvref_t<typename Problem::PDataType>;
using OaccDataType = remove_cvref_t<typename Problem::OaccDataType>;
using FmhaMask = remove_cvref_t<typename Problem::FmhaMask>;
using BlockFmhaShape = remove_cvref_t<typename Problem::BlockFmhaShape>;
using VLayout = remove_cvref_t<typename BlockFmhaShape::VLayout>;
static constexpr bool kQLoadOnce = true; // if q_tile load whole block length (hdim) at once
static_assert(kQLoadOnce == Policy::QLoadOnce);
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 bool kIsGroupMode = Problem::kIsGroupMode;
static constexpr bool kPadSeqLenQ = Problem::kPadSeqLenQ;
static constexpr bool kPadSeqLenK = Problem::kPadSeqLenK;
static constexpr bool kPadHeadDimQ = Problem::kPadHeadDimQ;
static constexpr bool kPadHeadDimV = Problem::kPadHeadDimV;
static constexpr auto BiasEnum = Problem::BiasEnum;
static constexpr bool kHasDropout = Problem::kHasDropout;
// last dimension vector length used to create tensor view(and decide buffer_load vector length)
// ... together with tensor distribution. tensor dist should able to overwrite this
static constexpr index_t kAlignmentQ =
kPadHeadDimQ ? 1 : Policy::template GetAlignmentQ<Problem>();
static constexpr index_t kAlignmentK =
kPadHeadDimQ ? 1 : Policy::template GetAlignmentK<Problem>();
static constexpr index_t kAlignmentV = []() {
if constexpr(std::is_same_v<VLayout, ck_tile::tensor_layout::gemm::RowMajor>)
return kPadHeadDimV ? 1 : Policy::template GetAlignmentV<Problem>();
else
return kPadSeqLenK ? 1 : Policy::template GetAlignmentV<Problem>();
}();
static constexpr index_t kAlignmentO =
kPadHeadDimV ? 1 : Policy::template GetAlignmentO<Problem>();
static constexpr index_t kAlignmentBias =
kPadSeqLenK ? 1 : Policy::template GetAlignmentBias<Problem>();
static constexpr index_t kBlockPerCu = []() {
if constexpr(Problem::kBlockPerCu != -1)
return Problem::kBlockPerCu;
else
{
if constexpr(kK0BlockLength <= 32)
{
return 2;
}
else if constexpr(kK0BlockLength <= 64)
{
return 3;
}
else if constexpr(kK0BlockLength <= 128)
{
if constexpr(BiasEnum == BlockAttentionBiasEnum::ELEMENTWISE_BIAS)
return 1;
else
return 2;
}
else if constexpr(kK0BlockLength <= 256)
{
return 1;
}
}
}();
static constexpr const char* name = "qr";
CK_TILE_HOST_DEVICE static constexpr ck_tile::index_t GetSmemSize()
{
return Policy::template GetSmemSize<Problem>();
}
template <typename QDramBlockWindowTmp,
typename KDramBlockWindowTmp,
typename VDramBlockWindowTmp,
typename BiasDramBlockWindowTmp,
typename RandValDramBlockWindowTmp,
typename LSEaccDramBlockWindowTmp,
typename QElementFunction,
typename KElementFunction,
typename VElementFunction,
typename BiasElementFunction,
typename LSEaccElementFunction,
typename SAccElementFunction,
typename PComputeElementFunction,
typename OAccElementFunction,
typename PositionEncoding>
CK_TILE_HOST_DEVICE auto
operator()(const QDramBlockWindowTmp& q_dram_block_window_tmp, // M0*K0 tile
const QElementFunction& q_element_func,
const KDramBlockWindowTmp& k_dram_block_window_tmp, // N0*K0 tile
const KElementFunction& k_element_func,
const VDramBlockWindowTmp& v_dram_block_window_tmp, // N1*K1 tile
const VElementFunction& v_element_func,
const BiasDramBlockWindowTmp& bias_dram_block_window_tmp, // M0*N0 tile
const BiasElementFunction& bias_element_func,
RandValDramBlockWindowTmp& randval_dram_block_window_tmp,
LSEaccDramBlockWindowTmp& lse_acc_dram_window_tmp, // M0*1 tile
const LSEaccElementFunction& lse_acc_element_func,
const SAccElementFunction& s_acc_element_func,
const PComputeElementFunction& p_compute_element_func,
const OAccElementFunction& o_acc_element_func,
FmhaMask mask,
PositionEncoding position_encoding,
float scale_s,
void* smem_ptr,
BlockDropout& dropout,
index_t i_split,
index_t num_splits) 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()[number<0>{}] &&
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>{}],
"wrong!");
// K tile in LDS
KDataType* k_lds_ptr = static_cast<KDataType*>(static_cast<void*>(
static_cast<char*>(smem_ptr) + Policy::template GetSmemSizeQ<Problem>()));
auto k_lds = make_tensor_view<address_space_enum::lds>(
k_lds_ptr, Policy::template MakeKLdsBlockDescriptor<Problem>());
auto k_lds_window =
make_tile_window(k_lds, make_tuple(number<kN0>{}, number<kK0>{}), {0, 0});
// V tile in LDS
auto v_lds = make_tensor_view<address_space_enum::lds>(
reinterpret_cast<VDataType*>(smem_ptr),
Policy::template MakeVLdsBlockDescriptor<Problem>());
auto v_lds_window = make_tile_window(
v_lds, Policy::template MakeVLdsBlockDescriptor<Problem>().get_lengths(), {0, 0});
// Block GEMM
constexpr auto gemm_0 = Policy::template GetQKBlockGemm<Problem>();
constexpr auto gemm_1 = Policy::template GetKVBlockGemm<Problem>();
auto q_dram_window = make_tile_window(
q_dram_block_window_tmp.get_bottom_tensor_view(),
q_dram_block_window_tmp.get_window_lengths(),
q_dram_block_window_tmp.get_window_origin(),
Policy::template MakeQDramTileDistribution<Problem, decltype(gemm_0)>());
auto q = load_tile(q_dram_window);
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; };
// infer Sacc, S, P, M, L, Oacc type
using SBlockTileType = decltype(cast_tile<SMPLComputeDataType>(s_acc));
using MLBlockTileType = decltype(block_tile_reduce<SMPLComputeDataType>(
SBlockTileType{}, sequence<1>{}, f_max, SMPLComputeDataType{0}));
using OaccBlockTileType = decltype(gemm_1.MakeCBlockTile());
// init Oacc, M, L
auto o_acc = OaccBlockTileType{};
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_window.get_window_origin();
const auto [seqlen_k_start, seqlen_k_end] = mask.GetTileRangeAlongX(
q_origin.at(number<0>{}), number<kM0>{}, number<kN0>{}, i_split, num_splits);
const auto num_total_loop = integer_divide_ceil(seqlen_k_end - seqlen_k_start, kN0);
// check early exit if masked and no work to do.
if constexpr(FmhaMask::IsMasking)
{
if(num_total_loop <= 0)
{
auto lse_acc =
make_static_distributed_tensor<LSEDataType>(m.get_tile_distribution());
set_tile(lse_acc, -numeric<SMPLComputeDataType>::infinity());
store_tile(lse_acc_dram_window_tmp,
tile_elementwise_in(lse_acc_element_func, lse_acc));
// Note: here occ are all cleard, return it
// Note: q loaded but no fence, ignore it.
return o_acc;
}
}
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});
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 randval_dram_window = dropout.MakeRandvalDramWindow<decltype(gemm_0)>(
randval_dram_block_window_tmp, seqlen_k_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?
Policy::template MakeVDramTileDistribution<Problem>());
auto q_tile = tile_elementwise_in(q_element_func, q);
// prefetch K tile
index_t i_total_loops = 0;
constexpr index_t k0_loops = kK0BlockLength / kK0;
constexpr index_t k1_loops = kN0 / kK1;
static_assert(2 <= k0_loops);
static_assert(1 <= k1_loops);
auto loop_body = [&, seqlen_k_start_ = seqlen_k_start, seqlen_k_end_ = seqlen_k_end](
auto is_last_iteration) {
// STAGE 1, QK gemm
auto k_dram_window = make_tile_window(
k_dram_block_window.get_bottom_tensor_view(),
k_dram_block_window.get_window_lengths(),
k_dram_block_window.get_window_origin(),
Policy::template MakeKDramTileDistribution<Problem>()); // K DRAM tile window for
// load
auto k_block_tile = load_tile(k_dram_window);
{
move_tile_window(k_dram_window, {0, kK0});
clear_tile(s_acc); // initialize C
store_tile(k_lds_window, tile_elementwise_in(k_element_func, k_block_tile));
k_block_tile = load_tile(k_dram_window);
}
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(k0_loops > 2)
{
static_for<0, k0_loops - 2, 1>{}([&](auto i_k0) {
block_sync_lds();
gemm_0(s_acc,
get_slice_tile(q_tile,
sequence<0, i_k0 * kK0>{},
sequence<kM0, (i_k0 + 1) * kK0>{}),
k_lds_window);
block_sync_lds();
move_tile_window(k_dram_window, {0, kK0});
store_tile(
k_lds_window,
tile_elementwise_in(k_element_func, k_block_tile)); // LDS write i + 1
k_block_tile = load_tile(k_dram_window); // global read i + 2
});
}
const auto v_prefetch = load_tile(v_dram_window); // prefetch load v tile
{ // tail
block_sync_lds();
gemm_0(s_acc,
get_slice_tile(q_tile,
sequence<0, (k0_loops - 2) * kK0>{},
sequence<kM0, (k0_loops - 1) * kK0>{}),
k_lds_window);
block_sync_lds();
store_tile(k_lds_window, tile_elementwise_in(k_element_func, k_block_tile));
block_sync_lds();
gemm_0(s_acc,
get_slice_tile(q_tile,
sequence<0, (k0_loops - 1) * kK0>{},
sequence<kM0, k0_loops * kK0>{}),
k_lds_window);
}
// STAGE 2, scale_s, add bias, mask, softmax
if constexpr(BiasEnum == BlockAttentionBiasEnum::ELEMENTWISE_BIAS)
{
s_acc = tile_elementwise_in(s_acc_element_func, s_acc);
tile_elementwise_inout([&scale_s](auto& x) { x = x * scale_s; }, s_acc);
tile_elementwise_inout(
[&](auto& x, const auto& y) {
#if !CK_TILE_FMHA_FWD_FAST_EXP2
x += type_convert<SaccDataType>(bias_element_func(y));
#else
x += log2e_v<SaccDataType> *
type_convert<SaccDataType>(bias_element_func(y));
#endif
},
s_acc,
bias_tile);
}
else if constexpr(BiasEnum == BlockAttentionBiasEnum::ALIBI)
{
const auto k_origin = k_dram_block_window.get_window_origin();
constexpr auto s_spans = decltype(s_acc)::get_distributed_spans();
s_acc = tile_elementwise_in(s_acc_element_func, s_acc);
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.update(s_acc(i_j_idx), row, col);
});
});
}
else
{
s_acc = tile_elementwise_in(s_acc_element_func, s_acc);
#if !CK_TILE_FMHA_FWD_FAST_EXP2
tile_elementwise_inout([&scale_s](auto& x) { x = x * scale_s; }, s_acc);
#endif
}
move_tile_window(bias_dram_window, {0, kN0});
if constexpr(is_last_iteration)
{
const auto k_origin = k_dram_block_window.get_window_origin();
set_tile_if(s_acc, -numeric<SMPLComputeDataType>::infinity(), [&](auto tile_idx) {
const auto col = k_origin.at(number<0>{}) + tile_idx.at(number<1>{});
return seqlen_k_end_ <= col;
});
}
if constexpr(kPadSeqLenK || FmhaMask::IsMasking)
{
const auto k_origin = k_dram_block_window.get_window_origin();
bool need_perpixel_check = mask.IsEdgeTile(q_origin.at(number<0>{}),
k_origin.at(number<0>{}),
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>{});
return mask.IsOutOfBound(row, col);
});
}
}
const auto s = cast_tile<SMPLComputeDataType>(s_acc); // S{j}
auto m_local = block_tile_reduce<SMPLComputeDataType>(
s,
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.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[number<0>{}], [&](auto idx0) {
constexpr auto i_idx = make_tuple(idx0);
#if CK_TILE_FMHA_FWD_FAST_EXP2
auto row_max = scale_s * get_validated_m(m[i_idx]);
#endif
sweep_tile_span(p_spans[number<1>{}], [&](auto idx1) {
constexpr auto i_j_idx = make_tuple(idx0, idx1);
#if CK_TILE_FMHA_FWD_FAST_EXP2
if constexpr(BiasEnum == BlockAttentionBiasEnum::ELEMENTWISE_BIAS ||
BiasEnum == BlockAttentionBiasEnum::ALIBI)
{
p_compute(i_j_idx) = exp2(s[i_j_idx] - get_validated_m(m[i_idx]));
}
else
{
p_compute(i_j_idx) = exp2(scale_s * s[i_j_idx] - row_max);
}
#else
p_compute(i_j_idx) = exp(s[i_j_idx] - get_validated_m(m[i_idx]));
#endif
});
});
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>{});
// l{j}, Oacc{j}
constexpr auto o_spans = decltype(o_acc)::get_distributed_spans();
sweep_tile_span(o_spans[number<0>{}], [&](auto idx0) {
constexpr auto i_idx = make_tuple(idx0);
#if CK_TILE_FMHA_FWD_FAST_EXP2
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
{
auto row_max = scale_s * get_validated_m(m[i_idx]);
return exp2(scale_s * m_old[i_idx] - row_max);
}
}();
#else
const auto tmp = exp(m_old[i_idx] - get_validated_m(m[i_idx]));
#endif
l(i_idx) = tmp * l[i_idx] + rowsum_p[i_idx];
sweep_tile_span(o_spans[number<1>{}], [&](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;
});
});
if constexpr(kHasDropout)
{
dropout.Run<decltype(gemm_0), SMPLComputeDataType, RandValOutputDataType>(
smem_ptr,
seqlen_k_start_ + i_total_loops * kN0,
p_compute,
randval_dram_window);
}
block_sync_lds();
if constexpr(std::is_same_v<VLayout, ck_tile::tensor_layout::gemm::RowMajor>)
{
auto v_shuffle_tmp = make_static_distributed_tensor<VDataType>(
Policy::template MakeShuffledVRegBlockDescriptor<Problem>());
shuffle_tile(v_shuffle_tmp, v_prefetch);
store_tile(
v_lds_window,
tile_elementwise_in(v_element_func, v_shuffle_tmp)); // store the prefetch
}
else
{
store_tile(v_lds_window,
tile_elementwise_in(v_element_func, v_prefetch)); // store the prefetch
}
move_tile_window(v_dram_window, {0, kK1});
const auto p =
cast_tile<PDataType>(tile_elementwise_in(p_compute_element_func, p_compute));
// STAGE 3, KV gemm
if constexpr(k1_loops > 1)
{
static_for<0, k1_loops - 1, 1>{}([&](auto i_k1) {
const auto v = load_tile(v_dram_window); // load next v
block_sync_lds();
gemm_1(o_acc,
get_slice_tile(
p, sequence<0, i_k1 * kK1>{}, sequence<kM0, (i_k1 + 1) * kK1>{}),
v_lds_window);
block_sync_lds();
if constexpr(std::is_same_v<VLayout, ck_tile::tensor_layout::gemm::RowMajor>)
{
auto v_shuffle_tmp = make_static_distributed_tensor<VDataType>(
Policy::template MakeShuffledVRegBlockDescriptor<Problem>());
shuffle_tile(v_shuffle_tmp, v);
store_tile(v_lds_window,
tile_elementwise_in(v_element_func,
v_shuffle_tmp)); // store the prefetch
}
else
{
store_tile(v_lds_window,
tile_elementwise_in(v_element_func, v)); // store next v
}
move_tile_window(v_dram_window, {0, kK1});
});
}
// move K tile windows
move_tile_window(k_dram_block_window, {kN0, 0});
// tail
{
block_sync_lds();
gemm_1(o_acc,
get_slice_tile(p, sequence<0, (k1_loops - 1) * kK1>{}, sequence<kM0, kN0>{}),
v_lds_window);
block_sync_lds();
}
};
while(++i_total_loops < num_total_loop - 1)
{
loop_body(std::false_type{});
}
loop_body(std::true_type{});
// 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[number<0>{}], [&, m_ = m, l_ = l](auto idx0) {
constexpr auto i_idx = make_tuple(idx0);
#if CK_TILE_FMHA_FWD_FAST_EXP2
if constexpr(BiasEnum == BlockAttentionBiasEnum::ELEMENTWISE_BIAS ||
BiasEnum == BlockAttentionBiasEnum::ALIBI)
{
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]);
}
#else
lse_acc(i_idx) = m_[i_idx] + log(l_[i_idx]);
#endif
});
store_tile(lse_acc_dram_window_tmp, tile_elementwise_in(lse_acc_element_func, lse_acc));
// finally, O
constexpr auto o_spans = decltype(o_acc)::get_distributed_spans();
sweep_tile_span(o_spans[number<0>{}], [&](auto idx0) {
constexpr auto i_idx = make_tuple(idx0);
const auto tmp = [&]() {
if constexpr(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[number<1>{}], [&](auto idx1) {
constexpr auto i_j_idx = make_tuple(idx0, idx1);
o_acc(i_j_idx) *= tmp;
});
});
o_acc = tile_elementwise_in(o_acc_element_func, o_acc);
return o_acc;
}
template <typename QDramBlockWindowTmp,
typename KDramBlockWindowTmp,
typename VDramBlockWindowTmp,
typename BiasDramBlockWindowTmp,
typename RandValDramBlockWindowTmp,
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
RandValDramBlockWindowTmp& randval_dram_block_window_tmp, // M0*N0 tile
LSEaccDramBlockWindowTmp& lse_acc_dram_block_window_tmp, // M0*1 tile
FmhaMask mask,
PositionEncoding position_encoding,
float scale_s,
void* smem_ptr,
BlockDropout& dropout,
index_t i_split,
index_t num_splits) const
{
return operator()(q_dram_block_window_tmp,
identity{},
k_dram_block_window_tmp,
identity{},
v_dram_block_window_tmp,
identity{},
bias_dram_block_window_tmp,
identity{},
randval_dram_block_window_tmp,
lse_acc_dram_block_window_tmp,
identity{},
identity{},
identity{},
identity{},
mask,
position_encoding,
scale_s,
smem_ptr,
dropout,
i_split,
num_splits);
}
};
} // namespace ck_tile

765
include/ck_tile/ops/fmha/pipeline/block_fmha_fwd_splitkv_pipeline_qr_ks_vs_async.hpp Normal file
View File

@@ -0,0 +1,765 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck_tile/core.hpp"
#include "ck_tile/ops/common/tensor_layout.hpp"
#include "ck_tile/ops/fmha/block/block_attention_bias_enum.hpp"
#include "ck_tile/ops/fmha/pipeline/block_fmha_fwd_splitkv_pipeline_qr_ks_vs_async_default_policy.hpp"
#include "ck_tile/ops/fmha/block/block_dropout.hpp"
#include "ck_tile/ops/reduce/block/block_reduce.hpp"
namespace ck_tile {
// a variation of qr/ks/vs, where we use async copy to load k (potentially v in the future)
template <typename Problem_, typename Policy_ = BlockFmhaFwdSplitKVPipelineQRKSVSAsyncDefaultPolicy>
struct BlockFmhaFwdSplitKVPipelineQRKSVSAsync
{
using Problem = remove_cvref_t<Problem_>;
using Policy = remove_cvref_t<Policy_>;
using QDataType = remove_cvref_t<typename Problem::QDataType>;
using KDataType = remove_cvref_t<typename Problem::KDataType>;
using VDataType = remove_cvref_t<typename Problem::VDataType>;
using SaccDataType = remove_cvref_t<typename Problem::SaccDataType>;
using SMPLComputeDataType = remove_cvref_t<typename Problem::SMPLComputeDataType>;
using BiasDataType = remove_cvref_t<typename Problem::BiasDataType>;
using RandValOutputDataType = remove_cvref_t<typename Problem::RandValOutputDataType>;
using LSEDataType = remove_cvref_t<typename Problem::LSEDataType>;
using PDataType = remove_cvref_t<typename Problem::PDataType>;
using OaccDataType = remove_cvref_t<typename Problem::OaccDataType>;
using FmhaMask = remove_cvref_t<typename Problem::FmhaMask>;
using BlockFmhaShape = remove_cvref_t<typename Problem::BlockFmhaShape>;
using VLayout = remove_cvref_t<typename BlockFmhaShape::VLayout>;
static constexpr bool kQLoadOnce = true; // if q_tile load whole block length (hdim) at once
static_assert(kQLoadOnce == Policy::QLoadOnce);
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 bool kIsGroupMode = Problem::kIsGroupMode;
// TODO: seq_q always support padding, hdim_q/v support multiple of vector(like 8x)
// only need special care about seq_k padding (oob need set -INF of p instead of zero)
static_assert(Problem::kPadSeqLenQ == true && Problem::kPadHeadDimQ == true &&
Problem::kPadHeadDimV == true);
static constexpr bool kPadSeqLenQ = true;
static constexpr bool kPadSeqLenK = Problem::kPadSeqLenK;
static constexpr bool kPadHeadDimQ = true; // support multiple of vector(like 8x)
static constexpr bool kPadHeadDimV = true; // support multiple of vector(like 8x)
static constexpr auto BiasEnum = Problem::BiasEnum;
static constexpr bool kHasDropout = Problem::kHasDropout;
// last dimension vector length used to create tensor view(and decide buffer_load vector length)
// ... together with tensor distribution. tensor dist should able to overwrite this
static constexpr index_t kAlignmentQ = Policy::template GetAlignmentQ<Problem>();
static constexpr index_t kAlignmentK = Policy::template GetAlignmentK<Problem>();
static constexpr index_t kAlignmentV = []() {
if constexpr(std::is_same_v<VLayout, ck_tile::tensor_layout::gemm::RowMajor>)
return Policy::template GetAlignmentV<Problem>();
else
return kPadSeqLenK ? 1 : Policy::template GetAlignmentV<Problem>();
}();
static constexpr index_t kAlignmentO = Policy::template GetAlignmentO<Problem>();
static constexpr index_t kAlignmentBias =
kPadSeqLenK ? 1 : Policy::template GetAlignmentBias<Problem>();
#if CK_TILE_FMHA_FWD_FAST_EXP2
static constexpr auto R_LOG2E = 1.0 / log2e_v<SaccDataType>;
#endif
static constexpr index_t kBlockPerCu = []() {
if constexpr(Problem::kBlockPerCu != -1)
return Problem::kBlockPerCu;
else
{
if constexpr(kK0BlockLength <= 32)
{
if constexpr(kPadSeqLenK && BiasEnum == BlockAttentionBiasEnum::ELEMENTWISE_BIAS &&
FmhaMask::IsMasking)
return 1;
else
return 2;
}
else if constexpr(kK0BlockLength <= 64)
{
if constexpr(kPadSeqLenK && BiasEnum == BlockAttentionBiasEnum::ELEMENTWISE_BIAS)
return 2;
else
return 3;
}
else if constexpr(kK0BlockLength <= 128)
{
if constexpr(kPadSeqLenK && BiasEnum == BlockAttentionBiasEnum::ELEMENTWISE_BIAS)
return 1;
else
return 2;
}
else if constexpr(kK0BlockLength <= 256)
{
return 1;
}
}
}();
static constexpr const char* name = "qr_async";
CK_TILE_HOST_DEVICE static constexpr ck_tile::index_t GetSmemSize()
{
return Policy::template GetSmemSize<Problem>();
}
template <typename QDramBlockWindowTmp,
typename KDramBlockWindowTmp,
typename VDramBlockWindowTmp,
typename BiasDramBlockWindowTmp,
typename RandValDramBlockWindowTmp,
typename LSEaccDramBlockWindowTmp,
typename QElementFunction,
typename KElementFunction,
typename VElementFunction,
typename BiasElementFunction,
typename LSEaccElementFunction,
typename SAccElementFunction,
typename PComputeElementFunction,
typename OAccElementFunction,
typename PositionEncoding>
CK_TILE_HOST_DEVICE auto
operator()(const QDramBlockWindowTmp& q_dram_block_window_tmp, // M0*K0 tile
const QElementFunction& q_element_func,
const KDramBlockWindowTmp& k_dram_block_window_tmp, // N0*K0 tile
const KElementFunction& /*k_element_func*/,
const VDramBlockWindowTmp& v_dram_block_window_tmp, // N1*K1 tile
const VElementFunction& v_element_func,
const BiasDramBlockWindowTmp& bias_dram_block_window_tmp, // M0*N0 tile
const BiasElementFunction& bias_element_func,
RandValDramBlockWindowTmp& randval_dram_block_window_tmp,
LSEaccDramBlockWindowTmp& lse_acc_dram_window_tmp, // M0*1 tile
const LSEaccElementFunction& lse_acc_element_func,
const SAccElementFunction& s_acc_element_func,
const PComputeElementFunction& p_compute_element_func,
const OAccElementFunction& o_acc_element_func,
FmhaMask mask,
PositionEncoding position_encoding,
float scale_s,
void* smem_ptr,
BlockDropout& dropout,
index_t i_split,
index_t num_splits) 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()[number<0>{}] &&
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>{}],
"wrong!");
constexpr auto LdsSeq = Policy::template GetLdsBufferSequence<Problem>();
// K tile in LDS
auto k_lds_ptr = reinterpret_cast<KDataType*>(smem_ptr);
auto k_lds_store = generate_tuple(
[&](auto i_buf) {
return make_tile_window(
make_tensor_view<address_space_enum::lds>(
k_lds_ptr, Policy::template MakeKLdsStoreBlockDescriptor<Problem>(i_buf)),
Policy::template MakeKLdsStoreBlockDescriptor<Problem>(i_buf).get_lengths(),
{0, 0, 0});
},
number<Policy::NumPrefetchK>{});
#if K_LDS_LOAD_USE_OFFSET_TRANSFORM
auto k_lds_load = generate_tuple(
[&](auto i_buf) {
return make_tile_window(
make_tensor_view<address_space_enum::lds>(
k_lds_ptr, Policy::template MakeKLdsLoadBlockDescriptor<Problem>(i_buf)),
Policy::template MakeKLdsLoadBlockDescriptor<Problem>(i_buf).get_lengths(),
{0, 0});
},
number<Policy::NumPrefetchK>{});
#else
auto k_lds_Load_view = make_tensor_view<address_space_enum::lds>(
k_lds_ptr, Policy::template MakeKLdsLoadBlockDescriptor<Problem>());
auto k_lds_load =
make_tile_window(k_lds_Load_view,
Policy::template MakeKLdsLoadBlockDescriptor<Problem>().get_lengths(),
{0, 0});
#endif
// V tile in LDS
auto v_lds = make_tensor_view<address_space_enum::lds>(
reinterpret_cast<VDataType*>(smem_ptr),
Policy::template MakeVLdsBlockDescriptor<Problem>());
auto v_lds_window = make_tile_window(
v_lds, Policy::template MakeVLdsBlockDescriptor<Problem>().get_lengths(), {0, 0});
// Block GEMM
constexpr auto gemm_0 = Policy::template GetQKBlockGemm<Problem>();
constexpr auto gemm_1 = Policy::template GetKVBlockGemm<Problem>();
auto q_dram_window = make_tile_window(
q_dram_block_window_tmp.get_bottom_tensor_view(),
q_dram_block_window_tmp.get_window_lengths(),
q_dram_block_window_tmp.get_window_origin(),
Policy::template MakeQDramTileDistribution<Problem, decltype(gemm_0)>());
// TODO: we use async Copy for K, which is inline asm
// a side effect is we have to use inline asm for q as well
auto q = decltype(load_tile(q_dram_window)){};
set_tile(q, number<0>{}); // use per-dword clear to avoid scratch
load_tile_raw(q, q_dram_window);
__builtin_amdgcn_sched_barrier(0);
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; };
// infer Sacc, S, P, M, L, Oacc type
using SBlockTileType = decltype(cast_tile<SMPLComputeDataType>(s_acc));
using MLBlockTileType = decltype(block_tile_reduce<SMPLComputeDataType>(
SBlockTileType{}, sequence<1>{}, f_max, SMPLComputeDataType{0}));
using OaccBlockTileType = decltype(gemm_1.MakeCBlockTile());
// init Oacc, M, L
auto o_acc = OaccBlockTileType{};
auto m = MLBlockTileType{};
auto l = MLBlockTileType{};
clear_tile(o_acc);
set_tile(m, -numeric<SMPLComputeDataType>::infinity());
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>{}, i_split, num_splits);
const auto num_total_loop = integer_divide_ceil(seqlen_k_end - seqlen_k_start, kN0);
// check early exit
if constexpr(FmhaMask::IsMasking || kPadSeqLenK)
{
if(num_total_loop <= 0)
{
auto lse_acc =
make_static_distributed_tensor<LSEDataType>(m.get_tile_distribution());
set_tile(lse_acc, -numeric<SMPLComputeDataType>::infinity());
store_tile(lse_acc_dram_window_tmp,
tile_elementwise_in(lse_acc_element_func, lse_acc));
buffer_load_fence(0); // rocm-6.1, if whole tile is masked out, need to fence(0)
// otherwise will have compute error(maybe compiler bug?)
// Note: here occ are all cleard, return it
return o_acc;
}
__builtin_amdgcn_sched_barrier(0); // make sure sched_barrier(0) for this check
}
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});
auto k_dram_window = make_tile_window(
k_dram_block_window.get_bottom_tensor_view(),
k_dram_block_window.get_window_lengths(),
k_dram_block_window.get_window_origin(),
Policy::template MakeKDramTileDistribution<Problem>()); // K DRAM tile window for
// load
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 randval_dram_window = dropout.MakeRandvalDramWindow<decltype(gemm_0)>(
randval_dram_block_window_tmp, seqlen_k_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?
Policy::template MakeVDramTileDistribution<Problem>());
// prefetch K tile
async_load_tile_raw(k_lds_store(LdsSeq.at(number<0>{})), k_dram_window);
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());
(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 k1_loops = kN0 / kK1;
static_assert(1 <= k0_loops);
static_assert(1 <= k1_loops);
// main loop
auto loop_body = [&, seqlen_k_start_ = seqlen_k_start, seqlen_k_end_ = seqlen_k_end](
auto is_last_iteration) {
// STAGE 1, QK gemm
clear_tile(s_acc); // initialize C
if constexpr(k0_loops > 1)
{
static_for<0, k0_loops - 1, 1>{}([&](auto i_k0) {
async_load_tile_raw(k_lds_store(number<LdsSeq.at(number<i_k0 + 1>{})>{}),
k_dram_window);
if constexpr(i_k0 < k0_loops - 1)
move_tile_window(k_dram_window, {0, kK0});
async_load_fence(k_dram_window.get_num_access());
__builtin_amdgcn_s_barrier();
__builtin_amdgcn_sched_barrier(0);
gemm_0(s_acc,
get_slice_tile(
q, sequence<0, i_k0 * kK0>{}, sequence<kM0, (i_k0 + 1) * kK0>{}),
#if K_LDS_LOAD_USE_OFFSET_TRANSFORM
k_lds_load[number<LdsSeq.at(number<i_k0>{})>{}]);
#else
get_slice_tile(k_lds_load,
sequence<(LdsSeq.at(number<i_k0>{})) * kN0, 0>{},
sequence<(LdsSeq.at(number<i_k0>{}) + 1) * kN0, kK0>{}));
#endif
});
}
// TODO: this to fix a bug when loop smaller than 2,
// the following fence/barrier will be scheduled inside 1st loop
if constexpr(k0_loops <= 2)
__builtin_amdgcn_sched_barrier(0);
async_load_fence();
__builtin_amdgcn_s_barrier();
const auto bias_tile = load_tile(bias_dram_window); // load bias tile
auto v_buf = load_tile(v_dram_window, bool_constant<false>{});
__builtin_amdgcn_sched_barrier(0);
{ // tail
gemm_0(s_acc,
get_slice_tile(
q, sequence<0, (k0_loops - 1) * kK0>{}, sequence<kM0, k0_loops * kK0>{}),
#if K_LDS_LOAD_USE_OFFSET_TRANSFORM
k_lds_load[number<LdsSeq.at(number<k0_loops - 1>{})>{}]);
#else
get_slice_tile(
k_lds_load,
sequence<(LdsSeq.at(number<k0_loops - 1>{})) * kN0, 0>{},
sequence<(LdsSeq.at(number<k0_loops - 1>{}) + 1) * kN0, kK0>{}));
#endif
}
__builtin_amdgcn_sched_barrier(1);
// STAGE 2, scale_s, add bias, mask, softmax
if constexpr(BiasEnum == BlockAttentionBiasEnum::ELEMENTWISE_BIAS)
{
s_acc = tile_elementwise_in(s_acc_element_func, s_acc);
tile_elementwise_inout([&scale_s](auto& x) { x = x * scale_s; }, s_acc);
tile_elementwise_inout(
[&](auto& x, const auto& y) {
#if !CK_TILE_FMHA_FWD_FAST_EXP2
x += type_convert<SaccDataType>(bias_element_func(y));
#else
x += log2e_v<SaccDataType> *
type_convert<SaccDataType>(bias_element_func(y));
#endif
},
s_acc,
bias_tile);
}
else if constexpr(BiasEnum == BlockAttentionBiasEnum::ALIBI)
{
const auto k_origin = k_dram_block_window.get_window_origin();
constexpr auto s_spans = decltype(s_acc)::get_distributed_spans();
s_acc = tile_elementwise_in(s_acc_element_func, s_acc);
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.update(s_acc(i_j_idx), row, col);
});
});
}
else
{
s_acc = tile_elementwise_in(s_acc_element_func, s_acc);
#if !CK_TILE_FMHA_FWD_FAST_EXP2
tile_elementwise_inout([&scale_s](auto& x) { x = x * scale_s; }, s_acc);
#endif
}
move_tile_window(bias_dram_window, {0, kN0});
if constexpr(is_last_iteration)
{
const auto k_origin = k_dram_block_window.get_window_origin();
set_tile_if(s_acc, -numeric<SMPLComputeDataType>::infinity(), [&](auto tile_idx) {
const auto col = k_origin.at(number<0>{}) + tile_idx.at(number<1>{});
return seqlen_k_end_ <= col;
});
}
if constexpr(kPadSeqLenK || FmhaMask::IsMasking)
{
const auto k_origin = k_dram_block_window.get_window_origin();
bool need_perpixel_check = mask.IsEdgeTile(q_origin.at(number<0>{}),
k_origin.at(number<0>{}),
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>{});
return mask.IsOutOfBound(row, col);
});
}
}
const auto s = cast_tile<SMPLComputeDataType>(s_acc); // S{j}
auto m_local = block_tile_reduce<SMPLComputeDataType>(
s,
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.get_tile_distribution()); // Pcompute{j}
__builtin_amdgcn_sched_barrier(0x7F);
// store & prefetch next v, after the max reduction
if constexpr(std::is_same_v<VLayout, ck_tile::tensor_layout::gemm::RowMajor>)
{
auto v_shuffle_tmp = make_static_distributed_tensor<VDataType>(
Policy::template MakeShuffledVRegBlockDescriptor<Problem>());
shuffle_tile(v_shuffle_tmp, v_buf);
auto v_lds_window_tmp =
get_slice_tile(v_lds_window,
sequence<(LdsSeq.at(number<k0_loops>{})) * kN1, 0>{},
sequence<(LdsSeq.at(number<k0_loops>{}) + 1) * kN1, kK1>{});
store_tile(
v_lds_window_tmp,
tile_elementwise_in(v_element_func, v_shuffle_tmp)); // store the prefetch
}
else
{
auto v_lds_window_tmp =
get_slice_tile(v_lds_window,
sequence<(LdsSeq.at(number<k0_loops>{})) * kN1, 0>{},
sequence<(LdsSeq.at(number<k0_loops>{}) + 1) * kN1, kK1>{});
store_tile(v_lds_window_tmp,
tile_elementwise_in(v_element_func, v_buf)); // store the prefetch
}
if constexpr(k1_loops > 1)
{
move_tile_window(
v_dram_window,
{0, kK1}); // will have scratch if move this right after load_tile(v_dram)...
v_buf = load_tile(v_dram_window, bool_constant<false>{}); // load next v_buf
}
__builtin_amdgcn_sched_barrier(0);
static const auto get_validated_m = [](SMPLComputeDataType raw_m) {
/// NOTICE: bias might be materialized mask including -inf values, need
/// consideration. alibi does not have this problem
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[number<0>{}], [&](auto idx0) {
constexpr auto i_idx = make_tuple(idx0);
#if CK_TILE_FMHA_FWD_FAST_EXP2
auto row_max = scale_s * get_validated_m(m[i_idx]);
#endif
sweep_tile_span(p_spans[number<1>{}], [&](auto idx1) {
constexpr auto i_j_idx = make_tuple(idx0, idx1);
#if CK_TILE_FMHA_FWD_FAST_EXP2
if constexpr(BiasEnum == BlockAttentionBiasEnum::ELEMENTWISE_BIAS ||
BiasEnum == BlockAttentionBiasEnum::ALIBI)
{
p_compute(i_j_idx) = exp2(s[i_j_idx] - get_validated_m(m[i_idx]));
}
else
{
p_compute(i_j_idx) = exp2(scale_s * s[i_j_idx] - row_max);
}
#else
p_compute(i_j_idx) = exp(s[i_j_idx] - get_validated_m(m[i_idx]));
#endif
});
});
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>{});
// l{j}, Oacc{j}
constexpr auto o_spans = decltype(o_acc)::get_distributed_spans();
sweep_tile_span(o_spans[number<0>{}], [&](auto idx0) {
constexpr auto i_idx = make_tuple(idx0);
#if CK_TILE_FMHA_FWD_FAST_EXP2
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
{
auto row_max = scale_s * get_validated_m(m[i_idx]);
return exp2(scale_s * m_old[i_idx] - row_max);
}
}();
#else
const auto tmp = exp(m_old[i_idx] - get_validated_m(m[i_idx]));
#endif
l(i_idx) = tmp * l[i_idx] + rowsum_p[i_idx];
sweep_tile_span(o_spans[number<1>{}], [&](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;
});
});
if constexpr(kHasDropout)
{
auto randval_ptr =
reinterpret_cast<char*>(smem_ptr) + Policy::template GetSmemSizeKV<Problem>();
dropout.Run<decltype(gemm_0), SMPLComputeDataType, RandValOutputDataType>(
randval_ptr,
seqlen_k_start_ + i_total_loops * kN0,
p_compute,
randval_dram_window);
}
const auto p =
cast_tile<PDataType>(tile_elementwise_in(p_compute_element_func, p_compute));
// STAGE 3, KV gemm
if constexpr(k1_loops > 1)
{
static_for<0, k1_loops - 1, 1>{}([&](auto i_k1) {
if constexpr(i_k1 != 0 && i_k1 < k1_loops - 1)
{
v_buf = load_tile(v_dram_window, bool_constant<false>{}); // load next v_buf
}
block_sync_lds();
gemm_1(o_acc,
get_slice_tile(
p, sequence<0, i_k1 * kK1>{}, sequence<kM0, (i_k1 + 1) * kK1>{}),
get_slice_tile(
v_lds_window,
sequence<(LdsSeq.at(number<k0_loops + i_k1>{})) * kN1, 0>{},
sequence<(LdsSeq.at(number<k0_loops + i_k1>{}) + 1) * kN1, kK1>{}));
if constexpr(std::is_same_v<VLayout, ck_tile::tensor_layout::gemm::RowMajor>)
{
auto v_shuffle_tmp = make_static_distributed_tensor<VDataType>(
Policy::template MakeShuffledVRegBlockDescriptor<Problem>());
shuffle_tile(v_shuffle_tmp, v_buf);
auto v_lds_window_tmp = get_slice_tile(
v_lds_window,
sequence<(LdsSeq.at(number<k0_loops + i_k1 + 1>{})) * kN1, 0>{},
sequence<(LdsSeq.at(number<k0_loops + i_k1 + 1>{}) + 1) * kN1, kK1>{});
store_tile(v_lds_window_tmp,
tile_elementwise_in(v_element_func,
v_shuffle_tmp)); // store the prefetch
}
else
{
auto v_lds_window_tmp = get_slice_tile(
v_lds_window,
sequence<(LdsSeq.at(number<k0_loops + i_k1 + 1>{})) * kN1, 0>{},
sequence<(LdsSeq.at(number<k0_loops + i_k1 + 1>{}) + 1) * kN1, kK1>{});
store_tile(v_lds_window_tmp,
tile_elementwise_in(v_element_func, v_buf)); // store next v_buf
}
if constexpr(i_k1 < k1_loops - 1)
move_tile_window(v_dram_window, {0, kK1});
});
}
i_total_loops++;
if(i_total_loops < num_total_loop)
{
// move K tile windows
move_tile_window(k_dram_block_window, {kN0, 0});
k_dram_window =
make_tile_window(k_dram_block_window.get_bottom_tensor_view(),
k_dram_block_window.get_window_lengths(),
k_dram_block_window.get_window_origin(),
Policy::template MakeKDramTileDistribution<Problem>());
if constexpr(k1_loops >= 2 &&
LdsSeq.at(number<0>{}) == LdsSeq.at(number<k0_loops + k1_loops - 2>{}))
__builtin_amdgcn_s_barrier();
async_load_tile_raw(k_lds_store(LdsSeq.at(number<0>{})), k_dram_window);
move_tile_window(k_dram_window, {0, kK0});
}
// tail
{
block_sync_lds();
gemm_1(
o_acc,
get_slice_tile(p, sequence<0, (k1_loops - 1) * kK1>{}, sequence<kM0, kN0>{}),
get_slice_tile(
v_lds_window,
sequence<(LdsSeq.at(number<k0_loops + k1_loops - 1>{})) * kN1, 0>{},
sequence<(LdsSeq.at(number<k0_loops + k1_loops - 1>{}) + 1) * kN1, kK1>{}));
}
};
while(i_total_loops < num_total_loop - 1)
{
loop_body(std::false_type{});
}
loop_body(std::true_type{});
// 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[number<0>{}], [&, m_ = m, l_ = l](auto idx0) {
constexpr auto i_idx = make_tuple(idx0);
#if CK_TILE_FMHA_FWD_FAST_EXP2
if constexpr(BiasEnum == BlockAttentionBiasEnum::ELEMENTWISE_BIAS ||
BiasEnum == BlockAttentionBiasEnum::ALIBI)
{
lse_acc(i_idx) = m_[i_idx] * R_LOG2E + log(l_[i_idx]);
}
else
{
lse_acc(i_idx) = m_[i_idx] * scale_s * R_LOG2E + log(l_[i_idx]);
}
#else
lse_acc(i_idx) = m_[i_idx] + log(l_[i_idx]);
#endif
});
store_tile(lse_acc_dram_window_tmp, tile_elementwise_in(lse_acc_element_func, lse_acc));
// finally, O
constexpr auto o_spans = decltype(o_acc)::get_distributed_spans();
sweep_tile_span(o_spans[number<0>{}], [&](auto idx0) {
constexpr auto i_idx = make_tuple(idx0);
const auto tmp = [&]() {
if constexpr(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[number<1>{}], [&](auto idx1) {
constexpr auto i_j_idx = make_tuple(idx0, idx1);
o_acc(i_j_idx) *= tmp;
});
});
o_acc = tile_elementwise_in(o_acc_element_func, o_acc);
return o_acc;
}
template <typename QDramBlockWindowTmp,
typename KDramBlockWindowTmp,
typename VDramBlockWindowTmp,
typename BiasDramBlockWindowTmp,
typename RandValDramBlockWindowTmp,
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
RandValDramBlockWindowTmp& randval_dram_block_window_tmp, // M0*N0 tile
LSEaccDramBlockWindowTmp& lse_acc_dram_block_window_tmp, // M0*1 tile
FmhaMask mask,
PositionEncoding position_encoding,
float scale_s,
void* smem_ptr,
BlockDropout& dropout,
index_t i_split,
index_t num_splits) const
{
return operator()(q_dram_block_window_tmp,
identity{},
k_dram_block_window_tmp,
identity{},
v_dram_block_window_tmp,
identity{},
bias_dram_block_window_tmp,
identity{},
randval_dram_block_window_tmp,
lse_acc_dram_block_window_tmp,
identity{},
identity{},
identity{},
identity{},
mask,
position_encoding,
scale_s,
smem_ptr,
dropout,
i_split,
num_splits);
}
};
} // namespace ck_tile

19
include/ck_tile/ops/fmha/pipeline/block_fmha_fwd_splitkv_pipeline_qr_ks_vs_async_default_policy.hpp Normal file
View File

@@ -0,0 +1,19 @@
// 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/fmha/pipeline/block_fmha_pipeline_qx_ks_vs_custom_policy.hpp"
namespace ck_tile {
// This pipeline is qkv all located in LDS
using BlockFmhaFwdSplitKVPipelineQRKSVSAsyncDefaultPolicy =
BlockFmhaPipelineQXKSVSCustomPolicy</* QLoadOnce = */ true,
/* AsyncCopyK = */ true,
/* AsyncCopyV = */ false,
/* NumPrefetchK = */ 3,
/* NumPrefetchV = */ 3>;
} // namespace ck_tile

19
include/ck_tile/ops/fmha/pipeline/block_fmha_fwd_splitkv_pipeline_qr_ks_vs_default_policy.hpp Normal file
View File

@@ -0,0 +1,19 @@
// 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/fmha/pipeline/block_fmha_pipeline_qx_ks_vs_custom_policy.hpp"
namespace ck_tile {
// This pipeline is qkv all located in LDS
using BlockFmhaFwdSplitKVPipelineQRKSVSDefaultPolicy =
BlockFmhaPipelineQXKSVSCustomPolicy</* QLoadOnce = */ true,
/* AsyncCopyK = */ false,
/* AsyncCopyV = */ false,
/* NumPrefetchK = */ 1,
/* NumPrefetchV = */ 1>;
} // namespace ck_tile

10
include/ck_tile/ops/fmha/pipeline/tile_fmha_traits.hpp
View File

@@ -32,6 +32,16 @@ struct TileFmhaTraits
static constexpr index_t kBlockPerCu = kBlockPerCu_;
};
template <bool kPadSeqLenQ_ /* padding for seqlen_q */,
bool kPadHeadDimV_ /* paddding for hdim_v */,
index_t kBlockPerCu_ = 2 /* hint to occupancy */>
struct TileFmhaFwdSplitKVCombineTraits
{
static constexpr bool kPadSeqLenQ = kPadSeqLenQ_;
static constexpr bool kPadHeadDimV = kPadHeadDimV_;
static constexpr index_t kBlockPerCu = kBlockPerCu_;
};
template <bool kPadSeqLenQ_ /* padding for seqlen_q */,
bool kPadHeadDimV_ /* paddding for hdim_v */,
index_t kBlockPerCu_ = 2 /* hint to occupancy */>