Skip to content

GitLab

Commit 7ea01f4d authored by AMD-dteng's avatar AMD-dteng
Browse files

1. enable bias feature that add bias before adding residual; 2. change block... 

1. enable bias feature that add bias before adding residual; 2. change block size from 128->64 when m<64 in fp16
parent 357a0b1c
Hide whitespace changes
Inline Side-by-side
Showing with 202 additions and 64 deletions
example/ck_tile/01_fmha/codegen/ops/fmha_fwd.py
View file @ 7ea01f4d
......@@ -410,7 +410,8 @@ def get_fmha_fwd_tile_dict_from_dtype(dtype : str) -> Optional[dict]:
if dtype == 'fp16' or dtype == 'bf16':
return {
'32' : FmhaFwdTileSize(128, 64, 16, 32, 32, 32, 2, 1, 1, 2, 1, 1, 32, 32, 16, -1),
'64' : FmhaFwdTileSize(128, 64, 32, 64, 32, 64, 4, 1, 1, 4, 1, 1, 32, 32, 16, -1),
'64' : FmhaFwdTileSize(64, 64, 32, 64, 32, 64, 4, 1, 1, 4, 1, 1, 16, 16, 16, -1),
#'64' : FmhaFwdTileSize(128, 64, 32, 64, 32, 64, 4, 1, 1, 4, 1, 1, 32, 32, 16, -1),
## '96' : FmhaFwdTileSize(128, 128, 32, 128, 32, 96, 4, 1, 1, 4, 1, 1, 32, 32, 16, -1),
'128' : FmhaFwdTileSize(128, 128, 32, 128, 32, 128, 4, 1, 1, 4, 1, 1, 32, 32, 16, -1),
'256' : FmhaFwdTileSize(128, 128, 32, 256, 32, 256, 4, 1, 1, 4, 1, 1, 32, 32, 16, -1),
......
example/ck_tile/02_layernorm2d/generate.py
View file @ 7ea01f4d
......@@ -23,6 +23,10 @@ def get_if_str(idx, total, lase_else = True):
else:
return 'else if'
BIAS_ENUM_STR_MAP = [
'no',
'bias'] # pre-norm add bias
FUSED_ADD_ENUM_STR_MAP = [
'no',
'pras', # pre-norm
......@@ -59,6 +63,7 @@ template <typename XDataType_,
bool kSaveMeanInvStd_,
bool kFastFDiv_,
bool kTwoPass_,
ck_tile::index_t kBias_ = 0,
ck_tile::index_t kFusedAdd_ = 0,
ck_tile::index_t kFusedQuant_ = 0>
struct layernorm2d_fwd_traits_
......@@ -121,6 +126,7 @@ struct layernorm2d_fwd_traits_
static constexpr bool kSaveMeanInvStd = kSaveMeanInvStd_;
static constexpr bool kFastFDiv = kFastFDiv_;
static constexpr bool kTwoPass = kTwoPass_;
static constexpr ck_tile::index_t kBias = kBias_;
static constexpr ck_tile::index_t kFusedAdd = kFusedAdd_;
static constexpr ck_tile::index_t kFusedQuant = kFusedQuant_;
};
......@@ -138,6 +144,7 @@ template <typename XDataType_,
bool kSaveMeanInvStd_,
bool kFastFDiv_,
bool kTwoPass_,
int kBias_,
int kFusedAdd_,
int kFusedQuant_>
using traits_ = layernorm2d_fwd_traits_<XDataType_,
......@@ -153,6 +160,7 @@ using traits_ = layernorm2d_fwd_traits_<XDataType_,
kSaveMeanInvStd_,
kFastFDiv_,
kTwoPass_,
kBias_,
kFusedAdd_,
kFusedQuant_>;
"""
......@@ -185,10 +193,12 @@ float layernorm2d_fwd_(const S& s, A a)
Traits_::kSaveMeanInvStd,
Traits_::kFastFDiv,
Traits_::kTwoPass,
static_cast<ck_tile::Layernorm2dBiasEnum>(Traits_::kBias),
static_cast<ck_tile::Layernorm2dFusedAddEnum>(Traits_::kFusedAdd),
static_cast<ck_tile::Layernorm2dFusedQuantEnum>(Traits_::kFusedQuant)>;
using PipelineProblem = ck_tile::Layernorm2dFwdPipelineProblem<
typename LayerNormTypeConfig<XDataType, YDataType, XScaleDataType, YScaleDataType>::XDataType,
typename LayerNormTypeConfig<XDataType, YDataType, XScaleDataType, YScaleDataType>::BiasDataType,
typename LayerNormTypeConfig<XDataType, YDataType, XScaleDataType, YScaleDataType>::GammaDataType,
typename LayerNormTypeConfig<XDataType, YDataType, XScaleDataType, YScaleDataType>::BetaDataType,
typename LayerNormTypeConfig<XDataType, YDataType, XScaleDataType, YScaleDataType>::ComputeDataType,
......@@ -274,7 +284,7 @@ float layernorm2d_fwd(layernorm2d_fwd_traits t,
#include "layernorm2d_fwd_api_common.hpp"
// clang-format off
// prec_i prec_o prec_sy rm rn tm tn vn pd mv rpcf 2p add sweep
// prec_i prec_o prec_sy rm rn tm tn vn pd mv rpcf 2p bias add sweep
{F_instance_def}
// clang-format on
......@@ -284,6 +294,10 @@ float layernorm2d_fwd(layernorm2d_fwd_traits t,
self.working_path = working_path
self.kernel_filter = kernel_filter
class k_bias_enum(IntEnum):
F_NO_BIAS = 0
F_ADD_BIAS = 1
class k_fuesd_add_enum(IntEnum):
F_NO_ADD = 0
F_PRE_ADD = 1
......@@ -299,6 +313,7 @@ float layernorm2d_fwd(layernorm2d_fwd_traits t,
F_kPadN : bool
F_kSaveMeanInvStd : bool
F_kTwoPass : bool
F_kBias : Any #: layernorm_fwd_codegen.k_bias_enum
F_kFusedAdd : Any #: layernorm_fwd_codegen.k_fuesd_add_enum
F_kFusedQuant : Any #: layernorm_fwd_codegen.k_fused_sweep_enum
......@@ -315,6 +330,7 @@ float layernorm2d_fwd(layernorm2d_fwd_traits t,
@dataclass
class k_problem:
F_XDataType : str
F_BiasDataType : str
F_GammaDataType : str
F_BetaDataType : str
F_ComputeDataType : str
......@@ -363,6 +379,7 @@ float layernorm2d_fwd(layernorm2d_fwd_traits t,
F_kSaveMeanInvStd_ : bool
F_kFastFDiv_ : bool
F_kTwoPass_ : bool
F_kBias_ : int
F_kFusedAdd : int
F_kFusedQuant : int
......@@ -370,7 +387,7 @@ float layernorm2d_fwd(layernorm2d_fwd_traits t,
def trait_name(self) ->str:
t_ = f'{DATA_TYPE_MAP[self.F_XDataType]}, {DATA_TYPE_MAP[self.F_YDataType]}, {DATA_TYPE_MAP[self.F_XScaleDataType]}, {DATA_TYPE_MAP[self.F_YScaleDataType]}, {self.F_Repeat_M:2}, {self.F_Repeat_N:2}, {self.F_ThreadPerBlock_M:2}, {self.F_ThreadPerBlock_N:4}'
t_ += f', {self.F_Vector_N:2}, {BOOL_MAP(self.F_kPadN):5}, {BOOL_MAP(self.F_kSaveMeanInvStd_):5}, {BOOL_MAP(self.F_kFastFDiv_):5}'
t_ += f', {BOOL_MAP(self.F_kTwoPass_):5}, {self.F_kFusedAdd:4}, {self.F_kFusedQuant:4}'
t_ += f', {BOOL_MAP(self.F_kTwoPass_):5}, {self.F_kBias:4}, {self.F_kFusedAdd:4}, {self.F_kFusedQuant:4}'
return t_
# string when calling this kernel
......@@ -388,6 +405,7 @@ float layernorm2d_fwd(layernorm2d_fwd_traits t,
class h_instance:
F_DataTypePair : str
F_N : str
F_bias : int
F_add : int
F_sweep : int
instance_list : List[Any] # List[h_traits]
......@@ -397,6 +415,8 @@ float layernorm2d_fwd(layernorm2d_fwd_traits t,
prec_i, prec_o = self.F_DataTypePair.split(',')
dtype_str = f'{prec_i}' if prec_i == prec_o else f'{prec_i}_{prec_o}'
nnn = f'layernorm2d_fwd_{dtype_str}_n{self.F_N}'
if self.F_bias != 0:
nnn = nnn + '_' + BIAS_ENUM_STR_MAP[self.F_bias]
if self.F_add != 0:
nnn = nnn + '_' + FUSED_ADD_ENUM_STR_MAP[self.F_add]
if self.F_sweep != 0:
......@@ -456,8 +476,8 @@ float layernorm2d_fwd(layernorm2d_fwd_traits t,
elif ins.F_kFusedQuant == 2:
_sweep_cond = 't.fused_quant == {f_fused_sweep} && (t.prec_sy == \"{f_sy_type}\")'.format(
f_fused_sweep = ins.F_kFusedQuant, f_sy_type=ins.F_YScaleDataType)
_cond = '((a.n % {f_vec_n} == 0) && (t.fused_add == {f_fused_add}) && ({f_sweep_cond}))'.format(
f_vec_n = ins.F_Vector_N, f_fused_add = ins.F_kFusedAdd,
_cond = '((a.n % {f_vec_n} == 0) && (t.bias == {f_bias}) && (t.fused_add == {f_fused_add}) && ({f_sweep_cond}))'.format(
f_vec_n = ins.F_Vector_N, f_bias = ins.F_kBias, f_fused_add = ins.F_kFusedAdd,
f_sweep_cond = _sweep_cond)
inner_str += self.API_INNER_CASE.format(F_if = get_if_str(idx_in_n, len_in_n, False),
F_VEC_COND = _cond, F_instance_func=ins.call_name)
......@@ -484,65 +504,67 @@ float layernorm2d_fwd(layernorm2d_fwd_traits t,
scale_list = [('fp32,fp32')]
dtype_list = [('fp16,fp16'), ('bf16,bf16'),
('fp16,int8'), ('bf16,int8')] # NOTE: only fused-dynamic-quant use int8 out
#bias_list = [0, 1]
#fused_add_list = [0, 1, 2]
#fused_sweep_list = [0, 1, 2] # NOTE: only single pass can use fused dynamic quant
bias_list = [0, 1]
fused_add_list = [0, 1]
fused_sweep_list = [0, 1] # NOTE: only single pass can use fused dynamic quant
# rm rn tm tn vn pd mv fdiv 2p add sweep
h_trait_dict = {'64' : [ h_traits('x', 'y', 'xs', 'ys', 1, 1, 8, 8, 8, True, False, True, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 1, 4, 16, 4, True, False, True, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 1, 4, 64, 1, True, False, True, False, 0, 0)],
'128' : [ h_traits('x', 'y', 'xs', 'ys', 1, 1, 4, 16, 8, True, False, True, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 1, 4, 64, 2, True, False, True, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 2, 4, 64, 1, True, False, True, False, 0, 0)],
'256' : [ h_traits('x', 'y', 'xs', 'ys', 1, 1, 4, 64, 4, True, False, True, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 2, 4, 64, 2, True, False, True, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 4, 4, 64, 1, True, False, True, False, 0, 0)],
'512' : [ h_traits('x', 'y', 'xs', 'ys', 1, 1, 4, 64, 8, True, False, True, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 2, 4, 64, 4, True, False, True, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 4, 4, 64, 2, True, False, True, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 8, 4, 64, 1, True, False, True, False, 0, 0)],
'768' : [ h_traits('x', 'y', 'xs', 'ys', 1, 3, 4, 64, 4, True, False, True, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 6, 4, 64, 2, True, False, True, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 12, 4, 64, 1, True, False, True, False, 0, 0)],
'1024' :[ h_traits('x', 'y', 'xs', 'ys', 1, 1, 2, 128, 8, True, False, True, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 2, 2, 128, 4, True, False, True, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 4, 2, 128, 2, True, False, True, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 4, 1, 256, 1, True, False, True, False, 0, 0)],
'1536' :[ h_traits('x', 'y', 'xs', 'ys', 1, 3, 4, 64, 8, True, False, True, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 3, 2, 128, 4, True, False, True, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 3, 1, 256, 2, True, False, True, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 6, 1, 256, 1, True, False, True, False, 0, 0)],
'2048' :[ h_traits('x', 'y', 'xs', 'ys', 1, 1, 1, 256, 8, True, False, True, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 2, 1, 256, 4, True, False, True, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 4, 1, 256, 2, True, False, True, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 8, 1, 256, 1, True, False, True, False, 0, 0)],
'3072' :[ h_traits('x', 'y', 'xs', 'ys', 1, 3, 1, 128, 8, True, False, True, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 3, 1, 256, 4, True, False, True, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 6, 1, 256, 2, True, False, True, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 3, 1,1024, 1, True, False, True, False, 0, 0)],
'4096' :[ h_traits('x', 'y', 'xs', 'ys', 1, 2, 1, 256, 8, True, False, True, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 4, 1, 256, 4, True, False, True, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 2, 1,1024, 2, True, False, True, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 4, 1,1024, 1, True, False, True, False, 0, 0)],
'6144' :[ h_traits('x', 'y', 'xs', 'ys', 1, 3, 1, 256, 8, True, False, True, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 3, 1, 512, 4, True, False, True, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 3, 1,1024, 2, True, False, True, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 6, 1,1024, 1, True, False, True, False, 0, 0)],
'8192' :[ h_traits('x', 'y', 'xs', 'ys', 1, 4, 1, 256, 8, True, False, True, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 4, 1, 512, 4, True, False, True, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 4, 1,1024, 2, True, False, True, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 8, 1,1024, 1, True, False, True, False, 0, 0)],
'big' :[ h_traits('x', 'y', 'xs', 'ys', 1, 2, 1, 256, 8, True, False, True, True, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 4, 1, 256, 4, True, False, True, True, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 2, 1,1024, 2, True, False, True, True, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 4, 1,1024, 1, True, False, True, True, 0, 0)]}
# rm rn tm tn vn pd mv fdiv 2p bias add sweep
h_trait_dict = {'64' : [ h_traits('x', 'y', 'xs', 'ys', 1, 1, 8, 8, 8, True, False, True, False, 0, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 1, 4, 16, 4, True, False, True, False, 0, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 1, 4, 64, 1, True, False, True, False, 0, 0, 0)],
'128' : [ h_traits('x', 'y', 'xs', 'ys', 1, 1, 4, 16, 8, True, False, True, False, 0, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 1, 4, 64, 2, True, False, True, False, 0, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 2, 4, 64, 1, True, False, True, False, 0, 0, 0)],
'256' : [ h_traits('x', 'y', 'xs', 'ys', 1, 1, 4, 64, 4, True, False, True, False, 0, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 2, 4, 64, 2, True, False, True, False, 0, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 4, 4, 64, 1, True, False, True, False, 0, 0, 0)],
'512' : [ h_traits('x', 'y', 'xs', 'ys', 1, 1, 4, 64, 8, True, False, True, False, 0, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 2, 4, 64, 4, True, False, True, False, 0, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 4, 4, 64, 2, True, False, True, False, 0, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 8, 4, 64, 1, True, False, True, False, 0, 0, 0)],
'768' : [ h_traits('x', 'y', 'xs', 'ys', 1, 3, 4, 64, 4, True, False, True, False, 0, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 6, 4, 64, 2, True, False, True, False, 0, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 12, 4, 64, 1, True, False, True, False, 0, 0, 0)],
'1024' :[ h_traits('x', 'y', 'xs', 'ys', 1, 1, 2, 128, 8, True, False, True, False, 0, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 2, 2, 128, 4, True, False, True, False, 0, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 4, 2, 128, 2, True, False, True, False, 0, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 4, 1, 256, 1, True, False, True, False, 0, 0, 0)],
'1536' :[ h_traits('x', 'y', 'xs', 'ys', 1, 3, 4, 64, 8, True, False, True, False, 0, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 3, 2, 128, 4, True, False, True, False, 0, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 3, 1, 256, 2, True, False, True, False, 0, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 6, 1, 256, 1, True, False, True, False, 0, 0, 0)],
'2048' :[ h_traits('x', 'y', 'xs', 'ys', 1, 1, 1, 256, 8, True, False, True, False, 0, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 2, 1, 256, 4, True, False, True, False, 0, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 4, 1, 256, 2, True, False, True, False, 0, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 8, 1, 256, 1, True, False, True, False, 0, 0, 0)],
'3072' :[ h_traits('x', 'y', 'xs', 'ys', 1, 3, 1, 128, 8, True, False, True, False, 0, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 3, 1, 256, 4, True, False, True, False, 0, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 6, 1, 256, 2, True, False, True, False, 0, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 3, 1,1024, 1, True, False, True, False, 0, 0, 0)],
'4096' :[ h_traits('x', 'y', 'xs', 'ys', 1, 2, 1, 256, 8, True, False, True, False, 0, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 4, 1, 256, 4, True, False, True, False, 0, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 2, 1,1024, 2, True, False, True, False, 0, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 4, 1,1024, 1, True, False, True, False, 0, 0, 0)],
'6144' :[ h_traits('x', 'y', 'xs', 'ys', 1, 3, 1, 256, 8, True, False, True, False, 0, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 3, 1, 512, 4, True, False, True, False, 0, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 3, 1,1024, 2, True, False, True, False, 0, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 6, 1,1024, 1, True, False, True, False, 0, 0, 0)],
'8192' :[ h_traits('x', 'y', 'xs', 'ys', 1, 4, 1, 256, 8, True, False, True, False, 0, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 4, 1, 512, 4, True, False, True, False, 0, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 4, 1,1024, 2, True, False, True, False, 0, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 8, 1,1024, 1, True, False, True, False, 0, 0, 0)],
'big' :[ h_traits('x', 'y', 'xs', 'ys', 1, 2, 1, 256, 8, True, False, True, True, 0, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 4, 1, 256, 4, True, False, True, True, 0, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 2, 1,1024, 2, True, False, True, True, 0, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 4, 1,1024, 1, True, False, True, True, 0, 0, 0)]}
total_blob = list()
for hs_key in h_trait_dict:
hs = h_trait_dict[hs_key]
current_n = hs[0].F_Repeat_N * hs[0].F_ThreadPerBlock_N * hs[0].F_Vector_N
for dtype, scale_type, fused_add, fused_quant in itertools.product(dtype_list, scale_list, fused_add_list, fused_sweep_list):
for dtype, scale_type, bias, fused_add, fused_quant in itertools.product(dtype_list, scale_list, bias_list, fused_add_list, fused_sweep_list):
prec_i, prec_o = dtype.split(',')
scale_x, scale_y = scale_type.split(',')
if prec_o in dynamic_quant_out_dtype and fused_quant != 1:
......@@ -556,12 +578,13 @@ float layernorm2d_fwd(layernorm2d_fwd_traits t,
h_.F_YDataType = prec_o
h_.F_XScaleDataType = scale_y
h_.F_YScaleDataType = scale_x
h_.F_kBias = bias
h_.F_kFusedAdd = fused_add
h_.F_kFusedQuant = fused_quant
current_hs.append(h_) # + "\n"
#f.write(str(f.parent / GEN_DIR / (blobs.api_common_header_
current_n_str = 'big' if hs_key == 'big' else current_n
total_blob.append(h_instance(dtype, current_n_str, fused_add, fused_quant, current_hs))
total_blob.append(h_instance(dtype, current_n_str, bias, fused_add, fused_quant, current_hs))
return total_blob
def list_blobs(self) -> None:
......
example/ck_tile/02_layernorm2d/layernorm2d_fwd.cpp
View file @ 7ea01f4d
......@@ -41,6 +41,7 @@ auto create_args(int argc, char* argv[])
.insert("prec_sy",
"auto",
"output quant scale type, set auto will use fp32. used when fquant=1 or 2")
.insert("bias", "0", "add bias, 0:no add, 1:add bias before fadd")
.insert("fadd", "0", "fused-add, 0:no fused add, 1:preadd+store, 2:preadd only")
.insert("fquant", "0", "fused-quant, 0:no, 1:smooth-dynamic-quant, 2:dynamic-quant")
.insert("warmup", "5", "cold iter")
......@@ -93,6 +94,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
int do_validation = arg_parser.get_int("v");
int warmup = arg_parser.get_int("warmup");
int repeat = arg_parser.get_int("repeat");
int bias = arg_parser.get_int("bias");
int fused_add = arg_parser.get_int("fadd");
int fused_quant = arg_parser.get_int("fquant");
if(fused_quant == 1 && prec_o != "int8")
......@@ -107,6 +109,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
using XDataType = typename TypeConfig::XDataType;
using YDataType = typename TypeConfig::YDataType;
using BiasDataType = typename TypeConfig::BiasDataType;
using GammaDataType = typename TypeConfig::GammaDataType;
using BetaDataType = typename TypeConfig::BetaDataType;
using XResidualDataType = XDataType;
......@@ -121,6 +124,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
// host verify
ck_tile::HostTensor<XDataType> x_host({m, n}, {x_stride, 1});
ck_tile::HostTensor<BiasDataType> bias_host({n});
ck_tile::HostTensor<GammaDataType> gamma_host({n});
ck_tile::HostTensor<BetaDataType> beta_host({n});
......@@ -141,10 +145,12 @@ bool run(const ck_tile::ArgParser& arg_parser)
ck_tile::FillUniformDistribution<XDataType>{-.5f, .5f}(x_host);
ck_tile::FillUniformDistribution<XResidualDataType>{-.5f, .5f}(x_residual_host);
ck_tile::FillUniformDistribution<XScaleDataType>{-1.f, 1.f}(x_scale_host);
ck_tile::FillUniformDistribution<BiasDataType>{-.5f, .5f}(bias_host);
ck_tile::FillUniformDistribution<GammaDataType>{-.5f, .5f}(gamma_host);
ck_tile::FillUniformDistribution<BetaDataType>{-.5f, .5f}(beta_host);
ck_tile::DeviceMem x_buf(x_host.get_element_space_size_in_bytes());
ck_tile::DeviceMem bias_buf(bias_host.get_element_space_size_in_bytes());
ck_tile::DeviceMem gamma_buf(gamma_host.get_element_space_size_in_bytes());
ck_tile::DeviceMem beta_buf(beta_host.get_element_space_size_in_bytes());
ck_tile::DeviceMem y_buf(y_host_dev.get_element_space_size_in_bytes());
......@@ -155,6 +161,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
ck_tile::DeviceMem y_residual_buf(y_residual_host.get_element_space_size_in_bytes());
x_buf.ToDevice(x_host.data());
bias_buf.ToDevice(bias_host.data());
gamma_buf.ToDevice(gamma_host.data());
beta_buf.ToDevice(beta_host.data());
x_residual_buf.ToDevice(x_residual_host.data());
......@@ -179,11 +186,12 @@ bool run(const ck_tile::ArgParser& arg_parser)
<< ", yr_stride:" << yr_stride << std::flush;
layernorm2d_fwd_traits traits{
prec_i, prec_o, prec_sx, prec_sy, SaveMeanVar, fused_add, fused_quant};
prec_i, prec_o, prec_sx, prec_sy, SaveMeanVar, bias, fused_add, fused_quant};
layernorm2d_fwd_args args{x_buf.GetDeviceBuffer(),
fused_add != 0 ? x_residual_buf.GetDeviceBuffer() : nullptr,
fused_quant == 1 ? x_scale_buf.GetDeviceBuffer() : nullptr,
bias_buf.GetDeviceBuffer(),
gamma_buf.GetDeviceBuffer(),
beta_buf.GetDeviceBuffer(),
......@@ -210,7 +218,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
return false;
}
std::size_t num_byte = sizeof(XDataType) * m * n + sizeof(GammaDataType) * n +
std::size_t num_byte = sizeof(XDataType) * m * n + sizeof(BiasDataType) * n + sizeof(GammaDataType) * n +
sizeof(BetaDataType) * n + sizeof(YDataType) * m * n;
float gb_per_sec = num_byte / 1.E6 / ave_time;
......@@ -223,6 +231,19 @@ bool run(const ck_tile::ArgParser& arg_parser)
// reference
if(fused_add != 0)
{
if(bias != 0)
{
// add bias before fadd
int M = x_host.mDesc.get_lengths()[0];
int N = x_host.mDesc.get_lengths()[1];
for(int idx_m = 0; idx_m < M; ++idx_m)
{
for(int idx_n = 0; idx_n < N; ++idx_n)
{
x_host(idx_m, idx_n) = ck_tile::type_convert<XDataType>(ck_tile::type_convert<ComputeDataType>(x_host(idx_m, idx_n)) + ck_tile::type_convert<ComputeDataType>(bias_host(idx_n)));
}
}
}
// fused pre_add/pre_add_store
// TODO we accumulate directly to x_host for simplcity here...
......
example/ck_tile/02_layernorm2d/layernorm2d_fwd.hpp
View file @ 7ea01f4d
......@@ -16,6 +16,7 @@ struct LayerNormTypeConfig<ck_tile::half_t, OutType, XScaleDataType_, YScaleData
{
using XDataType = ck_tile::half_t;
using YDataType = OutType;
using BiasDataType = ck_tile::half_t;
using GammaDataType = ck_tile::half_t;
using BetaDataType = ck_tile::half_t;
using MeanDataType = ck_tile::half_t;
......@@ -30,6 +31,7 @@ struct LayerNormTypeConfig<ck_tile::bf16_t, OutType, XScaleDataType_, YScaleData
{
using XDataType = ck_tile::bf16_t;
using YDataType = OutType;
using BiasDataType = ck_tile::bf16_t;
using GammaDataType = ck_tile::bf16_t;
using BetaDataType = ck_tile::bf16_t;
using MeanDataType = ck_tile::bf16_t;
......@@ -57,6 +59,7 @@ struct layernorm2d_fwd_traits
std::string prec_sy; // y-scale, used for [M*1] output for next layer
bool save_mean_var; //
int bias; // 0:no-bias, 1:add bias
int fused_add; // 0:no-add, 1:pre-add-store, 2:pre-add
int fused_quant; // 0:no-sweep, 1:smooth-dynamic-quant, 2:dynamic-quant
};
......
include/ck_tile/host/reference/reference_layernorm2d_fwd.hpp
View file @ 7ea01f4d
......@@ -59,6 +59,11 @@ void reference_layernorm2d_fwd(const HostTensor<XDataType>& x_m_n,
{
++count;
ComputeDataType x = ck_tile::type_convert<ComputeDataType>(x_m_n(m, n));
// printf("dteng_ck_print::x_pre[%lu][%d] = %f\n",m,n,x);
// printf("dteng_ck_print::gamma_n[0][%d] = %f\n",n,ck_tile::type_convert<ComputeDataType>(gamma_n(n)));
// x = x + ck_tile::type_convert<ComputeDataType>(bias_n(n));
// x_m_n(m, n) = ck_tile::type_convert<XDataType>(x);
// printf("dteng_ck_print::x_post[%lu][%d] = %f\n",m,n,x);
ComputeDataType delta = x - mean;
mean += delta / count;
ComputeDataType delta2 = x - mean;
......@@ -85,6 +90,10 @@ void reference_layernorm2d_fwd(const HostTensor<XDataType>& x_m_n,
a_ = a_ * gamma + beta;
acc(m, n) = a_;
// printf("dteng_ck_print::mean[%lu][%d] = %f\n",m,n,mean);
// printf("dteng_ck_print::gamma[%lu][%d] = %f\n",m,n,gamma);
// printf("dteng_ck_print::beta[%lu][%d] = %f\n",m,n,beta);
// printf("dteng_ck_print::acc[%lu][%d] = %f\n",m,n,a_);
}
epilogue_functor(m, y_m_n, acc);
......
include/ck_tile/ops/layernorm2d/kernel/layernorm2d_fwd_kernel.hpp
View file @ 7ea01f4d
......@@ -15,6 +15,7 @@ struct Layernorm2dFwdHostArgs
const void* p_x; // [m ,n], input, fp16/bf16
const void* p_x_residual; // [m ,n], shortcut input, prec same as input, nullptr if not used
const void* p_x_scale; // [1 ,n], smooth scale input, fp32, nullptr if not used
const void* p_bias; // [1, n], bias, prec same as input
const void* p_gamma; // [1, n], gamma, prec same as input
const void* p_beta; // [1, n], beta, prec same as input
......@@ -43,6 +44,7 @@ struct Layernorm2dFwd
using Problem = typename Pipeline::Problem;
using XDataType = remove_cvref_t<typename Problem::XDataType>;
using BiasDataType = remove_cvref_t<typename Problem::BiasDataType>;
using GammaDataType = remove_cvref_t<typename Problem::GammaDataType>;
using BetaDataType = remove_cvref_t<typename Problem::BetaDataType>;
using ComputeDataType = remove_cvref_t<typename Problem::ComputeDataType>;
......@@ -67,6 +69,7 @@ struct Layernorm2dFwd
static constexpr bool kPadM = false; // always no need to pad along M
static constexpr bool kPadN = Problem::Traits::kPadN;
static constexpr bool kTwoPass = Problem::Traits::kTwoPass;
static constexpr auto kBias = Problem::Traits::kBias;
static constexpr auto kFusedAdd = Problem::Traits::kFusedAdd;
static constexpr auto kFusedQuant = Problem::Traits::kFusedQuant;
......@@ -82,6 +85,7 @@ struct Layernorm2dFwd
const void* p_x; // [m ,n], input, fp16/bf16
const void* p_x_residual; // [m ,n], shortcut input, prec same as input, nullptr if not used
const void* p_x_scale; // [1 ,n], smooth scale input, fp32, nullptr if not used
const void* p_bias; // [1, n], bias, prec same as input
const void* p_gamma; // [1, n], gamma, prec same as input
const void* p_beta; // [1, n], beta, prec same as input
......@@ -108,6 +112,7 @@ struct Layernorm2dFwd
return Kargs{hargs.p_x,
hargs.p_x_residual,
hargs.p_x_scale,
hargs.p_bias,
hargs.p_gamma,
hargs.p_beta,
hargs.p_y,
......@@ -152,6 +157,7 @@ struct Layernorm2dFwd
using S_ = typename Problem::BlockShape;
auto surfix = [&] () {
std::string n;
if (kBias != Layernorm2dBiasEnum::NO_BIAS) n += _SS_("_") + Layernorm2dBiasEnumName<kBias>::name;
if (kFusedAdd != Layernorm2dFusedAddEnum::NO_ADD) n += _SS_("_") + Layernorm2dFusedAddEnumName<kFusedAdd>::name;
if (kFusedQuant != Layernorm2dFusedQuantEnum::NO_SWEEP) n += _SS_("_") + Layernorm2dFusedQuantEnumName<kFusedQuant>::name;
if (kPadN) n += "_pn";
......@@ -228,6 +234,27 @@ struct Layernorm2dFwd
}
}();
const auto bias_window = [&]() {
if constexpr(kBias == Layernorm2dBiasEnum::ADD_BIAS)
{
const auto tmp_ = make_naive_tensor_view<address_space_enum::global>(
static_cast<const BiasDataType*>(kargs.p_bias),
make_tuple(kargs.n),
make_tuple(1),
number<Vector_N>{},
number<1>{});
const auto tmp2_ =
pad_tensor_view(tmp_, make_tuple(number<Block_N>{}), sequence<false>{});
return make_tile_window(tmp2_, make_tuple(number<Block_N>{}), {0});
}
else
{
return make_null_tile_window(make_tuple(number<Block_N>{}));
}
}();
const auto gamma_window = [&]() {
const auto tmp_ = make_naive_tensor_view<address_space_enum::global>(
static_cast<const GammaDataType*>(kargs.p_gamma),
......@@ -371,6 +398,7 @@ struct Layernorm2dFwd
Pipeline{}(x_window,
x_residual_window,
bias_window,
gamma_window,
beta_window,
y_window,
......
include/ck_tile/ops/layernorm2d/pipeline/layernorm2d_fwd_pipeline_one_pass.hpp
View file @ 7ea01f4d
......@@ -18,6 +18,7 @@ struct Layernorm2dFwdPipelineOnePass
using Policy = ck_tile::remove_cvref_t<Policy_>;
using XDataType = ck_tile::remove_cvref_t<typename Problem::XDataType>;
using BiasDataType = ck_tile::remove_cvref_t<typename Problem::BiasDataType>;
using GammaDataType = ck_tile::remove_cvref_t<typename Problem::GammaDataType>;
using BetaDataType = ck_tile::remove_cvref_t<typename Problem::BetaDataType>;
using ComputeDataType = ck_tile::remove_cvref_t<typename Problem::ComputeDataType>;
......@@ -37,6 +38,7 @@ struct Layernorm2dFwdPipelineOnePass
static constexpr bool kPadM = false; // TODO - BlockLayernorm2dFwdProblem::kPadM
static constexpr bool kPadN = Problem::Traits::kPadN;
static constexpr bool kFastFDiv = Problem::Traits::kFastFDiv;
static constexpr auto kBias = Problem::Traits::kBias;
static constexpr auto kFusedAdd = Problem::Traits::kFusedAdd;
static constexpr auto kFusedQuant = Problem::Traits::kFusedQuant;
......@@ -54,6 +56,7 @@ struct Layernorm2dFwdPipelineOnePass
template <typename XWindow,
typename XResidualWindow,
typename BiasWindow,
typename GammaWindow,
typename BetaWindow,
typename YWindow,
......@@ -65,6 +68,7 @@ struct Layernorm2dFwdPipelineOnePass
typename Epilogue>
CK_TILE_DEVICE auto operator()(const XWindow& x_window_,
const XResidualWindow& x_residual_window_,
const BiasWindow& bias_window_,
const GammaWindow& gamma_window_,
const BetaWindow& beta_window_,
YWindow& y_window_,
......@@ -80,6 +84,8 @@ struct Layernorm2dFwdPipelineOnePass
{
const auto x_window =
make_tile_window(x_window_, Policy::template MakeXBlockTileDistribution<Problem>());
const auto bias_window = make_tile_window(
bias_window_, Policy::template MakeGammaBetaBlockTileDistribution<Problem>());
const auto gamma_window = make_tile_window(
gamma_window_, Policy::template MakeGammaBetaBlockTileDistribution<Problem>());
const auto beta_window = make_tile_window(
......@@ -89,8 +95,9 @@ struct Layernorm2dFwdPipelineOnePass
auto y_residual_window = make_tile_window(
y_residual_window_, Policy::template MakeXBlockTileDistribution<Problem>());
auto x = load_tile(x_window);
auto x_resi = load_tile(x_residual_window);
auto x = load_tile(x_window);
auto x_resi = load_tile(x_residual_window);
const auto bias = load_tile(bias_window);
int cur_count = 0;
int max_count =
......@@ -110,6 +117,12 @@ struct Layernorm2dFwdPipelineOnePass
kFusedAdd == Layernorm2dFusedAddEnum::PRE_ADD)
{
sweep_tile(x_resi, [&](auto idx) {
if constexpr(kBias == Layernorm2dBiasEnum::ADD_BIAS)
{
// compute x = bias + x
constexpr auto j_idx = make_tuple(idx[number<1>{}]);
acc(idx) = type_convert<ComputeDataType>(bias[j_idx]) + acc(idx);
}
// compute x = x_resi + x
acc(idx) = type_convert<ComputeDataType>(x_resi(idx)) + acc(idx);
});
......
include/ck_tile/ops/layernorm2d/pipeline/layernorm2d_fwd_pipeline_problem.hpp
View file @ 7ea01f4d
......@@ -8,6 +8,7 @@
namespace ck_tile {
template <typename XDataType_,
typename BiasDataType_,
typename GammaDataType_,
typename BetaDataType_,
typename ComputeDataType_,
......@@ -21,6 +22,7 @@ template <typename XDataType_,
struct Layernorm2dFwdPipelineProblem
{
using XDataType = remove_cvref_t<XDataType_>;
using BiasDataType = remove_cvref_t<BiasDataType_>;
using GammaDataType = remove_cvref_t<GammaDataType_>;
using BetaDataType = remove_cvref_t<BetaDataType_>;
using ComputeDataType = remove_cvref_t<ComputeDataType_>;
......
include/ck_tile/ops/layernorm2d/pipeline/layernorm2d_fwd_pipeline_two_pass.hpp
View file @ 7ea01f4d
......@@ -17,6 +17,7 @@ struct Layernorm2dFwdPipelineTwoPass
using Policy = ck_tile::remove_cvref_t<Policy_>;
using XDataType = ck_tile::remove_cvref_t<typename Problem::XDataType>;
using BiasDataType = ck_tile::remove_cvref_t<typename Problem::BiasDataType>;
using GammaDataType = ck_tile::remove_cvref_t<typename Problem::GammaDataType>;
using BetaDataType = ck_tile::remove_cvref_t<typename Problem::BetaDataType>;
using ComputeDataType = ck_tile::remove_cvref_t<typename Problem::ComputeDataType>;
......@@ -36,6 +37,7 @@ struct Layernorm2dFwdPipelineTwoPass
static constexpr bool kPadM = false; // TODO - BlockLayernorm2dFwdProblem::kPadM
static constexpr bool kPadN = Problem::Traits::kPadN;
static constexpr bool kFastFDiv = Problem::Traits::kFastFDiv;
static constexpr auto kBias = Problem::Traits::kBias;
static constexpr auto kFusedAdd = Problem::Traits::kFusedAdd;
static constexpr auto kFusedQuant = Problem::Traits::kFusedQuant;
......@@ -53,6 +55,7 @@ struct Layernorm2dFwdPipelineTwoPass
template <typename XWindow,
typename XResidualWindow,
typename BiasWindow,
typename GammaWindow,
typename BetaWindow,
typename YWindow,
......@@ -64,6 +67,7 @@ struct Layernorm2dFwdPipelineTwoPass
typename Epilogue>
CK_TILE_DEVICE auto operator()(const XWindow& x_window_,
const XResidualWindow& x_residual_window_,
const BiasWindow& bias_window_,
const GammaWindow& gamma_window_,
const BetaWindow& beta_window_,
YWindow& y_window,
......@@ -79,6 +83,8 @@ struct Layernorm2dFwdPipelineTwoPass
{
auto x_window =
make_tile_window(x_window_, Policy::template MakeXBlockTileDistribution<Problem>());
auto bias_window = make_tile_window(
bias_window_, Policy::template MakeGammaBetaBlockTileDistribution<Problem>());
auto gamma_window = make_tile_window(
gamma_window_, Policy::template MakeGammaBetaBlockTileDistribution<Problem>());
auto beta_window = make_tile_window(
......@@ -113,17 +119,25 @@ struct Layernorm2dFwdPipelineTwoPass
for(int iN = __builtin_amdgcn_readfirstlane(0); iN < num_n_tile_iteration; ++iN)
{
auto x = load_tile(x_window);
auto x_resi = load_tile(x_residual_window);
auto x = load_tile(x_window);
auto x_resi = load_tile(x_residual_window);
const auto bias = load_tile(bias_window);
move_tile_window(x_window, {0, Block_N});
move_tile_window(x_residual_window, {0, Block_N});
move_tile_window(bias_window, {Block_N});
auto acc = cast_tile<ComputeDataType>(x);
if constexpr(kFusedAdd == Layernorm2dFusedAddEnum::PRE_ADD_STORE ||
kFusedAdd == Layernorm2dFusedAddEnum::PRE_ADD)
{
sweep_tile(x_resi, [&](auto idx) {
if constexpr(kBias == Layernorm2dBiasEnum::ADD_BIAS)
{
// compute x = bias + x
constexpr auto j_idx = make_tuple(idx[number<1>{}]);
acc(idx) = type_convert<ComputeDataType>(bias[j_idx]) + acc(idx);
}
// compute x = x_resi + x
acc(idx) = type_convert<ComputeDataType>(x_resi(idx)) + acc(idx);
});
......@@ -165,6 +179,7 @@ struct Layernorm2dFwdPipelineTwoPass
move_tile_window(x_window, {0, -Block_N});
move_tile_window(x_residual_window, {0, -Block_N});
move_tile_window(bias_window, {-Block_N});
move_tile_window(gamma_window, {stride_to_right_most_window});
move_tile_window(beta_window, {stride_to_right_most_window});
move_tile_window(y_window, {0, stride_to_right_most_window});
......@@ -172,14 +187,21 @@ struct Layernorm2dFwdPipelineTwoPass
// layernorm computation
for(int iN = __builtin_amdgcn_readfirstlane(0); iN < num_n_tile_iteration; ++iN)
{
auto x = load_tile(x_window);
auto x_resi = load_tile(x_residual_window);
auto x = load_tile(x_window);
auto x_resi = load_tile(x_residual_window);
const auto bias = load_tile(bias_window);
auto acc = cast_tile<ComputeDataType>(x);
if constexpr(kFusedAdd == Layernorm2dFusedAddEnum::PRE_ADD_STORE ||
kFusedAdd == Layernorm2dFusedAddEnum::PRE_ADD)
{
sweep_tile(x_resi, [&](auto idx) {
if constexpr(kBias == Layernorm2dBiasEnum::ADD_BIAS)
{
// compute x = bias + x
constexpr auto j_idx = make_tuple(idx[number<1>{}]);
acc(idx) = type_convert<ComputeDataType>(bias[j_idx]) + acc(idx);
}
// compute x = x_resi + x
acc(idx) = type_convert<ComputeDataType>(x_resi(idx)) + acc(idx);
});
......@@ -207,6 +229,7 @@ struct Layernorm2dFwdPipelineTwoPass
move_tile_window(x_window, {0, -Block_N});
move_tile_window(x_residual_window, {0, -Block_N});
move_tile_window(bias_window, {-Block_N});
move_tile_window(gamma_window, {-Block_N});
move_tile_window(beta_window, {-Block_N});
move_tile_window(y_window, {0, -Block_N});
......
include/ck_tile/ops/layernorm2d/pipeline/layernorm2d_fwd_traits.hpp
View file @ 7ea01f4d
......@@ -7,6 +7,19 @@
namespace ck_tile {
enum class Layernorm2dBiasEnum
{
NO_BIAS = 0,
// add bias before fused add
ADD_BIAS = 1,
};
// clang-format off
template<Layernorm2dBiasEnum> struct Layernorm2dBiasEnumName;
template<> struct Layernorm2dBiasEnumName<Layernorm2dBiasEnum::NO_BIAS> { static constexpr const char * name = "no"; };
template<> struct Layernorm2dBiasEnumName<Layernorm2dBiasEnum::ADD_BIAS> { static constexpr const char * name = "bias"; };
// clang-format on
enum class Layernorm2dFusedAddEnum
{
NO_ADD = 0,
......@@ -41,6 +54,7 @@ template <bool kPadN_,
bool kSaveMeanInvStd_,
bool kFastFDiv_,
bool kTwoPass_,
Layernorm2dBiasEnum kBias_,
Layernorm2dFusedAddEnum kFusedAdd_,
Layernorm2dFusedQuantEnum kFusedQuant_>
struct Layernorm2dFwdTraits
......@@ -49,6 +63,7 @@ struct Layernorm2dFwdTraits
static constexpr bool kSaveMeanInvStd = kSaveMeanInvStd_;
static constexpr bool kFastFDiv = kFastFDiv_;
static constexpr bool kTwoPass = kTwoPass_;
static constexpr Layernorm2dBiasEnum kBias = kBias_;
static constexpr Layernorm2dFusedAddEnum kFusedAdd = kFusedAdd_;
static constexpr Layernorm2dFusedQuantEnum kFusedQuant = kFusedQuant_;
};
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment