Skip to content
GitLab
Menu
Projects
Groups
Snippets
Loading...
Help
Help
Support
Community forum
Keyboard shortcuts
?
Submit feedback
Contribute to GitLab
Sign in / Register
Toggle navigation
Menu
Open sidebar
gaoqiong
composable_kernel
Commits
ce82ed9b
Commit
ce82ed9b
authored
Apr 06, 2023
by
Rosty Geyyer
Browse files
Add a warp-level op
parent
ffde5d16
Changes
1
Hide whitespace changes
Inline
Side-by-side
Showing
1 changed file
with
226 additions
and
0 deletions
+226
-0
include/ck/tensor_operation/gpu/warp/dpp_gemm.hpp
include/ck/tensor_operation/gpu/warp/dpp_gemm.hpp
+226
-0
No files found.
include/ck/tensor_operation/gpu/warp/dpp_gemm.hpp
0 → 100644
View file @
ce82ed9b
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/utility/common_header.hpp"
#include "ck/utility/math.hpp"
#include "ck/utility/amd_gemm_dpp.hpp"
namespace
ck
{
enum
struct
DppGemmInstr
{
dpp_f32_8x8x8_f16
=
0
,
dpp_i32_8x8x8_i8
};
template
<
DppGemmInstr
Instr
,
index_t
WaveSize
,
typename
=
void
>
struct
dpp_gemm_type
{
};
template
<
index_t
WaveSize
>
struct
dpp_gemm_type
<
DppGemmInstr
::
dpp_f32_8x8x8_f16
,
WaveSize
,
typename
std
::
enable_if_t
<
WaveSize
==
64
>>
{
// * DPP GEMM setup
static
constexpr
index_t
waves_per_wg
=
4
;
static
constexpr
index_t
m_per_dpp
=
8
;
static
constexpr
index_t
n_per_dpp
=
8
;
static
constexpr
index_t
k_per_dpp
=
8
;
static
constexpr
index_t
dpp_per_wave
=
8
;
static
constexpr
index_t
src_a_data_size
=
2
;
static
constexpr
index_t
src_b_data_size
=
2
;
static
constexpr
index_t
acc_data_size
=
4
;
static
constexpr
index_t
wave_size
=
WaveSize
;
// * Thread mapping inside wave, num_thread_per_subgroups always alone N direction
static
constexpr
index_t
num_thread_per_dpp
=
n_per_dpp
;
template
<
index_t
MPerWave
,
index_t
NPerWave
,
index_t
KPerWave
,
class
FloatA
,
class
FloatB
,
class
FloatC
>
__device__
void
run
(
const
FloatA
&
a
,
const
FloatB
&
b
,
FloatC
&
reg_c
)
const
{
intrin_dpp_f32_8x8x8_f16
<
MPerWave
,
NPerWave
,
KPerWave
>::
Run
(
a
,
b
,
reg_c
);
}
};
template
<
index_t
WaveSize
>
struct
dpp_gemm_type
<
DppGemmInstr
::
dpp_i32_8x8x8_i8
,
WaveSize
,
typename
std
::
enable_if_t
<
WaveSize
==
64
>>
{
// * DPP GEMM setup
static
constexpr
index_t
waves_per_wg
=
4
;
static
constexpr
index_t
m_per_dpp
=
8
;
static
constexpr
index_t
n_per_dpp
=
8
;
static
constexpr
index_t
k_per_dpp
=
8
;
static
constexpr
index_t
dpp_per_wave
=
8
;
static
constexpr
index_t
src_a_data_size
=
1
;
static
constexpr
index_t
src_b_data_size
=
1
;
static
constexpr
index_t
acc_data_size
=
4
;
static
constexpr
index_t
wave_size
=
WaveSize
;
// * Thread mapping inside wave, num_thread_per_dpp always alone N direction
static
constexpr
index_t
num_thread_per_dpp
=
n_per_dpp
;
template
<
index_t
MPerWave
,
index_t
NPerWave
,
index_t
KPerWave
,
class
FloatA
,
class
FloatB
,
class
FloatC
>
__device__
void
run
(
const
FloatA
&
a
,
const
FloatB
&
b
,
FloatC
&
reg_c
)
const
{
intrin_dpp_i32_8x8x8_i8
<
MPerWave
,
NPerWave
,
KPerWave
>::
Run
(
a
,
b
,
reg_c
);
}
};
template
<
typename
src_type_a
,
typename
src_type_b
,
typename
dst_type
,
index_t
MPerWave
,
index_t
NPerWave
,
index_t
KPerWave
>
struct
DppGemmSelector
{
template
<
typename
src_type_a_
,
typename
src_type_b_
,
typename
dst_type_
,
index_t
MPerWave_
,
index_t
NPerWave_
,
index_t
KPerWave_
>
static
constexpr
auto
GetDppGemm
();
template
<
>
static
constexpr
auto
GetDppGemm
<
half_t
,
half_t
,
float
,
8
,
8
,
8
>
()
{
return
DppGemmInstr
::
dpp_f32_8x8x8_f16
;
}
template
<
>
static
constexpr
auto
GetDppGemm
<
int8_t
,
int8_t
,
int
,
8
,
8
,
8
>
()
{
return
DppGemmInstr
::
dpp_i32_8x8x8_i8
;
}
// get_warp_size do not return the correct wavesize, hardcode to 32 as workaround
static
constexpr
auto
selected_dpp_gemm
=
dpp_gemm_type
<
GetDppGemm
<
src_type_a
,
src_type_b
,
dst_type
,
MPerWave
,
NPerWave
,
KPerWave
>
(),
Number
<
64
>
{}
>
{};
__host__
__device__
constexpr
DppGemmSelector
()
{
static_assert
(
selected_dpp_gemm
.
m_per_wave
==
8
,
"Something went wrong, M per wave should be equal to 8"
);
static_assert
(
selected_dpp_gemm
.
n_per_wave
==
8
,
"Something went wrong, N per wave should be equal to 8"
);
static_assert
(
selected_dpp_gemm
.
k_per_wave
==
8
,
"Something went wrong, K per wave should be equal to 8"
);
}
};
template
<
typename
src_type_a
,
typename
src_type_b
,
typename
dst_type
,
index_t
MPerWave
,
index_t
NPerWave
,
index_t
KPerWave
,
bool
TransposeC
=
false
>
struct
DppGemm
{
static
constexpr
auto
I0
=
Number
<
0
>
{};
static
constexpr
auto
I1
=
Number
<
1
>
{};
static
constexpr
auto
I2
=
Number
<
2
>
{};
static
constexpr
auto
I3
=
Number
<
3
>
{};
static
constexpr
auto
I4
=
Number
<
4
>
{};
static
constexpr
auto
I5
=
Number
<
5
>
{};
// using CIndex = MultiIndex<2>;
// using CIndex4D = MultiIndex<4>;
__host__
__device__
constexpr
DppGemm
()
{
static_assert
(
MPerWave
==
8
&&
NPerWave
==
8
&&
KPerWave
==
8
,
"Only MPerWave == 8, NPerWave == 8 and KPerWave == 8 are supported"
);
}
// DPP output supporting C = A * B
// MPerDpp x KPerDpp @ KPerDpp x NPerDpp -> MPerDpp x NPerDpp
template
<
typename
CDesc_MBlockxRepeat_NBlockxRepeat_MWave_NWave_MPerDpp_NPerDpp
>
__host__
__device__
static
constexpr
auto
CDesc_MBlockxRepeat_NBlockxRepeat_MWave_NWave_MSubGroup_NThreadPerSubgroup
(
const
CDesc_MBlockxRepeat_NBlockxRepeat_MWave_NWave_MPerDpp_NPerDpp
&
c_desc_mblockxrepeat_nblockxRepeat_mwave_nwave_mperdpp_nperdpp
)
{
const
auto
MBlockxRepeat
=
c_desc_mblockxrepeat_nblockxRepeat_mwave_nwave_mperdpp_nperdpp
.
GetLength
(
I0
);
const
auto
NBlockxRepeat
=
c_desc_mblockxrepeat_nblockxRepeat_mwave_nwave_mperdpp_nperdpp
.
GetLength
(
I1
);
const
auto
MWave
=
c_desc_mblockxrepeat_nblockxRepeat_mwave_nwave_mperdpp_nperdpp
.
GetLength
(
I2
);
const
auto
NWave
=
c_desc_mblockxrepeat_nblockxRepeat_mwave_nwave_mperdpp_nperdpp
.
GetLength
(
I3
);
return
transform_tensor_descriptor
(
c_desc_mblockxrepeat_nblockxRepeat_mwave_nwave_mperdpp_nperdpp
,
make_tuple
(
make_pass_through_transform
(
MBlockxRepeat
),
make_pass_through_transform
(
NBlockxRepeat
),
make_pass_through_transform
(
MWave
),
make_pass_through_transform
(
NWave
),
make_pass_through_transform
(
Number
<
dpp_gemm_instr
.
dpp_per_wave
>
{}),
make_pass_through_transform
(
Number
<
dpp_gemm_instr
.
num_thread_per_dpp
>
{})),
make_tuple
(
Sequence
<
0
>
{},
Sequence
<
1
>
{},
Sequence
<
2
>
{},
Sequence
<
3
>
{},
Sequence
<
4
>
{},
Sequence
<
5
>
{}),
make_tuple
(
Sequence
<
0
>
{},
Sequence
<
1
>
{},
Sequence
<
2
>
{},
Sequence
<
3
>
{},
Sequence
<
4
>
{},
Sequence
<
5
>
{}));
}
__device__
static
constexpr
index_t
GetWaveSize
()
{
return
dpp_gemm_instr
.
wave_size
;
}
template
<
class
FloatA
,
class
FloatB
,
class
FloatC
>
__device__
void
Run
(
const
FloatA
&
p_a_wave
,
const
FloatB
&
p_b_wave
,
FloatC
&
p_c_thread
)
const
{
static_assert
(
(
is_same
<
src_type_a
,
half_t
>::
value
&&
is_same
<
src_type_b
,
half_t
>::
value
&&
is_same
<
dst_type
,
float
>::
value
)
||
(
is_same
<
src_type_a
,
int8_t
>::
value
&&
is_same
<
src_type_b
,
int8_t
>::
value
&&
is_same
<
dst_type
,
int32_t
>::
value
),
"base type couple must be (half, float) "
"or (int8, int32)!"
);
if
constexpr
(
!
TransposeC
)
{
dpp_gemm_instr
.
template
run
<
MPerWave
,
NPerWave
,
KPerWave
>(
p_a_wave
,
p_b_wave
,
p_c_thread
);
}
else
{
dpp_gemm_instr
.
template
run
<
MPerWave
,
NPerWave
,
KPerWave
>(
p_b_wave
,
p_a_wave
,
p_c_thread
);
}
}
__device__
static
auto
GetLaneId
()
{
return
get_thread_local_1d_id
()
%
dpp_gemm_instr
.
wave_size
;
}
__device__
static
auto
GetSubGroupId
()
{
return
(
GetLaneId
()
/
dpp_gemm_instr
.
num_thread_per_dpp
)
%
dpp_gemm_instr
.
dpp_per_wave
;
}
__device__
static
auto
GetLaneIdUnderSubGroup
()
{
return
GetLaneId
()
%
dpp_gemm_instr
.
num_thread_per_dpp
;
}
static
constexpr
auto
dpp_gemm
=
DppGemmSelector
<
src_type_a
,
src_type_b
,
dst_type
,
MPerWave
,
NPerWave
,
KPerWave
>
{};
static
constexpr
auto
dpp_gemm_instr
=
dpp_gemm
.
selected_dpp_gemm
;
};
}
// namespace ck
Write
Preview
Markdown
is supported
0%
Try again
or
attach a new file
.
Attach a file
Cancel
You are about to add
0
people
to the discussion. Proceed with caution.
Finish editing this message first!
Cancel
Please
register
or
sign in
to comment