ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-05 06:01:23 +00:00
Code Issues Packages Projects Releases Wiki Activity

[Ck_tile] smoothquant (#1617)

Browse Source 
* fix compile error

* fix typo of padding

* Add smoothquant op

* Add smoothquant instance library

* refine type

* add test script

* Re-generate smoothquant.hpp

* Always use 'current year' in copyright

* use Generic2dBlockShape instead

* Add vector = 8 instance back

* Find exe path automatically

* Simplify the api condition

* Remove debugging code

* update year

* Add blank line between function declaration

* explicitly cast return value to dim3

* refine return value

* Fix default warmup and repeat value

* Add comment

* refactor sommthquant cmake

* Add README

* Fix typo

---------

Co-authored-by: Po Yen, Chen <PoYen.Chen@amd.com>
This commit is contained in:
rocking
2024-11-01 13:51:56 +08:00
committed by GitHub
parent 550248deec
commit fbd654545a
62 changed files with 1758 additions and 219 deletions
Download Patch File Download Diff File Expand all files Collapse all files

22
example/ck_tile/12_smoothquant/instances/smoothquant_bf16_n1024_instance.cpp Normal file
View File

@@ -0,0 +1,22 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "smoothquant_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd 2p
#if 0
template float smoothquant_<trait_<ck_tile::bf16_t, 1, 2, 4, 64, 8, true, false>>(const S&, A);
template float smoothquant_<trait_<ck_tile::bf16_t, 1, 4, 4, 64, 4, true, false>>(const S&, A);
template float smoothquant_<trait_<ck_tile::bf16_t, 1, 8, 4, 64, 2, true, false>>(const S&, A);
template float smoothquant_<trait_<ck_tile::bf16_t, 1, 16, 4, 64, 1, true, false>>(const S&, A);
template float smoothquant_<trait_<ck_tile::bf16_t, 1, 1, 1, 256, 4, true, false>>(const S&, A);
#endif
template float smoothquant_<trait_<ck_tile::bf16_t, 1, 1, 2, 128, 8, true, false>>(const S&, A);
template float smoothquant_<trait_<ck_tile::bf16_t, 1, 2, 2, 128, 4, true, false>>(const S&, A);
template float smoothquant_<trait_<ck_tile::bf16_t, 1, 4, 2, 128, 2, true, false>>(const S&, A);
template float smoothquant_<trait_<ck_tile::bf16_t, 1, 4, 1, 256, 1, true, false>>(const S&, A);
// clang-format on

13
example/ck_tile/12_smoothquant/instances/smoothquant_bf16_n1536_instance.cpp Normal file
View File

@@ -0,0 +1,13 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "smoothquant_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd 2p
template float smoothquant_<trait_<ck_tile::bf16_t, 1, 3, 4, 64, 8, true, false>>(const S&, A);
template float smoothquant_<trait_<ck_tile::bf16_t, 1, 3, 2, 128, 4, true, false>>(const S&, A);
template float smoothquant_<trait_<ck_tile::bf16_t, 1, 3, 1, 256, 2, true, false>>(const S&, A);
template float smoothquant_<trait_<ck_tile::bf16_t, 1, 6, 1, 256, 1, true, false>>(const S&, A);
// clang-format on

14
example/ck_tile/12_smoothquant/instances/smoothquant_bf16_n2048_instance.cpp Normal file
View File

@@ -0,0 +1,14 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "smoothquant_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd 2p
template float smoothquant_<trait_<ck_tile::bf16_t, 1, 1, 1, 256, 8, true, false>>(const S&, A);
template float smoothquant_<trait_<ck_tile::bf16_t, 1, 2, 1, 256, 4, true, false>>(const S&, A);
template float smoothquant_<trait_<ck_tile::bf16_t, 1, 4, 1, 256, 2, true, false>>(const S&, A);
template float smoothquant_<trait_<ck_tile::bf16_t, 1, 8, 1, 256, 1, true, false>>(const S&, A);
// clang-format on

12
example/ck_tile/12_smoothquant/instances/smoothquant_bf16_n256_instance.cpp Normal file
View File

@@ -0,0 +1,12 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "smoothquant_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd 2p
template float smoothquant_<trait_<ck_tile::bf16_t, 1, 1, 4, 64, 4, true, false>>(const S&, A);
template float smoothquant_<trait_<ck_tile::bf16_t, 1, 2, 4, 64, 2, true, false>>(const S&, A);
template float smoothquant_<trait_<ck_tile::bf16_t, 1, 4, 4, 64, 1, true, false>>(const S&, A);
// clang-format on

14
example/ck_tile/12_smoothquant/instances/smoothquant_bf16_n3072_instance.cpp Normal file
View File

@@ -0,0 +1,14 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "smoothquant_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd 2p
template float smoothquant_<trait_<ck_tile::bf16_t, 1, 3, 1, 128, 8, true, false>>(const S&, A);
template float smoothquant_<trait_<ck_tile::bf16_t, 1, 3, 1, 256, 4, true, false>>(const S&, A);
template float smoothquant_<trait_<ck_tile::bf16_t, 1, 6, 1, 256, 2, true, false>>(const S&, A);
template float smoothquant_<trait_<ck_tile::bf16_t, 1, 3, 1, 1024, 1, true, false>>(const S&, A);
// clang-format on

14
example/ck_tile/12_smoothquant/instances/smoothquant_bf16_n4096_instance.cpp Normal file
View File

@@ -0,0 +1,14 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "smoothquant_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd 2p
template float smoothquant_<trait_<ck_tile::bf16_t, 1, 2, 1, 256, 8, true, false>>(const S&, A);
template float smoothquant_<trait_<ck_tile::bf16_t, 1, 4, 1, 256, 4, true, false>>(const S&, A);
template float smoothquant_<trait_<ck_tile::bf16_t, 1, 2, 1, 1024, 2, true, false>>(const S&, A);
template float smoothquant_<trait_<ck_tile::bf16_t, 1, 4, 1, 1024, 1, true, false>>(const S&, A);
// clang-format on

14
example/ck_tile/12_smoothquant/instances/smoothquant_bf16_n4096_tp_instance.cpp Normal file
View File

@@ -0,0 +1,14 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "smoothquant_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd 2p
template float smoothquant_<trait_<ck_tile::bf16_t, 1, 2, 1, 256, 8, true, true>>(const S&, A);
template float smoothquant_<trait_<ck_tile::bf16_t, 1, 4, 1, 256, 4, true, true>>(const S&, A);
template float smoothquant_<trait_<ck_tile::bf16_t, 1, 2, 1, 1024, 2, true, true>>(const S&, A);
template float smoothquant_<trait_<ck_tile::bf16_t, 1, 4, 1, 1024, 1, true, true>>(const S&, A);
// clang-format on

13
example/ck_tile/12_smoothquant/instances/smoothquant_bf16_n512_instance.cpp Normal file
View File

@@ -0,0 +1,13 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "smoothquant_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd 2p
template float smoothquant_<trait_<ck_tile::bf16_t, 1, 1, 4, 64, 8, true, false>>(const S&, A);
template float smoothquant_<trait_<ck_tile::bf16_t, 1, 2, 4, 64, 4, true, false>>(const S&, A);
template float smoothquant_<trait_<ck_tile::bf16_t, 1, 4, 4, 64, 2, true, false>>(const S&, A);
template float smoothquant_<trait_<ck_tile::bf16_t, 1, 8, 4, 64, 1, true, false>>(const S&, A);
// clang-format on

12
example/ck_tile/12_smoothquant/instances/smoothquant_bf16_n64_n128_instance.cpp Normal file
View File

@@ -0,0 +1,12 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "smoothquant_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd 2p
template float smoothquant_<trait_<ck_tile::bf16_t, 1, 1, 4, 64, 1, true , false>>(const S&, A);
template float smoothquant_<trait_<ck_tile::bf16_t, 1, 1, 4, 64, 2, true , false>>(const S&, A);
template float smoothquant_<trait_<ck_tile::bf16_t, 1, 2, 4, 64, 1, true , false>>(const S&, A);
// clang-format on

12
example/ck_tile/12_smoothquant/instances/smoothquant_bf16_n768_instance.cpp Normal file
View File

@@ -0,0 +1,12 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "smoothquant_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd 2p
template float smoothquant_<trait_<ck_tile::bf16_t, 1, 3, 4, 64, 4, true, false>>(const S&, A);
template float smoothquant_<trait_<ck_tile::bf16_t, 1, 6, 4, 64, 2, true, false>>(const S&, A);
template float smoothquant_<trait_<ck_tile::bf16_t, 1, 12, 4, 64, 1, true, false>>(const S&, A);
// clang-format on

22
example/ck_tile/12_smoothquant/instances/smoothquant_fp16_n1024_instance.cpp Normal file
View File

@@ -0,0 +1,22 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "smoothquant_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd 2p
#if 0
template float smoothquant_<trait_<ck_tile::fp16_t, 1, 2, 4, 64, 8, true ,false>>(const S&, A);
template float smoothquant_<trait_<ck_tile::fp16_t, 1, 4, 4, 64, 4, true ,false>>(const S&, A);
template float smoothquant_<trait_<ck_tile::fp16_t, 1, 8, 4, 64, 2, true ,false>>(const S&, A);
template float smoothquant_<trait_<ck_tile::fp16_t, 1, 16, 4, 64, 1, true ,false>>(const S&, A);
template float smoothquant_<trait_<ck_tile::fp16_t, 1, 1, 1, 256, 4, true ,false>>(const S&, A);
#endif
template float smoothquant_<trait_<ck_tile::fp16_t, 1, 1, 2, 128, 8, true, false>>(const S&, A);
template float smoothquant_<trait_<ck_tile::fp16_t, 1, 2, 2, 128, 4, true, false>>(const S&, A);
template float smoothquant_<trait_<ck_tile::fp16_t, 1, 4, 2, 128, 2, true, false>>(const S&, A);
template float smoothquant_<trait_<ck_tile::fp16_t, 1, 4, 1, 256, 1, true, false>>(const S&, A);
// clang-format on

13
example/ck_tile/12_smoothquant/instances/smoothquant_fp16_n1536_instance.cpp Normal file
View File

@@ -0,0 +1,13 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "smoothquant_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd 2p
template float smoothquant_<trait_<ck_tile::fp16_t, 1, 3, 4, 64, 8, true, false>>(const S&, A);
template float smoothquant_<trait_<ck_tile::fp16_t, 1, 3, 2, 128, 4, true, false>>(const S&, A);
template float smoothquant_<trait_<ck_tile::fp16_t, 1, 3, 1, 256, 2, true, false>>(const S&, A);
template float smoothquant_<trait_<ck_tile::fp16_t, 1, 6, 1, 256, 1, true, false>>(const S&, A);
// clang-format on

14
example/ck_tile/12_smoothquant/instances/smoothquant_fp16_n2048_instance.cpp Normal file
View File

@@ -0,0 +1,14 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "smoothquant_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd 2p
template float smoothquant_<trait_<ck_tile::fp16_t, 1, 1, 1, 256, 8, true, false>>(const S&, A);
template float smoothquant_<trait_<ck_tile::fp16_t, 1, 2, 1, 256, 4, true, false>>(const S&, A);
template float smoothquant_<trait_<ck_tile::fp16_t, 1, 4, 1, 256, 2, true, false>>(const S&, A);
template float smoothquant_<trait_<ck_tile::fp16_t, 1, 8, 1, 256, 1, true, false>>(const S&, A);
// clang-format on

12
example/ck_tile/12_smoothquant/instances/smoothquant_fp16_n256_instance.cpp Normal file
View File

@@ -0,0 +1,12 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "smoothquant_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd 2p
template float smoothquant_<trait_<ck_tile::fp16_t, 1, 1, 4, 64, 4, true, false>>(const S&, A);
template float smoothquant_<trait_<ck_tile::fp16_t, 1, 2, 4, 64, 2, true, false>>(const S&, A);
template float smoothquant_<trait_<ck_tile::fp16_t, 1, 4, 4, 64, 1, true, false>>(const S&, A);
// clang-format on

14
example/ck_tile/12_smoothquant/instances/smoothquant_fp16_n3072_instance.cpp Normal file
View File

@@ -0,0 +1,14 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "smoothquant_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd 2p
template float smoothquant_<trait_<ck_tile::fp16_t, 1, 3, 1, 128, 8,true, false>>(const S&, A);
template float smoothquant_<trait_<ck_tile::fp16_t, 1, 3, 1, 256, 4,true, false>>(const S&, A);
template float smoothquant_<trait_<ck_tile::fp16_t, 1, 6, 1, 256, 2,true, false>>(const S&, A);
template float smoothquant_<trait_<ck_tile::fp16_t, 1, 3, 1, 1024, 1,true, false>>(const S&, A);
// clang-format on

14
example/ck_tile/12_smoothquant/instances/smoothquant_fp16_n4096_instance.cpp Normal file
View File

@@ -0,0 +1,14 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "smoothquant_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd 2p
template float smoothquant_<trait_<ck_tile::fp16_t, 1, 2, 1, 256, 8, true, false>>(const S&, A);
template float smoothquant_<trait_<ck_tile::fp16_t, 1, 4, 1, 256, 4, true, false>>(const S&, A);
template float smoothquant_<trait_<ck_tile::fp16_t, 1, 2, 1, 1024, 2, true, false>>(const S&, A);
template float smoothquant_<trait_<ck_tile::fp16_t, 1, 4, 1, 1024, 1, true, false>>(const S&, A);
// clang-format on

14
example/ck_tile/12_smoothquant/instances/smoothquant_fp16_n4096_tp_instance.cpp Normal file
View File

@@ -0,0 +1,14 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "smoothquant_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd 2p
template float smoothquant_<trait_<ck_tile::fp16_t, 1, 2, 1, 256, 8, true, true>>(const S&, A);
template float smoothquant_<trait_<ck_tile::fp16_t, 1, 4, 1, 256, 4, true, true>>(const S&, A);
template float smoothquant_<trait_<ck_tile::fp16_t, 1, 2, 1, 1024, 2, true, true>>(const S&, A);
template float smoothquant_<trait_<ck_tile::fp16_t, 1, 4, 1, 1024, 1, true, true>>(const S&, A);
// clang-format on

13
example/ck_tile/12_smoothquant/instances/smoothquant_fp16_n512_instance.cpp Normal file
View File

@@ -0,0 +1,13 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "smoothquant_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd 2p
template float smoothquant_<trait_<ck_tile::fp16_t, 1, 1, 4, 64, 8, true , false>>(const S&, A);
template float smoothquant_<trait_<ck_tile::fp16_t, 1, 2, 4, 64, 4, true , false>>(const S&, A);
template float smoothquant_<trait_<ck_tile::fp16_t, 1, 4, 4, 64, 2, true , false>>(const S&, A);
template float smoothquant_<trait_<ck_tile::fp16_t, 1, 8, 4, 64, 1, true , false>>(const S&, A);
// clang-format on

12
example/ck_tile/12_smoothquant/instances/smoothquant_fp16_n64_n128_instance.cpp Normal file
View File

@@ -0,0 +1,12 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "smoothquant_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd 2p
template float smoothquant_<trait_<ck_tile::fp16_t, 1, 1, 4, 64, 1, true, false>>(const S&, A);
template float smoothquant_<trait_<ck_tile::fp16_t, 1, 1, 4, 64, 2, true, false>>(const S&, A);
template float smoothquant_<trait_<ck_tile::fp16_t, 1, 2, 4, 64, 1, true, false>>(const S&, A);
// clang-format on

12
example/ck_tile/12_smoothquant/instances/smoothquant_fp16_n768_instance.cpp Normal file
View File

@@ -0,0 +1,12 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "smoothquant_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd 2p
template float smoothquant_<trait_<ck_tile::fp16_t, 1, 3, 4, 64, 4, true , false>>(const S&, A);
template float smoothquant_<trait_<ck_tile::fp16_t, 1, 6, 4, 64, 2, true , false>>(const S&, A);
template float smoothquant_<trait_<ck_tile::fp16_t, 1, 12, 4, 64, 1, true , false>>(const S&, A);
// clang-format on

143
example/ck_tile/12_smoothquant/instances/smoothquant_fwd_api.cpp Normal file
View File

@@ -0,0 +1,143 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include <ck_tile/core.hpp>
#include "smoothquant.hpp"
template <typename DataType_,
ck_tile::index_t Repeat_M_, // each thread repeat along M
ck_tile::index_t Repeat_N_, // each thread repeat along N
ck_tile::index_t ThreadPerBlock_M_, // num threads along M
ck_tile::index_t ThreadPerBlock_N_, // num threads along N
ck_tile::index_t Vector_N_, // vector size along N
bool kPadN_,
bool kTwoPass_>
using trait_ = smoothquant_traits_<DataType_,
Repeat_M_,
Repeat_N_,
ThreadPerBlock_M_,
ThreadPerBlock_N_,
Vector_N_,
kPadN_,
kTwoPass_>;
template <typename data_type>
float smoothquant_dispatch(smoothquant_traits /*t*/,
smoothquant_args a,
const ck_tile::stream_config& s)
{
float r = -1;
// clang-format off
// rm rn tm tn vn pd 2p
if(a.n <= 64) {
r = smoothquant_<trait_<data_type, 1, 1, 4, 64, 1, true, false>>(s, a);
}
else if(a.n <= 128) {
if (a.n % 2 == 0)
r = smoothquant_<trait_<data_type, 1, 1, 4, 64, 2, true, false>>(s, a);
else
r = smoothquant_<trait_<data_type, 1, 2, 4, 64, 1, true, false>>(s, a);
}
else if(a.n <= 256) {
if (a.n % 4 == 0)
r = smoothquant_<trait_<data_type, 1, 1, 4, 64, 4, true, false>>(s, a);
else if (a.n % 2 == 0)
r = smoothquant_<trait_<data_type, 1, 2, 4, 64, 2, true, false>>(s, a);
else
r = smoothquant_<trait_<data_type, 1, 4, 4, 64, 1, true, false>>(s, a);
}
else if(a.n <= 512) {
if (a.n % 8 == 0)
r = smoothquant_<trait_<data_type, 1, 1, 4, 64, 8, true, false>>(s, a);
else if (a.n % 4 == 0)
r = smoothquant_<trait_<data_type, 1, 2, 4, 64, 4, true, false>>(s, a);
else if (a.n % 2 == 0)
r = smoothquant_<trait_<data_type, 1, 4, 4, 64, 2, true, false>>(s, a);
else
r = smoothquant_<trait_<data_type, 1, 8, 4, 64, 1, true, false>>(s, a);
}
else if(a.n <= 768) {
if (a.n % 4 == 0)
r = smoothquant_<trait_<data_type, 1, 3, 4, 64, 4, true, false>>(s, a);
else if (a.n % 2 == 0)
r = smoothquant_<trait_<data_type, 1, 6, 4, 64, 2, true, false>>(s, a);
else
r = smoothquant_<trait_<data_type, 1,12, 4, 64, 1, true, false>>(s, a);
}
else if(a.n <= 1024) {
if (a.n % 8 == 0)
r = smoothquant_<trait_<data_type, 1, 1, 2, 128, 8, true, false>>(s, a);
else if (a.n % 4 == 0)
r = smoothquant_<trait_<data_type, 1, 2, 2, 128, 4, true, false>>(s, a);
else if (a.n % 2 == 0)
r = smoothquant_<trait_<data_type, 1, 4, 2, 128, 2, true, false>>(s, a);
else
r = smoothquant_<trait_<data_type, 1, 4, 1, 256, 1, true, false>>(s, a);
}
else if(a.n <= 1536) {
if (a.n % 8 == 0)
r = smoothquant_<trait_<data_type, 1, 3, 4, 64, 8, true, false>>(s, a);
else if (a.n % 4 == 0)
r = smoothquant_<trait_<data_type, 1, 3, 2, 128, 4, true, false>>(s, a);
else if (a.n % 2 == 0)
r = smoothquant_<trait_<data_type, 1, 3, 1, 256, 2, true, false>>(s, a);
else
r = smoothquant_<trait_<data_type, 1, 6, 1, 256, 1, true, false>>(s, a);
}
else if(a.n <= 2048) {
if (a.n % 8 == 0)
r = smoothquant_<trait_<data_type, 1, 1, 1, 256, 8, true, false>>(s, a);
else if (a.n % 4 == 0)
r = smoothquant_<trait_<data_type, 1, 2, 1, 256, 4, true, false>>(s, a);
else if (a.n % 2 == 0)
r = smoothquant_<trait_<data_type, 1, 4, 1, 256, 2, true, false>>(s, a);
else
r = smoothquant_<trait_<data_type, 1, 8, 1, 256, 1, true, false>>(s, a);
}
else if(a.n <= 3072) {
if (a.n % 8 == 0)
r = smoothquant_<trait_<data_type, 1, 3, 1, 128, 8, true, false>>(s, a);
else if (a.n % 4 == 0)
r = smoothquant_<trait_<data_type, 1, 3, 1, 256, 4, true, false>>(s, a);
else if (a.n % 2 == 0)
r = smoothquant_<trait_<data_type, 1, 6, 1, 256, 2, true, false>>(s, a);
else
r = smoothquant_<trait_<data_type, 1, 3, 1, 1024, 1, true, false>>(s, a);
}
else if(a.n <= 4096) {
if (a.n % 8 == 0)
r = smoothquant_<trait_<data_type, 1, 2, 1, 256, 8, true, false>>(s, a);
else if (a.n % 4 == 0)
r = smoothquant_<trait_<data_type, 1, 4, 1, 256, 4, true, false>>(s, a);
else if (a.n % 2 == 0)
r = smoothquant_<trait_<data_type, 1, 2, 1, 1024, 2, true, false>>(s, a);
else
r = smoothquant_<trait_<data_type, 1, 4, 1, 1024, 1, true, false>>(s, a);
}
else if(a.n > 4096) {
if (a.n % 8 == 0)
r = smoothquant_<trait_<data_type, 1, 2, 1, 256, 8, true, true>>(s, a);
else if (a.n % 4 == 0)
r = smoothquant_<trait_<data_type, 1, 4, 1, 256, 4, true, true>>(s, a);
else if (a.n % 2 == 0)
r = smoothquant_<trait_<data_type, 1, 2, 1, 1024, 2, true, true>>(s, a);
else
r = smoothquant_<trait_<data_type, 1, 4, 1, 1024, 1, true, true>>(s, a);
}
return r;
// clang-format on
}
float smoothquant(smoothquant_traits t, smoothquant_args a, const ck_tile::stream_config& s)
{
if(t.data_type.compare("fp16") == 0)
{
return smoothquant_dispatch<ck_tile::fp16_t>(t, a, s);
}
else if(t.data_type.compare("bf16") == 0)
{
return smoothquant_dispatch<ck_tile::bf16_t>(t, a, s);
}
else
throw std::runtime_error("Without supported instances!");
}

62
example/ck_tile/12_smoothquant/instances/smoothquant_instance_common.hpp Normal file
View File

@@ -0,0 +1,62 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include <ck_tile/core.hpp>
#include "smoothquant.hpp"
#include <iostream>
#pragma once
using S = ck_tile::stream_config;
using A = smoothquant_args;
template <typename DataType_,
ck_tile::index_t Repeat_M_, // each thread repeat along M
ck_tile::index_t Repeat_N_, // each thread repeat along N
ck_tile::index_t ThreadPerBlock_M_, // num threads along M
ck_tile::index_t ThreadPerBlock_N_, // num threads along N
ck_tile::index_t Vector_N_, // vector size along N
bool kPadN_,
bool kTwoPass_>
using trait_ = smoothquant_traits_<DataType_,
Repeat_M_,
Repeat_N_,
ThreadPerBlock_M_,
ThreadPerBlock_N_,
Vector_N_,
kPadN_,
kTwoPass_>;
template <typename Traits_>
float smoothquant_(const S& s, A a)
{
using DataType = typename Traits_::DataType;
using PipelineProblem = ck_tile::SmoothquantPipelineProblem<
typename SmoothquantTypeConfig<DataType>::XDataType,
typename SmoothquantTypeConfig<DataType>::XScaleDataType,
typename SmoothquantTypeConfig<DataType>::ComputeDataType,
typename SmoothquantTypeConfig<DataType>::YScaleDataType,
typename SmoothquantTypeConfig<DataType>::QYDataType,
typename Traits_::Shape,
Traits_::kPadN,
Traits_::kTwoPass>;
using OnePassPipeline = ck_tile::SmoothquantPipelineOnePass<PipelineProblem>;
using TwoPassPipeline = ck_tile::SmoothquantPipelineTwoPass<PipelineProblem>;
using Pipeline = std::conditional_t<Traits_::kTwoPass, TwoPassPipeline, OnePassPipeline>;
using Kernel = ck_tile::Smoothquant<Pipeline>;
const dim3 grids = Kernel::GridSize(a);
constexpr dim3 blocks = Kernel::BlockSize();
constexpr ck_tile::index_t kBlockPerCu = 1;
auto kargs = Kernel::MakeKargs(a);
if(s.log_level_ > 0)
std::cout << ", " << Kernel::GetName() << std::flush;
return ck_tile::launch_kernel(
s, ck_tile::make_kernel<blocks.x, kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
}