ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-20 04:49:54 +00:00
Code Issues Packages Projects Releases Wiki Activity

update layernorm (#1570)

Browse Source 
* port layernorm

* change warp_welford.hpp

* Update warpshuffle

* 1. Add save mean and save std back
2. Move construction of tensor_view and tile_window to operator()

* refine welford max count calculation

* unify layernorm api

* Rename file

* Remove save mean and inv std

* Revert "refine welford max count calculation"

This reverts commit 022365802b.

* Fix order of parameter

* refine welford max count calculation again

* Remove fp32 instances

* Fix bug of padding

* refactor api

* Support bf16

* Extract common function

* Refine arg of operator()

* Add kMThreadPerBlock to template parameter

* clang format

* Refine variable name

* Refine file name

* remove redundant line

* refactor layernorm2d pipeline and add block-per-block utility

* fix name

* rename more

* add more block-per-tile instance

* remove duplicated define

* update instance for 2048, 1024 case

* support up to 2048 now

* opt loading

* add n1536

* Add two pass pipeline

* format

* Fix incorrect type

* parallel compilation

* Use smaller N

* fix 2p pass

* Support Repeat_M in distribution

* Refine nameing

* Add reduce example

---------

Co-authored-by: letaoqin <letaoqin@amd.com>
Co-authored-by: aska-0096 <haocwang@amd.com>
Co-authored-by: rocking <ChunYu.Lai@amd.com>
Co-authored-by: carlushuang <carlus.huang@amd.com>

[ROCm/composable_kernel commit: 0394f8a713]
This commit is contained in:
ltqin
2024-10-22 09:26:18 +08:00
committed by GitHub
parent 124989ade1
commit 45d7cc2f41
59 changed files with 2917 additions and 1042 deletions
Download Patch File Download Diff File Expand all files Collapse all files

223
example/66_complex_contraction_bilinear/run_complex_contraction_bilinear_example.inc Executable file → Normal file
View File

@@ -127,44 +127,47 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[])
switch(init_method)
{
case 0: break;
case 1:
case 0: break;
case 1:
a_ms_ks_re.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
b_ns_ks_re.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
d_ms_ns_re.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
a_ms_ks_re.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
b_ns_ks_re.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
d_ms_ns_re.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
a_ms_ks_img.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
b_ns_ks_img.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
d_ms_ns_img.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
break;
a_ms_ks_img.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
b_ns_ks_img.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
d_ms_ns_img.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
break;
default:
a_ms_ks_re.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_ns_ks_re.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
d_ms_ns_re.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
default:
a_ms_ks_re.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_ns_ks_re.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
d_ms_ns_re.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
a_ms_ks_img.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_ns_ks_img.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
d_ms_ns_img.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
a_ms_ks_img.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_ns_ks_img.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
d_ms_ns_img.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
break;
break;
}
DeviceMem a_device_buf_re(sizeof(ADataType) * a_ms_ks_re.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf_re(sizeof(BDataType) * b_ns_ks_re.mDesc.GetElementSpaceSize());
DeviceMem d_device_buf_re(sizeof(DDataType) * d_ms_ns_re.mDesc.GetElementSpaceSize());
DeviceMem e_device_buf_re(sizeof(EDataType) * e_ms_ns_device_result_re.mDesc.GetElementSpaceSize());
DeviceMem e_device_buf_re(sizeof(EDataType) *
e_ms_ns_device_result_re.mDesc.GetElementSpaceSize());
DeviceMem a_device_buf_img(sizeof(ADataType) * a_ms_ks_img.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf_img(sizeof(BDataType) * b_ns_ks_img.mDesc.GetElementSpaceSize());
DeviceMem d_device_buf_img(sizeof(DDataType) * d_ms_ns_img.mDesc.GetElementSpaceSize());
DeviceMem e_device_buf_img(sizeof(EDataType) * e_ms_ns_device_result_img.mDesc.GetElementSpaceSize());
DeviceMem e_device_buf_img(sizeof(EDataType) *
e_ms_ns_device_result_img.mDesc.GetElementSpaceSize());
// Intermediate Value For E Real and Img
DeviceMem e_device_buf_re1(sizeof(EDataType) * e_ms_ns_device_result_re.mDesc.GetElementSpaceSize());
DeviceMem e_device_buf_img1(sizeof(EDataType) * e_ms_ns_device_result_img.mDesc.GetElementSpaceSize());
DeviceMem e_device_buf_re1(sizeof(EDataType) *
e_ms_ns_device_result_re.mDesc.GetElementSpaceSize());
DeviceMem e_device_buf_img1(sizeof(EDataType) *
e_ms_ns_device_result_img.mDesc.GetElementSpaceSize());
a_device_buf_re.ToDevice(a_ms_ks_re.mData.data());
b_device_buf_re.ToDevice(b_ns_ks_re.mData.data());
@@ -181,7 +184,7 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[])
// set zero for intermediate values
e_device_buf_re1.SetZero();
e_device_buf_img1.SetZero();
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto cde_element_op = CDEElementOp{alpha, beta};
@@ -189,23 +192,24 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[])
// device operation
// For real Intermediate Value re_1
auto op = DeviceOpInstance{};
auto invoker = op.MakeInvoker();
auto argument_re1 = op.MakeArgument(a_device_buf_re.GetDeviceBuffer(),
b_device_buf_re.GetDeviceBuffer(),
std::array<const void*, 1>{d_device_buf_re.GetDeviceBuffer()},
e_device_buf_re1.GetDeviceBuffer(),
a_ms_ks_lengths,
a_ms_ks_strides,
b_ns_ks_lengths,
b_ns_ks_strides,
std::array<std::vector<ck::index_t>, 1>{d_ms_ns_lengths},
std::array<std::vector<ck::index_t>, 1>{d_ms_ns_strides},
e_ms_ns_lengths,
e_ms_ns_strides,
a_element_op,
b_element_op,
cde_element_op);
auto op = DeviceOpInstance{};
auto invoker = op.MakeInvoker();
auto argument_re1 =
op.MakeArgument(a_device_buf_re.GetDeviceBuffer(),
b_device_buf_re.GetDeviceBuffer(),
std::array<const void*, 1>{d_device_buf_re.GetDeviceBuffer()},
e_device_buf_re1.GetDeviceBuffer(),
a_ms_ks_lengths,
a_ms_ks_strides,
b_ns_ks_lengths,
b_ns_ks_strides,
std::array<std::vector<ck::index_t>, 1>{d_ms_ns_lengths},
std::array<std::vector<ck::index_t>, 1>{d_ms_ns_strides},
e_ms_ns_lengths,
e_ms_ns_strides,
a_element_op,
b_element_op,
cde_element_op);
if(!op.IsSupportedArgument(argument_re1))
{
@@ -216,7 +220,6 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[])
float ave_time_re1 = invoker.Run(argument_re1, StreamConfig{nullptr, time_kernel});
alpha = -1.f;
beta = 1.f;
@@ -228,21 +231,22 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[])
// For real Intermediate Value re_2
// auto op = DeviceOpInstance{};
// auto invoker = op.MakeInvoker();
auto argument_re2 = op.MakeArgument(a_device_buf_img.GetDeviceBuffer(),
b_device_buf_img.GetDeviceBuffer(),
std::array<const void*, 1>{e_device_buf_re1.GetDeviceBuffer()},
e_device_buf_re.GetDeviceBuffer(),
a_ms_ks_lengths,
a_ms_ks_strides,
b_ns_ks_lengths,
b_ns_ks_strides,
std::array<std::vector<ck::index_t>, 1>{d_ms_ns_lengths},
std::array<std::vector<ck::index_t>, 1>{d_ms_ns_strides},
e_ms_ns_lengths,
e_ms_ns_strides,
a_element_op,
b_element_op,
cde_element_op);
auto argument_re2 =
op.MakeArgument(a_device_buf_img.GetDeviceBuffer(),
b_device_buf_img.GetDeviceBuffer(),
std::array<const void*, 1>{e_device_buf_re1.GetDeviceBuffer()},
e_device_buf_re.GetDeviceBuffer(),
a_ms_ks_lengths,
a_ms_ks_strides,
b_ns_ks_lengths,
b_ns_ks_strides,
std::array<std::vector<ck::index_t>, 1>{d_ms_ns_lengths},
std::array<std::vector<ck::index_t>, 1>{d_ms_ns_strides},
e_ms_ns_lengths,
e_ms_ns_strides,
a_element_op,
b_element_op,
cde_element_op);
if(!op.IsSupportedArgument(argument_re2))
{
@@ -253,7 +257,6 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[])
float ave_time_re2 = invoker.Run(argument_re2, StreamConfig{nullptr, time_kernel});
alpha = 1.f;
beta = 1.f;
@@ -261,22 +264,22 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[])
b_element_op = BElementOp{};
cde_element_op = CDEElementOp{alpha, beta};
auto argument_img1 = op.MakeArgument(a_device_buf_re.GetDeviceBuffer(),
b_device_buf_img.GetDeviceBuffer(),
std::array<const void*, 1>{d_device_buf_img.GetDeviceBuffer()},
e_device_buf_img1.GetDeviceBuffer(),
a_ms_ks_lengths,
a_ms_ks_strides,
b_ns_ks_lengths,
b_ns_ks_strides,
std::array<std::vector<ck::index_t>, 1>{d_ms_ns_lengths},
std::array<std::vector<ck::index_t>, 1>{d_ms_ns_strides},
e_ms_ns_lengths,
e_ms_ns_strides,
a_element_op,
b_element_op,
cde_element_op);
auto argument_img1 =
op.MakeArgument(a_device_buf_re.GetDeviceBuffer(),
b_device_buf_img.GetDeviceBuffer(),
std::array<const void*, 1>{d_device_buf_img.GetDeviceBuffer()},
e_device_buf_img1.GetDeviceBuffer(),
a_ms_ks_lengths,
a_ms_ks_strides,
b_ns_ks_lengths,
b_ns_ks_strides,
std::array<std::vector<ck::index_t>, 1>{d_ms_ns_lengths},
std::array<std::vector<ck::index_t>, 1>{d_ms_ns_strides},
e_ms_ns_lengths,
e_ms_ns_strides,
a_element_op,
b_element_op,
cde_element_op);
if(!op.IsSupportedArgument(argument_img1))
{
@@ -290,23 +293,22 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[])
alpha = 1.f;
beta = 1.f;
auto argument_img2 = op.MakeArgument(a_device_buf_img.GetDeviceBuffer(),
b_device_buf_re.GetDeviceBuffer(),
std::array<const void*, 1>{e_device_buf_img1.GetDeviceBuffer()},
e_device_buf_img.GetDeviceBuffer(),
a_ms_ks_lengths,
a_ms_ks_strides,
b_ns_ks_lengths,
b_ns_ks_strides,
std::array<std::vector<ck::index_t>, 1>{d_ms_ns_lengths},
std::array<std::vector<ck::index_t>, 1>{d_ms_ns_strides},
e_ms_ns_lengths,
e_ms_ns_strides,
a_element_op,
b_element_op,
cde_element_op);
auto argument_img2 =
op.MakeArgument(a_device_buf_img.GetDeviceBuffer(),
b_device_buf_re.GetDeviceBuffer(),
std::array<const void*, 1>{e_device_buf_img1.GetDeviceBuffer()},
e_device_buf_img.GetDeviceBuffer(),
a_ms_ks_lengths,
a_ms_ks_strides,
b_ns_ks_lengths,
b_ns_ks_strides,
std::array<std::vector<ck::index_t>, 1>{d_ms_ns_lengths},
std::array<std::vector<ck::index_t>, 1>{d_ms_ns_strides},
e_ms_ns_lengths,
e_ms_ns_strides,
a_element_op,
b_element_op,
cde_element_op);
if(!op.IsSupportedArgument(argument_img2))
{
@@ -317,7 +319,6 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[])
float ave_time_img2 = invoker.Run(argument_img2, StreamConfig{nullptr, time_kernel});
ck::index_t M =
ck::accumulate_n<ck::index_t>(e_ms_ns_lengths.begin(), NumDimM, 1, std::multiplies<>{});
@@ -331,9 +332,9 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[])
std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
sizeof(DDataType) * M * N + sizeof(EDataType) * M * N * 2;
float ave_time = ave_time_img2 + ave_time_img1 + ave_time_re2 + ave_time_re1 ;
float ave_time = ave_time_img2 + ave_time_img1 + ave_time_re2 + ave_time_re1;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
@@ -343,7 +344,7 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[])
e_device_buf_img.FromDevice(e_ms_ns_device_result_img.mData.data());
auto isRealOk = 0;
auto isImgOk = 0;
auto isImgOk = 0;
if(do_verification)
{
@@ -366,17 +367,16 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[])
auto ref_op = ReferenceOpInstance{};
auto ref_invoker = ref_op.MakeInvoker();
auto ref_argument_re =
ref_op.MakeArgument(a_ms_ks_re, b_ns_ks_re, c_ms_ns_host_result_re, a_element_op, b_element_op);
auto ref_argument_re = ref_op.MakeArgument(
a_ms_ks_re, b_ns_ks_re, c_ms_ns_host_result_re, a_element_op, b_element_op);
ref_invoker.Run(ref_argument_re);
alpha = 1.f;
beta = 1.f;
cde_element_op = CDEElementOp{alpha, beta};
for(size_t m0 = 0; m0 < e_ms_ns_host_result_re.mDesc.GetLengths()[0]; ++m0)
{
for(size_t m1 = 0; m1 < e_ms_ns_host_result_re.mDesc.GetLengths()[1]; ++m1)
@@ -395,11 +395,11 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[])
alpha = 1.f;
beta = -1.f;
cde_element_op = CDEElementOp{alpha, beta};
auto ref_argument_re1 =
ref_op.MakeArgument(a_ms_ks_img, b_ns_ks_img, c_ms_ns_host_result_re1, a_element_op, b_element_op);
auto ref_argument_re1 = ref_op.MakeArgument(
a_ms_ks_img, b_ns_ks_img, c_ms_ns_host_result_re1, a_element_op, b_element_op);
ref_invoker.Run(ref_argument_re1);
@@ -419,23 +419,20 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[])
}
}
isRealOk = ck::utils::check_err(e_ms_ns_device_result_re, e_ms_ns_host_result_re) ? 0 : 1;
isRealOk = ck::utils::check_err(e_ms_ns_device_result_re, e_ms_ns_host_result_re) ? 0 : 1;
// Img Part Verification
Tensor<CShuffleDataType> c_ms_ns_host_result_img(e_ms_ns_lengths, e_ms_ns_strides);
Tensor<CShuffleDataType> c_ms_ns_host_result_img1(e_ms_ns_lengths, e_ms_ns_strides);
auto ref_argument_img =
ref_op.MakeArgument(a_ms_ks_re, b_ns_ks_img, c_ms_ns_host_result_img, a_element_op, b_element_op);
auto ref_argument_img = ref_op.MakeArgument(
a_ms_ks_re, b_ns_ks_img, c_ms_ns_host_result_img, a_element_op, b_element_op);
ref_invoker.Run(ref_argument_img);
alpha = 1.f;
beta = 1.f;
cde_element_op = CDEElementOp{alpha, beta};
for(size_t m0 = 0; m0 < e_ms_ns_host_result_img.mDesc.GetLengths()[0]; ++m0)
@@ -454,9 +451,9 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[])
}
}
auto ref_argument_img1 =
ref_op.MakeArgument(a_ms_ks_img, b_ns_ks_re, c_ms_ns_host_result_img1, a_element_op, b_element_op);
auto ref_argument_img1 = ref_op.MakeArgument(
a_ms_ks_img, b_ns_ks_re, c_ms_ns_host_result_img1, a_element_op, b_element_op);
ref_invoker.Run(ref_argument_img1);
for(size_t m0 = 0; m0 < e_ms_ns_host_result_img.mDesc.GetLengths()[0]; ++m0)
@@ -475,7 +472,7 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[])
}
}
isImgOk = ck::utils::check_err(e_ms_ns_device_result_re, e_ms_ns_host_result_re) ? 0 : 1;
isImgOk = ck::utils::check_err(e_ms_ns_device_result_re, e_ms_ns_host_result_re) ? 0 : 1;
return (isRealOk && isImgOk);
}

21
example/ck_tile/02_layernorm2d/CMakeLists.txt
View File

@@ -1,4 +1,21 @@
set(EXAMPLE_LAYERNORM2D_FWD "tile_example_layernorm2d_fwd")
# not using add_example_executable() to add this target, since we don't want this to have
# to be included in "make all/install/check"
add_executable(tile_example_layernorm2d_fwd EXCLUDE_FROM_ALL layernorm2d_fwd.cpp)
target_compile_options(tile_example_layernorm2d_fwd PRIVATE -DSAVE_MEAN_INV_STD)
message("adding example ${EXAMPLE_LAYERNORM2D_FWD}")
file(GLOB INSTANCE_SRCS instances/*.cpp)
add_executable(${EXAMPLE_LAYERNORM2D_FWD} EXCLUDE_FROM_ALL layernorm2d_fwd.cpp)
target_include_directories(${EXAMPLE_LAYERNORM2D_FWD} PRIVATE ${CMAKE_CURRENT_LIST_DIR})
target_sources(${EXAMPLE_LAYERNORM2D_FWD} PRIVATE ${INSTANCE_SRCS})
set(EXAMPLE_LAYERNORM2D_FWD_COMPILE_OPTIONS)
# NOTE: we turn off undefined-func-template to let source compile without explicit declare function specializations
list(APPEND EXAMPLE_LAYERNORM2D_FWD_COMPILE_OPTIONS -Wno-undefined-func-template -Wno-float-equal)
target_compile_options(${EXAMPLE_LAYERNORM2D_FWD} PRIVATE ${EXAMPLE_LAYERNORM2D_FWD_COMPILE_OPTIONS})
# TODO: we have to turn off this global prop, otherwise the progress bar generated
# by cmake will print too many files, execvp: /bin/sh: Argument list too long
# however, this property may affect global
# TODO: consider codegen a makefile by us
set_property(GLOBAL PROPERTY RULE_MESSAGES OFF)

5
example/ck_tile/02_layernorm2d/README.md
View File

@@ -6,8 +6,7 @@ This folder contains example for Layernorm2D forward using ck_tile tile-programm
```
# in the root of ck_tile
mkdir build && cd build
# you can replace <arch> with the appropriate architecture (for example gfx90a or gfx942) or leave it blank
sh ../script/cmake-ck-dev.sh ../ <arch>
sh ../script/cmake-ck-dev.sh ../ <arch> # you can replace this <arch> to gfx90a, gfx942...
make tile_example_layernorm2d_fwd -j
```
This will result in an executable `build/bin/tile_example_layernorm2d_fwd`
@@ -20,4 +19,4 @@ args:
-e epsilon (default:1e-5)
-v cpu validation or not (default:1)
-prec precision (default:fp16)
```
```

155
example/ck_tile/02_layernorm2d/instances/layernorm2d_fwd_api.cpp Normal file
View File

@@ -0,0 +1,155 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include <ck_tile/core.hpp>
#include "layernorm2d_fwd.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 kSaveMeanInvStd_,
bool kTwoPass_>
using trait_ = layernorm2d_fwd_traits_<DataType_,
Repeat_M_,
Repeat_N_,
ThreadPerBlock_M_,
ThreadPerBlock_N_,
Vector_N_,
kPadN_,
kSaveMeanInvStd_,
kTwoPass_>;
template <typename data_type>
float layernorm2d_fwd_b16_(layernorm2d_fwd_traits /*t*/,
layernorm2d_fwd_args a,
const ck_tile::stream_config& s)
{
#if 1
float r = -1;
// clang-format off
// rm rn tm tn vn pd mv 2p
if(a.n <= 64) {
r = layernorm2d_fwd_<trait_<data_type, 1, 1, 4, 64, 1, true, false, false>>(s, a);
}
else if(a.n <= 128) {
if (a.n % 2 == 0)
r = layernorm2d_fwd_<trait_<data_type, 1, 1, 4, 64, 2, true, false, false>>(s, a);
else
r = layernorm2d_fwd_<trait_<data_type, 1, 2, 4, 64, 1, true, false, false>>(s, a);
}
else if(a.n <= 256) {
if (a.n % 4 == 0)
r = layernorm2d_fwd_<trait_<data_type, 1, 1, 4, 64, 4, true, false, false>>(s, a);
else if (a.n % 2 == 0)
r = layernorm2d_fwd_<trait_<data_type, 1, 2, 4, 64, 2, true, false, false>>(s, a);
else
r = layernorm2d_fwd_<trait_<data_type, 1, 4, 4, 64, 1, true, false, false>>(s, a);
}
else if(a.n <= 512) {
if (a.n % 8 == 0)
r = layernorm2d_fwd_<trait_<data_type, 1, 1, 4, 64, 8, true, false, false>>(s, a);
else if (a.n % 4 == 0)
r = layernorm2d_fwd_<trait_<data_type, 1, 2, 4, 64, 4, true, false, false>>(s, a);
else if (a.n % 2 == 0)
r = layernorm2d_fwd_<trait_<data_type, 1, 4, 4, 64, 2, true, false, false>>(s, a);
else
r = layernorm2d_fwd_<trait_<data_type, 1, 8, 4, 64, 1, true, false, false>>(s, a);
}
else if(a.n <= 768) {
if (a.n % 4 == 0)
r = layernorm2d_fwd_<trait_<data_type, 1, 3, 4, 64, 4, true, false, false>>(s, a);
else if (a.n % 2 == 0)
r = layernorm2d_fwd_<trait_<data_type, 1, 6, 4, 64, 2, true, false, false>>(s, a);
else
r = layernorm2d_fwd_<trait_<data_type, 1,12, 4, 64, 1, true, false, false>>(s, a);
}
else if(a.n <= 1024) {
if (a.n % 8 == 0)
r = layernorm2d_fwd_<trait_<data_type, 1, 1, 2, 128, 8, true, false, false>>(s, a);
else if (a.n % 4 == 0)
r = layernorm2d_fwd_<trait_<data_type, 1, 2, 2, 128, 4, true, false, false>>(s, a);
else if (a.n % 2 == 0)
r = layernorm2d_fwd_<trait_<data_type, 1, 4, 2, 128, 2, true, false, false>>(s, a);
else
r = layernorm2d_fwd_<trait_<data_type, 1, 4, 1, 256, 1, true, false, false>>(s, a);
}
else if(a.n <= 1536) {
if (a.n % 8 == 0)
r = layernorm2d_fwd_<trait_<data_type, 1, 3, 4, 64, 8, true, false, false>>(s, a);
else if (a.n % 4 == 0)
r = layernorm2d_fwd_<trait_<data_type, 1, 3, 2, 128, 4, true, false, false>>(s, a);
else if (a.n % 2 == 0)
r = layernorm2d_fwd_<trait_<data_type, 1, 3, 1, 256, 2, true, false, false>>(s, a);
else
r = layernorm2d_fwd_<trait_<data_type, 1, 6, 1, 256, 1, true, false, false>>(s, a);
}
else if(a.n <= 2048) {
if (a.n % 8 == 0)
r = layernorm2d_fwd_<trait_<data_type, 1, 1, 1, 256, 8, true, false, false>>(s, a);
else if (a.n % 4 == 0)
r = layernorm2d_fwd_<trait_<data_type, 1, 2, 1, 256, 4, true, false, false>>(s, a);
else if (a.n % 2 == 0)
r = layernorm2d_fwd_<trait_<data_type, 1, 4, 1, 256, 2, true, false, false>>(s, a);
else
r = layernorm2d_fwd_<trait_<data_type, 1, 8, 1, 256, 1, true, false, false>>(s, a);
}
else if(a.n <= 3072) {
if (a.n % 8 == 0)
r = layernorm2d_fwd_<trait_<data_type, 1, 3, 1, 128, 8, true, false, false>>(s, a);
else if (a.n % 4 == 0)
r = layernorm2d_fwd_<trait_<data_type, 1, 3, 1, 256, 4, true, false, false>>(s, a);
else if (a.n % 2 == 0)
r = layernorm2d_fwd_<trait_<data_type, 1, 6, 1, 256, 2, true, false, false>>(s, a);
else
r = layernorm2d_fwd_<trait_<data_type, 1, 3, 1, 1024, 1, true, false, false>>(s, a);
}
else if(a.n <= 4096) {
if (a.n % 8 == 0)
r = layernorm2d_fwd_<trait_<data_type, 1, 2, 1, 256, 8, true, false, false>>(s, a);
else if (a.n % 4 == 0)
r = layernorm2d_fwd_<trait_<data_type, 1, 4, 1, 256, 4, true, false, false>>(s, a);
else if (a.n % 2 == 0)
r = layernorm2d_fwd_<trait_<data_type, 1, 2, 1, 1024, 2, true, false, false>>(s, a);
else
r = layernorm2d_fwd_<trait_<data_type, 1, 4, 1, 1024, 1, true, false, false>>(s, a);
}
else if(a.n > 4096) {
if (a.n % 8 == 0)
r = layernorm2d_fwd_<trait_<data_type, 1, 2, 1, 256, 8, true, false, true>>(s, a);
else if (a.n % 4 == 0)
r = layernorm2d_fwd_<trait_<data_type, 1, 4, 1, 256, 4, true, false, true>>(s, a);
else if (a.n % 2 == 0)
r = layernorm2d_fwd_<trait_<data_type, 1, 2, 1, 1024, 2, true, false, true>>(s, a);
else
r = layernorm2d_fwd_<trait_<data_type, 1, 4, 1, 1024, 1, true, false, true>>(s, a);
}
return r;
#else
return layernorm2d_fwd_<trait_<data_type, 1, 1, 1, 256, 4, true, false, false>>(s, a);
#endif
// clang-format on
}
float layernorm2d_fwd(layernorm2d_fwd_traits t,
layernorm2d_fwd_args a,
const ck_tile::stream_config& s)
{
float r = -1;
if(t.data_type.compare("fp16") == 0)
{
return layernorm2d_fwd_b16_<ck_tile::fp16_t>(t, a, s);
}
else if(t.data_type.compare("bf16") == 0)
{
return layernorm2d_fwd_b16_<ck_tile::bf16_t>(t, a, s);
}
if(r < 0)
throw std::runtime_error("Without supported instances!");
return r;
}

22
example/ck_tile/02_layernorm2d/instances/layernorm2d_fwd_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 "layernorm2d_fwd_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd mv 2p
#if 0
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 2, 4, 64, 8, true , false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 4, 4, 64, 4, true , false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 8, 4, 64, 2, true , false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 16, 4, 64, 1, true , false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 1, 1, 256, 4, true , false, false>>(const S&, A);
#endif
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 1, 2, 128, 8, true, false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 2, 2, 128, 4, true, false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 4, 2, 128, 2, true, false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 4, 1, 256, 1, true, false, false>>(const S&, A);
// clang-format on

13
example/ck_tile/02_layernorm2d/instances/layernorm2d_fwd_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 "layernorm2d_fwd_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd mv 2p
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 3, 4, 64, 8, true, false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 3, 2, 128, 4, true, false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 3, 1, 256, 2, true, false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 6, 1, 256, 1, true, false, false>>(const S&, A);
// clang-format on

14
example/ck_tile/02_layernorm2d/instances/layernorm2d_fwd_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 "layernorm2d_fwd_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd mv 2p
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 1, 1, 256, 8, true, false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 2, 1, 256, 4, true, false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 4, 1, 256, 2, true, false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 8, 1, 256, 1, true, false, false>>(const S&, A);
// clang-format on

12
example/ck_tile/02_layernorm2d/instances/layernorm2d_fwd_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 "layernorm2d_fwd_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd mv 2p
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 1, 4, 64, 4, true , false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 2, 4, 64, 2, true , false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 4, 4, 64, 1, true , false, false>>(const S&, A);
// clang-format on

14
example/ck_tile/02_layernorm2d/instances/layernorm2d_fwd_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 "layernorm2d_fwd_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd mv 2p
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 3, 1, 128, 8, true, false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 3, 1, 256, 4, true, false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 6, 1, 256, 2, true, false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 3, 1, 1024, 1, true, false, false>>(const S&, A);
// clang-format on

14
example/ck_tile/02_layernorm2d/instances/layernorm2d_fwd_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 "layernorm2d_fwd_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd mv 2p
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 2, 1, 256, 8, true, false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 4, 1, 256, 4, true, false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 2, 1, 1024, 2, true, false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 4, 1, 1024, 1, true, false, false>>(const S&, A);
// clang-format on

14
example/ck_tile/02_layernorm2d/instances/layernorm2d_fwd_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 "layernorm2d_fwd_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd mv 2p
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 2, 1, 256, 8, true, false, true>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 4, 1, 256, 4, true, false, true>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 2, 1, 1024, 2, true, false, true>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 4, 1, 1024, 1, true, false, true>>(const S&, A);
// clang-format on

13
example/ck_tile/02_layernorm2d/instances/layernorm2d_fwd_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 "layernorm2d_fwd_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd mv 2p
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 1, 4, 64, 8, true , false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 2, 4, 64, 4, true , false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 4, 4, 64, 2, true , false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 8, 4, 64, 1, true , false, false>>(const S&, A);
// clang-format on

12
example/ck_tile/02_layernorm2d/instances/layernorm2d_fwd_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 "layernorm2d_fwd_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd mv 2p
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 1, 4, 64, 1, true , false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 1, 4, 64, 2, true , false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 2, 4, 64, 1, true , false, false>>(const S&, A);
// clang-format on

12
example/ck_tile/02_layernorm2d/instances/layernorm2d_fwd_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 "layernorm2d_fwd_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd mv 2p
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 3, 4, 64, 4, true , false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 6, 4, 64, 2, true , false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 12, 4, 64, 1, true , false, false>>(const S&, A);
// clang-format on

22
example/ck_tile/02_layernorm2d/instances/layernorm2d_fwd_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 "layernorm2d_fwd_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd mv 2p
#if 0
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 2, 4, 64, 8, true , false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 4, 4, 64, 4, true , false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 8, 4, 64, 2, true , false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 16, 4, 64, 1, true , false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 1, 1, 256, 4, true , false, false>>(const S&, A);
#endif
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 1, 2, 128, 8, true, false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 2, 2, 128, 4, true, false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 4, 2, 128, 2, true, false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 4, 1, 256, 1, true, false, false>>(const S&, A);
// clang-format on

13
example/ck_tile/02_layernorm2d/instances/layernorm2d_fwd_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 "layernorm2d_fwd_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd mv 2p
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 3, 4, 64, 8, true, false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 3, 2, 128, 4, true, false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 3, 1, 256, 2, true, false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 6, 1, 256, 1, true, false, false>>(const S&, A);
// clang-format on

14
example/ck_tile/02_layernorm2d/instances/layernorm2d_fwd_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 "layernorm2d_fwd_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd mv 2p
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 1, 1, 256, 8, true, false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 2, 1, 256, 4, true, false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 4, 1, 256, 2, true, false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 8, 1, 256, 1, true, false, false>>(const S&, A);
// clang-format on

12
example/ck_tile/02_layernorm2d/instances/layernorm2d_fwd_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 "layernorm2d_fwd_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd mv 2p
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 1, 4, 64, 4, true , false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 2, 4, 64, 2, true , false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 4, 4, 64, 1, true , false, false>>(const S&, A);
// clang-format on

14
example/ck_tile/02_layernorm2d/instances/layernorm2d_fwd_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 "layernorm2d_fwd_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd mv 2p
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 3, 1, 128, 8, true, false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 3, 1, 256, 4, true, false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 6, 1, 256, 2, true, false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 3, 1, 1024, 1, true, false, false>>(const S&, A);
// clang-format on

14
example/ck_tile/02_layernorm2d/instances/layernorm2d_fwd_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 "layernorm2d_fwd_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd mv 2p
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 2, 1, 256, 8, true, false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 4, 1, 256, 4, true, false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 2, 1, 1024, 2, true, false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 4, 1, 1024, 1, true, false, false>>(const S&, A);
// clang-format on

14
example/ck_tile/02_layernorm2d/instances/layernorm2d_fwd_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 "layernorm2d_fwd_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd mv 2p
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 2, 1, 256, 8, true, false, true>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 4, 1, 256, 4, true, false, true>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 2, 1, 1024, 2, true, false, true>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 4, 1, 1024, 1, true, false, true>>(const S&, A);
// clang-format on

13
example/ck_tile/02_layernorm2d/instances/layernorm2d_fwd_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 "layernorm2d_fwd_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd mv 2p
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 1, 4, 64, 8, true , false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 2, 4, 64, 4, true , false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 4, 4, 64, 2, true , false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 8, 4, 64, 1, true , false, false>>(const S&, A);
// clang-format on

12
example/ck_tile/02_layernorm2d/instances/layernorm2d_fwd_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 "layernorm2d_fwd_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd mv 2p
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 1, 4, 64, 1, true , false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 1, 4, 64, 2, true , false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 2, 4, 64, 1, true , false, false>>(const S&, A);
// clang-format on

12
example/ck_tile/02_layernorm2d/instances/layernorm2d_fwd_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 "layernorm2d_fwd_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd mv 2p
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 3, 4, 64, 4, true , false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 6, 4, 64, 2, true , false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 12, 4, 64, 1, true , false, false>>(const S&, A);
// clang-format on

67
example/ck_tile/02_layernorm2d/instances/layernorm2d_fwd_instance_common.hpp Normal file
View File

@@ -0,0 +1,67 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include <ck_tile/core.hpp>
#include "layernorm2d_fwd.hpp"
#include <iostream>
#pragma once
using S = ck_tile::stream_config;
using A = layernorm2d_fwd_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 kSaveMeanInvStd_,
bool kTwoPass_>
using trait_ = layernorm2d_fwd_traits_<DataType_,
Repeat_M_,
Repeat_N_,
ThreadPerBlock_M_,
ThreadPerBlock_N_,
Vector_N_,
kPadN_,
kSaveMeanInvStd_,
kTwoPass_>;
template <typename Traits_>
float layernorm2d_fwd_(const S& s, A a)
{
using DataType = typename Traits_::DataType;
using PipelineProblem = ck_tile::Layernorm2dFwdPipelineProblem<
typename LayerNormTypeConfig<DataType>::XDataType,
typename LayerNormTypeConfig<DataType>::GammaDataType,
typename LayerNormTypeConfig<DataType>::BetaDataType,
typename LayerNormTypeConfig<DataType>::ComputeDataType,
typename LayerNormTypeConfig<DataType>::YDataType,
typename LayerNormTypeConfig<DataType>::MeanDataType,
typename LayerNormTypeConfig<DataType>::InvStdDataType,
typename Traits_::Shape,
Traits_::kPadN,
Traits_::kSaveMeanInvStd,
Traits_::kTwoPass>;
using OnePassPipeline = ck_tile::Layernorm2dFwdPipelineOnePass<PipelineProblem>;
using TwoPassPipeline = ck_tile::Layernorm2dFwdPipelineTwoPass<PipelineProblem>;
using Pipeline = std::conditional_t<Traits_::kTwoPass, TwoPassPipeline, OnePassPipeline>;
using Kernel = ck_tile::Layernorm2dFwd<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));
}

240
example/ck_tile/02_layernorm2d/layernorm2d_fwd.cpp
View File

@@ -2,161 +2,120 @@
#include "layernorm2d_fwd.hpp"
#include <cstring>
// Host API implementation
float layernorm2d_fwd(layernorm2d_fwd_traits t,
layernorm2d_fwd_args a,
const ck_tile::stream_config& s)
// different threshold for different dtype
template <typename DataType>
auto get_elimit()
{
if(t.data_type.compare("fp16") == 0)
{
using XDataType = ck_tile::half_t;
using YDataType = ck_tile::half_t;
using GammaDataType = ck_tile::half_t;
using BetaDataType = ck_tile::half_t;
#ifdef SAVE_MEAN_INV_STD
using MeanDataType = ck_tile::half_t;
using InvStdDataType = ck_tile::half_t;
#else
using MeanDataType = ck_tile::null_type;
using InvStdDataType = ck_tile::null_type;
#endif
using ComputeDataType = float;
double rtol = 1e-2;
double atol = 1e-2;
return ck_tile::make_tuple(rtol, atol);
}
using thread_tile = ck_tile::sequence<4, 4>;
using warp_tile = ck_tile::sequence<8, 128>;
using block_tile = ck_tile::sequence<32, 128>;
using Shape = ck_tile::TileLayernorm2dShape<thread_tile, warp_tile, block_tile>;
using PipelineProblem = ck_tile::BlockLayernorm2dFwdProblem<XDataType,
GammaDataType,
BetaDataType,
ComputeDataType,
YDataType,
MeanDataType,
InvStdDataType,
Shape,
true,
true>;
using Kernel = ck_tile::Layernorm2dFwd<PipelineProblem>;
auto kargs = Kernel::MakeKargs(
a.p_x, a.p_gamma, a.p_beta, a.p_y, a.p_mean, a.p_invStd, a.epsilon, a.M, a.N);
const dim3 grids = Kernel::GridSize(a.M);
constexpr dim3 blocks = Kernel::BlockSize();
constexpr ck_tile::index_t kBlockPerCu = Shape::kMWarpPerBlock * Shape::kNWarpPerBlock;
float ave_time = ck_tile::launch_kernel(
s, ck_tile::make_kernel<blocks.x, kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
return ave_time;
}
return 0;
template <>
auto get_elimit<ck_tile::bf16_t>()
{
double rtol = 1e-2;
double atol = 1e-2;
return ck_tile::make_tuple(rtol, atol);
}
auto create_args(int argc, char* argv[])
{
ck_tile::ArgParser arg_parser;
arg_parser.insert("m", "3328", "m dimension")
.insert("n", "4096", "m dimension")
.insert("n", "4096", "n dimension")
.insert("stride", "-1", "stride per row, if -1 then equal to n")
.insert("e", "1e-5", "epsilon")
.insert("save_mv", "0", "save mean/variance(invstd) or not. set to 1 in training case")
.insert("v", "1", "cpu validation or not")
.insert("prec", "fp16", "precision");
.insert("kname", "1", "print kernel name or not")
.insert("prec", "fp16", "precision")
.insert("warmup", "5", "cold iter")
.insert("repeat", "20", "hot iter");
bool result = arg_parser.parse(argc, argv);
return std::make_tuple(result, arg_parser);
}
int main(int argc, char* argv[])
template <typename DataType, bool SaveMeanVar>
bool run(const ck_tile::ArgParser& arg_parser)
{
auto [result, arg_parser] = create_args(argc, argv);
if(!result)
return -1;
ck_tile::index_t m = arg_parser.get_int("m");
ck_tile::index_t n = arg_parser.get_int("n");
ck_tile::index_t stride = arg_parser.get_int("stride");
if(stride < 0)
stride = n;
float epsilon = arg_parser.get_float("e");
ck_tile::index_t M = arg_parser.get_int("m");
ck_tile::index_t N = arg_parser.get_int("n");
std::string data_type = arg_parser.get_str("prec");
int kname = arg_parser.get_int("kname");
int do_validation = arg_parser.get_int("v");
int warmup = arg_parser.get_int("warmup");
int repeat = arg_parser.get_int("repeat");
using XDataType = ck_tile::half_t;
using YDataType = ck_tile::half_t;
using GammaDataType = ck_tile::half_t;
using BetaDataType = ck_tile::half_t;
#ifdef SAVE_MEAN_INV_STD
using MeanDataType = ck_tile::half_t;
using InvStdDataType = ck_tile::half_t;
#else
using MeanDataType = ck_tile::null_type;
using InvStdDataType = ck_tile::null_type;
#endif
using ComputeDataType = float;
assert(stride >= n);
using TypeConfig = LayerNormTypeConfig<DataType>;
using XDataType = typename TypeConfig::XDataType;
using YDataType = typename TypeConfig::YDataType;
using GammaDataType = typename TypeConfig::GammaDataType;
using BetaDataType = typename TypeConfig::BetaDataType;
using MeanDataType =
std::conditional_t<SaveMeanVar, typename TypeConfig::MeanDataType, ck_tile::null_type>;
using InvStdDataType =
std::conditional_t<SaveMeanVar, typename TypeConfig::InvStdDataType, ck_tile::null_type>;
using ComputeDataType = typename TypeConfig::ComputeDataType;
// host verify
ck_tile::HostTensor<XDataType> x_host({M, N});
ck_tile::HostTensor<GammaDataType> gamma_host({N});
ck_tile::HostTensor<BetaDataType> beta_host({N});
ck_tile::HostTensor<XDataType> x_host({m, n}, {stride, 1});
ck_tile::HostTensor<GammaDataType> gamma_host({n});
ck_tile::HostTensor<BetaDataType> beta_host({n});
ck_tile::HostTensor<YDataType> y_host_ref({M, N});
ck_tile::HostTensor<YDataType> y_host_dev({M, N});
ck_tile::HostTensor<YDataType> y_host_ref({m, n}, {stride, 1});
ck_tile::HostTensor<YDataType> y_host_dev({m, n}, {stride, 1});
ck_tile::HostTensor<MeanDataType> mean_host_ref({M});
ck_tile::HostTensor<InvStdDataType> invStd_host_ref({M});
ck_tile::HostTensor<MeanDataType> mean_host_ref({m});
ck_tile::HostTensor<InvStdDataType> invStd_host_ref({m});
#ifdef SAVE_MEAN_INV_STD
ck_tile::HostTensor<MeanDataType> mean_host_dev({M});
ck_tile::HostTensor<InvStdDataType> invStd_host_dev({M});
#endif
ck_tile::FillUniformDistribution<XDataType>{-5.f, 5.f}(x_host);
ck_tile::FillUniformDistribution<GammaDataType>{-5.f, 5.f}(gamma_host);
ck_tile::FillUniformDistribution<BetaDataType>{-5.f, 5.f}(beta_host);
ck_tile::FillUniformDistribution<XDataType>{-.5f, .5f}(x_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 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());
#ifdef SAVE_MEAN_INV_STD
ck_tile::DeviceMem mean_buf(mean_host_dev.get_element_space_size_in_bytes());
ck_tile::DeviceMem invStd_buf(invStd_host_dev.get_element_space_size_in_bytes());
#endif
x_buf.ToDevice(x_host.data());
gamma_buf.ToDevice(gamma_host.data());
beta_buf.ToDevice(beta_host.data());
layernorm2d_fwd_traits traits{data_type};
std::cout << "[" << data_type << "]"
<< " m:" << m << ", n:" << n << ", stride:" << stride << std::flush;
layernorm2d_fwd_traits traits{data_type, SaveMeanVar};
layernorm2d_fwd_args args{x_buf.GetDeviceBuffer(),
gamma_buf.GetDeviceBuffer(),
beta_buf.GetDeviceBuffer(),
y_buf.GetDeviceBuffer(),
#ifdef SAVE_MEAN_INV_STD
mean_buf.GetDeviceBuffer(),
invStd_buf.GetDeviceBuffer(),
#else
nullptr,
nullptr,
#endif
epsilon,
M,
N};
m,
n,
stride};
float ave_time = layernorm2d_fwd(traits, args, ck_tile::stream_config{nullptr, true});
float ave_time = layernorm2d_fwd(
traits, args, ck_tile::stream_config{nullptr, true, kname ? 1 : 0, warmup, repeat});
std::size_t num_byte = sizeof(XDataType) * M * N + sizeof(GammaDataType) * N +
sizeof(BetaDataType) * N + sizeof(YDataType) * M * N;
std::size_t num_byte = sizeof(XDataType) * m * n + sizeof(GammaDataType) * n +
sizeof(BetaDataType) * n + sizeof(YDataType) * m * n;
float gb_per_sec = num_byte / 1.E6 / ave_time;
std::cout << "[" << data_type << "]"
<< " m:" << M << ", n:" << N << ", " << ave_time << " ms, " << gb_per_sec << " GB/s"
<< std::flush;
std::cout << ", " << ave_time * 1.E3 << " us, " << gb_per_sec << " GB/s" << std::flush;
bool pass = true;
@@ -174,20 +133,59 @@ int main(int argc, char* argv[])
y_buf.FromDevice(y_host_dev.data());
pass = ck_tile::check_err(y_host_dev, y_host_ref);
auto [rtol, atol] = get_elimit<DataType>();
if(stride == n)
{
pass = ck_tile::check_err(
y_host_dev, y_host_ref, std::string("OUT Error: Incorrect results!"), rtol, atol);
}
else
{
for(int i_r = 0; i_r < m; i_r++)
{
std::vector<YDataType> y_host_dev_row(y_host_dev.begin() + i_r * stride,
y_host_dev.begin() + i_r * stride + n);
std::vector<YDataType> y_host_ref_row(y_host_ref.begin() + i_r * stride,
y_host_ref.begin() + i_r * stride + n);
pass &= ck_tile::check_err(y_host_dev_row,
y_host_ref_row,
std::string("OUT[") + std::to_string(i_r) +
std::string("] Error: Incorrect results!"),
rtol,
atol);
}
}
#ifdef SAVE_MEAN_INV_STD
mean_buf.FromDevice(mean_host_dev.data());
pass &= ck_tile::check_err(mean_host_dev, mean_host_ref);
invStd_buf.FromDevice(invStd_host_dev.data());
pass &= ck_tile::check_err(invStd_host_dev, invStd_host_ref);
#endif
std::cout << ", valid:" << (pass ? "y" : "n") << std::flush;
std::cout << ", valid:" << (pass ? "y" : "n") << std::flush << std::endl;
}
std::cout << std::endl << std::flush;
return !pass;
return pass;
}
int main(int argc, char* argv[])
{
auto [result, arg_parser] = create_args(argc, argv);
if(!result)
return -1;
const std::string data_type = arg_parser.get_str("prec");
int save_mv = arg_parser.get_int("save_mv");
if(data_type == "fp16" && save_mv)
{
return run<ck_tile::half_t, true>(arg_parser) ? 0 : -2;
}
else if(data_type == "fp16" && !save_mv)
{
return run<ck_tile::half_t, false>(arg_parser) ? 0 : -2;
}
else if(data_type == "bf16" && save_mv)
{
return run<ck_tile::bf16_t, true>(arg_parser) ? 0 : -2;
}
else if(data_type == "bf16" && !save_mv)
{
return run<ck_tile::bf16_t, true>(arg_parser) ? 0 : -2;
}
return -3;
}

119
example/ck_tile/02_layernorm2d/layernorm2d_fwd.hpp
View File

@@ -8,23 +8,114 @@
#include "ck_tile/ops/layernorm2d.hpp"
#include <string>
template <typename DataType>
struct LayerNormTypeConfig;
template <>
struct LayerNormTypeConfig<ck_tile::half_t>
{
using XDataType = ck_tile::half_t;
using YDataType = ck_tile::half_t;
using GammaDataType = ck_tile::half_t;
using BetaDataType = ck_tile::half_t;
using MeanDataType = ck_tile::half_t;
using InvStdDataType = ck_tile::half_t;
using ComputeDataType = float;
};
template <>
struct LayerNormTypeConfig<ck_tile::bf16_t>
{
using XDataType = ck_tile::bf16_t;
using YDataType = ck_tile::bf16_t;
using GammaDataType = ck_tile::bf16_t;
using BetaDataType = ck_tile::bf16_t;
using MeanDataType = ck_tile::bf16_t;
using InvStdDataType = ck_tile::bf16_t;
using ComputeDataType = float;
};
// runtime args
struct layernorm2d_fwd_args : public ck_tile::Layernorm2dFwdHostArgs
{
};
// this is used to pattern-match internl kernel implementation, not to instantiate kernel
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 kSaveMeanInvStd_,
bool kTwoPass_>
struct layernorm2d_fwd_traits_
{
using DataType = ck_tile::remove_cvref_t<DataType_>;
static constexpr bool is_warp_per_row = ThreadPerBlock_N_ <= warpSize;
static_assert((ThreadPerBlock_M_ * ThreadPerBlock_N_) % warpSize == 0);
static constexpr ck_tile::index_t total_warps =
(ThreadPerBlock_M_ * ThreadPerBlock_N_) / warpSize;
// num of warps along m
static constexpr ck_tile::index_t BlockWarps_M = []() {
if constexpr(is_warp_per_row)
{
static_assert(warpSize % ThreadPerBlock_N_ == 0);
return total_warps * (warpSize / ThreadPerBlock_N_);
}
else
{
// static_assert(warpSize % ThreadPerBlock_M_ == 0);
return total_warps / (ThreadPerBlock_N_ / warpSize);
}
}();
// num of warps along n
static constexpr ck_tile::index_t BlockWarps_N = []() {
if constexpr(is_warp_per_row)
{
static_assert(warpSize % ThreadPerBlock_N_ == 0);
return 1;
}
else
{
static_assert(ThreadPerBlock_N_ % warpSize == 0);
return ThreadPerBlock_N_ / warpSize;
}
}();
static constexpr ck_tile::index_t Repeat_M = Repeat_M_;
static constexpr ck_tile::index_t Repeat_N = Repeat_N_;
static constexpr ck_tile::index_t Block_M = Repeat_M_ * ThreadPerBlock_M_;
static constexpr ck_tile::index_t Block_N = Repeat_N_ * ThreadPerBlock_N_ * Vector_N_;
static constexpr ck_tile::index_t Warp_M = ThreadPerBlock_M_ / BlockWarps_M;
static constexpr ck_tile::index_t Warp_N = ThreadPerBlock_N_ / BlockWarps_N * Vector_N_;
using BlockTile = ck_tile::sequence<Block_M, Block_N>;
using BlockWarps = ck_tile::sequence<BlockWarps_M, BlockWarps_N>;
using WarpTile = ck_tile::sequence<Warp_M, Warp_N>;
using Vector = ck_tile::sequence<1, Vector_N_>;
using Shape = ck_tile::Layernorm2dShape<BlockTile, BlockWarps, WarpTile, Vector>;
static constexpr bool kPadN = kPadN_;
static constexpr bool kSaveMeanInvStd = kSaveMeanInvStd_;
static constexpr bool kTwoPass = kTwoPass_;
};
template <typename Traits_>
float layernorm2d_fwd_(const ck_tile::stream_config& s, layernorm2d_fwd_args a);
// This is the public API, will be generated by script
struct layernorm2d_fwd_traits
{
std::string data_type;
bool save_mean_var;
};
struct layernorm2d_fwd_args
{
const void* p_x;
const void* p_gamma;
const void* p_beta;
void* p_y;
void* p_mean;
void* p_invStd;
float epsilon;
ck_tile::index_t M;
ck_tile::index_t N;
};
// host API
float layernorm2d_fwd(layernorm2d_fwd_traits, layernorm2d_fwd_args, const ck_tile::stream_config&);

38
example/ck_tile/02_layernorm2d/script/perf_test.sh Executable file
View File

@@ -0,0 +1,38 @@
# run from top of ck folder
EXE=build/bin/tile_example_layernorm2d_fwd
$EXE -m=1 -n=1 -e=1e-12 -v=1 -prec=bf16 -repeat=1000
$EXE -m=700 -n=80 -e=1e-12 -v=1 -prec=bf16 -repeat=1000
$EXE -m=700 -n=128 -e=1e-12 -v=1 -prec=bf16 -repeat=1000
$EXE -m=700 -n=144 -e=1e-12 -v=1 -prec=bf16 -repeat=1000
$EXE -m=700 -n=168 -e=1e-12 -v=1 -prec=bf16 -repeat=1000
$EXE -m=700 -n=184 -e=1e-12 -v=1 -prec=bf16 -repeat=1000
$EXE -m=700 -n=256 -e=1e-12 -v=1 -prec=bf16 -repeat=1000
$EXE -m=700 -n=288 -e=1e-12 -v=1 -prec=bf16 -repeat=1000
$EXE -m=700 -n=344 -e=1e-12 -v=1 -prec=bf16 -repeat=1000
$EXE -m=700 -n=376 -e=1e-12 -v=1 -prec=bf16 -repeat=1000
$EXE -m=700 -n=448 -e=1e-12 -v=1 -prec=bf16 -repeat=1000
$EXE -m=700 -n=512 -e=1e-12 -v=1 -prec=bf16 -repeat=1000
$EXE -m=700 -n=924 -e=1e-12 -v=1 -prec=bf16 -repeat=1000
$EXE -m=700 -n=1024 -e=1e-12 -v=1 -prec=bf16 -repeat=1000
$EXE -m=700 -n=1078 -e=1e-12 -v=1 -prec=bf16 -repeat=1000
$EXE -m=700 -n=1996 -e=1e-12 -v=1 -prec=bf16 -repeat=1000
$EXE -m=700 -n=4080 -e=1e-12 -v=1 -prec=bf16 -repeat=1000
$EXE -m=700 -n=80 -e=1e-12 -v=1 -prec=fp16 -repeat=1000
$EXE -m=700 -n=128 -e=1e-12 -v=1 -prec=fp16 -repeat=1000
$EXE -m=700 -n=144 -e=1e-12 -v=1 -prec=fp16 -repeat=1000
$EXE -m=700 -n=168 -e=1e-12 -v=1 -prec=fp16 -repeat=1000
$EXE -m=700 -n=184 -e=1e-12 -v=1 -prec=fp16 -repeat=1000
$EXE -m=700 -n=256 -e=1e-12 -v=1 -prec=fp16 -repeat=1000
$EXE -m=700 -n=288 -e=1e-12 -v=1 -prec=fp16 -repeat=1000
$EXE -m=700 -n=344 -e=1e-12 -v=1 -prec=fp16 -repeat=1000
$EXE -m=700 -n=376 -e=1e-12 -v=1 -prec=fp16 -repeat=1000
$EXE -m=700 -n=448 -e=1e-12 -v=1 -prec=fp16 -repeat=1000
$EXE -m=700 -n=512 -e=1e-12 -v=1 -prec=fp16 -repeat=1000
$EXE -m=700 -n=924 -e=1e-12 -v=1 -prec=fp16 -repeat=1000
$EXE -m=700 -n=1024 -e=1e-12 -v=1 -prec=fp16 -repeat=1000
$EXE -m=700 -n=1078 -e=1e-12 -v=1 -prec=fp16 -repeat=1000
$EXE -m=700 -n=1996 -e=1e-12 -v=1 -prec=fp16 -repeat=1000
$EXE -m=700 -n=4080 -e=1e-12 -v=1 -prec=fp16 -repeat=1000

31
example/ck_tile/02_layernorm2d/script/smoke_test.sh Executable file
View File

@@ -0,0 +1,31 @@
#!/bin/sh
# call from top of CK folder
EXE=./build/bin/tile_example_layernorm2d_fwd
for pr_i in "fp16" "bf16" ; do
$EXE -prec=$pr_i -m=99 -n=13
$EXE -prec=$pr_i -m=17 -n=16
$EXE -prec=$pr_i -m=1 -n=100
$EXE -prec=$pr_i -m=4 -n=128
$EXE -prec=$pr_i -m=80 -n=127
$EXE -prec=$pr_i -m=22 -n=255 -stride=256
$EXE -prec=$pr_i -m=7 -n=599
$EXE -prec=$pr_i -m=19 -n=512
$EXE -prec=$pr_i -m=33 -n=313 -stride=1000
$EXE -prec=$pr_i -m=11 -n=510
$EXE -prec=$pr_i -m=171 -n=676 -stride=818
$EXE -prec=$pr_i -m=91 -n=636
$EXE -prec=$pr_i -m=12 -n=768 -stride=800
$EXE -prec=$pr_i -m=100 -n=766 -stride=812
$EXE -prec=$pr_i -m=31 -n=1024
$EXE -prec=$pr_i -m=64 -n=1000 -stride=1004
$EXE -prec=$pr_i -m=8 -n=1501
$EXE -prec=$pr_i -m=3 -n=1826
$EXE -prec=$pr_i -m=5 -n=2040
$EXE -prec=$pr_i -m=7 -n=2734
$EXE -prec=$pr_i -m=1 -n=3182
$EXE -prec=$pr_i -m=9 -n=4096
$EXE -prec=$pr_i -m=3 -n=8192
$EXE -prec=$pr_i -m=1 -n=10547
$EXE -prec=$pr_i -m=3 -n=17134
done

19
example/ck_tile/05_reduce/CMakeLists.txt Normal file
View File

@@ -0,0 +1,19 @@
set(EXAMPLE_REDUCE "tile_example_reduce")
# not using add_example_executable() to add this target, since we don't want this to have
# to be included in "make all/install/check"
message("adding example ${EXAMPLE_REDUCE}")
add_executable(${EXAMPLE_REDUCE} EXCLUDE_FROM_ALL reduce.cpp)
target_include_directories(${EXAMPLE_REDUCE} PRIVATE ${CMAKE_CURRENT_LIST_DIR})
set(EXAMPLE_REDUCE_COMPILE_OPTIONS)
# NOTE: we turn off undefined-func-template to let source compile without explicit declare function specializations
list(APPEND EXAMPLE_REDUCE_COMPILE_OPTIONS -Wno-undefined-func-template -Wno-float-equal)
target_compile_options(${EXAMPLE_REDUCE} PRIVATE ${EXAMPLE_REDUCE_COMPILE_OPTIONS})
# TODO: we have to turn off this global prop, otherwise the progress bar generated
# by cmake will print too many files, execvp: /bin/sh: Argument list too long
# however, this property may affect global
# TODO: consider codegen a makefile by us
set_property(GLOBAL PROPERTY RULE_MESSAGES OFF)

110
example/ck_tile/05_reduce/reduce.cpp Normal file
View File

@@ -0,0 +1,110 @@
#include "ck_tile/host.hpp"
#include "reduce.hpp"
#include <cstring>
auto create_args(int argc, char* argv[])
{
ck_tile::ArgParser arg_parser;
arg_parser.insert("m", "3328", "m dimension")
.insert("n", "4096", "n dimension")
.insert("v", "1", "cpu validation or not")
.insert("prec", "fp16", "precision")
.insert("warmup", "5", "cold iter")
.insert("repeat", "20", "hot iter");
bool result = arg_parser.parse(argc, argv);
return std::make_tuple(result, arg_parser);
}
template <typename DataType>
bool run(const ck_tile::ArgParser& arg_parser)
{
using ADataType = DataType;
using AccDataType = float;
using BDataType = DataType;
ck_tile::index_t m = arg_parser.get_int("m");
ck_tile::index_t n = arg_parser.get_int("n");
int do_validation = arg_parser.get_int("v");
int warmup = arg_parser.get_int("warmup");
int repeat = arg_parser.get_int("repeat");
ck_tile::HostTensor<ADataType> a_host({m, n});
ck_tile::HostTensor<BDataType> b_host_ref({m});
ck_tile::HostTensor<BDataType> b_host_dev({m});
ck_tile::FillUniformDistribution<ADataType>{-5.f, 5.f}(a_host);
ck_tile::DeviceMem a_buf(a_host.get_element_space_size_in_bytes());
ck_tile::DeviceMem b_buf(b_host_dev.get_element_space_size_in_bytes());
a_buf.ToDevice(a_host.data());
using BlockWarps = ck_tile::sequence<4, 1>;
using BlockTile = ck_tile::sequence<128, 128>;
using WarpTile = ck_tile::sequence<32, 128>;
using ThreadTile = ck_tile::sequence<8, 8>;
constexpr ck_tile::index_t kBlockSize = 256;
constexpr ck_tile::index_t kBlockPerCu = 1;
ck_tile::index_t kGridSize = (m / BlockTile::at(ck_tile::number<0>{}));
std::cout << "grid size " << kGridSize << std::endl;
using Kernel = ck_tile::Reduce<ADataType,
AccDataType,
BDataType,
kBlockSize,
BlockWarps,
BlockTile,
WarpTile,
ThreadTile>;
float ave_time = launch_kernel(ck_tile::stream_config{nullptr, true, 0, warmup, repeat},
ck_tile::make_kernel<kBlockSize, kBlockPerCu>(
Kernel{},
kGridSize,
kBlockSize,
0,
static_cast<ADataType*>(a_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_buf.GetDeviceBuffer()),
m,
n));
std::size_t num_btype = sizeof(ADataType) * m * n + sizeof(BDataType) * m;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << gb_per_sec << " GB/s" << std::endl;
bool pass = true;
if(do_validation)
{
// reference
ck_tile::reference_reduce<ADataType, AccDataType, BDataType>(a_host, b_host_ref);
b_buf.FromDevice(b_host_dev.mData.data());
pass = ck_tile::check_err(b_host_dev, b_host_ref);
std::cout << "valid:" << (pass ? "y" : "n") << std::flush << std::endl;
}
return pass;
}
int main(int argc, char* argv[])
{
auto [result, arg_parser] = create_args(argc, argv);
if(!result)
return -1;
const std::string data_type = arg_parser.get_str("prec");
if(data_type == "fp16")
{
return run<ck_tile::half_t>(arg_parser) ? 0 : -2;
}
if(data_type == "bf16")
{
return run<ck_tile::bf16_t>(arg_parser) ? 0 : -2;
}
}

118
example/ck_tile/05_reduce/reduce.hpp Normal file
View File

@@ -0,0 +1,118 @@
// 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/reduce/block/block_reduce.hpp"
namespace ck_tile {
template <typename ADataType,
typename AccDataType,
typename BDataType,
index_t kBlockSize,
typename BlockWarps, // num warps along seq<M, N>
typename BlockTile, // block size, seq<M, N>
typename WarpTile, // warp size, seq<M, N>
typename ThreadTile> // contiguous pixels(vector size) along seq<M, N>
struct Reduce
{
static constexpr index_t Block_M = BlockTile::at(number<0>{});
static constexpr index_t Block_N = BlockTile::at(number<1>{});
static constexpr index_t Warp_M = WarpTile::at(number<0>{});
static constexpr index_t Warp_N = WarpTile::at(number<1>{});
static constexpr index_t Thread_M = ThreadTile::at(number<0>{});
static constexpr index_t Thread_N = ThreadTile::at(number<1>{});
static constexpr index_t WarpPerBlock_M = BlockWarps::at(number<0>{});
static constexpr index_t WarpPerBlock_N = BlockWarps::at(number<1>{});
static constexpr index_t ThreadPerWarp_M = Warp_M / Thread_M;
static constexpr index_t ThreadPerWarp_N = Warp_N / Thread_N;
static constexpr index_t Repeat_M = Block_M / (WarpPerBlock_M * Warp_M);
static constexpr index_t Repeat_N = Block_N / (WarpPerBlock_N * Warp_N);
__device__ static constexpr auto MakeABlockTileDistribution()
{
return make_static_tile_distribution(
tile_distribution_encoding<
sequence<>,
tuple<sequence<Repeat_M, WarpPerBlock_M, ThreadPerWarp_M, Thread_M>,
sequence<Repeat_N, WarpPerBlock_N, ThreadPerWarp_N, Thread_N>>,
tuple<sequence<1, 2>, sequence<1, 2>>,
tuple<sequence<1, 1>, sequence<2, 2>>,
sequence<1, 1, 2, 2>,
sequence<0, 3, 0, 3>>{});
}
__device__ void operator()(const ADataType* p_a, BDataType* p_b, index_t M, index_t N) const
{
const auto a_m_n = make_naive_tensor_view<address_space_enum::global>(
p_a, make_tuple(M, N), make_tuple(N, 1), number<Thread_N>{}, number<1>{});
const auto iM = get_block_id() * Block_M;
// A window
auto a_block_window = make_tile_window(a_m_n,
make_tuple(number<Block_M>{}, number<Block_N>{}),
{iM, 0},
MakeABlockTileDistribution());
const auto f_reduce = [](const auto& v0, const auto& v1) { return v0 + v1; };
const ADataType reduce_init_value = 0;
constexpr auto reduce_dims = sequence<1>{};
// Acc tile
// TODO: support cross warp reduction
auto acc_block_tensor = decltype(block_tile_reduce<AccDataType>(
load_tile(a_block_window), reduce_dims, f_reduce, reduce_init_value)){};
// init Acc tile
tile_elementwise_inout(
[&](auto& acc) { acc = type_convert<AccDataType>(reduce_init_value); },
acc_block_tensor);
// loop
index_t iN = 0;
do
{
const auto a_block_tensor = load_tile(a_block_window);
// FIXME: support cross warp reduction
block_tile_reduce(acc_block_tensor, a_block_tensor, reduce_dims, f_reduce);
move_tile_window(a_block_window, {0, Block_N});
iN += Block_N;
} while(iN < N);
// FIXME: support cross warp reduction
block_tile_reduce_sync(acc_block_tensor, f_reduce);
// convert acc_block_tensor to b_block_tensor
const auto b_block_tensor = tile_elementwise_in(
[](const auto& acc) { return type_convert<BDataType>(acc); }, acc_block_tensor);
// B
const auto b_m = make_naive_tensor_view_packed<address_space_enum::global>(
p_b, make_tuple(M), number<32>{});
// B window
auto b_block_window = make_tile_window(b_m, make_tuple(number<Block_M>{}), {iM});
// store B tile
store_tile(b_block_window, b_block_tensor);
}
};
} // namespace ck_tile

1
example/ck_tile/CMakeLists.txt
View File

@@ -6,3 +6,4 @@ add_subdirectory(01_fmha)
add_subdirectory(02_layernorm2d)
add_subdirectory(03_gemm)
add_subdirectory(04_img2col)
add_subdirectory(05_reduce)