ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-14 02:02:46 +00:00
Code Issues Packages Projects Releases Wiki Activity

external api for gemm + layernorm (#285)

Browse Source 
* Extract base class for elementwise

* Refactor interface of DeviceGemmReduce. Do not use tuple in interface

* [What] Rename d into reduce in gemm + reduction related code
[Why] Prepare to add d term for add

* Unify base class of gemm + reduce and gemm + bias + add + reduce

* 1. Rename gemm_bias_add_reduce for external api
 2. Refine cmake

* Add normalize device operation

* [What] Reorder the argument
[Why] Because d0 is also the input of c.

* Add type string

* Add example of gemm_bias_add_layernorm  via external api

* Refactor example code

* clang-format

* Fix compile error

* clang-format

* Add external api for gemm_add_add_layernorm and normalize

* Add client example

* clang-format

[ROCm/composable_kernel commit: 12235112a1]
This commit is contained in:
rocking5566
2022-06-28 03:25:10 +08:00
committed by GitHub
parent 9096feca63
commit ff7c5cc9e3
47 changed files with 2577 additions and 1946 deletions
Download Patch File Download Diff File Expand all files Collapse all files

18
example/19_binary_elementwise/broadcast_add_2d_amn_bn.cpp
View File

@@ -99,15 +99,17 @@ int main()
a_m_n_device_buf.ToDevice(a_m_n.mData.data());
b_n_device_buf.ToDevice(b_n.mData.data());
std::array<const void*, 2> input = {a_m_n_device_buf.GetDeviceBuffer(),
b_n_device_buf.GetDeviceBuffer()};
std::array<void*, 1> output = {c_m_n_device_buf.GetDeviceBuffer()};
std::vector<ck::index_t> a_strides = {Stride, 1};
std::vector<ck::index_t> b_strides = {0, 1};
std::vector<ck::index_t> c_strides = {Stride, 1};
auto broadcastAdd = DeviceElementwiseAddInstance{};
auto argument = broadcastAdd.MakeArgumentPointer(a_m_n_device_buf.GetDeviceBuffer(),
b_n_device_buf.GetDeviceBuffer(),
c_m_n_device_buf.GetDeviceBuffer(),
{M, N},
{Stride, 1},
{0, 1}, // broadcast in first dimension
{Stride, 1},
Add{});
auto argument = broadcastAdd.MakeArgumentPointer(
input, output, {M, N}, {a_strides, b_strides}, {c_strides}, Add{});
if(!broadcastAdd.IsSupportedArgument(argument.get()))
{

28
example/19_binary_elementwise/broadcast_add_3d_am_bmnk.cpp
View File

@@ -81,18 +81,24 @@ int main()
a_m_device_buf.ToDevice(a_m.mData.data());
b_m_n_k_device_buf.ToDevice(b_m_n_k.mData.data());
std::array<const void*, 2> input = {a_m_device_buf.GetDeviceBuffer(),
b_m_n_k_device_buf.GetDeviceBuffer()};
std::array<void*, 1> output = {c_m_n_k_device_buf.GetDeviceBuffer()};
std::vector<ck::index_t> a_strides = {1, 0, 0};
std::vector<ck::index_t> b_strides{b_m_n_k.mDesc.GetStrides().begin(),
b_m_n_k.mDesc.GetStrides().end()};
std::vector<ck::index_t> c_strides{c_m_n_k.mDesc.GetStrides().begin(),
c_m_n_k.mDesc.GetStrides().end()};
auto broadcastAdd = DeviceElementwiseAddInstance{};
auto argument = broadcastAdd.MakeArgumentPointer(
a_m_device_buf.GetDeviceBuffer(),
b_m_n_k_device_buf.GetDeviceBuffer(),
c_m_n_k_device_buf.GetDeviceBuffer(),
std::vector<ck::index_t>{mnk.begin(), mnk.end()},
{1, 0, 0}, // broadcast A on second and third dimension
std::vector<ck::index_t>{b_m_n_k.mDesc.GetStrides().begin(),
b_m_n_k.mDesc.GetStrides().end()},
std::vector<ck::index_t>{c_m_n_k.mDesc.GetStrides().begin(),
c_m_n_k.mDesc.GetStrides().end()},
Add{});
auto argument =
broadcastAdd.MakeArgumentPointer(input,
output,
std::vector<ck::index_t>{mnk.begin(), mnk.end()},
{a_strides, b_strides},
{c_strides},
Add{});
if(!broadcastAdd.IsSupportedArgument(argument.get()))
{

18
example/19_binary_elementwise/elementwise_add_1d.cpp
View File

@@ -79,15 +79,17 @@ int main()
a_m_device_buf.ToDevice(a_m.mData.data());
b_m_device_buf.ToDevice(b_m.mData.data());
std::array<const void*, 2> input = {a_m_device_buf.GetDeviceBuffer(),
b_m_device_buf.GetDeviceBuffer()};
std::array<void*, 1> output = {c_m_device_buf.GetDeviceBuffer()};
std::vector<ck::index_t> a_strides = {1};
std::vector<ck::index_t> b_strides = {1};
std::vector<ck::index_t> c_strides = {1};
auto broadcastAdd = DeviceElementwiseAddInstance{};
auto argument = broadcastAdd.MakeArgumentPointer(a_m_device_buf.GetDeviceBuffer(),
b_m_device_buf.GetDeviceBuffer(),
c_m_device_buf.GetDeviceBuffer(),
{M},
{1},
{1},
{1},
Add{});
auto argument = broadcastAdd.MakeArgumentPointer(
input, output, {M}, {{a_strides}, b_strides}, {c_strides}, Add{});
if(!broadcastAdd.IsSupportedArgument(argument.get()))
{

24
example/19_binary_elementwise/elementwise_add_4d.cpp
View File

@@ -81,16 +81,22 @@ int main()
a_device_buf.ToDevice(a.mData.data());
b_device_buf.ToDevice(b.mData.data());
std::array<const void*, 2> input = {a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer()};
std::array<void*, 1> output = {c_device_buf.GetDeviceBuffer()};
std::vector<ck::index_t> a_strides{a.mDesc.GetStrides().begin(), a.mDesc.GetStrides().end()};
std::vector<ck::index_t> b_strides{b.mDesc.GetStrides().begin(), b.mDesc.GetStrides().end()};
std::vector<ck::index_t> c_strides{c.mDesc.GetStrides().begin(), c.mDesc.GetStrides().end()};
auto broadcastAdd = DeviceElementwiseAddInstance{};
auto argument = broadcastAdd.MakeArgumentPointer(
a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
c_device_buf.GetDeviceBuffer(),
std::vector<ck::index_t>{nchw.begin(), nchw.end()},
std::vector<ck::index_t>{a.mDesc.GetStrides().begin(), a.mDesc.GetStrides().end()},
std::vector<ck::index_t>{b.mDesc.GetStrides().begin(), b.mDesc.GetStrides().end()},
std::vector<ck::index_t>{c.mDesc.GetStrides().begin(), c.mDesc.GetStrides().end()},
Add{});
auto argument =
broadcastAdd.MakeArgumentPointer(input,
output,
std::vector<ck::index_t>{nchw.begin(), nchw.end()},
{{a_strides}, b_strides},
{c_strides},
Add{});
if(!broadcastAdd.IsSupportedArgument(argument.get()))
{