ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-13 17:55:48 +00:00
Code Issues Packages Projects Releases Wiki Activity

Hotfix binary elementwise (for broadcast on fastest axis) (#254)

Browse Source 
* Support different length of ScalarPerVector

* Add example of broadcast on fastest axis

* Typo

* Refine fastest example

* Add dimension check

* Modify fastest broadcast example to 3d

* Enforce users give scalarPerVector explicitely

* 1. Add CscalarPerVedctor
2. Not only broadcast on fastest need to set scalarPerVector to 1

* Rename var

* Move IsScalarPerVectorValid() inside IsSupportedArgument()

* Separate GridDesc_M0 into A, B and C

* rename var

* Rename var of length

Co-authored-by: rocking <chunylai@amd.com>
This commit is contained in:
rocking5566
2022-05-26 00:17:27 +08:00
committed by GitHub
parent e579c9e5c6
commit 82d7d9938f
7 changed files with 319 additions and 125 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
example/19_binary_elementwise/CMakeLists.txt
View File

@@ -1,3 +1,4 @@
add_example_executable(example_broadcast_add_2d broadcast_add_2d.cpp)
add_example_executable(example_broadcast_add_2d_amn_bn broadcast_add_2d_amn_bn.cpp)
add_example_executable(example_broadcast_add_3d_am_bmnk broadcast_add_3d_am_bmnk.cpp)
add_example_executable(example_elementwise_add_1d elementwise_add_1d.cpp)
add_example_executable(example_elementwise_add_4d elementwise_add_4d.cpp)

17
example/19_binary_elementwise/broadcast_add_2d.cpp → example/19_binary_elementwise/broadcast_add_2d_amn_bn.cpp
View File

@@ -19,8 +19,17 @@ using EltwiseComputeDataType = F32;
using Add = ck::tensor_operation::binary_element_wise::Add;
using DeviceElementwiseAddInstance = ck::tensor_operation::device::
DeviceBinaryElementwise<ABDataType, ABDataType, CDataType, EltwiseComputeDataType, Add, 2, 8>;
using DeviceElementwiseAddInstance =
ck::tensor_operation::device::DeviceBinaryElementwise<ABDataType,
ABDataType,
CDataType,
EltwiseComputeDataType,
Add,
2,
8,
8,
8,
8>;
template <typename HostTensorA,
typename HostTensorB,
@@ -100,7 +109,7 @@ int main()
if(!broadcastAdd.IsSupportedArgument(argument.get()))
{
throw std::runtime_error("The runtime parameters seems not supported by the "
"DeviceBinaryElementwise_2D instance, exiting!");
"DeviceBinaryElementwise instance, exiting!");
};
auto broadcastAdd_invoker_ptr = broadcastAdd.MakeInvokerPointer();
@@ -123,7 +132,7 @@ int main()
0>(host_c_m_n, a_m_n, b_n, M, N, Add{});
pass &= ck::utils::check_err(
c_m_n.mData, host_c_m_n.mData, "Error: Incorrect results d1", 1e-3, 1e-3);
c_m_n.mData, host_c_m_n.mData, "Error: Incorrect results c", 1e-3, 1e-3);
}
return pass ? 0 : 1;

123
example/19_binary_elementwise/broadcast_add_3d_am_bmnk.cpp Normal file
View File

@@ -0,0 +1,123 @@
#include <iostream>
#include <cstdlib>
#include "check_err.hpp"
#include "config.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "device_tensor.hpp"
#include "binary_element_wise_operation.hpp"
#include "device_binary_elementwise.hpp"
using F16 = ck::half_t;
using F32 = float;
using ABDataType = F16;
using CDataType = F16;
using EltwiseComputeDataType = F32;
using Add = ck::tensor_operation::binary_element_wise::Add;
using DeviceElementwiseAddInstance =
ck::tensor_operation::device::DeviceBinaryElementwise<ABDataType,
ABDataType,
CDataType,
EltwiseComputeDataType,
Add,
3,
8,
1,
8,
8>;
template <typename HostTensorA,
typename HostTensorB,
typename HostTensorC,
typename ComputeDataType,
typename Functor>
void host_broadcast3D_am_bmnk(HostTensorC& C,
const HostTensorA& A,
const HostTensorB& B,
const std::vector<std::size_t>& shape,
Functor functor)
{
using ctype = ck::remove_reference_t<decltype(C(0, 0))>;
for(std::size_t m = 0; m < shape[0]; ++m)
for(std::size_t n = 0; n < shape[1]; ++n)
for(std::size_t k = 0; k < shape[2]; ++k)
{
ComputeDataType a_val = static_cast<ComputeDataType>(A(m));
ComputeDataType b_val = static_cast<ComputeDataType>(B(m, n, k));
ComputeDataType c_val = 0;
functor(c_val, a_val, b_val);
C(m, n, k) = static_cast<ctype>(c_val);
}
}
int main()
{
bool do_verification = true;
bool time_kernel = false;
std::vector<std::size_t> mnk = {4, 16, 32};
ck::index_t M = mnk[0];
Tensor<ABDataType> a_m({M});
Tensor<ABDataType> b_m_n_k(mnk);
Tensor<CDataType> c_m_n_k(mnk);
a_m.GenerateTensorValue(GeneratorTensor_3<ABDataType>{0.0, 1.0});
b_m_n_k.GenerateTensorValue(GeneratorTensor_3<ABDataType>{0.0, 1.0});
DeviceMem a_m_device_buf(sizeof(ABDataType) * a_m.mDesc.GetElementSpace());
DeviceMem b_m_n_k_device_buf(sizeof(ABDataType) * b_m_n_k.mDesc.GetElementSpace());
DeviceMem c_m_n_k_device_buf(sizeof(CDataType) * c_m_n_k.mDesc.GetElementSpace());
a_m_device_buf.ToDevice(a_m.mData.data());
b_m_n_k_device_buf.ToDevice(b_m_n_k.mData.data());
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{});
if(!broadcastAdd.IsSupportedArgument(argument.get()))
{
throw std::runtime_error("The runtime parameters seems not supported by the "
"DeviceBinaryElementwise instance, exiting!");
};
auto broadcastAdd_invoker_ptr = broadcastAdd.MakeInvokerPointer();
float ave_time =
broadcastAdd_invoker_ptr->Run(argument.get(), StreamConfig{nullptr, time_kernel});
std::cout << "Perf: " << ave_time << " ms" << std::endl;
bool pass = true;
if(do_verification)
{
c_m_n_k_device_buf.FromDevice(c_m_n_k.mData.data());
Tensor<CDataType> host_c_m_n_k(mnk);
host_broadcast3D_am_bmnk<Tensor<ABDataType>,
Tensor<ABDataType>,
Tensor<CDataType>,
EltwiseComputeDataType,
Add>(host_c_m_n_k, a_m, b_m_n_k, mnk, Add{});
pass &= ck::utils::check_err(
c_m_n_k.mData, host_c_m_n_k.mData, "Error: Incorrect results c", 1e-3, 1e-3);
}
return pass ? 0 : 1;
}

17
example/19_binary_elementwise/elementwise_add_1d.cpp
View File

@@ -19,8 +19,17 @@ using EltwiseComputeDataType = F32;
using Add = ck::tensor_operation::binary_element_wise::Add;
using DeviceElementwiseAddInstance = ck::tensor_operation::device::
DeviceBinaryElementwise<ABDataType, ABDataType, CDataType, EltwiseComputeDataType, Add, 1, 8>;
using DeviceElementwiseAddInstance =
ck::tensor_operation::device::DeviceBinaryElementwise<ABDataType,
ABDataType,
CDataType,
EltwiseComputeDataType,
Add,
1,
8,
8,
8,
8>;
template <typename HostTensorA,
typename HostTensorB,
@@ -81,7 +90,7 @@ int main()
if(!broadcastAdd.IsSupportedArgument(argument.get()))
{
throw std::runtime_error("The runtime parameters seems not supported by the "
"DeviceBinaryElementwise_2D instance, exiting!");
"DeviceBinaryElementwise instance, exiting!");
};
auto broadcastAdd_invoker_ptr = broadcastAdd.MakeInvokerPointer();
@@ -103,7 +112,7 @@ int main()
Add>(host_c_m, a_m, b_m, M, Add{});
pass &= ck::utils::check_err(
c_m.mData, host_c_m.mData, "Error: Incorrect results d1", 1e-3, 1e-3);
c_m.mData, host_c_m.mData, "Error: Incorrect results c", 1e-3, 1e-3);
}
return pass ? 0 : 1;

17
example/19_binary_elementwise/elementwise_add_4d.cpp
View File

@@ -19,8 +19,17 @@ using EltwiseComputeDataType = F32;
using Add = ck::tensor_operation::binary_element_wise::Add;
using DeviceElementwiseAddInstance = ck::tensor_operation::device::
DeviceBinaryElementwise<ABDataType, ABDataType, CDataType, EltwiseComputeDataType, Add, 4, 8>;
using DeviceElementwiseAddInstance =
ck::tensor_operation::device::DeviceBinaryElementwise<ABDataType,
ABDataType,
CDataType,
EltwiseComputeDataType,
Add,
4,
8,
8,
8,
8>;
template <typename HostTensorA,
typename HostTensorB,
@@ -83,7 +92,7 @@ int main()
if(!broadcastAdd.IsSupportedArgument(argument.get()))
{
throw std::runtime_error("The runtime parameters seems not supported by the "
"DeviceBinaryElementwise_2D instance, exiting!");
"DeviceBinaryElementwise instance, exiting!");
};
auto broadcastAdd_invoker_ptr = broadcastAdd.MakeInvokerPointer();
@@ -105,7 +114,7 @@ int main()
Add>(host_c, a, b, nchw, Add{});
pass &=
ck::utils::check_err(c.mData, host_c.mData, "Error: Incorrect results d1", 1e-3, 1e-3);
ck::utils::check_err(c.mData, host_c.mData, "Error: Incorrect results c", 1e-3, 1e-3);
}
return pass ? 0 : 1;