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

Batchnorm-forward and Batchnorm-infer Implemented using generic kernels (#320)

Browse Source 
* Implement multiple-reduction in one kernel (kernels, device ops, examples)

* Add generic elementwise kernel and device interface

* Add generator for normal-distributed data initialization

* Add host refer implementation of batchnorm-forward and batchnorm-infer

* Add examples for implementing batchnorm-forward and batchnorm-infer using generic kernels

* Remove un-needed including in batchnorm example

* Renaming generic_elementwise to elementiwise in kernel and device classes/functions

* Change in gemm_layernorm examples to use DeviceElementwise instead of Device5AryElementwise

* Change in exampe 19_binary_elementwise to use DeviceElementwise instead of DeviceBinaryElementwise

* Change in device_cgemm_4gemm_xdl_cshuffle.hpp to use kernel_elementwise instead of kernel_binary_elementwise

* Add DeviceElementwiseBase and use it in device_normalize_instance.cpp

* Removing and renaming files

* Update to synchronize gemm_layernorm client example to the generic element-wise device op API

* Update to synchronize with the latest headers directory and HostTensorDescriptor interface renaming

* Merge two static member functions in device_elementwise.hpp

* Remove unary_elementwise_1d kernel and device

[ROCm/composable_kernel commit: 53ea4713af]
This commit is contained in:
Qianfeng
2022-08-15 23:11:02 +08:00
committed by GitHub
parent 215b32cd3f
commit 25cca7c462
47 changed files with 5195 additions and 1707 deletions
Download Patch File Download Diff File Expand all files Collapse all files

56
example/19_binary_elementwise/broadcast_add_2d_amn_bn.cpp
View File

@@ -6,7 +6,7 @@
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/device_binary_elementwise.hpp"
#include "ck/tensor_operation/gpu/device/device_elementwise.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
@@ -16,28 +16,23 @@
using F16 = ck::half_t;
using F32 = float;
using ABDataType = F16;
using CDataType = F16;
using EltwiseComputeDataType = F32;
using ABDataType = F16;
using CDataType = F16;
using Add = ck::tensor_operation::element_wise::Add;
using DeviceElementwiseAddInstance =
ck::tensor_operation::device::DeviceBinaryElementwise<ABDataType,
ABDataType,
CDataType,
EltwiseComputeDataType,
Add,
2,
8,
8,
8,
8>;
ck::tensor_operation::device::DeviceElementwise<ck::Tuple<ABDataType, ABDataType>,
ck::Tuple<CDataType>,
Add,
2,
8,
ck::Sequence<8, 8>,
ck::Sequence<8>>;
template <typename HostTensorA,
typename HostTensorB,
typename HostTensorC,
typename ComputeDataType,
typename Functor,
int broadcastDim>
void host_broadcast2D(
@@ -49,19 +44,19 @@ void host_broadcast2D(
{
for(int n = 0; n < N; ++n)
{
ComputeDataType Amn = ck::type_convert<ComputeDataType>(A(m, n));
ComputeDataType Cmn = 0;
auto Amn = A(m, n);
ctype Cmn = 0;
if constexpr(broadcastDim == 0)
{
ComputeDataType Bn = ck::type_convert<ComputeDataType>(B(n));
auto Bn = B(n);
functor(Cmn, Amn, Bn);
}
else
{
ComputeDataType Bm = ck::type_convert<ComputeDataType>(B(m));
auto Bm = B(m);
functor(Cmn, Amn, Bm);
}
C(m, n) = ck::type_convert<ctype>(Cmn);
C(m, n) = Cmn;
}
}
}
@@ -103,18 +98,19 @@ int main()
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};
std::array<ck::index_t, 2> abc_lengths = {M, N};
std::array<ck::index_t, 2> a_strides = {Stride, 1};
std::array<ck::index_t, 2> b_strides = {0, 1};
std::array<ck::index_t, 2> c_strides = {Stride, 1};
auto broadcastAdd = DeviceElementwiseAddInstance{};
auto argument = broadcastAdd.MakeArgumentPointer(
input, output, {M, N}, {a_strides, b_strides}, {c_strides}, Add{});
abc_lengths, {a_strides, b_strides}, {c_strides}, input, output, Add{});
if(!broadcastAdd.IsSupportedArgument(argument.get()))
{
throw std::runtime_error("The runtime parameters seems not supported by the "
"DeviceBinaryElementwise instance, exiting!");
throw std::runtime_error(
"The runtime parameters seems not supported by the device instance, exiting!");
};
auto broadcastAdd_invoker_ptr = broadcastAdd.MakeInvokerPointer();
@@ -129,12 +125,8 @@ int main()
c_m_n_device_buf.FromDevice(c_m_n.mData.data());
Tensor<CDataType> host_c_m_n(f_host_tensor_descriptor2d(M, N, Stride));
host_broadcast2D<Tensor<ABDataType>,
Tensor<ABDataType>,
Tensor<CDataType>,
EltwiseComputeDataType,
Add,
0>(host_c_m_n, a_m_n, b_n, M, N, Add{});
host_broadcast2D<Tensor<ABDataType>, Tensor<ABDataType>, Tensor<CDataType>, Add, 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 c", 1e-3, 1e-3);

73
example/19_binary_elementwise/broadcast_add_3d_am_bmnk.cpp
View File

@@ -6,7 +6,7 @@
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/device_binary_elementwise.hpp"
#include "ck/tensor_operation/gpu/device/device_elementwise.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
@@ -16,29 +16,21 @@
using F16 = ck::half_t;
using F32 = float;
using ABDataType = F16;
using CDataType = F16;
using EltwiseComputeDataType = F32;
using ABDataType = F16;
using CDataType = F16;
using Add = ck::tensor_operation::element_wise::Add;
using DeviceElementwiseAddInstance =
ck::tensor_operation::device::DeviceBinaryElementwise<ABDataType,
ABDataType,
CDataType,
EltwiseComputeDataType,
Add,
3,
8,
1,
8,
8>;
ck::tensor_operation::device::DeviceElementwise<ck::Tuple<ABDataType, ABDataType>,
ck::Tuple<CDataType>,
Add,
3,
8,
ck::Sequence<1, 8>,
ck::Sequence<8>>;
template <typename HostTensorA,
typename HostTensorB,
typename HostTensorC,
typename ComputeDataType,
typename Functor>
template <typename HostTensorA, typename HostTensorB, typename HostTensorC, typename Functor>
void host_broadcast3D_am_bmnk(HostTensorC& C,
const HostTensorA& A,
const HostTensorB& B,
@@ -51,11 +43,11 @@ void host_broadcast3D_am_bmnk(HostTensorC& C,
for(std::size_t n = 0; n < shape[1]; ++n)
for(std::size_t k = 0; k < shape[2]; ++k)
{
ComputeDataType a_val = ck::type_convert<ComputeDataType>(A(m));
ComputeDataType b_val = ck::type_convert<ComputeDataType>(B(m, n, k));
ComputeDataType c_val = 0;
auto a_val = A(m);
auto b_val = B(m, n, k);
ctype c_val = 0;
functor(c_val, a_val, b_val);
C(m, n, k) = ck::type_convert<ctype>(c_val);
C(m, n, k) = c_val;
}
}
@@ -85,25 +77,25 @@ int main()
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()};
std::array<ck::index_t, 3> abc_lengths;
std::array<ck::index_t, 3> a_strides = {1, 0, 0};
std::array<ck::index_t, 3> b_strides;
std::array<ck::index_t, 3> c_strides;
std::copy(mnk.begin(), mnk.end(), abc_lengths.begin());
std::copy(
b_m_n_k.mDesc.GetStrides().begin(), b_m_n_k.mDesc.GetStrides().end(), b_strides.begin());
std::copy(
c_m_n_k.mDesc.GetStrides().begin(), c_m_n_k.mDesc.GetStrides().end(), c_strides.begin());
auto broadcastAdd = DeviceElementwiseAddInstance{};
auto argument =
broadcastAdd.MakeArgumentPointer(input,
output,
std::vector<ck::index_t>{mnk.begin(), mnk.end()},
{a_strides, b_strides},
{c_strides},
Add{});
auto argument = broadcastAdd.MakeArgumentPointer(
abc_lengths, {a_strides, b_strides}, {c_strides}, input, output, Add{});
if(!broadcastAdd.IsSupportedArgument(argument.get()))
{
throw std::runtime_error("The runtime parameters seems not supported by the "
"DeviceBinaryElementwise instance, exiting!");
throw std::runtime_error(
"The runtime parameters seems not supported by the device instance, exiting!");
};
auto broadcastAdd_invoker_ptr = broadcastAdd.MakeInvokerPointer();
@@ -118,11 +110,8 @@ int main()
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{});
host_broadcast3D_am_bmnk<Tensor<ABDataType>, Tensor<ABDataType>, Tensor<CDataType>, 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);

58
example/19_binary_elementwise/elementwise_add_1d.cpp
View File

@@ -5,7 +5,7 @@
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/device_binary_elementwise.hpp"
#include "ck/tensor_operation/gpu/device/device_elementwise.hpp"
#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
@@ -15,29 +15,21 @@
using F16 = ck::half_t;
using F32 = float;
using ABDataType = F16;
using CDataType = F16;
using EltwiseComputeDataType = F32;
using ABDataType = F16;
using CDataType = F16;
using Add = ck::tensor_operation::element_wise::Add;
using DeviceElementwiseAddInstance =
ck::tensor_operation::device::DeviceBinaryElementwise<ABDataType,
ABDataType,
CDataType,
EltwiseComputeDataType,
Add,
1,
8,
8,
8,
8>;
ck::tensor_operation::device::DeviceElementwise<ck::Tuple<ABDataType, ABDataType>,
ck::Tuple<CDataType>,
Add,
1,
8,
ck::Sequence<8, 8>,
ck::Sequence<8>>;
template <typename HostTensorA,
typename HostTensorB,
typename HostTensorC,
typename ComputeDataType,
typename Functor>
template <typename HostTensorA, typename HostTensorB, typename HostTensorC, typename Functor>
void host_elementwise1D(
HostTensorC& C, const HostTensorA& A, const HostTensorB& B, int M, Functor functor)
{
@@ -45,11 +37,11 @@ void host_elementwise1D(
for(int m = 0; m < M; ++m)
{
ComputeDataType Am = ck::type_convert<ComputeDataType>(A(m));
ComputeDataType Bm = ck::type_convert<ComputeDataType>(B(m));
ComputeDataType Cm = 0;
auto Am = A(m);
auto Bm = B(m);
ctype Cm = 0;
functor(Cm, Am, Bm);
C(m) = ck::type_convert<ctype>(Cm);
C(m) = Cm;
}
}
@@ -83,18 +75,19 @@ int main()
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};
std::array<ck::index_t, 1> abc_lengths = {M};
std::array<ck::index_t, 1> a_strides = {1};
std::array<ck::index_t, 1> b_strides = {1};
std::array<ck::index_t, 1> c_strides = {1};
auto broadcastAdd = DeviceElementwiseAddInstance{};
auto argument = broadcastAdd.MakeArgumentPointer(
input, output, {M}, {{a_strides}, b_strides}, {c_strides}, Add{});
abc_lengths, {a_strides, b_strides}, {c_strides}, input, output, Add{});
if(!broadcastAdd.IsSupportedArgument(argument.get()))
{
throw std::runtime_error("The runtime parameters seems not supported by the "
"DeviceBinaryElementwise instance, exiting!");
throw std::runtime_error(
"The runtime parameters seems not supported by the device instance, exiting!");
};
auto broadcastAdd_invoker_ptr = broadcastAdd.MakeInvokerPointer();
@@ -109,11 +102,8 @@ int main()
c_m_device_buf.FromDevice(c_m.mData.data());
Tensor<CDataType> host_c_m(f_host_tensor_descriptor1d(M, 1));
host_elementwise1D<Tensor<ABDataType>,
Tensor<ABDataType>,
Tensor<CDataType>,
EltwiseComputeDataType,
Add>(host_c_m, a_m, b_m, M, Add{});
host_elementwise1D<Tensor<ABDataType>, Tensor<ABDataType>, Tensor<CDataType>, 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 c", 1e-3, 1e-3);

70
example/19_binary_elementwise/elementwise_add_4d.cpp
View File

@@ -6,7 +6,7 @@
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/device_binary_elementwise.hpp"
#include "ck/tensor_operation/gpu/device/device_elementwise.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
@@ -16,29 +16,21 @@
using F16 = ck::half_t;
using F32 = float;
using ABDataType = F16;
using CDataType = F16;
using EltwiseComputeDataType = F32;
using ABDataType = F16;
using CDataType = F16;
using Add = ck::tensor_operation::element_wise::Add;
using DeviceElementwiseAddInstance =
ck::tensor_operation::device::DeviceBinaryElementwise<ABDataType,
ABDataType,
CDataType,
EltwiseComputeDataType,
Add,
4,
8,
8,
8,
8>;
ck::tensor_operation::device::DeviceElementwise<ck::Tuple<ABDataType, ABDataType>,
ck::Tuple<CDataType>,
Add,
4,
8,
ck::Sequence<8, 8>,
ck::Sequence<8>>;
template <typename HostTensorA,
typename HostTensorB,
typename HostTensorC,
typename ComputeDataType,
typename Functor>
template <typename HostTensorA, typename HostTensorB, typename HostTensorC, typename Functor>
void host_elementwise4D(HostTensorC& C,
const HostTensorA& A,
const HostTensorB& B,
@@ -52,11 +44,11 @@ void host_elementwise4D(HostTensorC& C,
for(std::size_t h = 0; h < shape[2]; ++h)
for(std::size_t w = 0; w < shape[3]; ++w)
{
ComputeDataType a_val = ck::type_convert<ComputeDataType>(A(n, c, h, w));
ComputeDataType b_val = ck::type_convert<ComputeDataType>(B(n, c, h, w));
ComputeDataType c_val = 0;
auto a_val = A(n, c, h, w);
auto b_val = B(n, c, h, w);
ctype c_val = 0;
functor(c_val, a_val, b_val);
C(n, c, h, w) = ck::type_convert<ctype>(c_val);
C(n, c, h, w) = c_val;
}
}
@@ -85,23 +77,24 @@ int main()
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()};
std::array<ck::index_t, 4> abc_lengths;
std::array<ck::index_t, 4> a_strides;
std::array<ck::index_t, 4> b_strides;
std::array<ck::index_t, 4> c_strides;
std::copy(nchw.begin(), nchw.end(), abc_lengths.begin());
std::copy(a.mDesc.GetStrides().begin(), a.mDesc.GetStrides().end(), a_strides.begin());
std::copy(b.mDesc.GetStrides().begin(), b.mDesc.GetStrides().end(), b_strides.begin());
std::copy(c.mDesc.GetStrides().begin(), c.mDesc.GetStrides().end(), c_strides.begin());
auto broadcastAdd = DeviceElementwiseAddInstance{};
auto argument =
broadcastAdd.MakeArgumentPointer(input,
output,
std::vector<ck::index_t>{nchw.begin(), nchw.end()},
{{a_strides}, b_strides},
{c_strides},
Add{});
auto argument = broadcastAdd.MakeArgumentPointer(
abc_lengths, {a_strides, b_strides}, {c_strides}, input, output, Add{});
if(!broadcastAdd.IsSupportedArgument(argument.get()))
{
throw std::runtime_error("The runtime parameters seems not supported by the "
"DeviceBinaryElementwise instance, exiting!");
throw std::runtime_error(
"The runtime parameters seems not supported by the device instance, exiting!");
};
auto broadcastAdd_invoker_ptr = broadcastAdd.MakeInvokerPointer();
@@ -116,11 +109,8 @@ int main()
c_device_buf.FromDevice(c.mData.data());
Tensor<CDataType> host_c(nchw);
host_elementwise4D<Tensor<ABDataType>,
Tensor<ABDataType>,
Tensor<CDataType>,
EltwiseComputeDataType,
Add>(host_c, a, b, nchw, Add{});
host_elementwise4D<Tensor<ABDataType>, Tensor<ABDataType>, Tensor<CDataType>, Add>(
host_c, a, b, nchw, Add{});
pass &=
ck::utils::check_err(c.mData, host_c.mData, "Error: Incorrect results c", 1e-3, 1e-3);