mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-04-19 22:39:03 +00:00
Transpose 3d (#984)
* added working example for 5D input using 1D kernel * example with 5D input tensor and 2d kernel - not working: issues with arguments * added updated version of 3d device op - changed descriptors/dims * added example file to check kernel * fixed descriptor and isSupportedArgument stride problem * added and modified kernel for 3d - updated tids/loop * adding some more 5d example files * fixed some issues * changes made for testing * working version: fixed error in stride for A, still a bit inefficient * cleaned up formatting/comments * updating formatting * more formatting fixes * fixing cmake, adding back gpu targets in cmake script * adding client example * added instances for client example * fixed errors in client example * implemented client ex with device_elementwise.hpp and device_elementwise_3d_impl.hpp * removed extra files * minor formatting and naming fixes * adding test files and profiler * fixing minor error * minor fix * removed unneccesary comments, renamed files * updated instance list for client example, added different layout example * removing instances * fixed error in instance generation * remove comments * update profiler and client example tensor layouts * fixed errors in test/profiler * updated vector dim access to enable vector load * updated test/profiler files * updated example with 1d kernel * updating profiler * renamed files --------- Co-authored-by: Jing Zhang <jizha@amd.com>
This commit is contained in:
arai713
@@ -1,2 +1,4 @@
|
||||
add_example_executable(example_elementwise_permute_4D_fp16 elementwise_permute_4D_fp16.cpp)
|
||||
add_example_executable(example_elementwise_permute_4D_fp16_2d elementwise_permute_4D_fp16_2d.cpp)
|
||||
add_example_executable(example_elementwise_permute elementwise_permute.cpp)
|
||||
add_example_executable(example_elementwise_permute_3d elementwise_permute_3d.cpp)
|
||||
|
||||
135
example/44_elementwise_permute/elementwise_permute.cpp
Normal file
135
example/44_elementwise_permute/elementwise_permute.cpp
Normal file
@@ -0,0 +1,135 @@
|
||||
#include <iostream>
|
||||
#include <cstdlib>
|
||||
|
||||
#include "ck/ck.hpp"
|
||||
#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/impl/device_elementwise_impl.hpp"
|
||||
|
||||
#include "ck/library/utility/algorithm.hpp"
|
||||
#include "ck/library/utility/check_err.hpp"
|
||||
#include "ck/library/utility/device_memory.hpp"
|
||||
#include "ck/library/utility/host_tensor.hpp"
|
||||
#include "ck/library/utility/host_tensor_generator.hpp"
|
||||
|
||||
using F16 = ck::half_t;
|
||||
using F32 = float;
|
||||
|
||||
using ADataType = F16;
|
||||
using BDataType = F16;
|
||||
|
||||
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
|
||||
using DeviceElementwisePermuteInstance =
|
||||
ck::tensor_operation::device::DeviceElementwiseImpl<ck::Tuple<ADataType>, // InDataTypeTuple
|
||||
ck::Tuple<BDataType>, // OutDataTypeTuple
|
||||
PassThrough, // ElementwiseOp
|
||||
5, // NumDim
|
||||
8, // MPerThread
|
||||
ck::Sequence<1>, // InScalarPerVectorSeq
|
||||
ck::Sequence<1>>; // OutScalarPerVectorSeq
|
||||
|
||||
template <typename HostTensorA, typename HostTensorB, typename Functor>
|
||||
void host_elementwise4D(HostTensorB& B_ndhwc, const HostTensorA& A_ncdhw, Functor functor)
|
||||
{
|
||||
for(std::size_t n = 0; n < A_ncdhw.mDesc.GetLengths()[0]; ++n)
|
||||
for(std::size_t c = 0; c < A_ncdhw.mDesc.GetLengths()[1]; ++c)
|
||||
for(std::size_t d = 0; d < A_ncdhw.mDesc.GetLengths()[2]; ++d)
|
||||
for(std::size_t h = 0; h < A_ncdhw.mDesc.GetLengths()[3]; ++h)
|
||||
for(std::size_t w = 0; w < A_ncdhw.mDesc.GetLengths()[4]; ++w)
|
||||
{
|
||||
auto a_val = A_ncdhw(n, c, d, h, w);
|
||||
functor(B_ndhwc(n, d, h, w, c), a_val);
|
||||
}
|
||||
}
|
||||
|
||||
int main()
|
||||
{
|
||||
bool do_verification = true;
|
||||
bool time_kernel = true;
|
||||
|
||||
std::vector<std::size_t> ncdhw = {16, 8, 8, 8, 8};
|
||||
std::vector<std::size_t> ndhwc = {16, 8, 8, 8, 8};
|
||||
Tensor<ADataType> a(ncdhw);
|
||||
Tensor<BDataType> b(ndhwc);
|
||||
|
||||
a.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
|
||||
|
||||
DeviceMem a_device_buf(sizeof(ADataType) * a.mDesc.GetElementSpaceSize());
|
||||
DeviceMem b_device_buf(sizeof(BDataType) * b.mDesc.GetElementSpaceSize());
|
||||
|
||||
a_device_buf.ToDevice(a.mData.data());
|
||||
|
||||
std::array<const void*, 1> input = {a_device_buf.GetDeviceBuffer()};
|
||||
std::array<void*, 1> output = {b_device_buf.GetDeviceBuffer()};
|
||||
|
||||
std::array<ck::index_t, 5> ab_lengths;
|
||||
/**std::array<ck::index_t, 5> a_strides = {
|
||||
static_cast<int>(ncdhw[1] * ncdhw[2] * ncdhw[3] * ncdhw[4]),
|
||||
static_cast<int>(ncdhw[2] * ncdhw[3] * ncdhw[4]),
|
||||
static_cast<int>(ncdhw[3] * ncdhw[4]),
|
||||
static_cast<int>(ncdhw[4]),
|
||||
1};
|
||||
std::array<ck::index_t, 5> b_strides = {
|
||||
static_cast<int>(ndhwc[1] * ndhwc[2] * ndhwc[3] * ndhwc[4]),
|
||||
static_cast<int>(ndhwc[2] * ndhwc[3] * ndhwc[4]),
|
||||
1,
|
||||
static_cast<int>(ndhwc[3] * ndhwc[4]),
|
||||
static_cast<int>(ndhwc[4])};**/
|
||||
|
||||
std::array<ck::index_t, 5> a_strides = {
|
||||
static_cast<int>(ncdhw[1] * ncdhw[2] * ncdhw[3] * ncdhw[4]),
|
||||
static_cast<int>(ncdhw[3] * ncdhw[4]),
|
||||
static_cast<int>(ncdhw[4]),
|
||||
1,
|
||||
static_cast<int>(ncdhw[2] * ncdhw[3] * ncdhw[4])};
|
||||
|
||||
std::array<ck::index_t, 5> b_strides = {
|
||||
static_cast<int>(ndhwc[1] * ndhwc[2] * ndhwc[3] * ndhwc[4]),
|
||||
static_cast<int>(ndhwc[2] * ndhwc[3] * ndhwc[4]),
|
||||
static_cast<int>(ndhwc[3] * ndhwc[4]),
|
||||
static_cast<int>(ndhwc[4]),
|
||||
1};
|
||||
ck::ranges::copy(ncdhw, ab_lengths.begin());
|
||||
|
||||
auto broadcastPermute = DeviceElementwisePermuteInstance{};
|
||||
auto argument = broadcastPermute.MakeArgumentPointer(
|
||||
ab_lengths, {a_strides}, {b_strides}, input, output, PassThrough{});
|
||||
|
||||
if(!broadcastPermute.IsSupportedArgument(argument.get()))
|
||||
{
|
||||
throw std::runtime_error(
|
||||
"The runtime parameters seems not supported by the device instance, exiting!");
|
||||
};
|
||||
|
||||
std::cout << "A (ncdhw): " << a.mDesc << std::endl;
|
||||
std::cout << "B (ndhwc): " << b.mDesc << std::endl;
|
||||
|
||||
auto broadcastPermute_invoker_ptr = broadcastPermute.MakeInvokerPointer();
|
||||
float ave_time =
|
||||
broadcastPermute_invoker_ptr->Run(argument.get(), StreamConfig{nullptr, time_kernel});
|
||||
std::size_t flop = std::size_t(2) * ncdhw[0] * ncdhw[1] * ncdhw[2] * ncdhw[3] * ncdhw[4];
|
||||
|
||||
std::size_t num_btype =
|
||||
sizeof(ADataType) * (ncdhw[0] * ncdhw[1] * ncdhw[2] * ncdhw[3] * ncdhw[4]) +
|
||||
sizeof(BDataType) * (ncdhw[0] * ncdhw[1] * ncdhw[2] * ncdhw[3] * ncdhw[4]);
|
||||
|
||||
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"
|
||||
<< std::endl;
|
||||
|
||||
bool pass = true;
|
||||
|
||||
if(do_verification)
|
||||
{
|
||||
b_device_buf.FromDevice(b.mData.data());
|
||||
Tensor<BDataType> host_b(ndhwc);
|
||||
host_elementwise4D(host_b, a, PassThrough{});
|
||||
|
||||
pass &=
|
||||
ck::utils::check_err(b.mData, host_b.mData, "Error: Incorrect results b", 1e-3, 1e-3);
|
||||
}
|
||||
|
||||
return pass ? 0 : 1;
|
||||
}
|
||||
120
example/44_elementwise_permute/elementwise_permute_3d.cpp
Normal file
120
example/44_elementwise_permute/elementwise_permute_3d.cpp
Normal file
@@ -0,0 +1,120 @@
|
||||
#include <iostream>
|
||||
#include <cstdlib>
|
||||
|
||||
#include "ck/ck.hpp"
|
||||
#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/impl/device_elementwise_3d_impl.hpp"
|
||||
|
||||
#include "ck/library/utility/algorithm.hpp"
|
||||
#include "ck/library/utility/check_err.hpp"
|
||||
#include "ck/library/utility/device_memory.hpp"
|
||||
#include "ck/library/utility/host_tensor.hpp"
|
||||
#include "ck/library/utility/host_tensor_generator.hpp"
|
||||
|
||||
using F16 = ck::half_t;
|
||||
using F32 = float;
|
||||
|
||||
using ADataType = F16;
|
||||
using BDataType = F16;
|
||||
|
||||
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
|
||||
using DeviceElementwisePermuteInstance =
|
||||
ck::tensor_operation::device::DeviceElementwise3dImpl<ck::Tuple<ADataType>, // InDataTypeTuple
|
||||
ck::Tuple<BDataType>, // OutDataTypeTuple
|
||||
PassThrough, // ElementwiseOp
|
||||
2, // NumDim_m, {N, C}
|
||||
2, // NumDim_n, {H, W}
|
||||
1, // NumDim_k, {D}
|
||||
8, // MPerThread
|
||||
8, // NPerThread
|
||||
8, // KPerThread
|
||||
ck::Sequence<8>, // InScalarPerVectorSeq
|
||||
ck::Sequence<4>>; // OutScalarPerVectorSeq
|
||||
|
||||
template <typename HostTensorA, typename HostTensorB, typename Functor>
|
||||
void host_elementwise4D(HostTensorB& B_ndhwc, const HostTensorA& A_ncdhw, Functor functor)
|
||||
{
|
||||
for(std::size_t n = 0; n < A_ncdhw.mDesc.GetLengths()[0]; ++n)
|
||||
for(std::size_t c = 0; c < A_ncdhw.mDesc.GetLengths()[1]; ++c)
|
||||
for(std::size_t d = 0; d < A_ncdhw.mDesc.GetLengths()[2]; ++d)
|
||||
for(std::size_t h = 0; h < A_ncdhw.mDesc.GetLengths()[3]; ++h)
|
||||
for(std::size_t w = 0; w < A_ncdhw.mDesc.GetLengths()[4]; ++w)
|
||||
{
|
||||
auto a_val = A_ncdhw(n, c, d, h, w);
|
||||
functor(B_ndhwc(n, d, h, w, c), a_val);
|
||||
}
|
||||
}
|
||||
|
||||
int main()
|
||||
{
|
||||
bool do_verification = true;
|
||||
bool time_kernel = true;
|
||||
|
||||
const int N = 4;
|
||||
const int C = 16;
|
||||
const int H = 32;
|
||||
const int W = 5;
|
||||
const int D = 16;
|
||||
|
||||
std::vector<std::size_t> ncdhw = {N, C, D, H, W};
|
||||
std::vector<std::size_t> ndhwc = {N, D, H, W, C};
|
||||
Tensor<ADataType> a(ncdhw);
|
||||
Tensor<BDataType> b(ndhwc);
|
||||
|
||||
a.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
|
||||
|
||||
DeviceMem a_device_buf(sizeof(ADataType) * a.mDesc.GetElementSpaceSize());
|
||||
DeviceMem b_device_buf(sizeof(BDataType) * b.mDesc.GetElementSpaceSize());
|
||||
|
||||
a_device_buf.ToDevice(a.mData.data());
|
||||
|
||||
std::array<const void*, 1> input = {a_device_buf.GetDeviceBuffer()};
|
||||
std::array<void*, 1> output = {b_device_buf.GetDeviceBuffer()};
|
||||
|
||||
std::array<ck::index_t, 5> ab_lengths{N, C, H, W, D};
|
||||
std::array<ck::index_t, 5> a_strides = {C * D * H * W, H * W, W, 1, D * H * W}; // N, C, D, H, W
|
||||
std::array<ck::index_t, 5> b_strides = {C * H * W * D, H * W * D, W * D, D, 1}; // N, D, H, W, C
|
||||
|
||||
auto broadcastPermute = DeviceElementwisePermuteInstance{};
|
||||
auto argument = broadcastPermute.MakeArgumentPointer(
|
||||
ab_lengths, {a_strides}, {b_strides}, input, output, PassThrough{});
|
||||
|
||||
if(!broadcastPermute.IsSupportedArgument(argument.get()))
|
||||
{
|
||||
throw std::runtime_error(
|
||||
"The runtime parameters seems not supported by the device instance, exiting!");
|
||||
};
|
||||
|
||||
std::cout << "A (ncdhw): " << a.mDesc << std::endl;
|
||||
std::cout << "B (ndhwc): " << b.mDesc << std::endl;
|
||||
|
||||
auto broadcastPermute_invoker_ptr = broadcastPermute.MakeInvokerPointer();
|
||||
float ave_time =
|
||||
broadcastPermute_invoker_ptr->Run(argument.get(), StreamConfig{nullptr, time_kernel});
|
||||
std::size_t flop = std::size_t(2) * ncdhw[0] * ncdhw[1] * ncdhw[2] * ncdhw[3] * ncdhw[4];
|
||||
|
||||
std::size_t num_btype =
|
||||
sizeof(ADataType) * (ncdhw[0] * ncdhw[1] * ncdhw[2] * ncdhw[3] * ncdhw[4]) +
|
||||
sizeof(BDataType) * (ncdhw[0] * ncdhw[1] * ncdhw[2] * ncdhw[3] * ncdhw[4]);
|
||||
|
||||
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"
|
||||
<< std::endl;
|
||||
|
||||
bool pass = true;
|
||||
|
||||
if(do_verification)
|
||||
{
|
||||
b_device_buf.FromDevice(b.mData.data());
|
||||
Tensor<BDataType> host_b(ndhwc);
|
||||
host_elementwise4D(host_b, a, PassThrough{});
|
||||
|
||||
pass &=
|
||||
ck::utils::check_err(b.mData, host_b.mData, "Error: Incorrect results b", 1e-3, 1e-3);
|
||||
}
|
||||
|
||||
return pass ? 0 : 1;
|
||||
}
|
||||
@@ -19,13 +19,13 @@ using BDataType = F16;
|
||||
|
||||
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
|
||||
using DeviceElementwisePermuteInstance =
|
||||
ck::tensor_operation::device::DeviceElementwiseImpl<ck::Tuple<ADataType>,
|
||||
ck::Tuple<BDataType>,
|
||||
PassThrough,
|
||||
4,
|
||||
8,
|
||||
ck::Sequence<8>,
|
||||
ck::Sequence<1>>;
|
||||
ck::tensor_operation::device::DeviceElementwiseImpl<ck::Tuple<ADataType>, // InDataTypeTuple
|
||||
ck::Tuple<BDataType>, // OutDataTypeTuple
|
||||
PassThrough, // Elementwise op
|
||||
4, // NumDim
|
||||
8, // MPerThread
|
||||
ck::Sequence<8>, // InScalarPerVectorSeq
|
||||
ck::Sequence<1>>; // OutScalarPerVectorSeq
|
||||
|
||||
template <typename HostTensorA, typename HostTensorB, typename Functor>
|
||||
void host_elementwise4D(HostTensorB& B_nhwc, const HostTensorA& A_nchw, Functor functor)
|
||||
@@ -99,7 +99,6 @@ int main()
|
||||
|
||||
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
|
||||
<< std::endl;
|
||||
|
||||
bool pass = true;
|
||||
|
||||
if(do_verification)
|
||||
|
||||
@@ -17,15 +17,15 @@ using BDataType = F16;
|
||||
|
||||
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
|
||||
using DeviceElementwisePermuteInstance =
|
||||
ck::tensor_operation::device::DeviceElementwise2dImpl<ck::Tuple<ADataType>,
|
||||
ck::Tuple<BDataType>,
|
||||
PassThrough,
|
||||
3, // NumDim_M
|
||||
1, // NumDim_N
|
||||
8,
|
||||
8,
|
||||
ck::Sequence<8>,
|
||||
ck::Sequence<8>>;
|
||||
ck::tensor_operation::device::DeviceElementwise2dImpl<ck::Tuple<ADataType>, // InDataTypeTuple
|
||||
ck::Tuple<BDataType>, // OutDataTypeTuple
|
||||
PassThrough, // Elementwise op
|
||||
3, // NumDim_M
|
||||
1, // NumDim_N
|
||||
1, // MPerThread
|
||||
1, // NPerThread
|
||||
ck::Sequence<1>, // InScalarPerVectorSeq
|
||||
ck::Sequence<1>>; // OutScalarPerVectorSeq
|
||||
|
||||
template <typename HostTensorA, typename HostTensorB, typename Functor>
|
||||
void host_elementwise4D(HostTensorB& B_nhwc,
|
||||
@@ -53,12 +53,6 @@ int main()
|
||||
const int H = 32;
|
||||
const int W = 1024;
|
||||
|
||||
/**const int N = 120;
|
||||
const int H = 32;
|
||||
const int W = 64;
|
||||
|
||||
const int C = 128;**/
|
||||
|
||||
std::vector<std::size_t> nchw = {N, C, H, W};
|
||||
std::vector<std::size_t> nhwc = {N, H, W, C};
|
||||
|
||||
@@ -71,7 +65,6 @@ int main()
|
||||
DeviceMem b_device_buf(sizeof(BDataType) * b.mDesc.GetElementSpaceSize());
|
||||
|
||||
a_device_buf.ToDevice(a.mData.data());
|
||||
// LogRangeAsType<float>(std::cout << "Tensor a : ", a.mData, ",") << std::endl;
|
||||
|
||||
std::array<const void*, 1> input = {a_device_buf.GetDeviceBuffer()};
|
||||
std::array<void*, 1> output = {b_device_buf.GetDeviceBuffer()};
|
||||
@@ -115,13 +108,10 @@ int main()
|
||||
if(do_verification)
|
||||
{
|
||||
b_device_buf.FromDevice(b.mData.data());
|
||||
// LogRangeAsType<float>(std::cout << "Tensor b : ", b.mData, ",") << std::endl;
|
||||
|
||||
Tensor<BDataType> host_b(nhwc);
|
||||
host_elementwise4D<Tensor<ADataType>, Tensor<BDataType>, PassThrough>(
|
||||
host_b, a, nchw, PassThrough{});
|
||||
|
||||
// LogRangeAsType<float>(std::cout << "Host b : ", host_b.mData, ",") << std::endl;
|
||||
pass &=
|
||||
ck::utils::check_err(b.mData, host_b.mData, "Error: Incorrect results b", 1e-3, 1e-3);
|
||||
}
|
||||
|
||||
Reference in New Issue
Block a user