Skip to content

GitLab

Commit 6fcaeada authored by Astha Rai's avatar Astha Rai
Browse files

fixed merge conflict after merge with develop

parents fc7a1825 d02a92cc
Hide whitespace changes
Inline Side-by-side
Showing with 1296 additions and 49 deletions
example/01_gemm/gemm_xdl_fp8_bf8.cpp
View file @ 6fcaeada
// SPDX-License-Identifier: MIT // SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved. // Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp" #include "common.hpp"
...@@ -44,6 +44,17 @@ using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataTyp ...@@ -44,6 +44,17 @@ using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataTyp
ComputeTypeA, ComputeTypeA,
ComputeTypeB>; ComputeTypeB>;
using ReferenceGemmInstanceGPU = ck::tensor_operation::device::ReferenceGemm<ALayout,
BLayout,
CLayout,
ADataType,
BDataType,
CDataType,
AccDataType,
AElementOp,
BElementOp,
CElementOp>;
#include "run_gemm_example.inc" #include "run_gemm_example.inc"
int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); } int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); }
example/01_gemm/gemm_xdl_int8.cpp
View file @ 6fcaeada
// SPDX-License-Identifier: MIT // SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved. // Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp" #include "common.hpp"
...@@ -33,6 +33,17 @@ using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle ...@@ -33,6 +33,17 @@ using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle
using ReferenceGemmInstance = ck::tensor_operation::host:: using ReferenceGemmInstance = ck::tensor_operation::host::
ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>; ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
using ReferenceGemmInstanceGPU = ck::tensor_operation::device::ReferenceGemm<ALayout,
BLayout,
CLayout,
ADataType,
BDataType,
CDataType,
AccDataType,
AElementOp,
BElementOp,
CElementOp>;
#include "run_gemm_example.inc" #include "run_gemm_example.inc"
int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); } int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); }
example/01_gemm/gemm_xdl_lds_direct_load_fp16.cpp
View file @ 6fcaeada
// SPDX-License-Identifier: MIT // SPDX-License-Identifier: MIT
// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved. // Copyright (c) 2023-2024, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream> #include <iostream>
...@@ -53,6 +53,17 @@ using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle ...@@ -53,6 +53,17 @@ using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle
using ReferenceGemmInstance = ck::tensor_operation::host:: using ReferenceGemmInstance = ck::tensor_operation::host::
ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>; ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
using ReferenceGemmInstanceGPU = ck::tensor_operation::device::ReferenceGemm<ALayout,
BLayout,
CLayout,
ADataType,
BDataType,
CDataType,
AccDataType,
AElementOp,
BElementOp,
CElementOp>;
#include "run_gemm_example.inc" #include "run_gemm_example.inc"
int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); } int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); }
example/01_gemm/gemm_xdl_lds_direct_load_fp32.cpp
View file @ 6fcaeada
// SPDX-License-Identifier: MIT // SPDX-License-Identifier: MIT
// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved. // Copyright (c) 2023-2024, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream> #include <iostream>
...@@ -52,6 +52,17 @@ using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle ...@@ -52,6 +52,17 @@ using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle
using ReferenceGemmInstance = ck::tensor_operation::host:: using ReferenceGemmInstance = ck::tensor_operation::host::
ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>; ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
using ReferenceGemmInstanceGPU = ck::tensor_operation::device::ReferenceGemm<ALayout,
BLayout,
CLayout,
ADataType,
BDataType,
CDataType,
AccDataType,
AElementOp,
BElementOp,
CElementOp>;
#include "run_gemm_example.inc" #include "run_gemm_example.inc"
int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); } int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); }
example/01_gemm/gemm_xdl_streamk.cpp
View file @ 6fcaeada
// SPDX-License-Identifier: MIT // SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. // Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp" #include "common.hpp"
...@@ -44,6 +44,17 @@ using DeviceGemmInstance = DeviceGemmStreamK; ...@@ -44,6 +44,17 @@ using DeviceGemmInstance = DeviceGemmStreamK;
using ReferenceGemmInstance = ck::tensor_operation::host:: using ReferenceGemmInstance = ck::tensor_operation::host::
ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>; ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
using ReferenceGemmInstanceGPU = ck::tensor_operation::device::ReferenceGemm<ALayout,
BLayout,
CLayout,
ADataType,
BDataType,
CDataType,
AccDataType,
AElementOp,
BElementOp,
CElementOp>;
#include "run_gemm_example.inc" #include "run_gemm_example.inc"
int main(int argc, char* argv[]) { return !run_gemm_streamk_example(argc, argv); } int main(int argc, char* argv[]) { return !run_gemm_streamk_example(argc, argv); }
example/01_gemm/gemm_xdl_wavelet_fp16.cpp
View file @ 6fcaeada
// SPDX-License-Identifier: MIT // SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved. // Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp" #include "common.hpp"
...@@ -37,6 +37,17 @@ using DeviceGemmInstance = DeviceGemmInstance; ...@@ -37,6 +37,17 @@ using DeviceGemmInstance = DeviceGemmInstance;
using ReferenceGemmInstance = ck::tensor_operation::host:: using ReferenceGemmInstance = ck::tensor_operation::host::
ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>; ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
using ReferenceGemmInstanceGPU = ck::tensor_operation::device::ReferenceGemm<ALayout,
BLayout,
CLayout,
ADataType,
BDataType,
CDataType,
AccDataType,
AElementOp,
BElementOp,
CElementOp>;
#include "run_gemm_example.inc" #include "run_gemm_example.inc"
int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); } int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); }
example/01_gemm/run_gemm_example.inc
View file @ 6fcaeada
...@@ -173,6 +173,7 @@ bool run_gemm(const ProblemType& problem_size, const ExecutionConfig& config) ...@@ -173,6 +173,7 @@ bool run_gemm(const ProblemType& problem_size, const ExecutionConfig& config)
Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{})); Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{})); Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
Tensor<CDataType> c_m_n_device_ref_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
std::cout << "a_m_k: " << a_m_k.mDesc << std::endl; std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl; std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
...@@ -193,6 +194,8 @@ bool run_gemm(const ProblemType& problem_size, const ExecutionConfig& config) ...@@ -193,6 +194,8 @@ bool run_gemm(const ProblemType& problem_size, const ExecutionConfig& config)
DeviceMem a_m_k_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize()); DeviceMem a_m_k_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
DeviceMem b_k_n_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize()); DeviceMem b_k_n_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
DeviceMem c_m_n_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpaceSize()); DeviceMem c_m_n_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpaceSize());
DeviceMem c_m_n_device_ref_buf(sizeof(CDataType) *
c_m_n_device_ref_result.mDesc.GetElementSpaceSize());
a_m_k_device_buf.ToDevice(a_m_k.mData.data()); a_m_k_device_buf.ToDevice(a_m_k.mData.data());
b_k_n_device_buf.ToDevice(b_k_n.mData.data()); b_k_n_device_buf.ToDevice(b_k_n.mData.data());
...@@ -325,14 +328,18 @@ bool run_gemm(const ProblemType& problem_size, const ExecutionConfig& config) ...@@ -325,14 +328,18 @@ bool run_gemm(const ProblemType& problem_size, const ExecutionConfig& config)
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, " std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
<< gemm.GetTypeString() << std::endl; << gemm.GetTypeString() << std::endl;
bool pass = true;
if(config.do_verification) if(config.do_verification)
{ {
// CPU verification
auto ref_gemm = ReferenceGemmInstance{}; auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker(); auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument( auto ref_argument = ref_gemm.MakeArgument(
a_m_k, b_k_n, c_m_n_host_result, a_element_op, b_element_op, c_element_op); a_m_k, b_k_n, c_m_n_host_result, a_element_op, b_element_op, c_element_op);
std::cout << "Running verification on CPU." << std::endl;
ref_invoker.Run(ref_argument); ref_invoker.Run(ref_argument);
#ifdef BUILD_INT4_EXAMPLE #ifdef BUILD_INT4_EXAMPLE
...@@ -346,15 +353,42 @@ bool run_gemm(const ProblemType& problem_size, const ExecutionConfig& config) ...@@ -346,15 +353,42 @@ bool run_gemm(const ProblemType& problem_size, const ExecutionConfig& config)
#else #else
c_m_n_device_buf.FromDevice(c_m_n_device_result.mData.data()); c_m_n_device_buf.FromDevice(c_m_n_device_result.mData.data());
return ck::utils::check_err(c_m_n_device_result, pass &= !ck::utils::check_err(c_m_n_device_result,
c_m_n_host_result, c_m_n_host_result,
"Error: Incorrect results!", "Error: Incorrect results!",
get_rtol<CDataType>(), get_rtol<CDataType>(),
get_atol<CDataType>()); get_atol<CDataType>());
#endif #endif
// GPU verification
auto ref_gemm_gpu = ReferenceGemmInstanceGPU{};
auto ref_invoker_gpu = ref_gemm_gpu.MakeInvoker();
auto ref_argument_gpu = ref_gemm_gpu.MakeArgument(
static_cast<ADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_m_n_device_ref_buf.GetDeviceBuffer()),
M,
N,
K,
a_element_op,
b_element_op,
c_element_op);
std::cout << "Running verification on GPU." << std::endl;
ref_invoker_gpu.Run(ref_argument_gpu, StreamConfig{});
c_m_n_device_ref_buf.FromDevice(c_m_n_device_ref_result.mData.data());
c_m_n_device_buf.FromDevice(c_m_n_device_result.mData.data());
pass &= !ck::utils::check_err(c_m_n_device_result,
c_m_n_device_ref_result,
"Error: Incorrect results!",
get_rtol<CDataType>(),
get_atol<CDataType>());
} }
return true; return !pass;
} }
bool run_gemm_example(int argc, char* argv[]) bool run_gemm_example(int argc, char* argv[])
......
example/01_gemm/run_gemm_example_streamk_v2.inc
View file @ 6fcaeada
...@@ -117,9 +117,9 @@ bool run_gemm(const ProblemType& problem_size, const ExecutionConfig& config) ...@@ -117,9 +117,9 @@ bool run_gemm(const ProblemType& problem_size, const ExecutionConfig& config)
auto f_get_default_stride = auto f_get_default_stride =
[](std::size_t row, std::size_t col, ck::index_t stride, auto layout) { [](std::size_t row, std::size_t col, ck::index_t stride, auto layout) {
if(stride == -1) if(stride == 0)
{ {
// give a chance if stride is -1, return a default packed stride // give a chance if stride is 0, return a default packed stride
if constexpr(std::is_same_v<decltype(layout), ck::tensor_layout::gemm::RowMajor>) if constexpr(std::is_same_v<decltype(layout), ck::tensor_layout::gemm::RowMajor>)
{ {
return static_cast<std::size_t>(col); return static_cast<std::size_t>(col);
......
example/44_elementwise_permute/CMakeLists.txt
View file @ 6fcaeada
...@@ -5,3 +5,4 @@ add_example_executable(example_elementwise_permute_4D_fp32_col elementwise_permu ...@@ -5,3 +5,4 @@ add_example_executable(example_elementwise_permute_4D_fp32_col elementwise_permu
add_example_executable(example_elementwise_permute_4D_fp16_col elementwise_permute_4D_fp16_col.cpp) add_example_executable(example_elementwise_permute_4D_fp16_col elementwise_permute_4D_fp16_col.cpp)
add_example_executable(example_elementwise_binary_4D_fp16 elementwise_binary_4D_fp16.cpp) add_example_executable(example_elementwise_binary_4D_fp16 elementwise_binary_4D_fp16.cpp)
add_example_executable(example_elementwise_trinary_4D_fp16 elementwise_trinary_4D_fp16.cpp) add_example_executable(example_elementwise_trinary_4D_fp16 elementwise_trinary_4D_fp16.cpp)
add_example_executable(elementwise_scale_permute_amax_2D_fp16_fp8 elementwise_scale_permute_amax_2D_fp16_fp8.cpp)
example/44_elementwise_permute/elementwise_scale_permute_amax_2D_fp16_fp8.cpp 0 → 100644
View file @ 6fcaeada
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#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_dynamic_vector_dims_impl.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_reduce_multiblock.hpp"
#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_elementwise.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"
#include "ck/utility/reduction_enums.hpp"
using F16 = ck::half_t;
using F32 = float;
using F8 = ck::f8_t;
using InputDataType = F16;
using ScaleDataType = F32;
using OutputDataType = F8;
static constexpr ck::index_t NumDim = 2;
constexpr ck::ReduceTensorOp ReduceOpId = ck::ReduceTensorOp::MAX;
constexpr bool PropagateNan = true;
constexpr bool OutputIndex = false;
using ReduceOperation = typename ck::reduce_binary_operator<ReduceOpId>::opType;
struct ScalePassThrough
{
ScalePassThrough(const float alpha = 1.f) : alpha_(alpha) {}
__host__ __device__ constexpr void
operator()(OutputDataType& y0, OutputDataType& y1, const InputDataType& x0) const
{
y0 = ck::type_convert<OutputDataType>(ck::type_convert<ScaleDataType>(x0) * alpha_);
y1 = y0;
}
const ScaleDataType alpha_;
};
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using UnaryAbs = ck::tensor_operation::element_wise::UnaryAbs;
using DeviceElementwisePermuteInstance = ck::tensor_operation::device::DeviceElementwiseImpl<
ck::Tuple<InputDataType>, // InDataTypeTuple
ck::Tuple<OutputDataType, OutputDataType>, // OutDataTypeTuple
ScalePassThrough, // Elementwise
NumDim, // NumDim
256, // BlockSize
128, // M0PerBlock
128, // M1PerBlock
8, // M0PerThread
8, // M1PerThread
ck::Sequence<1, 0>, // ThreadClusterArrangeOrder
ck::Sequence<8>, // InScalarPerVectorSeq
ck::Sequence<8, 1>>; // OutScalarPerVectorSeq
using DeviceReduceInstance =
ck::tensor_operation::device::DeviceReduceMultiBlock<OutputDataType,
OutputDataType,
OutputDataType,
NumDim,
NumDim,
ReduceOperation,
UnaryAbs,
PassThrough,
ck::InMemoryDataOperationEnum::Set,
PropagateNan,
OutputIndex,
false, // HaveIndexInputIfOutputIndex
1024, // BlockSize
1, // MThreadClusterSize
1024, // KThreadClusterSize
1, // MThreadSliceSize
16, // KThreadSliceSize
1, // InSrcVectorDim
16, // InSrceVectorSize
1>; // OutDstVectorSize
void reference_scale_permute_amax(Tensor<InputDataType>& input,
Tensor<OutputDataType>& host_output_scaled_casted_transposed,
Tensor<OutputDataType>& host_output_scaled_casted,
Tensor<OutputDataType>& host_output_amax,
const float scale)
{
ScalePassThrough out_element_op(scale);
const ck::index_t M = input.GetLengths()[0];
const ck::index_t K = input.GetLengths()[1];
for(ck::index_t m = 0; m < M; m++)
{
for(ck::index_t k = 0; k < K; k++)
{
OutputDataType y0, y1;
out_element_op(y0, y1, input(m, k));
host_output_scaled_casted(m, k) = y0;
host_output_scaled_casted_transposed(m, k) = y1;
const OutputDataType y_fabs =
ck::type_convert<OutputDataType>(ck::math::abs(ck::type_convert<float>(y0)));
host_output_amax(0) = ck::math::max(y_fabs, host_output_amax(0));
}
}
}
int main(int argc, char* argv[])
{
bool do_verification = true;
bool time_kernel = true;
const float scale = 2.f;
ck::index_t M = 1024;
ck::index_t K = 1024;
if(argc == 3)
{
M = std::stoi(argv[1]);
K = std::stoi(argv[2]);
}
std::array<ck::index_t, 2> dims = {M, K};
std::array<ck::index_t, 2> in_strides = {K, 1};
std::array<ck::index_t, 2> out_strides = {1, M};
Tensor<InputDataType> input(dims, in_strides);
Tensor<OutputDataType> output_scaled_casted_transposed(dims, out_strides);
Tensor<OutputDataType> output_scaled_casted(dims, in_strides);
Tensor<OutputDataType> output_amax({1});
input.GenerateTensorValue(GeneratorTensor_3<InputDataType>{0.0, 1.0});
DeviceMem input_dev_buf(sizeof(InputDataType) * input.mDesc.GetElementSpaceSize());
DeviceMem output_scaled_casted_transposed_dev_buf(
sizeof(OutputDataType) * output_scaled_casted_transposed.mDesc.GetElementSpaceSize());
DeviceMem output_scaled_casted_dev_buf(sizeof(OutputDataType) *
output_scaled_casted.mDesc.GetElementSpaceSize());
DeviceMem output_amax_dev_buf(sizeof(OutputDataType) * output_amax.mDesc.GetElementSpaceSize());
input_dev_buf.ToDevice(input.mData.data());
std::array<const void*, 1> inputs = {input_dev_buf.GetDeviceBuffer()};
std::array<void*, 2> outputs = {output_scaled_casted_transposed_dev_buf.GetDeviceBuffer(),
output_scaled_casted_dev_buf.GetDeviceBuffer()};
std::cout << "Input: " << input.mDesc << std::endl;
std::cout << "Scale: " << scale << std::endl;
std::cout << "Output scaled casted transposed: " << output_scaled_casted_transposed.mDesc
<< std::endl;
std::cout << "Output scaled casted: " << output_scaled_casted.mDesc << std::endl;
std::cout << "Output amax: " << output_amax.mDesc << std::endl;
auto launch_transpose_scale = [&]() {
auto transposeScale = DeviceElementwisePermuteInstance{};
auto argument = transposeScale.MakeArgumentPointer(dims,
{in_strides},
{out_strides, in_strides},
inputs,
outputs,
ScalePassThrough{scale});
if(!transposeScale.IsSupportedArgument(argument.get()))
{
throw std::runtime_error(
"The runtime parameters seems not supported by the device instance, exiting!");
};
auto transposeScale_invoker_ptr = transposeScale.MakeInvokerPointer();
return transposeScale_invoker_ptr->Run(argument.get(), StreamConfig{nullptr, time_kernel});
};
auto launch_reduce = [&]() {
auto reduce = DeviceReduceInstance{};
auto reduce_argument_ptr =
reduce.MakeArgumentPointer(dims,
in_strides,
{1}, // Output Lengths
{1}, // Output Strides
{0, 1}, // Reduce Dims
static_cast<double>(1.f),
static_cast<double>(0.f),
output_scaled_casted_dev_buf.GetDeviceBuffer(),
nullptr,
output_amax_dev_buf.GetDeviceBuffer(),
nullptr,
UnaryAbs{},
PassThrough{});
if(!reduce.IsSupportedArgument(reduce_argument_ptr.get()))
{
throw std::runtime_error(
"The runtime parameters seems not supported by the device instance, exiting!");
};
auto invoker_ptr = reduce.MakeInvokerPointer();
return invoker_ptr->Run(reduce_argument_ptr.get(), StreamConfig{nullptr, time_kernel});
};
float ave_time = launch_transpose_scale();
ave_time += launch_reduce();
std::cout << "Perf: " << ave_time << " ms" << std::endl;
bool pass = true;
if(do_verification)
{
Tensor<OutputDataType> host_output_scaled_casted_transposed(dims, out_strides);
Tensor<OutputDataType> host_output_scaled_casted(dims, in_strides);
Tensor<OutputDataType> host_output_amax({1});
reference_scale_permute_amax(input,
host_output_scaled_casted_transposed,
host_output_scaled_casted,
host_output_amax,
scale);
output_scaled_casted_transposed_dev_buf.FromDevice(
output_scaled_casted_transposed.mData.data());
output_scaled_casted_dev_buf.FromDevice(output_scaled_casted.mData.data());
output_amax_dev_buf.FromDevice(output_amax.mData.data());
pass &= ck::utils::check_err(output_scaled_casted_transposed.mData,
host_output_scaled_casted_transposed.mData,
"Error: Incorrect results scaled transposed",
1e-3,
1e-3);
pass &= ck::utils::check_err(output_scaled_casted.mData,
host_output_scaled_casted.mData,
"Error: Incorrect results scaled",
1e-3,
1e-3);
pass &= ck::utils::check_err(
output_amax.mData, host_output_amax.mData, "Error: Incorrect results amax", 1e-3, 1e-3);
}
return pass ? 0 : 1;
}
example/66_complex_contraction_bilinear/CMakeLists.txt 0 → 100755
View file @ 6fcaeada
add_example_executable(example_complex_contraction_bilinear_xdl_fp32 complex_contraction_bilinear_xdl_fp32.cpp)
add_example_executable(example_complex_contraction_bilinear_xdl_fp64 complex_contraction_bilinear_xdl_fp64.cpp)
example/66_complex_contraction_bilinear/README.md 0 → 100755
View file @ 6fcaeada
# Instructions for ```example_complex_contraction_bilinear_xdl_fp32```
## Run
```bash
#arg1: verification (0=no, 1=yes)
#arg2: initialization (0=no init, 1=integer value, 2=decimal value)
#arg3: time kernel (0=no, 1=yes)
./bin/example_contraction_bilinear_xdl_fp32 1 1 1
```
example/66_complex_contraction_bilinear/common_instances.hpp 0 → 100644
View file @ 6fcaeada
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_contraction_multiple_d_xdl_cshuffle.hpp"
using F16 = ck::half_t;
using BF16 = ck::bhalf_t;
using F32 = float;
using F64 = double;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
// Generic instances for fp32, fp16 and bf16 data types.
template <ck::index_t NumDimM,
ck::index_t NumDimN,
ck::index_t NumDimK,
typename ADataType,
typename BDataType,
typename AccDataType,
typename CShuffleDataType,
typename DsDataType,
typename EDataType,
typename ComputeDataType,
typename AElementOp,
typename BElementOp,
typename CDEElementOp>
// clang-format off
using DeviceOpInstanceKK_Generic = ck::tensor_operation::device::
//#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Compute|
//#####################################| | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| Data|
//#####################################| | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| Type|
//#####################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceContractionMultipleD_Xdl_CShuffle< NumDimM, NumDimN, NumDimK, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementOp, BElementOp, CDEElementOp, GemmSpec, 1, 256, 256, 128, 16, 4, 4, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>;
// clang-format on
template <ck::index_t NumDimM,
ck::index_t NumDimN,
ck::index_t NumDimK,
typename ADataType,
typename BDataType,
typename AccDataType,
typename CShuffleDataType,
typename DsDataType,
typename EDataType,
typename ComputeDataType,
typename AElementOp,
typename BElementOp,
typename CDEElementOp>
// clang-format off
using DeviceOpInstanceKN_Generic = ck::tensor_operation::device::
//#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Compute|
//#####################################| | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| Data|
//#####################################| | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| Type|
//#####################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceContractionMultipleD_Xdl_CShuffle< NumDimM, NumDimN, NumDimK, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementOp, BElementOp, CDEElementOp, GemmSpec, 1, 256, 256, 128, 16, 4, 1, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>;
// clang-format on
template <ck::index_t NumDimM,
ck::index_t NumDimN,
ck::index_t NumDimK,
typename ADataType,
typename BDataType,
typename AccDataType,
typename CShuffleDataType,
typename DsDataType,
typename EDataType,
typename ComputeDataType,
typename AElementOp,
typename BElementOp,
typename CDEElementOp>
// clang-format off
using DeviceOpInstanceMK_Generic = ck::tensor_operation::device::
//#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Compute|
//#####################################| | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| Data|
//#####################################| | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| Type|
//#####################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceContractionMultipleD_Xdl_CShuffle< NumDimM, NumDimN, NumDimK, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementOp, BElementOp, CDEElementOp, GemmSpec, 1, 256, 256, 128, 16, 1, 4, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>;
// clang-format on
template <ck::index_t NumDimM,
ck::index_t NumDimN,
ck::index_t NumDimK,
typename ADataType,
typename BDataType,
typename AccDataType,
typename CShuffleDataType,
typename DsDataType,
typename EDataType,
typename ComputeDataType,
typename AElementOp,
typename BElementOp,
typename CDEElementOp>
// clang-format off
using DeviceOpInstanceMN_Generic = ck::tensor_operation::device::
//#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Compute|
//#####################################| | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| Data|
//#####################################| | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| Type|
//#####################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceContractionMultipleD_Xdl_CShuffle< NumDimM, NumDimN, NumDimK, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementOp, BElementOp, CDEElementOp, GemmSpec, 1, 256, 256, 128, 16, 1, 1, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>;
// clang-format on
// Fp64 instances.
template <ck::index_t NumDimM,
ck::index_t NumDimN,
ck::index_t NumDimK,
typename ADataType,
typename BDataType,
typename AccDataType,
typename CShuffleDataType,
typename DsDataType,
typename EDataType,
typename ComputeDataType,
typename AElementOp,
typename BElementOp,
typename CDEElementOp>
// clang-format off
using DeviceOpInstanceKK_FP64 = ck::tensor_operation::device::
//#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Compute|
//#####################################| | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| Data|
//#####################################| | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| Type|
//#####################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceContractionMultipleD_Xdl_CShuffle< NumDimM, NumDimN, NumDimK, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementOp, BElementOp, CDEElementOp, GemmSpec, 1, 256, 128, 128, 16, 2, 2, 16, 16, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 2, 1, 1, 1, S<1, 16, 1, 16>, 1, ComputeDataType>;
// clang-format on
template <ck::index_t NumDimM,
ck::index_t NumDimN,
ck::index_t NumDimK,
typename ADataType,
typename BDataType,
typename AccDataType,
typename CShuffleDataType,
typename DsDataType,
typename EDataType,
typename ComputeDataType,
typename AElementOp,
typename BElementOp,
typename CDEElementOp>
// clang-format off
using DeviceOpInstanceKN_FP64 = ck::tensor_operation::device::
//#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Compute|
//#####################################| | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| Data|
//#####################################| | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| Type|
//#####################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceContractionMultipleD_Xdl_CShuffle< NumDimM, NumDimN, NumDimK, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementOp, BElementOp, CDEElementOp, GemmSpec, 1, 256, 128, 128, 16, 2, 1, 16, 16, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 2, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 1, 0, 1, 1, S<1, 16, 1, 16>, 1, ComputeDataType>;
// clang-format on
template <ck::index_t NumDimM,
ck::index_t NumDimN,
ck::index_t NumDimK,
typename ADataType,
typename BDataType,
typename AccDataType,
typename CShuffleDataType,
typename DsDataType,
typename EDataType,
typename ComputeDataType,
typename AElementOp,
typename BElementOp,
typename CDEElementOp>
// clang-format off
using DeviceOpInstanceMK_FP64 = ck::tensor_operation::device::
//#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Compute|
//#####################################| | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| Data|
//#####################################| | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| Type|
//#####################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceContractionMultipleD_Xdl_CShuffle< NumDimM, NumDimN, NumDimK, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementOp, BElementOp, CDEElementOp, GemmSpec, 1, 256, 128, 128, 16, 1, 2, 16, 16, 4, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 1, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 2, 1, 1, 1, S<1, 16, 1, 16>, 1, ComputeDataType>;
// clang-format on
template <ck::index_t NumDimM,
ck::index_t NumDimN,
ck::index_t NumDimK,
typename ADataType,
typename BDataType,
typename AccDataType,
typename CShuffleDataType,
typename DsDataType,
typename EDataType,
typename ComputeDataType,
typename AElementOp,
typename BElementOp,
typename CDEElementOp>
// clang-format off
using DeviceOpInstanceMN_FP64 = ck::tensor_operation::device::
//#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Compute|
//#####################################| | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| Data|
//#####################################| | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| Type|
//#####################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceContractionMultipleD_Xdl_CShuffle< NumDimM, NumDimN, NumDimK, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementOp, BElementOp, CDEElementOp, GemmSpec, 1, 256, 128, 128, 16, 1, 1, 16, 16, 4, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 1, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 1, 0, 1, 1, S<1, 16, 1, 16>, 1, ComputeDataType>;
// clang-format on
example/66_complex_contraction_bilinear/complex_contraction_bilinear_xdl_fp32.cpp 0 → 100755
View file @ 6fcaeada
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "common_instances.hpp"
using ADataType = F32;
using BDataType = F32;
using AccDataType = F32;
using CShuffleDataType = F32;
using DDataType = F32;
using DsDataType = ck::Tuple<DDataType>;
using EDataType = F32;
using ComputeDataType = F32;
static constexpr ck::index_t NumDimM = 2;
static constexpr ck::index_t NumDimN = 2;
static constexpr ck::index_t NumDimK = 2;
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CDEElementOp = ck::tensor_operation::element_wise::Bilinear;
using DeviceOpInstanceKKNN = DeviceOpInstanceKK_Generic<NumDimM,
NumDimN,
NumDimK,
ADataType,
BDataType,
AccDataType,
CShuffleDataType,
DsDataType,
EDataType,
ComputeDataType,
AElementOp,
BElementOp,
CDEElementOp>;
using DeviceOpInstanceKNNN = DeviceOpInstanceKN_Generic<NumDimM,
NumDimN,
NumDimK,
ADataType,
BDataType,
AccDataType,
CShuffleDataType,
DsDataType,
EDataType,
ComputeDataType,
AElementOp,
BElementOp,
CDEElementOp>;
using DeviceOpInstanceMKNN = DeviceOpInstanceMK_Generic<NumDimM,
NumDimN,
NumDimK,
ADataType,
BDataType,
AccDataType,
CShuffleDataType,
DsDataType,
EDataType,
ComputeDataType,
AElementOp,
BElementOp,
CDEElementOp>;
using DeviceOpInstanceMNNN = DeviceOpInstanceMN_Generic<NumDimM,
NumDimN,
NumDimK,
ADataType,
BDataType,
AccDataType,
CShuffleDataType,
DsDataType,
EDataType,
ComputeDataType,
AElementOp,
BElementOp,
CDEElementOp>;
using DeviceOpInstance = DeviceOpInstanceKKNN;
#include "run_complex_contraction_bilinear_example.inc"
int main(int argc, char* argv[]) { return run_complex_contraction_bilinear_example(argc, argv); }
example/66_complex_contraction_bilinear/complex_contraction_bilinear_xdl_fp64.cpp 0 → 100755
View file @ 6fcaeada
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "common_instances.hpp"
using ADataType = F64;
using BDataType = F64;
using AccDataType = F64;
using CShuffleDataType = F64;
using DDataType = F64;
using DsDataType = ck::Tuple<DDataType>;
using EDataType = F64;
using ComputeDataType = F64;
static constexpr ck::index_t NumDimM = 2;
static constexpr ck::index_t NumDimN = 2;
static constexpr ck::index_t NumDimK = 2;
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CDEElementOp = ck::tensor_operation::element_wise::Bilinear;
using DeviceOpInstanceKKNN = DeviceOpInstanceKK_FP64<NumDimM,
NumDimN,
NumDimK,
ADataType,
BDataType,
AccDataType,
CShuffleDataType,
DsDataType,
EDataType,
ComputeDataType,
AElementOp,
BElementOp,
CDEElementOp>;
using DeviceOpInstanceKNNN = DeviceOpInstanceKN_FP64<NumDimM,
NumDimN,
NumDimK,
ADataType,
BDataType,
AccDataType,
CShuffleDataType,
DsDataType,
EDataType,
ComputeDataType,
AElementOp,
BElementOp,
CDEElementOp>;
using DeviceOpInstanceMKNN = DeviceOpInstanceMK_FP64<NumDimM,
NumDimN,
NumDimK,
ADataType,
BDataType,
AccDataType,
CShuffleDataType,
DsDataType,
EDataType,
ComputeDataType,
AElementOp,
BElementOp,
CDEElementOp>;
using DeviceOpInstanceMNNN = DeviceOpInstanceMN_FP64<NumDimM,
NumDimN,
NumDimK,
ADataType,
BDataType,
AccDataType,
CShuffleDataType,
DsDataType,
EDataType,
ComputeDataType,
AElementOp,
BElementOp,
CDEElementOp>;
using DeviceOpInstance = DeviceOpInstanceKKNN;
#include "run_complex_contraction_bilinear_example.inc"
int main(int argc, char* argv[]) { return run_complex_contraction_bilinear_example(argc, argv); }
example/66_complex_contraction_bilinear/run_complex_contraction_bilinear_example.inc 0 → 100755
View file @ 6fcaeada
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <cstdlib>
#include <iostream>
#include <string>
#include <vector>
#include "ck/ck.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/numeric.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_contraction.hpp"
int run_complex_contraction_bilinear_example(int argc, char* argv[])
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
// A[M0, M1, K0, K1]
std::vector<ck::index_t> a_ms_ks_lengths{30, 128, 32, 64};
std::vector<ck::index_t> a_ms_ks_strides{524288, 4096, 128, 1};
// B[N0, N1, K0, K1]
std::vector<ck::index_t> b_ns_ks_lengths{32, 64, 32, 64};
std::vector<ck::index_t> b_ns_ks_strides{524288, 4096, 128, 1};
// D[M0, M1, N0, N1]
std::vector<ck::index_t> d_ms_ns_lengths{30, 128, 32, 64};
std::vector<ck::index_t> d_ms_ns_strides{524288, 4096, 128, 1};
// E[M0, M1, N0, N1]
std::vector<ck::index_t> e_ms_ns_lengths{30, 128, 32, 64};
std::vector<ck::index_t> e_ms_ns_strides{524288, 4096, 128, 1};
float alpha = 1.f;
float beta = 1.f;
if(argc == 1)
{
// use default case
}
else if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else if(argc == 28)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
const ck::index_t M0 = std::stoi(argv[4]);
const ck::index_t M1 = std::stoi(argv[5]);
const ck::index_t N0 = std::stoi(argv[6]);
const ck::index_t N1 = std::stoi(argv[7]);
const ck::index_t K0 = std::stoi(argv[8]);
const ck::index_t K1 = std::stoi(argv[9]);
a_ms_ks_lengths = {M0, M1, K0, K1};
a_ms_ks_strides = {
std::stoi(argv[10]), std::stoi(argv[11]), std::stoi(argv[12]), std::stoi(argv[13])};
b_ns_ks_lengths = {N0, N1, K0, K1};
b_ns_ks_strides = {
std::stoi(argv[14]), std::stoi(argv[15]), std::stoi(argv[16]), std::stoi(argv[17])};
d_ms_ns_lengths = {M0, M1, N0, N1};
d_ms_ns_strides = {
std::stoi(argv[18]), std::stoi(argv[19]), std::stoi(argv[20]), std::stoi(argv[21])};
e_ms_ns_lengths = {M0, M1, N0, N1};
e_ms_ns_strides = {
std::stoi(argv[22]), std::stoi(argv[23]), std::stoi(argv[24]), std::stoi(argv[25])};
alpha = std::stof(argv[26]);
beta = std::stof(argv[27]);
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: time kernel (0=no, 1=yes)\n");
printf("arg4 to 9: M0, M1, N0, N1, K0, K1\n");
printf("arg10 to 13: Stride_A_M0, Stride_A_M1, Stride_A_K0, Stride_A_K1\n");
printf("arg14 to 17: Stride_B_N0, Stride_B_N1, Stride_B_K0, Stride_B_K1\n");
printf("arg18 to 21: Stride_D_M0, Stride_D_M1, Stride_D_N0, Stride_D_N1\n");
printf("arg22 to 25: Stride_E_M0, Stride_E_M1, Stride_E_N0, Stride_E_N1\n");
printf("arg26 to 27: alpha, beta\n");
exit(0);
}
// For Real Part of Complex Tensor
Tensor<ADataType> a_ms_ks_re(a_ms_ks_lengths, a_ms_ks_strides);
Tensor<BDataType> b_ns_ks_re(b_ns_ks_lengths, b_ns_ks_strides);
Tensor<EDataType> d_ms_ns_re(d_ms_ns_lengths, d_ms_ns_strides);
Tensor<EDataType> e_ms_ns_host_result_re(e_ms_ns_lengths, e_ms_ns_strides);
Tensor<EDataType> e_ms_ns_device_result_re(e_ms_ns_lengths, e_ms_ns_strides);
// For Imaginary Part of Complex Tensor
Tensor<ADataType> a_ms_ks_img(a_ms_ks_lengths, a_ms_ks_strides);
Tensor<BDataType> b_ns_ks_img(b_ns_ks_lengths, b_ns_ks_strides);
Tensor<EDataType> d_ms_ns_img(d_ms_ns_lengths, d_ms_ns_strides);
Tensor<EDataType> e_ms_ns_host_result_img(e_ms_ns_lengths, e_ms_ns_strides);
Tensor<EDataType> e_ms_ns_device_result_img(e_ms_ns_lengths, e_ms_ns_strides);
// Intermediate E tensor Definition
Tensor<EDataType> e_ms_ns_device_result_re1(e_ms_ns_lengths, e_ms_ns_strides);
Tensor<EDataType> e_ms_ns_device_result_img1(e_ms_ns_lengths, e_ms_ns_strides);
std::cout << "a_ms_ks_re: " << a_ms_ks_re.mDesc << std::endl;
std::cout << "b_ns_ks_re: " << b_ns_ks_re.mDesc << std::endl;
std::cout << "d_ms_ns_re: " << d_ms_ns_re.mDesc << std::endl;
std::cout << "e_ms_ns_re: " << e_ms_ns_host_result_re.mDesc << std::endl;
std::cout << "a_ms_ks_img: " << a_ms_ks_img.mDesc << std::endl;
std::cout << "b_ns_ks_img: " << b_ns_ks_img.mDesc << std::endl;
std::cout << "d_ms_ns_img: " << d_ms_ns_img.mDesc << std::endl;
std::cout << "e_ms_ns_img: " << e_ms_ns_host_result_img.mDesc << std::endl;
switch(init_method)
{
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_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});
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;
}
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 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());
// 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());
a_device_buf_re.ToDevice(a_ms_ks_re.mData.data());
b_device_buf_re.ToDevice(b_ns_ks_re.mData.data());
d_device_buf_re.ToDevice(d_ms_ns_re.mData.data());
a_device_buf_img.ToDevice(a_ms_ks_img.mData.data());
b_device_buf_img.ToDevice(b_ns_ks_img.mData.data());
d_device_buf_img.ToDevice(d_ms_ns_img.mData.data());
// set zero
e_device_buf_re.SetZero();
e_device_buf_img.SetZero();
// 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};
// 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);
if(!op.IsSupportedArgument(argument_re1))
{
std::cout << op.GetTypeString() << " does not support this problem" << std::endl;
return 0;
}
float ave_time_re1 = invoker.Run(argument_re1, StreamConfig{nullptr, time_kernel});
alpha = -1.f;
beta = 1.f;
a_element_op = AElementOp{};
b_element_op = BElementOp{};
cde_element_op = CDEElementOp{alpha, beta};
// device operation
// 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);
if(!op.IsSupportedArgument(argument_re2))
{
std::cout << op.GetTypeString() << " does not support this problem" << std::endl;
return 0;
}
float ave_time_re2 = invoker.Run(argument_re2, StreamConfig{nullptr, time_kernel});
alpha = 1.f;
beta = 1.f;
a_element_op = AElementOp{};
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);
if(!op.IsSupportedArgument(argument_img1))
{
std::cout << op.GetTypeString() << " does not support this problem" << std::endl;
return 0;
}
float ave_time_img1 = invoker.Run(argument_img1, StreamConfig{nullptr, time_kernel});
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);
if(!op.IsSupportedArgument(argument_img2))
{
std::cout << op.GetTypeString() << " does not support this problem" << std::endl;
return 0;
}
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<>{});
ck::index_t N = ck::accumulate_n<ck::index_t>(
e_ms_ns_lengths.begin() + NumDimM, NumDimN, 1, std::multiplies<>{});
ck::index_t K = ck::accumulate_n<ck::index_t>(
a_ms_ks_lengths.begin() + NumDimM, NumDimK, 1, std::multiplies<>{});
std::size_t flop = std::size_t(2) * M * N * K * 2;
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 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, "
<< op.GetTypeString() << std::endl;
e_device_buf_re.FromDevice(e_ms_ns_device_result_re.mData.data());
e_device_buf_img.FromDevice(e_ms_ns_device_result_img.mData.data());
auto isRealOk = 0;
auto isImgOk = 0;
if(do_verification)
{
// Real Part Verification
Tensor<CShuffleDataType> c_ms_ns_host_result_re(e_ms_ns_lengths, e_ms_ns_strides);
Tensor<CShuffleDataType> c_ms_ns_host_result_re1(e_ms_ns_lengths, e_ms_ns_strides);
using ReferenceOpInstance =
ck::tensor_operation::host::ReferenceContraction_M2_N2_K2<NumDimM,
NumDimN,
NumDimK,
ADataType,
BDataType,
CShuffleDataType,
AccDataType,
F32,
AElementOp,
BElementOp>;
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);
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)
{
for(size_t n0 = 0; n0 < e_ms_ns_host_result_re.mDesc.GetLengths()[2]; ++n0)
{
for(size_t n1 = 0; n1 < e_ms_ns_host_result_re.mDesc.GetLengths()[3]; ++n1)
{
cde_element_op(e_ms_ns_host_result_re(m0, m1, n0, n1),
c_ms_ns_host_result_re(m0, m1, n0, n1),
d_ms_ns_re(m0, m1, n0, n1));
}
}
}
}
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);
ref_invoker.Run(ref_argument_re1);
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)
{
for(size_t n0 = 0; n0 < e_ms_ns_host_result_re.mDesc.GetLengths()[2]; ++n0)
{
for(size_t n1 = 0; n1 < e_ms_ns_host_result_re.mDesc.GetLengths()[3]; ++n1)
{
cde_element_op(e_ms_ns_host_result_re(m0, m1, n0, n1),
e_ms_ns_host_result_re(m0, m1, n0, n1),
c_ms_ns_host_result_re1(m0, m1, n0, n1));
}
}
}
}
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);
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)
{
for(size_t m1 = 0; m1 < e_ms_ns_host_result_img.mDesc.GetLengths()[1]; ++m1)
{
for(size_t n0 = 0; n0 < e_ms_ns_host_result_img.mDesc.GetLengths()[2]; ++n0)
{
for(size_t n1 = 0; n1 < e_ms_ns_host_result_img.mDesc.GetLengths()[3]; ++n1)
{
cde_element_op(e_ms_ns_host_result_img(m0, m1, n0, n1),
c_ms_ns_host_result_img(m0, m1, n0, n1),
d_ms_ns_img(m0, m1, n0, n1));
}
}
}
}
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)
{
for(size_t m1 = 0; m1 < e_ms_ns_host_result_img.mDesc.GetLengths()[1]; ++m1)
{
for(size_t n0 = 0; n0 < e_ms_ns_host_result_img.mDesc.GetLengths()[2]; ++n0)
{
for(size_t n1 = 0; n1 < e_ms_ns_host_result_img.mDesc.GetLengths()[3]; ++n1)
{
cde_element_op(e_ms_ns_host_result_img(m0, m1, n0, n1),
e_ms_ns_host_result_img(m0, m1, n0, n1),
c_ms_ns_host_result_img1(m0, m1, n0, n1));
}
}
}
}
isImgOk = ck::utils::check_err(e_ms_ns_device_result_re, e_ms_ns_host_result_re) ? 0 : 1;
return (isRealOk && isImgOk);
}
return 0;
}
example/CMakeLists.txt
View file @ 6fcaeada
...@@ -45,11 +45,7 @@ function(add_example_executable EXAMPLE_NAME FILE_NAME) ...@@ -45,11 +45,7 @@ function(add_example_executable EXAMPLE_NAME FILE_NAME)
endforeach() endforeach()
endif() endif()
if(INSTANCES_ONLY) set(EX_TARGETS ${SUPPORTED_GPU_TARGETS})
set(EX_TARGETS ${DEFAULT_GPU_TARGETS})
else()
set(EX_TARGETS ${GPU_TARGETS})
endif()
#Do not build any DL examples if DL_KERNELS not set #Do not build any DL examples if DL_KERNELS not set
foreach(source IN LISTS FILE_NAME) foreach(source IN LISTS FILE_NAME)
...@@ -147,11 +143,8 @@ function(add_example_executable_no_testing EXAMPLE_NAME FILE_NAME) ...@@ -147,11 +143,8 @@ function(add_example_executable_no_testing EXAMPLE_NAME FILE_NAME)
endforeach() endforeach()
endif() endif()
if(INSTANCES_ONLY) set(EX_TARGETS ${SUPPORTED_GPU_TARGETS})
set(EX_TARGETS ${DEFAULT_GPU_TARGETS})
else()
set(EX_TARGETS ${GPU_TARGETS})
endif()
#Do not build any DL examples if DL_KERNELS not set #Do not build any DL examples if DL_KERNELS not set
foreach(source IN LISTS FILE_NAME) foreach(source IN LISTS FILE_NAME)
if(NOT DEFINED DL_KERNELS AND source MATCHES "_dl") if(NOT DEFINED DL_KERNELS AND source MATCHES "_dl")
......
example/ck_tile/01_fmha/README.md
View file @ 6fcaeada
...@@ -6,7 +6,8 @@ This folder contains example for fmha(fused multi-head attention) using ck_tile ...@@ -6,7 +6,8 @@ This folder contains example for fmha(fused multi-head attention) using ck_tile
``` ```
# in the root of ck_tile # in the root of ck_tile
mkdir build && cd build mkdir build && cd build
sh ../script/cmake-ck-dev.sh ../ <arch> # you can replace this <arch> to gfx90a, gfx942... # you can replace <arch> with the appropriate architecture (for example gfx90a or gfx942) or leave it blank
sh ../script/cmake-ck-dev.sh ../ <arch>
make tile_example_fmha_fwd -j make tile_example_fmha_fwd -j
``` ```
This will result in an executable `build/bin/tile_example_fmha_fwd` This will result in an executable `build/bin/tile_example_fmha_fwd`
...@@ -23,7 +24,7 @@ There are 3 template parameters for this kernel template. ...@@ -23,7 +24,7 @@ There are 3 template parameters for this kernel template.
To speed up compile time, we instantiate the kernels into separate file. In this way we can benefit from parallel building from CMake/Make system. This is achieved by `generate.py` script. Besides, you can look into this script to learn how to instantiate a kernel instance step by step, which is described in `FMHA_FWD_KERNEL_BODY` variable. To speed up compile time, we instantiate the kernels into separate file. In this way we can benefit from parallel building from CMake/Make system. This is achieved by `generate.py` script. Besides, you can look into this script to learn how to instantiate a kernel instance step by step, which is described in `FMHA_FWD_KERNEL_BODY` variable.
## executable ## executable
`tile_example_fmha_fwd` is the example executable, implemented in `fmha_fwd.cpp`. You can type `./bin/tile_example_fmha_fwd -?` to list all supported args. Below is an example of the output (may subject to change) `tile_example_fmha_fwd` is the example executable, implemented in `fmha_fwd.cpp`. You can type `./bin/tile_example_fmha_fwd -?` to list all the arguments. Below is an example of the output (may subject to change)
``` ```
args: args:
-v weather do CPU validation or not (default:1) -v weather do CPU validation or not (default:1)
...@@ -31,47 +32,52 @@ args: ...@@ -31,47 +32,52 @@ args:
-b batch size (default:2) -b batch size (default:2)
-h num of head, for q (default:8) -h num of head, for q (default:8)
-h_k num of head, for k/v, -1 means equal to h (default:-1) -h_k num of head, for k/v, -1 means equal to h (default:-1)
if not equal to h, then this is GQA/MQA case if not equal to h, then this is GQA/MQA case
-s seqlen_q. if group-mode, means the average value of seqlen_q (default:3328) -s seqlen_q. if group-mode, means the average value of seqlen_q (default:3328)
total_seqlen_q = seqlen_q * batch, and seqlen_q per batch may vary total_seqlen_q = seqlen_q * batch, and seqlen_q per batch may vary
also with "-s=s0,s1,s2..." comma seperated int to set per batch seqlen(group-mode) also with "-s=s0,s1,s2..." comma seperated int to set per batch seqlen(group-mode)
-s_k seqlen_k, -1 means equal to s (default:-1) -s_k seqlen_k (including new key/value), -1 means equal to s (default:-1)
-d head dim for q, k (default:128) -d head dim for q, k (default:128)
-d_v head dim for v, -1 means equal to d (default:-1) -d_v head dim for v, -1 means equal to d (default:-1)
-scale_s scale factor of S. 0 means equal to 1/sqrt(hdim). (default:0) -scale_s scale factor of S. 0 means equal to 1/sqrt(hdim). (default:0)
note when squant=1, this value will be modified by range_q/k note when squant=1, this value will be modified by range_q/k
-range_q per-tensor quantization range of q. used if squant=1. (default:16) -range_q per-tensor quantization range of q. used if squant=1. (default:16)
-range_k per-tensor quantization range of k. used if squant=1. (default:16) -range_k per-tensor quantization range of k. used if squant=1. (default:16)
-range_v per-tensor quantization range of v. used if squant=1. (default:16) -range_v per-tensor quantization range of v. used if squant=1. (default:16)
-range_p per-tensor quantization range of p [e^(s-m)]. used if squant=1. (default:1) -range_p per-tensor quantization range of p [e^(s-m)]. used if squant=1. (default:1)
-range_o per-tensor quantization range of o (p*v). used if squant=1. (default:16) -range_o per-tensor quantization range of o (p*v). used if squant=1. (default:16)
-squant if using static quantization fusion or not. auto: fp8 will default use squant, other will not (default:auto) -squant if using static quantization fusion or not. auto: fp8 will default use squant, other will not (default:auto)
0: no static quant(not implemented) 1: apply scale_p and scale_o with respect to P and O. 0: no static quant(not implemented) 1: apply scale_p and scale_o with respect to P and O.
calculate scale_s, scale_p, scale_o according to range_q, range_k, range_v, range_p, range_o calculate scale_s, scale_p, scale_o according to range_q, range_k, range_v, range_p, range_o
-iperm permute input (default:1) -iperm permute input (default:1)
if true, will be b*h*s*d, else b*s*h*d if true, will be b*h*s*d, else b*s*h*d
-operm permute output (default:1) -operm permute output (default:1)
-bias n or 0, no bias (default:n) -bias n or 0, no bias (default:n)
e(lementwise) or 1, elementwise bias with 1*1*s*s. e:1, 1*h*s*s. e:2, b*h*s*s e(lementwise) or 1, elementwise bias with 1*1*s*s. e:1, 1*h*s*s. e:2, b*h*s*s
a(libi) or 2, alibi with 1*h. a:1, b*h a(libi) or 2, alibi with 1*h. a:1, b*h
-prec data type. fp16/bf16/fp8/bf8 (default:fp16) -prec data type. fp16/bf16/fp8/bf8 (default:fp16)
-mask 0: no mask, 1: top-left(same as 't'), 2:bottom-right(same as 'b') (default:0) -mask 0: no mask, 1: top-left(same as 't'), 2:bottom-right(same as 'b') (default:0)
't', top-left causal mask, 'b', bottom-r causal mask 't', top-left causal mask, 'b', bottom-r causal mask
't:l,r', top-left sliding window attn(swa) with FA style left right size 't:l,r', top-left sliding window attn(swa) with FA style left right size
'b:l,r', bottom-r sliding window attn(swa) with FA style left right size 'b:l,r', bottom-r sliding window attn(swa) with FA style left right size
'xt:window_size', xformer style masking from top-left, window_size negative is causal, positive is swa 'xt:window_size', xformer style masking from top-left, window_size negative is causal, positive is swa
'xb:window_size', xformer style masking from bottom-r, window_size negative is causal, positive is swa 'xb:window_size', xformer style masking from bottom-r, window_size negative is causal, positive is swa
'g:y,x', generic attention mask coordinate with y/x size (only debug purpose for now) 'g:y,x', generic attention mask coordinate with y/x size (only debug purpose for now)
-vlayout r for row-major(seqlen*hdim), c for col-major(hdim*seqlen) (default:r) -vlayout r for row-major(seqlen*hdim), c for col-major(hdim*seqlen) (default:r)
-lse 0 not store lse, 1 store lse (default:0) -lse 0 not store lse, 1 store lse (default:0)
-kname if set to 1 will print kernel name (default:0) -kname if set to 1 will print kernel name (default:0)
-init init method. ui, uniform random int, ni, normalized random int (default:uf) -init init method. ui, uniform random int, ni, normalized random int (default:uf)
uf, uniform random float, nf, normalized random float, tf, trig float, uf:q, quantization uf, uniform random float, nf, normalized random float, tf, trig float, uf:q, quantization
-seed random seed used for initializing input tensors. 0 for non-deterministic seed (default:11939) -seed random seed used for initializing input tensors. 0 for non-deterministic seed (default:11939)
-drop_seed seed for random number generator (default:1)
-drop_offset offset for random number generator (default:0)
-drop_prefs seed and offset values are present on GPU; 0 - host, 1 - device/GPU (default:0)
-warmup number of iterations before benchmark the kernel (default:5) -warmup number of iterations before benchmark the kernel (default:5)
-repeat number of iterations to benchmark the kernel (default:20) -repeat number of iterations to benchmark the kernel (default:20)
``` ```
Example: `./bin/tile_example_fmha_fwd -b=1 -h=16 -s=16384 -d=128` will run a fmha case with batch=1, nhead=16, sequence length=16384, hdim=128, fp16 case. Example 1: `./bin/tile_example_fmha_fwd -b=1 -h=16 -s=16384 -d=128` will run a fmha case with batch=1, nhead=16, sequence length=16384, hdim=128, fp16 case.
Example 2: `./bin/tile_example_fmha_fwd -b=1 -h=8 -s=16384 -d=64 -drop_prefs=1 -drop_seed=10 -drop_offset=1234` will run a fmha case with
batch=1, nhead=8, sequence length=16384, hdim=64, drop_seed=0 (in GPU memory), drop_offset=1234 (in GPU memory) fp16 case
## support features ## support features
Currently we are still in rapid development stage, so more features/optimizations will be coming soon. Currently we are still in rapid development stage, so more features/optimizations will be coming soon.
......
example/ck_tile/01_fmha/codegen/ops/fmha_fwd_splitkv.py
View file @ 6fcaeada
...@@ -600,8 +600,8 @@ def get_fwd_splitkv_blobs(kernel_filter : Optional[str], receipt, mask_impl) -> ...@@ -600,8 +600,8 @@ def get_fwd_splitkv_blobs(kernel_filter : Optional[str], receipt, mask_impl) ->
# TODO: use async pipeline when compiler is more stable # TODO: use async pipeline when compiler is more stable
if hdim == 256 or hdim in [32, 64, 128]: if hdim == 256 or hdim in [32, 64, 128]:
# if True: # if True:
pipelines.append(Pipeline('qr', 'row', 'f', 'f', 'f', 'f', bias, lse, squant, pagedkv, mask)) pipelines.append(Pipeline('qr', 'row', 'f', 't', 'f', 'f', bias, lse, squant, pagedkv, mask))
pipelines.append(Pipeline('qr', 'col', 'f', 'f', 'f', 'f', bias, lse, squant, pagedkv, mask)) pipelines.append(Pipeline('qr', 'col', 'f', 't', 'f', 'f', bias, lse, squant, pagedkv, mask))
pipelines.append(Pipeline('qr', 'row', 't', 't', 't', 't', bias, lse, squant, pagedkv, mask)) pipelines.append(Pipeline('qr', 'row', 't', 't', 't', 't', bias, lse, squant, pagedkv, mask))
pipelines.append(Pipeline('qr', 'col', 't', 't', 't', 't', bias, lse, squant, pagedkv, mask)) pipelines.append(Pipeline('qr', 'col', 't', 't', 't', 't', bias, lse, squant, pagedkv, mask))
......
example/ck_tile/01_fmha/fmha_bwd.cpp
View file @ 6fcaeada
...@@ -85,6 +85,9 @@ auto create_args(int argc, char* argv[]) ...@@ -85,6 +85,9 @@ auto create_args(int argc, char* argv[])
.insert("p_drop", "0", "0~1 probability of dropout") .insert("p_drop", "0", "0~1 probability of dropout")
.insert("drop_seed", "1", "seed for random number generator") .insert("drop_seed", "1", "seed for random number generator")
.insert("drop_offset", "0", "offset for random number generator") .insert("drop_offset", "0", "offset for random number generator")
.insert("drop_prefs",
"0",
"seed and offset values are present on GPU; 0 - host, 1 - device/GPU")
.insert("timer", "gpu", "gpu:gpu timer, cpu:cpu timer") .insert("timer", "gpu", "gpu:gpu timer, cpu:cpu timer")
.insert("warmup", "5", "number of iterations before benchmark the kernel") .insert("warmup", "5", "number of iterations before benchmark the kernel")
.insert("repeat", "20", "number of iterations to benchmark the kernel") .insert("repeat", "20", "number of iterations to benchmark the kernel")
...@@ -99,10 +102,23 @@ auto create_args(int argc, char* argv[]) ...@@ -99,10 +102,23 @@ auto create_args(int argc, char* argv[])
// different threshold for different dtype // different threshold for different dtype
template <typename DataType> template <typename DataType>
auto get_elimit(int /*init_method*/) auto get_elimit(ck_tile::index_t /*hdim_q*/, ck_tile::index_t /*hdim_v*/)
{
double rtol = 1e-2;
double atol = 1e-2;
return ck_tile::make_tuple(rtol, atol);
}
template <>
auto get_elimit<ck_tile::bf16_t>(ck_tile::index_t hdim_q, ck_tile::index_t hdim_v)
{ {
double rtol = 1e-2; double rtol = 1e-2;
double atol = 1e-2; double atol = 1e-2;
if(hdim_q > 128 && hdim_v > 128) // 3.2 for RTZ/1.5 for RTN
{
rtol = 3.2e-2;
atol = 3.2e-2;
}
return ck_tile::make_tuple(rtol, atol); return ck_tile::make_tuple(rtol, atol);
} }
...@@ -145,6 +161,8 @@ bool run(const ck_tile::ArgParser& arg_parser) ...@@ -145,6 +161,8 @@ bool run(const ck_tile::ArgParser& arg_parser)
float p_drop = arg_parser.get_float("p_drop"); float p_drop = arg_parser.get_float("p_drop");
uint64_t drop_seed = arg_parser.get_uint64("drop_seed"); uint64_t drop_seed = arg_parser.get_uint64("drop_seed");
uint64_t drop_offset = arg_parser.get_uint64("drop_offset"); uint64_t drop_offset = arg_parser.get_uint64("drop_offset");
bool drop_prefs = arg_parser.get_bool("drop_prefs");
if(use_dbias && bias.type != bias_enum::elementwise_bias) if(use_dbias && bias.type != bias_enum::elementwise_bias)
{ {
std::cerr << "dbias only exists when bias type is elementwise" << std::endl; std::cerr << "dbias only exists when bias type is elementwise" << std::endl;
...@@ -368,6 +386,8 @@ bool run(const ck_tile::ArgParser& arg_parser) ...@@ -368,6 +386,8 @@ bool run(const ck_tile::ArgParser& arg_parser)
ck_tile::DeviceMem dbias_buf(dbias_host.get_element_space_size_in_bytes()); ck_tile::DeviceMem dbias_buf(dbias_host.get_element_space_size_in_bytes());
ck_tile::DeviceMem seqstart_q(seqstart_q_host.size() * sizeof(int32_t)); ck_tile::DeviceMem seqstart_q(seqstart_q_host.size() * sizeof(int32_t));
ck_tile::DeviceMem seqstart_k(seqstart_k_host.size() * sizeof(int32_t)); ck_tile::DeviceMem seqstart_k(seqstart_k_host.size() * sizeof(int32_t));
ck_tile::DeviceMem drop_seed_buf(drop_prefs ? sizeof(uint64_t) : 0);
ck_tile::DeviceMem drop_offset_buf(drop_prefs ? sizeof(uint64_t) : 0);
ck_tile::DeviceMem alibi_slope_buf(alibi_slope_host.get_element_space_size_in_bytes()); ck_tile::DeviceMem alibi_slope_buf(alibi_slope_host.get_element_space_size_in_bytes());
ck_tile::DeviceMem dq_acc_buf(dq_acc_host.get_element_space_size_in_bytes()); ck_tile::DeviceMem dq_acc_buf(dq_acc_host.get_element_space_size_in_bytes());
...@@ -378,6 +398,8 @@ bool run(const ck_tile::ArgParser& arg_parser) ...@@ -378,6 +398,8 @@ bool run(const ck_tile::ArgParser& arg_parser)
do_buf.ToDevice(do_host.data()); do_buf.ToDevice(do_host.data());
seqstart_q.ToDevice(seqstart_q_host.data()); seqstart_q.ToDevice(seqstart_q_host.data());
seqstart_k.ToDevice(seqstart_k_host.data()); seqstart_k.ToDevice(seqstart_k_host.data());
drop_seed_buf.ToDevice(drop_prefs ? &drop_seed : nullptr);
drop_offset_buf.ToDevice(drop_prefs ? &drop_offset : nullptr);
alibi_slope_buf.ToDevice(alibi_slope_host.data()); alibi_slope_buf.ToDevice(alibi_slope_host.data());
// clang-format off // clang-format off
...@@ -459,6 +481,18 @@ bool run(const ck_tile::ArgParser& arg_parser) ...@@ -459,6 +481,18 @@ bool run(const ck_tile::ArgParser& arg_parser)
const ck_tile::index_t split_stride_dq_acc = const ck_tile::index_t split_stride_dq_acc =
(shape_batch * nhead * shape_seqlen_q * hdim_q); (shape_batch * nhead * shape_seqlen_q * hdim_q);
const auto drop_seed_offset = [&]() -> decltype(fmha_bwd_args::drop_seed_offset) {
if(drop_prefs)
{
return std::make_pair(drop_seed_buf.GetDeviceBuffer(),
drop_offset_buf.GetDeviceBuffer());
}
else
{
return std::make_pair(drop_seed, drop_offset);
}
}();
return fmha_bwd_args{q_buf.GetDeviceBuffer(), return fmha_bwd_args{q_buf.GetDeviceBuffer(),
k_buf.GetDeviceBuffer(), k_buf.GetDeviceBuffer(),
v_buf.GetDeviceBuffer(), v_buf.GetDeviceBuffer(),
...@@ -532,7 +566,7 @@ bool run(const ck_tile::ArgParser& arg_parser) ...@@ -532,7 +566,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
static_cast<ck_tile::index_t>(mask.type), static_cast<ck_tile::index_t>(mask.type),
p_drop, p_drop,
p_undrop, p_undrop,
{drop_seed, drop_offset}}; drop_seed_offset};
}(); }();
float ave_time = fmha_bwd(fmha_traits, fmha_args, stream_config); float ave_time = fmha_bwd(fmha_traits, fmha_args, stream_config);
...@@ -899,7 +933,7 @@ bool run(const ck_tile::ArgParser& arg_parser) ...@@ -899,7 +933,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
} }
// clang-format on // clang-format on
auto [rtol, atol] = get_elimit<DataType>(init_method); auto [rtol, atol] = get_elimit<DataType>(hdim_q, hdim_v);
bool dq_cur_pass = ck_tile::check_err(dq_host_result, bool dq_cur_pass = ck_tile::check_err(dq_host_result,
dq_host_ref, dq_host_ref,
std::string("Error: QGrad Incorrect results!"), std::string("Error: QGrad Incorrect results!"),
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment