mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-14 02:02:46 +00:00
173 implement device grouped gemm fixed nk for rdna4 (#4299)
## Proposed changes This PR adds an RDNA4 implementation of the device_grouped_gemm_fixed_nk instance library using for WMMA. The implementation is based on the existing DeviceGroupedGemm_Xdl_Fixed_NK design and reuses the same high-level structure, but replaces the XDL kernel with a WMMA-based one. It uses the GridwiseGemm_wmma_cshuffle_v3 kernel. At this stage, the focus is functional correctness and compatibility, not performance tuning. ## Technical Details - Device struct for grouped gemm fixed NK - Example code for the WMMA version - Unit tests for both new wmma implementation and the reference XDL code (previously missing) - Generic ck profiler interface with the purpose of calling unit tests. ## Checklist Please put an into the boxes that apply. You can also fill these out after creating the PR. If you're not sure, please don't hesitate to ask. - [x] I have added tests relevant to the introduced functionality, and the unit tests are passing locally - [x] I have added the test to REGRESSION_TESTS list defined at the top of CMakeLists.txt in tests/CMakeLists.txt, **IF** the test takes more than 30 seconds to run. - [ ] I have added inline documentation which enables the maintainers with understanding the motivation - [ ] I have removed the stale documentation which is no longer relevant after this pull request - [x] (If this change is user-facing) I have added release notes which provide the end users with a brief summary of the improvement from this pull request - [x] I have run on all changed files - [x] Any dependent changes have been merged ## Discussion If this is a relatively large or complex change, feel free to start a discussion by explaining why you chose the solution you did and what alternatives you considered --- 🔁 Imported from [ROCm/composable_kernel#3668](https://github.com/ROCm/composable_kernel/pull/3668) 🧑💻 Originally authored by @bidlekm --------- Co-authored-by: Marton Bidlek <marton.bidlek@streamhpc.com> Co-authored-by: Erwin Terpstra <erwin.terpstra@streamhpc.com> Co-authored-by: bidlekm <bidlekmarton@gmail.com> Co-authored-by: assistant-librarian[bot] <assistant-librarian[bot]@users.noreply.github.com> Co-authored-by: Illia Silin <98187287+illsilin@users.noreply.github.com> Co-authored-by: illsilin_amdeng <Illia.Silin@amd.com>
This commit is contained in:
assistant-librarian[bot]
committed by
GitHub
GitHub
parent
33f83b99ed
commit
9c8d3a39ac
@@ -47,6 +47,10 @@ add_example_dependencies(example_grouped_gemm_wmma example_grouped_gemm_wmma_spl
|
||||
add_example_executable(example_grouped_gemm_multiple_d_wmma_fp16 grouped_gemm_multiple_d_wmma_fp16.cpp)
|
||||
add_example_dependencies(example_grouped_gemm_wmma example_grouped_gemm_multiple_d_wmma_fp16)
|
||||
|
||||
add_example_executable(example_grouped_gemm_wmma_fixed_nk_fp16 grouped_gemm_wmma_fixed_nk_fp16.cpp)
|
||||
add_example_dependencies(example_grouped_gemm_wmma example_grouped_gemm_wmma_fixed_nk_fp16)
|
||||
|
||||
|
||||
list(APPEND gpu_list_tf32 gfx942 gfx950)
|
||||
set(target 0)
|
||||
foreach(gpu IN LISTS GPU_TARGETS)
|
||||
|
||||
351
example/15_grouped_gemm/grouped_gemm_wmma_fixed_nk_fp16.cpp
Normal file
351
example/15_grouped_gemm/grouped_gemm_wmma_fixed_nk_fp16.cpp
Normal file
@@ -0,0 +1,351 @@
|
||||
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
|
||||
// SPDX-License-Identifier: MIT
|
||||
|
||||
#include <iostream>
|
||||
#include <numeric>
|
||||
#include <initializer_list>
|
||||
#include <cstdlib>
|
||||
|
||||
#include "ck/ck.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/impl/device_grouped_gemm_wmma_fixed_nk.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/device_grouped_gemm.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/library/utility/literals.hpp"
|
||||
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
|
||||
|
||||
using ::ck::DeviceMem;
|
||||
using ::ck::hip_check_error;
|
||||
using ::ck::HostTensorDescriptor;
|
||||
using ::ck::Tensor;
|
||||
|
||||
template <ck::index_t... Is>
|
||||
using S = ck::Sequence<Is...>;
|
||||
|
||||
using F16 = ck::half_t;
|
||||
using F32 = float;
|
||||
|
||||
using Row = ck::tensor_layout::gemm::RowMajor;
|
||||
using Col = ck::tensor_layout::gemm::ColumnMajor;
|
||||
|
||||
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
|
||||
|
||||
using ADataType = F16;
|
||||
using BDataType = F16;
|
||||
using AccDataType = F32;
|
||||
using CShuffleDataType = F32;
|
||||
using DsDataType = ck::Tuple<>;
|
||||
using EDataType = F16;
|
||||
|
||||
using ALayout = Row;
|
||||
using BLayout = Col;
|
||||
using DsLayout = ck::Tuple<>;
|
||||
using ELayout = Row;
|
||||
|
||||
using AElementOp = PassThrough;
|
||||
using BElementOp = PassThrough;
|
||||
using CDEElementOp = PassThrough;
|
||||
|
||||
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
|
||||
|
||||
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGroupedGemm_Wmma_Fixed_Nk
|
||||
// clang-format off
|
||||
//######| ALayout| BLayout| DsLayout| ELayout| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
|
||||
//######| | | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Spacialization| Size| Block| Block| Block| | | Wmma| Wmma| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MRepeat| NRepeat| _MBlock_MRepeat| ScalarPerVector|
|
||||
//######| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NRepeat| _NRepeat|
|
||||
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
< ALayout, BLayout, DsLayout, ELayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementOp, BElementOp, CDEElementOp, GemmDefault, 256, 128, 128, 64, 8, 8, 16, 16, 2, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 64, 1, 4>, 8>;
|
||||
// clang-format on
|
||||
|
||||
struct ProblemSize final
|
||||
{
|
||||
std::vector<ck::index_t> Ms;
|
||||
std::vector<ck::index_t> Ns;
|
||||
std::vector<ck::index_t> Ks;
|
||||
|
||||
std::vector<ck::index_t> stride_As;
|
||||
std::vector<ck::index_t> stride_Bs;
|
||||
std::vector<ck::index_t> stride_Cs;
|
||||
|
||||
ck::index_t group_count;
|
||||
};
|
||||
|
||||
struct ExecutionConfig final
|
||||
{
|
||||
bool do_verification = true;
|
||||
int init_method = 1;
|
||||
int k_batch = 1;
|
||||
bool time_kernel = false;
|
||||
};
|
||||
|
||||
bool run_grouped_gemm(const ProblemSize& problem_size, const ExecutionConfig& config)
|
||||
{
|
||||
auto group_count = problem_size.group_count;
|
||||
|
||||
// GEMM shape
|
||||
std::vector<ck::tensor_operation::device::GemmDesc> gemm_descs;
|
||||
std::vector<void*> p_Cs;
|
||||
|
||||
gemm_descs.reserve(group_count);
|
||||
|
||||
int sum_of_m = 0;
|
||||
|
||||
auto f_host_tensor_descriptor =
|
||||
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
|
||||
using namespace ck::literals;
|
||||
|
||||
if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
|
||||
{
|
||||
return HostTensorDescriptor({row, col}, {stride, 1_uz});
|
||||
}
|
||||
else
|
||||
{
|
||||
return HostTensorDescriptor({row, col}, {1_uz, stride});
|
||||
}
|
||||
};
|
||||
|
||||
std::vector<Tensor<ADataType>> a_tensors;
|
||||
std::vector<Tensor<BDataType>> b_tensors;
|
||||
std::vector<Tensor<EDataType>> c_host_tensors;
|
||||
std::vector<Tensor<EDataType>> c_device_tensors;
|
||||
|
||||
a_tensors.reserve(group_count);
|
||||
b_tensors.reserve(group_count);
|
||||
c_host_tensors.reserve(group_count);
|
||||
c_device_tensors.reserve(group_count);
|
||||
|
||||
using DeviceMemPtr = std::unique_ptr<DeviceMem>;
|
||||
|
||||
std::vector<DeviceMemPtr> a_tensors_device, b_tensors_device, c_tensors_device;
|
||||
|
||||
a_tensors_device.reserve(group_count);
|
||||
b_tensors_device.reserve(group_count);
|
||||
c_tensors_device.reserve(group_count);
|
||||
|
||||
std::size_t flop = 0, num_btype = 0;
|
||||
|
||||
for(int i = 0; i < group_count; i++)
|
||||
{
|
||||
sum_of_m += problem_size.Ms[i];
|
||||
a_tensors.push_back(Tensor<ADataType>(f_host_tensor_descriptor(
|
||||
problem_size.Ms[i], problem_size.Ks[i], problem_size.stride_As[i], ALayout{})));
|
||||
b_tensors.push_back(Tensor<BDataType>(f_host_tensor_descriptor(
|
||||
problem_size.Ks[i], problem_size.Ns[i], problem_size.stride_Bs[i], BLayout{})));
|
||||
c_host_tensors.push_back(Tensor<EDataType>(f_host_tensor_descriptor(
|
||||
problem_size.Ms[i], problem_size.Ns[i], problem_size.stride_Cs[i], ELayout{})));
|
||||
c_device_tensors.push_back(Tensor<EDataType>(f_host_tensor_descriptor(
|
||||
problem_size.Ms[i], problem_size.Ns[i], problem_size.stride_Cs[i], ELayout{})));
|
||||
std::cout << "gemm[" << i << "] a_m_k: " << a_tensors[i].mDesc
|
||||
<< " b_k_n: " << b_tensors[i].mDesc << " c_m_n: " << c_device_tensors[i].mDesc
|
||||
<< std::endl;
|
||||
|
||||
flop += std::size_t(2) * problem_size.Ms[i] * problem_size.Ks[i] * problem_size.Ns[i];
|
||||
num_btype += sizeof(ADataType) * a_tensors[i].mDesc.GetElementSize() +
|
||||
sizeof(BDataType) * b_tensors[i].mDesc.GetElementSize() +
|
||||
sizeof(EDataType) * c_device_tensors[i].mDesc.GetElementSize();
|
||||
|
||||
switch(config.init_method)
|
||||
{
|
||||
case 0: break;
|
||||
case 1:
|
||||
a_tensors[i].GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
|
||||
b_tensors[i].GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
|
||||
break;
|
||||
case 2:
|
||||
a_tensors[i].GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
|
||||
b_tensors[i].GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
|
||||
break;
|
||||
default:
|
||||
a_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<ADataType, 0>{});
|
||||
b_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<BDataType, 1>{});
|
||||
}
|
||||
}
|
||||
|
||||
std::cout << "Sum of M: " << sum_of_m << std::endl;
|
||||
|
||||
using GroupedGemmKernelArgument = ck::tensor_operation::device::GroupedGemmKernelArgument<>;
|
||||
|
||||
std::vector<GroupedGemmKernelArgument> grouped_gemm_kernel_args_;
|
||||
grouped_gemm_kernel_args_.reserve(group_count);
|
||||
|
||||
for(int i = 0; i < group_count; i++)
|
||||
{
|
||||
a_tensors_device.emplace_back(std::make_unique<DeviceMem>(
|
||||
sizeof(ADataType) * problem_size.Ms[i] * problem_size.Ks[i]));
|
||||
|
||||
b_tensors_device.emplace_back(std::make_unique<DeviceMem>(
|
||||
sizeof(BDataType) * problem_size.Ns[i] * problem_size.Ks[i]));
|
||||
|
||||
c_tensors_device.emplace_back(std::make_unique<DeviceMem>(
|
||||
sizeof(EDataType) * problem_size.Ms[i] * problem_size.Ns[i]));
|
||||
|
||||
a_tensors_device[i]->ToDevice(a_tensors[i].mData.data(),
|
||||
a_tensors[i].mDesc.GetElementSpaceSize() * sizeof(ADataType));
|
||||
b_tensors_device[i]->ToDevice(b_tensors[i].mData.data(),
|
||||
b_tensors[i].mDesc.GetElementSpaceSize() * sizeof(BDataType));
|
||||
c_tensors_device[i]->SetZero();
|
||||
|
||||
p_Cs.push_back(c_tensors_device[i]->GetDeviceBuffer());
|
||||
|
||||
gemm_descs.push_back({sum_of_m,
|
||||
problem_size.Ns[i],
|
||||
problem_size.Ks[i],
|
||||
1,
|
||||
problem_size.stride_Bs[i],
|
||||
1,
|
||||
{}});
|
||||
|
||||
grouped_gemm_kernel_args_.push_back({a_tensors_device[i]->GetDeviceBuffer(),
|
||||
b_tensors_device[i]->GetDeviceBuffer(),
|
||||
{},
|
||||
c_tensors_device[i]->GetDeviceBuffer(),
|
||||
problem_size.Ms[i],
|
||||
problem_size.Ns[i],
|
||||
problem_size.Ks[i],
|
||||
problem_size.stride_As[i],
|
||||
problem_size.stride_Bs[i],
|
||||
{},
|
||||
problem_size.stride_Cs[i]});
|
||||
}
|
||||
|
||||
auto a_element_op = AElementOp{};
|
||||
auto b_element_op = BElementOp{};
|
||||
auto c_element_op = CDEElementOp{};
|
||||
|
||||
auto gemm = DeviceGemmInstance{};
|
||||
auto invoker = gemm.MakeInvoker();
|
||||
|
||||
std::vector<const void*> p_As = {};
|
||||
std::vector<const void*> p_Bs = {};
|
||||
std::vector<std::array<const void*, 0>> p_Ds = {};
|
||||
|
||||
// do GEMM
|
||||
auto argument = gemm.MakeArgument(
|
||||
p_As, p_Bs, p_Ds, p_Cs, gemm_descs, a_element_op, b_element_op, c_element_op);
|
||||
|
||||
DeviceMem gemm_arg_dev_mem(gemm.GetDeviceKernelArgSize(&argument));
|
||||
DeviceMem gemm_workspace_dev(gemm.GetWorkSpaceSize(&argument));
|
||||
|
||||
gemm.SetWorkSpacePointer(&argument, gemm_workspace_dev.GetDeviceBuffer());
|
||||
|
||||
hip_check_error(hipMemcpy(gemm_arg_dev_mem.GetDeviceBuffer(),
|
||||
grouped_gemm_kernel_args_.data(),
|
||||
gemm.GetDeviceKernelArgSize(&argument),
|
||||
hipMemcpyHostToDevice));
|
||||
|
||||
if(!gemm.IsSupportedArgument(argument))
|
||||
{
|
||||
throw std::runtime_error(
|
||||
"wrong! device_gemm with the specified compilation parameters does "
|
||||
"not support this GEMM problem");
|
||||
}
|
||||
|
||||
gemm.SetDeviceKernelArgs(&argument, gemm_arg_dev_mem.GetDeviceBuffer());
|
||||
gemm.SetKBatch(argument, config.k_batch);
|
||||
|
||||
invoker.Run(argument, StreamConfig{nullptr, false});
|
||||
|
||||
if(config.time_kernel)
|
||||
{
|
||||
float ave_time = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
|
||||
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, " << gemm.GetTypeString() << std::endl;
|
||||
}
|
||||
|
||||
bool pass = true;
|
||||
if(config.do_verification)
|
||||
{
|
||||
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
|
||||
BDataType,
|
||||
EDataType,
|
||||
AccDataType,
|
||||
AElementOp,
|
||||
BElementOp,
|
||||
CDEElementOp>;
|
||||
|
||||
for(std::size_t i = 0; i < gemm_descs.size(); i++)
|
||||
{
|
||||
c_tensors_device[i]->FromDevice(c_device_tensors[i].mData.data(),
|
||||
c_device_tensors[i].mDesc.GetElementSize() *
|
||||
sizeof(EDataType));
|
||||
auto ref_gemm = ReferenceGemmInstance{};
|
||||
auto ref_invoker = ref_gemm.MakeInvoker();
|
||||
|
||||
auto ref_argument = ref_gemm.MakeArgument(a_tensors[i],
|
||||
b_tensors[i],
|
||||
c_host_tensors[i],
|
||||
a_element_op,
|
||||
b_element_op,
|
||||
c_element_op);
|
||||
|
||||
ref_invoker.Run(ref_argument);
|
||||
|
||||
pass &= ck::utils::check_err(c_device_tensors[i], c_host_tensors[i]);
|
||||
}
|
||||
// Copy device tensors back to host
|
||||
for(std::size_t i = 0; i < c_device_tensors.size(); i++)
|
||||
{
|
||||
c_tensors_device[i]->FromDevice(c_device_tensors[i].mData.data(),
|
||||
c_device_tensors[i].mDesc.GetElementSize() *
|
||||
sizeof(EDataType));
|
||||
}
|
||||
}
|
||||
|
||||
std::cout << "Verification: " << (pass ? "SUCCESS" : "FAILURE") << "!" << std::endl;
|
||||
return pass;
|
||||
}
|
||||
|
||||
int main(int argc, char* argv[])
|
||||
{
|
||||
ProblemSize problem_size;
|
||||
ExecutionConfig config;
|
||||
|
||||
problem_size.group_count = 16;
|
||||
|
||||
if(argc == 5)
|
||||
{
|
||||
config.do_verification = std::stoi(argv[1]);
|
||||
config.init_method = std::stoi(argv[2]);
|
||||
config.time_kernel = std::stoi(argv[3]);
|
||||
config.k_batch = std::stoi(argv[4]);
|
||||
}
|
||||
else if(argc == 6)
|
||||
{
|
||||
config.do_verification = std::stoi(argv[1]);
|
||||
config.init_method = std::stoi(argv[2]);
|
||||
config.time_kernel = std::stoi(argv[3]);
|
||||
config.k_batch = std::stoi(argv[4]);
|
||||
problem_size.group_count = std::stoi(argv[5]);
|
||||
}
|
||||
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=n0, 1=yes)\n");
|
||||
printf("arg4: k_batch (> 0)\n");
|
||||
printf("arg5: group count (default=16)");
|
||||
|
||||
exit(0);
|
||||
}
|
||||
|
||||
for(int i = 0; i < problem_size.group_count; i++)
|
||||
{
|
||||
problem_size.Ms.push_back(128 + i * 128);
|
||||
problem_size.Ns.push_back(1024);
|
||||
problem_size.Ks.push_back(1024);
|
||||
|
||||
problem_size.stride_As.push_back(problem_size.Ks[i]);
|
||||
problem_size.stride_Bs.push_back(problem_size.Ks[i]);
|
||||
problem_size.stride_Cs.push_back(problem_size.Ns[i]);
|
||||
}
|
||||
|
||||
return !run_grouped_gemm(problem_size, config);
|
||||
}
|
||||
@@ -291,6 +291,7 @@ bool run_grouped_gemm(const ProblemSize& problem_size, const ExecutionConfig& co
|
||||
}
|
||||
}
|
||||
|
||||
std::cout << "Verification: " << (pass ? "SUCCESS" : "FAILURE") << "!" << std::endl;
|
||||
return pass;
|
||||
}
|
||||
|
||||
@@ -329,7 +330,7 @@ int main(int argc, char* argv[])
|
||||
|
||||
for(int i = 0; i < problem_size.group_count; i++)
|
||||
{
|
||||
problem_size.Ms.push_back(128 + rand() % 128);
|
||||
problem_size.Ms.push_back(128 + i * 128);
|
||||
problem_size.Ns.push_back(1024);
|
||||
problem_size.Ks.push_back(1024);
|
||||
|
||||
|
||||
@@ -272,6 +272,8 @@ bool run_grouped_gemm(const ProblemSize& problem_size, const ExecutionConfig& co
|
||||
ComputeDataType>(c_device_tensors[i], c_host_tensors[i]);
|
||||
#endif
|
||||
}
|
||||
|
||||
std::cout << "Verification: " << (pass ? "SUCCESS" : "FAILURE") << "!" << std::endl;
|
||||
}
|
||||
|
||||
if(config.time_kernel)
|
||||
|
||||
File diff suppressed because it is too large
Load Diff
@@ -775,8 +775,8 @@ struct GridwiseGemmMultipleD_xdl_splitk_cshuffle
|
||||
ds_grid_desc_mblock_mperblock_nblock_nperblock,
|
||||
e_grid_desc_mblock_mperblock_nblock_nperblock,
|
||||
c_thread_buf,
|
||||
block_work_idx[I0],
|
||||
block_work_idx[I1],
|
||||
block_work_idx[I2],
|
||||
p_shared,
|
||||
p_ds_grid,
|
||||
p_e_grid,
|
||||
|
||||
@@ -344,6 +344,8 @@ struct GridwiseGemm_wmma_cshuffle_v3_base
|
||||
|
||||
using ThisThreadBlock = ThisThreadBlock<BlockSize>;
|
||||
|
||||
using GemmSpecialization = ck::tensor_operation::device::GemmSpecialization;
|
||||
|
||||
static constexpr index_t APackedSize = []() {
|
||||
if constexpr(is_same_v<remove_cvref_t<LDSTypeA>, pk_i4_t>)
|
||||
return 2;
|
||||
@@ -627,8 +629,7 @@ struct GridwiseGemm_wmma_cshuffle_v3_base
|
||||
const std::array<index_t, NumATensor>& StrideAs,
|
||||
const index_t AK0)
|
||||
{
|
||||
using GemmSpecialization = tensor_operation::device::GemmSpecialization;
|
||||
constexpr bool padM = GemmSpec == GemmSpecialization::MKPadding ||
|
||||
constexpr bool padM = GemmSpec == GemmSpecialization::MKPadding ||
|
||||
GemmSpec == GemmSpecialization::MNKPadding ||
|
||||
GemmSpec == GemmSpecialization::MPadding ||
|
||||
GemmSpec == GemmSpecialization::MNPadding;
|
||||
@@ -696,8 +697,7 @@ struct GridwiseGemm_wmma_cshuffle_v3_base
|
||||
const std::array<index_t, NumBTensor>& StrideBs,
|
||||
const index_t BK0)
|
||||
{
|
||||
using GemmSpecialization = tensor_operation::device::GemmSpecialization;
|
||||
constexpr bool padN = GemmSpec == GemmSpecialization::NKPadding ||
|
||||
constexpr bool padN = GemmSpec == GemmSpecialization::NKPadding ||
|
||||
GemmSpec == GemmSpecialization::MNKPadding ||
|
||||
GemmSpec == GemmSpecialization::NPadding ||
|
||||
GemmSpec == GemmSpecialization::MNPadding;
|
||||
@@ -794,7 +794,6 @@ struct GridwiseGemm_wmma_cshuffle_v3_base
|
||||
// TODO: Investigate why this path is not used in the original
|
||||
// gridwise_gemm_xdl_cshuffle_v3.hpp
|
||||
#if 0
|
||||
using GemmSpecialization = tensor_operation::device::GemmSpecialization;
|
||||
|
||||
if constexpr(GemmSpec == GemmSpecialization::MNPadding ||
|
||||
GemmSpec == GemmSpecialization::MNKPadding)
|
||||
@@ -1033,6 +1032,49 @@ struct GridwiseGemm_wmma_cshuffle_v3_base
|
||||
}
|
||||
}
|
||||
|
||||
__host__ __device__ static constexpr bool
|
||||
CheckValidity(const index_t M,
|
||||
const index_t N,
|
||||
const index_t K,
|
||||
const index_t StrideA,
|
||||
const index_t StrideB,
|
||||
const std::array<index_t, NumDTensor> StrideDs,
|
||||
const index_t StrideE,
|
||||
const index_t KBatch)
|
||||
{
|
||||
|
||||
ignore = StrideDs;
|
||||
const auto M_padded = CalculateMPadded(M);
|
||||
const auto N_padded = CalculateMPadded(N);
|
||||
const auto K_padded = CalculateKPadded(K, KBatch);
|
||||
|
||||
const auto e_grid_desc_m_n =
|
||||
MakeDEGridDescriptor_M_N<ELayout>(M, M_padded, N, N_padded, StrideE);
|
||||
|
||||
const index_t AK0 = CalculateAK0Padded(K, KBatch);
|
||||
const index_t BK0 = CalculateBK0Padded(K, KBatch);
|
||||
|
||||
const auto a_grid_desc_ak0_m_ak1 = MakeAsGridDescriptor_AK0_M_AK1(
|
||||
M, M_padded, K, K_padded, std::array<index_t, 1>{StrideA}, AK0);
|
||||
|
||||
const auto b_grid_desc_bk0_n_bk1 = MakeBsGridDescriptor_BK0_N_BK1(
|
||||
K, K_padded, N, N_padded, std::array<index_t, 1>{StrideB}, BK0);
|
||||
|
||||
constexpr long_index_t TwoGB = (long_index_t{1} << 31);
|
||||
|
||||
const auto& a_desc = a_grid_desc_ak0_m_ak1.At(I0);
|
||||
const auto& b_desc = b_grid_desc_bk0_n_bk1.At(I0);
|
||||
|
||||
if(!(a_desc.GetElementSpaceSize() * sizeof(LDSTypeA) <= TwoGB &&
|
||||
b_desc.GetElementSpaceSize() * sizeof(LDSTypeB) <= TwoGB &&
|
||||
e_grid_desc_m_n.GetElementSpaceSize() * sizeof(EDataType) <= TwoGB))
|
||||
{
|
||||
return false;
|
||||
}
|
||||
|
||||
return true;
|
||||
}
|
||||
|
||||
// block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
|
||||
template <typename Argument>
|
||||
__host__ static constexpr bool CheckValidity(const Argument& karg,
|
||||
@@ -1089,9 +1131,11 @@ struct GridwiseGemm_wmma_cshuffle_v3_base
|
||||
{
|
||||
if(ck::EnvIsEnabled(CK_ENV(CK_LOGGING)))
|
||||
{
|
||||
std::cout << "Arg K value is not a multiple of K_Batch * K0PerBlock * K1! K: "
|
||||
<< karg.K << " " << __FILE__ << ":" << __LINE__
|
||||
<< ", in function: " << __func__ << std::endl;
|
||||
std::cout << "Arg K value is not a multiple of K_Batch * K0PerBlock * K1! "
|
||||
"K_Batch:"
|
||||
<< karg.KBatch << " " << "K0PerBlock:" << KPerBlock << " "
|
||||
<< "K1:" << AK1Number << " " << "K:" << karg.K << " " << __FILE__
|
||||
<< ":" << __LINE__ << ", in function: " << __func__ << std::endl;
|
||||
}
|
||||
return false;
|
||||
}
|
||||
|
||||
@@ -0,0 +1,268 @@
|
||||
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
|
||||
// SPDX-License-Identifier: MIT
|
||||
|
||||
#pragma once
|
||||
|
||||
#include <cstdlib>
|
||||
|
||||
#include "ck/ck.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/impl/device_grouped_gemm_wmma_fixed_nk.hpp"
|
||||
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
|
||||
#include "ck/utility/scheduler_enum.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
|
||||
using F8 = ck::f8_t;
|
||||
using F16 = ck::half_t;
|
||||
using BF16 = ck::bhalf_t;
|
||||
using F32 = float;
|
||||
using I8 = int8_t;
|
||||
|
||||
using Row = ck::tensor_layout::gemm::RowMajor;
|
||||
using Col = ck::tensor_layout::gemm::ColumnMajor;
|
||||
|
||||
template <ck::index_t... Is>
|
||||
using S = ck::Sequence<Is...>;
|
||||
|
||||
using Empty_Tuple = ck::Tuple<>;
|
||||
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
|
||||
using FastGelu = ck::tensor_operation::element_wise::FastGelu;
|
||||
|
||||
using AccDataType = F32;
|
||||
using DsDataType = Empty_Tuple;
|
||||
|
||||
using DsLayout = Empty_Tuple;
|
||||
using ELayout = Row;
|
||||
|
||||
static constexpr auto PipelineV1 = BlockGemmPipelineVersion::v1;
|
||||
static constexpr auto PipelineV3 = BlockGemmPipelineVersion::v3;
|
||||
static constexpr auto IntrawaveScheduler = BlockGemmPipelineScheduler::Intrawave;
|
||||
static constexpr auto InterwaveScheduler = BlockGemmPipelineScheduler::Interwave;
|
||||
static constexpr auto GemmMNKPadding = device::GemmSpecialization::MNKPadding;
|
||||
static constexpr auto GemmDefault = device::GemmSpecialization::Default;
|
||||
|
||||
// Instances for 2 byte datatypes in RRR layout with ADataType = BDataType = EDataType
|
||||
template <typename T,
|
||||
device::GemmSpecialization GemmSpec,
|
||||
BlockGemmPipelineScheduler BlkGemmPipeSched,
|
||||
BlockGemmPipelineVersion BlkGemmPipelineVer,
|
||||
typename AElementOp,
|
||||
typename BElementOp,
|
||||
typename CDEElementOp,
|
||||
enable_if_t<sizeof(T) == 2, bool> = false>
|
||||
using device_grouped_gemm_wmma_fixed_nk_mk_kn_mn_instances = std::tuple<
|
||||
// clang-format off
|
||||
//#############################| ALayout| BLayout| DsLayout| ELayout| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
|
||||
//#############################| | | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Spacialization| Size| Block| Block| Block| | | Wmma| Wmma| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MRepeat| NRepeat| _MBlock_MRepeat| ScalarPerVector|
|
||||
//#############################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NRepeat| _NRepeat|
|
||||
//#############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGroupedGemm_Wmma_Fixed_Nk< Row, Row, DsLayout, ELayout, T, T, AccDataType, T, DsDataType, T, AElementOp, BElementOp, CDEElementOp, GemmSpec, 256, 128, 128, 64, 8, 8, 16, 16, 2, 4, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 2, 8, 8, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 64, 1, 4>, 8, BlkGemmPipeSched, BlkGemmPipelineVer>,
|
||||
DeviceGroupedGemm_Wmma_Fixed_Nk< Row, Row, DsLayout, ELayout, T, T, AccDataType, T, DsDataType, T, AElementOp, BElementOp, CDEElementOp, GemmSpec, 256, 128, 128, 64, 2, 2, 16, 16, 2, 4, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 2, 2, 2, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 1, 1, 1, S<1, 64, 1, 4>, 8, BlkGemmPipeSched, BlkGemmPipelineVer>,
|
||||
DeviceGroupedGemm_Wmma_Fixed_Nk< Row, Row, DsLayout, ELayout, T, T, AccDataType, T, DsDataType, T, AElementOp, BElementOp, CDEElementOp, GemmSpec, 256, 128, 128, 32, 8, 8, 16, 16, 2, 4, S<4, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 2, 8, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 64, 1, 4>, 8, BlkGemmPipeSched, BlkGemmPipelineVer>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
// Instances for 2 byte datatypes in RCR layout with ADataType = BDataType = EDataType
|
||||
template <typename T,
|
||||
device::GemmSpecialization GemmSpec,
|
||||
BlockGemmPipelineScheduler BlkGemmPipeSched,
|
||||
BlockGemmPipelineVersion BlkGemmPipelineVer,
|
||||
typename AElementOp,
|
||||
typename BElementOp,
|
||||
typename CDEElementOp,
|
||||
enable_if_t<sizeof(T) == 2, bool> = false>
|
||||
using device_grouped_gemm_wmma_fixed_nk_mk_nk_mn_instances = std::tuple<
|
||||
// clang-format off
|
||||
//#############################| ALayout| BLayout| DsLayout| ELayout| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
|
||||
//#############################| | | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Spacialization| Size| Block| Block| Block| | | Wmma| Wmma| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MRepeat| NRepeat| _MBlock_MRepeat| ScalarPerVector|
|
||||
//#############################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NRepeat| _NRepeat|
|
||||
//#############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGroupedGemm_Wmma_Fixed_Nk< Row, Col, DsLayout, ELayout, T, T, AccDataType, T, DsDataType, T, AElementOp, BElementOp, CDEElementOp, GemmSpec, 256, 128, 128, 64, 8, 8, 16, 16, 2, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 64, 1, 4>, 8, BlkGemmPipeSched, BlkGemmPipelineVer>,
|
||||
DeviceGroupedGemm_Wmma_Fixed_Nk< Row, Col, DsLayout, ELayout, T, T, AccDataType, T, DsDataType, T, AElementOp, BElementOp, CDEElementOp, GemmSpec, 256, 128, 128, 64, 2, 2, 16, 16, 2, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 2, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 2, 1, 1, 1, S<1, 64, 1, 4>, 8, BlkGemmPipeSched, BlkGemmPipelineVer>,
|
||||
DeviceGroupedGemm_Wmma_Fixed_Nk< Row, Col, DsLayout, ELayout, T, T, AccDataType, T, DsDataType, T, AElementOp, BElementOp, CDEElementOp, GemmSpec, 256, 128, 128, 32, 8, 8, 16, 16, 2, 4, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 64, 1, 4>, 8, BlkGemmPipeSched, BlkGemmPipelineVer>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
template <typename TA,
|
||||
typename TB,
|
||||
device::GemmSpecialization GemmSpec,
|
||||
BlockGemmPipelineScheduler BlkGemmPipeSched,
|
||||
BlockGemmPipelineVersion BlkGemmPipelineVer,
|
||||
typename AElementOp,
|
||||
typename BElementOp,
|
||||
typename CDEElementOp>
|
||||
using device_grouped_gemm_wmma_fixed_nk_mk_kn_mn_irregular_instances = std::tuple<
|
||||
// clang-format off
|
||||
//##############################| ALayout| BLayout| DsLayout| ELayout| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
|
||||
//##############################| | | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Spacialization| Size| Block| Block| Block| | | Wmma| Wmma| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MRepeat| NRepeat| _MBlock_MRepeat| ScalarPerVector|
|
||||
//##############################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NRepeat| _NRepeat|
|
||||
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGroupedGemm_Wmma_Fixed_Nk< Row, Row, DsLayout, ELayout, TA, TB, AccDataType, AccDataType, DsDataType, TA, AElementOp, BElementOp, CDEElementOp, GemmSpec, 256, 128, 128, 64, 8, 8, 16, 16, 2, 4, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 2, 8, 8, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 64, 1, 4>, 8, BlkGemmPipeSched, BlkGemmPipelineVer>,
|
||||
DeviceGroupedGemm_Wmma_Fixed_Nk< Row, Row, DsLayout, ELayout, TA, TB, AccDataType, AccDataType, DsDataType, TA, AElementOp, BElementOp, CDEElementOp, GemmSpec, 256, 128, 128, 64, 2, 2, 16, 16, 2, 4, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 2, 2, 2, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 1, 1, 1, S<1, 64, 1, 4>, 8, BlkGemmPipeSched, BlkGemmPipelineVer>,
|
||||
DeviceGroupedGemm_Wmma_Fixed_Nk< Row, Row, DsLayout, ELayout, TA, TB, AccDataType, AccDataType, DsDataType, TA, AElementOp, BElementOp, CDEElementOp, GemmSpec, 256, 128, 128, 32, 8, 8, 16, 16, 2, 4, S<4, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 2, 8, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 64, 1, 4>, 8, BlkGemmPipeSched, BlkGemmPipelineVer>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
template <typename TA,
|
||||
typename TB,
|
||||
device::GemmSpecialization GemmSpec,
|
||||
BlockGemmPipelineScheduler BlkGemmPipeSched,
|
||||
BlockGemmPipelineVersion BlkGemmPipelineVer,
|
||||
typename AElementOp,
|
||||
typename BElementOp,
|
||||
typename CDEElementOp>
|
||||
using device_grouped_gemm_wmma_fixed_nk_mk_nk_mn_irregular_instances = std::tuple<
|
||||
// clang-format off
|
||||
//##############################| ALayout| BLayout| DsLayout| ELayout| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
|
||||
//##############################| | | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Spacialization| Size| Block| Block| Block| | | Wmma| Wmma| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MRepeat| NRepeat| _MBlock_MRepeat| ScalarPerVector|
|
||||
//##############################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NRepeat| _NRepeat|
|
||||
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGroupedGemm_Wmma_Fixed_Nk< Row, Col, DsLayout, ELayout, TA, TB, AccDataType, AccDataType, DsDataType, TA, AElementOp, BElementOp, CDEElementOp, GemmSpec, 256, 128, 128, 64, 8, 8, 16, 16, 2, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 64, 1, 4>, 8, BlkGemmPipeSched, BlkGemmPipelineVer>,
|
||||
DeviceGroupedGemm_Wmma_Fixed_Nk< Row, Col, DsLayout, ELayout, TA, TB, AccDataType, AccDataType, DsDataType, TA, AElementOp, BElementOp, CDEElementOp, GemmSpec, 256, 128, 128, 64, 2, 2, 16, 16, 2, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 2, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 2, 1, 1, 1, S<1, 64, 1, 4>, 8, BlkGemmPipeSched, BlkGemmPipelineVer>,
|
||||
DeviceGroupedGemm_Wmma_Fixed_Nk< Row, Col, DsLayout, ELayout, TA, TB, AccDataType, AccDataType, DsDataType, TA, AElementOp, BElementOp, CDEElementOp, GemmSpec, 256, 128, 128, 32, 8, 8, 16, 16, 2, 4, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 64, 1, 4>, 8, BlkGemmPipeSched, BlkGemmPipelineVer>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
// List of instance variants to add (pipeline/scheduler/padding combinations)
|
||||
// Some are disabled now, can be re-enabled if needed
|
||||
using InstanceVariant =
|
||||
ck::Tuple<device::GemmSpecialization, BlockGemmPipelineScheduler, BlockGemmPipelineVersion>;
|
||||
static constexpr InstanceVariant InstanceVariants[] = {
|
||||
|
||||
make_tuple(GemmDefault, IntrawaveScheduler, PipelineV1),
|
||||
make_tuple(GemmDefault, IntrawaveScheduler, PipelineV3),
|
||||
|
||||
make_tuple(GemmMNKPadding, IntrawaveScheduler, PipelineV1)};
|
||||
|
||||
template <typename ALayout,
|
||||
typename BLayout,
|
||||
template <device::GemmSpecialization GemmSpec,
|
||||
BlockGemmPipelineScheduler BlkGemmPipeSched,
|
||||
BlockGemmPipelineVersion BlkGemmPipelineVer,
|
||||
typename AElementOp,
|
||||
typename BElementOp,
|
||||
typename CDEElementOp>
|
||||
typename LayoutInstances,
|
||||
typename ADataType,
|
||||
typename BDataType,
|
||||
typename EDataType,
|
||||
typename AElementOp,
|
||||
typename BElementOp,
|
||||
typename CDEElementOp>
|
||||
void add_device_grouped_gemm_wmma_fixed_nk_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedGemmFixedNK<ALayout,
|
||||
BLayout,
|
||||
DsLayout,
|
||||
ELayout,
|
||||
ADataType,
|
||||
BDataType,
|
||||
DsDataType,
|
||||
EDataType,
|
||||
AElementOp,
|
||||
BElementOp,
|
||||
CDEElementOp>>>& instances)
|
||||
{
|
||||
static_for<0, std::size(InstanceVariants), 1>{}([&](auto i) {
|
||||
constexpr auto instance = InstanceVariants[i];
|
||||
add_device_operation_instances(instances,
|
||||
LayoutInstances<instance.At(Number<0>{}),
|
||||
instance.At(Number<1>{}),
|
||||
instance.At(Number<2>{}),
|
||||
AElementOp,
|
||||
BElementOp,
|
||||
CDEElementOp>{});
|
||||
});
|
||||
}
|
||||
|
||||
template <typename T,
|
||||
typename ALayout,
|
||||
typename BLayout,
|
||||
template <typename T2,
|
||||
device::GemmSpecialization GemmSpec,
|
||||
BlockGemmPipelineScheduler BlkGemmPipeSched,
|
||||
BlockGemmPipelineVersion BlkGemmPipelineVer,
|
||||
typename AElementOp,
|
||||
typename BElementOp,
|
||||
typename CDEElementOp>
|
||||
typename LayoutInstances,
|
||||
typename AElementOp,
|
||||
typename BElementOp,
|
||||
typename CDEElementOp>
|
||||
void add_device_grouped_gemm_wmma_fixed_nk_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedGemmFixedNK<ALayout,
|
||||
BLayout,
|
||||
DsLayout,
|
||||
ELayout,
|
||||
T,
|
||||
T,
|
||||
DsDataType,
|
||||
T,
|
||||
AElementOp,
|
||||
BElementOp,
|
||||
CDEElementOp>>>& instances)
|
||||
{
|
||||
static_for<0, std::size(InstanceVariants), 1>{}([&](auto i) {
|
||||
constexpr auto instance = InstanceVariants[i];
|
||||
add_device_operation_instances(instances,
|
||||
LayoutInstances<T,
|
||||
instance.At(Number<0>{}),
|
||||
instance.At(Number<1>{}),
|
||||
instance.At(Number<2>{}),
|
||||
AElementOp,
|
||||
BElementOp,
|
||||
CDEElementOp>{});
|
||||
});
|
||||
}
|
||||
|
||||
template <typename TA,
|
||||
typename TB,
|
||||
typename ALayout,
|
||||
typename BLayout,
|
||||
template <typename T2,
|
||||
typename T3,
|
||||
device::GemmSpecialization GemmSpec,
|
||||
BlockGemmPipelineScheduler BlkGemmPipeSched,
|
||||
BlockGemmPipelineVersion BlkGemmPipelineVer,
|
||||
typename AElementOp,
|
||||
typename BElementOp,
|
||||
typename CDEElementOp>
|
||||
typename LayoutInstances,
|
||||
typename AElementOp,
|
||||
typename BElementOp,
|
||||
typename CDEElementOp>
|
||||
void add_device_grouped_gemm_wmma_fixed_nk_irregular_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedGemmFixedNK<ALayout,
|
||||
BLayout,
|
||||
DsLayout,
|
||||
ELayout,
|
||||
TA,
|
||||
TB,
|
||||
DsDataType,
|
||||
TA,
|
||||
AElementOp,
|
||||
BElementOp,
|
||||
CDEElementOp>>>& instances)
|
||||
{
|
||||
static_for<0, std::size(InstanceVariants), 1>{}([&](auto i) {
|
||||
constexpr auto instance = InstanceVariants[i];
|
||||
add_device_operation_instances(instances,
|
||||
LayoutInstances<TA,
|
||||
TB,
|
||||
instance.At(Number<0>{}),
|
||||
instance.At(Number<1>{}),
|
||||
instance.At(Number<2>{}),
|
||||
AElementOp,
|
||||
BElementOp,
|
||||
CDEElementOp>{});
|
||||
});
|
||||
}
|
||||
|
||||
} // namespace instance
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
@@ -17,6 +17,7 @@ namespace device {
|
||||
namespace instance {
|
||||
|
||||
// fp16_output
|
||||
#if defined(CK_USE_XDL)
|
||||
void add_device_grouped_gemm_xdl_fixed_nk_f16_f16_f16_mk_kn_mn_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedGemmFixedNK<Row,
|
||||
Row,
|
||||
@@ -96,9 +97,91 @@ void add_device_grouped_gemm_xdl_fixed_nk_f16_i8_f16_mk_nk_mn_instances(
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
PassThrough>>>& instances);
|
||||
#endif
|
||||
|
||||
#if defined(CK_USE_WMMA)
|
||||
void add_device_grouped_gemm_wmma_fixed_nk_f16_f16_f16_mk_kn_mn_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedGemmFixedNK<Row,
|
||||
Row,
|
||||
Empty_Tuple,
|
||||
Row,
|
||||
F16,
|
||||
F16,
|
||||
Empty_Tuple,
|
||||
F16,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
PassThrough>>>& instances);
|
||||
|
||||
void add_device_grouped_gemm_wmma_fixed_nk_f16_f16_f16_mk_nk_mn_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedGemmFixedNK<Row,
|
||||
Col,
|
||||
Empty_Tuple,
|
||||
Row,
|
||||
F16,
|
||||
F16,
|
||||
Empty_Tuple,
|
||||
F16,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
PassThrough>>>& instances);
|
||||
|
||||
void add_device_grouped_gemm_wmma_fixed_nk_f16_fp8_f16_mk_kn_mn_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedGemmFixedNK<Row,
|
||||
Row,
|
||||
Empty_Tuple,
|
||||
Row,
|
||||
F16,
|
||||
F8,
|
||||
Empty_Tuple,
|
||||
F16,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
PassThrough>>>& instances);
|
||||
|
||||
void add_device_grouped_gemm_wmma_fixed_nk_f16_fp8_f16_mk_nk_mn_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedGemmFixedNK<Row,
|
||||
Col,
|
||||
Empty_Tuple,
|
||||
Row,
|
||||
F16,
|
||||
F8,
|
||||
Empty_Tuple,
|
||||
F16,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
PassThrough>>>& instances);
|
||||
|
||||
void add_device_grouped_gemm_wmma_fixed_nk_f16_i8_f16_mk_kn_mn_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedGemmFixedNK<Row,
|
||||
Row,
|
||||
Empty_Tuple,
|
||||
Row,
|
||||
F16,
|
||||
I8,
|
||||
Empty_Tuple,
|
||||
F16,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
PassThrough>>>& instances);
|
||||
|
||||
void add_device_grouped_gemm_wmma_fixed_nk_f16_i8_f16_mk_nk_mn_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedGemmFixedNK<Row,
|
||||
Col,
|
||||
Empty_Tuple,
|
||||
Row,
|
||||
F16,
|
||||
I8,
|
||||
Empty_Tuple,
|
||||
F16,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
PassThrough>>>& instances);
|
||||
#endif
|
||||
|
||||
// bf16_inputA i8_inputB
|
||||
#if defined(CK_ENABLE_BF16) && defined(CK_ENABLE_INT8)
|
||||
#if defined(CK_USE_XDL)
|
||||
void add_device_grouped_gemm_xdl_fixed_nk_bf16_i8_bf16_mk_kn_mn_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedGemmFixedNK<Row,
|
||||
Row,
|
||||
@@ -125,9 +208,38 @@ void add_device_grouped_gemm_xdl_fixed_nk_bf16_i8_bf16_mk_nk_mn_instances(
|
||||
PassThrough,
|
||||
PassThrough>>>& instances);
|
||||
#endif
|
||||
#if defined(CK_USE_WMMA)
|
||||
void add_device_grouped_gemm_wmma_fixed_nk_bf16_i8_bf16_mk_nk_mn_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedGemmFixedNK<Row,
|
||||
Col,
|
||||
Empty_Tuple,
|
||||
Row,
|
||||
BF16,
|
||||
I8,
|
||||
Empty_Tuple,
|
||||
BF16,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
PassThrough>>>& instances);
|
||||
|
||||
void add_device_grouped_gemm_wmma_fixed_nk_bf16_i8_bf16_mk_kn_mn_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedGemmFixedNK<Row,
|
||||
Row,
|
||||
Empty_Tuple,
|
||||
Row,
|
||||
BF16,
|
||||
I8,
|
||||
Empty_Tuple,
|
||||
BF16,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
PassThrough>>>& instances);
|
||||
#endif
|
||||
#endif
|
||||
|
||||
// bf16_inputA bf16_inputB
|
||||
#if defined(CK_ENABLE_BF16)
|
||||
#if defined(CK_USE_XDL)
|
||||
void add_device_grouped_gemm_xdl_fixed_nk_bf16_bf16_bf16_mk_kn_mn_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedGemmFixedNK<Row,
|
||||
Row,
|
||||
@@ -153,6 +265,34 @@ void add_device_grouped_gemm_xdl_fixed_nk_bf16_bf16_bf16_mk_nk_mn_instances(
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
PassThrough>>>& instances);
|
||||
#endif
|
||||
#if defined(CK_USE_WMMA)
|
||||
void add_device_grouped_gemm_wmma_fixed_nk_bf16_bf16_bf16_mk_kn_mn_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedGemmFixedNK<Row,
|
||||
Row,
|
||||
Empty_Tuple,
|
||||
Row,
|
||||
BF16,
|
||||
BF16,
|
||||
Empty_Tuple,
|
||||
BF16,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
PassThrough>>>& instances);
|
||||
|
||||
void add_device_grouped_gemm_wmma_fixed_nk_bf16_bf16_bf16_mk_nk_mn_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedGemmFixedNK<Row,
|
||||
Col,
|
||||
Empty_Tuple,
|
||||
Row,
|
||||
BF16,
|
||||
BF16,
|
||||
Empty_Tuple,
|
||||
BF16,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
PassThrough>>>& instances);
|
||||
#endif
|
||||
#endif // CK_ENABLE_BF16
|
||||
|
||||
template <typename ALayout,
|
||||
@@ -190,6 +330,7 @@ struct DeviceOperationInstanceFactory<
|
||||
{
|
||||
std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
|
||||
|
||||
#if defined(CK_USE_XDL)
|
||||
// fp16_output
|
||||
if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
|
||||
is_same_v<EDataType, half_t>)
|
||||
@@ -273,6 +414,93 @@ struct DeviceOperationInstanceFactory<
|
||||
}
|
||||
}
|
||||
#endif // CK_ENABLE_BF16
|
||||
#endif // CK_USE_XDL
|
||||
|
||||
#if defined(CK_USE_WMMA)
|
||||
// fp16_output
|
||||
if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
|
||||
is_same_v<EDataType, half_t>)
|
||||
{
|
||||
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
|
||||
is_same_v<ELayout, Row>)
|
||||
{
|
||||
add_device_grouped_gemm_wmma_fixed_nk_f16_f16_f16_mk_kn_mn_instances(op_ptrs);
|
||||
}
|
||||
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
|
||||
is_same_v<ELayout, Row>)
|
||||
{
|
||||
add_device_grouped_gemm_wmma_fixed_nk_f16_f16_f16_mk_nk_mn_instances(op_ptrs);
|
||||
}
|
||||
}
|
||||
|
||||
// fp8_input
|
||||
if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, f8_t> &&
|
||||
is_same_v<EDataType, half_t>)
|
||||
{
|
||||
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
|
||||
is_same_v<ELayout, Row>)
|
||||
{
|
||||
add_device_grouped_gemm_wmma_fixed_nk_f16_fp8_f16_mk_kn_mn_instances(op_ptrs);
|
||||
}
|
||||
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
|
||||
is_same_v<ELayout, Row>)
|
||||
{
|
||||
add_device_grouped_gemm_wmma_fixed_nk_f16_fp8_f16_mk_nk_mn_instances(op_ptrs);
|
||||
}
|
||||
}
|
||||
|
||||
// i8_input
|
||||
if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, int8_t> &&
|
||||
is_same_v<EDataType, half_t>)
|
||||
{
|
||||
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
|
||||
is_same_v<ELayout, Row>)
|
||||
{
|
||||
add_device_grouped_gemm_wmma_fixed_nk_f16_i8_f16_mk_kn_mn_instances(op_ptrs);
|
||||
}
|
||||
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
|
||||
is_same_v<ELayout, Row>)
|
||||
{
|
||||
add_device_grouped_gemm_wmma_fixed_nk_f16_i8_f16_mk_nk_mn_instances(op_ptrs);
|
||||
}
|
||||
}
|
||||
|
||||
// bf16_i8_input
|
||||
#if defined(CK_ENABLE_BF16) && defined(CK_ENABLE_INT8)
|
||||
if constexpr(is_same_v<ADataType, bhalf_t> && is_same_v<BDataType, int8_t> &&
|
||||
is_same_v<EDataType, bhalf_t>)
|
||||
{
|
||||
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
|
||||
is_same_v<ELayout, Row>)
|
||||
{
|
||||
add_device_grouped_gemm_wmma_fixed_nk_bf16_i8_bf16_mk_kn_mn_instances(op_ptrs);
|
||||
}
|
||||
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
|
||||
is_same_v<ELayout, Row>)
|
||||
{
|
||||
add_device_grouped_gemm_wmma_fixed_nk_bf16_i8_bf16_mk_nk_mn_instances(op_ptrs);
|
||||
}
|
||||
}
|
||||
#endif
|
||||
|
||||
// bf16_inputA bf16_inputB
|
||||
#if defined(CK_ENABLE_BF16)
|
||||
if constexpr(is_same_v<ADataType, bhalf_t> && is_same_v<BDataType, bhalf_t> &&
|
||||
is_same_v<EDataType, bhalf_t>)
|
||||
{
|
||||
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
|
||||
is_same_v<ELayout, Row>)
|
||||
{
|
||||
add_device_grouped_gemm_wmma_fixed_nk_bf16_bf16_bf16_mk_kn_mn_instances(op_ptrs);
|
||||
}
|
||||
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
|
||||
is_same_v<ELayout, Row>)
|
||||
{
|
||||
add_device_grouped_gemm_wmma_fixed_nk_bf16_bf16_bf16_mk_nk_mn_instances(op_ptrs);
|
||||
}
|
||||
}
|
||||
#endif // CK_ENABLE_BF16
|
||||
#endif // CK_USE_WMMA
|
||||
|
||||
return op_ptrs;
|
||||
}
|
||||
|
||||
@@ -1,18 +1,31 @@
|
||||
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
|
||||
# SPDX-License-Identifier: MIT
|
||||
|
||||
# ONLY XDL_KERNELS
|
||||
# ONLY XDL_AND_WMMA_KERNELS
|
||||
set(GROUPED_GEMM_FIXED_NK_INSTANCES)
|
||||
|
||||
list(APPEND GROUPED_GEMM_FIXED_NK_INSTANCES device_grouped_gemm_xdl_fixed_nk_f16_f16_f16_mk_kn_mn_instance.cpp
|
||||
device_grouped_gemm_xdl_fixed_nk_f16_f16_f16_mk_nk_mn_instance.cpp
|
||||
device_grouped_gemm_xdl_fixed_nk_f16_fp8_f16_mk_kn_mn_instance.cpp
|
||||
device_grouped_gemm_xdl_fixed_nk_f16_fp8_f16_mk_nk_mn_instance.cpp
|
||||
device_grouped_gemm_xdl_fixed_nk_f16_i8_f16_mk_kn_mn_instance.cpp
|
||||
device_grouped_gemm_xdl_fixed_nk_f16_i8_f16_mk_nk_mn_instance.cpp
|
||||
device_grouped_gemm_xdl_fixed_nk_bf16_i8_bf16_mk_kn_mn_instance.cpp
|
||||
device_grouped_gemm_xdl_fixed_nk_bf16_i8_bf16_mk_nk_mn_instance.cpp
|
||||
device_grouped_gemm_xdl_fixed_nk_bf16_bf16_bf16_mk_kn_mn_instance.cpp
|
||||
device_grouped_gemm_xdl_fixed_nk_bf16_bf16_bf16_mk_nk_mn_instance.cpp)
|
||||
list(APPEND GROUPED_GEMM_FIXED_NK_INSTANCES
|
||||
device_grouped_gemm_xdl_fixed_nk_f16_f16_f16_mk_kn_mn_instance.cpp
|
||||
device_grouped_gemm_xdl_fixed_nk_f16_f16_f16_mk_nk_mn_instance.cpp
|
||||
device_grouped_gemm_xdl_fixed_nk_f16_fp8_f16_mk_kn_mn_instance.cpp
|
||||
device_grouped_gemm_xdl_fixed_nk_f16_fp8_f16_mk_nk_mn_instance.cpp
|
||||
device_grouped_gemm_xdl_fixed_nk_f16_i8_f16_mk_kn_mn_instance.cpp
|
||||
device_grouped_gemm_xdl_fixed_nk_f16_i8_f16_mk_nk_mn_instance.cpp
|
||||
device_grouped_gemm_xdl_fixed_nk_bf16_i8_bf16_mk_kn_mn_instance.cpp
|
||||
device_grouped_gemm_xdl_fixed_nk_bf16_i8_bf16_mk_nk_mn_instance.cpp
|
||||
device_grouped_gemm_xdl_fixed_nk_bf16_bf16_bf16_mk_kn_mn_instance.cpp
|
||||
device_grouped_gemm_xdl_fixed_nk_bf16_bf16_bf16_mk_nk_mn_instance.cpp
|
||||
|
||||
device_grouped_gemm_wmma_fixed_nk_bf16_bf16_bf16_mk_kn_mn_instance.cpp
|
||||
device_grouped_gemm_wmma_fixed_nk_bf16_bf16_bf16_mk_nk_mn_instance.cpp
|
||||
device_grouped_gemm_wmma_fixed_nk_bf16_i8_bf16_mk_kn_mn_instance.cpp
|
||||
device_grouped_gemm_wmma_fixed_nk_bf16_i8_bf16_mk_nk_mn_instance.cpp
|
||||
device_grouped_gemm_wmma_fixed_nk_f16_f16_f16_mk_kn_mn_instance.cpp
|
||||
device_grouped_gemm_wmma_fixed_nk_f16_f16_f16_mk_nk_mn_instance.cpp
|
||||
device_grouped_gemm_wmma_fixed_nk_f16_fp8_f16_mk_kn_mn_instance.cpp
|
||||
device_grouped_gemm_wmma_fixed_nk_f16_fp8_f16_mk_nk_mn_instance.cpp
|
||||
device_grouped_gemm_wmma_fixed_nk_f16_i8_f16_mk_kn_mn_instance.cpp
|
||||
device_grouped_gemm_wmma_fixed_nk_f16_i8_f16_mk_nk_mn_instance.cpp
|
||||
)
|
||||
|
||||
add_instance_library(device_grouped_gemm_fixed_nk_instance ${GROUPED_GEMM_FIXED_NK_INSTANCES})
|
||||
@@ -0,0 +1,37 @@
|
||||
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
|
||||
// SPDX-License-Identifier: MIT
|
||||
|
||||
#include <cstdlib>
|
||||
|
||||
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
|
||||
#include "ck/library/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_wmma_fixed_nk_instance.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
|
||||
void add_device_grouped_gemm_wmma_fixed_nk_bf16_bf16_bf16_mk_kn_mn_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedGemmFixedNK<Row,
|
||||
Row,
|
||||
DsLayout,
|
||||
Row,
|
||||
BF16,
|
||||
BF16,
|
||||
DsDataType,
|
||||
BF16,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
PassThrough>>>& instances)
|
||||
{
|
||||
add_device_grouped_gemm_wmma_fixed_nk_instances<
|
||||
BF16,
|
||||
Row,
|
||||
Row,
|
||||
device_grouped_gemm_wmma_fixed_nk_mk_kn_mn_instances>(instances);
|
||||
}
|
||||
|
||||
} // namespace instance
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
@@ -0,0 +1,37 @@
|
||||
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
|
||||
// SPDX-License-Identifier: MIT
|
||||
|
||||
#include <cstdlib>
|
||||
|
||||
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
|
||||
#include "ck/library/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_wmma_fixed_nk_instance.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
|
||||
void add_device_grouped_gemm_wmma_fixed_nk_bf16_bf16_bf16_mk_nk_mn_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedGemmFixedNK<Row,
|
||||
Col,
|
||||
DsLayout,
|
||||
Row,
|
||||
BF16,
|
||||
BF16,
|
||||
DsDataType,
|
||||
BF16,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
PassThrough>>>& instances)
|
||||
{
|
||||
add_device_grouped_gemm_wmma_fixed_nk_instances<
|
||||
BF16,
|
||||
Row,
|
||||
Col,
|
||||
device_grouped_gemm_wmma_fixed_nk_mk_nk_mn_instances>(instances);
|
||||
}
|
||||
|
||||
} // namespace instance
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
@@ -0,0 +1,38 @@
|
||||
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
|
||||
// SPDX-License-Identifier: MIT
|
||||
|
||||
#include <cstdlib>
|
||||
|
||||
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
|
||||
#include "ck/library/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_wmma_fixed_nk_instance.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
|
||||
void add_device_grouped_gemm_wmma_fixed_nk_bf16_i8_bf16_mk_kn_mn_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedGemmFixedNK<Row,
|
||||
Row,
|
||||
DsLayout,
|
||||
Row,
|
||||
BF16,
|
||||
I8,
|
||||
DsDataType,
|
||||
BF16,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
PassThrough>>>& instances)
|
||||
{
|
||||
add_device_grouped_gemm_wmma_fixed_nk_irregular_instances<
|
||||
BF16,
|
||||
I8,
|
||||
Row,
|
||||
Row,
|
||||
device_grouped_gemm_wmma_fixed_nk_mk_kn_mn_irregular_instances>(instances);
|
||||
}
|
||||
|
||||
} // namespace instance
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
@@ -0,0 +1,38 @@
|
||||
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
|
||||
// SPDX-License-Identifier: MIT
|
||||
|
||||
#include <cstdlib>
|
||||
|
||||
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
|
||||
#include "ck/library/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_wmma_fixed_nk_instance.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
|
||||
void add_device_grouped_gemm_wmma_fixed_nk_bf16_i8_bf16_mk_nk_mn_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedGemmFixedNK<Row,
|
||||
Col,
|
||||
DsLayout,
|
||||
Row,
|
||||
BF16,
|
||||
I8,
|
||||
DsDataType,
|
||||
BF16,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
PassThrough>>>& instances)
|
||||
{
|
||||
add_device_grouped_gemm_wmma_fixed_nk_irregular_instances<
|
||||
BF16,
|
||||
I8,
|
||||
Row,
|
||||
Col,
|
||||
device_grouped_gemm_wmma_fixed_nk_mk_nk_mn_irregular_instances>(instances);
|
||||
}
|
||||
|
||||
} // namespace instance
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
@@ -0,0 +1,37 @@
|
||||
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
|
||||
// SPDX-License-Identifier: MIT
|
||||
|
||||
#include <cstdlib>
|
||||
|
||||
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
|
||||
#include "ck/library/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_wmma_fixed_nk_instance.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
|
||||
void add_device_grouped_gemm_wmma_fixed_nk_f16_f16_f16_mk_kn_mn_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedGemmFixedNK<Row,
|
||||
Row,
|
||||
DsLayout,
|
||||
Row,
|
||||
F16,
|
||||
F16,
|
||||
DsDataType,
|
||||
F16,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
PassThrough>>>& instances)
|
||||
{
|
||||
add_device_grouped_gemm_wmma_fixed_nk_instances<
|
||||
F16,
|
||||
Row,
|
||||
Row,
|
||||
device_grouped_gemm_wmma_fixed_nk_mk_kn_mn_instances>(instances);
|
||||
}
|
||||
|
||||
} // namespace instance
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
@@ -0,0 +1,37 @@
|
||||
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
|
||||
// SPDX-License-Identifier: MIT
|
||||
|
||||
#include <cstdlib>
|
||||
|
||||
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
|
||||
#include "ck/library/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_wmma_fixed_nk_instance.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
|
||||
void add_device_grouped_gemm_wmma_fixed_nk_f16_f16_f16_mk_nk_mn_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedGemmFixedNK<Row,
|
||||
Col,
|
||||
DsLayout,
|
||||
Row,
|
||||
F16,
|
||||
F16,
|
||||
DsDataType,
|
||||
F16,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
PassThrough>>>& instances)
|
||||
{
|
||||
add_device_grouped_gemm_wmma_fixed_nk_instances<
|
||||
F16,
|
||||
Row,
|
||||
Col,
|
||||
device_grouped_gemm_wmma_fixed_nk_mk_nk_mn_instances>(instances);
|
||||
}
|
||||
|
||||
} // namespace instance
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
@@ -0,0 +1,38 @@
|
||||
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
|
||||
// SPDX-License-Identifier: MIT
|
||||
|
||||
#include <cstdlib>
|
||||
|
||||
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
|
||||
#include "ck/library/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_wmma_fixed_nk_instance.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
|
||||
void add_device_grouped_gemm_wmma_fixed_nk_f16_fp8_f16_mk_kn_mn_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedGemmFixedNK<Row,
|
||||
Row,
|
||||
DsLayout,
|
||||
Row,
|
||||
F16,
|
||||
F8,
|
||||
DsDataType,
|
||||
F16,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
PassThrough>>>& instances)
|
||||
{
|
||||
add_device_grouped_gemm_wmma_fixed_nk_irregular_instances<
|
||||
F16,
|
||||
F8,
|
||||
Row,
|
||||
Row,
|
||||
device_grouped_gemm_wmma_fixed_nk_mk_kn_mn_irregular_instances>(instances);
|
||||
}
|
||||
|
||||
} // namespace instance
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
@@ -0,0 +1,38 @@
|
||||
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
|
||||
// SPDX-License-Identifier: MIT
|
||||
|
||||
#include <cstdlib>
|
||||
|
||||
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
|
||||
#include "ck/library/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_wmma_fixed_nk_instance.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
|
||||
void add_device_grouped_gemm_wmma_fixed_nk_f16_fp8_f16_mk_nk_mn_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedGemmFixedNK<Row,
|
||||
Col,
|
||||
DsLayout,
|
||||
Row,
|
||||
F16,
|
||||
F8,
|
||||
DsDataType,
|
||||
F16,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
PassThrough>>>& instances)
|
||||
{
|
||||
add_device_grouped_gemm_wmma_fixed_nk_irregular_instances<
|
||||
F16,
|
||||
F8,
|
||||
Row,
|
||||
Col,
|
||||
device_grouped_gemm_wmma_fixed_nk_mk_nk_mn_irregular_instances>(instances);
|
||||
}
|
||||
|
||||
} // namespace instance
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
@@ -0,0 +1,38 @@
|
||||
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
|
||||
// SPDX-License-Identifier: MIT
|
||||
|
||||
#include <cstdlib>
|
||||
|
||||
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
|
||||
#include "ck/library/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_wmma_fixed_nk_instance.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
|
||||
void add_device_grouped_gemm_wmma_fixed_nk_f16_i8_f16_mk_kn_mn_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedGemmFixedNK<Row,
|
||||
Row,
|
||||
DsLayout,
|
||||
Row,
|
||||
F16,
|
||||
I8,
|
||||
DsDataType,
|
||||
F16,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
PassThrough>>>& instances)
|
||||
{
|
||||
add_device_grouped_gemm_wmma_fixed_nk_irregular_instances<
|
||||
F16,
|
||||
I8,
|
||||
Row,
|
||||
Row,
|
||||
device_grouped_gemm_wmma_fixed_nk_mk_kn_mn_irregular_instances>(instances);
|
||||
}
|
||||
|
||||
} // namespace instance
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
@@ -0,0 +1,38 @@
|
||||
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
|
||||
// SPDX-License-Identifier: MIT
|
||||
|
||||
#include <cstdlib>
|
||||
|
||||
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
|
||||
#include "ck/library/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_wmma_fixed_nk_instance.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
|
||||
void add_device_grouped_gemm_wmma_fixed_nk_f16_i8_f16_mk_nk_mn_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedGemmFixedNK<Row,
|
||||
Col,
|
||||
DsLayout,
|
||||
Row,
|
||||
F16,
|
||||
I8,
|
||||
DsDataType,
|
||||
F16,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
PassThrough>>>& instances)
|
||||
{
|
||||
add_device_grouped_gemm_wmma_fixed_nk_irregular_instances<
|
||||
F16,
|
||||
I8,
|
||||
Row,
|
||||
Col,
|
||||
device_grouped_gemm_wmma_fixed_nk_mk_nk_mn_irregular_instances>(instances);
|
||||
}
|
||||
|
||||
} // namespace instance
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
@@ -30,7 +30,8 @@ using DsLayout = ck::Tuple<>;
|
||||
|
||||
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
|
||||
|
||||
static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
|
||||
// static constexpr auto GemmMNKPadding =
|
||||
// ck::tensor_operation::device::GemmSpecialization::MNKPadding;
|
||||
|
||||
using device_grouped_gemm_xdl_fixed_nk_f16_f16_f16_mk_kn_mn_irregular_tile_instances = std::tuple<
|
||||
// clang-format off
|
||||
@@ -38,16 +39,16 @@ using device_grouped_gemm_xdl_fixed_nk_f16_f16_f16_mk_kn_mn_irregular_tile_insta
|
||||
//############################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Spacialization| 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|
|
||||
//############################| | | | | | | | | | | 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|
|
||||
//############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGroupedGemm_Xdl_Fixed_NK< Row, Row, DsLayout, Row, F16, F16, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S< 1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedGemm_Xdl_Fixed_NK< Row, Row, DsLayout, Row, F16, F16, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S< 1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedGemm_Xdl_Fixed_NK< Row, Row, DsLayout, Row, F16, F16, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 2, 32, 32, 2, 1, S< 1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 16,16, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedGemm_Xdl_Fixed_NK< Row, Row, DsLayout, Row, F16, F16, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S< 1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedGemm_Xdl_Fixed_NK< Row, Row, DsLayout, Row, F16, F16, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 2, 32, 32, 1, 2, S< 1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedGemm_Xdl_Fixed_NK< Row, Row, DsLayout, Row, F16, F16, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S< 1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedGemm_Xdl_Fixed_NK< Row, Row, DsLayout, Row, F16, F16, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 2, 32, 32, 2, 2, S< 1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 16, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 2, 0, 1, 1, S<1, 32, 1, 4>, 8>,
|
||||
DeviceGroupedGemm_Xdl_Fixed_NK< Row, Row, DsLayout, Row, F16, F16, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S< 1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
|
||||
DeviceGroupedGemm_Xdl_Fixed_NK< Row, Row, DsLayout, Row, F16, F16, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 64, 128, 32, 8, 2, 32, 32, 2, 2, S< 1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 2, 0, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
DeviceGroupedGemm_Xdl_Fixed_NK< Row, Row, DsLayout, Row, F16, F16, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S< 1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>
|
||||
// DeviceGroupedGemm_Xdl_Fixed_NK< Row, Row, DsLayout, Row, F16, F16, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S< 1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
// DeviceGroupedGemm_Xdl_Fixed_NK< Row, Row, DsLayout, Row, F16, F16, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S< 1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
// DeviceGroupedGemm_Xdl_Fixed_NK< Row, Row, DsLayout, Row, F16, F16, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 2, 32, 32, 2, 1, S< 1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 16,16, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
// DeviceGroupedGemm_Xdl_Fixed_NK< Row, Row, DsLayout, Row, F16, F16, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S< 1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
// DeviceGroupedGemm_Xdl_Fixed_NK< Row, Row, DsLayout, Row, F16, F16, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 2, 32, 32, 1, 2, S< 1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
// DeviceGroupedGemm_Xdl_Fixed_NK< Row, Row, DsLayout, Row, F16, F16, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S< 1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
// DeviceGroupedGemm_Xdl_Fixed_NK< Row, Row, DsLayout, Row, F16, F16, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 2, 32, 32, 2, 2, S< 1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 16, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 2, 0, 1, 1, S<1, 32, 1, 4>, 8>,
|
||||
// DeviceGroupedGemm_Xdl_Fixed_NK< Row, Row, DsLayout, Row, F16, F16, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S< 1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
|
||||
// DeviceGroupedGemm_Xdl_Fixed_NK< Row, Row, DsLayout, Row, F16, F16, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 64, 128, 32, 8, 2, 32, 32, 2, 2, S< 1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 2, 0, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
// DeviceGroupedGemm_Xdl_Fixed_NK< Row, Row, DsLayout, Row, F16, F16, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S< 1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
|
||||
@@ -30,27 +30,28 @@ using DsLayout = ck::Tuple<>;
|
||||
|
||||
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
|
||||
|
||||
static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
|
||||
// static constexpr auto GemmMNKPadding =
|
||||
// ck::tensor_operation::device::GemmSpecialization::MNKPadding;
|
||||
|
||||
using device_grouped_gemm_xdl_fixed_nk_f16_f16_f16_mk_nk_mn_irregular_tile_instances = std::tuple<
|
||||
// clang-format off
|
||||
//############################| A| B| Ds| E| 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|
|
||||
//############################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Spacialization| 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|
|
||||
//############################| | | | | | | | | | | 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|
|
||||
//############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, F16, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 128, 256, 64, 8, 8, 32, 32, 2, 4, S<1, 8, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, F16, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 128, 128, 64, 8, 8, 32, 32, 2, 2, S<1, 8, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, F16, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 128, 64, 64, 8, 8, 32, 32, 2, 1, S<1, 8, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, F16, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 64, 128, 64, 8, 8, 32, 32, 1, 2, S<1, 8, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, F16, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 128, 128, 64, 8, 8, 32, 32, 4, 2, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, F16, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 128, 64, 64, 8, 8, 32, 32, 2, 2, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
|
||||
DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, F16, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 64, 128, 64, 8, 8, 32, 32, 2, 2, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, F16, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 128, 32, 64, 8, 8, 32, 32, 2, 1, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
|
||||
DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, F16, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 32, 128, 64, 8, 8, 32, 32, 1, 2, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, F16, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 32, 256, 64, 8, 8, 32, 32, 1, 4, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, F16, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 64, 64, 64, 64, 8, 8, 32, 32, 2, 2, S<1, 8, 8, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 8, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
|
||||
DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, F16, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 64, 64, 32, 64, 8, 8, 32, 32, 2, 1, S<1, 8, 8, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 8, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
|
||||
DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, F16, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 64, 32, 64, 64, 8, 8, 32, 32, 1, 2, S<1, 8, 8, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 8, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>
|
||||
// //############################| A| B| Ds| E| 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|
|
||||
// //############################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Spacialization| 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|
|
||||
// //############################| | | | | | | | | | | 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|
|
||||
// //############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
// DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, F16, F32, F16, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 128, 256, 64, 8, 8, 32, 32, 2, 4, S<1, 8, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
// DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, F16, F32, F16, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 128, 128, 64, 8, 8, 32, 32, 2, 2, S<1, 8, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
// DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, F16, F32, F16, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 128, 64, 64, 8, 8, 32, 32, 2, 1, S<1, 8, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
// DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, F16, F32, F16, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 64, 128, 64, 8, 8, 32, 32, 1, 2, S<1, 8, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
// DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, F16, F32, F16, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 128, 128, 64, 8, 8, 32, 32, 4, 2, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
// DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, F16, F32, F16, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 128, 64, 64, 8, 8, 32, 32, 2, 2, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
|
||||
// DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, F16, F32, F16, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 64, 128, 64, 8, 8, 32, 32, 2, 2, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
// DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, F16, F32, F16, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 128, 32, 64, 8, 8, 32, 32, 2, 1, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
|
||||
// DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, F16, F32, F16, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 32, 128, 64, 8, 8, 32, 32, 1, 2, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
// DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, F16, F32, F16, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 32, 256, 64, 8, 8, 32, 32, 1, 4, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
// DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, F16, F32, F16, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 64, 64, 64, 64, 8, 8, 32, 32, 2, 2, S<1, 8, 8, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 8, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
|
||||
// DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, F16, F32, F16, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 64, 64, 32, 64, 8, 8, 32, 32, 2, 1, S<1, 8, 8, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 8, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
|
||||
// DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, F16, F32, F16, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 64, 32, 64, 64, 8, 8, 32, 32, 1, 2, S<1, 8, 8, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 8, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
|
||||
@@ -31,7 +31,8 @@ using DsLayout = ck::Tuple<>;
|
||||
|
||||
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
|
||||
|
||||
static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
|
||||
// static constexpr auto GemmMNKPadding =
|
||||
// ck::tensor_operation::device::GemmSpecialization::MNKPadding;
|
||||
|
||||
using device_grouped_gemm_xdl_fixed_nk_f16_f8_f16_mk_kn_mn_irregular_tile_instances = std::tuple<
|
||||
// clang-format off
|
||||
@@ -39,16 +40,16 @@ using device_grouped_gemm_xdl_fixed_nk_f16_f8_f16_mk_kn_mn_irregular_tile_instan
|
||||
//############################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Spacialization| 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|
|
||||
//############################| | | | | | | | | | | 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|
|
||||
//############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGroupedGemm_Xdl_Fixed_NK< Row, Row, DsLayout, Row, F16, F8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S< 1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedGemm_Xdl_Fixed_NK< Row, Row, DsLayout, Row, F16, F8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S< 1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedGemm_Xdl_Fixed_NK< Row, Row, DsLayout, Row, F16, F8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 2, 32, 32, 2, 1, S< 1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 16,16, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedGemm_Xdl_Fixed_NK< Row, Row, DsLayout, Row, F16, F8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S< 1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedGemm_Xdl_Fixed_NK< Row, Row, DsLayout, Row, F16, F8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 2, 32, 32, 1, 2, S< 1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedGemm_Xdl_Fixed_NK< Row, Row, DsLayout, Row, F16, F8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S< 1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedGemm_Xdl_Fixed_NK< Row, Row, DsLayout, Row, F16, F8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 2, 32, 32, 2, 2, S< 1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 16, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 2, 0, 1, 1, S<1, 32, 1, 4>, 8>,
|
||||
DeviceGroupedGemm_Xdl_Fixed_NK< Row, Row, DsLayout, Row, F16, F8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S< 1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
|
||||
DeviceGroupedGemm_Xdl_Fixed_NK< Row, Row, DsLayout, Row, F16, F8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 64, 128, 32, 8, 2, 32, 32, 2, 2, S< 1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 2, 0, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
DeviceGroupedGemm_Xdl_Fixed_NK< Row, Row, DsLayout, Row, F16, F8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S< 1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>
|
||||
// DeviceGroupedGemm_Xdl_Fixed_NK< Row, Row, DsLayout, Row, F16, F8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S< 1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
// DeviceGroupedGemm_Xdl_Fixed_NK< Row, Row, DsLayout, Row, F16, F8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S< 1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
// DeviceGroupedGemm_Xdl_Fixed_NK< Row, Row, DsLayout, Row, F16, F8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 2, 32, 32, 2, 1, S< 1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 16,16, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
// DeviceGroupedGemm_Xdl_Fixed_NK< Row, Row, DsLayout, Row, F16, F8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S< 1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
// DeviceGroupedGemm_Xdl_Fixed_NK< Row, Row, DsLayout, Row, F16, F8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 2, 32, 32, 1, 2, S< 1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
// DeviceGroupedGemm_Xdl_Fixed_NK< Row, Row, DsLayout, Row, F16, F8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S< 1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
// DeviceGroupedGemm_Xdl_Fixed_NK< Row, Row, DsLayout, Row, F16, F8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 2, 32, 32, 2, 2, S< 1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 16, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 2, 0, 1, 1, S<1, 32, 1, 4>, 8>,
|
||||
// DeviceGroupedGemm_Xdl_Fixed_NK< Row, Row, DsLayout, Row, F16, F8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S< 1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
|
||||
// DeviceGroupedGemm_Xdl_Fixed_NK< Row, Row, DsLayout, Row, F16, F8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 64, 128, 32, 8, 2, 32, 32, 2, 2, S< 1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 2, 0, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
// DeviceGroupedGemm_Xdl_Fixed_NK< Row, Row, DsLayout, Row, F16, F8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S< 1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
|
||||
@@ -31,27 +31,28 @@ using DsLayout = ck::Tuple<>;
|
||||
|
||||
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
|
||||
|
||||
static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
|
||||
// static constexpr auto GemmMNKPadding =
|
||||
// ck::tensor_operation::device::GemmSpecialization::MNKPadding;
|
||||
|
||||
using device_grouped_gemm_xdl_fixed_nk_f16_f8_f16_mk_nk_mn_irregular_tile_instances = std::tuple<
|
||||
// clang-format off
|
||||
//############################| A| B| Ds| E| 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|
|
||||
//############################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Spacialization| 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|
|
||||
//############################| | | | | | | | | | | 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|
|
||||
//############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, F8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 128, 256, 64, 8, 8, 32, 32, 2, 4, S<1, 8, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, F8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 128, 128, 64, 8, 8, 32, 32, 2, 2, S<1, 8, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, F8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 128, 64, 64, 8, 8, 32, 32, 2, 1, S<1, 8, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, F8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 64, 128, 64, 8, 8, 32, 32, 1, 2, S<1, 8, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, F8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 128, 128, 64, 8, 8, 32, 32, 4, 2, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, F8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 128, 64, 64, 8, 8, 32, 32, 2, 2, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
|
||||
DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, F8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 64, 128, 64, 8, 8, 32, 32, 2, 2, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, F8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 128, 32, 64, 8, 8, 32, 32, 2, 1, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
|
||||
DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, F8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 32, 128, 64, 8, 8, 32, 32, 1, 2, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, F8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 32, 256, 64, 8, 8, 32, 32, 1, 4, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, F8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 64, 64, 64, 64, 8, 8, 32, 32, 2, 2, S<1, 8, 8, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 8, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
|
||||
DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, F8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 64, 64, 32, 64, 8, 8, 32, 32, 2, 1, S<1, 8, 8, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 8, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
|
||||
DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, F8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 64, 32, 64, 64, 8, 8, 32, 32, 1, 2, S<1, 8, 8, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 8, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>
|
||||
// //############################| A| B| Ds| E| 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|
|
||||
// //############################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Spacialization| 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|
|
||||
// //############################| | | | | | | | | | | 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|
|
||||
// //############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
// DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, F8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 128, 256, 64, 8, 8, 32, 32, 2, 4, S<1, 8, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
// DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, F8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 128, 128, 64, 8, 8, 32, 32, 2, 2, S<1, 8, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
// DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, F8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 128, 64, 64, 8, 8, 32, 32, 2, 1, S<1, 8, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
// DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, F8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 64, 128, 64, 8, 8, 32, 32, 1, 2, S<1, 8, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
// DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, F8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 128, 128, 64, 8, 8, 32, 32, 4, 2, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
// DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, F8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 128, 64, 64, 8, 8, 32, 32, 2, 2, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
|
||||
// DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, F8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 64, 128, 64, 8, 8, 32, 32, 2, 2, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
// DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, F8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 128, 32, 64, 8, 8, 32, 32, 2, 1, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
|
||||
// DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, F8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 32, 128, 64, 8, 8, 32, 32, 1, 2, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
// DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, F8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 32, 256, 64, 8, 8, 32, 32, 1, 4, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
// DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, F8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 64, 64, 64, 64, 8, 8, 32, 32, 2, 2, S<1, 8, 8, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 8, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
|
||||
// DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, F8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 64, 64, 32, 64, 8, 8, 32, 32, 2, 1, S<1, 8, 8, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 8, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
|
||||
// DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, F8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 64, 32, 64, 64, 8, 8, 32, 32, 1, 2, S<1, 8, 8, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 8, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
|
||||
@@ -31,7 +31,8 @@ using DsLayout = ck::Tuple<>;
|
||||
|
||||
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
|
||||
|
||||
static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
|
||||
// static constexpr auto GemmMNKPadding =
|
||||
// ck::tensor_operation::device::GemmSpecialization::MNKPadding;
|
||||
|
||||
using device_grouped_gemm_xdl_fixed_nk_f16_i8_f16_mk_kn_mn_irregular_tile_instances = std::tuple<
|
||||
// clang-format off
|
||||
@@ -39,16 +40,16 @@ using device_grouped_gemm_xdl_fixed_nk_f16_i8_f16_mk_kn_mn_irregular_tile_instan
|
||||
//############################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Spacialization| 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|
|
||||
//############################| | | | | | | | | | | 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|
|
||||
//############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGroupedGemm_Xdl_Fixed_NK< Row, Row, DsLayout, Row, F16, I8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S< 1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedGemm_Xdl_Fixed_NK< Row, Row, DsLayout, Row, F16, I8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S< 1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedGemm_Xdl_Fixed_NK< Row, Row, DsLayout, Row, F16, I8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 2, 32, 32, 2, 1, S< 1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 16,16, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedGemm_Xdl_Fixed_NK< Row, Row, DsLayout, Row, F16, I8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S< 1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedGemm_Xdl_Fixed_NK< Row, Row, DsLayout, Row, F16, I8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 2, 32, 32, 1, 2, S< 1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedGemm_Xdl_Fixed_NK< Row, Row, DsLayout, Row, F16, I8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S< 1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedGemm_Xdl_Fixed_NK< Row, Row, DsLayout, Row, F16, I8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 2, 32, 32, 2, 2, S< 1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 16, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 2, 0, 1, 1, S<1, 32, 1, 4>, 8>,
|
||||
DeviceGroupedGemm_Xdl_Fixed_NK< Row, Row, DsLayout, Row, F16, I8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S< 1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
|
||||
DeviceGroupedGemm_Xdl_Fixed_NK< Row, Row, DsLayout, Row, F16, I8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 64, 128, 32, 8, 2, 32, 32, 2, 2, S< 1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 2, 0, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
DeviceGroupedGemm_Xdl_Fixed_NK< Row, Row, DsLayout, Row, F16, I8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S< 1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>
|
||||
// DeviceGroupedGemm_Xdl_Fixed_NK< Row, Row, DsLayout, Row, F16, I8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S< 1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
// DeviceGroupedGemm_Xdl_Fixed_NK< Row, Row, DsLayout, Row, F16, I8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S< 1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
// DeviceGroupedGemm_Xdl_Fixed_NK< Row, Row, DsLayout, Row, F16, I8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 2, 32, 32, 2, 1, S< 1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 16,16, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
// DeviceGroupedGemm_Xdl_Fixed_NK< Row, Row, DsLayout, Row, F16, I8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S< 1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
// DeviceGroupedGemm_Xdl_Fixed_NK< Row, Row, DsLayout, Row, F16, I8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 2, 32, 32, 1, 2, S< 1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
// DeviceGroupedGemm_Xdl_Fixed_NK< Row, Row, DsLayout, Row, F16, I8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S< 1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
// DeviceGroupedGemm_Xdl_Fixed_NK< Row, Row, DsLayout, Row, F16, I8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 2, 32, 32, 2, 2, S< 1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 16, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 2, 0, 1, 1, S<1, 32, 1, 4>, 8>,
|
||||
// DeviceGroupedGemm_Xdl_Fixed_NK< Row, Row, DsLayout, Row, F16, I8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S< 1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
|
||||
// DeviceGroupedGemm_Xdl_Fixed_NK< Row, Row, DsLayout, Row, F16, I8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 64, 128, 32, 8, 2, 32, 32, 2, 2, S< 1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 2, 0, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
// DeviceGroupedGemm_Xdl_Fixed_NK< Row, Row, DsLayout, Row, F16, I8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S< 1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
|
||||
@@ -31,27 +31,28 @@ using DsLayout = ck::Tuple<>;
|
||||
|
||||
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
|
||||
|
||||
static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
|
||||
// static constexpr auto GemmMNKPadding =
|
||||
// ck::tensor_operation::device::GemmSpecialization::MNKPadding;
|
||||
|
||||
using device_grouped_gemm_xdl_fixed_nk_f16_i8_f16_mk_nk_mn_irregular_tile_instances = std::tuple<
|
||||
// clang-format off
|
||||
//############################| A| B| Ds| E| 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|
|
||||
//############################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Spacialization| 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|
|
||||
//############################| | | | | | | | | | | 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|
|
||||
//############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, I8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 128, 256, 64, 8, 8, 32, 32, 2, 4, S<1, 8, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, I8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 128, 128, 64, 8, 8, 32, 32, 2, 2, S<1, 8, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, I8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 128, 64, 64, 8, 8, 32, 32, 2, 1, S<1, 8, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, I8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 64, 128, 64, 8, 8, 32, 32, 1, 2, S<1, 8, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, I8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 128, 128, 64, 8, 8, 32, 32, 4, 2, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, I8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 128, 64, 64, 8, 8, 32, 32, 2, 2, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
|
||||
DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, I8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 64, 128, 64, 8, 8, 32, 32, 2, 2, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, I8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 128, 32, 64, 8, 8, 32, 32, 2, 1, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
|
||||
DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, I8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 32, 128, 64, 8, 8, 32, 32, 1, 2, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, I8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 32, 256, 64, 8, 8, 32, 32, 1, 4, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, I8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 64, 64, 64, 64, 8, 8, 32, 32, 2, 2, S<1, 8, 8, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 8, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
|
||||
DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, I8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 64, 64, 32, 64, 8, 8, 32, 32, 2, 1, S<1, 8, 8, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 8, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
|
||||
DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, I8, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 64, 32, 64, 64, 8, 8, 32, 32, 1, 2, S<1, 8, 8, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 8, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>
|
||||
// //############################| A| B| Ds| E| 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|
|
||||
// //############################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Spacialization| 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|
|
||||
// //############################| | | | | | | | | | | 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|
|
||||
// //############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
// DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, I8, F32, F16, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 128, 256, 64, 8, 8, 32, 32, 2, 4, S<1, 8, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
// DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, I8, F32, F16, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 128, 128, 64, 8, 8, 32, 32, 2, 2, S<1, 8, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
// DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, I8, F32, F16, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 128, 64, 64, 8, 8, 32, 32, 2, 1, S<1, 8, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
// DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, I8, F32, F16, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 64, 128, 64, 8, 8, 32, 32, 1, 2, S<1, 8, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
// DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, I8, F32, F16, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 128, 128, 64, 8, 8, 32, 32, 4, 2, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
// DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, I8, F32, F16, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 128, 64, 64, 8, 8, 32, 32, 2, 2, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
|
||||
// DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, I8, F32, F16, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 64, 128, 64, 8, 8, 32, 32, 2, 2, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
// DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, I8, F32, F16, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 128, 32, 64, 8, 8, 32, 32, 2, 1, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
|
||||
// DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, I8, F32, F16, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 32, 128, 64, 8, 8, 32, 32, 1, 2, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
// DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, I8, F32, F16, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 32, 256, 64, 8, 8, 32, 32, 1, 4, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
// DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, I8, F32, F16, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 64, 64, 64, 64, 8, 8, 32, 32, 2, 2, S<1, 8, 8, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 8, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
|
||||
// DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, I8, F32, F16, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 64, 64, 32, 64, 8, 8, 32, 32, 2, 1, S<1, 8, 8, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 8, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
|
||||
// DeviceGroupedGemm_Xdl_Fixed_NK< Row, Col, DsLayout, Row, F16, I8, F32, F16, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 64, 32, 64, 64, 8, 8, 32, 32, 1, 2, S<1, 8, 8, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 8, 8, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
|
||||
@@ -46,7 +46,8 @@ bool profile_grouped_gemm_fixed_nk_impl(int do_verification,
|
||||
int n_warmup = 1,
|
||||
int n_iter = 10)
|
||||
{
|
||||
bool pass = true;
|
||||
bool pass = true;
|
||||
using ComputeDataType = ADataType;
|
||||
|
||||
auto f_host_tensor_descriptor =
|
||||
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
|
||||
@@ -54,11 +55,11 @@ bool profile_grouped_gemm_fixed_nk_impl(int do_verification,
|
||||
|
||||
if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
|
||||
{
|
||||
return HostTensorDescriptor({row, col}, {stride, 1_uz});
|
||||
return HostTensorDescriptor({row, col}, {stride, 1_uz}, layout);
|
||||
}
|
||||
else
|
||||
{
|
||||
return HostTensorDescriptor({row, col}, {1_uz, stride});
|
||||
return HostTensorDescriptor({row, col}, {1_uz, stride}, layout);
|
||||
}
|
||||
};
|
||||
|
||||
@@ -74,8 +75,8 @@ bool profile_grouped_gemm_fixed_nk_impl(int do_verification,
|
||||
std::vector<Tensor<BDataType>> b_k_n;
|
||||
std::vector<Tensor<CDataType>> c_m_n_host_results;
|
||||
std::vector<Tensor<CDataType>> c_m_n_device_results;
|
||||
int sum_of_m = 0;
|
||||
|
||||
double max_abs_in_val = 0.f;
|
||||
int sum_of_m = 0;
|
||||
for(std::size_t i = 0; i < group_count; i++)
|
||||
{
|
||||
sum_of_m += Ms[i];
|
||||
@@ -95,17 +96,18 @@ bool profile_grouped_gemm_fixed_nk_impl(int do_verification,
|
||||
<< i << "]:" << b_k_n[i].mDesc << ", c_m_n_device_results[" << i
|
||||
<< "]:" << c_m_n_device_results[i].mDesc << std::endl;
|
||||
}
|
||||
std::size_t num_thread = 1;
|
||||
switch(init_method)
|
||||
{
|
||||
case 0: break;
|
||||
case 1:
|
||||
a_m_k[i].GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5}, num_thread);
|
||||
b_k_n[i].GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5}, num_thread);
|
||||
ck::utils::FillUniformDistributionIntegerValue<ADataType>{-5.f, 5.f}(a_m_k[i]);
|
||||
ck::utils::FillUniformDistributionIntegerValue<BDataType>{-5.f, 5.f}(b_k_n[i]);
|
||||
max_abs_in_val = 5.f;
|
||||
break;
|
||||
default:
|
||||
a_m_k[i].GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0}, num_thread);
|
||||
b_k_n[i].GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5}, num_thread);
|
||||
ck::utils::FillUniformDistribution<ADataType>{0.0f, 1.0f}(a_m_k[i]);
|
||||
ck::utils::FillUniformDistribution<BDataType>{-0.5f, 0.5f}(b_k_n[i]);
|
||||
max_abs_in_val = 1.0f;
|
||||
}
|
||||
}
|
||||
|
||||
@@ -282,23 +284,18 @@ bool profile_grouped_gemm_fixed_nk_impl(int do_verification,
|
||||
bool instance_pass = true;
|
||||
for(std::size_t i = 0; i < gemm_descs.size(); i++)
|
||||
{
|
||||
|
||||
c_device_buf[i]->FromDevice(c_m_n_device_results[i].mData.data());
|
||||
auto atol = ck::utils::get_absolute_threshold<ComputeDataType, CDataType>(
|
||||
max_abs_in_val, gemm_descs[i].K_);
|
||||
auto rtol = ck::utils::get_relative_threshold<ComputeDataType, CDataType>(
|
||||
gemm_descs[i].K_);
|
||||
|
||||
if(std::is_same_v<CDataType, ck::half_t> && kbatch_curr > 1)
|
||||
{
|
||||
instance_pass =
|
||||
instance_pass && ck::utils::check_err(c_m_n_device_results[i],
|
||||
c_m_n_host_results[i],
|
||||
"Error: Incorrect results!",
|
||||
0.06);
|
||||
}
|
||||
else
|
||||
{
|
||||
instance_pass =
|
||||
instance_pass && ck::utils::check_err(c_m_n_device_results[i],
|
||||
c_m_n_host_results[i]);
|
||||
}
|
||||
instance_pass =
|
||||
instance_pass && ck::utils::check_err(c_m_n_device_results[i],
|
||||
c_m_n_host_results[i],
|
||||
"Error: Incorrect results!",
|
||||
rtol,
|
||||
atol);
|
||||
|
||||
if(do_log)
|
||||
{
|
||||
@@ -315,7 +312,7 @@ bool profile_grouped_gemm_fixed_nk_impl(int do_verification,
|
||||
}
|
||||
}
|
||||
|
||||
std::cout << "Instance: " << gemm_name << " verification "
|
||||
std::cout << "Instance: " << gemm_name << "; KBatch: " << kbatch_curr << " "
|
||||
<< (instance_pass ? "SUCCEED" : "FAILED") << std::endl;
|
||||
|
||||
pass = pass && instance_pass;
|
||||
@@ -355,7 +352,8 @@ bool profile_grouped_gemm_fixed_nk_impl(int do_verification,
|
||||
}
|
||||
else
|
||||
{
|
||||
std::cout << "Instance: " << gemm_name << ", does not support this GEMM problem"
|
||||
std::cout << "Instance: " << gemm_name
|
||||
<< ", does not support this GEMM problem (KBatch: " << kbatch_curr << ")"
|
||||
<< std::endl;
|
||||
}
|
||||
}
|
||||
|
||||
@@ -305,4 +305,4 @@ endforeach()
|
||||
message(VERBOSE "ckProfiler libs: ${PROFILER_LIBS}")
|
||||
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE ${PROFILER_LIBS})
|
||||
|
||||
rocm_install(TARGETS ${PROFILER_EXECUTABLE} COMPONENT profiler)
|
||||
rocm_install(TARGETS ${PROFILER_EXECUTABLE} COMPONENT profiler)
|
||||
@@ -18,6 +18,12 @@ if (CK_USE_XDL OR CK_USE_WMMA)
|
||||
target_link_libraries(test_grouped_gemm_fastgelu PRIVATE utility device_grouped_gemm_fastgelu_instance)
|
||||
add_dependencies(test_grouped_gemm test_grouped_gemm_fastgelu)
|
||||
endif()
|
||||
|
||||
add_gtest_executable(test_grouped_gemm_fixed_nk test_grouped_gemm_fixed_nk.cpp)
|
||||
if(result EQUAL 0)
|
||||
target_link_libraries(test_grouped_gemm_fixed_nk PRIVATE utility device_grouped_gemm_fixed_nk_instance)
|
||||
add_dependencies(test_grouped_gemm test_grouped_gemm_fixed_nk)
|
||||
endif()
|
||||
endif()
|
||||
|
||||
add_gtest_executable(test_grouped_gemm_interface test_grouped_gemm_interface_xdl.cpp)
|
||||
|
||||
84
test/grouped_gemm/test_grouped_gemm_fixed_nk.cpp
Normal file
84
test/grouped_gemm/test_grouped_gemm_fixed_nk.cpp
Normal file
@@ -0,0 +1,84 @@
|
||||
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
|
||||
// SPDX-License-Identifier: MIT
|
||||
|
||||
#include <tuple>
|
||||
#include <vector>
|
||||
|
||||
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
|
||||
#include "ck/utility/data_type.hpp"
|
||||
#include "ck/utility/tuple.hpp"
|
||||
|
||||
#include "gtest/gtest.h"
|
||||
#include "test_grouped_gemm_util.hpp"
|
||||
|
||||
ck::index_t param_mask = 0xffffff;
|
||||
ck::index_t instance_index = -1;
|
||||
|
||||
using F16 = ck::half_t;
|
||||
using BF16 = ck::bhalf_t;
|
||||
using F8 = ck::f8_t;
|
||||
using I8 = int8_t;
|
||||
|
||||
using Row = ck::tensor_layout::gemm::RowMajor;
|
||||
using Col = ck::tensor_layout::gemm::ColumnMajor;
|
||||
|
||||
template <typename Tuple>
|
||||
class TestGroupedGemm
|
||||
: public ck::test::TestGroupedGemm<Tuple, false, ck::test::FixedNKGroupedGemmProfiler>
|
||||
{
|
||||
public:
|
||||
void SetUp() override
|
||||
{
|
||||
ck::test::TestGroupedGemm<Tuple, false, ck::test::FixedNKGroupedGemmProfiler>::SetUp();
|
||||
|
||||
#if defined(CK_USE_WMMA)
|
||||
// The old XDL tests didn't fail if instances were not supported, so we want to keep that
|
||||
// behaviour When compiling WMMA instances and WMMA is supported, then we'll fail if a
|
||||
// specific case is not supported
|
||||
this->fail_if_no_supported_instances_ =
|
||||
ck::is_gfx11_supported() || ck::is_gfx12_supported();
|
||||
#endif
|
||||
}
|
||||
};
|
||||
|
||||
// clang-format off
|
||||
using KernelTypes = ::testing::Types<
|
||||
#if CK_USE_OCP_FP8 || CK_USE_FNUZ_FP8 || defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH) || \
|
||||
defined(CK_USE_WMMA_FP8)
|
||||
ck::Tuple<Row, Row, Row, F16, F8, F16>,
|
||||
ck::Tuple<Row, Col, Row, F16, F8, F16>,
|
||||
#endif
|
||||
|
||||
ck::Tuple<Row, Row, Row, F16, F16, F16>,
|
||||
ck::Tuple<Row, Col, Row, F16, F16, F16>,
|
||||
|
||||
ck::Tuple<Row, Row, Row, BF16, BF16, BF16>,
|
||||
ck::Tuple<Row, Col, Row, BF16, BF16, BF16>,
|
||||
ck::Tuple<Row, Row, Row, BF16, I8, BF16>,
|
||||
ck::Tuple<Row, Col, Row, BF16, I8, BF16>,
|
||||
|
||||
ck::Tuple<Row, Row, Row, F16, I8, F16>,
|
||||
ck::Tuple<Row, Col, Row, F16, I8, F16>
|
||||
>;
|
||||
|
||||
// clang-format on
|
||||
|
||||
TYPED_TEST_SUITE(TestGroupedGemm, KernelTypes);
|
||||
|
||||
#include "test_grouped_gemm_fixed_nk_cases.inc"
|
||||
int main(int argc, char** argv)
|
||||
{
|
||||
testing::InitGoogleTest(&argc, argv);
|
||||
if(argc == 1) {}
|
||||
else if(argc == 3)
|
||||
{
|
||||
param_mask = strtol(argv[1], nullptr, 0);
|
||||
instance_index = atoi(argv[2]);
|
||||
}
|
||||
else
|
||||
{
|
||||
std::cout << "Usage of " << argv[0] << std::endl;
|
||||
std::cout << "Arg1,2: param_mask instance_index(-1 means all)" << std::endl;
|
||||
}
|
||||
return RUN_ALL_TESTS();
|
||||
}
|
||||
84
test/grouped_gemm/test_grouped_gemm_fixed_nk_cases.inc
Normal file
84
test/grouped_gemm/test_grouped_gemm_fixed_nk_cases.inc
Normal file
@@ -0,0 +1,84 @@
|
||||
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
|
||||
// SPDX-License-Identifier: MIT
|
||||
|
||||
#pragma once
|
||||
|
||||
TYPED_TEST(TestGroupedGemm, TinyCases)
|
||||
{
|
||||
const std::vector<int> Ms{2, 1};
|
||||
constexpr int N = 768;
|
||||
constexpr int K = 544;
|
||||
|
||||
const std::vector<int> Ns(Ms.size(), N);
|
||||
const std::vector<int> Ks(Ms.size(), K);
|
||||
|
||||
this->Run(Ms, Ns, Ks);
|
||||
}
|
||||
|
||||
TYPED_TEST(TestGroupedGemm, SmallCases)
|
||||
{
|
||||
const std::vector<int> Ms{2, 1, 3, 4, 5};
|
||||
constexpr int N = 768;
|
||||
constexpr int K = 544;
|
||||
|
||||
const std::vector<int> Ns(Ms.size(), N);
|
||||
const std::vector<int> Ks(Ms.size(), K);
|
||||
|
||||
this->Run(Ms, Ns, Ks);
|
||||
}
|
||||
|
||||
TYPED_TEST(TestGroupedGemm, MidCases)
|
||||
{
|
||||
const std::vector<int> Ms{167, 183, 177, 153, 139, 204};
|
||||
constexpr int N = 768;
|
||||
constexpr int K = 544;
|
||||
|
||||
const std::vector<int> Ns(Ms.size(), N);
|
||||
const std::vector<int> Ks(Ms.size(), K);
|
||||
|
||||
this->Run(Ms, Ns, Ks);
|
||||
}
|
||||
|
||||
TYPED_TEST(TestGroupedGemm, Regular)
|
||||
{
|
||||
const std::vector<int> Ms{64, 128, 256};
|
||||
constexpr int N = 768;
|
||||
constexpr int K = 320;
|
||||
|
||||
const std::vector<int> Ns(Ms.size(), N);
|
||||
const std::vector<int> Ks(Ms.size(), K);
|
||||
|
||||
this->Run(Ms, Ns, Ks);
|
||||
}
|
||||
|
||||
TYPED_TEST(TestGroupedGemm, MNKPadded)
|
||||
{
|
||||
const std::vector<int> Ms{127, 150, 188, 210};
|
||||
constexpr int N = 136;
|
||||
constexpr int K = 280;
|
||||
|
||||
const std::vector<int> Ns(Ms.size(), N);
|
||||
const std::vector<int> Ks(Ms.size(), K);
|
||||
|
||||
this->Run(Ms, Ns, Ks);
|
||||
}
|
||||
|
||||
TYPED_TEST(TestGroupedGemm, TestLargeKBatch)
|
||||
{
|
||||
// In some cases Split K is not supported. Running this test would fail since no instance will
|
||||
// be supported, so we skip the test
|
||||
if(!this->IsSplitKSupported())
|
||||
GTEST_SKIP() << "Split-K not supported for for the current configuration (FP16/BF16 on "
|
||||
"GFX11, or using CDE element-wise operation)";
|
||||
|
||||
const std::vector<int> Ms{188, 210};
|
||||
constexpr int N = 768;
|
||||
constexpr int K = 4096;
|
||||
|
||||
const std::vector<int> Ns(Ms.size(), N);
|
||||
const std::vector<int> Ks(Ms.size(), K);
|
||||
|
||||
this->k_batches_ = {32, 64};
|
||||
|
||||
this->Run(Ms, Ns, Ks);
|
||||
}
|
||||
@@ -16,6 +16,7 @@
|
||||
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
|
||||
#include "ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp"
|
||||
#include "profiler/profile_grouped_gemm_impl.hpp"
|
||||
#include "profiler/profile_grouped_gemm_fixed_nk_impl.hpp"
|
||||
|
||||
extern ck::index_t param_mask;
|
||||
extern ck::index_t instance_index;
|
||||
@@ -23,7 +24,124 @@ extern ck::index_t instance_index;
|
||||
namespace ck {
|
||||
namespace test {
|
||||
|
||||
template <typename Tuple, bool FailIfNoSupportedInstances = false>
|
||||
struct DefaultGroupedGemmProfiler
|
||||
{
|
||||
template <typename ADataType,
|
||||
typename BDataType,
|
||||
typename EDataType,
|
||||
typename AccDataType,
|
||||
typename ALayout,
|
||||
typename BLayout,
|
||||
typename ELayout,
|
||||
typename AElementOp,
|
||||
typename BElementOp,
|
||||
typename CDEElementOp>
|
||||
static bool Run(bool verify,
|
||||
int init_method,
|
||||
bool log,
|
||||
bool bench,
|
||||
const std::vector<int>& Ms,
|
||||
const std::vector<int>& Ns,
|
||||
const std::vector<int>& Ks,
|
||||
const std::vector<int>& StrideAs,
|
||||
const std::vector<int>& StrideBs,
|
||||
const std::vector<int>& StrideCs,
|
||||
const std::vector<int>& kbatches,
|
||||
int n_warmup,
|
||||
int n_iter,
|
||||
int instance_index,
|
||||
bool fail_if_no_supported_instances)
|
||||
{
|
||||
return ck::profiler::profile_grouped_gemm_impl<ADataType,
|
||||
BDataType,
|
||||
EDataType,
|
||||
AccDataType,
|
||||
ALayout,
|
||||
BLayout,
|
||||
ELayout,
|
||||
AElementOp,
|
||||
BElementOp,
|
||||
CDEElementOp>(
|
||||
verify,
|
||||
init_method,
|
||||
log,
|
||||
bench,
|
||||
Ms,
|
||||
Ns,
|
||||
Ks,
|
||||
StrideAs,
|
||||
StrideBs,
|
||||
StrideCs,
|
||||
kbatches,
|
||||
n_warmup,
|
||||
n_iter,
|
||||
instance_index,
|
||||
fail_if_no_supported_instances);
|
||||
}
|
||||
};
|
||||
|
||||
struct FixedNKGroupedGemmProfiler
|
||||
{
|
||||
template <typename ADataType,
|
||||
typename BDataType,
|
||||
typename EDataType,
|
||||
typename AccDataType,
|
||||
typename ALayout,
|
||||
typename BLayout,
|
||||
typename CLayout>
|
||||
static bool Run(bool verify,
|
||||
int init_method,
|
||||
bool log,
|
||||
bool bench,
|
||||
const std::vector<int>& Ms,
|
||||
const std::vector<int>& Ns,
|
||||
const std::vector<int>& Ks,
|
||||
const std::vector<int>& StrideAs,
|
||||
const std::vector<int>& StrideBs,
|
||||
const std::vector<int>& StrideCs,
|
||||
const std::vector<int>& kbatches,
|
||||
int n_warmup,
|
||||
int n_iter,
|
||||
int /*instance_index*/,
|
||||
bool /*fail_if_no_supported_instances*/)
|
||||
{
|
||||
bool pass = true;
|
||||
for(int kbatch : kbatches)
|
||||
{
|
||||
try
|
||||
{
|
||||
pass &= ck::profiler::profile_grouped_gemm_fixed_nk_impl<ADataType,
|
||||
BDataType,
|
||||
EDataType,
|
||||
AccDataType,
|
||||
ALayout,
|
||||
BLayout,
|
||||
CLayout>(verify,
|
||||
init_method,
|
||||
log,
|
||||
bench,
|
||||
Ms,
|
||||
Ns,
|
||||
Ks,
|
||||
StrideAs,
|
||||
StrideBs,
|
||||
StrideCs,
|
||||
kbatch,
|
||||
n_warmup,
|
||||
n_iter);
|
||||
}
|
||||
catch(const std::exception& e)
|
||||
{
|
||||
std::cerr << e.what() << std::endl;
|
||||
}
|
||||
}
|
||||
return pass;
|
||||
}
|
||||
};
|
||||
|
||||
template <typename Tuple,
|
||||
bool FailIfNoSupportedInstances = false,
|
||||
typename Profiler = ck::test::DefaultGroupedGemmProfiler>
|
||||
class TestGroupedGemm : public testing::Test
|
||||
{
|
||||
protected:
|
||||
@@ -76,7 +194,7 @@ class TestGroupedGemm : public testing::Test
|
||||
}
|
||||
else
|
||||
{
|
||||
k_batches_ = {1, 2, 3, 5, 8};
|
||||
k_batches_ = {1, 2, 3, 4, 8};
|
||||
}
|
||||
}
|
||||
|
||||
@@ -146,31 +264,61 @@ class TestGroupedGemm : public testing::Test
|
||||
const std::vector<int>& StrideCs,
|
||||
const std::vector<int>& kbatches)
|
||||
{
|
||||
bool pass =
|
||||
ck::profiler::profile_grouped_gemm_impl<ADataType,
|
||||
BDataType,
|
||||
EDataType,
|
||||
float,
|
||||
ALayout,
|
||||
BLayout,
|
||||
ELayout,
|
||||
AElementOp,
|
||||
BElementOp,
|
||||
CDEElementOp>(verify_,
|
||||
init_method_,
|
||||
log_,
|
||||
bench_,
|
||||
Ms,
|
||||
Ns,
|
||||
Ks,
|
||||
StrideAs,
|
||||
StrideBs,
|
||||
StrideCs,
|
||||
kbatches,
|
||||
n_warmup_,
|
||||
n_iter_,
|
||||
instance_index,
|
||||
fail_if_no_supported_instances_);
|
||||
bool pass = false;
|
||||
using AccDataType = float;
|
||||
|
||||
if constexpr(std::is_same_v<Profiler, FixedNKGroupedGemmProfiler>)
|
||||
{
|
||||
pass = Profiler::template Run<ADataType,
|
||||
BDataType,
|
||||
EDataType,
|
||||
AccDataType,
|
||||
ALayout,
|
||||
BLayout,
|
||||
ELayout>(verify_,
|
||||
init_method_,
|
||||
log_,
|
||||
bench_,
|
||||
Ms,
|
||||
Ns,
|
||||
Ks,
|
||||
StrideAs,
|
||||
StrideBs,
|
||||
StrideCs,
|
||||
kbatches,
|
||||
n_warmup_,
|
||||
n_iter_,
|
||||
instance_index,
|
||||
fail_if_no_supported_instances_);
|
||||
}
|
||||
else
|
||||
{
|
||||
pass = Profiler::template Run<ADataType,
|
||||
BDataType,
|
||||
EDataType,
|
||||
AccDataType,
|
||||
ALayout,
|
||||
BLayout,
|
||||
ELayout,
|
||||
AElementOp,
|
||||
BElementOp,
|
||||
CDEElementOp>(verify_,
|
||||
init_method_,
|
||||
log_,
|
||||
bench_,
|
||||
Ms,
|
||||
Ns,
|
||||
Ks,
|
||||
StrideAs,
|
||||
StrideBs,
|
||||
StrideCs,
|
||||
kbatches,
|
||||
n_warmup_,
|
||||
n_iter_,
|
||||
instance_index,
|
||||
fail_if_no_supported_instances_);
|
||||
}
|
||||
|
||||
EXPECT_TRUE(pass);
|
||||
}
|
||||
};
|
||||
|
||||
Reference in New Issue
Block a user