mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-14 02:02:46 +00:00
Add fp16/fp8 support into Grouped gemm FixedNK (#874)
* move all arguments into device
* add b2c_tile_map
* add examples
* add SetDeviceKernelArgs
* dedicated fixed_nk solution
* init client api
* add grouped_gemm_bias example
* add a instance
* add instances
* formatting
* fixed cmake
* Update EnableCompilerWarnings.cmake
* Update cmake-ck-dev.sh
* clean; fixed comments
* fixed comment
* add instances for fp32 output
* add instances for fp32 output
* add fp32 out client example
* fixed CI
* init commit for kbatch
* add splitk gridwise
* format
* fixed
* clean deviceop
* clean code
* finish splitk
* fixed instances
* change m_loops to tile_loops
* add setkbatch
* clean code
* add splitK+bias
* add instances
* opt mk_nk instances
* clean examples
* fixed CI
* remove zero
* finished non-zero
* clean
* clean code
* optimized global_barrier
* fixed ci
* fixed CI
* instance and client
* removed AddBias
* format
* fixed CI
* fixed CI
* move 20_grouped_gemm to 21_grouped_gemm
* clean
* formatting
* clean
* clean
* fixed computeType
---------
Co-authored-by: Jing Zhang <jizha@amd.com>
[ROCm/composable_kernel commit: f9d0eddb90]
This commit is contained in:
zjing14
8
client_example/22_grouped_gemm/CMakeLists.txt
Normal file
8
client_example/22_grouped_gemm/CMakeLists.txt
Normal file
@@ -0,0 +1,8 @@
|
||||
add_executable(client_grouped_gemm_fixed_nk_fp16 grouped_gemm_fixed_nk_fp16.cpp)
|
||||
target_link_libraries(client_grouped_gemm_fixed_nk_fp16 PRIVATE composable_kernel::device_operations)
|
||||
|
||||
add_executable(client_grouped_gemm_fixed_nk_fp8 grouped_gemm_fixed_nk_fp8.cpp)
|
||||
target_link_libraries(client_grouped_gemm_fixed_nk_fp8 PRIVATE composable_kernel::device_operations)
|
||||
|
||||
add_executable(client_grouped_gemm_fixed_nk_i8 grouped_gemm_fixed_nk_i8.cpp)
|
||||
target_link_libraries(client_grouped_gemm_fixed_nk_i8 PRIVATE composable_kernel::device_operations)
|
||||
238
client_example/22_grouped_gemm/grouped_gemm_fixed_nk_fp16.cpp
Normal file
238
client_example/22_grouped_gemm/grouped_gemm_fixed_nk_fp16.cpp
Normal file
@@ -0,0 +1,238 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#include <iomanip>
|
||||
#include <iostream>
|
||||
#include <vector>
|
||||
#include <random>
|
||||
|
||||
#include "ck/ck.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/device_grouped_gemm_fixed_nk.hpp"
|
||||
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
|
||||
|
||||
#include "ck/library/tensor_operation_instance/gpu/grouped_gemm_fixed_nk.hpp"
|
||||
|
||||
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 DsDataType = ck::Tuple<>;
|
||||
using EDataType = F16;
|
||||
|
||||
using ALayout = Row;
|
||||
using BLayout = Row;
|
||||
using DsLayout = ck::Tuple<>;
|
||||
using ELayout = Row;
|
||||
|
||||
using AElementOp = PassThrough;
|
||||
using BElementOp = PassThrough;
|
||||
using CDEElementOp = PassThrough;
|
||||
|
||||
struct SimpleDeviceMem
|
||||
{
|
||||
SimpleDeviceMem() = delete;
|
||||
|
||||
SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
|
||||
{
|
||||
(void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
|
||||
}
|
||||
|
||||
void* GetDeviceBuffer() { return p_mem_; }
|
||||
|
||||
~SimpleDeviceMem() { (void)hipFree(p_mem_); }
|
||||
|
||||
void* p_mem_;
|
||||
};
|
||||
|
||||
int main()
|
||||
{
|
||||
std::vector<int> Ms, Ns, Ks, StrideAs, StrideBs, StrideEs;
|
||||
|
||||
int sum_of_m = 0;
|
||||
|
||||
// Ms = {167, 183, 177, 181, 153, 139, 156, 173, 163, 150, 204, 184, 168, 156, 168, 148};
|
||||
Ms = {0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0};
|
||||
|
||||
int group_count = Ms.size();
|
||||
|
||||
for(int i = 0; i < group_count; ++i)
|
||||
{
|
||||
Ns.push_back(768);
|
||||
Ks.push_back(4608);
|
||||
|
||||
StrideAs.push_back(std::is_same<Row, ALayout>::value ? Ks[i] : Ms[i]);
|
||||
StrideBs.push_back(std::is_same<Row, BLayout>::value ? Ns[i] : Ks[i]);
|
||||
StrideEs.push_back(std::is_same<Row, ELayout>::value ? Ns[i] : Ms[i]);
|
||||
|
||||
sum_of_m += Ms[i];
|
||||
}
|
||||
|
||||
auto f_matrix_space_size =
|
||||
[](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
|
||||
using Layout = decltype(layout);
|
||||
|
||||
if constexpr(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
|
||||
{
|
||||
return (nRow - 1) * stride + nCol;
|
||||
}
|
||||
else
|
||||
{
|
||||
return (nCol - 1) * stride + nRow;
|
||||
}
|
||||
};
|
||||
|
||||
std::vector<SimpleDeviceMem> a_dev_bufs, b_dev_bufs, e_dev_bufs;
|
||||
|
||||
a_dev_bufs.reserve(group_count);
|
||||
b_dev_bufs.reserve(group_count);
|
||||
e_dev_bufs.reserve(group_count);
|
||||
|
||||
std::vector<void*> p_e;
|
||||
|
||||
p_e.reserve(group_count);
|
||||
|
||||
std::vector<ck::tensor_operation::device::GemmDesc> gemm_descs;
|
||||
|
||||
gemm_descs.reserve(group_count);
|
||||
|
||||
std::vector<ck::tensor_operation::device::GroupedGemmKernelArgument<1>>
|
||||
grouped_gemm_kernel_args_;
|
||||
grouped_gemm_kernel_args_.reserve(group_count);
|
||||
|
||||
for(int i = 0; i < group_count; ++i)
|
||||
{
|
||||
a_dev_bufs.emplace_back(sizeof(ADataType) *
|
||||
f_matrix_space_size(Ms[i], Ks[i], StrideAs[i], ALayout{}));
|
||||
b_dev_bufs.emplace_back(sizeof(BDataType) *
|
||||
f_matrix_space_size(Ks[i], Ns[i], StrideBs[i], BLayout{}));
|
||||
e_dev_bufs.emplace_back(sizeof(EDataType) *
|
||||
f_matrix_space_size(Ms[i], Ns[i], StrideEs[i], ELayout{}));
|
||||
|
||||
gemm_descs.push_back({sum_of_m, Ns[i], Ks[i], 1, StrideBs[i], 1, {0}});
|
||||
|
||||
p_e.push_back(e_dev_bufs[i].GetDeviceBuffer());
|
||||
|
||||
grouped_gemm_kernel_args_.push_back({a_dev_bufs[i].GetDeviceBuffer(),
|
||||
b_dev_bufs[i].GetDeviceBuffer(),
|
||||
{},
|
||||
e_dev_bufs[i].GetDeviceBuffer(),
|
||||
Ms[i],
|
||||
Ns[i],
|
||||
Ks[i],
|
||||
StrideAs[i],
|
||||
StrideBs[i],
|
||||
{},
|
||||
StrideEs[i]});
|
||||
}
|
||||
|
||||
using DeviceOp = ck::tensor_operation::device::DeviceGroupedGemmFixedNK<ALayout,
|
||||
BLayout,
|
||||
DsLayout,
|
||||
ELayout,
|
||||
ADataType,
|
||||
BDataType,
|
||||
DsDataType,
|
||||
EDataType,
|
||||
AElementOp,
|
||||
BElementOp,
|
||||
CDEElementOp>;
|
||||
|
||||
// get device op instances
|
||||
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
|
||||
DeviceOp>::GetInstances();
|
||||
|
||||
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
|
||||
|
||||
const auto a_element_op = AElementOp{};
|
||||
const auto b_element_op = BElementOp{};
|
||||
const auto cde_element_op = CDEElementOp{};
|
||||
|
||||
std::string best_op_name;
|
||||
bool found = false;
|
||||
int best_op_id = -1;
|
||||
float best_ave_time = 0;
|
||||
float best_tflops = 0;
|
||||
float best_gb_per_sec = 0;
|
||||
|
||||
// profile device operation instances
|
||||
std::cout << "Run all instances and do timing" << std::endl;
|
||||
|
||||
std::vector<const void*> p_a = {}, p_b = {};
|
||||
std::vector<std::array<const void*, 0>> p_ds = {};
|
||||
|
||||
for(int i = 0; i < op_ptrs.size(); ++i)
|
||||
{
|
||||
auto& op_ptr = op_ptrs[i];
|
||||
|
||||
auto argument_ptr = op_ptr->MakeArgumentPointer(
|
||||
p_a, p_b, p_ds, p_e, gemm_descs, a_element_op, b_element_op, cde_element_op);
|
||||
|
||||
auto invoker_ptr = op_ptr->MakeInvokerPointer();
|
||||
|
||||
SimpleDeviceMem grouped_gemm_kernel_args_dev(
|
||||
op_ptr->GetDeviceKernelArgSize(argument_ptr.get()));
|
||||
|
||||
SimpleDeviceMem grouped_gemm_workspace_dev(op_ptr->GetWorkSpaceSize(argument_ptr.get()));
|
||||
|
||||
std::string op_name = op_ptr->GetTypeString();
|
||||
|
||||
hipGetErrorString(hipMemcpy(grouped_gemm_kernel_args_dev.GetDeviceBuffer(),
|
||||
grouped_gemm_kernel_args_.data(),
|
||||
op_ptr->GetDeviceKernelArgSize(argument_ptr.get()),
|
||||
hipMemcpyHostToDevice));
|
||||
|
||||
op_ptr->SetWorkSpacePointer(argument_ptr.get(),
|
||||
grouped_gemm_workspace_dev.GetDeviceBuffer());
|
||||
|
||||
op_ptr->SetDeviceKernelArgs(argument_ptr.get(),
|
||||
grouped_gemm_kernel_args_dev.GetDeviceBuffer());
|
||||
|
||||
op_ptr->SetKBatch(argument_ptr.get(), 32);
|
||||
|
||||
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
|
||||
{
|
||||
float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
|
||||
|
||||
std::size_t flop = 0, num_btype = 0;
|
||||
for(std::size_t j = 0; j < gemm_descs.size(); ++j)
|
||||
{
|
||||
flop += std::size_t(2) * Ms[j] * Ns[j] * Ks[j];
|
||||
|
||||
num_btype += sizeof(ADataType) * Ms[j] * Ks[j] + sizeof(BDataType) * Ks[j] * Ns[j] +
|
||||
sizeof(EDataType) * Ms[j] * Ns[j];
|
||||
}
|
||||
|
||||
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
|
||||
float gb_per_sec = num_btype / 1.E6 / ave_time;
|
||||
|
||||
std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
|
||||
<< gb_per_sec << " GB/s, " << op_name << std::endl;
|
||||
|
||||
if(tflops > best_tflops)
|
||||
{
|
||||
found = true;
|
||||
best_op_id = i;
|
||||
best_op_name = op_name;
|
||||
best_tflops = tflops;
|
||||
best_ave_time = ave_time;
|
||||
best_gb_per_sec = gb_per_sec;
|
||||
}
|
||||
}
|
||||
else
|
||||
{
|
||||
std::cout << op_name << " does not support this problem" << std::endl;
|
||||
}
|
||||
}
|
||||
|
||||
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
|
||||
<< best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
|
||||
|
||||
return 0;
|
||||
}
|
||||
238
client_example/22_grouped_gemm/grouped_gemm_fixed_nk_fp8.cpp
Normal file
238
client_example/22_grouped_gemm/grouped_gemm_fixed_nk_fp8.cpp
Normal file
@@ -0,0 +1,238 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#include <iomanip>
|
||||
#include <iostream>
|
||||
#include <vector>
|
||||
#include <random>
|
||||
|
||||
#include "ck/ck.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/device_grouped_gemm_fixed_nk.hpp"
|
||||
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
|
||||
|
||||
#include "ck/library/tensor_operation_instance/gpu/grouped_gemm_fixed_nk.hpp"
|
||||
|
||||
using F8 = ck::f8_t;
|
||||
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 = F8;
|
||||
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;
|
||||
|
||||
struct SimpleDeviceMem
|
||||
{
|
||||
SimpleDeviceMem() = delete;
|
||||
|
||||
SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
|
||||
{
|
||||
(void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
|
||||
}
|
||||
|
||||
void* GetDeviceBuffer() { return p_mem_; }
|
||||
|
||||
~SimpleDeviceMem() { (void)hipFree(p_mem_); }
|
||||
|
||||
void* p_mem_;
|
||||
};
|
||||
|
||||
int main()
|
||||
{
|
||||
std::vector<int> Ms, Ns, Ks, StrideAs, StrideBs, StrideEs;
|
||||
|
||||
int sum_of_m = 0;
|
||||
|
||||
Ms = {167, 183, 177, 181, 153, 139, 156, 173, 163, 150, 204, 184, 168, 156, 168, 148};
|
||||
|
||||
int group_count = Ms.size();
|
||||
|
||||
for(int i = 0; i < group_count; ++i)
|
||||
{
|
||||
Ns.push_back(768);
|
||||
Ks.push_back(4608);
|
||||
|
||||
StrideAs.push_back(std::is_same<Row, ALayout>::value ? Ks[i] : Ms[i]);
|
||||
StrideBs.push_back(std::is_same<Row, BLayout>::value ? Ns[i] : Ks[i]);
|
||||
StrideEs.push_back(std::is_same<Row, ELayout>::value ? Ns[i] : Ms[i]);
|
||||
|
||||
sum_of_m += Ms[i];
|
||||
}
|
||||
|
||||
auto f_matrix_space_size =
|
||||
[](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
|
||||
using Layout = decltype(layout);
|
||||
|
||||
if constexpr(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
|
||||
{
|
||||
return (nRow - 1) * stride + nCol;
|
||||
}
|
||||
else
|
||||
{
|
||||
return (nCol - 1) * stride + nRow;
|
||||
}
|
||||
};
|
||||
|
||||
std::vector<SimpleDeviceMem> a_dev_bufs, b_dev_bufs, e_dev_bufs;
|
||||
|
||||
a_dev_bufs.reserve(group_count);
|
||||
b_dev_bufs.reserve(group_count);
|
||||
e_dev_bufs.reserve(group_count);
|
||||
|
||||
std::vector<void*> p_e;
|
||||
|
||||
p_e.reserve(group_count);
|
||||
|
||||
std::vector<ck::tensor_operation::device::GemmDesc> gemm_descs;
|
||||
|
||||
gemm_descs.reserve(group_count);
|
||||
|
||||
std::vector<ck::tensor_operation::device::GroupedGemmKernelArgument<1>>
|
||||
grouped_gemm_kernel_args_;
|
||||
grouped_gemm_kernel_args_.reserve(group_count);
|
||||
|
||||
for(int i = 0; i < group_count; ++i)
|
||||
{
|
||||
a_dev_bufs.emplace_back(sizeof(ADataType) *
|
||||
f_matrix_space_size(Ms[i], Ks[i], StrideAs[i], ALayout{}));
|
||||
b_dev_bufs.emplace_back(sizeof(BDataType) *
|
||||
f_matrix_space_size(Ks[i], Ns[i], StrideBs[i], BLayout{}));
|
||||
e_dev_bufs.emplace_back(sizeof(EDataType) *
|
||||
f_matrix_space_size(Ms[i], Ns[i], StrideEs[i], ELayout{}));
|
||||
|
||||
gemm_descs.push_back({sum_of_m, Ns[i], Ks[i], 1, StrideBs[i], 1, {0}});
|
||||
|
||||
p_e.push_back(e_dev_bufs[i].GetDeviceBuffer());
|
||||
|
||||
grouped_gemm_kernel_args_.push_back({a_dev_bufs[i].GetDeviceBuffer(),
|
||||
b_dev_bufs[i].GetDeviceBuffer(),
|
||||
{},
|
||||
e_dev_bufs[i].GetDeviceBuffer(),
|
||||
Ms[i],
|
||||
Ns[i],
|
||||
Ks[i],
|
||||
StrideAs[i],
|
||||
StrideBs[i],
|
||||
{},
|
||||
StrideEs[i]});
|
||||
}
|
||||
|
||||
using DeviceOp = ck::tensor_operation::device::DeviceGroupedGemmFixedNK<ALayout,
|
||||
BLayout,
|
||||
DsLayout,
|
||||
ELayout,
|
||||
ADataType,
|
||||
BDataType,
|
||||
DsDataType,
|
||||
EDataType,
|
||||
AElementOp,
|
||||
BElementOp,
|
||||
CDEElementOp>;
|
||||
|
||||
// get device op instances
|
||||
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
|
||||
DeviceOp>::GetInstances();
|
||||
|
||||
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
|
||||
|
||||
const auto a_element_op = AElementOp{};
|
||||
const auto b_element_op = BElementOp{};
|
||||
const auto cde_element_op = CDEElementOp{};
|
||||
|
||||
std::string best_op_name;
|
||||
bool found = false;
|
||||
int best_op_id = -1;
|
||||
float best_ave_time = 0;
|
||||
float best_tflops = 0;
|
||||
float best_gb_per_sec = 0;
|
||||
|
||||
// profile device operation instances
|
||||
std::cout << "Run all instances and do timing" << std::endl;
|
||||
|
||||
std::vector<const void*> p_a = {}, p_b = {};
|
||||
std::vector<std::array<const void*, 0>> p_ds = {};
|
||||
|
||||
for(int i = 0; i < op_ptrs.size(); ++i)
|
||||
{
|
||||
auto& op_ptr = op_ptrs[i];
|
||||
|
||||
auto argument_ptr = op_ptr->MakeArgumentPointer(
|
||||
p_a, p_b, p_ds, p_e, gemm_descs, a_element_op, b_element_op, cde_element_op);
|
||||
|
||||
auto invoker_ptr = op_ptr->MakeInvokerPointer();
|
||||
|
||||
SimpleDeviceMem grouped_gemm_kernel_args_dev(
|
||||
op_ptr->GetDeviceKernelArgSize(argument_ptr.get()));
|
||||
|
||||
SimpleDeviceMem grouped_gemm_workspace_dev(op_ptr->GetWorkSpaceSize(argument_ptr.get()));
|
||||
|
||||
std::string op_name = op_ptr->GetTypeString();
|
||||
|
||||
hipGetErrorString(hipMemcpy(grouped_gemm_kernel_args_dev.GetDeviceBuffer(),
|
||||
grouped_gemm_kernel_args_.data(),
|
||||
op_ptr->GetDeviceKernelArgSize(argument_ptr.get()),
|
||||
hipMemcpyHostToDevice));
|
||||
|
||||
op_ptr->SetWorkSpacePointer(argument_ptr.get(),
|
||||
grouped_gemm_workspace_dev.GetDeviceBuffer());
|
||||
|
||||
op_ptr->SetDeviceKernelArgs(argument_ptr.get(),
|
||||
grouped_gemm_kernel_args_dev.GetDeviceBuffer());
|
||||
|
||||
op_ptr->SetKBatch(argument_ptr.get(), 16);
|
||||
|
||||
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
|
||||
{
|
||||
float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
|
||||
|
||||
std::size_t flop = 0, num_btype = 0;
|
||||
for(std::size_t j = 0; j < gemm_descs.size(); ++j)
|
||||
{
|
||||
flop += std::size_t(2) * Ms[j] * Ns[j] * Ks[j];
|
||||
|
||||
num_btype += sizeof(ADataType) * Ms[j] * Ks[j] + sizeof(BDataType) * Ks[j] * Ns[j] +
|
||||
sizeof(EDataType) * Ms[j] * Ns[j];
|
||||
}
|
||||
|
||||
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
|
||||
float gb_per_sec = num_btype / 1.E6 / ave_time;
|
||||
|
||||
std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
|
||||
<< gb_per_sec << " GB/s, " << op_name << std::endl;
|
||||
|
||||
if(tflops > best_tflops)
|
||||
{
|
||||
found = true;
|
||||
best_op_id = i;
|
||||
best_op_name = op_name;
|
||||
best_tflops = tflops;
|
||||
best_ave_time = ave_time;
|
||||
best_gb_per_sec = gb_per_sec;
|
||||
}
|
||||
}
|
||||
else
|
||||
{
|
||||
std::cout << op_name << " does not support this problem" << std::endl;
|
||||
}
|
||||
}
|
||||
|
||||
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
|
||||
<< best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
|
||||
|
||||
return 0;
|
||||
}
|
||||
238
client_example/22_grouped_gemm/grouped_gemm_fixed_nk_i8.cpp
Normal file
238
client_example/22_grouped_gemm/grouped_gemm_fixed_nk_i8.cpp
Normal file
@@ -0,0 +1,238 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#include <iomanip>
|
||||
#include <iostream>
|
||||
#include <vector>
|
||||
#include <random>
|
||||
|
||||
#include "ck/ck.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/device_grouped_gemm_fixed_nk.hpp"
|
||||
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
|
||||
|
||||
#include "ck/library/tensor_operation_instance/gpu/grouped_gemm_fixed_nk.hpp"
|
||||
|
||||
using I8 = int8_t;
|
||||
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 = I8;
|
||||
using DsDataType = ck::Tuple<>;
|
||||
using EDataType = F16;
|
||||
|
||||
using ALayout = Row;
|
||||
using BLayout = Row;
|
||||
using DsLayout = ck::Tuple<>;
|
||||
using ELayout = Row;
|
||||
|
||||
using AElementOp = PassThrough;
|
||||
using BElementOp = PassThrough;
|
||||
using CDEElementOp = PassThrough;
|
||||
|
||||
struct SimpleDeviceMem
|
||||
{
|
||||
SimpleDeviceMem() = delete;
|
||||
|
||||
SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
|
||||
{
|
||||
(void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
|
||||
}
|
||||
|
||||
void* GetDeviceBuffer() { return p_mem_; }
|
||||
|
||||
~SimpleDeviceMem() { (void)hipFree(p_mem_); }
|
||||
|
||||
void* p_mem_;
|
||||
};
|
||||
|
||||
int main()
|
||||
{
|
||||
std::vector<int> Ms, Ns, Ks, StrideAs, StrideBs, StrideEs;
|
||||
|
||||
int sum_of_m = 0;
|
||||
|
||||
Ms = {167, 183, 177, 181, 153, 139, 156, 173, 163, 150, 204, 184, 168, 156, 168, 148};
|
||||
|
||||
int group_count = Ms.size();
|
||||
|
||||
for(int i = 0; i < group_count; ++i)
|
||||
{
|
||||
Ns.push_back(768);
|
||||
Ks.push_back(4608);
|
||||
|
||||
StrideAs.push_back(std::is_same<Row, ALayout>::value ? Ks[i] : Ms[i]);
|
||||
StrideBs.push_back(std::is_same<Row, BLayout>::value ? Ns[i] : Ks[i]);
|
||||
StrideEs.push_back(std::is_same<Row, ELayout>::value ? Ns[i] : Ms[i]);
|
||||
|
||||
sum_of_m += Ms[i];
|
||||
}
|
||||
|
||||
auto f_matrix_space_size =
|
||||
[](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
|
||||
using Layout = decltype(layout);
|
||||
|
||||
if constexpr(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
|
||||
{
|
||||
return (nRow - 1) * stride + nCol;
|
||||
}
|
||||
else
|
||||
{
|
||||
return (nCol - 1) * stride + nRow;
|
||||
}
|
||||
};
|
||||
|
||||
std::vector<SimpleDeviceMem> a_dev_bufs, b_dev_bufs, e_dev_bufs;
|
||||
|
||||
a_dev_bufs.reserve(group_count);
|
||||
b_dev_bufs.reserve(group_count);
|
||||
e_dev_bufs.reserve(group_count);
|
||||
|
||||
std::vector<void*> p_e;
|
||||
|
||||
p_e.reserve(group_count);
|
||||
|
||||
std::vector<ck::tensor_operation::device::GemmDesc> gemm_descs;
|
||||
|
||||
gemm_descs.reserve(group_count);
|
||||
|
||||
std::vector<ck::tensor_operation::device::GroupedGemmKernelArgument<1>>
|
||||
grouped_gemm_kernel_args_;
|
||||
grouped_gemm_kernel_args_.reserve(group_count);
|
||||
|
||||
for(int i = 0; i < group_count; ++i)
|
||||
{
|
||||
a_dev_bufs.emplace_back(sizeof(ADataType) *
|
||||
f_matrix_space_size(Ms[i], Ks[i], StrideAs[i], ALayout{}));
|
||||
b_dev_bufs.emplace_back(sizeof(BDataType) *
|
||||
f_matrix_space_size(Ks[i], Ns[i], StrideBs[i], BLayout{}));
|
||||
e_dev_bufs.emplace_back(sizeof(EDataType) *
|
||||
f_matrix_space_size(Ms[i], Ns[i], StrideEs[i], ELayout{}));
|
||||
|
||||
gemm_descs.push_back({sum_of_m, Ns[i], Ks[i], 1, StrideBs[i], 1, {0}});
|
||||
|
||||
p_e.push_back(e_dev_bufs[i].GetDeviceBuffer());
|
||||
|
||||
grouped_gemm_kernel_args_.push_back({a_dev_bufs[i].GetDeviceBuffer(),
|
||||
b_dev_bufs[i].GetDeviceBuffer(),
|
||||
{},
|
||||
e_dev_bufs[i].GetDeviceBuffer(),
|
||||
Ms[i],
|
||||
Ns[i],
|
||||
Ks[i],
|
||||
StrideAs[i],
|
||||
StrideBs[i],
|
||||
{},
|
||||
StrideEs[i]});
|
||||
}
|
||||
|
||||
using DeviceOp = ck::tensor_operation::device::DeviceGroupedGemmFixedNK<ALayout,
|
||||
BLayout,
|
||||
DsLayout,
|
||||
ELayout,
|
||||
ADataType,
|
||||
BDataType,
|
||||
DsDataType,
|
||||
EDataType,
|
||||
AElementOp,
|
||||
BElementOp,
|
||||
CDEElementOp>;
|
||||
|
||||
// get device op instances
|
||||
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
|
||||
DeviceOp>::GetInstances();
|
||||
|
||||
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
|
||||
|
||||
const auto a_element_op = AElementOp{};
|
||||
const auto b_element_op = BElementOp{};
|
||||
const auto cde_element_op = CDEElementOp{};
|
||||
|
||||
std::string best_op_name;
|
||||
bool found = false;
|
||||
int best_op_id = -1;
|
||||
float best_ave_time = 0;
|
||||
float best_tflops = 0;
|
||||
float best_gb_per_sec = 0;
|
||||
|
||||
// profile device operation instances
|
||||
std::cout << "Run all instances and do timing" << std::endl;
|
||||
|
||||
std::vector<const void*> p_a = {}, p_b = {};
|
||||
std::vector<std::array<const void*, 0>> p_ds = {};
|
||||
|
||||
for(int i = 0; i < op_ptrs.size(); ++i)
|
||||
{
|
||||
auto& op_ptr = op_ptrs[i];
|
||||
|
||||
auto argument_ptr = op_ptr->MakeArgumentPointer(
|
||||
p_a, p_b, p_ds, p_e, gemm_descs, a_element_op, b_element_op, cde_element_op);
|
||||
|
||||
auto invoker_ptr = op_ptr->MakeInvokerPointer();
|
||||
|
||||
SimpleDeviceMem grouped_gemm_kernel_args_dev(
|
||||
op_ptr->GetDeviceKernelArgSize(argument_ptr.get()));
|
||||
|
||||
SimpleDeviceMem grouped_gemm_workspace_dev(op_ptr->GetWorkSpaceSize(argument_ptr.get()));
|
||||
|
||||
std::string op_name = op_ptr->GetTypeString();
|
||||
|
||||
hipGetErrorString(hipMemcpy(grouped_gemm_kernel_args_dev.GetDeviceBuffer(),
|
||||
grouped_gemm_kernel_args_.data(),
|
||||
op_ptr->GetDeviceKernelArgSize(argument_ptr.get()),
|
||||
hipMemcpyHostToDevice));
|
||||
|
||||
op_ptr->SetWorkSpacePointer(argument_ptr.get(),
|
||||
grouped_gemm_workspace_dev.GetDeviceBuffer());
|
||||
|
||||
op_ptr->SetDeviceKernelArgs(argument_ptr.get(),
|
||||
grouped_gemm_kernel_args_dev.GetDeviceBuffer());
|
||||
|
||||
op_ptr->SetKBatch(argument_ptr.get(), 32);
|
||||
|
||||
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
|
||||
{
|
||||
float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
|
||||
|
||||
std::size_t flop = 0, num_btype = 0;
|
||||
for(std::size_t j = 0; j < gemm_descs.size(); ++j)
|
||||
{
|
||||
flop += std::size_t(2) * Ms[j] * Ns[j] * Ks[j];
|
||||
|
||||
num_btype += sizeof(ADataType) * Ms[j] * Ks[j] + sizeof(BDataType) * Ks[j] * Ns[j] +
|
||||
sizeof(EDataType) * Ms[j] * Ns[j];
|
||||
}
|
||||
|
||||
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
|
||||
float gb_per_sec = num_btype / 1.E6 / ave_time;
|
||||
|
||||
std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
|
||||
<< gb_per_sec << " GB/s, " << op_name << std::endl;
|
||||
|
||||
if(tflops > best_tflops)
|
||||
{
|
||||
found = true;
|
||||
best_op_id = i;
|
||||
best_op_name = op_name;
|
||||
best_tflops = tflops;
|
||||
best_ave_time = ave_time;
|
||||
best_gb_per_sec = gb_per_sec;
|
||||
}
|
||||
}
|
||||
else
|
||||
{
|
||||
std::cout << op_name << " does not support this problem" << std::endl;
|
||||
}
|
||||
}
|
||||
|
||||
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
|
||||
<< best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
|
||||
|
||||
return 0;
|
||||
}
|
||||
@@ -25,6 +25,11 @@ if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
|
||||
add_example_executable(example_grouped_gemm_xdl_int8 grouped_gemm_xdl_int8.cpp)
|
||||
add_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_int8)
|
||||
endif()
|
||||
if(DTYPES MATCHES "f8" OR NOT DEFINED DTYPES)
|
||||
add_example_executable(example_grouped_gemm_xdl_fixed_nk_fp8 grouped_gemm_xdl_fixed_nk_fp8.cpp)
|
||||
add_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_fixed_nk_fp8)
|
||||
endif()
|
||||
|
||||
if(USE_BITINT_EXTENSION_INT4)
|
||||
add_example_executable(example_grouped_gemm_xdl_int4 grouped_gemm_xdl_int4.cpp)
|
||||
add_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_int4)
|
||||
|
||||
330
example/15_grouped_gemm/grouped_gemm_xdl_fixed_nk_fp8.cpp
Normal file
330
example/15_grouped_gemm/grouped_gemm_xdl_fixed_nk_fp8.cpp
Normal file
@@ -0,0 +1,330 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#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_xdl_fixed_nk.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/device_grouped_gemm.hpp"
|
||||
#include "ck/tensor_operation/gpu/element/element_wise_operation.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"
|
||||
|
||||
template <ck::index_t... Is>
|
||||
using S = ck::Sequence<Is...>;
|
||||
|
||||
using F8 = ck::f8_t;
|
||||
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 = F8;
|
||||
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::MNPadding;
|
||||
|
||||
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGroupedGemm_Xdl_Fixed_NK
|
||||
// clang-format off
|
||||
//######| ALayout| BLayout| DsLayout| ELayout| 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|
|
||||
//######| | | | | 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|
|
||||
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
< ALayout, BLayout, DsLayout, ELayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementOp, BElementOp, CDEElementOp, GemmDefault, 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, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 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<0>{});
|
||||
b_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<1>{});
|
||||
}
|
||||
}
|
||||
|
||||
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) * sum_of_m * 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) * sum_of_m * 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]);
|
||||
}
|
||||
}
|
||||
|
||||
return pass;
|
||||
}
|
||||
|
||||
int main(int argc, char* argv[])
|
||||
{
|
||||
ProblemSize problem_size;
|
||||
ExecutionConfig config;
|
||||
|
||||
problem_size.group_count = 16;
|
||||
|
||||
problem_size.Ms = {
|
||||
167, 183, 177, 181, 153, 139, 156, 173, 163, 150, 204, 184, 168, 156, 168, 148};
|
||||
|
||||
for(int i = 0; i < problem_size.group_count; i++)
|
||||
{
|
||||
problem_size.Ns.push_back(768);
|
||||
problem_size.Ks.push_back(4608);
|
||||
|
||||
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]);
|
||||
}
|
||||
|
||||
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
|
||||
{
|
||||
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");
|
||||
exit(0);
|
||||
}
|
||||
|
||||
return !run_grouped_gemm(problem_size, config);
|
||||
}
|
||||
@@ -193,6 +193,7 @@ template <typename ALayout,
|
||||
index_t CShuffleNXdlPerWavePerShuffle,
|
||||
typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
|
||||
index_t CDEBlockTransferScalarPerVector_NPerBlock,
|
||||
typename ComputeType = ADataType,
|
||||
LoopScheduler LoopSched = make_default_loop_scheduler()>
|
||||
struct DeviceGroupedGemm_Xdl_Fixed_NK : public DeviceGroupedGemmFixedNK<ALayout,
|
||||
BLayout,
|
||||
@@ -217,6 +218,8 @@ struct DeviceGroupedGemm_Xdl_Fixed_NK : public DeviceGroupedGemmFixedNK<ALayout,
|
||||
// GridwiseGemm
|
||||
using GridwiseGemm = GridwiseGemmMultipleD_xdl_splitk_cshuffle<
|
||||
ADataType, // TODO: distinguish A/B datatype
|
||||
BDataType,
|
||||
ComputeType,
|
||||
AccDataType,
|
||||
CShuffleDataType,
|
||||
DsDataType,
|
||||
|
||||
@@ -75,12 +75,24 @@ struct PassThrough
|
||||
y = type_convert<bhalf_t>(x);
|
||||
}
|
||||
|
||||
template <>
|
||||
__host__ __device__ void operator()<float, half_t>(float& y, const half_t& x) const
|
||||
{
|
||||
y = type_convert<float>(x);
|
||||
}
|
||||
|
||||
template <>
|
||||
__host__ __device__ void operator()<int8_t, int8_t>(int8_t& y, const int8_t& x) const
|
||||
{
|
||||
y = x;
|
||||
}
|
||||
|
||||
template <>
|
||||
__host__ __device__ void operator()<half_t, int8_t>(half_t& y, const int8_t& x) const
|
||||
{
|
||||
y = type_convert<half_t>(x);
|
||||
}
|
||||
|
||||
template <>
|
||||
__host__ __device__ void operator()<int8_t, int32_t>(int8_t& y, const int32_t& x) const
|
||||
{
|
||||
|
||||
@@ -29,7 +29,9 @@ namespace ck {
|
||||
// E = cde_op(C, D0, D1, ...)
|
||||
// Assume:
|
||||
// D0, D1, ... and E have the same layout
|
||||
template <typename ABDataType, // FIXME: don't assume A/B have same datatype
|
||||
template <typename ADataType,
|
||||
typename BDataType,
|
||||
typename ComputeType,
|
||||
typename AccDataType,
|
||||
typename CShuffleDataType,
|
||||
typename DsDataType,
|
||||
@@ -96,17 +98,6 @@ struct GridwiseGemmMultipleD_xdl_splitk_cshuffle
|
||||
using GridwiseGemmPipe = remove_cvref_t<
|
||||
decltype(GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage, LoopSched>())>;
|
||||
|
||||
// denorm test fix, required to work around fp16 mfma issue
|
||||
// we convert fp16->fp32->bf16 and execute bf16 mfma instruction
|
||||
// when mfma if fixed, remove this section and update
|
||||
// ABDataTypeAdjusted -> ABDataType throughout this file
|
||||
#if CK_WORKAROUND_DENORM_FIX
|
||||
using ABDataTypeAdjusted =
|
||||
conditional_t<is_same_v<ABDataType, ck::half_t>, ck::bhalf_t, ABDataType>;
|
||||
#else
|
||||
using ABDataTypeAdjusted = ABDataType;
|
||||
#endif
|
||||
|
||||
__host__ __device__ static constexpr auto GetABlockDescriptor_KBatch_AK0PerBlock_MPerBlock_AK1()
|
||||
{
|
||||
// A matrix in LDS memory, dst of blockwise copy
|
||||
@@ -196,7 +187,7 @@ struct GridwiseGemmMultipleD_xdl_splitk_cshuffle
|
||||
c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize();
|
||||
|
||||
return math::max((a_block_space_size_aligned + b_block_space_size_aligned) *
|
||||
sizeof(ABDataType),
|
||||
sizeof(ComputeType),
|
||||
c_block_size * sizeof(CShuffleDataType));
|
||||
}
|
||||
|
||||
@@ -401,8 +392,8 @@ struct GridwiseGemmMultipleD_xdl_splitk_cshuffle
|
||||
// check tensor size: cannot be larger than 2GB each
|
||||
constexpr long_index_t TwoGB = (long_index_t{1} << 31);
|
||||
|
||||
if(!(a_grid_desc_kbatch_ak0_m_ak1.GetElementSpaceSize() * sizeof(ABDataType) <= TwoGB &&
|
||||
b_grid_desc_kbatch_bk0_n_bk1.GetElementSpaceSize() * sizeof(ABDataType) <= TwoGB &&
|
||||
if(!(a_grid_desc_kbatch_ak0_m_ak1.GetElementSpaceSize() * sizeof(ADataType) <= TwoGB &&
|
||||
b_grid_desc_kbatch_bk0_n_bk1.GetElementSpaceSize() * sizeof(BDataType) <= TwoGB &&
|
||||
e_grid_desc_m_n.GetElementSpaceSize() * sizeof(EDataType) <= TwoGB))
|
||||
{
|
||||
return false;
|
||||
@@ -470,8 +461,8 @@ struct GridwiseGemmMultipleD_xdl_splitk_cshuffle
|
||||
typename EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
|
||||
typename CDEElementwiseOperation_,
|
||||
typename Block2ETileMap>
|
||||
__device__ static void Run(const ABDataType* __restrict__ p_a_grid,
|
||||
const ABDataType* __restrict__ p_b_grid,
|
||||
__device__ static void Run(const ADataType* __restrict__ p_a_grid,
|
||||
const BDataType* __restrict__ p_b_grid,
|
||||
DsGridPointer p_ds_grid,
|
||||
EDataType* __restrict__ p_e_grid,
|
||||
void* __restrict__ p_shared,
|
||||
@@ -538,8 +529,8 @@ struct GridwiseGemmMultipleD_xdl_splitk_cshuffle
|
||||
Sequence<1, AK0PerBlock, MPerBlock, AK1>,
|
||||
ABlockTransferThreadClusterLengths_KBatch_AK0_M_AK1,
|
||||
ABlockTransferThreadClusterArrangeOrder,
|
||||
ABDataType,
|
||||
ABDataTypeAdjusted,
|
||||
ADataType,
|
||||
ComputeType,
|
||||
decltype(a_grid_desc_kbatch_ak0_m_ak1),
|
||||
decltype(a_block_desc_kbatch_ak0_m_ak1),
|
||||
ABlockTransferSrcAccessOrder,
|
||||
@@ -569,8 +560,8 @@ struct GridwiseGemmMultipleD_xdl_splitk_cshuffle
|
||||
Sequence<1, BK0PerBlock, NPerBlock, BK1>,
|
||||
BBlockTransferThreadClusterLengths_KBatch_BK0_N_BK1,
|
||||
BBlockTransferThreadClusterArrangeOrder,
|
||||
ABDataType,
|
||||
ABDataTypeAdjusted,
|
||||
BDataType,
|
||||
ComputeType,
|
||||
decltype(b_grid_desc_kbatch_bk0_n_bk1),
|
||||
decltype(b_block_desc_kbatch_bk0_n_bk1),
|
||||
BBlockTransferSrcAccessOrder,
|
||||
@@ -606,11 +597,11 @@ struct GridwiseGemmMultipleD_xdl_splitk_cshuffle
|
||||
// sanity check
|
||||
constexpr index_t KPack =
|
||||
math::max(math::lcm(AK1, BK1),
|
||||
MfmaSelector<ABDataTypeAdjusted, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
|
||||
MfmaSelector<ComputeType, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
|
||||
|
||||
auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
|
||||
BlockSize,
|
||||
ABDataTypeAdjusted,
|
||||
ComputeType,
|
||||
AccDataType,
|
||||
decltype(a_block_desc_ak0_m_ak1),
|
||||
decltype(b_block_desc_bk0_n_bk1),
|
||||
@@ -683,11 +674,10 @@ struct GridwiseGemmMultipleD_xdl_splitk_cshuffle
|
||||
a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
|
||||
|
||||
auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
|
||||
static_cast<ABDataTypeAdjusted*>(p_shared),
|
||||
a_block_desc_ak0_m_ak1.GetElementSpaceSize());
|
||||
static_cast<ComputeType*>(p_shared), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
|
||||
|
||||
auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
|
||||
static_cast<ABDataTypeAdjusted*>(p_shared) + a_block_space_size_aligned,
|
||||
static_cast<ComputeType*>(p_shared) + a_block_space_size_aligned,
|
||||
b_block_desc_bk0_n_bk1.GetElementSpaceSize());
|
||||
|
||||
constexpr auto a_block_slice_copy_step = make_multi_index(0, KPerBlock / AK1, 0, 0);
|
||||
@@ -999,8 +989,8 @@ struct GridwiseGemmMultipleD_xdl_splitk_cshuffle
|
||||
const index_t KBatch,
|
||||
const Block2ETileMap& block_2_etile_map)
|
||||
{
|
||||
const auto p_a_grid = reinterpret_cast<const ABDataType*>(p_a_grid_);
|
||||
const auto p_b_grid = reinterpret_cast<const ABDataType*>(p_b_grid_);
|
||||
const auto p_a_grid = reinterpret_cast<const ADataType*>(p_a_grid_);
|
||||
const auto p_b_grid = reinterpret_cast<const BDataType*>(p_b_grid_);
|
||||
const auto p_e_grid = reinterpret_cast<EDataType*>(p_e_grid_);
|
||||
|
||||
using DsGridDesc_M_N =
|
||||
|
||||
@@ -0,0 +1,190 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include <vector>
|
||||
#include <memory>
|
||||
#include "ck/ck.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/device_grouped_gemm_fixed_nk.hpp"
|
||||
|
||||
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
|
||||
// fp16_output
|
||||
void add_device_grouped_gemm_xdl_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_xdl_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);
|
||||
|
||||
// fp8_inputB
|
||||
void add_device_grouped_gemm_xdl_fixed_nk_f16_f8_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_xdl_fixed_nk_f16_f8_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);
|
||||
|
||||
// i8_inputB
|
||||
void add_device_grouped_gemm_xdl_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_xdl_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);
|
||||
|
||||
template <typename ALayout,
|
||||
typename BLayout,
|
||||
typename ELayout,
|
||||
typename ADataType,
|
||||
typename BDataType,
|
||||
typename EDataType>
|
||||
struct DeviceOperationInstanceFactory<
|
||||
ck::tensor_operation::device::DeviceGroupedGemmFixedNK<ALayout,
|
||||
BLayout,
|
||||
Empty_Tuple,
|
||||
ELayout,
|
||||
ADataType,
|
||||
BDataType,
|
||||
Empty_Tuple,
|
||||
EDataType,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
PassThrough>>
|
||||
{
|
||||
using DeviceOp = DeviceGroupedGemmFixedNK<ALayout,
|
||||
BLayout,
|
||||
Empty_Tuple,
|
||||
ELayout,
|
||||
ADataType,
|
||||
BDataType,
|
||||
Empty_Tuple,
|
||||
EDataType,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
PassThrough>;
|
||||
|
||||
static auto GetInstances()
|
||||
{
|
||||
std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
|
||||
|
||||
// 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_xdl_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_xdl_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_xdl_fixed_nk_f16_f8_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_xdl_fixed_nk_f16_f8_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_xdl_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_xdl_fixed_nk_f16_i8_f16_mk_nk_mn_instances(op_ptrs);
|
||||
}
|
||||
}
|
||||
|
||||
return op_ptrs;
|
||||
}
|
||||
};
|
||||
|
||||
} // namespace instance
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
@@ -0,0 +1,10 @@
|
||||
add_instance_library(device_grouped_gemm_fixed_nk_instance
|
||||
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_f8_f16_mk_kn_mn_instance.cpp
|
||||
device_grouped_gemm_xdl_fixed_nk_f16_f8_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
|
||||
)
|
||||
@@ -0,0 +1,75 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#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_xdl_fixed_nk.hpp"
|
||||
|
||||
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
|
||||
using F16 = ck::half_t;
|
||||
using F32 = float;
|
||||
|
||||
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 DsDataType = ck::Tuple<>;
|
||||
|
||||
using DsLayout = ck::Tuple<>;
|
||||
|
||||
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
|
||||
|
||||
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
|
||||
//############################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| C| 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, 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
|
||||
>;
|
||||
|
||||
void add_device_grouped_gemm_xdl_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_operation_instances(
|
||||
instances,
|
||||
device_grouped_gemm_xdl_fixed_nk_f16_f16_f16_mk_kn_mn_irregular_tile_instances{});
|
||||
}
|
||||
|
||||
} // namespace instance
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
@@ -0,0 +1,78 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#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_xdl_fixed_nk.hpp"
|
||||
|
||||
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
|
||||
using F16 = ck::half_t;
|
||||
using F32 = float;
|
||||
|
||||
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 DsDataType = ck::Tuple<>;
|
||||
|
||||
using DsLayout = ck::Tuple<>;
|
||||
|
||||
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
|
||||
|
||||
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>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
void add_device_grouped_gemm_xdl_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_operation_instances(
|
||||
instances,
|
||||
device_grouped_gemm_xdl_fixed_nk_f16_f16_f16_mk_nk_mn_irregular_tile_instances{});
|
||||
}
|
||||
|
||||
} // namespace instance
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
@@ -0,0 +1,75 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#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_xdl_fixed_nk.hpp"
|
||||
|
||||
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
|
||||
using F8 = ck::f8_t;
|
||||
using F16 = ck::half_t;
|
||||
using F32 = float;
|
||||
|
||||
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 DsDataType = ck::Tuple<>;
|
||||
|
||||
using DsLayout = ck::Tuple<>;
|
||||
|
||||
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
|
||||
|
||||
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
|
||||
//############################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| C| 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, 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
|
||||
>;
|
||||
|
||||
void add_device_grouped_gemm_xdl_fixed_nk_f16_f8_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_operation_instances(
|
||||
instances, device_grouped_gemm_xdl_fixed_nk_f16_f8_f16_mk_kn_mn_irregular_tile_instances{});
|
||||
}
|
||||
|
||||
} // namespace instance
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
@@ -0,0 +1,78 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#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_xdl_fixed_nk.hpp"
|
||||
|
||||
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
|
||||
using F8 = ck::f8_t;
|
||||
using F16 = ck::half_t;
|
||||
using F32 = float;
|
||||
|
||||
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 DsDataType = ck::Tuple<>;
|
||||
|
||||
using DsLayout = ck::Tuple<>;
|
||||
|
||||
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
|
||||
|
||||
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>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
void add_device_grouped_gemm_xdl_fixed_nk_f16_f8_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_operation_instances(
|
||||
instances, device_grouped_gemm_xdl_fixed_nk_f16_f8_f16_mk_nk_mn_irregular_tile_instances{});
|
||||
}
|
||||
|
||||
} // namespace instance
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
@@ -0,0 +1,75 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#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_xdl_fixed_nk.hpp"
|
||||
|
||||
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
|
||||
using I8 = int8_t;
|
||||
using F16 = ck::half_t;
|
||||
using F32 = float;
|
||||
|
||||
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 DsDataType = ck::Tuple<>;
|
||||
|
||||
using DsLayout = ck::Tuple<>;
|
||||
|
||||
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
|
||||
|
||||
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
|
||||
//############################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| C| 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, 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
|
||||
>;
|
||||
|
||||
void add_device_grouped_gemm_xdl_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_operation_instances(
|
||||
instances, device_grouped_gemm_xdl_fixed_nk_f16_i8_f16_mk_kn_mn_irregular_tile_instances{});
|
||||
}
|
||||
|
||||
} // namespace instance
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
@@ -0,0 +1,78 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#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_xdl_fixed_nk.hpp"
|
||||
|
||||
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
|
||||
using I8 = int8_t;
|
||||
using F16 = ck::half_t;
|
||||
using F32 = float;
|
||||
|
||||
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 DsDataType = ck::Tuple<>;
|
||||
|
||||
using DsLayout = ck::Tuple<>;
|
||||
|
||||
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
|
||||
|
||||
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>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
void add_device_grouped_gemm_xdl_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_operation_instances(
|
||||
instances, device_grouped_gemm_xdl_fixed_nk_f16_i8_f16_mk_nk_mn_irregular_tile_instances{});
|
||||
}
|
||||
|
||||
} // namespace instance
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
Reference in New Issue
Block a user