Skip to content

GitLab

Commit 036c5234 authored by Adam Osewski's avatar Adam Osewski
Browse files

Merge remote-tracking branch 'origin/develop' into aosewski/ggemm_multi_d2

parents 22995e9a 7843a8a7
Hide whitespace changes
Inline Side-by-side
Showing with 1112 additions and 131 deletions
example/01_gemm/CMakeLists.txt
View file @ 036c5234
......@@ -28,6 +28,8 @@ add_example_executable(example_gemm_xdl_fp8_v3 gemm_xdl_fp8_v3.cpp)
add_example_dependencies(example_gemm_xdl example_gemm_xdl_fp8_v3)
add_example_executable(example_gemm_xdl_fp16_fp8_v3 gemm_xdl_fp16_fp8_v3.cpp)
add_example_dependencies(example_gemm_xdl example_gemm_xdl_fp16_fp8_v3)
add_example_executable(example_gemm_xdl_bf16_v3 gemm_xdl_bf16_v3.cpp)
add_example_dependencies(example_gemm_xdl example_gemm_xdl_bf16_v3)
add_example_executable(example_gemm_xdl_wavelet_fp16 gemm_xdl_wavelet_fp16.cpp)
add_example_dependencies(example_gemm_xdl example_gemm_xdl_wavelet_fp16)
......
example/01_gemm/README.md
View file @ 036c5234
......@@ -7,17 +7,3 @@
#arg3: run kernel # of times (>1)
./bin/example_gemm_xdl 0 1 5
```
Result (MI100 @ 1087Mhz, 133.5TFlops peak FP16)
```
a_m_k: dim 2, lengths {3840, 4096}, strides {4096, 1}
b_k_n: dim 2, lengths {4096, 4096}, strides {1, 4096}
c_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
arg.a_grid_desc_k0_m_k1_{512, 3840, 8}
arg.b_grid_desc_k0_n_k1_{512, 4096, 8}
arg.c_grid_desc_m_n_{ 3840, 4096}
launch_and_time_kernel: grid_dim {480, 1, 1}, block_dim {256, 1, 1}
Warm up
Start running 5 times...
Perf: 1.19685 ms, 107.657 TFlops, 78.8501 GB/s
```
example/01_gemm/gemm_xdl_bf16_v3.cpp 0 → 100644
View file @ 036c5234
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_v3.hpp"
using ADataType = ck::bhalf_t;
using BDataType = ck::bhalf_t;
using AccDataType = float;
using CShuffleDataType = ck::bhalf_t;
using CDataType = ck::bhalf_t;
using ALayout = Row;
using BLayout = Col;
using CLayout = Row;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CElementOp = PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// clang-format off
using DeviceGemmV2Instance =
ck::tensor_operation::device::DeviceGemm_Xdl_CShuffleV3<
ALayout, BLayout, CLayout,
ADataType, BDataType, CDataType, AccDataType, CShuffleDataType,
PassThrough, PassThrough, PassThrough, GemmDefault,
256,
128, 128,
64, 8, 8,
16, 16,
4, 4,
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>,
2, 8, 8, 0,
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>,
2, 8, 8, 0,
1, 2, S<1, 32, 1, 8>, 8,
ck::BlockGemmPipelineScheduler::Intrawave,ck::BlockGemmPipelineVersion::v3>;
// clang-format on
using ReferenceGemmInstance = ck::tensor_operation::host::
ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
#include "run_gemm_example_v2.inc"
int main(int argc, char* argv[]) { return !run_gemm_splitk_example(argc, argv); }
example/01_gemm/run_gemm_example_v2.inc
View file @ 036c5234
......@@ -3,6 +3,88 @@
#pragma once
template <typename DataType>
inline __host__ __device__ constexpr double get_rtol()
{
if constexpr(std::is_same_v<DataType, float>)
{
return 1e-3;
}
else if constexpr(std::is_same_v<DataType, double>)
{
return 1e-6;
}
else if constexpr(std::is_same_v<DataType, ck::half_t>)
{
return 1e-3;
}
else if constexpr(std::is_same_v<DataType, ck::bhalf_t>)
{
return 5e-2;
}
else if constexpr(std::is_same_v<DataType, int32_t>)
{
return 1e-1;
}
else if constexpr(std::is_same_v<DataType, int8_t>)
{
return 1e-1;
}
else if constexpr(std::is_same_v<DataType, ck::f8_t>)
{
return 1e-1; // 240 and 224 are acceptable
}
else if constexpr(std::is_same_v<DataType, ck::bf8_t>)
{
return 1.5e-1; // 57344 and 49152 are acceptable
}
else
{
return 1e-3;
}
}
template <typename DataType>
inline __host__ __device__ constexpr double get_atol()
{
if constexpr(std::is_same_v<DataType, float>)
{
return 1e-3;
}
else if constexpr(std::is_same_v<DataType, double>)
{
return 1e-6;
}
else if constexpr(std::is_same_v<DataType, ck::half_t>)
{
return 1e-3;
}
else if constexpr(std::is_same_v<DataType, ck::bhalf_t>)
{
return 5e-2;
}
else if constexpr(std::is_same_v<DataType, int32_t>)
{
return 1e-1;
}
else if constexpr(std::is_same_v<DataType, int8_t>)
{
return 1e-1;
}
else if constexpr(std::is_same_v<DataType, ck::f8_t>)
{
return 16.1; // 240 and 224 are acceptable
}
else if constexpr(std::is_same_v<DataType, ck::bf8_t>)
{
return 8192.1; // 57344 and 49152 are acceptable
}
else
{
return 1e-3;
}
}
template <typename ProblemType>
bool run_gemm(const ProblemType& problem_size, const ExecutionConfig& config)
{
......@@ -180,7 +262,11 @@ bool run_gemm(const ProblemType& problem_size, const ExecutionConfig& config)
#else
c_m_n_device_buf.FromDevice(c_m_n_device_result.mData.data());
pass &= ck::utils::check_err(c_m_n_device_result, c_m_n_host_result);
pass &= ck::utils::check_err(c_m_n_device_result,
c_m_n_host_result,
"Error: Incorrect results!",
get_rtol<CDataType>(),
get_atol<CDataType>());
#endif
}
......
example/02_gemm_bilinear/README.md
View file @ 036c5234
......@@ -9,20 +9,3 @@
#arg11 to 12: alpha, beta
./bin/example_gemm_bilinear_xdl_fp16 1 1 1 3840 4096 4096 4096 4096 4096 4096 0.5 0.5
```
Result (MI100 @ 1502Mhz, 184.6TFlops peak FP16)
```
a_m_k: dim 2, lengths {3840, 4096}, strides {4096, 1}
b_k_n: dim 2, lengths {4096, 4096}, strides {1, 4096}
c0_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
c_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
arg.a_grid_desc_k0_m_k1_{512, 3840, 8}
arg.b_grid_desc_k0_n_k1_{512, 4096, 8}
arg.c0_grid_desc_m_n_{ 3840, 4096}
arg.c_grid_desc_m_n_{ 3840, 4096}
launch_and_time_kernel: grid_dim {480, 1, 1}, block_dim {256, 1, 1}
Warm up
Start running 1 times...
Perf: 0.936965 ms, 137.517 TFlops, 102.959 GB/s
error: 0
max_diff: 0, 558.5, 558.5
```
example/04_gemm_add_add_fastgelu/README.md
View file @ 036c5234
......@@ -8,16 +8,3 @@
#arg4 to 11: M (256x), N(128x), K(32x), StrideA, StrideB, StrideD0, StrideD1, StrideE"
./bin/example_gemm_add_add_fastgelu_xdl_fp16 1 1 1
```
Result (MI100 @ 1087Mhz, 133.5TFlops peak FP16)
```
a_m_k: dim 2, lengths {3840, 4096}, strides {4096, 1}
b_k_n: dim 2, lengths {4096, 4096}, strides {1, 4096}
d0_m_n: dim 2, lengths {3840, 4096}, strides {0, 1}
d1_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
e_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
launch_and_time_kernel: grid_dim {480, 1, 1}, block_dim {256, 1, 1}
Warm up 1 time
Start running 10 times...
Perf: 1.26914 ms, 101.525 TFlops, 100.804 GB/s, DeviceGemmMultipleD_Xdl_CShuffle<256, 256, 128, 32, 8, 8>
```
example/09_convnd_fwd/CMakeLists.txt
View file @ 036c5234
......@@ -3,8 +3,7 @@ add_example_executable(example_convnd_fwd_xdl_fp16 convnd_fwd_xdl_fp16.cpp)
add_example_executable(example_convnd_fwd_xdl_bf16 convnd_fwd_xdl_bf16.cpp)
add_example_executable(example_convnd_fwd_xdl_int8 convnd_fwd_xdl_int8.cpp)
add_example_executable(example_convnd_fwd_xdl_fp8 convnd_fwd_xdl_fp8.cpp)
# FIXME: re-enable this exampe as test when SWDEV-335738 is fixed
add_example_executable_no_testing(example_convnd_fwd_xdl_fp64 convnd_fwd_xdl_fp64.cpp)
add_example_executable(example_convnd_fwd_xdl_fp64 convnd_fwd_xdl_fp64.cpp)
add_example_executable(example_convnd_fwd_xdl_bf8 convnd_fwd_xdl_bf8.cpp)
add_example_executable(example_convnd_fwd_xdl_fp16_comp_fp8 convnd_fwd_xdl_fp16_comp_fp8.cpp)
add_example_executable(example_convnd_fwd_xdl_fp8_bf8 convnd_fwd_xdl_fp8_bf8.cpp)
......
example/09_convnd_fwd/README.md
View file @ 036c5234
......@@ -16,17 +16,3 @@
# <right padding>, (ie RightPy, RightPx for 2D)
./bin/example_convnd_fwd_xdl 0 1 100
```
Result (MI100 @ 1087Mhz, 33.4TFlops peak FP32)
```
input: dim 4, lengths {128, 192, 71, 71}, strides {967872, 1, 13632, 192}
weights: dim 4, lengths {256, 192, 3, 3}, strides {1728, 1, 576, 192}
output: dim 4, lengths {128, 256, 36, 36}, strides {331776, 1, 9216, 256}
arg.a_grid_desc_k0_m_k1_{432, 165888, 4}
arg.b_grid_desc_k0_n_k1_{432, 256, 4}
arg.c_grid_desc_m_n_{ 165888, 256}
launch_and_time_kernel: grid_dim {1296, 1, 1}, block_dim {256, 1, 1}
Warm up
Start running 100 times...
Perf: 4.43736 ms, 33.0753 TFlops, 150.357 GB/s
```
example/15_grouped_gemm/CMakeLists.txt
View file @ 036c5234
......@@ -31,6 +31,9 @@ add_example_executable(example_grouped_gemm_multiple_d_splitk_xdl_single_kernel_
grouped_gemm_multiple_d_splitk_xdl_single_kernel_fp16.cpp)
add_example_dependencies(example_grouped_gemm_xdl example_grouped_gemm_multiple_d_splitk_xdl_single_kernel_fp16)
add_example_executable(example_grouped_gemm_multiple_d_xdl_fp16 grouped_gemm_multiple_d_xdl_fp16.cpp)
add_example_dependencies(example_grouped_gemm_xdl example_grouped_gemm_multiple_d_xdl_fp16)
if(USE_BITINT_EXTENSION_INT4)
add_example_executable(example_grouped_gemm_xdl_int4 grouped_gemm_xdl_int4.cpp)
add_example_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_int4)
......
example/15_grouped_gemm/README.md
View file @ 036c5234
......@@ -7,19 +7,3 @@
#arg3: run kernel # of times (>1)
./bin/example_grouped_gemm_xdl_fp16 0 1 5
```
Result (MI100 @ 1087Mhz, 133.5TFlops peak FP16)
```
gemm[0] a_m_k: dim 2, lengths {256, 64}, strides {64, 1} b_k_n: dim 2, lengths {64, 128}, strides {1, 64} c_m_n: dim 2, lengths {256, 128}, strides {128, 1}
gemm[1] a_m_k: dim 2, lengths {512, 128}, strides {128, 1} b_k_n: dim 2, lengths {128, 256}, strides {1, 128} c_m_n: dim 2, lengths {512, 256}, strides {256, 1}
gemm[2] a_m_k: dim 2, lengths {768, 192}, strides {192, 1} b_k_n: dim 2, lengths {192, 384}, strides {1, 192} c_m_n: dim 2, lengths {768, 384}, strides {384, 1}
gemm[3] a_m_k: dim 2, lengths {1024, 256}, strides {256, 1} b_k_n: dim 2, lengths {256, 512}, strides {1, 256} c_m_n: dim 2, lengths {1024, 512}, strides {512, 1}
group: 0 arg.a_grid_desc_k0_m_k1_{8, 256, 8}, arg.b_grid_desc_k0_n_k1_{8, 128, 8}, arg.c_grid_desc_m_n_{ 256, 128}
group: 1 arg.a_grid_desc_k0_m_k1_{16, 512, 8}, arg.b_grid_desc_k0_n_k1_{16, 256, 8}, arg.c_grid_desc_m_n_{ 512, 256}
group: 2 arg.a_grid_desc_k0_m_k1_{24, 768, 8}, arg.b_grid_desc_k0_n_k1_{24, 384, 8}, arg.c_grid_desc_m_n_{ 768, 384}
group: 3 arg.a_grid_desc_k0_m_k1_{32, 1024, 8}, arg.b_grid_desc_k0_n_k1_{32, 512, 8}, arg.c_grid_desc_m_n_{ 1024, 512}
launch_and_time_kernel: grid_dim {30, 1, 1}, block_dim {256, 1, 1}
Warm up
Start running 5 times...
Perf: 0.037887 ms, 11.0706 TFlops, 90.8132 GB/s, DeviceGroupedGemmXdl<256, 256, 128, 4, 8, 32, 32, 4, 2>
```
example/15_grouped_gemm/grouped_gemm_multiple_d_xdl_fp16.cpp 0 → 100644
View file @ 036c5234
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, 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_multiple_d_xdl_cshuffle_tile_loop.hpp"
#include "ck/tensor_operation/gpu/device/device_grouped_gemm_tile_loop.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include <ck/utility/data_type.hpp>
#include <ck/utility/tuple.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_multiple_d.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using F16 = ck::half_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using AddAdd = ck::tensor_operation::element_wise::AddAdd;
using ADataType = F16;
using BDataType = F16;
using AccDataType = F32;
using CShuffleDataType = F32;
using DDataType = F16;
using DsDataType = ck::Tuple<DDataType, DDataType>;
using EDataType = F16;
using ALayout = Row;
using BLayout = Col;
using DLayout = Row;
using DsLayout = ck::Tuple<DLayout, DLayout>;
using ELayout = Row;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = AddAdd;
static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
static constexpr int NumDs = 2;
using DeviceGemmInstance =
ck::tensor_operation::device::DeviceGroupedGemmMultipleDXdlCShuffleTileLoop
// 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, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 4>;
// 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<std::vector<ck::index_t>> stride_Ds;
std::vector<ck::index_t> stride_Cs;
ck::index_t group_count;
};
struct ExecutionConfig final
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = true;
};
bool run_grouped_gemm(const ProblemSize& problem_size, const ExecutionConfig& config)
{
auto group_count = problem_size.group_count;
using KernelArguments = ck::tensor_operation::device::GroupedGemmTileLoopKernelArguments<NumDs>;
using GemmDesc = ck::tensor_operation::device::GemmDesc;
// GEMM shape
std::vector<GemmDesc> gemm_descs;
std::vector<KernelArguments> ggemm_kargs;
std::vector<void*> p_Cs;
std::vector<const void*> p_As;
std::vector<const void*> p_Bs;
std::vector<std::array<const void*, NumDs>> p_Ds = {};
gemm_descs.reserve(group_count);
ggemm_kargs.reserve(group_count);
p_As.reserve(group_count);
p_Bs.reserve(group_count);
p_Ds.reserve(group_count);
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<std::array<Tensor<DDataType>, NumDs>> d_tensors;
std::vector<Tensor<EDataType>> c_host_tensors;
std::vector<Tensor<EDataType>> c_device_result_tensors;
a_tensors.reserve(group_count);
b_tensors.reserve(group_count);
d_tensors.reserve(group_count);
c_host_tensors.reserve(group_count);
c_device_result_tensors.reserve(group_count);
using DeviceMemPtr = std::unique_ptr<DeviceMem>;
std::vector<DeviceMemPtr> a_tensors_device, b_tensors_device, c_tensors_device;
std::vector<std::vector<DeviceMemPtr>> d_tensors_device;
a_tensors_device.reserve(group_count);
b_tensors_device.reserve(group_count);
d_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++)
{
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{})));
auto d0_tensor = Tensor<DDataType>(f_host_tensor_descriptor(
problem_size.Ms[i], problem_size.Ns[i], problem_size.stride_Cs[i], DLayout{}));
auto d1_tensor = Tensor<DDataType>(f_host_tensor_descriptor(
problem_size.Ms[i], problem_size.Ns[i], problem_size.stride_Cs[i], DLayout{}));
std::array<Tensor<DDataType>, NumDs> d_tens = {d0_tensor, d1_tensor};
d_tensors.push_back(d_tens);
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_result_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_result_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].GetElementSize() +
sizeof(BDataType) * b_tensors[i].GetElementSize() +
sizeof(DDataType) * d_tensors[i][0].GetElementSize() * NumDs +
sizeof(EDataType) * c_device_result_tensors[i].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});
for(int j = 0; j < NumDs; ++j)
{
d_tensors[i][j].GenerateTensorValue(GeneratorTensor_2<DDataType>{-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});
for(int j = 0; j < NumDs; ++j)
{
d_tensors[i][j].GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
}
break;
default:
a_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<0>{});
b_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<1>{});
for(int j = 0; j < NumDs; ++j)
{
d_tensors[i][j].GenerateTensorValue(GeneratorTensor_Sequential<0>{});
}
}
}
for(int i = 0; i < group_count; i++)
{
a_tensors_device.emplace_back(
std::make_unique<DeviceMem>(a_tensors[i].GetElementSpaceSize() * sizeof(ADataType)));
b_tensors_device.emplace_back(
std::make_unique<DeviceMem>(b_tensors[i].GetElementSpaceSize() * sizeof(BDataType)));
c_tensors_device.emplace_back(std::make_unique<DeviceMem>(
c_device_result_tensors[i].GetElementSpaceSize() * sizeof(EDataType)));
for(int j = 0; j < NumDs; ++j)
{
d_tensors_device[i].emplace_back(std::make_unique<DeviceMem>(
d_tensors[i][j].GetElementSpaceSize() * sizeof(DDataType)));
}
a_tensors_device[i]->ToDevice(a_tensors[i].mData.data());
b_tensors_device[i]->ToDevice(b_tensors[i].mData.data());
for(int j = 0; j < NumDs; ++j)
{
d_tensors_device[i][j]->ToDevice(d_tensors[i][j].mData.data());
}
c_tensors_device[i]->SetZero();
p_As.push_back(a_tensors_device[i]->GetDeviceBuffer());
p_Bs.push_back(b_tensors_device[i]->GetDeviceBuffer());
p_Ds.push_back(
{d_tensors_device[i][0]->GetDeviceBuffer(), d_tensors_device[i][1]->GetDeviceBuffer()});
p_Cs.push_back(c_tensors_device[i]->GetDeviceBuffer());
// The device op does not have to know M problem size at lunch time.
gemm_descs.push_back({0,
problem_size.Ns[i],
problem_size.Ks[i],
problem_size.stride_As[i],
problem_size.stride_Bs[i],
problem_size.stride_Cs[i],
{problem_size.stride_Cs[i], problem_size.stride_Cs[i]}});
ggemm_kargs.push_back(
{a_tensors_device[i]->GetDeviceBuffer(),
b_tensors_device[i]->GetDeviceBuffer(),
{d_tensors_device[i][0]->GetDeviceBuffer(), d_tensors_device[i][1]->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], problem_size.stride_Cs[i]},
problem_size.stride_Cs[i]});
}
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto cde_element_op = CDEElementOp{};
auto gemm = DeviceGemmInstance{};
auto invoker = gemm.MakeInvoker();
// do GEMM
auto argument = gemm.MakeArgument(
p_As, p_Bs, p_Ds, p_Cs, gemm_descs, a_element_op, b_element_op, cde_element_op);
if(!gemm.IsSupportedArgument(argument))
{
throw std::runtime_error(
"wrong! device_gemm with the specified compilation parameters does "
"not support this GEMM problem");
}
DeviceMem gemm_arg_dev_mem(gemm.GetDeviceKernelArgSize(&argument));
hip_check_error(hipMemcpy(gemm_arg_dev_mem.GetDeviceBuffer(),
ggemm_kargs.data(),
gemm.GetDeviceKernelArgSize(&argument),
hipMemcpyHostToDevice));
gemm.SetDeviceKernelArgs(argument, gemm_arg_dev_mem.GetDeviceBuffer());
invoker.Run(argument, StreamConfig{nullptr, false, 1});
bool pass = true;
if(config.do_verification)
{
using ReferenceGemmInstance =
ck::tensor_operation::host::ReferenceGemmMultipleD<ADataType,
BDataType,
DsDataType,
EDataType,
AccDataType,
AElementOp,
BElementOp,
CDEElementOp>;
for(std::size_t i = 0; i < gemm_descs.size(); i++)
{
auto karg = ggemm_kargs[i];
auto dev_res_tensor =
Tensor<float>(f_host_tensor_descriptor(karg.M, karg.N, karg.StrideE, ELayout{}));
c_tensors_device[i]->FromDevice(c_device_result_tensors[i].mData.data());
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument(a_tensors[i],
b_tensors[i],
d_tensors[i],
c_host_tensors[i],
a_element_op,
b_element_op,
cde_element_op);
ref_invoker.Run(ref_argument);
pass &= ck::utils::check_err(c_device_result_tensors[i], c_host_tensors[i]);
}
std::cout << "Verification: " << (pass ? "SUCCESS" : "FAILURE") << "!" << std::endl;
}
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;
}
return pass;
}
std::vector<int> argToIntArray(char* input)
{
std::vector<int> out;
std::istringstream in(input);
std::string item;
while(std::getline(in, item, ','))
{
out.push_back(std::stoi(item));
}
return out;
}
int main(int argc, char* argv[])
{
ProblemSize problem_size;
ExecutionConfig config;
if(argc < 10)
{
std::vector<ck::index_t> Ms{64, 127, 255, 129, 260, 190, 77};
problem_size.group_count = Ms.size();
for(int i = 0; i < problem_size.group_count; i++)
{
problem_size.Ms.push_back(Ms[i]);
problem_size.Ns.push_back(252);
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]);
problem_size.stride_Ds.push_back({});
for(int j = 0; j < NumDs; ++j)
{
problem_size.stride_Ds[i].push_back(problem_size.Ns[i]);
}
}
std::cout
<< "Usage:\n"
<< "arg1: verification (0=no, 1=yes)\n"
<< "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
<< "arg3: time kernel (0=n0, 1=yes)\n"
<< "arg4 to 9: Ms, Ns, Ks, StrideAs, StrideBs, StrideCs (e.g., 256,256 128,128 64,64 "
"64,64 64,64 128,128)\n"
<< "... setting default values." << std::endl;
}
else
{
config.do_verification = std::stoi(argv[1]);
config.init_method = std::stoi(argv[2]);
config.time_kernel = std::stoi(argv[3]);
problem_size.Ms = argToIntArray(argv[4]);
problem_size.Ns = argToIntArray(argv[5]);
problem_size.Ks = argToIntArray(argv[6]);
problem_size.stride_As = argToIntArray(argv[7]);
problem_size.stride_Bs = argToIntArray(argv[8]);
problem_size.stride_Cs = argToIntArray(argv[9]);
for(int j = 0; j < NumDs; ++j)
{
problem_size.stride_Ds.push_back(problem_size.stride_Cs);
}
problem_size.group_count = problem_size.Ms.size();
}
return !run_grouped_gemm(problem_size, config);
}
example/26_contraction/README.md
View file @ 036c5234
......@@ -7,14 +7,3 @@
#arg3: time kernel (0=no, 1=yes)
./bin/example_contraction_bilinear_xdl_fp32 1 1 1
```
Result (MI100 @ dynammic freq, 46TFlops peak FP32)
```
a_ms_ks: dim 4, lengths {30, 128, 32, 64}, strides {524288, 4096, 128, 1}
b_ks_ns: dim 4, lengths {32, 64, 32, 64}, strides {128, 1, 524288, 4096}
c_ms_ns: dim 4, lengths {30, 128, 32, 64}, strides {524288, 4096, 128, 1}
launch_and_time_kernel: grid_dim {240, 1, 1}, block_dim {256, 1, 1}
Warm up 1 time
Start running 10 times...
Perf: 0.843286 ms, 38.1985 TFlops, 94.5014 GB/s, DeviceContractionMultipleD_Xdl_CShuffle<256, 256, 128, 16, 4, 4>
```
example/30_grouped_conv_fwd_multiple_d/README.md
View file @ 036c5234
......@@ -16,15 +16,3 @@ Following arguments (depending on number of spatial dims):
./bin/example_grouped_conv_fwd_bias_relu_add_xdl_fp16 1 1 1
```
Result (MI100)
```
in: dim 5, lengths {1, 128, 192, 71, 71}, strides {192, 967872, 1, 13632, 192}
wei: dim 5, lengths {1, 256, 192, 3, 3}, strides {442368, 1728, 1, 576, 192}
bias: dim 5, lengths {1, 128, 256, 36, 36}, strides {256, 0, 1, 0, 0}
residual: dim 5, lengths {1, 128, 256, 36, 36}, strides {256, 0, 1, 0, 0}
out: dim 5, lengths {1, 128, 256, 36, 36}, strides {256, 331776, 1, 9216, 256}
launch_and_time_kernel: grid_dim {1296, 1, 1}, block_dim {256, 1, 1}
Warm up 1 time
Start running 10 times...
Perf: 1.55981 ms, 94.0927 TFlops, 213.868 GB/s, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<256, 128, 256, 16, Default>
```
example/46_gemm_add_multiply/README.md
View file @ 036c5234
......@@ -8,19 +8,3 @@
#arg4 to 11: M (256x), N(128x), K(32x), StrideA, StrideB, StrideD0, StrideD1, StrideE"
./bin/example_gemm_add_multiply_dl_fp16 1 1 1
```
Result (MI100 @ 1087Mhz, 133.5TFlops peak FP16)
```
a_m_k: dim 2, lengths {3840, 4096}, strides {4096, 1}
b_k_n: dim 2, lengths {4096, 4096}, strides {4096, 1}
d0_m_n: dim 2, lengths {3840, 4096}, strides {0, 1}
d1_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
e_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
arg.a_grid_desc_k0_m0_m1_k1_{2048, 3840, 2}
arg.b_grid_desc_k0_n0_n1_k1_{2048, 4096, 2}
arg.e_grid_desc_m_n_{ 3840, 4096}
launch_and_time_kernel: grid_dim {960, 1, 1}, block_dim {256, 1, 1}
Warm up 1 time
Start running 10 times...
Perf: 3.99904 ms, 32.22 TFlops, 31.9913 GB/s, DeviceGemmMultipleD_Dl<256, 128, 128, 16, 2, 4, 4, 1>
```
example/59_grouped_gemm_multi_ABD/grouped_gemm_multi_abd_xdl_fixed_nk_bias_bf16_i8.cpp
View file @ 036c5234
......@@ -52,12 +52,12 @@ using BsLayout = ck::Tuple<B0Layout, B1Layout>;
using DsLayout = ck::Tuple<Row>;
using ELayout = Row;
using Scales = ck::tensor_operation::element_wise::Scales;
using Multiply = ck::tensor_operation::element_wise::Multiply;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using AddFastGelu = ck::tensor_operation::element_wise::AddFastGelu;
using AElementOp = PassThrough;
using BElementOp = Scales;
using BElementOp = Multiply;
using CDEElementOp = AddFastGelu;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
......
example/60_gemm_multi_ABD/CMakeLists.txt
View file @ 036c5234
add_example_executable(example_gemm_multi_ABD_xdl_fp16 gemm_multi_ABD_xdl_fp16.cpp)
add_example_executable(example_gemm_multi_ABD_xdl_bf16_i8 gemm_multi_ABD_xdl_bf16_i8.cpp)
\ No newline at end of file
add_example_executable(example_gemm_multi_ABD_xdl_bias_fastgelu_bf16_i8 gemm_multi_ABD_xdl_bias_fastgelu_bf16_i8.cpp)
add_example_executable(example_gemm_multi_ABD_xdl_multiply_bias_fastgelu_bf16_i8 gemm_multi_ABD_xdl_multiply_bias_fastgelu_bf16_i8.cpp)
add_example_executable(example_gemm_multi_ABD_xdl_fastgelu_bf16_i8 gemm_multi_ABD_xdl_fastgelu_bf16_i8.cpp)
example/60_gemm_multi_ABD/gemm_multi_ABD_xdl_bf16_i8.cpp → example/60_gemm_multi_ABD/gemm_multi_ABD_xdl_bias_fastgelu_bf16_i8.cpp
View file @ 036c5234
......@@ -18,9 +18,12 @@
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/utility/blkgemmpipe_scheduler.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using F16 = ck::half_t;
using BF16 = ck::bhalf_t;
using I8 = int8_t;
using F32 = float;
......@@ -41,22 +44,22 @@ using EDataType = BF16;
using A0Layout = Row;
using AsLayout = ck::Tuple<A0Layout>;
using B0Layout = Col;
using B0Layout = Row;
using B1Layout = B0Layout;
using BsLayout = ck::Tuple<B0Layout, B1Layout>;
using D0Layout = Row;
using DsLayout = ck::Tuple<D0Layout>;
using ELayout = Row;
using Scales = ck::tensor_operation::element_wise::Scales;
using Multiply = ck::tensor_operation::element_wise::Multiply;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using AddFastGelu = ck::tensor_operation::element_wise::AddFastGelu;
using AElementOp = PassThrough;
using BElementOp = Scales;
using BElementOp = Multiply;
using CDEElementOp = AddFastGelu;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default;
using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleABD_Xdl_CShuffle
// clang-format off
......@@ -64,9 +67,9 @@ using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleABD_Xdl
///######| | | | | 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|
///######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementOp, BElementOp, CDEElementOp, GemmSpec, 1, 128, 16, 128, 32, 8, 8, 16, 16, 1, 4, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 16, 1, 8>, 1>;
< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementOp, BElementOp, CDEElementOp, GemmSpec, 1, 256, 128, 128, 64, 8, 4, 32, 32, 2, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 32, 1, 8>, 8, ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v4>;
// clang-format on
int main(int argc, char* argv[])
{
bool do_verification = true;
......@@ -74,13 +77,13 @@ int main(int argc, char* argv[])
bool time_kernel = false;
// GEMM shape
ck::index_t M = 64;
ck::index_t N = 1024;
ck::index_t K = 512;
ck::index_t M = 4096;
ck::index_t N = 768;
ck::index_t K = 6144;
ck::index_t StrideA = K;
ck::index_t StrideB = K;
ck::index_t StrideD = N;
ck::index_t StrideB = N;
ck::index_t StrideD = 0;
ck::index_t StrideE = N;
if(argc == 1)
......
example/60_gemm_multi_ABD/gemm_multi_ABD_xdl_fastgelu_bf16_i8.cpp 0 → 100644
View file @ 036c5234
// 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/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_abd_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.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"
#include "ck/library/utility/check_err.hpp"
#include "ck/utility/blkgemmpipe_scheduler.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using F16 = ck::half_t;
using BF16 = ck::bhalf_t;
using I8 = int8_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using A0DataType = BF16;
using AsDataType = ck::Tuple<A0DataType>;
using B0DataType = I8;
using B1DataType = BF16;
using BsDataType = ck::Tuple<B0DataType, B1DataType>;
using AccDataType = F32;
using CShuffleDataType = F32;
using D0DataType = BF16;
using DsDataType = ck::Tuple<>;
using EDataType = BF16;
using A0Layout = Row;
using AsLayout = ck::Tuple<A0Layout>;
using B0Layout = Row;
using B1Layout = B0Layout;
using BsLayout = ck::Tuple<B0Layout, B1Layout>;
using D0Layout = Row;
using DsLayout = ck::Tuple<>;
using ELayout = Row;
using Multiply = ck::tensor_operation::element_wise::Multiply;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using FastGelu = ck::tensor_operation::element_wise::FastGelu;
using AElementOp = PassThrough;
using BElementOp = Multiply;
using CDEElementOp = FastGelu;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default;
using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleABD_Xdl_CShuffle
// 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|
///######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementOp, BElementOp, CDEElementOp, GemmSpec, 1, 256, 128, 128, 64, 8, 4, 32, 32, 2, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 32, 1, 8>, 8, ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v4>;
// clang-format on
int main(int argc, char* argv[])
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
// GEMM shape
ck::index_t M = 4096;
ck::index_t N = 768;
ck::index_t K = 6144;
ck::index_t StrideA = K;
ck::index_t StrideB = N;
ck::index_t StrideD = 0;
ck::index_t StrideE = N;
if(argc == 1)
{
// use default case
}
else if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else if(argc == 11)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
M = std::stoi(argv[4]);
N = std::stoi(argv[5]);
K = std::stoi(argv[6]);
StrideA = std::stoi(argv[7]);
StrideB = std::stoi(argv[8]);
StrideD = std::stoi(argv[9]);
StrideE = std::stoi(argv[10]);
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: time kernel (0=no, 1=yes)\n");
printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideD, StrideE\n");
exit(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});
}
};
Tensor<A0DataType> a0_m_k(f_host_tensor_descriptor(M, K, StrideA, A0Layout{}));
Tensor<B0DataType> b0_k_n(f_host_tensor_descriptor(K, N, StrideB, B0Layout{}));
Tensor<B1DataType> b1_k_n(f_host_tensor_descriptor(K, N, 0, B1Layout{}));
Tensor<D0DataType> d_m_n(f_host_tensor_descriptor(M, N, StrideD, D0Layout{}));
Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
std::cout << "a0_m_k: " << a0_m_k.mDesc << std::endl;
std::cout << "b0_k_n: " << b0_k_n.mDesc << std::endl;
std::cout << "b1_k_n: " << b1_k_n.mDesc << std::endl;
std::cout << "d_m_n: " << d_m_n.mDesc << std::endl;
std::cout << "e_m_n: " << e_m_n_host_result.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
a0_m_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-5, 5});
b0_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-5, 5});
b1_k_n.GenerateTensorValue(GeneratorTensor_2<B1DataType>{0, 5});
d_m_n.GenerateTensorValue(GeneratorTensor_2<D0DataType>{-5, 5});
break;
default:
a0_m_k.GenerateTensorValue(GeneratorTensor_3<A0DataType>{0.0, 1.0});
b0_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-5, 5});
b1_k_n.GenerateTensorValue(GeneratorTensor_3<B1DataType>{0, 5});
d_m_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{-0.5, 0.5});
}
DeviceMem a0_device_buf(sizeof(A0DataType) * a0_m_k.mDesc.GetElementSpaceSize());
DeviceMem b0_device_buf(sizeof(B0DataType) * b0_k_n.mDesc.GetElementSpaceSize());
DeviceMem b1_device_buf(sizeof(B1DataType) * b1_k_n.mDesc.GetElementSpaceSize());
DeviceMem d_device_buf(sizeof(D0DataType) * d_m_n.mDesc.GetElementSpaceSize());
DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
a0_device_buf.ToDevice(a0_m_k.mData.data());
b0_device_buf.ToDevice(b0_k_n.mData.data());
b1_device_buf.ToDevice(b1_k_n.mData.data());
d_device_buf.ToDevice(d_m_n.mData.data());
e_device_buf.ToDevice(e_m_n_device_result.mData.data());
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto cde_element_op = CDEElementOp{};
constexpr ck::index_t NumATensor = 1;
constexpr ck::index_t NumBTensor = 2;
constexpr ck::index_t NumDTensor = 0;
// do GEMM
auto device_op = DeviceOpInstance{};
auto invoker = device_op.MakeInvoker();
auto argument =
device_op.MakeArgument(std::array<const void*, NumATensor>{a0_device_buf.GetDeviceBuffer()},
std::array<const void*, NumBTensor>{b0_device_buf.GetDeviceBuffer(),
b1_device_buf.GetDeviceBuffer()},
std::array<const void*, NumDTensor>{},
e_device_buf.GetDeviceBuffer(),
M,
N,
K,
std::array<ck::index_t, NumATensor>{StrideA},
std::array<ck::index_t, NumBTensor>{StrideB, 0},
std::array<ck::index_t, NumDTensor>{},
StrideE,
a_element_op,
b_element_op,
cde_element_op);
if(!device_op.IsSupportedArgument(argument))
{
throw std::runtime_error(
"wrong! device_gemm with the specified compilation parameters does "
"not support this GEMM problem");
}
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype =
sizeof(A0DataType) * M * K + sizeof(B0DataType) * K * N + sizeof(EDataType) * M * N;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
<< std::endl;
e_device_buf.FromDevice(e_m_n_device_result.mData.data());
if(do_verification)
{
Tensor<CShuffleDataType> c_m_n({M, N});
Tensor<A0DataType> a_m_k({M, K});
Tensor<B1DataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, B0Layout{}));
for(int n = 0; n < N; ++n)
{
for(int k = 0; k < K; ++k)
{
b_element_op(b_k_n(k, n), b0_k_n(k, n), b1_k_n(k, n));
}
}
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<A0DataType,
B1DataType,
CShuffleDataType,
AccDataType,
PassThrough,
PassThrough,
PassThrough>;
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument(
a0_m_k, b_k_n, c_m_n, PassThrough{}, PassThrough{}, PassThrough{});
ref_invoker.Run(ref_argument);
for(int m = 0; m < M; ++m)
{
for(int n = 0; n < N; ++n)
{
cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n));
}
}
e_device_buf.FromDevice(e_m_n_device_result.mData.data());
return ck::utils::check_err(e_m_n_device_result, e_m_n_host_result) ? 0 : 1;
}
return 0;
}
example/60_gemm_multi_ABD/gemm_multi_ABD_xdl_multiply_bias_fastgelu_bf16_i8.cpp 0 → 100644
View file @ 036c5234
// 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/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_abd_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.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"
#include "ck/library/utility/check_err.hpp"
#include "ck/utility/blkgemmpipe_scheduler.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using F16 = ck::half_t;
using BF16 = ck::bhalf_t;
using I8 = int8_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using A0DataType = BF16;
using AsDataType = ck::Tuple<A0DataType>;
using B0DataType = I8;
using B1DataType = BF16;
using BsDataType = ck::Tuple<B0DataType>;
using AccDataType = F32;
using CShuffleDataType = F32;
using D0DataType = BF16;
using DsDataType = ck::Tuple<B1DataType, D0DataType>;
using EDataType = BF16;
using A0Layout = Row;
using AsLayout = ck::Tuple<A0Layout>;
using B0Layout = Row;
using B1Layout = B0Layout;
using BsLayout = ck::Tuple<B0Layout>;
using D0Layout = Row;
using DsLayout = ck::Tuple<B1Layout, D0Layout>;
using ELayout = Row;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using MultiplyAddFastGelu = ck::tensor_operation::element_wise::MultiplyAddFastGelu;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = MultiplyAddFastGelu;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default;
using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleABD_Xdl_CShuffle
// 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|
///######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementOp, BElementOp, CDEElementOp, GemmSpec, 1, 256, 128, 128, 64, 8, 4, 32, 32, 2, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 32, 1, 8>, 8, ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v4>;
// clang-format on
int main(int argc, char* argv[])
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
// GEMM shape
ck::index_t M = 4096;
ck::index_t N = 768;
ck::index_t K = 6144;
ck::index_t StrideA = K;
ck::index_t StrideB = N;
ck::index_t StrideD = 0;
ck::index_t StrideE = N;
if(argc == 1)
{
// use default case
}
else if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else if(argc == 11)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
M = std::stoi(argv[4]);
N = std::stoi(argv[5]);
K = std::stoi(argv[6]);
StrideA = std::stoi(argv[7]);
StrideB = std::stoi(argv[8]);
StrideD = std::stoi(argv[9]);
StrideE = std::stoi(argv[10]);
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: time kernel (0=no, 1=yes)\n");
printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideD, StrideE\n");
exit(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});
}
};
Tensor<A0DataType> a0_m_k(f_host_tensor_descriptor(M, K, StrideA, A0Layout{}));
Tensor<B0DataType> b0_k_n(f_host_tensor_descriptor(K, N, StrideB, B0Layout{}));
Tensor<B1DataType> b1_k_n(f_host_tensor_descriptor(K, N, 0, B1Layout{}));
Tensor<D0DataType> d_m_n(f_host_tensor_descriptor(M, N, StrideD, D0Layout{}));
Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
std::cout << "a0_m_k: " << a0_m_k.mDesc << std::endl;
std::cout << "b0_k_n: " << b0_k_n.mDesc << std::endl;
std::cout << "b1_k_n: " << b1_k_n.mDesc << std::endl;
std::cout << "d_m_n: " << d_m_n.mDesc << std::endl;
std::cout << "e_m_n: " << e_m_n_host_result.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
a0_m_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-5, 5});
b0_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-5, 5});
b1_k_n.GenerateTensorValue(GeneratorTensor_2<B1DataType>{0, 5});
d_m_n.GenerateTensorValue(GeneratorTensor_2<D0DataType>{-5, 5});
break;
default:
a0_m_k.GenerateTensorValue(GeneratorTensor_3<A0DataType>{0.0, 1.0});
b0_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-5, 5});
b1_k_n.GenerateTensorValue(GeneratorTensor_3<B1DataType>{0, 5});
d_m_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{-0.5, 0.5});
}
DeviceMem a0_device_buf(sizeof(A0DataType) * a0_m_k.mDesc.GetElementSpaceSize());
DeviceMem b0_device_buf(sizeof(B0DataType) * b0_k_n.mDesc.GetElementSpaceSize());
DeviceMem b1_device_buf(sizeof(B1DataType) * b1_k_n.mDesc.GetElementSpaceSize());
DeviceMem d_device_buf(sizeof(D0DataType) * d_m_n.mDesc.GetElementSpaceSize());
DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
a0_device_buf.ToDevice(a0_m_k.mData.data());
b0_device_buf.ToDevice(b0_k_n.mData.data());
b1_device_buf.ToDevice(b1_k_n.mData.data());
d_device_buf.ToDevice(d_m_n.mData.data());
e_device_buf.ToDevice(e_m_n_device_result.mData.data());
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto cde_element_op = CDEElementOp{};
constexpr ck::index_t NumATensor = 1;
constexpr ck::index_t NumBTensor = 1;
constexpr ck::index_t NumDTensor = 2;
// do GEMM
auto device_op = DeviceOpInstance{};
auto invoker = device_op.MakeInvoker();
auto argument =
device_op.MakeArgument(std::array<const void*, NumATensor>{a0_device_buf.GetDeviceBuffer()},
std::array<const void*, NumBTensor>{b0_device_buf.GetDeviceBuffer()},
std::array<const void*, NumDTensor>{b1_device_buf.GetDeviceBuffer(),
d_device_buf.GetDeviceBuffer()},
e_device_buf.GetDeviceBuffer(),
M,
N,
K,
std::array<ck::index_t, NumATensor>{StrideA},
std::array<ck::index_t, NumBTensor>{StrideB},
std::array<ck::index_t, NumDTensor>{0, StrideD},
StrideE,
a_element_op,
b_element_op,
cde_element_op);
if(!device_op.IsSupportedArgument(argument))
{
throw std::runtime_error(
"wrong! device_gemm with the specified compilation parameters does "
"not support this GEMM problem");
}
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype =
sizeof(A0DataType) * M * K + sizeof(B0DataType) * K * N + sizeof(EDataType) * M * N;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
<< std::endl;
e_device_buf.FromDevice(e_m_n_device_result.mData.data());
if(do_verification)
{
Tensor<CShuffleDataType> c_m_n({M, N});
Tensor<A0DataType> a_m_k({M, K});
Tensor<B1DataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, B0Layout{}));
#if 0
for(int n = 0; n < N; ++n)
{
for(int k = 0; k < K; ++k)
{
b_element_op(b_k_n(k, n), b0_k_n(k, n), b1_k_n(k, n));
}
}
#endif
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<A0DataType,
B0DataType,
CShuffleDataType,
AccDataType,
PassThrough,
PassThrough,
PassThrough>;
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument(
a0_m_k, b0_k_n, c_m_n, PassThrough{}, PassThrough{}, PassThrough{});
ref_invoker.Run(ref_argument);
for(int m = 0; m < M; ++m)
{
for(int n = 0; n < N; ++n)
{
cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n), b1_k_n(0, n), d_m_n(m, n));
}
}
e_device_buf.FromDevice(e_m_n_device_result.mData.data());
return ck::utils::check_err(e_m_n_device_result, e_m_n_host_result) ? 0 : 1;
}
return 0;
}
example/ck_tile/01_fmha/CMakeLists.txt
View file @ 036c5234
......@@ -17,7 +17,7 @@ add_custom_command(
set(EXAMPLE_FMHA_FWD "tile_example_fmha_fwd")
# not using add_example_executable() to add this target, since we don't want this to have
# to be included in "make all/install/check"
message("adding tile_example ${EXAMPLE_NAME}")
message("adding example ${EXAMPLE_FMHA_FWD}")
add_executable(${EXAMPLE_FMHA_FWD} EXCLUDE_FROM_ALL fmha_fwd.cpp)
target_include_directories(${EXAMPLE_FMHA_FWD} PRIVATE ${CMAKE_CURRENT_LIST_DIR})
target_sources(${EXAMPLE_FMHA_FWD} PRIVATE ${FMHA_FWD_GEN_BLOBS})
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment