ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-04-19 22:39:03 +00:00
Code Issues Packages Projects Releases Wiki Activity

GEMM+Bias+ReLU+Add (#76)

Browse Source 
* tweak conv for odd C

* update script

* clean up elementwise op

* fix build

* clean up

* added example for gemm+bias+relu+add

* added example for gemm+bias+relu

* add profiler for gemm_s_shuffle; re-org files

* add profiler

* fix build

* clean up

* clean up

* clean up

* fix build
This commit is contained in:
Chao Liu
2022-02-06 22:32:47 -06:00
committed by GitHub
parent 690c75a7eb
commit 823657ed12
77 changed files with 3865 additions and 932 deletions
Download Patch File Download Diff File Expand all files Collapse all files

53
example/1_gemm_xdl/gemm_xdl.cpp
View File

@@ -11,9 +11,9 @@
#include "host_tensor_generator.hpp"
#include "host_gemm.hpp"
#include "device_tensor.hpp"
#include "device_base.hpp"
#include "device_gemm_xdl_c_shuffle.hpp"
#include "element_wise_operation.hpp"
#include "reference_gemm.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
@@ -72,37 +72,8 @@ using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdl_C_Shuffle
8>; // CBlockTransferScalarPerVector_NWaveNPerXdl
// clang-format on
template <typename AType,
typename BType,
typename CType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation>
static void host_verify(const Tensor<AType>& a_m_k,
const Tensor<BType>& b_k_n,
Tensor<CType>& c_m_n,
const AElementwiseOperation& a_element_op,
const BElementwiseOperation& b_element_op,
const CElementwiseOperation& c_element_op)
{
auto f_mk_kn_mn = [&](auto m, auto n) {
const int K = a_m_k.mDesc.GetLengths()[1];
double v = 0;
for(int k = 0; k < K; ++k)
{
v += static_cast<const double>(a_element_op(a_m_k(m, k))) *
static_cast<const double>(b_element_op(b_k_n(k, n)));
}
c_m_n(m, n) = c_element_op(v);
};
make_ParallelTensorFunctor(f_mk_kn_mn,
c_m_n.mDesc.GetLengths()[0],
c_m_n.mDesc.GetLengths()[1])(std::thread::hardware_concurrency());
}
using ReferenceGemmInstance = ck::tensor_operation::host::
ReferenceGemm<ADataType, BDataType, CDataType, AElementOp, BElementOp, CElementOp>;
int main(int argc, char* argv[])
{
@@ -191,6 +162,10 @@ int main(int argc, char* argv[])
b_k_n_device_buf.ToDevice(b_k_n.mData.data());
c_m_n_device_buf.ToDevice(c_m_n_device_result.mData.data());
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto c_element_op = CElementOp{};
// do GEMM
auto gemm = DeviceGemmInstance{};
auto invoker = gemm.MakeInvoker();
@@ -203,9 +178,9 @@ int main(int argc, char* argv[])
StrideA,
StrideB,
StrideC,
AElementOp{},
BElementOp{},
CElementOp{});
a_element_op,
b_element_op,
c_element_op);
if(!gemm.IsSupportedArgument(argument))
{
@@ -231,7 +206,13 @@ int main(int argc, char* argv[])
if(do_verification)
{
host_verify(a_m_k, b_k_n, c_m_n_host_result, AElementOp{}, BElementOp{}, CElementOp{});
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument(
a_m_k, b_k_n, c_m_n_host_result, a_element_op, b_element_op, c_element_op);
ref_invoker.Run(ref_argument);
check_error(c_m_n_host_result, c_m_n_device_result);
}

61
example/2_gemm_xdl_bias_relu/README.md Normal file
View File

@@ -0,0 +1,61 @@
# Instructions for ```gemm_xdl_bias_relu_add``` Example
## Docker script
```bash
docker run \
-it \
--rm \
--privileged \
--group-add sudo \
-w /root/workspace \
-v ${PATH_TO_LOCAL_WORKSPACE}:/root/workspace \
rocm/tensorflow:rocm4.3.1-tf2.6-dev \
/bin/bash
```
## Build ```gemm_xdl_bias_relu_add```
```bash
mkdir build && cd build
```
```bash
# Need to specify target ID, example below is gfx908
cmake \
-D BUILD_DEV=OFF \
-D CMAKE_BUILD_TYPE=Release \
-D CMAKE_CXX_FLAGS="-DCK_AMD_GPU_GFX908 --amdgpu-target=gfx908 -O3 " \
-D CMAKE_CXX_COMPILER=/opt/rocm/bin/hipcc \
-D CMAKE_PREFIX_PATH=/opt/rocm \
..
```
```bash
make -j gemm_xdl_bias_relu_add
```
## Run ```gemm_xdl_bias_relu_add```
```bash
#arg1: verification (0=no, 1=yes)
#arg2: initialization (0=no init, 1=integer value, 2=decimal value)
#arg3: run kernel # of times (>1)
#arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC
./example/gemm_xdl_bias_relu_add 0 1 5 3840 4096 4096 4096 4096 4096
```
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}
c0_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
c1_m_n: dim 2, lengths {3840, 4096}, strides {1, 0}
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}
arg.c0_grid_desc_m_n_{ 3840, 4096}
arg.c1_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: 1.27583 ms, 100.992 TFlops, 73.9688 GB/s
```

235
example/2_gemm_xdl_bias_relu/gemm_xdl_bias_relu.cpp Normal file
View File

@@ -0,0 +1,235 @@
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include <stdlib.h>
#include <half.hpp>
#include "config.hpp"
#include "print.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "host_gemm.hpp"
#include "device_tensor.hpp"
#include "element_wise_operation.hpp"
#include "device_gemm_xdl_c_shuffle_bias_activation.hpp"
#include "reference_gemm_bias_activation.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using ADataType = ck::half_t;
using BDataType = ck::half_t;
using CDataType = ck::half_t;
using AccDataType = float;
using ALayout = ck::tensor_layout::gemm::RowMajor;
using BLayout = ck::tensor_layout::gemm::ColumnMajor;
using CLayout = ck::tensor_layout::gemm::RowMajor;
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::AddRelu;
// clang-format off
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdl_C_Shuffle_Bias_Activation<
ADataType, // ADataType
BDataType, // BDataType
CDataType, // CDataType
AccDataType, // AccDataType
ALayout, // ALayout
BLayout, // BLayout
CLayout, // CLayout
AElementOp, // AElementwiseOperation
BElementOp, // BElementwiseOperation
CElementOp, // CElementwiseOperation
256, // BlockSize
256, // MPerBlock
128, // NPerBlock
4, // K0PerBlock
8, // K1
32, // MPerXDL
32, // NPerXDL
4, // MXdlPerWave
2, // NXdlPerWave
S<4, 64, 1>, // ABlockTransferThreadClusterLengths_K0_M_K1
S<1, 0, 2>, // ABlockTransferThreadClusterArrangeOrder
S<1, 0, 2>, // ABlockTransferSrcAccessOrder
2, // ABlockTransferSrcVectorDim
8, // ABlockTransferSrcScalarPerVector
8, // ABlockTransferDstScalarPerVector_K1
true, // ABlockLdsAddExtraM
S<4, 64, 1>, // BBlockTransferThreadClusterLengths_K0_N_K1
S<1, 0, 2>, // BBlockTransferThreadClusterArrangeOrder
S<1, 0, 2>, // BBlockTransferSrcAccessOrder
2, // BBlockTransferSrcVectorDim
8, // BBlockTransferSrcScalarPerVector
8, // BBlockTransferDstScalarPerVector_K1
true, // BBlockLdsAddExtraN
1, // CShuffleMXdlPerWavePerShuffle
1, // CShuffleNXdlPerWavePerShuffle
S<1, 1, 32, 1, 1, 8>, // CBlockTransferClusterLengths_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl
8>; // CBlockTransferScalarPerVector_NWaveNPerXdl
// clang-format on
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemmBiasActivation<ADataType,
BDataType,
CDataType,
AElementOp,
BElementOp,
CElementOp>;
int main(int argc, char* argv[])
{
bool do_verification = 0;
int init_method = 0;
int nrepeat = 5;
// GEMM shape
ck::index_t M = 3840;
ck::index_t N = 4096;
ck::index_t K = 4096;
ck::index_t StrideA = 4096;
ck::index_t StrideB = 4096;
ck::index_t StrideC = 4096;
if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]);
}
else if(argc == 10)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = 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]);
StrideC = std::stoi(argv[9]);
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: run kernel # of times (>1)\n");
printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n");
exit(0);
}
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({stride, 1}));
}
else
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({1, stride}));
}
};
Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
Tensor<BDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
Tensor<BDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
// c0_n[n]
Tensor<CDataType> c0_n(HostTensorDescriptor(
std::vector<std::size_t>({static_cast<std::size_t>(N)}), std::vector<std::size_t>({1})));
std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
std::cout << "c_m_n: " << c_m_n_host_result.mDesc << std::endl;
std::cout << "c0_n: " << c0_n.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
c0_n.GenerateTensorValue(GeneratorTensor_2<CDataType>{-5, 5});
break;
default:
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
c0_n.GenerateTensorValue(GeneratorTensor_3<CDataType>{0.0, 1.0});
}
DeviceMem a_m_k_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpace());
DeviceMem b_k_n_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpace());
DeviceMem c_m_n_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpace());
DeviceMem c0_n_device_buf(sizeof(CDataType) * c0_n.mDesc.GetElementSpace());
a_m_k_device_buf.ToDevice(a_m_k.mData.data());
b_k_n_device_buf.ToDevice(b_k_n.mData.data());
c_m_n_device_buf.ToDevice(c_m_n_device_result.mData.data());
c0_n_device_buf.ToDevice(c0_n.mData.data());
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto c_element_op = CElementOp{};
// do GEMM
auto gemm = DeviceGemmInstance{};
auto invoker = gemm.MakeInvoker();
auto argument = gemm.MakeArgument(static_cast<ADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_m_n_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c0_n_device_buf.GetDeviceBuffer()),
M,
N,
K,
StrideA,
StrideB,
StrideC,
a_element_op,
b_element_op,
c_element_op);
if(!gemm.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, nrepeat);
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * M +
sizeof(CDataType) * M * N + sizeof(CDataType) * 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;
c_m_n_device_buf.FromDevice(c_m_n_device_result.mData.data());
if(do_verification)
{
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument(
a_m_k, b_k_n, c_m_n_host_result, c0_n, a_element_op, b_element_op, c_element_op);
ref_invoker.Run(ref_argument);
check_error(c_m_n_host_result, c_m_n_device_result);
}
}

276
example/3_gemm_xdl_bias_relu_add/gemm_xdl_bias_relu_add.cpp
View File

@@ -11,107 +11,9 @@
#include "host_tensor_generator.hpp"
#include "host_gemm.hpp"
#include "device_tensor.hpp"
#include "device_base.hpp"
#include "example/3_gemm_xdl_bias_relu_add/include/device_gemm_xdl_two_extra_source_reduce.hpp"
// C[m, n] = Relu(A[m, k] * B[k, n] + C0[m]) + C1[m, n]
// assume C0 is contiguous in memory
// C0 resides in memory as 1d vector [m], but is represented as 2D matrix [m, n], with stride =
// 0 in the "n" dimension
// assume C1 and C have same layout C
struct BiasReluAdd
{
template <typename T1, typename T2>
__host__ constexpr float operator()(float v0, T1 v1, T2 v2) const
{
float b = v0 + v1;
float c = b > 0 ? b : 0;
float d = c + v2;
return d;
}
template <typename T1, typename T2>
__device__ constexpr float operator()(float v0, T1 v1, T2 v2) const
{
#if 0
float a = v1 + v0;
float b = a > 0 ? a : 0;
float c = b + v2;
return c;
#else
float a = v1 + v2;
float b = v2;
float c = (v0 > -v1) ? a + v0 : v2;
return c;
#endif
}
};
struct DoSomething
{
#if 1
// correct result
// no scratch memory, good VGPR allocation (59)
// good perf (101Tflops @ 1089Mhz)
__host__ __device__ constexpr float operator()(float v0, ck::half_t v1, ck::half_t v2) const
{
constexpr float alpha = 0.1;
constexpr float beta = 0.2;
constexpr float gamma = 0.3;
// compiler seems very volatile to the order of these calculation:
// compiler is very eager to read AccVgpr (v0) out prematurely, resulting in register
// over-allocation. Therefore, move v0 calculation to the very end
float a = ck::half_t(beta) * v1 + ck::half_t(gamma) * v2;
float b = a + float(alpha) * v0;
return b;
}
#elif 0
float alpha = 0.1;
float beta = 0.2;
float gamma = 0.3;
// wrong result
// lots of scratch memory
// huge perf drop
__host__ __device__ constexpr float operator()(float v0, ck::half_t v1, ck::half_t v2) const
{
return alpha * v0 + beta * v1 + gamma * v2;
}
#elif 0
// correct result
// some scratch memory (68 dword)
// some perf drop (94Tflops @ 1089MHz)
// fp64 instructions are used
__host__ __device__ constexpr auto operator()(float v0, ck::half_t v1, ck::half_t v2) const
{
return 0.1 * v0 + 0.2 * v1 + 0.3 * v2;
}
#elif 1
// wrong result
// lots of scratch memory
// huge perf drop
__host__ __device__ constexpr auto operator()(float v0, ck::half_t v1, ck::half_t v2) const
{
return float(0.1) * v0 + float(0.2) * v1 + float(0.3) * v2;
}
#endif
};
struct PassThrough
{
template <typename T>
__host__ __device__ constexpr T operator()(T v) const
{
return v;
}
};
#include "element_wise_operation.hpp"
#include "device_gemm_xdl_c_shuffle_bias_activation_add.hpp"
#include "reference_gemm_bias_activation_add.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
@@ -125,58 +27,58 @@ using ALayout = ck::tensor_layout::gemm::RowMajor;
using BLayout = ck::tensor_layout::gemm::ColumnMajor;
using CLayout = ck::tensor_layout::gemm::RowMajor;
using AOp = PassThrough;
using BOp = PassThrough;
#if 1
using COp = BiasReluAdd;
#else
using COp = DoSomething;
#endif
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::AddReluAdd;
// Compilation parameters for NT problem
// clang-format off
using DeviceGemmInstance =
//#################################################################| AData| BData| CData| AccData| ALayout| BLayout| CLayout| AElementwise| BElementwise| CElementwise| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
//#################################################################| Type| Type| Type| Type| | | | Operation| Operation| Operation| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| SrcDstVectorDim| DstScalar|
//#################################################################| | | | | | | | | | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | | PerVector|
//#################################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
ck::tensor_operation::device::DeviceGemmXdl_two_extra_source_reduce< ADataType, BDataType, CDataType, AccDataType, ALayout, BLayout, CLayout, AOp, BOp, COp, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 7, 1>;
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdl_C_Shuffle_Bias_Activation_Add<
ADataType, // ADataType
BDataType, // BDataType
CDataType, // CDataType
AccDataType, // AccDataType
ALayout, // ALayout
BLayout, // BLayout
CLayout, // CLayout
AElementOp, // AElementwiseOperation
BElementOp, // BElementwiseOperation
CElementOp, // CElementwiseOperation
256, // BlockSize
256, // MPerBlock
128, // NPerBlock
4, // K0PerBlock
8, // K1
32, // MPerXDL
32, // NPerXDL
4, // MXdlPerWave
2, // NXdlPerWave
S<4, 64, 1>, // ABlockTransferThreadClusterLengths_K0_M_K1
S<1, 0, 2>, // ABlockTransferThreadClusterArrangeOrder
S<1, 0, 2>, // ABlockTransferSrcAccessOrder
2, // ABlockTransferSrcVectorDim
8, // ABlockTransferSrcScalarPerVector
8, // ABlockTransferDstScalarPerVector_K1
true, // ABlockLdsAddExtraM
S<4, 64, 1>, // BBlockTransferThreadClusterLengths_K0_N_K1
S<1, 0, 2>, // BBlockTransferThreadClusterArrangeOrder
S<1, 0, 2>, // BBlockTransferSrcAccessOrder
2, // BBlockTransferSrcVectorDim
8, // BBlockTransferSrcScalarPerVector
8, // BBlockTransferDstScalarPerVector_K1
true, // BBlockLdsAddExtraN
1, // CShuffleMXdlPerWavePerShuffle
1, // CShuffleNXdlPerWavePerShuffle
S<1, 1, 32, 1, 1, 8>, // CBlockTransferClusterLengths_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl
8>; // CBlockTransferScalarPerVector_NWaveNPerXdl
// clang-format on
template <typename AType,
typename BType,
typename CType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation>
static void host_verify(const Tensor<AType>& a_m_k,
const Tensor<BType>& b_k_n,
Tensor<CType>& c_m_n,
const Tensor<CType>& c0_m_n,
const Tensor<CType>& c1_m_n,
const AElementwiseOperation& a_element_op,
const BElementwiseOperation& b_element_op,
const CElementwiseOperation& c_element_op)
{
auto f_mk_kn_mn = [&](auto m, auto n) {
const int K = a_m_k.mDesc.GetLengths()[1];
float acc = 0;
for(int k = 0; k < K; ++k)
{
acc += static_cast<const double>(a_element_op(a_m_k(m, k))) *
static_cast<const double>(b_element_op(b_k_n(k, n)));
}
c_m_n(m, n) = c_element_op(acc, c0_m_n(m, n), c1_m_n(m, n));
};
make_ParallelTensorFunctor(f_mk_kn_mn,
c_m_n.mDesc.GetLengths()[0],
c_m_n.mDesc.GetLengths()[1])(std::thread::hardware_concurrency());
}
using ReferenceGemmInstance =
ck::tensor_operation::host::ReferenceGemmBiasActivationAdd<ADataType,
BDataType,
CDataType,
AElementOp,
BElementOp,
CElementOp>;
int main(int argc, char* argv[])
{
bool do_verification = 0;
@@ -188,9 +90,10 @@ int main(int argc, char* argv[])
ck::index_t N = 4096;
ck::index_t K = 4096;
ck::index_t StrideA = 4096;
ck::index_t StrideB = 4096;
ck::index_t StrideC = 4096;
ck::index_t StrideA = 4096;
ck::index_t StrideB = 4096;
ck::index_t StrideC = 4096;
ck::index_t StrideC1 = 4096;
if(argc == 4)
{
@@ -198,7 +101,7 @@ int main(int argc, char* argv[])
init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]);
}
else if(argc == 10)
else if(argc == 11)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
@@ -208,16 +111,17 @@ int main(int argc, char* argv[])
N = std::stoi(argv[5]);
K = std::stoi(argv[6]);
StrideA = std::stoi(argv[7]);
StrideB = std::stoi(argv[8]);
StrideC = std::stoi(argv[9]);
StrideA = std::stoi(argv[7]);
StrideB = std::stoi(argv[8]);
StrideC = std::stoi(argv[9]);
StrideC1 = 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: run kernel # of times (>1)\n");
printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n");
printf("arg4 to 10: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC, StrideC1\n");
exit(0);
}
@@ -240,18 +144,17 @@ int main(int argc, char* argv[])
Tensor<BDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
Tensor<BDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
// C0[m]
Tensor<CDataType> c1_m_n(HostTensorDescriptor(
std::vector<std::size_t>({static_cast<std::size_t>(M), static_cast<std::size_t>(N)}),
std::vector<std::size_t>({1, 0})));
// c0_n[n]
Tensor<CDataType> c0_n(HostTensorDescriptor(
std::vector<std::size_t>({static_cast<std::size_t>(N)}), std::vector<std::size_t>({1})));
// C1[m ,n]
Tensor<BDataType> c0_m_n(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
// c1_m_n[m ,n]
Tensor<BDataType> c1_m_n(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
std::cout << "c_m_n: " << c_m_n_host_result.mDesc << std::endl;
std::cout << "c0_m_n: " << c0_m_n.mDesc << std::endl;
std::cout << "c0_n: " << c0_n.mDesc << std::endl;
std::cout << "c1_m_n: " << c1_m_n.mDesc << std::endl;
switch(init_method)
@@ -260,31 +163,31 @@ int main(int argc, char* argv[])
case 1:
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
c0_m_n.GenerateTensorValue(GeneratorTensor_2<CDataType>{-5, 5});
c0_n.GenerateTensorValue(GeneratorTensor_2<CDataType>{-5, 5});
c1_m_n.GenerateTensorValue(GeneratorTensor_2<CDataType>{-5, 5});
break;
default:
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
c0_m_n.GenerateTensorValue(GeneratorTensor_3<CDataType>{0.0, 1.0});
c0_n.GenerateTensorValue(GeneratorTensor_3<CDataType>{0.0, 1.0});
c1_m_n.GenerateTensorValue(GeneratorTensor_3<CDataType>{0.0, 1.0});
}
DeviceMem a_m_k_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpace());
DeviceMem b_k_n_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpace());
DeviceMem c_m_n_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpace());
DeviceMem c0_m_n_device_buf(sizeof(CDataType) * c0_m_n.mDesc.GetElementSpace());
DeviceMem c0_n_device_buf(sizeof(CDataType) * c0_n.mDesc.GetElementSpace());
DeviceMem c1_m_n_device_buf(sizeof(CDataType) * c1_m_n.mDesc.GetElementSpace());
a_m_k_device_buf.ToDevice(a_m_k.mData.data());
b_k_n_device_buf.ToDevice(b_k_n.mData.data());
c_m_n_device_buf.ToDevice(c_m_n_device_result.mData.data());
c0_m_n_device_buf.ToDevice(c0_m_n.mData.data());
c0_n_device_buf.ToDevice(c0_n.mData.data());
c1_m_n_device_buf.ToDevice(c1_m_n.mData.data());
auto a_element_op = AOp{};
auto b_element_op = BOp{};
auto c_element_op = COp{};
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto c_element_op = CElementOp{};
// do GEMM
auto gemm = DeviceGemmInstance{};
@@ -293,7 +196,7 @@ int main(int argc, char* argv[])
auto argument = gemm.MakeArgument(static_cast<ADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_m_n_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c0_m_n_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c0_n_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c1_m_n_device_buf.GetDeviceBuffer()),
M,
N,
@@ -301,6 +204,7 @@ int main(int argc, char* argv[])
StrideA,
StrideB,
StrideC,
StrideC1,
a_element_op,
b_element_op,
c_element_op);
@@ -314,9 +218,10 @@ int main(int argc, char* argv[])
float ave_time = invoker.Run(argument, nrepeat);
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype =
sizeof(ADataType) * M * K + sizeof(BDataType) * K * M + sizeof(CDataType) * M * N;
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * M +
sizeof(CDataType) * M * N + sizeof(CDataType) * N +
sizeof(CDataType) * M * N;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
@@ -329,14 +234,19 @@ int main(int argc, char* argv[])
if(do_verification)
{
host_verify(a_m_k,
b_k_n,
c_m_n_host_result,
c0_m_n,
c1_m_n,
PassThrough{},
PassThrough{},
c_element_op);
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument(a_m_k,
b_k_n,
c_m_n_host_result,
c0_n,
c1_m_n,
a_element_op,
b_element_op,
c_element_op);
ref_invoker.Run(ref_argument);
check_error(c_m_n_host_result, c_m_n_device_result);
}

537
example/3_gemm_xdl_bias_relu_add/include/device_gemm_xdl_two_extra_source_reduce.hpp
View File

@@ -1,537 +0,0 @@
#ifndef DEVICE_GEMM_XDL_TWO_EXTRA_SOURCE_REDUCE_HPP
#define DEVICE_GEMM_XDL_TWO_EXTRA_SOURCE_REDUCE_HPP
#include <iostream>
#include <sstream>
#include "device.hpp"
#include "device_base.hpp"
#include "device_gemm.hpp"
#include "common_header.hpp"
#include "tensor_layout.hpp"
#include "tensor_descriptor.hpp"
#include "tensor_descriptor_helper.hpp"
#include "gridwise_gemm_xdlops_v2r5.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
template <typename ADataType,
typename BDataType,
typename CDataType,
typename AccDataType,
typename ALayout,
typename BLayout,
typename CLayout,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation,
ck::index_t BlockSize,
ck::index_t MPerBlock,
ck::index_t NPerBlock,
ck::index_t K0PerBlock,
ck::index_t K1,
ck::index_t MPerXDL,
ck::index_t NPerXDL,
ck::index_t MXdlPerWave,
ck::index_t NXdlPerWave,
typename ABlockTransferThreadClusterLengths_K0_M_K1,
typename ABlockTransferThreadClusterArrangeOrder,
typename ABlockTransferSrcAccessOrder,
ck::index_t ABlockTransferSrcVectorDim,
ck::index_t ABlockTransferSrcScalarPerVector,
ck::index_t ABlockTransferDstScalarPerVector_K1,
bool ABlockLdsAddExtraM,
typename BBlockTransferThreadClusterLengths_K0_N_K1,
typename BBlockTransferThreadClusterArrangeOrder,
typename BBlockTransferSrcAccessOrder,
ck::index_t BBlockTransferSrcVectorDim,
ck::index_t BBlockTransferSrcScalarPerVector,
ck::index_t BBlockTransferDstScalarPerVector_K1,
bool BBlockLdsAddExtraN,
ck::index_t CThreadTransferSrcDstVectorDim,
ck::index_t CThreadTransferDstScalarPerVector>
struct DeviceGemmXdl_two_extra_source_reduce : public BaseOperator
{
static constexpr auto I0 = Number<0>{};
static constexpr auto I1 = Number<1>{};
static constexpr auto I2 = Number<2>{};
static constexpr auto K1Number = Number<K1>{};
static auto MakeAGridDescriptor_K0_M_K1(index_t M, index_t K, index_t StrideA)
{
assert(K % K1 == 0);
const index_t K0 = K / K1;
const auto a_grid_desc_m_k = [&]() {
if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(StrideA, I1));
}
else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ALayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(I1, StrideA));
}
}();
const auto a_grid_desc_k0_m_k1 =
transform_tensor_descriptor(a_grid_desc_m_k,
make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
make_pass_through_transform(M)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
return a_grid_desc_k0_m_k1;
}
static auto MakeBGridDescriptor_K0_N_K1(index_t K, index_t N, index_t StrideB)
{
assert(K % K1 == 0);
const index_t K0 = K / K1;
const auto b_grid_desc_k_n = [&]() {
if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(StrideB, I1));
}
else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(I1, StrideB));
}
}();
const auto b_grid_desc_k0_n_k1 =
transform_tensor_descriptor(b_grid_desc_k_n,
make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
make_pass_through_transform(N)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
return b_grid_desc_k0_n_k1;
}
static auto MakeCGridDescriptor_M_N(index_t M, index_t N, index_t StrideC)
{
if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(StrideC, I1));
}
else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, CLayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(I1, StrideC));
}
}
using AGridDesc_K0_M_K1 = decltype(MakeAGridDescriptor_K0_M_K1(1, 1, 1));
using BGridDesc_K0_N_K1 = decltype(MakeBGridDescriptor_K0_N_K1(1, 1, 1));
using CGridDesc_M_N = decltype(MakeCGridDescriptor_M_N(1, 1, 1));
using C0GridDesc_M_N = decltype(MakeCGridDescriptor_M_N(1, 1, 1));
// hardcoding
// TODO: fix this
using C1GridDesc_M_N =
decltype(make_naive_tensor_descriptor(make_tuple(1, 1), make_tuple(I1, I0)));
// GridwiseGemm
using GridwiseGemm = GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r5<
BlockSize,
ADataType, // TODO: distinguish A/B datatype
AccDataType,
CDataType,
InMemoryDataOperationEnum_t::Set,
AGridDesc_K0_M_K1,
BGridDesc_K0_N_K1,
CGridDesc_M_N,
C0GridDesc_M_N,
C1GridDesc_M_N,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation,
MPerBlock,
NPerBlock,
K0PerBlock,
MPerXDL,
NPerXDL,
K1,
MXdlPerWave,
NXdlPerWave,
ABlockTransferThreadClusterLengths_K0_M_K1,
ABlockTransferThreadClusterArrangeOrder,
ABlockTransferSrcAccessOrder,
ABlockTransferSrcVectorDim,
ABlockTransferSrcScalarPerVector,
ABlockTransferDstScalarPerVector_K1,
false, // AThreadTransferSrcResetCoordinateAfterRun,
ABlockLdsAddExtraM,
BBlockTransferThreadClusterLengths_K0_N_K1,
BBlockTransferThreadClusterArrangeOrder,
BBlockTransferSrcAccessOrder,
BBlockTransferSrcVectorDim,
BBlockTransferSrcScalarPerVector,
BBlockTransferDstScalarPerVector_K1,
false, // BThreadTransferSrcResetCoordinateAfterRun,
BBlockLdsAddExtraN,
Sequence<0, 2, 4, 5, 6, 1, 3, 7>, // CThreadTransferSrcDstAccessOrder,
CThreadTransferSrcDstVectorDim,
CThreadTransferDstScalarPerVector>;
using CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2 =
decltype(GridwiseGemm::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(CGridDesc_M_N{}));
using C0GridDesc_M0_N0_M1_N1_M2_M3_M4_N2 =
decltype(GridwiseGemm::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(C0GridDesc_M_N{}));
using C1GridDesc_M0_N0_M1_N1_M2_M3_M4_N2 =
decltype(GridwiseGemm::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(C1GridDesc_M_N{}));
using Block2CTileMap = decltype(GridwiseGemm::MakeBlock2CTileMap(CGridDesc_M_N{}, 1, 1));
// Argument
struct Argument : public BaseArgument
{
Argument(const ADataType* p_a_grid,
const BDataType* p_b_grid,
CDataType* p_c_grid,
const CDataType* p_c0_grid,
const CDataType* p_c1_grid,
index_t M,
index_t N,
index_t K,
index_t StrideA,
index_t StrideB,
index_t StrideC,
index_t M01,
index_t N01,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op)
: p_a_grid_{p_a_grid},
p_b_grid_{p_b_grid},
p_c_grid_{p_c_grid},
p_c0_grid_{p_c0_grid},
p_c1_grid_{p_c1_grid},
a_grid_desc_k0_m_k1_{},
b_grid_desc_k0_n_k1_{},
c_grid_desc_m_n_{},
c0_grid_desc_m_n_{},
c1_grid_desc_m_n_{},
c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_{},
c0_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_{},
c1_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_{},
block_2_ctile_map_{},
M01_{M01},
N01_{N01},
a_element_op_{a_element_op},
b_element_op_{b_element_op},
c_element_op_{c_element_op}
{
a_grid_desc_k0_m_k1_ =
DeviceGemmXdl_two_extra_source_reduce::MakeAGridDescriptor_K0_M_K1(M, K, StrideA);
b_grid_desc_k0_n_k1_ =
DeviceGemmXdl_two_extra_source_reduce::MakeBGridDescriptor_K0_N_K1(K, N, StrideB);
c_grid_desc_m_n_ =
DeviceGemmXdl_two_extra_source_reduce::MakeCGridDescriptor_M_N(M, N, StrideC);
// assume C0 has same layout as C
// TODO: fix this
c0_grid_desc_m_n_ =
DeviceGemmXdl_two_extra_source_reduce::MakeCGridDescriptor_M_N(M, N, StrideC);
// hardcoding C1 layout
// TODO: fix this
c1_grid_desc_m_n_ = make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(I1, I0));
if(GridwiseGemm::CheckValidity(
a_grid_desc_k0_m_k1_, b_grid_desc_k0_n_k1_, c_grid_desc_m_n_, M01_, N01_))
{
c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_ =
GridwiseGemm::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_grid_desc_m_n_);
c0_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_ =
GridwiseGemm::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c0_grid_desc_m_n_);
c1_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_ =
GridwiseGemm::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c1_grid_desc_m_n_);
block_2_ctile_map_ = GridwiseGemm::MakeBlock2CTileMap(c_grid_desc_m_n_, M01, N01);
}
}
// private:
const ADataType* p_a_grid_;
const BDataType* p_b_grid_;
CDataType* p_c_grid_;
const CDataType* p_c0_grid_;
const CDataType* p_c1_grid_;
AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1_;
BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1_;
CGridDesc_M_N c_grid_desc_m_n_;
C0GridDesc_M_N c0_grid_desc_m_n_;
C1GridDesc_M_N c1_grid_desc_m_n_;
CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2 c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_;
C0GridDesc_M0_N0_M1_N1_M2_M3_M4_N2 c0_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_;
C1GridDesc_M0_N0_M1_N1_M2_M3_M4_N2 c1_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_;
Block2CTileMap block_2_ctile_map_;
index_t M01_;
index_t N01_;
AElementwiseOperation a_element_op_;
BElementwiseOperation b_element_op_;
CElementwiseOperation c_element_op_;
};
// Invoker
struct Invoker : public BaseInvoker
{
using Argument = DeviceGemmXdl_two_extra_source_reduce::Argument;
float Run(const Argument& arg, int nrepeat = 1)
{
{
std::cout << "arg.a_grid_desc_k0_m_k1_{" << arg.a_grid_desc_k0_m_k1_.GetLength(I0)
<< ", " << arg.a_grid_desc_k0_m_k1_.GetLength(I1) << ", "
<< arg.a_grid_desc_k0_m_k1_.GetLength(I2) << "}" << std::endl;
std::cout << "arg.b_grid_desc_k0_n_k1_{" << arg.b_grid_desc_k0_n_k1_.GetLength(I0)
<< ", " << arg.b_grid_desc_k0_n_k1_.GetLength(I1) << ", "
<< arg.b_grid_desc_k0_n_k1_.GetLength(I2) << "}" << std::endl;
std::cout << "arg.c_grid_desc_m_n_{ " << arg.c_grid_desc_m_n_.GetLength(I0) << ", "
<< arg.c_grid_desc_m_n_.GetLength(I1) << "}" << std::endl;
std::cout << "arg.c0_grid_desc_m_n_{ " << arg.c0_grid_desc_m_n_.GetLength(I0)
<< ", " << arg.c0_grid_desc_m_n_.GetLength(I1) << "}" << std::endl;
std::cout << "arg.c1_grid_desc_m_n_{ " << arg.c1_grid_desc_m_n_.GetLength(I0)
<< ", " << arg.c1_grid_desc_m_n_.GetLength(I1) << "}" << std::endl;
}
if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_,
arg.b_grid_desc_k0_n_k1_,
arg.c_grid_desc_m_n_,
arg.M01_,
arg.N01_))
{
throw std::runtime_error(
"wrong! GridwiseGemm_km_kn_m0m1n0n1_xdlops_v2r5 has invalid setting");
}
const index_t grid_size = GridwiseGemm::CalculateGridSize(arg.c_grid_desc_m_n_);
const auto K0 = arg.a_grid_desc_k0_m_k1_.GetLength(I0);
const bool has_main_k0_block_loop = GridwiseGemm::CalculateHasMainK0BlockLoop(K0);
float ave_time = 0;
if(has_main_k0_block_loop)
{
const auto kernel = kernel_gemm_xdlops_v2r5<
GridwiseGemm,
ADataType, // TODO: distiguish A/B datatype
CDataType,
remove_reference_t<DeviceGemmXdl_two_extra_source_reduce::AGridDesc_K0_M_K1>,
remove_reference_t<DeviceGemmXdl_two_extra_source_reduce::BGridDesc_K0_N_K1>,
remove_reference_t<
DeviceGemmXdl_two_extra_source_reduce::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>,
remove_reference_t<
DeviceGemmXdl_two_extra_source_reduce::C0GridDesc_M0_N0_M1_N1_M2_M3_M4_N2>,
remove_reference_t<
DeviceGemmXdl_two_extra_source_reduce::C1GridDesc_M0_N0_M1_N1_M2_M3_M4_N2>,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation,
remove_reference_t<DeviceGemmXdl_two_extra_source_reduce::Block2CTileMap>,
true>;
ave_time = launch_and_time_kernel(kernel,
nrepeat,
dim3(grid_size),
dim3(BlockSize),
0,
arg.p_a_grid_,
arg.p_b_grid_,
arg.p_c_grid_,
arg.p_c0_grid_,
arg.p_c1_grid_,
arg.a_grid_desc_k0_m_k1_,
arg.b_grid_desc_k0_n_k1_,
arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_,
arg.c0_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_,
arg.c1_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_,
arg.a_element_op_,
arg.b_element_op_,
arg.c_element_op_,
arg.block_2_ctile_map_);
}
else
{
const auto kernel = kernel_gemm_xdlops_v2r5<
GridwiseGemm,
ADataType, // TODO: distiguish A/B datatype
CDataType,
remove_reference_t<DeviceGemmXdl_two_extra_source_reduce::AGridDesc_K0_M_K1>,
remove_reference_t<DeviceGemmXdl_two_extra_source_reduce::BGridDesc_K0_N_K1>,
remove_reference_t<
DeviceGemmXdl_two_extra_source_reduce::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>,
remove_reference_t<
DeviceGemmXdl_two_extra_source_reduce::C0GridDesc_M0_N0_M1_N1_M2_M3_M4_N2>,
remove_reference_t<
DeviceGemmXdl_two_extra_source_reduce::C1GridDesc_M0_N0_M1_N1_M2_M3_M4_N2>,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation,
remove_reference_t<DeviceGemmXdl_two_extra_source_reduce::Block2CTileMap>,
false>;
ave_time = launch_and_time_kernel(kernel,
nrepeat,
dim3(grid_size),
dim3(BlockSize),
0,
arg.p_a_grid_,
arg.p_b_grid_,
arg.p_c_grid_,
arg.p_c0_grid_,
arg.p_c1_grid_,
arg.a_grid_desc_k0_m_k1_,
arg.b_grid_desc_k0_n_k1_,
arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_,
arg.c0_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_,
arg.c1_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_,
arg.a_element_op_,
arg.b_element_op_,
arg.c_element_op_,
arg.block_2_ctile_map_);
}
return ave_time;
}
// polymorphic
float Run(const BaseArgument* p_arg, int nrepeat = 1) override
{
return Run(*dynamic_cast<const Argument*>(p_arg), nrepeat);
}
};
static constexpr bool IsValidCompilationParameter()
{
// TODO: properly implement this check
return true;
}
static bool IsSupportedArgument(const Argument& arg)
{
return GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_,
arg.b_grid_desc_k0_n_k1_,
arg.c_grid_desc_m_n_,
arg.M01_,
arg.N01_);
}
// polymorphic
bool IsSupportedArgument(const BaseArgument* p_arg) override
{
return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
}
static auto MakeArgument(const ADataType* p_a,
const BDataType* p_b,
CDataType* p_c,
const CDataType* p_c0,
const CDataType* p_c1,
index_t M,
index_t N,
index_t K,
index_t StrideA,
index_t StrideB,
index_t StrideC,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op)
{
return Argument{p_a,
p_b,
p_c,
p_c0,
p_c1,
M,
N,
K,
StrideA,
StrideB,
StrideC,
1,
1,
a_element_op,
b_element_op,
c_element_op};
}
static auto MakeInvoker() { return Invoker{}; }
// polymorphic
std::unique_ptr<BaseArgument> MakeArgumentPointer(const void* p_a,
const void* p_b,
void* p_c,
const void* p_c0,
const void* p_c1,
index_t M,
index_t N,
index_t K,
index_t StrideA,
index_t StrideB,
index_t StrideC,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op)
{
return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
static_cast<const BDataType*>(p_b),
static_cast<CDataType*>(p_c),
static_cast<const CDataType*>(p_c0),
static_cast<const CDataType*>(p_c1),
M,
N,
K,
StrideA,
StrideB,
StrideC,
1,
1,
a_element_op,
b_element_op,
c_element_op);
}
// polymorphic
std::unique_ptr<BaseInvoker> MakeInvokerPointer()
{
return std::make_unique<Invoker>(Invoker{});
}
std::string GetTypeString() const override
{
auto str = std::stringstream();
// clang-format off
str << "DeviceGemmXdl_two_extra_source_reduce"
<< "<"
<< BlockSize << ", "
<< MPerBlock << ", "
<< NPerBlock << ", "
<< K0PerBlock
<< ">";
// clang-format on
return str.str();
}
};
} // namespace device
} // namespace tensor_operation
} // namespace ck
#endif

36
example/4_conv2d_fwd_xdl/conv2d_fwd_xdl.cpp
View File

@@ -74,13 +74,8 @@ using DeviceConvFwdInstance = ck::tensor_operation::device::
8>; // CBlockTransferScalarPerVector_NWaveNPerXdl
// clang-format on
using ReferenceConvFwdInstance = ck::tensor_operation::host::ReferenceConvFwd<InDataType,
WeiDataType,
OutDataType,
AccDataType,
InElementOp,
WeiElementOp,
OutElementOp>;
using ReferenceConvFwdInstance = ck::tensor_operation::host::
ReferenceConvFwd<InDataType, WeiDataType, OutDataType, InElementOp, WeiElementOp, OutElementOp>;
int main(int argc, char* argv[])
{
@@ -254,20 +249,21 @@ int main(int argc, char* argv[])
if(do_verification)
{
auto refConv = ReferenceConvFwdInstance{};
auto refInvoker = refConv.MakeInvoker();
auto ref_conv = ReferenceConvFwdInstance{};
auto ref_invoker = ref_conv.MakeInvoker();
auto refArgument = refConv.MakeArgument(in_n_c_hi_wi,
wei_k_c_y_x,
out_n_k_ho_wo_host_result,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
InElementOp{},
WeiElementOp{},
OutElementOp{});
refInvoker.Run(refArgument);
auto ref_argument = ref_conv.MakeArgument(in_n_c_hi_wi,
wei_k_c_y_x,
out_n_k_ho_wo_host_result,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
InElementOp{},
WeiElementOp{},
OutElementOp{});
ref_invoker.Run(ref_argument);
out_device_buf.FromDevice(out_n_k_ho_wo_device_result.mData.data());

29
example/5_conv2d_fwd_xdl_bias_relu/conv2d_fwd_xdl_bias_relu.cpp
View File

@@ -81,7 +81,6 @@ using ReferenceConvFwdInstance =
ck::tensor_operation::host::ReferenceConvFwd_Bias_Activation<InDataType,
WeiDataType,
OutDataType,
AccDataType,
InElementOp,
WeiElementOp,
OutElementOp>;
@@ -268,21 +267,21 @@ int main(int argc, char* argv[])
if(do_verification)
{
auto refConv = ReferenceConvFwdInstance{};
auto refInvoker = refConv.MakeInvoker();
auto ref_conv = ReferenceConvFwdInstance{};
auto ref_invoker = ref_conv.MakeInvoker();
auto refArgument = refConv.MakeArgument(in_n_c_hi_wi,
wei_k_c_y_x,
out_n_k_ho_wo_host_result,
bias_k,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
InElementOp{},
WeiElementOp{},
OutElementOp{});
refInvoker.Run(refArgument);
auto ref_argument = ref_conv.MakeArgument(in_n_c_hi_wi,
wei_k_c_y_x,
out_n_k_ho_wo_host_result,
bias_k,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
InElementOp{},
WeiElementOp{},
OutElementOp{});
ref_invoker.Run(ref_argument);
out_device_buf.FromDevice(out_n_k_ho_wo_device_result.mData.data());

42
example/6_conv2d_fwd_xdl_bias_relu_add/conv2d_fwd_xdl_bias_relu_add.cpp
View File

@@ -78,7 +78,6 @@ using ReferenceConvFwdInstance =
ck::tensor_operation::host::ReferenceConvFwd_Bias_Activation_Add<InDataType,
WeiDataType,
OutDataType,
AccDataType,
InElementOp,
WeiElementOp,
OutElementOp>;
@@ -228,6 +227,10 @@ int main(int argc, char* argv[])
bias_device_buf.ToDevice(bias_k.mData.data());
resi_device_buf.ToDevice(resi_n_k_ho_wo.mData.data());
const auto in_element_op = InElementOp{};
const auto wei_element_op = WeiElementOp{};
const auto out_element_op = OutElementOp{};
auto conv = DeviceConvFwdInstance{};
auto invoker = conv.MakeInvoker();
auto argument =
@@ -246,9 +249,9 @@ int main(int argc, char* argv[])
conv_filter_dilations,
input_left_pads,
input_right_pads,
InElementOp{},
WeiElementOp{},
OutElementOp{});
in_element_op,
wei_element_op,
out_element_op);
if(!conv.IsSupportedArgument(argument))
{
@@ -275,22 +278,23 @@ int main(int argc, char* argv[])
if(do_verification)
{
auto refConv = ReferenceConvFwdInstance{};
auto refInvoker = refConv.MakeInvoker();
auto ref_conv = ReferenceConvFwdInstance{};
auto ref_invoker = ref_conv.MakeInvoker();
auto refArgument = refConv.MakeArgument(in_n_c_hi_wi,
wei_k_c_y_x,
out_n_k_ho_wo_host_result,
bias_k,
resi_n_k_ho_wo,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
InElementOp{},
WeiElementOp{},
OutElementOp{});
refInvoker.Run(refArgument);
auto ref_argument = ref_conv.MakeArgument(in_n_c_hi_wi,
wei_k_c_y_x,
out_n_k_ho_wo_host_result,
bias_k,
resi_n_k_ho_wo,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
in_element_op,
wei_element_op,
out_element_op);
ref_invoker.Run(ref_argument);
out_device_buf.FromDevice(out_n_k_ho_wo_device_result.mData.data());

18
example/7_conv2d_fwd_xdl_bias_relu_atomic_add/conv2d_fwd_xdl_bias_relu_atomic_add.cpp
View File

@@ -65,7 +65,8 @@ void host_reference_calculation(const Tensor<TIn>& in_n_c_hi_wi,
const OutElementOp& out_element_op)
{
auto f_nchw = [&](auto n, auto k, auto ho, auto wo) {
double v = 0;
float v_acc = 0;
for(int c = 0; c < wei_k_c_y_x.mDesc.GetLengths()[1]; ++c)
{
for(int y = 0; y < wei_k_c_y_x.mDesc.GetLengths()[2]; ++y)
@@ -77,14 +78,23 @@ void host_reference_calculation(const Tensor<TIn>& in_n_c_hi_wi,
if(hi >= 0 && hi < in_n_c_hi_wi.mDesc.GetLengths()[2] && wi >= 0 &&
wi < in_n_c_hi_wi.mDesc.GetLengths()[3])
{
v += in_element_op(static_cast<const double>(in_n_c_hi_wi(n, c, hi, wi))) *
wei_element_op(static_cast<const double>(wei_k_c_y_x(k, c, y, x)));
float v_in;
float v_wei;
in_element_op(v_in, static_cast<const float>(in_n_c_hi_wi(n, c, hi, wi)));
wei_element_op(v_wei, static_cast<const float>(wei_k_c_y_x(k, c, y, x)));
v_acc += v_in * v_wei;
}
}
}
}
out_n_k_ho_wo(n, k, ho, wo) += out_element_op(v, bias_k(k));
float v_out;
out_element_op(v_out, v_acc, static_cast<float>(bias_k(k)));
out_n_k_ho_wo(n, k, ho, wo) += v_out;
};
make_ParallelTensorFunctor(f_nchw,

5
example/CMakeLists.txt
View File

@@ -2,8 +2,8 @@ include_directories(BEFORE
${PROJECT_SOURCE_DIR}
${PROJECT_SOURCE_DIR}/host/host_tensor/include
${PROJECT_SOURCE_DIR}/host/device/include
${PROJECT_SOURCE_DIR}/host/include
${PROJECT_SOURCE_DIR}/device_operation/include
${PROJECT_SOURCE_DIR}/reference_operation/include
${PROJECT_SOURCE_DIR}/composable_kernel/include
${PROJECT_SOURCE_DIR}/composable_kernel/include/utility
${PROJECT_SOURCE_DIR}/composable_kernel/include/tensor_description
@@ -13,6 +13,7 @@ include_directories(BEFORE
)
set(GEMM_XDL_SOURCE 1_gemm_xdl/gemm_xdl.cpp)
set(GEMM_XDL_BIAS_RELU_SOURCE 2_gemm_xdl_bias_relu/gemm_xdl_bias_relu.cpp)
set(GEMM_XDL_BIAS_RELU_ADD_SOURCE 3_gemm_xdl_bias_relu_add/gemm_xdl_bias_relu_add.cpp)
set(CONV2D_FWD_XDL_SOURCE 4_conv2d_fwd_xdl/conv2d_fwd_xdl.cpp)
set(CONV2D_FWD_XDL_BIAS_RELU_SOURCE 5_conv2d_fwd_xdl_bias_relu/conv2d_fwd_xdl_bias_relu.cpp)
@@ -20,6 +21,7 @@ set(CONV2D_FWD_XDL_BIAS_RELU_ADD_SOURCE 6_conv2d_fwd_xdl_bias_relu_add/conv2d_fw
set(CONV2D_FWD_XDL_BIAS_RELU_ATOMIC_ADD_SOURCE 7_conv2d_fwd_xdl_bias_relu_atomic_add/conv2d_fwd_xdl_bias_relu_atomic_add.cpp)
add_executable(gemm_xdl ${GEMM_XDL_SOURCE})
add_executable(gemm_xdl_bias_relu ${GEMM_XDL_BIAS_RELU_SOURCE})
add_executable(gemm_xdl_bias_relu_add ${GEMM_XDL_BIAS_RELU_ADD_SOURCE})
add_executable(conv2d_fwd_xdl ${CONV2D_FWD_XDL_SOURCE})
add_executable(conv2d_fwd_xdl_bias_relu ${CONV2D_FWD_XDL_BIAS_RELU_SOURCE})
@@ -27,6 +29,7 @@ add_executable(conv2d_fwd_xdl_bias_relu_add ${CONV2D_FWD_XDL_BIAS_RELU_ADD_SOURC
add_executable(conv2d_fwd_xdl_bias_relu_atomic_add ${CONV2D_FWD_XDL_BIAS_RELU_ATOMIC_ADD_SOURCE})
target_link_libraries(gemm_xdl PRIVATE host_tensor)
target_link_libraries(gemm_xdl_bias_relu PRIVATE host_tensor)
target_link_libraries(gemm_xdl_bias_relu_add PRIVATE host_tensor)
target_link_libraries(conv2d_fwd_xdl PRIVATE host_tensor)
target_link_libraries(conv2d_fwd_xdl_bias_relu PRIVATE host_tensor)