Skip to content

GitLab

Unverified Commit 5eb2f81e authored by zjing14's avatar zjing14 Committed by GitHub
Browse files

Merge branch 'develop' into lib_gemm_softmax_gemm_type

parents e7a56ac6 ed3a2e52
Show whitespace changes
Inline Side-by-side
Showing with 1064 additions and 515 deletions
client_example/18_groupnorm/CMakeLists.txt 0 → 100644
View file @ 5eb2f81e
add_executable(client_groupnorm_swish groupnorm_swish.cpp)
target_link_libraries(client_groupnorm_swish PRIVATE composable_kernel::device_operations)
client_example/18_groupnorm/groupnorm_swish.cpp 0 → 100644
View file @ 5eb2f81e
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iomanip>
#include <vector>
#include <iostream>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_normalization.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/gpu/normalization_swish.hpp"
using XDataType = ck::half_t;
using GammaDataType = ck::half_t;
using BetaDataType = ck::half_t;
using YDataType = ck::half_t;
using ComputeDataType = float;
using Swish = ck::tensor_operation::element_wise::Swish;
constexpr int Rank = 5;
constexpr int NumReduceDim = 3;
struct SimpleDeviceMem
{
SimpleDeviceMem() = delete;
SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
{
(void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
}
void* GetDeviceBuffer() { return p_mem_; }
~SimpleDeviceMem() { (void)hipFree(p_mem_); }
void* p_mem_;
};
int main(int argc, char* argv[])
{
ck::index_t N = 32;
ck::index_t H = 16;
ck::index_t W = 16;
ck::index_t G = 64;
ck::index_t C = 128;
std::size_t xy_size = N * H * W * G * C;
std::size_t gamma_beta_size = G * C;
std::vector<ck::index_t> xy_strides = {H * W * G * C, W * G * C, G * C, C, 1};
std::vector<ck::index_t> gamma_beta_strides = {0, 0, 0, C, 1};
SimpleDeviceMem x_device_buf(sizeof(XDataType) * xy_size);
SimpleDeviceMem gamma_device_buf(sizeof(GammaDataType) * gamma_beta_size);
SimpleDeviceMem beta_device_buf(sizeof(BetaDataType) * gamma_beta_size);
SimpleDeviceMem y_device_buf(sizeof(YDataType) * xy_size);
using DeviceOp = ck::tensor_operation::device::DeviceNormalization<XDataType,
GammaDataType,
BetaDataType,
ComputeDataType,
YDataType,
Swish,
Rank,
NumReduceDim>;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
std::string best_op_name;
bool found = false;
int best_op_id = -1;
float best_ave_time = std::numeric_limits<float>::max();
float best_gb_per_sec = 0;
// profile device operation instances
std::cout << "Run all instances and do timing" << std::endl;
for(int i = 0; i < op_ptrs.size(); ++i)
{
auto& op_ptr = op_ptrs[i];
auto argument_ptr = op_ptr->MakeArgumentPointer({N, H, W, G, C}, // lengths
xy_strides, // xStrides
gamma_beta_strides, // gammaStrides
gamma_beta_strides, // betaStrides
xy_strides, // yStrides
{1, 2, 4}, // reduceDims
1e-6,
x_device_buf.GetDeviceBuffer(),
gamma_device_buf.GetDeviceBuffer(),
beta_device_buf.GetDeviceBuffer(),
y_device_buf.GetDeviceBuffer(),
nullptr,
nullptr,
Swish{});
auto invoker_ptr = op_ptr->MakeInvokerPointer();
std::string op_name = op_ptr->GetTypeString();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
std::size_t num_byte =
sizeof(XDataType) * xy_size + sizeof(GammaDataType) * gamma_beta_size +
sizeof(BetaDataType) * gamma_beta_size + sizeof(YDataType) * xy_size;
float gb_per_sec = num_byte / 1.E6 / ave_time;
std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << gb_per_sec << " GB/s, "
<< op_name << std::endl;
if(ave_time < best_ave_time)
{
found = true;
best_op_id = i;
best_op_name = op_name;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
}
}
else
{
std::cout << op_name << " does not support this problem" << std::endl;
}
}
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_gb_per_sec << " GB/s, "
<< best_op_name << std::endl;
// run the best intance
{
auto& op_ptr = op_ptrs[best_op_id];
std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
<< std::endl;
auto argument_ptr = op_ptr->MakeArgumentPointer({N, H, W, G, C}, // lengths
xy_strides, // xStrides
gamma_beta_strides, // gammaStrides
gamma_beta_strides, // betaStrides
xy_strides, // yStrides
{1, 2, 4}, // reduceDims
1e-6,
x_device_buf.GetDeviceBuffer(),
gamma_device_buf.GetDeviceBuffer(),
beta_device_buf.GetDeviceBuffer(),
y_device_buf.GetDeviceBuffer(),
nullptr,
nullptr,
Swish{});
auto invoker_ptr = op_ptr->MakeInvokerPointer();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
}
std::cout << "Done" << std::endl;
}
return 0;
}
example/42_groupnorm/CMakeLists.txt
View file @ 5eb2f81e
add_example_executable(example_groupnorm_sigmoid_fp16 groupnorm_sigmoid_fp16.cpp) add_example_executable(example_groupnorm_sigmoid_mul_fp16 groupnorm_sigmoid_mul_fp16.cpp)
add_example_executable(example_groupnorm_swish_fp16 groupnorm_swish_fp16.cpp)
example/42_groupnorm/common.hpp 0 → 100644
View file @ 5eb2f81e
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include <getopt.h>
#include "ck/ck.hpp"
#include "ck/utility/reduction_enums.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_normalization_impl.hpp"
#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
#include "ck/library/utility/fill.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_common_util.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_groupnorm.hpp"
example/42_groupnorm/groupnorm_sigmoid_mul_fp16.cpp 0 → 100644
View file @ 5eb2f81e
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
constexpr int Rank = 5;
constexpr int NumReduceDim = 3;
using XDataType = ck::half_t;
using GammaDataType = ck::half_t;
using BetaDataType = ck::half_t;
using YDataType = ck::half_t;
using ComputeDataType = float;
struct YElementOp
{
template <typename T>
__host__ __device__ void operator()(T& y, const T& x) const
{
static_assert(ck::is_same<T, float>::value || ck::is_same<T, double>::value ||
ck::is_same<T, ck::half_t>::value,
"Data type is not supported by this operation!");
T a;
ck::tensor_operation::element_wise::Sigmoid{}(a, x);
y = x * a;
};
};
using DeviceInstance =
ck::tensor_operation::device::DeviceNormalizationImpl<XDataType,
GammaDataType,
BetaDataType,
ComputeDataType,
YDataType,
YElementOp,
Rank,
NumReduceDim,
1024, // BlockSize
1, // ClusterM
1024, // ClusterK
1, // SliceM
32, // SliceK
1, // SrcVecDim (0=M, 1=K)
2, // SrcScalarPerVector
1, // GammaVecDim (0=M, 1=K)
2, // GammaScalarPerVector
1, // BetaVecDim (0=M, 1=K)
2, // BetaScalarPerVector
2>; // OutScalarPerVector
#include "run_groupnorm_example.inc"
int main(int argc, char* argv[]) { run_groupnorm_example(argc, argv); }
example/42_groupnorm/groupnorm_swish_fp16.cpp 0 → 100644
View file @ 5eb2f81e
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
constexpr int Rank = 5;
constexpr int NumReduceDim = 3;
using XDataType = ck::half_t;
using GammaDataType = ck::half_t;
using BetaDataType = ck::half_t;
using YDataType = ck::half_t;
using ComputeDataType = float;
using YElementOp = ck::tensor_operation::element_wise::Swish;
using DeviceInstance =
ck::tensor_operation::device::DeviceNormalizationImpl<XDataType,
GammaDataType,
BetaDataType,
ComputeDataType,
YDataType,
YElementOp,
Rank,
NumReduceDim,
1024, // BlockSize
1, // ClusterM
1024, // ClusterK
1, // SliceM
32, // SliceK
1, // SrcVecDim (0=M, 1=K)
2, // SrcScalarPerVector
1, // GammaVecDim (0=M, 1=K)
2, // GammaScalarPerVector
1, // BetaVecDim (0=M, 1=K)
2, // BetaScalarPerVector
2>; // OutScalarPerVector
#include "run_groupnorm_example.inc"
int main(int argc, char* argv[]) { run_groupnorm_example(argc, argv); }
example/42_groupnorm/groupnorm_sigmoid_fp16.cpp → example/42_groupnorm/run_groupnorm_example.inc
View file @ 5eb2f81e
// SPDX-License-Identifier: MIT // SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. // Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream> #pragma once
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include <getopt.h>
#include "ck/ck.hpp"
#include "ck/utility/reduction_enums.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_normalization_impl.hpp"
#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
#include "ck/library/utility/fill.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_common_util.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_groupnorm.hpp"
constexpr int Rank = 5;
constexpr int NumReduceDim = 3;
using XDataType = ck::half_t;
using GammaDataType = ck::half_t;
using BetaDataType = ck::half_t;
using YDataType = ck::half_t;
using ComputeDataType = float;
struct YElementOp
{
template <typename T>
__host__ __device__ void operator()(T& y, const T& x) const
{
static_assert(ck::is_same<T, float>::value || ck::is_same<T, double>::value ||
ck::is_same<T, ck::half_t>::value,
"Data type is not supported by this operation!");
T a;
ck::tensor_operation::element_wise::Sigmoid{}(a, x);
y = x * a;
};
};
using DeviceInstance = int run_groupnorm_example(int argc, char* argv[])
ck::tensor_operation::device::DeviceNormalizationImpl<XDataType,
GammaDataType,
BetaDataType,
ComputeDataType,
YDataType,
YElementOp,
Rank,
NumReduceDim,
1024, // BlockSize
1, // ClusterM
1024, // ClusterK
1, // SliceM
32, // SliceK
1, // SrcVecDim (0=M, 1=K)
2, // SrcScalarPerVector
1, // GammaVecDim (0=M, 1=K)
2, // GammaScalarPerVector
1, // BetaVecDim (0=M, 1=K)
2, // BetaScalarPerVector
2>; // OutScalarPerVector
int main(int argc, char* argv[])
{ {
ck::index_t N = 2; ck::index_t N = 32;
ck::index_t H = 32; ck::index_t H = 16;
ck::index_t W = 32; ck::index_t W = 16;
ck::index_t G = 32; ck::index_t G = 64;
ck::index_t C = 30; ck::index_t C = 128;
if(argc == 1) if(argc == 1)
{ {
......
include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp
View file @ 5eb2f81e
...@@ -73,18 +73,199 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout, ...@@ -73,18 +73,199 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout,
static constexpr auto I2 = Number<2>{}; static constexpr auto I2 = Number<2>{};
static constexpr auto I3 = Number<3>{}; static constexpr auto I3 = Number<3>{};
static constexpr auto K1Number = Number<K1>{};
static auto
MakeAGridDescriptor_KBatch_K0_M_K1(index_t M, index_t K, index_t StrideA, int KBatch, int KPad)
{
assert(KPad % (K1 * KBatch) == 0);
const index_t K0 = KPad / (K1 * KBatch);
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_m_kpad = transform_tensor_descriptor(
a_grid_desc_m_k,
make_tuple(make_pass_through_transform(M), make_right_pad_transform(K, KPad - K)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
if constexpr(GemmSpec == GemmSpecialization::MNPadding)
{
const auto PadM = (MPerBlock - M % MPerBlock) % MPerBlock;
return transform_tensor_descriptor(
a_grid_desc_m_kpad,
make_tuple(make_unmerge_transform(make_tuple(KBatch, K0, K1Number)),
make_right_pad_transform(M, PadM)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
}
else
{
return transform_tensor_descriptor(
a_grid_desc_m_kpad,
make_tuple(make_unmerge_transform(make_tuple(KBatch, K0, K1Number)),
make_pass_through_transform(M)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
}
}
static auto
MakeBGridDescriptor_KBatch_K0_N_K1(index_t K, index_t N, index_t StrideB, int KBatch, int KPad)
{
assert(KPad % (K1 * KBatch) == 0);
const index_t K0 = KPad / (K1 * KBatch);
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_kpad_n = transform_tensor_descriptor(
b_grid_desc_k_n,
make_tuple(make_right_pad_transform(K, KPad - K), make_pass_through_transform(N)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
if constexpr(GemmSpec == GemmSpecialization::MNPadding)
{
const auto PadN = (NPerBlock - N % NPerBlock) % NPerBlock;
return transform_tensor_descriptor(
b_grid_desc_kpad_n,
make_tuple(make_unmerge_transform(make_tuple(KBatch, K0, K1Number)),
make_right_pad_transform(N, PadN)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
}
else
{
return transform_tensor_descriptor(
b_grid_desc_kpad_n,
make_tuple(make_unmerge_transform(make_tuple(KBatch, K0, K1Number)),
make_pass_through_transform(N)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
}
}
static auto MakeCGridDescriptor_M_N(index_t M, index_t N, index_t StrideC)
{
const auto c_grid_desc_m_n = [&]() {
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));
}
}();
if constexpr(GemmSpec == GemmSpecialization::MNPadding)
{
const auto PadM = (MPerBlock - M % MPerBlock) % MPerBlock;
const auto PadN = (NPerBlock - N % NPerBlock) % NPerBlock;
return transform_tensor_descriptor(
c_grid_desc_m_n,
make_tuple(make_right_pad_transform(M, PadM), make_right_pad_transform(N, PadN)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
}
else
{
return transform_tensor_descriptor(
c_grid_desc_m_n,
make_tuple(make_pass_through_transform(M), make_pass_through_transform(N)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
}
}
static auto GetKPad(index_t K, index_t KBatch)
{
const index_t K0 = math::integer_divide_ceil(K, K1 * K0PerBlock * KBatch) * K0PerBlock;
const index_t KPad = KBatch * K0 * K1;
return KPad;
}
using AGridDesc_K0_M_K1 = decltype(MakeAGridDescriptor_KBatch_K0_M_K1(1, 1, 1, 1, 1));
using BGridDesc_K0_N_K1 = decltype(MakeBGridDescriptor_KBatch_K0_N_K1(1, 1, 1, 1, 1));
using CGridDesc_M_N = decltype(MakeCGridDescriptor_M_N(1, 1, 1));
// GridwiseGemm
using GridwiseGemm = GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2< using GridwiseGemm = GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2<
BlockSize, BlockSize,
ADataType, // TODO: distinguish A/B datatype ADataType, // TODO: distinguish A/B datatype
AccDataType, AccDataType,
CDataType, CDataType,
ALayout, InMemoryDataOperationEnum::Set,
BLayout, AGridDesc_K0_M_K1,
CLayout, BGridDesc_K0_N_K1,
CGridDesc_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,
CShuffleMRepeatPerShuffle,
CShuffleNRepeatPerShuffle,
CBlockTransferScalarPerVector_NWaveNPerXDL,
CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock>;
// GridwiseGemm
using GridwiseGemmAtomicAdd = GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2<
BlockSize,
ADataType, // TODO: distinguish A/B datatype
AccDataType,
CDataType,
InMemoryDataOperationEnum::AtomicAdd,
AGridDesc_K0_M_K1,
BGridDesc_K0_N_K1,
CGridDesc_M_N,
AElementwiseOperation, AElementwiseOperation,
BElementwiseOperation, BElementwiseOperation,
CElementwiseOperation, CElementwiseOperation,
GemmSpec,
MPerBlock, MPerBlock,
NPerBlock, NPerBlock,
K0PerBlock, K0PerBlock,
...@@ -114,64 +295,194 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout, ...@@ -114,64 +295,194 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout,
CBlockTransferScalarPerVector_NWaveNPerXDL, CBlockTransferScalarPerVector_NWaveNPerXDL,
CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock>; CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock>;
using Argument = typename GridwiseGemm::Argument; using CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock =
decltype(GridwiseGemm::MakeCGridDesc_MBlock_MPerBlock_NBlock_NPerBlock(CGridDesc_M_N{}));
using Block2CTileMap = typename GridwiseGemm::CBlockClusterAdaptor;
// Argument
struct Argument : public BaseArgument
{
Argument(const ADataType* p_a_grid,
const BDataType* p_b_grid,
CDataType* p_c_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,
index_t k_batch)
: p_a_grid_{p_a_grid},
p_b_grid_{p_b_grid},
p_c_grid_{p_c_grid},
a_grid_desc_kbatch_k0_m_k1_{},
b_grid_desc_kbatch_k0_n_k1_{},
c_grid_desc_m_n_{},
c_grid_desc_mblock_mperblock_nblock_nperblock_{},
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},
k_batch_{k_batch}
{
int KPad = DeviceGemmXdlSplitKCShuffle::GetKPad(K, k_batch_);
a_grid_desc_kbatch_k0_m_k1_ =
DeviceGemmXdlSplitKCShuffle::MakeAGridDescriptor_KBatch_K0_M_K1(
M, K, StrideA, k_batch_, KPad);
b_grid_desc_kbatch_k0_n_k1_ =
DeviceGemmXdlSplitKCShuffle::MakeBGridDescriptor_KBatch_K0_N_K1(
K, N, StrideB, k_batch_, KPad);
c_grid_desc_m_n_ = DeviceGemmXdlSplitKCShuffle::MakeCGridDescriptor_M_N(M, N, StrideC);
block_2_ctile_map_ =
GridwiseGemm::MakeCBlockClusterAdaptor(c_grid_desc_m_n_, M01, N01, k_batch_);
if(GridwiseGemm::CheckValidity(a_grid_desc_kbatch_k0_m_k1_,
b_grid_desc_kbatch_k0_n_k1_,
c_grid_desc_m_n_,
block_2_ctile_map_))
{
c_grid_desc_mblock_mperblock_nblock_nperblock_ =
GridwiseGemm::MakeCGridDesc_MBlock_MPerBlock_NBlock_NPerBlock(c_grid_desc_m_n_);
}
}
// private:
const ADataType* p_a_grid_;
const BDataType* p_b_grid_;
CDataType* p_c_grid_;
AGridDesc_K0_M_K1 a_grid_desc_kbatch_k0_m_k1_;
BGridDesc_K0_N_K1 b_grid_desc_kbatch_k0_n_k1_;
CGridDesc_M_N c_grid_desc_m_n_;
CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock c_grid_desc_mblock_mperblock_nblock_nperblock_;
Block2CTileMap block_2_ctile_map_;
index_t M01_;
index_t N01_;
AElementwiseOperation a_element_op_;
BElementwiseOperation b_element_op_;
CElementwiseOperation c_element_op_;
index_t k_batch_;
};
// Invoker // Invoker
struct Invoker : public BaseInvoker struct Invoker : public BaseInvoker
{ {
using Argument = DeviceGemmXdlSplitKCShuffle::Argument;
void Print(const Argument& karg) { karg.Print(); } void Print(const Argument& arg)
{
std::cout << "arg.a_grid_desc_kbatch_k0_m_k1_{"
<< arg.a_grid_desc_kbatch_k0_m_k1_.GetLength(I0) << ", "
<< arg.a_grid_desc_kbatch_k0_m_k1_.GetLength(I1) << ", "
<< arg.a_grid_desc_kbatch_k0_m_k1_.GetLength(I2) << ", "
<< arg.a_grid_desc_kbatch_k0_m_k1_.GetLength(I3) << "}" << std::endl;
std::cout << "arg.b_grid_desc_kbatch_k0_n_k1_{"
<< arg.b_grid_desc_kbatch_k0_n_k1_.GetLength(I0) << ", "
<< arg.b_grid_desc_kbatch_k0_n_k1_.GetLength(I1) << ", "
<< arg.b_grid_desc_kbatch_k0_n_k1_.GetLength(I2) << ", "
<< arg.b_grid_desc_kbatch_k0_n_k1_.GetLength(I3) << "}" << 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;
}
float Run(const Argument& karg, const StreamConfig& stream_config = StreamConfig{}) float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
{ {
if(stream_config.log_level_ > 0) if(stream_config.log_level_ > 0)
{ {
Print(karg); Print(arg);
} }
const auto kbatch = karg.k_batch; const auto kbatch = arg.a_grid_desc_kbatch_k0_m_k1_.GetLength(I0);
if(!GridwiseGemm::CheckValidity(karg)) if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_kbatch_k0_m_k1_,
arg.b_grid_desc_kbatch_k0_n_k1_,
arg.c_grid_desc_m_n_,
arg.block_2_ctile_map_))
{ {
throw std::runtime_error( throw std::runtime_error(
"wrong! GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2 has invalid " "wrong! GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2 has invalid setting");
"setting");
} }
index_t gdx, gdy, gdz; const index_t grid_size =
std::tie(gdx, gdy, gdz) = GridwiseGemm::CalculateGridSize(karg); arg.block_2_ctile_map_.CalculateGridSize(arg.c_grid_desc_m_n_);
const auto K0 = karg.K0;
const auto K0 = arg.a_grid_desc_kbatch_k0_m_k1_.GetLength(I1);
const bool has_main_k0_block_loop = GridwiseGemm::CalculateHasMainK0BlockLoop(K0); const bool has_main_k0_block_loop = GridwiseGemm::CalculateHasMainK0BlockLoop(K0);
float ave_time = 0; float ave_time = 0;
const auto Run = [&](const auto& kernel) { const auto Run = [&](const auto& kernel) {
if(kbatch > 1) hipGetErrorString(hipMemset(
hipGetErrorString( arg.p_c_grid_,
hipMemset(karg.p_c_grid, 0, karg.M * karg.N * sizeof(CDataType))); 0,
arg.c_grid_desc_mblock_mperblock_nblock_nperblock_.GetElementSpaceSize() *
ave_time = launch_and_time_kernel( sizeof(CDataType)));
stream_config, kernel, dim3(gdx, gdy, gdz), dim3(BlockSize), 0, karg);
ave_time =
launch_and_time_kernel(stream_config,
kernel,
dim3(grid_size),
dim3(BlockSize),
0,
arg.p_a_grid_,
arg.p_b_grid_,
arg.p_c_grid_,
arg.a_grid_desc_kbatch_k0_m_k1_,
arg.b_grid_desc_kbatch_k0_n_k1_,
arg.c_grid_desc_mblock_mperblock_nblock_nperblock_,
arg.a_element_op_,
arg.b_element_op_,
arg.c_element_op_,
arg.block_2_ctile_map_);
}; };
if(has_main_k0_block_loop) if(has_main_k0_block_loop)
{ {
if(kbatch == 1) if(kbatch == 1)
{ {
const auto kernel = const auto kernel = kernel_gemm_xdlops_v2r4r2<
kernel_gemm_xdlops_v2r4r2_simplified<GridwiseGemm, GridwiseGemm,
true, ADataType, // TODO: distiguish A/B datatype
InMemoryDataOperationEnum::Set>; CDataType,
remove_reference_t<DeviceGemmXdlSplitKCShuffle::AGridDesc_K0_M_K1>,
remove_reference_t<DeviceGemmXdlSplitKCShuffle::BGridDesc_K0_N_K1>,
remove_reference_t<DeviceGemmXdlSplitKCShuffle::
CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock>,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation,
remove_reference_t<DeviceGemmXdlSplitKCShuffle::Block2CTileMap>,
true>;
Run(kernel); Run(kernel);
} }
else else
{ {
const auto kernel = const auto kernel = kernel_gemm_xdlops_v2r4r2<
kernel_gemm_xdlops_v2r4r2_simplified<GridwiseGemm, GridwiseGemmAtomicAdd,
true, ADataType, // TODO: distiguish A/B datatype
InMemoryDataOperationEnum::AtomicAdd>; CDataType,
remove_reference_t<DeviceGemmXdlSplitKCShuffle::AGridDesc_K0_M_K1>,
remove_reference_t<DeviceGemmXdlSplitKCShuffle::BGridDesc_K0_N_K1>,
remove_reference_t<DeviceGemmXdlSplitKCShuffle::
CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock>,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation,
remove_reference_t<DeviceGemmXdlSplitKCShuffle::Block2CTileMap>,
true>;
Run(kernel); Run(kernel);
} }
...@@ -180,19 +491,37 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout, ...@@ -180,19 +491,37 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout,
{ {
if(kbatch == 1) if(kbatch == 1)
{ {
const auto kernel = const auto kernel = kernel_gemm_xdlops_v2r4r2<
kernel_gemm_xdlops_v2r4r2_simplified<GridwiseGemm, GridwiseGemm,
false, ADataType, // TODO: distiguish A/B datatype
InMemoryDataOperationEnum::Set>; CDataType,
remove_reference_t<DeviceGemmXdlSplitKCShuffle::AGridDesc_K0_M_K1>,
remove_reference_t<DeviceGemmXdlSplitKCShuffle::BGridDesc_K0_N_K1>,
remove_reference_t<DeviceGemmXdlSplitKCShuffle::
CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock>,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation,
remove_reference_t<DeviceGemmXdlSplitKCShuffle::Block2CTileMap>,
false>;
Run(kernel); Run(kernel);
} }
else else
{ {
const auto kernel = const auto kernel = kernel_gemm_xdlops_v2r4r2<
kernel_gemm_xdlops_v2r4r2_simplified<GridwiseGemm, GridwiseGemmAtomicAdd,
false, ADataType, // TODO: distiguish A/B datatype
InMemoryDataOperationEnum::AtomicAdd>; CDataType,
remove_reference_t<DeviceGemmXdlSplitKCShuffle::AGridDesc_K0_M_K1>,
remove_reference_t<DeviceGemmXdlSplitKCShuffle::BGridDesc_K0_N_K1>,
remove_reference_t<DeviceGemmXdlSplitKCShuffle::
CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock>,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation,
remove_reference_t<DeviceGemmXdlSplitKCShuffle::Block2CTileMap>,
false>;
Run(kernel); Run(kernel);
} }
...@@ -215,9 +544,12 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout, ...@@ -215,9 +544,12 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout,
return true; return true;
} }
static bool IsSupportedArgument(const Argument& karg) static bool IsSupportedArgument(const Argument& arg)
{ {
return GridwiseGemm::CheckValidity(karg); return GridwiseGemm::CheckValidity(arg.a_grid_desc_kbatch_k0_m_k1_,
arg.b_grid_desc_kbatch_k0_n_k1_,
arg.c_grid_desc_m_n_,
arg.block_2_ctile_map_);
} }
// polymorphic // polymorphic
...@@ -235,9 +567,9 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout, ...@@ -235,9 +567,9 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout,
index_t StrideA, index_t StrideA,
index_t StrideB, index_t StrideB,
index_t StrideC, index_t StrideC,
AElementwiseOperation, AElementwiseOperation a_element_op,
BElementwiseOperation, BElementwiseOperation b_element_op,
CElementwiseOperation, CElementwiseOperation c_element_op,
index_t KBatch) index_t KBatch)
{ {
return Argument{p_a, return Argument{p_a,
...@@ -249,10 +581,11 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout, ...@@ -249,10 +581,11 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout,
StrideA, StrideA,
StrideB, StrideB,
StrideC, StrideC,
GridwiseGemm::CalculateMPadded(M), 1,
GridwiseGemm::CalculateNPadded(N), 1,
GridwiseGemm::CalculateKPadded(K), a_element_op,
GridwiseGemm::CalculateK0(K, KBatch), b_element_op,
c_element_op,
KBatch}; KBatch};
} }
...@@ -268,9 +601,9 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout, ...@@ -268,9 +601,9 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout,
index_t StrideA, index_t StrideA,
index_t StrideB, index_t StrideB,
index_t StrideC, index_t StrideC,
AElementwiseOperation, AElementwiseOperation a_element_op,
BElementwiseOperation, BElementwiseOperation b_element_op,
CElementwiseOperation, CElementwiseOperation c_element_op,
ck::index_t KBatch = 1) override ck::index_t KBatch = 1) override
{ {
return std::make_unique<Argument>(static_cast<const ADataType*>(p_a), return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
...@@ -282,10 +615,11 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout, ...@@ -282,10 +615,11 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout,
StrideA, StrideA,
StrideB, StrideB,
StrideC, StrideC,
GridwiseGemm::CalculateMPadded(M), 1,
GridwiseGemm::CalculateNPadded(N), 1,
GridwiseGemm::CalculateKPadded(K), a_element_op,
GridwiseGemm::CalculateK0(K, KBatch), b_element_op,
c_element_op,
KBatch); KBatch);
} }
...@@ -296,7 +630,31 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout, ...@@ -296,7 +630,31 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout,
} }
// polymorphic // polymorphic
std::string GetTypeString() const override { return GridwiseGemm::GetTypeString(); } std::string GetTypeString() const override
{
auto str = std::stringstream();
// clang-format off
str << "DeviceGemmXdlSplitKCShuffle"
<< "<"
<< BlockSize << ", "
<< MPerBlock << ", "
<< NPerBlock << ", "
<< K0PerBlock << ", "
<< K1 << ", "
<< MPerXDL << ", "
<< NPerXDL << ", "
<< MXdlPerWave << ", "
<< NXdlPerWave << ", "
<< ABlockTransferSrcScalarPerVector << ", "
<< ABlockTransferDstScalarPerVector_K1 << ", "
<< BBlockTransferSrcScalarPerVector << ", "
<< BBlockTransferDstScalarPerVector_K1
<< ">";
// clang-format on
return str.str();
}
}; };
} // namespace device } // namespace device
......
include/ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp
View file @ 5eb2f81e
...@@ -316,8 +316,6 @@ struct Sigmoid ...@@ -316,8 +316,6 @@ struct Sigmoid
y = 1 / (ck::type_convert<T>(1) + exp(-x)); y = 1 / (ck::type_convert<T>(1) + exp(-x));
}; };
int32_t divider_ = 1;
}; };
struct TanH struct TanH
...@@ -333,6 +331,23 @@ struct TanH ...@@ -333,6 +331,23 @@ struct TanH
}; };
}; };
struct Swish
{
Swish(float beta = 1.0f) : beta_(beta) {}
template <typename T>
__host__ __device__ void operator()(T& y, const T& x) const
{
static_assert(is_same<T, float>::value || is_same<T, double>::value ||
is_same<T, ck::half_t>::value,
"Data type is not supported by this operation!");
y = x / (ck::type_convert<T>(1) + ck::math::exp(-beta_ * x));
};
float beta_ = 1.0f;
};
} // namespace element_wise } // namespace element_wise
} // namespace tensor_operation } // namespace tensor_operation
} // namespace ck } // namespace ck
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4r2.hpp
View file @ 5eb2f81e
...@@ -18,23 +18,60 @@ ...@@ -18,23 +18,60 @@
namespace ck { namespace ck {
template <typename GridwiseGemm, template <typename GridwiseGemm,
bool HasMainKBlockLoop, typename FloatAB,
InMemoryDataOperationEnum CGlobalMemoryDataOperation> typename FloatC,
typename AGridDesc_B_K0_M_K1,
typename BGridDesc_B_K0_N_K1,
typename CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation,
typename CBlockClusterAdaptor,
bool HasMainKBlockLoop>
__global__ void __global__ void
#if CK_USE_LAUNCH_BOUNDS #if CK_USE_LAUNCH_BOUNDS
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU) __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
#endif #endif
kernel_gemm_xdlops_v2r4r2_simplified(typename GridwiseGemm::Argument karg) kernel_gemm_xdlops_v2r4r2(const FloatAB* __restrict__ p_a_grid,
const FloatAB* __restrict__ p_b_grid,
FloatC* __restrict__ p_c_grid,
const AGridDesc_B_K0_M_K1 a_b_k0_m_k1_grid_desc,
const BGridDesc_B_K0_N_K1 b_b_k0_n_k1_grid_desc,
const CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
c_grid_desc_mblock_mperblock_nblock_nperblock,
const AElementwiseOperation a_element_op,
const BElementwiseOperation b_element_op,
const CElementwiseOperation c_element_op,
const CBlockClusterAdaptor c_block_cluster_adaptor)
{ {
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__)) #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
constexpr index_t shared_size = GridwiseGemm::GetSharedMemoryNumberOfByte(); constexpr index_t shared_block_size =
GridwiseGemm::GetSharedMemoryNumberOfByte() / sizeof(FloatAB);
__shared__ uint8_t p_shared[shared_size]; __shared__ FloatAB p_shared_block[shared_block_size];
GridwiseGemm::template Run<HasMainKBlockLoop, CGlobalMemoryDataOperation>( GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
karg, static_cast<void*>(p_shared)); p_b_grid,
p_c_grid,
static_cast<void*>(p_shared_block),
a_b_k0_m_k1_grid_desc,
b_b_k0_n_k1_grid_desc,
c_grid_desc_mblock_mperblock_nblock_nperblock,
a_element_op,
b_element_op,
c_element_op,
c_block_cluster_adaptor);
#else #else
ignore = karg; ignore = p_a_grid;
ignore = p_b_grid;
ignore = p_c_grid;
ignore = a_b_k0_m_k1_grid_desc;
ignore = b_b_k0_n_k1_grid_desc;
ignore = c_grid_desc_mblock_mperblock_nblock_nperblock;
ignore = a_element_op;
ignore = b_element_op;
ignore = c_element_op;
ignore = c_block_cluster_adaptor;
#endif // end of if (defined(__gfx908__) || defined(__gfx90a__)) #endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
} }
...@@ -42,13 +79,13 @@ template <index_t BlockSize, ...@@ -42,13 +79,13 @@ template <index_t BlockSize,
typename FloatAB, typename FloatAB,
typename FloatAcc, typename FloatAcc,
typename FloatC, typename FloatC,
typename ALayout, InMemoryDataOperationEnum CGlobalMemoryDataOperation,
typename BLayout, typename AGridDesc_B_K0_M_K1,
typename CLayout, typename BGridDesc_B_K0_N_K1,
typename CMNGridDesc,
typename AElementwiseOperation, typename AElementwiseOperation,
typename BElementwiseOperation, typename BElementwiseOperation,
typename CElementwiseOperation, typename CElementwiseOperation,
tensor_operation::device::GemmSpecialization GemmSpec,
index_t MPerBlock, index_t MPerBlock,
index_t NPerBlock, index_t NPerBlock,
index_t K0PerBlock, index_t K0PerBlock,
...@@ -90,237 +127,9 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2 ...@@ -90,237 +127,9 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
// K1 should be Number<...> // K1 should be Number<...>
static constexpr auto K1 = Number<K1Value>{}; static constexpr auto K1 = Number<K1Value>{};
static constexpr auto M01 = 1;
static constexpr auto N01 = 1;
using ThisThreadBlock = ThisThreadBlock<BlockSize>; using ThisThreadBlock = ThisThreadBlock<BlockSize>;
struct Argument : public ck::tensor_operation::device::BaseArgument
{
const FloatAB* p_a_grid;
const FloatAB* p_b_grid;
FloatC* p_c_grid;
index_t M;
index_t N;
index_t K;
index_t StrideA;
index_t StrideB;
index_t StrideC;
index_t MPadded;
index_t NPadded;
index_t KPadded;
index_t K0;
index_t k_batch;
Argument(const FloatAB* p_a_grid_,
const FloatAB* p_b_grid_,
FloatC* p_c_grid_,
index_t M_,
index_t N_,
index_t K_,
index_t StrideA_,
index_t StrideB_,
index_t StrideC_,
index_t MPadded_,
index_t NPadded_,
index_t KPadded_,
index_t K0_,
index_t k_batch_)
: p_a_grid(p_a_grid_),
p_b_grid(p_b_grid_),
p_c_grid(p_c_grid_),
M(M_),
N(N_),
K(K_),
StrideA(StrideA_),
StrideB(StrideB_),
StrideC(StrideC_),
MPadded(MPadded_),
NPadded(NPadded_),
KPadded(KPadded_),
K0(K0_),
k_batch(k_batch_)
{
}
void Print() const
{
std::cout << "arg {"
<< "M:" << M << ", "
<< "N:" << N << ", "
<< "K:" << K << ", "
<< "SA:" << StrideA << ", "
<< "SB:" << StrideB << ", "
<< "SC:" << StrideC << ", "
<< "MP:" << MPadded << ", "
<< "NP:" << NPadded << ", "
<< "KP:" << KPadded << ", "
<< "K0:" << K0 << ", "
<< "KB:" << k_batch << "}" << std::endl;
}
};
__host__ __device__ static auto CalculateGridSize(const Argument& karg)
{
return std::make_tuple(math::integer_divide_ceil(karg.N, NPerBlock),
math::integer_divide_ceil(karg.M, MPerBlock),
karg.k_batch);
}
// prefer this to be called on host
__host__ __device__ static auto CalculateMPadded(index_t M)
{
return (M + MPerBlock - 1) / MPerBlock * MPerBlock;
}
__host__ __device__ static auto CalculateNPadded(index_t N)
{
return (N + NPerBlock - 1) / NPerBlock * NPerBlock;
}
__host__ __device__ static auto CalculateK0(index_t K, index_t K_Batch = 1)
{
// k_batch * k0 * k0_per_block * k1
auto K_t = K_Batch * K0PerBlock * K1;
return (K + K_t - 1) / K_t * K0PerBlock;
}
__host__ __device__ static auto CalculateKPadded(index_t K, index_t K_Batch = 1)
{
auto K0 = CalculateK0(K, K_Batch);
return K_Batch * K0 * K1;
}
__host__ __device__ static auto MakeAGridDescriptor_KBatch_K0_M_K1(index_t M,
index_t MPad,
index_t K,
index_t StrideA,
index_t KBatch,
index_t K0,
index_t KPad)
{
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_m_kpad = transform_tensor_descriptor(
a_grid_desc_m_k,
make_tuple(make_pass_through_transform(M), make_right_pad_transform(K, KPad - K)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
if constexpr(GemmSpec == tensor_operation::device::GemmSpecialization::MPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::MKPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding)
{
// const auto PadM = (MPerBlock - M % MPerBlock) % MPerBlock;
return transform_tensor_descriptor(
a_grid_desc_m_kpad,
make_tuple(make_unmerge_transform(make_tuple(KBatch, K0, K1)),
make_right_pad_transform(M, MPad - M)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
}
else
{
return transform_tensor_descriptor(
a_grid_desc_m_kpad,
make_tuple(make_unmerge_transform(make_tuple(KBatch, K0, K1)),
make_pass_through_transform(M)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
}
}
__host__ __device__ static auto MakeBGridDescriptor_KBatch_K0_N_K1(index_t K,
index_t NPad,
index_t N,
index_t StrideB,
index_t KBatch,
index_t K0,
index_t KPad)
{
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_kpad_n = transform_tensor_descriptor(
b_grid_desc_k_n,
make_tuple(make_right_pad_transform(K, KPad - K), make_pass_through_transform(N)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
if constexpr(GemmSpec == tensor_operation::device::GemmSpecialization::NPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::NKPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding)
{
// const auto PadN = (NPerBlock - N % NPerBlock) % NPerBlock;
return transform_tensor_descriptor(
b_grid_desc_kpad_n,
make_tuple(make_unmerge_transform(make_tuple(KBatch, K0, K1)),
make_right_pad_transform(N, NPad - N)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
}
else
{
return transform_tensor_descriptor(
b_grid_desc_kpad_n,
make_tuple(make_unmerge_transform(make_tuple(KBatch, K0, K1)),
make_pass_through_transform(N)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
}
}
__host__ __device__ static auto
MakeCGridDescriptor_M_N(index_t M, index_t N, index_t MPad, index_t NPad, index_t StrideC)
{
const auto c_grid_desc_m_n = [&]() {
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));
}
}();
if constexpr(GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding)
{
return transform_tensor_descriptor(c_grid_desc_m_n,
make_tuple(make_right_pad_transform(M, MPad - M),
make_right_pad_transform(N, NPad - N)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
}
else
{
return transform_tensor_descriptor(
c_grid_desc_m_n,
make_tuple(make_pass_through_transform(M), make_pass_through_transform(N)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
}
}
__host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte() __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
{ {
constexpr auto max_lds_align = K1; constexpr auto max_lds_align = K1;
...@@ -369,68 +178,45 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2 ...@@ -369,68 +178,45 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
c_block_size * sizeof(FloatC)); c_block_size * sizeof(FloatC));
} }
__host__ __device__ static constexpr bool CheckValidity(const Argument& karg) // block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
{ template <typename Block2CTileMap>
if constexpr(!(GemmSpec == tensor_operation::device::GemmSpecialization::MPadding || __host__ __device__ static constexpr bool
GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding || CheckValidity(const AGridDesc_B_K0_M_K1& a_b_k0_m_k1_grid_desc,
GemmSpec == tensor_operation::device::GemmSpecialization::MKPadding || const BGridDesc_B_K0_N_K1& b_b_k0_n_k1_grid_desc,
GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding)) const CMNGridDesc& c_m_n_grid_desc,
{ const Block2CTileMap& block_2_ctile_map)
if(!(karg.M % MPerBlock == 0))
return false;
}
if constexpr(!(GemmSpec == tensor_operation::device::GemmSpecialization::NPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::NKPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding))
{ {
if(!(karg.N % NPerBlock == 0)) static_assert(is_known_at_compile_time<remove_cv_t<decltype(K1)>>::value,
return false; "wrong! K1 need to be known at compile-time");
}
if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value) static_assert((MPerBlock % (MPerXDL * MRepeat) == 0) &&
{ (NPerBlock % (NRepeat * NPerXDL)) == 0,
if(karg.K % ABlockTransferSrcScalarPerVector != 0) "Invalid tuning param!");
return false;
}
else
{
if(karg.M % ABlockTransferSrcScalarPerVector != 0)
return false;
}
if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value) const auto M = a_b_k0_m_k1_grid_desc.GetLength(I2);
{ const auto N = b_b_k0_n_k1_grid_desc.GetLength(I2);
if(karg.N % BBlockTransferSrcScalarPerVector != 0) const auto K0 = a_b_k0_m_k1_grid_desc.GetLength(I1);
return false; const auto KBatch = a_b_k0_m_k1_grid_desc.GetLength(I0);
}
else if(!(M == c_m_n_grid_desc.GetLength(I0) && N == c_m_n_grid_desc.GetLength(I1) &&
{ K0 == b_b_k0_n_k1_grid_desc.GetLength(I1) &&
if(karg.K % BBlockTransferSrcScalarPerVector != 0) K1 == a_b_k0_m_k1_grid_desc.GetLength(I3) &&
K1 == b_b_k0_n_k1_grid_desc.GetLength(I3) &&
KBatch == b_b_k0_n_k1_grid_desc.GetLength(I0)))
return false; return false;
}
if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value) if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K0 % K0PerBlock == 0))
{
if(karg.N % CBlockTransferScalarPerVector_NWaveNPerXDL != 0)
return false; return false;
}
else if(!block_2_ctile_map.CheckValidity(c_m_n_grid_desc))
{ {
if(karg.M % CBlockTransferScalarPerVector_NWaveNPerXDL != 0)
return false; return false;
} }
// TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
return true; return true;
} }
__host__ __device__ static auto GetKPad(index_t K, index_t KBatch)
{
const index_t K0 = math::integer_divide_ceil(K, K1 * K0PerBlock * KBatch) * K0PerBlock;
const index_t KPad = KBatch * K0 * K1;
return KPad;
}
__host__ __device__ static constexpr bool CalculateHasMainK0BlockLoop(index_t K0) __host__ __device__ static constexpr bool CalculateHasMainK0BlockLoop(index_t K0)
{ {
const bool has_main_k0_block_loop = K0 > K0PerBlock; const bool has_main_k0_block_loop = K0 > K0PerBlock;
...@@ -438,9 +224,8 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2 ...@@ -438,9 +224,8 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
return has_main_k0_block_loop; return has_main_k0_block_loop;
} }
template <typename CGridDesc>
__host__ __device__ static constexpr auto __host__ __device__ static constexpr auto
MakeCGridDesc_MBlock_MPerBlock_NBlock_NPerBlock(const CGridDesc& c_m_n_grid_desc) MakeCGridDesc_MBlock_MPerBlock_NBlock_NPerBlock(const CMNGridDesc& c_m_n_grid_desc)
{ {
const auto M = c_m_n_grid_desc.GetLength(I0); const auto M = c_m_n_grid_desc.GetLength(I0);
const auto N = c_m_n_grid_desc.GetLength(I1); const auto N = c_m_n_grid_desc.GetLength(I1);
...@@ -457,11 +242,10 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2 ...@@ -457,11 +242,10 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
} }
// return block_id to C matrix tile idx (m0, n0) mapping // return block_id to C matrix tile idx (m0, n0) mapping
template <typename CGridDesc>
__host__ __device__ static constexpr auto MakeCBlockClusterAdaptor( __host__ __device__ static constexpr auto MakeCBlockClusterAdaptor(
const CGridDesc& c_m_n_grid_desc, index_t /* M01 */, index_t /* N01 */, index_t KBatch) const CMNGridDesc& c_m_n_grid_desc, index_t /* M01 */, index_t /* N01 */, index_t KBatch)
{ {
return BlockToCTileMap_KSplit_M00_N0_M01Adapt<MPerBlock, NPerBlock, CGridDesc>( return BlockToCTileMap_KSplit_M00_N0_M01Adapt<MPerBlock, NPerBlock, CMNGridDesc>(
c_m_n_grid_desc, 8, KBatch); c_m_n_grid_desc, 8, KBatch);
} }
...@@ -478,25 +262,24 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2 ...@@ -478,25 +262,24 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
Number<CShuffleNRepeatPerShuffle * NWave * NPerXDL>{})); Number<CShuffleNRepeatPerShuffle * NWave * NPerXDL>{}));
} }
template <bool HasMainKBlockLoop, InMemoryDataOperationEnum CGlobalMemoryDataOperation> using CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock =
__device__ static void Run(const Argument& karg, void* __restrict__ p_shared_block) decltype(MakeCGridDesc_MBlock_MPerBlock_NBlock_NPerBlock(CMNGridDesc{}));
{ using CBlockClusterAdaptor = decltype(MakeCBlockClusterAdaptor(CMNGridDesc{}, 1, 1, 1));
const FloatAB* p_a_grid = karg.p_a_grid;
const FloatAB* p_b_grid = karg.p_b_grid;
FloatC* p_c_grid = karg.p_c_grid;
const auto a_b_k0_m_k1_grid_desc = MakeAGridDescriptor_KBatch_K0_M_K1(
karg.M, karg.MPadded, karg.K, karg.StrideA, karg.k_batch, karg.K0, karg.KPadded);
const auto b_b_k0_n_k1_grid_desc = MakeBGridDescriptor_KBatch_K0_N_K1(
karg.K, karg.NPadded, karg.N, karg.StrideB, karg.k_batch, karg.K0, karg.KPadded);
const auto c_grid_desc_m_n =
MakeCGridDescriptor_M_N(karg.M, karg.N, karg.MPadded, karg.NPadded, karg.StrideC);
const auto c_grid_desc_mblock_mperblock_nblock_nperblock =
MakeCGridDesc_MBlock_MPerBlock_NBlock_NPerBlock(c_grid_desc_m_n);
const AElementwiseOperation a_element_op = AElementwiseOperation{};
const BElementwiseOperation b_element_op = BElementwiseOperation{};
const CElementwiseOperation c_element_op = CElementwiseOperation{};
template <bool HasMainKBlockLoop>
__device__ static void Run(const FloatAB* __restrict__ p_a_grid,
const FloatAB* __restrict__ p_b_grid,
FloatC* __restrict__ p_c_grid,
void* __restrict__ p_shared_block,
const AGridDesc_B_K0_M_K1& a_b_k0_m_k1_grid_desc,
const BGridDesc_B_K0_N_K1& b_b_k0_n_k1_grid_desc,
const CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock&
c_grid_desc_mblock_mperblock_nblock_nperblock,
const AElementwiseOperation& a_element_op,
const BElementwiseOperation& b_element_op,
const CElementwiseOperation& c_element_op,
const CBlockClusterAdaptor& c_block_cluster_adaptor)
{
const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>( const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_a_grid, a_b_k0_m_k1_grid_desc.GetElementSpaceSize()); p_a_grid, a_b_k0_m_k1_grid_desc.GetElementSpaceSize());
const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>( const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
...@@ -506,16 +289,26 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2 ...@@ -506,16 +289,26 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
const auto K0 = a_b_k0_m_k1_grid_desc.GetLength(I1); const auto K0 = a_b_k0_m_k1_grid_desc.GetLength(I1);
const index_t block_m_id = __builtin_amdgcn_readfirstlane(blockIdx.y); // divide block work by [M, N]
const index_t block_n_id = __builtin_amdgcn_readfirstlane(blockIdx.x); const auto block_work_idx =
const index_t k_batch_id = __builtin_amdgcn_readfirstlane(blockIdx.z); c_block_cluster_adaptor.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
if(!c_block_cluster_adaptor.ValidCTileIndex(
make_tuple(block_work_idx[I1], block_work_idx[I2]),
make_tuple(c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I0),
c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I2))))
{
return;
}
const index_t k_batch_id = block_work_idx[I0];
// HACK: this force m/n_block_data_idx_on_grid into SGPR // HACK: this force m/n_block_data_idx_on_grid into SGPR
const index_t m_block_data_idx_on_grid = const index_t m_block_data_idx_on_grid =
__builtin_amdgcn_readfirstlane(block_m_id * MPerBlock); __builtin_amdgcn_readfirstlane(block_work_idx[I1] * MPerBlock);
const index_t n_block_data_idx_on_grid = const index_t n_block_data_idx_on_grid =
__builtin_amdgcn_readfirstlane(block_n_id * NPerBlock); __builtin_amdgcn_readfirstlane(block_work_idx[I2] * NPerBlock);
// lds max alignment // lds max alignment
constexpr auto max_lds_align = K1; constexpr auto max_lds_align = K1;
...@@ -651,6 +444,7 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2 ...@@ -651,6 +444,7 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
// c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in // c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
// register // register
// sanity check // sanity check
auto blockwise_gemm = auto blockwise_gemm =
BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize, BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
FloatAB, FloatAB,
...@@ -853,7 +647,7 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2 ...@@ -853,7 +647,7 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
{c_block_desc_mblock_mperblock_nblock_nperblock, {c_block_desc_mblock_mperblock_nblock_nperblock,
make_multi_index(0, 0, 0, 0), make_multi_index(0, 0, 0, 0),
c_grid_desc_mblock_mperblock_nblock_nperblock, c_grid_desc_mblock_mperblock_nblock_nperblock,
make_multi_index(block_m_id, 0, block_n_id, 0), make_multi_index(block_work_idx[I1], 0, block_work_idx[I2], 0),
c_element_op}; c_element_op};
constexpr auto mxdlperwave_forward_step = constexpr auto mxdlperwave_forward_step =
...@@ -922,48 +716,6 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2 ...@@ -922,48 +716,6 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
}); });
} }
} }
template <typename Layout>
struct LStr
{
static std::string Get() { return ""; }
};
template <>
struct LStr<ck::tensor_layout::gemm::RowMajor>
{
static std::string Get() { return "R"; }
};
template <>
struct LStr<ck::tensor_layout::gemm::ColumnMajor>
{
static std::string Get() { return "C"; }
};
static std::string GetTypeString()
{
auto str = std::stringstream();
// clang-format off
str << "GemmXdlSplitKCShuffle_"
<< getGemmSpecializationString(GemmSpec) << "_"
<< std::string(ALayout::name)[0]
<< std::string(BLayout::name)[0]
<< std::string(CLayout::name)[0]
<< "_"
<< "B" << BlockSize << "_"
<< "Vec" << ABlockTransferSrcScalarPerVector << "x"
<< BBlockTransferSrcScalarPerVector << "x"
<< CBlockTransferScalarPerVector_NWaveNPerXDL << "_"
<< MPerBlock << "x"
<< NPerBlock << "x"
<< K0PerBlock << "x"
<< K1 ;
// clang-format on
return str.str();
}
}; };
} // namespace ck } // namespace ck
include/ck/utility/math.hpp
View file @ 5eb2f81e
...@@ -168,6 +168,10 @@ __device__ double exp<double>(double x) ...@@ -168,6 +168,10 @@ __device__ double exp<double>(double x)
return exp(x); return exp(x);
} }
static inline __host__ float exp(float x) { return std::expf(x); }
static inline __host__ double exp(double x) { return std::exp(x); }
// greatest common divisor, aka highest common factor // greatest common divisor, aka highest common factor
__host__ __device__ constexpr index_t gcd(index_t x, index_t y) __host__ __device__ constexpr index_t gcd(index_t x, index_t y)
{ {
......
library/include/ck/library/tensor_operation_instance/device_operation_instance_factory.hpp
View file @ 5eb2f81e
...@@ -96,6 +96,7 @@ using FastGelu = ck::tensor_operation::element_wise::FastGelu; ...@@ -96,6 +96,7 @@ using FastGelu = ck::tensor_operation::element_wise::FastGelu;
using AddMultiply = ck::tensor_operation::element_wise::AddMultiply; using AddMultiply = ck::tensor_operation::element_wise::AddMultiply;
using ScaleAdd = ck::tensor_operation::element_wise::ScaleAdd; using ScaleAdd = ck::tensor_operation::element_wise::ScaleAdd;
using Gelu = ck::tensor_operation::element_wise::Gelu; using Gelu = ck::tensor_operation::element_wise::Gelu;
using Swish = ck::tensor_operation::element_wise::Swish;
template <typename Activation> template <typename Activation>
using Activation_Mul_Clamp = ck::tensor_operation::element_wise::Activation_Mul_Clamp<Activation>; using Activation_Mul_Clamp = ck::tensor_operation::element_wise::Activation_Mul_Clamp<Activation>;
......
library/include/ck/library/tensor_operation_instance/gpu/normalization_swish.hpp 0 → 100644
View file @ 5eb2f81e
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_normalization.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
// FP16
void add_device_normalization_rank_5_3_swish_f16_instances(
std::vector<std::unique_ptr<DeviceNormalization<F16, F16, F16, F32, F16, Swish, 5, 3>>>&);
// FP32
void add_device_normalization_rank_5_3_swish_f32_instances(
std::vector<std::unique_ptr<DeviceNormalization<F32, F32, F32, F32, F32, Swish, 5, 3>>>&);
template <typename XDataType,
typename GammaDataType,
typename BetaDataType,
typename YDataType,
index_t Rank,
index_t NumReduceDim>
struct DeviceOperationInstanceFactory<
ck::tensor_operation::device::DeviceNormalization<XDataType,
GammaDataType,
BetaDataType,
F32,
YDataType,
ck::tensor_operation::element_wise::Swish,
Rank,
NumReduceDim>>
{
using DeviceOp = DeviceNormalization<XDataType,
GammaDataType,
BetaDataType,
F32,
YDataType,
ck::tensor_operation::element_wise::Swish,
Rank,
NumReduceDim>;
static auto GetInstances()
{
std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
if constexpr(is_same_v<XDataType, F16> && is_same_v<GammaDataType, F16> &&
is_same_v<BetaDataType, F16> && is_same_v<YDataType, F16>)
{
if constexpr(Rank == 5 && NumReduceDim == 3)
{
add_device_normalization_rank_5_3_swish_f16_instances(op_ptrs);
}
}
else if constexpr(is_same_v<XDataType, F32> && is_same_v<GammaDataType, F32> &&
is_same_v<BetaDataType, F32> && is_same_v<YDataType, F32>)
{
if constexpr(Rank == 5 && NumReduceDim == 3)
{
add_device_normalization_rank_5_3_swish_f32_instances(op_ptrs);
}
}
return op_ptrs;
}
};
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instance.cpp
View file @ 5eb2f81e
...@@ -26,8 +26,7 @@ using S = ck::Sequence<Is...>; ...@@ -26,8 +26,7 @@ using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough; using PassThrough = ck::tensor_operation::element_wise::PassThrough;
// static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default; static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
static constexpr auto GemmMNPadding = ck::tensor_operation::device::GemmSpecialization::MNPadding;
// Compilation parameters for a[m, k] * b[k, n] = c[m, n] // Compilation parameters for a[m, k] * b[k, n] = c[m, n]
using device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instances = std::tuple< using device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instances = std::tuple<
...@@ -36,22 +35,14 @@ using device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instances = std::tuple< ...@@ -36,22 +35,14 @@ using device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instances = std::tuple<
//#########################| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise|Specialization| 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_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| //#########################| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise|Specialization| 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_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
//#########################| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| //#########################| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl|
//#########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | //#########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8>, DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, true, 1, 1, S<1, 32, 1, 8>, 8>, DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, true, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmMNPadding, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, true, 1, 1, S<1, 16, 1, 8>, 8>, DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, true, 1, 1, S<1, 16, 1, 8>, 8>,
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 64, 192, 4, 8, 32, 32, 1, 3, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 48, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8>, DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 192, 64, 4, 8, 32, 32, 3, 1, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8>, DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, true, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8>, DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, true, 1, 1, S<1, 16, 1, 8>, 8>,
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmMNPadding, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, true, 1, 1, S<1, 32, 1, 4>, 8>, DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, true, 1, 1, S<1, 16, 1, 4>, 8>,
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmMNPadding, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, true, 1, 1, S<1, 16, 1, 8>, 8>, DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8>
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, true, 1, 1, S<1, 16, 1, 4>, 8>,
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmMNPadding, 128, 32, 192, 4, 8, 32, 32, 1, 3, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 24, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8>,
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmMNPadding, 128, 192, 32, 4, 8, 32, 32, 3, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, true, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmMNPadding, 128, 32, 64, 4, 8, 32, 32, 1, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, true, 1, 1, S<1, 16, 1, 8>, 8>,
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmMNPadding, 128, 64, 32, 4, 8, 32, 32, 1, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, true, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmMNPadding, 128, 32, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, true, 1, 1, S<1, 16, 1, 8>, 8>,
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmMNPadding, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, true, 1, 1, S<1, 32, 1, 4>, 8>
// clang-format on // clang-format on
>; >;
......
library/src/tensor_operation_instance/gpu/normalization/CMakeLists.txt
View file @ 5eb2f81e
add_instance_library(device_normalization_instance add_instance_library(device_normalization_instance
device_normalization_f16_instance.cpp device_layernorm2d_f16_instance.cpp
device_normalization_f32_instance.cpp device_layernorm2d_f32_instance.cpp
device_layernorm4d_f16_instance.cpp
device_layernorm4d_f32_instance.cpp
device_groupnorm_f16_instance.cpp
device_groupnorm_f32_instance.cpp
device_groupnorm_swish_f16_instance.cpp
device_groupnorm_swish_f32_instance.cpp
) )
library/src/tensor_operation_instance/gpu/normalization/device_groupnorm_f16_instance.cpp 0 → 100644
View file @ 5eb2f81e
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "normalization_instance_common.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using Pass = ck::tensor_operation::element_wise::PassThrough;
void add_device_normalization_rank_5_3_f16_instances(
std::vector<std::unique_ptr<DeviceNormalization<F16, F16, F16, F32, F16, Pass, 5, 3>>>&
instances)
{
add_device_operation_instances(instances, device_normalization_f16_instances<Pass, 5, 3>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/normalization/device_groupnorm_f32_instance.cpp 0 → 100644
View file @ 5eb2f81e
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "normalization_instance_common.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using Pass = ck::tensor_operation::element_wise::PassThrough;
void add_device_normalization_rank_5_3_f32_instances(
std::vector<std::unique_ptr<DeviceNormalization<F32, F32, F32, F32, F32, Pass, 5, 3>>>&
instances)
{
add_device_operation_instances(instances, device_normalization_f32_instances<Pass, 5, 3>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/normalization/device_groupnorm_swish_f16_instance.cpp 0 → 100644
View file @ 5eb2f81e
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "normalization_instance_common.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using Swish = ck::tensor_operation::element_wise::Swish;
void add_device_normalization_rank_5_3_swish_f16_instances(
std::vector<std::unique_ptr<DeviceNormalization<F16, F16, F16, F32, F16, Swish, 5, 3>>>&
instances)
{
add_device_operation_instances(instances, device_normalization_f16_instances<Swish, 5, 3>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/normalization/device_groupnorm_swish_f32_instance.cpp 0 → 100644
View file @ 5eb2f81e
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "normalization_instance_common.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using Swish = ck::tensor_operation::element_wise::Swish;
void add_device_normalization_rank_5_3_swish_f32_instances(
std::vector<std::unique_ptr<DeviceNormalization<F32, F32, F32, F32, F32, Swish, 5, 3>>>&
instances)
{
add_device_operation_instances(instances, device_normalization_f32_instances<Swish, 5, 3>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/normalization/device_layernorm2d_f16_instance.cpp 0 → 100644
View file @ 5eb2f81e
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "normalization_instance_common.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using Pass = ck::tensor_operation::element_wise::PassThrough;
void add_device_normalization_rank_2_1_f16_instances(
std::vector<std::unique_ptr<DeviceNormalization<F16, F16, F16, F32, F16, Pass, 2, 1>>>&
instances)
{
add_device_operation_instances(instances, device_normalization_f16_instances<Pass, 2, 1>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
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