Skip to content

GitLab

"vscode:/vscode.git/clone" did not exist on "276892778219cb7903fd277cd54940ff11787105"
Commit 250a89f3 authored by Mirza Halilcevic's avatar Mirza Halilcevic
Browse files

Replace gemm_gemm with gemm_multiple_d_gemm_multiple_d.

parent d1e9682a
Hide whitespace changes
Inline Side-by-side
Showing with 638 additions and 43 deletions
codegen/include/ck/host/device_gemm_elementwise_gemm/operation.hpp → codegen/include/ck/host/device_batched_gemm_multiple_d_gemm_multiple_d/operation.hpp
View file @ 250a89f3
......@@ -8,13 +8,13 @@
#include <string>
#include "ck/host/types.hpp"
#include "ck/host/operation/gemm.hpp"
#include "ck/host/device_gemm_elementwise_gemm/problem.hpp"
#include "ck/host/device_batched_gemm_multiple_d_gemm_multiple_d/problem.hpp"
namespace ck {
namespace host {
namespace device_gemm_elementwise_gemm {
namespace device_batched_gemm_multiple_d_gemm_multiple_d {
// defines all values need for an instance of fwd conv
// defines all values needed for an instance
struct Operation_Xdl_CShuffle
{
// returns a vector of instances, only given fusion operators: will use default problem spec
......@@ -23,36 +23,40 @@ struct Operation_Xdl_CShuffle
// returns a vector of instances, given a problem spec and fusion operators
static std::vector<Operation_Xdl_CShuffle>
CreateOperations(const Problem& prob, const std::string& prologue, const std::string& epilogue);
TensorDesc A{};
TensorDesc A0{};
TensorDesc B0{};
std::vector<TensorDesc> D0s = {};
TensorDesc B1{};
TensorDesc C{};
DataType acc = DataType::Float;
DataType cs_type = DataType::Half;
std::string a_elem_op = PassThrough;
std::string b0_elem_op = PassThrough;
std::string acc0_elem_op = PassThrough;
std::string b1_elem_op = PassThrough;
std::string c_elem_op = PassThrough;
std::string prologue = "";
std::string epilogue = "";
std::string gemm_specialization = "ck::tensor_operation::device::GemmSpecialization::Default";
std::vector<TensorDesc> D1s = {};
TensorDesc E1{};
DataType acc_type = DataType::Float;
DataType cshuffle_type = DataType::Float;
std::string a0_elem_op = PassThrough;
std::string b0_elem_op = PassThrough;
std::string cde0_elem_op = PassThrough;
std::string b1_elem_op = PassThrough;
std::string cde1_elem_op = PassThrough;
std::string prologue = "";
std::string epilogue = "";
// tuning parameters
operation::TileDescGemmElementwiseGemm tile_desc{};
operation::BlockTransferDesc a_block_transfer{};
operation::PaddingDesc padding_desc{};
operation::TileDescGemmGemm tile_desc{};
operation::BlockTransferDesc a0_block_transfer{};
operation::BlockTransferDesc b0_block_transfer{};
operation::BlockTransferDesc cde0_block_transfer{};
operation::BlockTransferDesc b1_block_transfer{};
operation::CShuffleDesc cshuffle{};
operation::CBlockTransferDesc c_block_transfer{};
operation::CBlockTransferDesc cde1_block_transfer{};
// functions to update fusion operators if provided
void update_prologue(const std::string& prologue);
void update_epilogue(const std::string& epilogue);
/**constexpr**/ bool IsSupported(std::size_t MRaw_, std::size_t NRaw_, std::size_t KRaw_);
/**constexpr**/ bool
IsSupported(std::size_t MRaw_, std::size_t NRaw_, std::size_t KRaw_, std::size_t Gemm1NRaw_);
// returns a templated instance
Solution ToSolution() const;
};
} // namespace device_gemm_elementwise_gemm
} // namespace device_batched_gemm_multiple_d_gemm_multiple_d
} // namespace host
} // namespace ck
codegen/include/ck/host/device_gemm_elementwise_gemm/problem.hpp → codegen/include/ck/host/device_batched_gemm_multiple_d_gemm_multiple_d/problem.hpp
View file @ 250a89f3
......@@ -10,28 +10,32 @@
namespace ck {
namespace host {
namespace device_gemm_elementwise_gemm {
namespace device_batched_gemm_multiple_d_gemm_multiple_d {
// defines the problem specification for a GEMM operation
// defines the problem specification for a GEMM_ELEMENTWISE_GEMM operation
struct Problem
{
std::size_t M = 0;
std::size_t N = 0;
std::size_t K = 0;
std::size_t O = 0;
bool TransA = false;
bool TransB0 = false;
bool TransB1 = false;
bool TransC = false;
DataType ADataType = DataType::Half;
DataType B0DataType = DataType::Half;
DataType B1DataType = DataType::Half;
DataType CDataType = DataType::Half;
std::string AElementOp = PassThrough;
std::string B0ElementOp = PassThrough;
std::string Acc0ElementOp = PassThrough;
std::string B1ElementOp = PassThrough;
std::string CElementOp = PassThrough;
std::size_t M = 0;
std::size_t N = 0;
std::size_t K = 0;
std::size_t O = 0;
bool TransA0 = false;
bool TransB0 = false;
std::vector<bool> D0sTrans = {};
bool TransB1 = false;
std::vector<bool> D1sTrans = {};
bool TransE1 = false;
DataType A0DataType = DataType::Half;
DataType B0DataType = DataType::Half;
std::vector<DataType> D0sDataType = {};
DataType B1DataType = DataType::Half;
std::vector<DataType> D1sDataType = {};
DataType E1DataType = DataType::Half;
std::string A0ElementOp = PassThrough;
std::string B0ElementOp = PassThrough;
std::string CDE0ElementOp = PassThrough;
std::string B1ElementOp = PassThrough;
std::string CDE1ElementOp = PassThrough;
// returns the correct device op file for the operation
std::string GetIncludeHeader() const;
......@@ -42,6 +46,6 @@ struct Problem
const std::string& epilogue) const;
};
} // namespace device_gemm_elementwise_gemm
} // namespace device_batched_gemm_multiple_d_gemm_multiple_d
} // namespace host
} // namespace ck
codegen/include/ck/host/operation/gemm.hpp
View file @ 250a89f3
......@@ -9,6 +9,15 @@ namespace ck {
namespace host {
namespace operation {
struct PaddingDesc
{
bool pad_gemm0_m = 0;
bool pad_gemm0_n = 0;
bool pad_gemm0_k = 0;
bool pad_gemm1_n = 0;
bool pad_gemm1_k = 0;
};
struct TileDesc
{
int block_size = 0;
......@@ -24,23 +33,23 @@ struct TileDesc
int num_gemmk_prefetch_stage = 0;
};
struct TileDescGemmElementwiseGemm
struct TileDescGemmGemm
{
int block_size = 0;
int gemm01_m_per_block = 0;
int gemm0_n_per_block = 0;
int gemm0_k_per_block = 0;
int gemm1_n_per_block = 0;
int gemm1_k_per_block = 0;
int ak1 = 0;
int bk1 = 0;
int b1k1 = 0;
int m_per_XDL = 0;
int n_per_XDL = 0;
int gemm0_m_Xdl_per_wave = 0;
int gemm0_n_Xdl_per_wave = 0;
int gemm1_n_Xdl_per_wave = 0;
int num_gemmk_prefetch_stage = 0;
int block_size = 0;
int gemm0_m_per_block = 0;
int gemm0_n_per_block = 0;
int gemm0_k_per_block = 0;
int gemm1_n_per_block = 0;
int gemm1_k_per_block = 0;
int a0k1 = 0;
int b0k1 = 0;
int b1k1 = 0;
int m_per_XDL = 0;
int n_per_XDL = 0;
int gemm0_m_Xdl_per_wave = 0;
int gemm0_n_Xdl_per_wave = 0;
int gemm1_n_Xdl_per_wave = 0;
int num_gemm0k_prefetch_stage = 0;
};
struct BlockTransferDesc
......@@ -53,11 +62,13 @@ struct BlockTransferDesc
int dst_scalar_per_vector_k1 = 0;
int lds_add_extra_dim = 0;
};
struct CShuffleDesc
{
int m_Xdl_per_wave_per_shuffle = 0;
int n_Xdl_per_wave_per_shuffle = 0;
};
struct CBlockTransferDesc
{
std::string cluster_lengths_m_block_m_wave_m_per_Xdl_n_block_n_wave_n_per_Xdl = "";
......
codegen/src/device_gemm_elementwise_gemm.cpp → codegen/src/device_batched_gemm_multiple_d_gemm_multiple_d.cpp
View file @ 250a89f3
......@@ -2,19 +2,20 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/host/device_gemm_elementwise_gemm/problem.hpp"
#include "ck/host/device_gemm_elementwise_gemm/operation.hpp"
#include "ck/host/device_batched_gemm_multiple_d_gemm_multiple_d/problem.hpp"
#include "ck/host/device_batched_gemm_multiple_d_gemm_multiple_d/operation.hpp"
#include "ck/host/utils.hpp"
#include <algorithm>
namespace ck {
namespace host {
namespace device_gemm_elementwise_gemm {
namespace device_batched_gemm_multiple_d_gemm_multiple_d {
// return the relevant device op file based on the operation
std::string Problem::GetIncludeHeader() const
{
return "ck/tensor_operation/gpu/device/impl/device_batched_gemm_gemm_xdl_cshuffle.hpp";
return "ck/tensor_operation/gpu/device/impl/"
"device_batched_gemm_multiple_d_gemm_multiple_d_xdl_cshuffle.hpp";
}
// returns templated instances when provided with a problem specification
......@@ -24,8 +25,8 @@ std::vector<Solution> Problem::GetSolutions(const std::string& arch,
{
if(get_xdlop_archs().count(arch) == 0)
return {};
auto ops = ck::host::device_gemm_elementwise_gemm::Operation_Xdl_CShuffle::CreateOperations(
*this, prologue, epilogue); // obtains vector of instances
auto ops = ck::host::device_batched_gemm_multiple_d_gemm_multiple_d::Operation_Xdl_CShuffle::
CreateOperations(*this, prologue, epilogue); // obtains vector of instances
std::vector<Solution> result;
std::transform(ops.begin(), ops.end(), std::back_inserter(result), [&](const auto& op) {
return op.ToSolution(); // template instance with correct values
......@@ -33,6 +34,6 @@ std::vector<Solution> Problem::GetSolutions(const std::string& arch,
return result;
}
} // namespace device_gemm_elementwise_gemm
} // namespace device_batched_gemm_multiple_d_gemm_multiple_d
} // namespace host
} // namespace ck
codegen/src/device_gemm_elementwise_gemm_operation_xdl_cshuffle.cpp → codegen/src/device_batched_gemm_multiple_d_gemm_multiple_d_operation_xdl_cshuffle.cpp
View file @ 250a89f3
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/host/device_gemm_elementwise_gemm/operation.hpp"
#include "ck/host/device_batched_gemm_multiple_d_gemm_multiple_d/operation.hpp"
#include "ck/host/stringutils.hpp"
#include "ck/host/utils.hpp"
#include <cassert>
namespace ck {
namespace host {
namespace device_gemm_elementwise_gemm {
namespace device_batched_gemm_multiple_d_gemm_multiple_d {
// calculate appropriate Gemm Specification based on input tensor dimensions
std::string GetGemmSpec(const std::size_t m,
const std::size_t n,
const std::size_t k,
const std::size_t n1,
const std::size_t m_per_block,
const std::size_t n_per_block,
const std::size_t k_per_block,
const std::size_t n1_per_block)
operation::PaddingDesc GetPaddingDesc(const std::size_t m,
const std::size_t n,
const std::size_t k,
const std::size_t n1,
const std::size_t m_per_block,
const std::size_t n_per_block,
const std::size_t k_per_block,
const std::size_t n1_per_block,
const std::size_t k1_per_block)
{
std::string spec = "";
operation::PaddingDesc desc;
if(integer_divide_ceil(m, m_per_block) * m_per_block - m != 0)
spec += "M";
desc.pad_gemm0_m = true;
if(integer_divide_ceil(n, n_per_block) * n_per_block - n != 0)
spec += "N";
desc.pad_gemm0_n = true;
if(integer_divide_ceil(k, k_per_block) * k_per_block - k != 0)
spec += "K";
desc.pad_gemm0_k = true;
if(integer_divide_ceil(n1, n1_per_block) * n1_per_block - n1 != 0)
spec += "O";
if(spec == "")
return "ck::tensor_operation::device::GemmSpecialization::Default";
desc.pad_gemm1_n = true;
if(integer_divide_ceil(n, k1_per_block) * k1_per_block - n != 0) // TODO is n == k1 ?
desc.pad_gemm1_k = true;
return "ck::tensor_operation::device::GemmSpecialization::" + spec + "Padding";
return desc;
}
// function to update prologue/epilogue with user provided operation
......@@ -41,6 +42,9 @@ void Operation_Xdl_CShuffle::update_prologue(const std::string& pro)
if(!prologue.empty())
{
this->prologue = pro;
// TODO is this right?
this->cde0_elem_op = "CDE0ElementOp";
this->cde1_elem_op = "CDE1ElementOp";
}
else
{
......@@ -53,6 +57,9 @@ void Operation_Xdl_CShuffle::update_epilogue(const std::string& epi)
if(!epilogue.empty())
{
this->epilogue = epi;
// TODO is this right?
this->cde0_elem_op = "CDE0ElementOp";
this->cde1_elem_op = "CDE1ElementOp";
}
else
{
......@@ -68,24 +75,19 @@ static Layout ToLayout(bool Trans) { return Trans ? Layout::Column : Layout::Row
std::vector<Operation_Xdl_CShuffle> Operation_Xdl_CShuffle::CreateOperations(
const Problem& prob, const std::string& prologue, const std::string& epilogue)
{
assert(prob.TransA == false);
assert(prob.TransB0 == true);
assert(prob.TransC == false);
std::vector<Operation_Xdl_CShuffle> result;
const auto b1k1 = prob.TransB1 ? 4 : 2;
std::vector<Operation_Xdl_CShuffle> result;
std::vector<operation::TileDescGemmElementwiseGemm> tile_descriptions = {
std::vector<operation::TileDescGemmGemm> tile_descriptions = {
// clang-format off
// Block| Gemm01| Gemm0| Gemm0| Gemm1| Gemm1| AK1| BK1| B1K1| MPer| NPer| Gemm0| Gemm0| Gemm1| NumGemmK|
// Block| Gemm0| Gemm0| Gemm0| Gemm1| Gemm1|A0K1|B0K1| B1K1| MPer| NPer| Gemm0| Gemm0| Gemm1|NumGemm0K|
// Size| MPer| NPer| KPer| NPer| KPer| | | | XDL| XDL| MXdl| NXdl| NXdl| Prefetch|
// | Block| Block| Block| Block| Block| | | | | | Per| Per| Per| Stage|
// | | | | | | | | | | | Wave| Wave| Wave| |
{ 256, 256, 128, 32, 64, 32, 8, 8, b1k1, 32, 32, 2, 4, 2, 1},
{ 256, 256, 128, 32, 128, 32, 8, 8, b1k1, 32, 32, 2, 4, 4, 1},
{ 256, 128, 256, 32, 64, 32, 8, 8, b1k1, 32, 32, 1, 8, 2, 1},
{ 256, 128, 256, 32, 128, 32, 8, 8, b1k1, 32, 32, 1, 8, 4, 1},
//generic
{ 256, 128, 64, 32, 128, 32, 8, 8, b1k1, 32, 32, 1, 2, 4, 1},
// no padding
{ 256, 128, 128, 64, 64, 32, 8, 8, b1k1, 32, 32, 1, 4, 2, 1},
{ 256, 128, 128, 32, 64, 32, 8, 8, b1k1, 32, 32, 1, 4, 2, 1},
{ 256, 128, 128, 64, 128, 32, 8, 8, b1k1, 32, 32, 1, 4, 4, 1},
......@@ -94,22 +96,29 @@ std::vector<Operation_Xdl_CShuffle> Operation_Xdl_CShuffle::CreateOperations(
{ 256, 64, 256, 32, 64, 32, 8, 8, b1k1, 16, 16, 1, 16, 4, 1},
{ 256, 64, 256, 64, 128, 32, 8, 8, b1k1, 16, 16, 1, 16, 8, 1},
{ 256, 64, 256, 64, 64, 32, 8, 8, b1k1, 16, 16, 1, 16, 4, 1},
// Padded fallback kerne
// Padded fallback kernel
{ 256, 128, 128, 64, 128, 32, 8, 8, b1k1, 32, 32, 1, 4, 4, 1},
{ 256, 128, 64, 32, 128, 32, 8, 8, b1k1, 32, 32, 1, 2, 4, 1},
// clang-format on
};
if(prob.TransB1)
{
// clang-format off
tile_descriptions.push_back(
{ 256, 256, 128, 32, 128, 32, 8, 8, 4, 32, 32, 2, 4, 4, 1}
);
// clang-format on
}
const std::vector<operation::BlockTransferDesc> a_block_descriptions = {
std::vector<operation::BlockTransferDesc> a0_block_descriptions = {
// clang-format off
// ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|
// A0BlockTransfer|A0BlockTransfer|A0BlockTransfer|A0BlockTransfer|A0BlockTransfer|A0BlockTransfer|A0BlockLds|
// ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM|
// Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| |
// Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_AK1| |
// | | | | | | |
//generic
{ 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},
{ 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},
// no padding
{ S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false},
{ S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true},
{ S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false},
......@@ -123,90 +132,111 @@ std::vector<Operation_Xdl_CShuffle> Operation_Xdl_CShuffle::CreateOperations(
{ S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true},
// clang-format on
};
const auto& b0_block_descriptions_rowmajor = a_block_descriptions;
const std::vector<operation::BlockTransferDesc> b0_block_descriptions_colmajor = {
if(prob.TransB1)
{
// clang-format off
// B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockLds|
// ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN|
// Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| |
// | | | | | | |
{ 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},
{ 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},
{ S<8, 32, 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},
{ S<8, 32, 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},
{ 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},
{ S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true},
{ S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true},
{ S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false},
{ S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true},
a0_block_descriptions.push_back(
{ S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true}
);
// clang-format on
};
}
auto b0_block_descriptions = a0_block_descriptions;
if(prob.TransB1)
{
b0_block_descriptions[1].lds_add_extra_dim = true;
b0_block_descriptions[3].lds_add_extra_dim = true;
}
const std::vector<operation::BlockTransferDesc> b1_block_descriptions_rowmajor = {
std::vector<operation::BlockTransferDesc> cde0_block_descriptions = {
// clang-format off
// B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockLds|
// ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN|
// Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| |
// | | | | | | |
{ S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false},
{ S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false},
{ S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false},
{ S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false},
{ S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false},
{ S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false},
{ S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false},
{ S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false},
{ S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false},
{ S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false},
{ S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false},
{ S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false},
// ... | CDE0BlockTransfer| CDE0BlockTransfer| ... |
// ... | SrcVectorDim| SrcScalar| ... |
// ... | | PerVector| ... |
// | | | |
//generic
{"", "", "", 9, 1, 0, 0},
// no padding
{"", "", "", 9, 4, 0, 0},
{"", "", "", 9, 4, 0, 0},
{"", "", "", 9, 4, 0, 0},
{"", "", "", 9, 4, 0, 0},
{"", "", "", 9, 4, 0, 0},
{"", "", "", 9, 4, 0, 0},
{"", "", "", 9, 4, 0, 0},
{"", "", "", 9, 4, 0, 0},
// Padded fallback kernel
{ S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false},
{ S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false},
{"", "", "", 9, 4, 0, 0},
{"", "", "", 9, 4, 0, 0},
// clang-format on
};
const std::vector<operation::BlockTransferDesc> b1_block_descriptions_colmajor = {
if(prob.TransB1)
{
// clang-format off
// B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockLds|
// ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN|
// Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| |
// | | | | | | |
{ S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true},
{ S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true},
{ S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true},
{ S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true},
{ S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, false},
{ S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true},
{ S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, false},
{ S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true},
{ S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true},
{ S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true},
{ S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true},
{ S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true},
// Padded fallback kernel
{ S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, false},
{ S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true},
cde0_block_descriptions.push_back(
{"", "", "", 9, 4, 0, 0}
);
// clang-format on
};
}
const std::vector<operation::BlockTransferDesc> b1_block_descriptions_rowmajor =
{
// clang-format off
// B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockLds|
// ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN|
// Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| |
// | | | | | | |
//generic
{ S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false},
// no padding
{ S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false},
{ S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false},
{ S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false},
{ S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false},
{ S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false},
{ S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false},
{ S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false},
{ S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false},
// Padded fallback kernel
{ S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false},
{ S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false},
// clang-format on
};
const std::vector<operation::BlockTransferDesc> b1_block_descriptions_colmajor =
{
// clang-format off
// B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockLds|
// ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN|
// Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| |
// | | | | | | |
//generic
{ S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true},
// no padding
{ S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true},
{ S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, false},
{ S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true},
{ S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, false},
{ S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true},
{ S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true},
{ S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true},
{ S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true},
{ S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true},
// Padded fallback kernel
{ S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, false},
{ S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true},
// clang-format on
};
std::vector<operation::CShuffleDesc> cshuffle_descriptions = {
// clang-format off
// CShuffle| CShuffle|
// C1Shuffle| C1Shuffle|
// MXdlPerWave| NXdlPerWave|
// PerShuffle| PerShuffle|
// | |
// generic
{ 1, 2},
{ 1, 2},
{ 1, 2},
{ 1, 2},
// no padding
{ 1, 2},
{ 1, 2},
{ 1, 2},
......@@ -220,69 +250,92 @@ std::vector<Operation_Xdl_CShuffle> Operation_Xdl_CShuffle::CreateOperations(
{ 1, 2},
// clang-format on
};
if(prob.TransB1)
{
// clang-format off
cshuffle_descriptions.push_back(
{ 1, 2}
);
// clang-format on
}
std::vector<operation::CBlockTransferDesc> c_block_descriptions = {
std::vector<operation::CBlockTransferDesc> cde1_block_descriptions = {
// clang-format off
// CBlockTransferClusterLengths| CBlockTransfer
// _MBlock_MWaveMPerXdl| ScalarPerVector
// _NBlock_NWaveNPerXdl| _NWaveNPerXdl
// |
{ S<1, 32, 1, 8>, 8},
{ S<1, 32, 1, 8>, 8},
{ S<1, 32, 1, 8>, 8},
{ S<1, 32, 1, 8>, 8},
{ S<1, 32, 1, 8>, 8},
{ S<1, 32, 1, 8>, 8},
{ S<1, 32, 1, 8>, 8},
{ S<1, 32, 1, 8>, 8},
{ S<1, 16, 1,16>, 8},
{ S<1, 32, 1, 8>, 8},
{ S<1, 16, 1,16>, 8},
{ S<1, 32, 1, 8>, 8},
// CDE1BlockTransferClusterLengths| CDE1BlockTransfer|
// _MBlock_MWaveMPerXdl| ScalarPerVector|
// _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
// | |
// generic
{ S<1, 32, 1, 8>, 8},
// no padding
{ S<1, 32, 1, 8>, 8},
{ S<1, 32, 1, 8>, 8},
{ S<1, 32, 1, 8>, 8},
{ S<1, 32, 1, 8>, 8},
{ S<1, 16, 1,16>, 8},
{ S<1, 32, 1, 8>, 8},
{ S<1, 16, 1,16>, 8},
{ S<1, 32, 1, 8>, 8},
// Padded fallback kernel
{ S<1, 32, 1, 8>, 8},
{ S<1, 32, 1, 8>, 8},
{ S<1, 32, 1, 8>, 8},
{ S<1, 32, 1, 8>, 8},
// clang-format on
};
if(prob.TransB1)
{
// clang-format off
cde1_block_descriptions.push_back(
{ S<1, 32, 1, 8>, 8}
);
// clang-format on
}
// choose correct arrangement of tuning parameters based on the layout of each tensor
const auto& b0_block_descriptions =
prob.TransB1 ? b0_block_descriptions_colmajor : b0_block_descriptions_rowmajor;
const auto& b1_block_descriptions =
prob.TransB1 ? b1_block_descriptions_colmajor : b1_block_descriptions_rowmajor;
assert(tile_descriptions.size() == a_block_descriptions.size());
assert(tile_descriptions.size() == a0_block_descriptions.size());
assert(tile_descriptions.size() == b0_block_descriptions.size());
assert(tile_descriptions.size() == cde0_block_descriptions.size());
assert(tile_descriptions.size() == b1_block_descriptions.size());
assert(tile_descriptions.size() == cshuffle_descriptions.size());
assert(tile_descriptions.size() == c_block_descriptions.size());
assert(tile_descriptions.size() == cde1_block_descriptions.size());
// Put all values together into a single operation > store into the result vector
for(std::size_t i = 0; i < tile_descriptions.size(); i++)
{
Operation_Xdl_CShuffle x;
x.tile_desc = tile_descriptions[i];
x.a_block_transfer = a_block_descriptions[i];
x.a0_block_transfer = a0_block_descriptions[i];
x.b0_block_transfer = b0_block_descriptions[i];
x.cde0_block_transfer = cde0_block_descriptions[i];
x.b1_block_transfer = b1_block_descriptions[i];
x.cshuffle = cshuffle_descriptions[i];
x.c_block_transfer = c_block_descriptions[i];
x.A = TensorDesc{prob.ADataType, ToLayout(prob.TransA)};
x.cde1_block_transfer = cde1_block_descriptions[i];
x.A0 = TensorDesc{prob.A0DataType, ToLayout(prob.TransA0)};
x.B0 = TensorDesc{prob.B0DataType, ToLayout(prob.TransB0)};
x.D0s = Transform(prob.D0sTrans, prob.D0sDataType, [](auto trans, auto dt) {
return TensorDesc{dt, ToLayout(trans)};
});
x.B1 = TensorDesc{prob.B1DataType, ToLayout(prob.TransB1)};
x.C = TensorDesc{prob.CDataType, ToLayout(prob.TransC)};
x.a_elem_op = prob.AElementOp;
x.D1s = Transform(prob.D1sTrans, prob.D1sDataType, [](auto trans, auto dt) {
return TensorDesc{dt, ToLayout(trans)};
});
x.E1 = TensorDesc{prob.E1DataType, ToLayout(prob.TransE1)};
x.a0_elem_op = prob.A0ElementOp;
x.b0_elem_op = prob.B0ElementOp;
x.cde0_elem_op = prob.CDE0ElementOp;
x.b1_elem_op = prob.B1ElementOp;
x.c_elem_op = prob.CElementOp;
x.acc0_elem_op = prob.Acc0ElementOp;
x.gemm_specialization = GetGemmSpec(prob.M,
prob.N,
prob.K,
prob.O,
x.tile_desc.gemm01_m_per_block,
x.tile_desc.gemm0_n_per_block,
x.tile_desc.gemm0_k_per_block,
x.tile_desc.gemm1_n_per_block);
x.cde1_elem_op = prob.CDE1ElementOp;
x.padding_desc = GetPaddingDesc(prob.M,
prob.N,
prob.K,
prob.O,
x.tile_desc.gemm0_m_per_block,
x.tile_desc.gemm0_n_per_block,
x.tile_desc.gemm0_k_per_block,
x.tile_desc.gemm1_n_per_block,
x.tile_desc.gemm1_k_per_block);
x.update_prologue(prologue);
x.update_epilogue(epilogue);
result.push_back(x);
......@@ -298,10 +351,7 @@ Operation_Xdl_CShuffle::CreateOperations(const std::string& prologue, const std:
std::vector<std::vector<Operation_Xdl_CShuffle>> operations;
Problem prob;
prob.TransA = false;
prob.TransB0 = true;
prob.TransB1 = false;
prob.TransC = false;
operations.push_back(CreateOperations(prob, prologue, epilogue));
prob.TransB1 = true;
......@@ -310,29 +360,43 @@ Operation_Xdl_CShuffle::CreateOperations(const std::string& prologue, const std:
return operations;
}
static const char* const DeviceBatchedGemmGemm_Xdl_CShuffleTemplate =
"ck::tensor_operation::device::DeviceBatchedGemmGemm_Xdl_CShuffle<${LayoutA}, "
"${LayoutB0}, ${LayoutB1}, ${LayoutC}, ${ADataType}, ${B0DataType}, ${B1DataType}, "
"${CDataType}, ${AccDataType}, ${CShuffleDataType}, ${AElementwiseOperation}, "
"${B0ElementwiseOperation}, ${Acc0ElementwiseOperation}, ${B1ElementwiseOperation}, "
"${CElementwiseOperation}, ${GemmSpecialization}, ${NumGemmkPrefetchStage}, ${BlockSize}, "
"${Gemm01MPerBlock}, ${Gemm0NPerBlock}, ${Gemm0KPerBlock}, ${Gemm1NPerBlock}, "
"${Gemm1KPerBlock}, ${AK1}, ${BK1}, ${B1K1}, ${MPerXDL}, ${NPerXDL}, ${Gemm0MXdlPerWave}, "
"${Gemm0NXdlPerWave}, ${Gemm1NXdlPerWave}, ${ABlockTransferThreadClusterLengths_AK0_M_AK1}, "
"${ABlockTransferThreadClusterArrangeOrder}, ${ABlockTransferSrcAccessOrder}, "
"${ABlockTransferSrcVectorDim}, ${ABlockTransferSrcScalarPerVector}, "
"${ABlockTransferDstScalarPerVector_AK1}, ${ABlockLdsExtraM}, "
static const char* const DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffleTemplate =
"ck::tensor_operation::device::DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle<"
"${A0Layout}, ${B0Layout}, ${D0sLayout}, ${B1Layout}, ${D1sLayout}, ${E1Layout}, "
"${A0DataType}, ${B0DataType}, ${Acc0DataType}, ${D0sDataType}, ${B1DataType}, "
"${Acc1DataType}, ${C1ShuffleDataType}, ${D1sDataType}, ${E1DataType}, "
"${A0ElementwiseOperation}, ${B0ElementwiseOperation}, ${CDE0ElementwiseOperation}, "
"${B1ElementwiseOperation}, ${CDE1ElementwiseOperation}, "
"${PadGemm0M}, ${PadGemm0N}, ${PadGemm0K}, ${PadGemm1N}, ${PadGemm1K}, "
"${NumGemm0KPrefetchStage}, ${BlockSize}, ${Gemm0MPerBlock}, ${Gemm0NPerBlock}, "
"${Gemm0KPerBlock}, ${Gemm1NPerBlock}, ${Gemm1KPerBlock}, ${A0K1}, ${B0K1}, ${B1K1}, "
"${MPerXDL}, ${NPerXDL}, ${Gemm0MXdlPerWave}, ${Gemm0NXdlPerWave}, ${Gemm1NXdlPerWave}, "
"${A0BlockTransferThreadClusterLengths_AK0_M_AK1}, "
"${A0BlockTransferThreadClusterArrangeOrder}, ${A0BlockTransferSrcAccessOrder}, "
"${A0BlockTransferSrcVectorDim}, ${A0BlockTransferSrcScalarPerVector}, "
"${A0BlockTransferDstScalarPerVector_AK1}, ${A0BlockLdsExtraM}, "
"${B0BlockTransferThreadClusterLengths_BK0_N_BK1}, "
"${B0BlockTransferThreadClusterArrangeOrder}, ${B0BlockTransferSrcAccessOrder}, "
"${B0BlockTransferSrcVectorDim}, ${B0BlockTransferSrcScalarPerVector}, "
"${B0BlockTransferDstScalarPerVector_BK1}, ${B0BlockLdsExtraN}, "
"${CDE0BlockTransferSrcVectorDim}, ${CDE0BlockTransferSrcScalarPerVector}, "
"${B1BlockTransferThreadClusterLengths_BK0_N_BK1}, "
"${B1BlockTransferThreadClusterArrangeOrder}, ${B1BlockTransferSrcAccessOrder}, "
"${B1BlockTransferSrcVectorDim}, ${B1BlockTransferSrcScalarPerVector}, "
"${B1BlockTransferDstScalarPerVector_BK1}, ${B1BlockLdsExtraN}, "
"${CShuffleMXdlPerWavePerShuffle}, ${CShuffleNXdlPerWavePerShuffle}, "
"${CBlockTransferClusterLengths_MBlock_MWaveMPerXdl_NBlock_NWaveNPerXdl}, "
"${CBlockTransferScalarPerVector_NWaveNPerXdl}>";
"${C1ShuffleMXdlPerWavePerShuffle}, ${C1ShuffleGemm0NXdlPerWavePerShuffle}, "
"${CDE1ShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock}, "
"${CDE1ShuffleBlockTransferScalarPerVector_NPerBlock}>";
// use hardcoded instances from vector of operations to substitute values into instance template
Solution Operation_Xdl_CShuffle::ToSolution() const
......@@ -340,60 +404,80 @@ Solution Operation_Xdl_CShuffle::ToSolution() const
std::unordered_map<std::string, std::string> values = {
{"name",
std::to_string(this->tile_desc.block_size) + "_" +
std::to_string(this->tile_desc.gemm01_m_per_block) + "_" +
std::to_string(this->tile_desc.gemm0_m_per_block) + "_" +
std::to_string(this->tile_desc.gemm0_n_per_block) + "_" +
std::to_string(this->tile_desc.gemm0_k_per_block) + "_" +
std::to_string(this->tile_desc.gemm1_n_per_block) + "_" +
std::to_string(this->tile_desc.gemm1_k_per_block) + "_" +
std::to_string(this->tile_desc.ak1) + "_" + std::to_string(this->tile_desc.bk1) + "_" +
std::to_string(this->tile_desc.b1k1) + "_" +
std::to_string(this->tile_desc.a0k1) + "_" + std::to_string(this->tile_desc.b0k1) +
"_" + std::to_string(this->tile_desc.b1k1) + "_" +
std::to_string(this->tile_desc.m_per_XDL) + "_" +
std::to_string(this->tile_desc.n_per_XDL) + "_" +
std::to_string(this->tile_desc.gemm0_m_Xdl_per_wave) + "_" +
std::to_string(this->tile_desc.gemm0_n_Xdl_per_wave) + "_" +
std::to_string(this->tile_desc.gemm1_n_Xdl_per_wave)},
{"LayoutA", ToString(this->A.layout)},
{"LayoutB0", ToString(this->B0.layout)},
{"LayoutB1", ToString(this->B1.layout)},
{"LayoutC", ToString(this->C.layout)},
{"ADataType", ToString(this->A.element)},
{"A0Layout", ToString(this->A0.layout)},
{"B0Layout", ToString(this->B0.layout)},
{"D0sLayout",
MakeTuple(Transform(this->D0s, [](auto tensor) { return ToString(tensor.layout); }))},
{"B1Layout", ToString(this->B1.layout)},
{"D1sLayout",
MakeTuple(Transform(this->D1s, [](auto tensor) { return ToString(tensor.layout); }))},
{"E1Layout", ToString(this->E1.layout)},
{"ADataType", ToString(this->A0.element)},
{"B0DataType", ToString(this->B0.element)},
{"Acc0DataType", ToString(this->acc_type)},
{"D0sDataType",
MakeTuple(Transform(this->D0s, [](auto tensor) { return ToString(tensor.element); }))},
{"B1DataType", ToString(this->B1.element)},
{"CDataType", ToString(this->C.element)},
{"AccDataType", ToString(this->acc)},
{"CShuffleDataType", ToString(this->cs_type)},
{"AElementwiseOperation", this->a_elem_op},
{"Acc1DataType", ToString(this->acc_type)},
{"C1ShuffleDataType", ToString(this->cshuffle_type)},
{"D1sDataType",
MakeTuple(Transform(this->D1s, [](auto tensor) { return ToString(tensor.element); }))},
{"E1DataType", ToString(this->E1.element)},
{"A0ElementwiseOperation", this->a0_elem_op},
{"B0ElementwiseOperation", this->b0_elem_op},
{"Acc0ElementwiseOperation", this->acc0_elem_op},
{"CDE0ElementwiseOperation", this->cde0_elem_op},
{"B1ElementwiseOperation", this->b1_elem_op},
{"CElementwiseOperation", this->c_elem_op},
{"GemmSpecialization", this->gemm_specialization},
{"NumGemmkPrefetchStage", std::to_string(this->tile_desc.num_gemmk_prefetch_stage)},
{"CDE1ElementwiseOperation", this->cde1_elem_op},
{"PadGemm0M", std::to_string(this->padding_desc.pad_gemm0_m)},
{"PadGemm0N", std::to_string(this->padding_desc.pad_gemm0_n)},
{"PadGemm0K", std::to_string(this->padding_desc.pad_gemm0_k)},
{"PadGemm1N", std::to_string(this->padding_desc.pad_gemm1_n)},
{"PadGemm1K", std::to_string(this->padding_desc.pad_gemm1_k)},
{"NumGemm0KPrefetchStage", std::to_string(this->tile_desc.num_gemm0k_prefetch_stage)},
{"BlockSize", std::to_string(this->tile_desc.block_size)},
{"Gemm01MPerBlock", std::to_string(this->tile_desc.gemm01_m_per_block)},
{"Gemm0MPerBlock", std::to_string(this->tile_desc.gemm0_m_per_block)},
{"Gemm0NPerBlock", std::to_string(this->tile_desc.gemm0_n_per_block)},
{"Gemm0KPerBlock", std::to_string(this->tile_desc.gemm0_k_per_block)},
{"Gemm1NPerBlock", std::to_string(this->tile_desc.gemm1_n_per_block)},
{"Gemm1KPerBlock", std::to_string(this->tile_desc.gemm1_k_per_block)},
{"AK1", std::to_string(this->tile_desc.ak1)},
{"BK1", std::to_string(this->tile_desc.bk1)},
{"A0K1", std::to_string(this->tile_desc.a0k1)},
{"B0K1", std::to_string(this->tile_desc.b0k1)},
{"B1K1", std::to_string(this->tile_desc.b1k1)},
{"MPerXDL", std::to_string(this->tile_desc.m_per_XDL)},
{"NPerXDL", std::to_string(this->tile_desc.n_per_XDL)},
{"Gemm0MXdlPerWave", std::to_string(this->tile_desc.gemm0_m_Xdl_per_wave)},
{"Gemm0NXdlPerWave", std::to_string(this->tile_desc.gemm0_n_Xdl_per_wave)},
{"Gemm1NXdlPerWave", std::to_string(this->tile_desc.gemm1_n_Xdl_per_wave)},
{"ABlockTransferThreadClusterLengths_AK0_M_AK1",
this->a_block_transfer.thread_cluster_length},
{"ABlockTransferThreadClusterArrangeOrder",
this->a_block_transfer.thread_cluster_arrange_order},
{"ABlockTransferSrcAccessOrder", this->a_block_transfer.src_access_order},
{"ABlockTransferSrcVectorDim", std::to_string(this->a_block_transfer.src_vec_dim)},
{"ABlockTransferSrcScalarPerVector",
std::to_string(this->a_block_transfer.src_scalar_per_vector)},
{"ABlockTransferDstScalarPerVector_AK1",
std::to_string(this->a_block_transfer.dst_scalar_per_vector_k1)},
{"ABlockLdsExtraM", std::to_string(this->a_block_transfer.lds_add_extra_dim)},
{"A0BlockTransferThreadClusterLengths_AK0_M_AK1",
this->a0_block_transfer.thread_cluster_length},
{"A0BlockTransferThreadClusterArrangeOrder",
this->a0_block_transfer.thread_cluster_arrange_order},
{"A0BlockTransferSrcAccessOrder", this->a0_block_transfer.src_access_order},
{"A0BlockTransferSrcVectorDim", std::to_string(this->a0_block_transfer.src_vec_dim)},
{"A0BlockTransferSrcScalarPerVector",
std::to_string(this->a0_block_transfer.src_scalar_per_vector)},
{"A0BlockTransferDstScalarPerVector_AK1",
std::to_string(this->a0_block_transfer.dst_scalar_per_vector_k1)},
{"A0BlockLdsExtraM", std::to_string(this->a0_block_transfer.lds_add_extra_dim)},
{"B0BlockTransferThreadClusterLengths_BK0_N_BK1",
this->b0_block_transfer.thread_cluster_length},
{"B0BlockTransferThreadClusterArrangeOrder",
......@@ -405,6 +489,11 @@ Solution Operation_Xdl_CShuffle::ToSolution() const
{"B0BlockTransferDstScalarPerVector_BK1",
std::to_string(this->b0_block_transfer.dst_scalar_per_vector_k1)},
{"B0BlockLdsExtraN", std::to_string(this->b0_block_transfer.lds_add_extra_dim)},
{"CDE0BlockTransferSrcVectorDim", std::to_string(this->cde0_block_transfer.src_vec_dim)},
{"CDE0BlockTransferSrcScalarPerVector",
std::to_string(this->cde0_block_transfer.src_scalar_per_vector)},
{"B1BlockTransferThreadClusterLengths_BK0_N_BK1",
this->b1_block_transfer.thread_cluster_length},
{"B1BlockTransferThreadClusterArrangeOrder",
......@@ -416,20 +505,24 @@ Solution Operation_Xdl_CShuffle::ToSolution() const
{"B1BlockTransferDstScalarPerVector_BK1",
std::to_string(this->b1_block_transfer.dst_scalar_per_vector_k1)},
{"B1BlockLdsExtraN", std::to_string(this->b1_block_transfer.lds_add_extra_dim)},
{"CShuffleMXdlPerWavePerShuffle",
{"C1ShuffleMXdlPerWavePerShuffle",
std::to_string(this->cshuffle.m_Xdl_per_wave_per_shuffle)},
{"CShuffleNXdlPerWavePerShuffle",
{"C1ShuffleGemm0NXdlPerWavePerShuffle",
std::to_string(this->cshuffle.n_Xdl_per_wave_per_shuffle)},
{"CBlockTransferClusterLengths_MBlock_MWaveMPerXdl_NBlock_NWaveNPerXdl",
this->c_block_transfer.cluster_lengths_m_block_m_wave_m_per_Xdl_n_block_n_wave_n_per_Xdl},
{"CBlockTransferScalarPerVector_NWaveNPerXdl",
std::to_string(this->c_block_transfer.scalar_per_vector_n_wave_n_per_Xdl)},
{"CDE1ShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock",
this->cde1_block_transfer
.cluster_lengths_m_block_m_wave_m_per_Xdl_n_block_n_wave_n_per_Xdl},
{"CDE1ShuffleBlockTransferScalarPerVector_NPerBlock",
std::to_string(this->cde1_block_transfer.scalar_per_vector_n_wave_n_per_Xdl)},
};
return Solution{InterpolateString(DeviceBatchedGemmGemm_Xdl_CShuffleTemplate, values),
std::move(values)};
return Solution{
InterpolateString(DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffleTemplate, values),
std::move(values)};
}
} // namespace device_gemm_elementwise_gemm
} // namespace device_batched_gemm_multiple_d_gemm_multiple_d
} // namespace host
} // 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