Skip to content

GitLab

"vscode:/vscode.git/clone" did not exist on "ec25eea8f8237cf86c30703f59747e42f34b6f75"
Unverified Commit 29deceb6 authored by Illia Silin's avatar Illia Silin Committed by GitHub
Browse files

Merge pull request #18 from ROCmSoftwarePlatform/merge-from-public 

Merge from public
parents 91c1d147 c997bbf6
Hide whitespace changes
Inline Side-by-side
Showing with 2728 additions and 429 deletions
include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp
View file @ 29deceb6
...@@ -7,6 +7,7 @@ ...@@ -7,6 +7,7 @@
#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v1.hpp" #include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v1.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v2.hpp" #include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v2.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v4_direct_load.hpp"
namespace ck { namespace ck {
...@@ -14,6 +15,8 @@ enum struct PipelineVersion ...@@ -14,6 +15,8 @@ enum struct PipelineVersion
{ {
v1, v1,
v2, v2,
// v3 is only used in the Stream-K implementation.
v4,
}; };
template <PipelineVersion PipelineVer, template <PipelineVersion PipelineVer,
...@@ -36,6 +39,10 @@ constexpr auto GridwiseGemmPipeline_Selector() ...@@ -36,6 +39,10 @@ constexpr auto GridwiseGemmPipeline_Selector()
{ {
return GridwiseGemmPipeline_v2{}; return GridwiseGemmPipeline_v2{};
} }
else if constexpr(PipelineVer == PipelineVersion::v4)
{
return GridwiseGemmPipeline_v4<NumPrefetch>{};
}
else else
{ {
std::cerr << "GridwiseGemmPipeline configuration is not available" << std::endl; std::cerr << "GridwiseGemmPipeline configuration is not available" << std::endl;
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v4_direct_load.hpp 0 → 100644
View file @ 29deceb6
// SPDX-License-Identifier: MIT
// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/utility/common_header.hpp"
#include "ck/utility/loop_scheduler.hpp"
#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
namespace ck {
template <index_t NumPrefetch>
struct GridwiseGemmPipeline_v4;
// 1-stage prefetch
template <>
struct GridwiseGemmPipeline_v4<1>
{
static constexpr auto I0 = Number<0>{};
static constexpr auto I1 = Number<1>{};
__host__ __device__ static constexpr bool IsSupported(index_t /* num_loop */) { return true; }
__host__ __device__ static constexpr bool CalculateHasMainLoop(index_t num_loop)
{
return num_loop > 1;
}
template <bool HasMainLoop,
typename AGridDesc,
typename ABlockDesc,
typename ABlockTransfer,
typename AGridBuffer,
typename ABlockBuffer,
typename ABlockTransferStep,
typename BGridDesc,
typename BBlockDesc,
typename BBlockTransfer,
typename BGridBuffer,
typename BBlockBuffer,
typename BBlockTransferStep,
typename BlockwiseGemm,
typename CThreadBuffer>
__device__ static void Run(const AGridDesc& a_grid_desc,
const ABlockDesc& a_block_desc,
ABlockTransfer& a_blockwise_copy,
const AGridBuffer& a_grid_buf,
ABlockBuffer& a_block_buf,
const ABlockTransferStep& a_block_copy_step,
const BGridDesc& b_grid_desc,
const BBlockDesc& b_block_desc,
BBlockTransfer& b_blockwise_copy,
const BGridBuffer& b_grid_buf,
BBlockBuffer& b_block_buf,
const BBlockTransferStep& b_block_copy_step,
const BlockwiseGemm& blockwise_gemm,
CThreadBuffer& c_thread_buf,
index_t num_loop)
{
a_blockwise_copy.Run(a_grid_desc, a_grid_buf, a_block_desc, a_block_buf);
b_blockwise_copy.Run(b_grid_desc, b_grid_buf, b_block_desc, b_block_buf);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
// Initialize C
c_thread_buf.Clear();
// main body
if constexpr(HasMainLoop)
{
index_t i = 0;
do
{
block_sync_lds_direct_load();
blockwise_gemm.Run(a_block_buf, b_block_buf, c_thread_buf);
block_sync_lds_direct_load();
a_blockwise_copy.Run(a_grid_desc, a_grid_buf, a_block_desc, a_block_buf);
b_blockwise_copy.Run(b_grid_desc, b_grid_buf, b_block_desc, b_block_buf);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
++i;
} while(i < (num_loop - 1));
}
// tail
{
block_sync_lds_direct_load();
blockwise_gemm.Run(a_block_buf, b_block_buf, c_thread_buf);
}
}
};
} // namespace ck
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r3.hpp
View file @ 29deceb6
...@@ -996,6 +996,17 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3_ext ...@@ -996,6 +996,17 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3_ext
} }
} }
if constexpr(!(GemmSpec == tensor_operation::device::GemmSpecialization::KPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::MKPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::NKPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding))
{
if(!(problem.K0 % K0PerBlock == 0))
{
return false;
}
}
if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value) if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
{ {
if(problem.K % ABlockTransferSrcScalarPerVector != 0) if(problem.K % ABlockTransferSrcScalarPerVector != 0)
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4r2.hpp
View file @ 29deceb6
...@@ -136,7 +136,7 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2 ...@@ -136,7 +136,7 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
index_t MPadded; index_t MPadded;
index_t NPadded; index_t NPadded;
index_t KPadded; index_t KPadded;
index_t K0; index_t K0Padded;
index_t k_batch; index_t k_batch;
Argument(const FloatA* p_a_grid_, Argument(const FloatA* p_a_grid_,
...@@ -151,7 +151,7 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2 ...@@ -151,7 +151,7 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
index_t MPadded_, index_t MPadded_,
index_t NPadded_, index_t NPadded_,
index_t KPadded_, index_t KPadded_,
index_t K0_, index_t K0Padded_,
index_t k_batch_) index_t k_batch_)
: p_a_grid(p_a_grid_), : p_a_grid(p_a_grid_),
p_b_grid(p_b_grid_), p_b_grid(p_b_grid_),
...@@ -165,7 +165,7 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2 ...@@ -165,7 +165,7 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
MPadded(MPadded_), MPadded(MPadded_),
NPadded(NPadded_), NPadded(NPadded_),
KPadded(KPadded_), KPadded(KPadded_),
K0(K0_), K0Padded(K0Padded_),
k_batch(k_batch_) k_batch(k_batch_)
{ {
} }
...@@ -182,7 +182,7 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2 ...@@ -182,7 +182,7 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
<< "MP:" << MPadded << ", " << "MP:" << MPadded << ", "
<< "NP:" << NPadded << ", " << "NP:" << NPadded << ", "
<< "KP:" << KPadded << ", " << "KP:" << KPadded << ", "
<< "K0:" << K0 << ", " << "K0Padded:" << K0Padded << ", "
<< "KB:" << k_batch << "}" << std::endl; << "KB:" << k_batch << "}" << std::endl;
} }
}; };
...@@ -205,7 +205,7 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2 ...@@ -205,7 +205,7 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
return math::integer_least_multiple(N, NPerBlock); return math::integer_least_multiple(N, NPerBlock);
} }
__host__ __device__ static auto CalculateK0(index_t K, index_t K_Batch = 1) __host__ __device__ static auto CalculateK0Padded(index_t K, index_t K_Batch = 1)
{ {
// k_batch * k0 * k0_per_block * k1 // k_batch * k0 * k0_per_block * k1
auto K_t = K_Batch * K0PerBlock * K1; auto K_t = K_Batch * K0PerBlock * K1;
...@@ -214,8 +214,8 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2 ...@@ -214,8 +214,8 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
__host__ __device__ static auto CalculateKPadded(index_t K, index_t K_Batch = 1) __host__ __device__ static auto CalculateKPadded(index_t K, index_t K_Batch = 1)
{ {
auto K0 = CalculateK0(K, K_Batch); auto K0Padded = CalculateK0Padded(K, K_Batch);
return K_Batch * K0 * K1; return K_Batch * K0Padded * K1;
} }
__host__ __device__ static auto MakeAGridDescriptor_KBatch_K0_M_K1(index_t M, __host__ __device__ static auto MakeAGridDescriptor_KBatch_K0_M_K1(index_t M,
...@@ -223,7 +223,7 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2 ...@@ -223,7 +223,7 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
index_t K, index_t K,
index_t StrideA, index_t StrideA,
index_t KBatch, index_t KBatch,
index_t K0, index_t K0Padded,
index_t KPad) index_t KPad)
{ {
const auto a_grid_desc_m_k = [&]() { const auto a_grid_desc_m_k = [&]() {
...@@ -237,21 +237,33 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2 ...@@ -237,21 +237,33 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
} }
}(); }();
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 || if constexpr(GemmSpec == tensor_operation::device::GemmSpecialization::MPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding || GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::MKPadding || GemmSpec == tensor_operation::device::GemmSpecialization::MKPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding) GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding)
{ {
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>{}));
// const auto PadM = (MPerBlock - M % MPerBlock) % MPerBlock; // const auto PadM = (MPerBlock - M % MPerBlock) % MPerBlock;
return transform_tensor_descriptor( return transform_tensor_descriptor(
a_grid_desc_m_kpad, a_grid_desc_m_kpad,
make_tuple(make_unmerge_transform(make_tuple(KBatch, K0, K1)), make_tuple(make_unmerge_transform(make_tuple(KBatch, K0Padded, K1)),
make_right_pad_transform(M, MPad - M)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
}
else if constexpr(GemmSpec == tensor_operation::device::GemmSpecialization::MPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding)
{
// const auto PadM = (MPerBlock - M % MPerBlock) % MPerBlock;
return transform_tensor_descriptor(
a_grid_desc_m_k,
make_tuple(make_unmerge_transform(make_tuple(KBatch, K0Padded, K1)),
make_right_pad_transform(M, MPad - M)), make_right_pad_transform(M, MPad - M)),
make_tuple(Sequence<1>{}, Sequence<0>{}), make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{})); make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
...@@ -259,8 +271,8 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2 ...@@ -259,8 +271,8 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
else else
{ {
return transform_tensor_descriptor( return transform_tensor_descriptor(
a_grid_desc_m_kpad, a_grid_desc_m_k,
make_tuple(make_unmerge_transform(make_tuple(KBatch, K0, K1)), make_tuple(make_unmerge_transform(make_tuple(KBatch, K0Padded, K1)),
make_pass_through_transform(M)), make_pass_through_transform(M)),
make_tuple(Sequence<1>{}, Sequence<0>{}), make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{})); make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
...@@ -272,7 +284,7 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2 ...@@ -272,7 +284,7 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
index_t N, index_t N,
index_t StrideB, index_t StrideB,
index_t KBatch, index_t KBatch,
index_t K0, index_t K0Padded,
index_t KPad) index_t KPad)
{ {
const auto b_grid_desc_k_n = [&]() { const auto b_grid_desc_k_n = [&]() {
...@@ -286,21 +298,33 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2 ...@@ -286,21 +298,33 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
} }
}(); }();
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 || if constexpr(GemmSpec == tensor_operation::device::GemmSpecialization::NPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding || GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::NKPadding || GemmSpec == tensor_operation::device::GemmSpecialization::NKPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding) GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding)
{ {
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>{}));
// const auto PadN = (NPerBlock - N % NPerBlock) % NPerBlock; // const auto PadN = (NPerBlock - N % NPerBlock) % NPerBlock;
return transform_tensor_descriptor( return transform_tensor_descriptor(
b_grid_desc_kpad_n, b_grid_desc_kpad_n,
make_tuple(make_unmerge_transform(make_tuple(KBatch, K0, K1)), make_tuple(make_unmerge_transform(make_tuple(KBatch, K0Padded, K1)),
make_right_pad_transform(N, NPad - N)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
}
else if constexpr(GemmSpec == tensor_operation::device::GemmSpecialization::NPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding)
{
// const auto PadN = (NPerBlock - N % NPerBlock) % NPerBlock;
return transform_tensor_descriptor(
b_grid_desc_k_n,
make_tuple(make_unmerge_transform(make_tuple(KBatch, K0Padded, K1)),
make_right_pad_transform(N, NPad - N)), make_right_pad_transform(N, NPad - N)),
make_tuple(Sequence<0>{}, Sequence<1>{}), make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{})); make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
...@@ -308,8 +332,8 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2 ...@@ -308,8 +332,8 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
else else
{ {
return transform_tensor_descriptor( return transform_tensor_descriptor(
b_grid_desc_kpad_n, b_grid_desc_k_n,
make_tuple(make_unmerge_transform(make_tuple(KBatch, K0, K1)), make_tuple(make_unmerge_transform(make_tuple(KBatch, K0Padded, K1)),
make_pass_through_transform(N)), make_pass_through_transform(N)),
make_tuple(Sequence<0>{}, Sequence<1>{}), make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{})); make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
...@@ -398,6 +422,7 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2 ...@@ -398,6 +422,7 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
return false; return false;
} }
} }
if constexpr(!(GemmSpec == tensor_operation::device::GemmSpecialization::NPadding || if constexpr(!(GemmSpec == tensor_operation::device::GemmSpecialization::NPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding || GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::NKPadding || GemmSpec == tensor_operation::device::GemmSpecialization::NKPadding ||
...@@ -410,6 +435,25 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2 ...@@ -410,6 +435,25 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
<< __FILE__ << ":" << __LINE__ << ", in function: " << __func__ << __FILE__ << ":" << __LINE__ << ", in function: " << __func__
<< std::endl; << std::endl;
#endif // DEBUG_LOG
return false;
}
}
if constexpr(!(GemmSpec == tensor_operation::device::GemmSpecialization::KPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::MKPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::NKPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding))
{
auto K_t = karg.k_batch * K0PerBlock * K1;
if(!(karg.K % K_t == 0))
{
#if DEBUG_LOG
std::cout << "Arg K value is not a multiple of K_Batch * K0PerBlock * K1! K: "
<< karg.K << " " << __FILE__ << ":" << __LINE__
<< ", in function: " << __func__ << std::endl;
#endif // DEBUG_LOG #endif // DEBUG_LOG
return false; return false;
} }
...@@ -478,11 +522,11 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2 ...@@ -478,11 +522,11 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
if(karg.N % CBlockTransferScalarPerVector_NWaveNPerXDL != 0) if(karg.N % CBlockTransferScalarPerVector_NWaveNPerXDL != 0)
{ {
#if DEBUG_LOG #if DEBUG_LOG
std::cout std::cout << "Arg N (" << karg.N
<< "Arg N (" << karg.N << ") value is not a multiple of "
<< ") value is not a multiple of CBlockTransferScalarPerVector_NWaveNPerXDL (" "CBlockTransferScalarPerVector_NWaveNPerXDL ("
<< CBlockTransferScalarPerVector_NWaveNPerXDL << " )! " << __FILE__ << ":" << CBlockTransferScalarPerVector_NWaveNPerXDL << " )! " << __FILE__ << ":"
<< __LINE__ << ", in function: " << __func__ << std::endl; << __LINE__ << ", in function: " << __func__ << std::endl;
#endif // DEBUG_LOG #endif // DEBUG_LOG
return false; return false;
...@@ -493,25 +537,25 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2 ...@@ -493,25 +537,25 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
if(karg.M % CBlockTransferScalarPerVector_NWaveNPerXDL != 0) if(karg.M % CBlockTransferScalarPerVector_NWaveNPerXDL != 0)
{ {
#if DEBUG_LOG #if DEBUG_LOG
std::cout std::cout << "Arg M (" << karg.M
<< "Arg M (" << karg.M << ") value is not a multiple of "
<< ") value is not a multiple of CBlockTransferScalarPerVector_NWaveNPerXDL (" "CBlockTransferScalarPerVector_NWaveNPerXDL ("
<< CBlockTransferScalarPerVector_NWaveNPerXDL << " )! " << __FILE__ << ":" << CBlockTransferScalarPerVector_NWaveNPerXDL << " )! " << __FILE__ << ":"
<< __LINE__ << ", in function: " << __func__ << std::endl; << __LINE__ << ", in function: " << __func__ << std::endl;
#endif // DEBUG_LOG #endif // DEBUG_LOG
return false; return false;
} }
} }
const auto num_k_loop = karg.K0 / K0PerBlock; const auto num_k_loop = karg.K0Padded / K0PerBlock;
if(!GridwiseGemmPipe::IsSupported(num_k_loop)) if(!GridwiseGemmPipe::IsSupported(num_k_loop))
{ {
#if DEBUG_LOG #if DEBUG_LOG
std::cout << "The number of k loops (" << num_k_loop std::cout << "The number of k loops (" << num_k_loop
<< ") value is not supported by GridwiseGemm Pipeline." << ") value is not supported by GridwiseGemm Pipeline."
<< " K0: " << karg.K0 << ", K0PerBlock: " << K0PerBlock << " " << __FILE__ << " K0Padded: " << karg.K0Padded << ", K0PerBlock: " << K0PerBlock << " "
<< ":" << __LINE__ << ", in function: " << __func__ << std::endl; << __FILE__ << ":" << __LINE__ << ", in function: " << __func__ << std::endl;
#endif // DEBUG_LOG #endif // DEBUG_LOG
return false; return false;
} }
...@@ -521,14 +565,15 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2 ...@@ -521,14 +565,15 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
__host__ __device__ static auto GetKPad(index_t K, index_t KBatch) __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 K0Padded =
const index_t KPad = KBatch * K0 * K1; math::integer_divide_ceil(K, K1 * K0PerBlock * KBatch) * K0PerBlock;
const index_t KPad = KBatch * K0Padded * K1;
return KPad; return KPad;
} }
__host__ __device__ static constexpr bool CalculateHasMainK0BlockLoop(index_t K0) __host__ __device__ static constexpr bool CalculateHasMainK0BlockLoop(index_t K0Padded)
{ {
const index_t num_loop = K0 / K0PerBlock; const index_t num_loop = K0Padded / K0PerBlock;
return GridwiseGemmPipe::CalculateHasMainLoop(num_loop); return GridwiseGemmPipe::CalculateHasMainLoop(num_loop);
} }
...@@ -595,9 +640,9 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2 ...@@ -595,9 +640,9 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
const FloatB* p_b_grid = karg.p_b_grid; const FloatB* p_b_grid = karg.p_b_grid;
FloatC* p_c_grid = karg.p_c_grid; FloatC* p_c_grid = karg.p_c_grid;
const auto a_b_k0_m_k1_grid_desc = MakeAGridDescriptor_KBatch_K0_M_K1( 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); karg.M, karg.MPadded, karg.K, karg.StrideA, karg.k_batch, karg.K0Padded, karg.KPadded);
const auto b_b_k0_n_k1_grid_desc = MakeBGridDescriptor_KBatch_K0_N_K1( 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); karg.K, karg.NPadded, karg.N, karg.StrideB, karg.k_batch, karg.K0Padded, karg.KPadded);
const auto c_grid_desc_m_n = MakeCGridDescriptor_M_N(karg.M, karg.N, karg.StrideC); const auto c_grid_desc_m_n = MakeCGridDescriptor_M_N(karg.M, karg.N, karg.StrideC);
const auto c_grid_desc_mblock_mperblock_nblock_nperblock = const auto c_grid_desc_mblock_mperblock_nblock_nperblock =
......
include/ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_bwd_gamma_beta.hpp 0 → 100644
View file @ 29deceb6
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/utility/data_type.hpp"
#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/block/reduction_functions_blockwise.hpp"
namespace ck {
// dgamma = reduce_sum(dy * (x - mean) * inv_std)
// dbeta = reduce_sum(dy)
template <typename DYDataType,
typename XDataType,
typename MeanInvStdDataType,
typename ComputeDataType,
typename DGammaDataType,
typename DBetaDataType,
typename GridDesc_M_K,
typename GridDesc_M,
index_t BlockSize,
index_t MThreadClusterSize,
index_t KThreadClusterSize,
index_t MThreadSliceSize,
index_t KThreadSliceSize,
index_t DYSrcVectorDim,
index_t DYSrcVectorSize,
index_t XSrcVectorDim,
index_t XSrcVectorSize,
index_t MeanInvStdSrcVectorDim,
index_t MeanInvStdSrcVectorSize,
index_t DGammaDstVectorSize,
index_t DBetaDstVectorSize>
struct GridwiseNormalizationBwdGammaBeta_mk_to_k
{
// if we just check ThreadSliceSize & VectorSize == 0, the performance may be poor
static_assert(((DYSrcVectorDim == 0 && MThreadSliceSize == DYSrcVectorSize) ||
(DYSrcVectorDim == 1 && KThreadSliceSize == DYSrcVectorSize)),
"Invalid thread slice sizes and/or dy vector sizes configuration, please check!");
static_assert(((XSrcVectorDim == 0 && MThreadSliceSize == XSrcVectorSize) ||
(XSrcVectorDim == 1 && KThreadSliceSize == XSrcVectorSize)),
"Invalid thread slice sizes and/or x vector sizes configuration, please check!");
using ThreadClusterLengths_M_K = Sequence<MThreadClusterSize, KThreadClusterSize>;
using DYThreadBufferDimAccessOrder =
typename conditional<DYSrcVectorDim == 0, Sequence<1, 0>, Sequence<0, 1>>::type;
using XThreadBufferDimAccessOrder =
typename conditional<XSrcVectorDim == 0, Sequence<1, 0>, Sequence<0, 1>>::type;
using MeanInvStdThreadBufferDimAccessOrder =
typename conditional<MeanInvStdSrcVectorDim == 0, Sequence<1, 0>, Sequence<0, 1>>::type;
using ThreadClusterArrangeOrder = DYThreadBufferDimAccessOrder;
static constexpr auto thread_cluster_desc =
make_cluster_descriptor(ThreadClusterLengths_M_K{}, ThreadClusterArrangeOrder{});
using ThreadBufferLengths_M_K = Sequence<MThreadSliceSize, KThreadSliceSize>;
using ThreadBufferLengths_M = Sequence<MThreadSliceSize>;
static constexpr auto thread_buffer_desc_m_k = make_naive_tensor_descriptor_packed(
make_tuple(Number<MThreadSliceSize>{}, Number<KThreadSliceSize>{}));
static constexpr auto thread_buffer_desc_m =
make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{}));
using PassThroughOp = tensor_operation::element_wise::PassThrough;
using BlockwiseSumReduce = PartitionedBlockwiseReduction<ComputeDataType,
BlockSize,
ThreadClusterLengths_M_K,
ThreadClusterArrangeOrder,
reduce::Add,
true>;
static constexpr auto I0 = Number<0>{};
static constexpr auto I1 = Number<1>{};
static constexpr index_t M_BlockTileSize = MThreadClusterSize * MThreadSliceSize;
static constexpr index_t K_BlockTileSize = KThreadClusterSize * KThreadSliceSize;
__device__ static void Run(const GridDesc_M_K& dy_grid_desc_m_k,
const GridDesc_M_K& x_grid_desc_m_k,
const GridDesc_M_K& mean_grid_desc_m_k,
const GridDesc_M_K& inv_std_grid_desc_m_k,
const GridDesc_M& dgamma_grid_desc_m,
const GridDesc_M& dbeta_grid_desc_m,
index_t num_k_block_tile_iteration,
const DYDataType* const __restrict__ p_dy_global,
const XDataType* const __restrict__ p_x_global,
const MeanInvStdDataType* const __restrict__ p_mean_global,
const MeanInvStdDataType* const __restrict__ p_inv_std_global,
DGammaDataType* const __restrict__ p_dgamma_global,
DBetaDataType* const __restrict__ p_dbeta_global)
{
// LDS
__shared__ ComputeDataType p_reduce_work_buffer[BlockSize];
auto reduce_work_buf =
make_dynamic_buffer<AddressSpaceEnum::Lds>(p_reduce_work_buffer, BlockSize);
// Global
const auto dy_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_dy_global, dy_grid_desc_m_k.GetElementSpaceSize());
const auto x_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_x_global, x_grid_desc_m_k.GetElementSpaceSize());
const auto mean_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_mean_global, mean_grid_desc_m_k.GetElementSpaceSize());
const auto inv_std_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_inv_std_global, inv_std_grid_desc_m_k.GetElementSpaceSize());
auto dgamma_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_dgamma_global, dgamma_grid_desc_m.GetElementSpaceSize());
auto dbeta_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_dbeta_global, dbeta_grid_desc_m.GetElementSpaceSize());
// VGPR
auto dy_thread_buf = StaticBuffer<AddressSpaceEnum::Vgpr,
ComputeDataType,
MThreadSliceSize * KThreadSliceSize,
true>{};
auto x_thread_buf = StaticBuffer<AddressSpaceEnum::Vgpr,
ComputeDataType,
MThreadSliceSize * KThreadSliceSize,
true>{};
auto mean_thread_buf = StaticBuffer<AddressSpaceEnum::Vgpr,
ComputeDataType,
MThreadSliceSize * KThreadSliceSize,
true>{};
auto inv_std_thread_buf = StaticBuffer<AddressSpaceEnum::Vgpr,
ComputeDataType,
MThreadSliceSize * KThreadSliceSize,
true>{};
auto dgamma_thread_buf =
StaticBuffer<AddressSpaceEnum::Vgpr, ComputeDataType, MThreadSliceSize, true>{};
auto dbeta_thread_buf =
StaticBuffer<AddressSpaceEnum::Vgpr, ComputeDataType, MThreadSliceSize, true>{};
const index_t thread_local_id = get_thread_local_1d_id();
const index_t block_global_id = get_block_1d_id();
const auto thread_cluster_idx =
thread_cluster_desc.CalculateBottomIndex(make_multi_index(thread_local_id));
const auto thread_m_cluster_id = thread_cluster_idx[I0];
const auto thread_k_cluster_id = thread_cluster_idx[I1];
// IO
auto threadwise_dy_load = ThreadwiseTensorSliceTransfer_v2<DYDataType,
ComputeDataType,
GridDesc_M_K,
decltype(thread_buffer_desc_m_k),
ThreadBufferLengths_M_K,
DYThreadBufferDimAccessOrder,
DYSrcVectorDim,
DYSrcVectorSize,
1,
true>(
dy_grid_desc_m_k,
make_multi_index(block_global_id * M_BlockTileSize +
thread_m_cluster_id * MThreadSliceSize,
thread_k_cluster_id * KThreadSliceSize));
auto threadwise_x_load = ThreadwiseTensorSliceTransfer_v2<XDataType,
ComputeDataType,
GridDesc_M_K,
decltype(thread_buffer_desc_m_k),
ThreadBufferLengths_M_K,
XThreadBufferDimAccessOrder,
XSrcVectorDim,
XSrcVectorSize,
1,
true>(
x_grid_desc_m_k,
make_multi_index(block_global_id * M_BlockTileSize +
thread_m_cluster_id * MThreadSliceSize,
thread_k_cluster_id * KThreadSliceSize));
auto threadwise_mean_load =
ThreadwiseTensorSliceTransfer_v2<MeanInvStdDataType,
ComputeDataType,
GridDesc_M_K,
decltype(thread_buffer_desc_m_k),
ThreadBufferLengths_M_K,
MeanInvStdThreadBufferDimAccessOrder,
MeanInvStdSrcVectorDim,
MeanInvStdSrcVectorSize,
1,
true>(
mean_grid_desc_m_k,
make_multi_index(block_global_id * M_BlockTileSize +
thread_m_cluster_id * MThreadSliceSize,
thread_k_cluster_id * KThreadSliceSize));
auto threadwise_inv_std_load =
ThreadwiseTensorSliceTransfer_v2<MeanInvStdDataType,
ComputeDataType,
GridDesc_M_K,
decltype(thread_buffer_desc_m_k),
ThreadBufferLengths_M_K,
MeanInvStdThreadBufferDimAccessOrder,
MeanInvStdSrcVectorDim,
MeanInvStdSrcVectorSize,
1,
true>(
inv_std_grid_desc_m_k,
make_multi_index(block_global_id * M_BlockTileSize +
thread_m_cluster_id * MThreadSliceSize,
thread_k_cluster_id * KThreadSliceSize));
auto threadwise_dgamma_store =
ThreadwiseTensorSliceTransfer_v1r3<ComputeDataType,
DGammaDataType,
decltype(thread_buffer_desc_m),
GridDesc_M,
PassThroughOp,
ThreadBufferLengths_M,
Sequence<0>,
0,
DGammaDstVectorSize,
InMemoryDataOperationEnum::Set,
1,
true>(
dgamma_grid_desc_m,
make_multi_index(block_global_id * M_BlockTileSize +
thread_m_cluster_id * MThreadSliceSize),
PassThroughOp{});
auto threadwise_dbeta_store =
ThreadwiseTensorSliceTransfer_v1r3<ComputeDataType,
DBetaDataType,
decltype(thread_buffer_desc_m),
GridDesc_M,
PassThroughOp,
ThreadBufferLengths_M,
Sequence<0>,
0,
DBetaDstVectorSize,
InMemoryDataOperationEnum::Set,
1,
true>(
dbeta_grid_desc_m,
make_multi_index(block_global_id * M_BlockTileSize +
thread_m_cluster_id * MThreadSliceSize),
PassThroughOp{});
static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
dgamma_thread_buf(I) = type_convert<ComputeDataType>(0.0f);
dbeta_thread_buf(I) = type_convert<ComputeDataType>(0.0f);
});
constexpr auto thread_copy_fwd_step_m_k = make_multi_index(0, K_BlockTileSize);
for(index_t reducedTiles = 0; reducedTiles < num_k_block_tile_iteration; ++reducedTiles)
{
threadwise_dy_load.Run(dy_grid_desc_m_k,
dy_global_val_buf,
thread_buffer_desc_m_k,
make_tuple(I0, I0),
dy_thread_buf);
threadwise_x_load.Run(x_grid_desc_m_k,
x_global_val_buf,
thread_buffer_desc_m_k,
make_tuple(I0, I0),
x_thread_buf);
threadwise_mean_load.Run(mean_grid_desc_m_k,
mean_global_val_buf,
thread_buffer_desc_m_k,
make_tuple(I0, I0),
mean_thread_buf);
threadwise_inv_std_load.Run(inv_std_grid_desc_m_k,
inv_std_global_val_buf,
thread_buffer_desc_m_k,
make_tuple(I0, I0),
inv_std_thread_buf);
threadwise_dy_load.MoveSrcSliceWindow(dy_grid_desc_m_k, thread_copy_fwd_step_m_k);
threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_fwd_step_m_k);
threadwise_mean_load.MoveSrcSliceWindow(mean_grid_desc_m_k, thread_copy_fwd_step_m_k);
threadwise_inv_std_load.MoveSrcSliceWindow(inv_std_grid_desc_m_k,
thread_copy_fwd_step_m_k);
static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
constexpr auto offset_m =
Number<thread_buffer_desc_m.CalculateOffset(make_tuple(iM))>{};
static_for<0, KThreadSliceSize, 1>{}([&](auto iK) {
constexpr auto offset_m_k =
Number<thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK))>{};
dgamma_thread_buf(offset_m) +=
dy_thread_buf[offset_m_k] * inv_std_thread_buf[offset_m_k] *
(x_thread_buf[offset_m_k] - mean_thread_buf[offset_m_k]);
dbeta_thread_buf(offset_m) += dy_thread_buf[offset_m_k];
});
});
}
static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
if constexpr(I > 0)
block_sync_lds();
BlockwiseSumReduce::Reduce(reduce_work_buf, dbeta_thread_buf(I));
block_sync_lds();
BlockwiseSumReduce::Reduce(reduce_work_buf, dgamma_thread_buf(I));
});
if(thread_k_cluster_id == 0)
{
threadwise_dgamma_store.Run(thread_buffer_desc_m,
make_tuple(I0),
dgamma_thread_buf,
dgamma_grid_desc_m,
dgamma_global_val_buf);
threadwise_dbeta_store.Run(thread_buffer_desc_m,
make_tuple(I0),
dbeta_thread_buf,
dbeta_grid_desc_m,
dbeta_global_val_buf);
}
}
};
} // namespace ck
include/ck/utility/amd_buffer_addressing.hpp
View file @ 29deceb6
...@@ -944,4 +944,41 @@ amd_buffer_atomic_max(const typename vector_type_maker<T, N>::type::type src_thr ...@@ -944,4 +944,41 @@ amd_buffer_atomic_max(const typename vector_type_maker<T, N>::type::type src_thr
#endif #endif
} }
// Direct loads from global to LDS.
__device__ void
llvm_amdgcn_raw_buffer_load_lds(int32x4_t rsrc,
__attribute__((address_space(3))) uint32_t* lds_ptr,
index_t size,
index_t voffset,
index_t soffset,
index_t offset,
index_t aux) __asm("llvm.amdgcn.raw.buffer.load.lds");
template <typename T, index_t NumElemsPerThread>
__device__ void amd_direct_load_global_to_lds(const T* global_base_ptr,
const index_t global_offset,
T* lds_base_ptr,
const index_t lds_offset,
const bool is_valid,
const index_t src_element_space_size)
{
// Direct loads require that each thread reads and writes exactly a single DWORD.
constexpr auto dword_bytes = 4;
constexpr auto bytes_per_thread = sizeof(T) * NumElemsPerThread;
static_assert(bytes_per_thread == dword_bytes);
const uint32_t* global_ptr =
reinterpret_cast<uint32_t*>(reinterpret_cast<uintptr_t>(global_base_ptr));
const int32x4_t src_resource = make_wave_buffer_resource(global_ptr, src_element_space_size);
const index_t global_offset_bytes = is_valid ? global_offset * sizeof(T) : 0x80000000;
// LDS pointer must be attributed with the LDS address space.
__attribute__((address_space(3))) uint32_t* lds_ptr =
reinterpret_cast<__attribute__((address_space(3))) uint32_t*>(
reinterpret_cast<uintptr_t>(lds_base_ptr + lds_offset));
llvm_amdgcn_raw_buffer_load_lds(
src_resource, lds_ptr, sizeof(uint32_t), global_offset_bytes, 0, 0, 0);
}
} // namespace ck } // namespace ck
include/ck/utility/dynamic_buffer.hpp
View file @ 29deceb6
...@@ -173,6 +173,26 @@ struct DynamicBuffer ...@@ -173,6 +173,26 @@ struct DynamicBuffer
} }
} }
template <typename DstBuffer, index_t NumElemsPerThread>
__host__ __device__ void DirectCopyToLds(DstBuffer& dst_buf,
index_t src_offset,
index_t dst_offset,
bool is_valid_element) const
{
// Copy data from global to LDS memory using direct loads.
static_assert(GetAddressSpace() == AddressSpaceEnum::Global,
"Source data must come from a global memory buffer.");
static_assert(DstBuffer::GetAddressSpace() == AddressSpaceEnum::Lds,
"Destination data must be stored in an LDS memory buffer.");
amd_direct_load_global_to_lds<T, NumElemsPerThread>(p_data_,
src_offset,
dst_buf.p_data_,
dst_offset,
is_valid_element,
element_space_size_);
}
template <typename X, template <typename X,
typename enable_if<is_same<typename scalar_type<remove_cvref_t<X>>::type, typename enable_if<is_same<typename scalar_type<remove_cvref_t<X>>::type,
typename scalar_type<remove_cvref_t<T>>::type>::value, typename scalar_type<remove_cvref_t<T>>::type>::value,
......
include/ck/utility/synchronization.hpp
View file @ 29deceb6
...@@ -19,6 +19,15 @@ __device__ void block_sync_lds() ...@@ -19,6 +19,15 @@ __device__ void block_sync_lds()
#endif #endif
} }
__device__ void block_sync_lds_direct_load()
{
asm volatile("\
s_waitcnt vmcnt(0) \n \
s_waitcnt lgkmcnt(0) \n \
s_barrier \
" ::);
}
__device__ void s_nop() __device__ void s_nop()
{ {
#if 1 #if 1
......
include/ck/utility/type_convert.hpp
View file @ 29deceb6
...@@ -95,10 +95,16 @@ inline __host__ __device__ constexpr bhalf_t type_convert<bhalf_t, int8_t>(int8_ ...@@ -95,10 +95,16 @@ inline __host__ __device__ constexpr bhalf_t type_convert<bhalf_t, int8_t>(int8_
return type_convert<bhalf_t>(x_fp32); return type_convert<bhalf_t>(x_fp32);
} }
// convert fp32 to fp8 // Declare a template function for fp8 conversion using SR
template <typename Y, typename X>
__host__ __device__ constexpr Y f8_convert_sr(X x);
// convert fp32 to fp8 with stochastic rounding
template <> template <>
inline __host__ __device__ f8_t type_convert<f8_t, float>(float x) inline __host__ __device__ f8_t f8_convert_sr<f8_t, float>(float x)
{ {
constexpr int seed = 42;
uint32_t rng = prand_generator<float, seed>(reinterpret_cast<uintptr_t>(&x), x);
#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__) #if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
float max_fp8 = 240.0f; float max_fp8 = 240.0f;
x = x > max_fp8 ? max_fp8 : (x < -max_fp8 ? -max_fp8 : x); x = x > max_fp8 ? max_fp8 : (x < -max_fp8 ? -max_fp8 : x);
...@@ -110,70 +116,139 @@ inline __host__ __device__ f8_t type_convert<f8_t, float>(float x) ...@@ -110,70 +116,139 @@ inline __host__ __device__ f8_t type_convert<f8_t, float>(float x)
} val; } val;
val.fval = x; val.fval = x;
uint32_t ival = 0; uint32_t ival = 0;
ival = __builtin_amdgcn_cvt_pk_fp8_f32(val.fval, val.fval, ival, false); // false -> WORD0 ival = __builtin_amdgcn_cvt_sr_fp8_f32(val.fval, rng, ival, 0); // 0 pos
val.i32val = ival; val.i32val = ival;
return val.i8val[0]; return val.i8val[0]; // little endian
#else #else
constexpr bool negative_zero_nan = true; constexpr bool negative_zero_nan = true;
constexpr bool clip = true; constexpr bool clip = true;
constexpr f8_rounding_mode rm = f8_rounding_mode::standard; constexpr f8_rounding_mode rm = f8_rounding_mode::stochastic;
constexpr uint32_t rng = 0;
return utils:: return utils::
cast_to_f8<float, f8_t, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>(x, cast_to_f8<float, f8_t, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>(x,
rng); rng);
#endif #endif
} }
// convert fp8 to fp32 // convert fp16 to fp8 with stochastic rounding
template <> template <>
inline __host__ __device__ float type_convert<float, f8_t>(f8_t x) inline __host__ __device__ f8_t f8_convert_sr<f8_t, half_t>(half_t x)
{ {
#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__) #if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
float fval; // convert to float and use native converion
uint32_t i32val = static_cast<uint32_t>(x); return f8_convert_sr<f8_t>(type_convert<float>(x));
fval = __builtin_amdgcn_cvt_f32_fp8(i32val, 0);
// asm volatile("v_cvt_f32_fp8 %0, %1 src0_sel:BYTE_0" : "=v"(fval) : "v"(i32val));
return fval;
#else #else
constexpr bool negative_zero_nan = true; constexpr bool negative_zero_nan = true;
return utils::cast_from_f8<f8_t, float, negative_zero_nan>(x); constexpr bool clip = true;
constexpr f8_rounding_mode rm = f8_rounding_mode::stochastic;
constexpr int seed = 42;
uint32_t rng = prand_generator<half_t, seed>(reinterpret_cast<uintptr_t>(&x), x);
return utils::
cast_to_f8<half_t, f8_t, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>(
x, rng);
#endif #endif
} }
// convert fp16 to fp8 // convert fp32 to bf8 with stochastic rounding
template <> template <>
inline __host__ __device__ f8_t type_convert<f8_t, half_t>(half_t x) inline __host__ __device__ bf8_t f8_convert_sr<bf8_t, float>(float x)
{
constexpr int seed = 42;
uint32_t rng = prand_generator<float, seed>(reinterpret_cast<uintptr_t>(&x), x);
#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
union
{
float fval;
uint32_t i32val;
uint8_t i8val[4]; // not endian independent
} val;
val.fval = x;
uint32_t ival = 0;
ival = __builtin_amdgcn_cvt_sr_bf8_f32(val.fval, rng, ival, 0); // 0 pos
val.i32val = ival;
return val.i8val[0]; // little endian
#else
constexpr bool negative_zero_nan = true;
constexpr bool clip = true;
constexpr f8_rounding_mode rm = f8_rounding_mode::stochastic;
return utils::
cast_to_f8<float, bf8_t, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>(
x, rng);
#endif
}
// convert fp16 to bf8 with stochastic rounding
template <>
inline __host__ __device__ bf8_t f8_convert_sr<bf8_t, half_t>(half_t x)
{ {
#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__) #if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
// convert to float and use native converion // convert to float and use native converion
return type_convert<f8_t>(type_convert<float>(x)); return f8_convert_sr<f8_t>(type_convert<float>(x));
#else
constexpr bool negative_zero_nan = true;
constexpr bool clip = true;
constexpr f8_rounding_mode rm = f8_rounding_mode::stochastic;
constexpr int seed = 42;
uint32_t rng = prand_generator<half_t, seed>(reinterpret_cast<uintptr_t>(&x), x);
return utils::
cast_to_f8<half_t, bf8_t, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>(
x, rng);
#endif
}
// Declare a template function for fp8 conversion using RNE
template <typename Y, typename X>
__host__ __device__ constexpr Y f8_convert_rne(X x);
// convert fp32 to fp8 with rounding to nearest even
template <>
inline __host__ __device__ f8_t f8_convert_rne<f8_t, float>(float x)
{
#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
float max_fp8 = 240.0f;
x = x > max_fp8 ? max_fp8 : (x < -max_fp8 ? -max_fp8 : x);
union
{
float fval;
uint32_t i32val;
uint8_t i8val[4]; // not endian independent
} val;
val.fval = x;
uint32_t ival = 0;
ival = __builtin_amdgcn_cvt_pk_fp8_f32(val.fval, val.fval, ival, false); // false -> WORD0
val.i32val = ival;
return val.i8val[0];
#else #else
constexpr bool negative_zero_nan = true; constexpr bool negative_zero_nan = true;
constexpr bool clip = true; constexpr bool clip = true;
constexpr f8_rounding_mode rm = f8_rounding_mode::standard; constexpr f8_rounding_mode rm = f8_rounding_mode::standard;
constexpr uint32_t rng = 0; constexpr uint32_t rng = 0;
return utils:: return utils::
cast_to_f8<half_t, f8_t, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>( cast_to_f8<float, f8_t, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>(x,
x, rng); rng);
#endif #endif
} }
// convert fp8 to fp16 // convert fp16 to fp8 with rounding to nearest even
template <> template <>
inline __host__ __device__ half_t type_convert<half_t, f8_t>(f8_t x) inline __host__ __device__ f8_t f8_convert_rne<f8_t, half_t>(half_t x)
{ {
#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__) #if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
// use native conversion to float and convert to fp16 // convert to float and use native converion
return type_convert<half_t>(type_convert<float>(x)); return f8_convert_rne<f8_t>(type_convert<float>(x));
#else #else
constexpr bool negative_zero_nan = true; constexpr bool negative_zero_nan = true;
return utils::cast_from_f8<f8_t, half_t, negative_zero_nan>(x); constexpr bool clip = true;
constexpr f8_rounding_mode rm = f8_rounding_mode::standard;
constexpr uint32_t rng = 0;
return utils::
cast_to_f8<half_t, f8_t, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>(
x, rng);
#endif #endif
} }
// convert fp32 to bf8 // convert fp32 to bf8 with rounding to nearest even
template <> template <>
inline __host__ __device__ bf8_t type_convert<bf8_t, float>(float x) inline __host__ __device__ bf8_t f8_convert_rne<bf8_t, float>(float x)
{ {
#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__) #if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
union union
...@@ -198,6 +273,116 @@ inline __host__ __device__ bf8_t type_convert<bf8_t, float>(float x) ...@@ -198,6 +273,116 @@ inline __host__ __device__ bf8_t type_convert<bf8_t, float>(float x)
#endif #endif
} }
// convert fp16 to bf8 with rounding to nearest even
template <>
inline __host__ __device__ bf8_t f8_convert_rne<bf8_t, half_t>(half_t x)
{
#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
// convert to float and use native converion
return f8_convert_rne<bf8_t>(type_convert<float>(x));
#else
constexpr bool negative_zero_nan = true;
constexpr bool clip = true;
constexpr f8_rounding_mode rm = f8_rounding_mode::standard;
constexpr uint32_t rng = 0;
return utils::
cast_to_f8<half_t, bf8_t, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>(
x, rng);
#endif
}
// convert fp32 to fp8
template <>
inline __host__ __device__ f8_t type_convert<f8_t, float>(float x)
{
#if defined CK_USE_SR_F8_CONVERSION
return f8_convert_sr<f8_t>(x);
#else
return f8_convert_rne<f8_t>(x);
#endif
}
// convert fp8 to fp32
template <>
inline __host__ __device__ float type_convert<float, f8_t>(f8_t x)
{
#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
float fval;
uint32_t i32val = static_cast<uint32_t>(x);
fval = __builtin_amdgcn_cvt_f32_fp8(i32val, 0);
// asm volatile("v_cvt_f32_fp8 %0, %1 src0_sel:BYTE_0" : "=v"(fval) : "v"(i32val));
return fval;
#else
constexpr bool negative_zero_nan = true;
return utils::cast_from_f8<f8_t, float, negative_zero_nan>(x);
#endif
}
template <>
inline __host__ __device__ float2_t type_convert<float2_t, f8x2_t>(f8x2_t x)
{
#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
const auto i16val = bit_cast<uint16_t>(x);
return __builtin_amdgcn_cvt_pk_f32_fp8(i16val, 0);
#else
constexpr bool negative_zero_nan = true;
const auto f8x2_v = vector_type<f8_t, 2>(x);
vector_type<float, 2> f32x2_v;
f32x2_v.template AsType<float>()(Number<0>{}) =
utils::cast_from_f8<f8_t, float, negative_zero_nan>(
f8x2_v.template AsType<f8_t>()[Number<0>{}]);
f32x2_v.template AsType<float>()(Number<1>{}) =
utils::cast_from_f8<f8_t, float, negative_zero_nan>(
f8x2_v.template AsType<f8_t>()[Number<1>{}]);
return f32x2_v.template AsType<float2_t>()[Number<0>{}];
#endif
}
template <>
inline __host__ __device__ half2_t type_convert<half2_t, float2_t>(float2_t x)
{
const vector_type<float, 2> f32x2_v(x);
const auto y = __builtin_amdgcn_cvt_pkrtz(f32x2_v.template AsType<float>()[Number<0>{}],
f32x2_v.template AsType<float>()[Number<1>{}]);
return bit_cast<half2_t>(y);
}
// convert fp16 to fp8
template <>
inline __host__ __device__ f8_t type_convert<f8_t, half_t>(half_t x)
{
#if defined CK_USE_SR_F8_CONVERSION
return f8_convert_sr<f8_t>(x);
#else
return f8_convert_nre<f8_t>(x);
#endif
}
// convert fp8 to fp16
template <>
inline __host__ __device__ half_t type_convert<half_t, f8_t>(f8_t x)
{
#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
// use native conversion to float and convert to fp16
return type_convert<half_t>(type_convert<float>(x));
#else
constexpr bool negative_zero_nan = true;
return utils::cast_from_f8<f8_t, half_t, negative_zero_nan>(x);
#endif
}
// convert fp32 to bf8
template <>
inline __host__ __device__ bf8_t type_convert<bf8_t, float>(float x)
{
#if defined CK_USE_SR_F8_CONVERSION
return f8_convert_sr<bf8_t>(x);
#else
return f8_convert_rne<bf8_t>(x);
#endif
}
// convert bf8 to fp32 // convert bf8 to fp32
template <> template <>
inline __host__ __device__ float type_convert<float, bf8_t>(bf8_t x) inline __host__ __device__ float type_convert<float, bf8_t>(bf8_t x)
...@@ -218,17 +403,10 @@ inline __host__ __device__ float type_convert<float, bf8_t>(bf8_t x) ...@@ -218,17 +403,10 @@ inline __host__ __device__ float type_convert<float, bf8_t>(bf8_t x)
template <> template <>
inline __host__ __device__ bf8_t type_convert<bf8_t, half_t>(half_t x) inline __host__ __device__ bf8_t type_convert<bf8_t, half_t>(half_t x)
{ {
#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__) #if defined CK_USE_SR_F8_CONVERSION
// convert to float and use native converion return f8_convert_sr<bf8_t>(x);
return type_convert<bf8_t>(type_convert<float>(x));
#else #else
constexpr bool negative_zero_nan = true; return f8_convert_rne<bf8_t>(x);
constexpr bool clip = true;
constexpr f8_rounding_mode rm = f8_rounding_mode::standard;
constexpr uint32_t rng = 0;
return utils::
cast_to_f8<half_t, bf8_t, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>(
x, rng);
#endif #endif
} }
...@@ -301,104 +479,4 @@ inline __host__ __device__ constexpr bhalf_t bf16_convert_rtn<bhalf_t, half_t>(h ...@@ -301,104 +479,4 @@ inline __host__ __device__ constexpr bhalf_t bf16_convert_rtn<bhalf_t, half_t>(h
return bf16_convert_rtn<bhalf_t>(x_fp32); return bf16_convert_rtn<bhalf_t>(x_fp32);
} }
// Declare a template function for fp8 conversion using SR
template <typename Y, typename X>
__host__ __device__ constexpr Y f8_convert_sr(X x);
// convert fp32 to fp8 with stochastic rounding
template <>
inline __host__ __device__ f8_t f8_convert_sr<f8_t, float>(float x)
{
constexpr int seed = 42;
uint32_t rng = prand_generator<float, seed>(reinterpret_cast<uintptr_t>(&x), x);
#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
union
{
float fval;
uint32_t i32val;
uint8_t i8val[4]; // not endian independent
} val;
val.fval = x;
uint32_t ival = 0;
ival = __builtin_amdgcn_cvt_sr_fp8_f32(val.fval, rng, ival, 0); // 0 pos
val.i32val = ival;
return val.i8val[0]; // little endian
#else
constexpr bool negative_zero_nan = true;
constexpr bool clip = true;
constexpr f8_rounding_mode rm = f8_rounding_mode::stochastic;
return utils::
cast_to_f8<float, f8_t, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>(x,
rng);
#endif
}
// convert fp16 to fp8 with stochastic rounding
template <>
inline __host__ __device__ f8_t f8_convert_sr<f8_t, half_t>(half_t x)
{
#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
// convert to float and use native converion
return f8_convert_sr<f8_t>(type_convert<float>(x));
#else
constexpr bool negative_zero_nan = true;
constexpr bool clip = true;
constexpr f8_rounding_mode rm = f8_rounding_mode::stochastic;
constexpr int seed = 42;
uint32_t rng = prand_generator<half_t, seed>(reinterpret_cast<uintptr_t>(&x), x);
return utils::
cast_to_f8<half_t, f8_t, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>(
x, rng);
#endif
}
// convert fp32 to bf8 with stochastic rounding
template <>
inline __host__ __device__ bf8_t f8_convert_sr<bf8_t, float>(float x)
{
constexpr int seed = 42;
uint32_t rng = prand_generator<float, seed>(reinterpret_cast<uintptr_t>(&x), x);
#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
union
{
float fval;
uint32_t i32val;
uint8_t i8val[4]; // not endian independent
} val;
val.fval = x;
uint32_t ival = 0;
ival = __builtin_amdgcn_cvt_sr_bf8_f32(val.fval, rng, ival, 0); // 0 pos
val.i32val = ival;
return val.i8val[0]; // little endian
#else
constexpr bool negative_zero_nan = true;
constexpr bool clip = true;
constexpr f8_rounding_mode rm = f8_rounding_mode::stochastic;
return utils::
cast_to_f8<float, bf8_t, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>(
x, rng);
#endif
}
// convert fp16 to bf8 with stochastic rounding
template <>
inline __host__ __device__ bf8_t f8_convert_sr<bf8_t, half_t>(half_t x)
{
#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
// convert to float and use native converion
return f8_convert_sr<f8_t>(type_convert<float>(x));
#else
constexpr bool negative_zero_nan = true;
constexpr bool clip = true;
constexpr f8_rounding_mode rm = f8_rounding_mode::stochastic;
constexpr int seed = 42;
// as thread id is not available on host, use 0 for prn generation
uint32_t rng = prand_generator<half_t, seed>(reinterpret_cast<uintptr_t>(&x), x);
return utils::
cast_to_f8<half_t, bf8_t, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>(
x, rng);
#endif
}
} // namespace ck } // namespace ck
library/include/ck/library/reference_tensor_operation/cpu/reference_contraction.hpp
View file @ 29deceb6
...@@ -23,6 +23,7 @@ template <ck::index_t NumDimM, ...@@ -23,6 +23,7 @@ template <ck::index_t NumDimM,
typename BDataType, typename BDataType,
typename CDataType, typename CDataType,
typename AccDataType, typename AccDataType,
typename ComputeDataType,
typename AElementwiseOperation, typename AElementwiseOperation,
typename BElementwiseOperation, typename BElementwiseOperation,
ck::enable_if_t<NumDimM == 2 && NumDimN == 2 && NumDimK == 2, bool> = false> ck::enable_if_t<NumDimM == 2 && NumDimN == 2 && NumDimK == 2, bool> = false>
...@@ -69,19 +70,24 @@ struct ReferenceContraction_M2_N2_K2 : public ck::tensor_operation::device::Base ...@@ -69,19 +70,24 @@ struct ReferenceContraction_M2_N2_K2 : public ck::tensor_operation::device::Base
{ {
for(ck::index_t k1 = 0; k1 < K1; ++k1) for(ck::index_t k1 = 0; k1 < K1; ++k1)
{ {
// Simulate the possible casting when ComputeDataType is different than the
// A/B data types
ComputeDataType v_a_compute_input =
ck::type_convert<ComputeDataType>(arg.a_ms_ks_(m0, m1, k0, k1));
ComputeDataType v_b_compute_input =
ck::type_convert<ComputeDataType>(arg.b_ns_ks_(n0, n1, k0, k1));
AccDataType v_a; AccDataType v_a;
AccDataType v_b; AccDataType v_b;
arg.a_element_op_( arg.a_element_op_(v_a, ck::type_convert<AccDataType>(v_a_compute_input));
v_a, ck::type_convert<const AccDataType>(arg.a_ms_ks_(m0, m1, k0, k1))); arg.b_element_op_(v_b, ck::type_convert<AccDataType>(v_b_compute_input));
arg.b_element_op_(
v_b, ck::type_convert<const AccDataType>(arg.b_ns_ks_(n0, n1, k0, k1)));
v_acc += v_a * v_b; v_acc += v_a * v_b;
} }
} }
arg.c_ms_ns_(m0, m1, n0, n1) = v_acc; arg.c_ms_ns_(m0, m1, n0, n1) = ck::type_convert<CDataType>(v_acc);
}; };
make_ParallelTensorFunctor(f_ms_ns, make_ParallelTensorFunctor(f_ms_ns,
......
library/include/ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp
View file @ 29deceb6
...@@ -3,12 +3,23 @@ ...@@ -3,12 +3,23 @@
#pragma once #pragma once
#include <iostream> #include <cmath>
#include <cstdlib>
#include <numeric>
#include <type_traits> #include <type_traits>
#include <sstream> #include <vector>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_base.hpp" #include "ck/tensor_operation/gpu/device/device_base.hpp"
#include "ck/library/utility/algorithm.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/fill.hpp"
#include "ck/library/utility/host_tensor.hpp" #include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/convolution_parameter.hpp"
#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
namespace ck { namespace ck {
namespace tensor_operation { namespace tensor_operation {
...@@ -22,6 +33,7 @@ namespace host { ...@@ -22,6 +33,7 @@ namespace host {
// Supports both GNCHW/NGCHW as well as GNHWC/NHWGC physical layout // Supports both GNCHW/NGCHW as well as GNHWC/NHWGC physical layout
// as long as dimensions in tensor descriptor is in GNCHW order // as long as dimensions in tensor descriptor is in GNCHW order
// //
// @tparam NDimSpatial Number of spatial dimensions.
// @tparam InDataType Input tensor data type. // @tparam InDataType Input tensor data type.
// @tparam WeiDataType Weights tensor data type. // @tparam WeiDataType Weights tensor data type.
// @tparam OutDataType Output tensor data type. // @tparam OutDataType Output tensor data type.
...@@ -29,7 +41,9 @@ namespace host { ...@@ -29,7 +41,9 @@ namespace host {
// operation. // operation.
// @tparam WeiElementwiseOperation Functor for weights tensor elementwise // @tparam WeiElementwiseOperation Functor for weights tensor elementwise
// operation. // operation.
// @tparam NDimSpatial Number of spatial dimensions. // @tparam NumAElementwiseTensor Number of A elementwise tensors.
// @tparam NumBElementwiseTensor Number of B elementwise tensors.
// @tparam NumDElementwiseTensor Number of D elementwise tensors.
// //
// input descriptor in [G, N, C, Do, Ho, Wo] order // input descriptor in [G, N, C, Do, Ho, Wo] order
// weight descriptor in [G, K, C, Z, Y, X] order // weight descriptor in [G, K, C, Z, Y, X] order
...@@ -42,25 +56,35 @@ template <ck::index_t NDimSpatial, ...@@ -42,25 +56,35 @@ template <ck::index_t NDimSpatial,
typename InElementwiseOperation, typename InElementwiseOperation,
typename WeiElementwiseOperation, typename WeiElementwiseOperation,
typename OutElementwiseOperation, typename OutElementwiseOperation,
ck::index_t NumAElementwiseTensor = 0,
ck::index_t NumBElementwiseTensor = 0,
ck::index_t NumDElementwiseTensor = 0,
typename std::enable_if<NDimSpatial >= 1 && NDimSpatial <= 3, bool>::type = false> typename std::enable_if<NDimSpatial >= 1 && NDimSpatial <= 3, bool>::type = false>
struct ReferenceConvFwd : public device::BaseOperator struct ReferenceConvFwd : public device::BaseOperator
{ {
// Argument // Argument
struct Argument : public device::BaseArgument struct Argument : public device::BaseArgument
{ {
Argument(const Tensor<InDataType>& input, Argument(
const Tensor<WeiDataType>& weight, const Tensor<InDataType>& input,
Tensor<OutDataType>& output, const Tensor<WeiDataType>& weight,
std::vector<ck::index_t> conv_filter_strides, Tensor<OutDataType>& output,
std::vector<ck::index_t> conv_filter_dilations, std::vector<ck::index_t> conv_filter_strides,
std::vector<ck::index_t> input_left_pads, std::vector<ck::index_t> conv_filter_dilations,
std::vector<ck::index_t> input_right_pads, std::vector<ck::index_t> input_left_pads,
InElementwiseOperation in_element_op, std::vector<ck::index_t> input_right_pads,
WeiElementwiseOperation wei_element_op, InElementwiseOperation in_element_op,
OutElementwiseOperation out_element_op) WeiElementwiseOperation wei_element_op,
OutElementwiseOperation out_element_op,
const std::array<Tensor<InDataType>, NumAElementwiseTensor>& elementwise_a_tensors,
const std::array<Tensor<WeiDataType>, NumBElementwiseTensor>& elementwise_b_tensors,
const std::array<Tensor<OutDataType>, NumDElementwiseTensor>& elementwise_d_tensors)
: input_{input}, : input_{input},
weight_{weight}, weight_{weight},
output_{output}, output_{output},
elementwise_a_tensors_{elementwise_a_tensors},
elementwise_b_tensors_{elementwise_b_tensors},
elementwise_d_tensors_{elementwise_d_tensors},
conv_strides_{conv_filter_strides}, conv_strides_{conv_filter_strides},
conv_dilations_{conv_filter_dilations}, conv_dilations_{conv_filter_dilations},
in_left_pads_{input_left_pads}, in_left_pads_{input_left_pads},
...@@ -75,6 +99,10 @@ struct ReferenceConvFwd : public device::BaseOperator ...@@ -75,6 +99,10 @@ struct ReferenceConvFwd : public device::BaseOperator
const Tensor<WeiDataType>& weight_; const Tensor<WeiDataType>& weight_;
Tensor<OutDataType>& output_; Tensor<OutDataType>& output_;
const std::array<Tensor<InDataType>, NumAElementwiseTensor>& elementwise_a_tensors_;
const std::array<Tensor<WeiDataType>, NumBElementwiseTensor>& elementwise_b_tensors_;
const std::array<Tensor<OutDataType>, NumDElementwiseTensor>& elementwise_d_tensors_;
std::vector<index_t> conv_strides_; std::vector<index_t> conv_strides_;
std::vector<index_t> conv_dilations_; std::vector<index_t> conv_dilations_;
std::vector<index_t> in_left_pads_; std::vector<index_t> in_left_pads_;
...@@ -114,23 +142,43 @@ struct ReferenceConvFwd : public device::BaseOperator ...@@ -114,23 +142,43 @@ struct ReferenceConvFwd : public device::BaseOperator
if(wi >= 0 && if(wi >= 0 &&
ck::type_convert<std::size_t>(wi) < arg.input_.GetLengths()[3]) ck::type_convert<std::size_t>(wi) < arg.input_.GetLengths()[3])
{ {
float v_in; InDataType v_in;
float v_wei; WeiDataType v_wei;
arg.in_element_op_( ExecuteElementwiseOp(arg.in_element_op_,
v_in, ck::type_convert<float>(arg.input_(g, n, c, wi))); arg.elementwise_a_tensors_,
Number<NumAElementwiseTensor>{},
arg.wei_element_op_( v_in,
v_wei, ck::type_convert<float>(arg.weight_(g, k, c, x))); arg.input_(g, n, c, wi),
g,
v_acc += v_in * v_wei; n,
c,
wi);
ExecuteElementwiseOp(arg.wei_element_op_,
arg.elementwise_b_tensors_,
Number<NumBElementwiseTensor>{},
v_wei,
arg.weight_(g, k, c, x),
g,
k,
c,
x);
v_acc +=
ck::type_convert<float>(v_in) * ck::type_convert<float>(v_wei);
} }
} }
} }
OutDataType v_acc_converted = ck::type_convert<OutDataType>(v_acc);
OutDataType v_out; OutDataType& v_out = arg.output_(g, n, k, wo);
arg.out_element_op_(v_out, ck::type_convert<OutDataType>(v_acc)); ExecuteElementwiseOp(arg.out_element_op_,
arg.output_(g, n, k, wo) = v_out; arg.elementwise_d_tensors_,
Number<NumDElementwiseTensor>{},
v_out,
v_acc_converted,
g,
n,
k,
wo);
}; };
make_ParallelTensorFunctor(func, make_ParallelTensorFunctor(func,
...@@ -167,24 +215,47 @@ struct ReferenceConvFwd : public device::BaseOperator ...@@ -167,24 +215,47 @@ struct ReferenceConvFwd : public device::BaseOperator
wi >= 0 && wi >= 0 &&
ck::type_convert<std::size_t>(wi) < arg.input_.GetLengths()[4]) ck::type_convert<std::size_t>(wi) < arg.input_.GetLengths()[4])
{ {
float v_in; InDataType v_in;
float v_wei; WeiDataType v_wei;
arg.in_element_op_( ExecuteElementwiseOp(arg.in_element_op_,
v_in, ck::type_convert<float>(arg.input_(g, n, c, hi, wi))); arg.elementwise_a_tensors_,
Number<NumAElementwiseTensor>{},
arg.wei_element_op_( v_in,
v_wei, ck::type_convert<float>(arg.weight_(g, k, c, y, x))); arg.input_(g, n, c, hi, wi),
g,
v_acc += v_in * v_wei; n,
c,
hi,
wi);
ExecuteElementwiseOp(arg.wei_element_op_,
arg.elementwise_b_tensors_,
Number<NumBElementwiseTensor>{},
v_wei,
arg.weight_(g, k, c, y, x),
g,
k,
c,
y,
x);
v_acc += ck::type_convert<float>(v_in) *
ck::type_convert<float>(v_wei);
} }
} }
} }
} }
OutDataType v_acc_converted = ck::type_convert<OutDataType>(v_acc);
OutDataType v_out; OutDataType& v_out = arg.output_(g, n, k, ho, wo);
arg.out_element_op_(v_out, ck::type_convert<OutDataType>(v_acc)); ExecuteElementwiseOp(arg.out_element_op_,
arg.output_(g, n, k, ho, wo) = v_out; arg.elementwise_d_tensors_,
Number<NumDElementwiseTensor>{},
v_out,
v_acc_converted,
g,
n,
k,
ho,
wo);
}; };
make_ParallelTensorFunctor(func, make_ParallelTensorFunctor(func,
...@@ -231,27 +302,51 @@ struct ReferenceConvFwd : public device::BaseOperator ...@@ -231,27 +302,51 @@ struct ReferenceConvFwd : public device::BaseOperator
ck::type_convert<std::size_t>(wi) < ck::type_convert<std::size_t>(wi) <
arg.input_.GetLengths()[5]) arg.input_.GetLengths()[5])
{ {
float v_in; InDataType v_in;
float v_wei; WeiDataType v_wei;
arg.in_element_op_(v_in, ExecuteElementwiseOp(arg.in_element_op_,
ck::type_convert<float>( arg.elementwise_a_tensors_,
arg.input_(g, n, c, di, hi, wi))); Number<NumAElementwiseTensor>{},
v_in,
arg.wei_element_op_( arg.input_(g, n, c, di, hi, wi),
v_wei, g,
ck::type_convert<float>(arg.weight_(g, k, c, z, y, x))); n,
c,
v_acc += v_in * v_wei; di,
hi,
wi);
ExecuteElementwiseOp(arg.wei_element_op_,
arg.elementwise_b_tensors_,
Number<NumBElementwiseTensor>{},
v_wei,
arg.weight_(g, k, c, z, y, x),
g,
k,
c,
z,
y,
x);
v_acc += ck::type_convert<float>(v_in) *
ck::type_convert<float>(v_wei);
} }
} }
} }
} }
} }
OutDataType v_acc_converted = ck::type_convert<OutDataType>(v_acc);
OutDataType v_out; OutDataType& v_out = arg.output_(g, n, k, d_o, ho, wo);
arg.out_element_op_(v_out, ck::type_convert<OutDataType>(v_acc)); ExecuteElementwiseOp(arg.out_element_op_,
arg.output_(g, n, k, d_o, ho, wo) = v_out; arg.elementwise_d_tensors_,
Number<NumDElementwiseTensor>{},
v_out,
v_acc_converted,
g,
n,
k,
d_o,
ho,
wo);
}; };
make_ParallelTensorFunctor(func, make_ParallelTensorFunctor(func,
...@@ -274,6 +369,36 @@ struct ReferenceConvFwd : public device::BaseOperator ...@@ -274,6 +369,36 @@ struct ReferenceConvFwd : public device::BaseOperator
} }
}; };
template <typename... Args,
typename ElementwiseOp,
typename ElementwiseTensor,
typename NumTensor,
typename T>
static void ExecuteElementwiseOp(ElementwiseOp& elementwise_op,
ElementwiseTensor& elementwise_tensors,
NumTensor,
T& y,
const T& x,
Args... dims)
{
if constexpr(NumTensor::value == 0)
{
elementwise_op(y, x);
}
else if constexpr(NumTensor::value == 1)
{
elementwise_op(y, x, elementwise_tensors[0](dims...));
}
else if constexpr(NumTensor::value == 2)
{
elementwise_op(y, x, elementwise_tensors[0](dims...), elementwise_tensors[1](dims...));
}
else
{
throw std::runtime_error("ElementOp not supported in reference.");
}
}
static constexpr bool IsValidCompilationParameter() static constexpr bool IsValidCompilationParameter()
{ {
// TODO: properly implement this check // TODO: properly implement this check
...@@ -285,16 +410,20 @@ struct ReferenceConvFwd : public device::BaseOperator ...@@ -285,16 +410,20 @@ struct ReferenceConvFwd : public device::BaseOperator
return NDimSpatial >= 1 && NDimSpatial <= 3; return NDimSpatial >= 1 && NDimSpatial <= 3;
} }
static auto MakeArgument(const Tensor<InDataType>& input, static auto MakeArgument(
const Tensor<WeiDataType>& weight, const Tensor<InDataType>& input,
Tensor<OutDataType>& output, const Tensor<WeiDataType>& weight,
std::vector<ck::index_t> conv_filter_strides, Tensor<OutDataType>& output,
std::vector<ck::index_t> conv_filter_dilations, std::vector<ck::index_t> conv_filter_strides,
std::vector<ck::index_t> input_left_pads, std::vector<ck::index_t> conv_filter_dilations,
std::vector<ck::index_t> input_right_pads, std::vector<ck::index_t> input_left_pads,
InElementwiseOperation in_element_op, std::vector<ck::index_t> input_right_pads,
WeiElementwiseOperation wei_element_op, InElementwiseOperation in_element_op,
OutElementwiseOperation out_element_op) WeiElementwiseOperation wei_element_op,
OutElementwiseOperation out_element_op,
const std::array<Tensor<InDataType>, NumAElementwiseTensor>& elementwise_a_tensors = {},
const std::array<Tensor<WeiDataType>, NumBElementwiseTensor>& elementwise_b_tensors = {},
const std::array<Tensor<OutDataType>, NumDElementwiseTensor>& elementwise_d_tensors = {})
{ {
return Argument{input, return Argument{input,
weight, weight,
...@@ -305,7 +434,10 @@ struct ReferenceConvFwd : public device::BaseOperator ...@@ -305,7 +434,10 @@ struct ReferenceConvFwd : public device::BaseOperator
input_right_pads, input_right_pads,
in_element_op, in_element_op,
wei_element_op, wei_element_op,
out_element_op}; out_element_op,
elementwise_a_tensors,
elementwise_b_tensors,
elementwise_d_tensors};
} }
static auto MakeInvoker() { return Invoker{}; } static auto MakeInvoker() { return Invoker{}; }
......
library/include/ck/library/reference_tensor_operation/cpu/reference_groupnorm_bwd.hpp 0 → 100644
View file @ 29deceb6
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iostream>
#include <sstream>
#include <vector>
#include <algorithm>
#include "ck/tensor_operation/gpu/device/device_base.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
namespace ck {
namespace tensor_operation {
namespace host {
template <typename DYDataType,
typename XDataType,
typename GammaDataType,
typename MeanInvStdDataType,
typename DGammaDataType,
typename DBetaDataType,
typename DXDataType,
typename ComputeDataType>
struct ReferenceGroupnormBwd : public device::BaseOperator
{
// Argument
struct Argument : public device::BaseArgument
{
Argument(const Tensor<DYDataType>& dy_nhwgc,
const Tensor<XDataType>& x_nhwgc,
const Tensor<GammaDataType>& gamma_gc,
const Tensor<MeanInvStdDataType>& mean_ng,
const Tensor<MeanInvStdDataType>& inv_std_ng,
Tensor<DGammaDataType>& dgamma_gc,
Tensor<DBetaDataType>& dbeta_gc,
Tensor<DXDataType>& dx_nhwgc,
const std::vector<index_t> lengths)
: dy_nhwgc_(dy_nhwgc),
x_nhwgc_(x_nhwgc),
gamma_gc_(gamma_gc),
mean_ng_(mean_ng),
inv_std_ng_(inv_std_ng),
dgamma_gc_(dgamma_gc),
dbeta_gc_(dbeta_gc),
dx_nhwgc_(dx_nhwgc),
lengths_(lengths)
{
}
const Tensor<DYDataType>& dy_nhwgc_;
const Tensor<XDataType>& x_nhwgc_;
const Tensor<GammaDataType>& gamma_gc_;
const Tensor<MeanInvStdDataType>& mean_ng_;
const Tensor<MeanInvStdDataType>& inv_std_ng_;
Tensor<DGammaDataType>& dgamma_gc_;
Tensor<DBetaDataType>& dbeta_gc_;
Tensor<DXDataType>& dx_nhwgc_;
std::vector<index_t> lengths_;
};
// Invoker
struct Invoker : public device::BaseInvoker
{
float Run(const Argument& arg)
{
int N = arg.lengths_[0];
int H = arg.lengths_[1];
int W = arg.lengths_[2];
int G = arg.lengths_[3];
int C = arg.lengths_[4];
// Calculate dgamma and dbeta
for(int g = 0; g < G; ++g)
for(int c = 0; c < C; ++c)
{
ComputeDataType dgamma = 0;
ComputeDataType dbeta = 0;
for(int n = 0; n < N; ++n)
for(int h = 0; h < H; ++h)
for(int w = 0; w < W; ++w)
{
ComputeDataType dy =
ck::type_convert<ComputeDataType>(arg.dy_nhwgc_(n, h, w, g, c));
ComputeDataType x =
ck::type_convert<ComputeDataType>(arg.x_nhwgc_(n, h, w, g, c));
ComputeDataType mean =
ck::type_convert<ComputeDataType>(arg.mean_ng_(n, g));
ComputeDataType rstd =
ck::type_convert<ComputeDataType>(arg.inv_std_ng_(n, g));
dgamma += dy * rstd * (x - mean);
dbeta += dy;
}
arg.dgamma_gc_(g, c) = ck::type_convert<DGammaDataType>(dgamma);
arg.dbeta_gc_(g, c) = ck::type_convert<DBetaDataType>(dbeta);
}
// Calculate dx
int reduce_size = H * W * C;
for(int n = 0; n < N; ++n)
for(int g = 0; g < G; ++g)
{
ComputeDataType ds = 0;
ComputeDataType db = 0;
ComputeDataType mean = ck::type_convert<ComputeDataType>(arg.mean_ng_(n, g));
ComputeDataType rstd = ck::type_convert<ComputeDataType>(arg.inv_std_ng_(n, g));
for(int h = 0; h < H; ++h)
for(int w = 0; w < W; ++w)
for(int c = 0; c < C; ++c)
{
ComputeDataType dy =
ck::type_convert<ComputeDataType>(arg.dy_nhwgc_(n, h, w, g, c));
ComputeDataType x =
ck::type_convert<ComputeDataType>(arg.x_nhwgc_(n, h, w, g, c));
ComputeDataType gamma =
ck::type_convert<ComputeDataType>(arg.gamma_gc_(g, c));
ds += dy * gamma * x;
db += dy * gamma;
}
for(int h = 0; h < H; ++h)
for(int w = 0; w < W; ++w)
for(int c = 0; c < C; ++c)
{
ComputeDataType dy =
ck::type_convert<ComputeDataType>(arg.dy_nhwgc_(n, h, w, g, c));
ComputeDataType x =
ck::type_convert<ComputeDataType>(arg.x_nhwgc_(n, h, w, g, c));
ComputeDataType gamma =
ck::type_convert<ComputeDataType>(arg.gamma_gc_(g, c));
ComputeDataType b =
(db * mean - ds) * rstd * rstd * rstd / reduce_size;
ComputeDataType c1 = -b * mean - db * rstd / reduce_size;
arg.dx_nhwgc_(n, h, w, g, c) =
ck::type_convert<DXDataType>(dy * gamma * rstd + b * x + c1);
}
}
return 0;
}
float Run(const device::BaseArgument* p_arg,
const StreamConfig& /* stream_config */ = StreamConfig{}) override
{
return Run(*dynamic_cast<const Argument*>(p_arg));
}
};
static constexpr bool IsValidCompilationParameter()
{
// TODO: properly implement this check
return true;
}
bool IsSupportedArgument(const device::BaseArgument*) override { return true; }
static auto MakeArgument(const Tensor<DYDataType>& dy_nhwgc,
const Tensor<XDataType>& x_nhwgc,
const Tensor<GammaDataType>& gamma_gc,
const Tensor<MeanInvStdDataType>& mean_ng,
const Tensor<MeanInvStdDataType>& inv_std_ng,
Tensor<DGammaDataType>& dgamma_gc,
Tensor<DBetaDataType>& dbeta_gc,
Tensor<DXDataType>& dx_nhwgc,
const std::vector<index_t> lengths)
{
return Argument{dy_nhwgc,
x_nhwgc,
gamma_gc,
mean_ng,
inv_std_ng,
dgamma_gc,
dbeta_gc,
dx_nhwgc,
lengths};
}
static auto MakeInvoker() { return Invoker{}; }
virtual std::unique_ptr<device::BaseInvoker> MakeInvokerPointer()
{
return std::make_unique<Invoker>(Invoker{});
}
std::string GetTypeString() const override
{
auto str = std::stringstream();
// clang-format off
str << "ReferenceGroupnormBwd"
<< std::endl;
// clang-format on
return str.str();
}
};
} // namespace host
} // namespace tensor_operation
} // namespace ck
library/include/ck/library/reference_tensor_operation/cpu/reference_layernorm.hpp
View file @ 29deceb6
...@@ -28,7 +28,8 @@ template <typename XDataType, ...@@ -28,7 +28,8 @@ template <typename XDataType,
struct ReferenceLayernorm : public device::BaseOperator struct ReferenceLayernorm : public device::BaseOperator
{ {
// TODO - support generic layernorm // TODO - support generic layernorm
static_assert((Rank == 2 && NumReduceDim == 1), "Only support 2D version so far"); static_assert((Rank == 2 && NumReduceDim == 1) || (Rank == 4 && NumReduceDim == 3),
"Only support 2D & 4D version so far");
// Argument // Argument
struct Argument : public device::BaseArgument struct Argument : public device::BaseArgument
...@@ -71,7 +72,7 @@ struct ReferenceLayernorm : public device::BaseOperator ...@@ -71,7 +72,7 @@ struct ReferenceLayernorm : public device::BaseOperator
// Invoker // Invoker
struct Invoker : public device::BaseInvoker struct Invoker : public device::BaseInvoker
{ {
float Run(const Argument& arg) float Run2D(const Argument& arg)
{ {
int M = arg.lengths_[0]; int M = arg.lengths_[0];
int N = arg.lengths_[1]; int N = arg.lengths_[1];
...@@ -117,6 +118,71 @@ struct ReferenceLayernorm : public device::BaseOperator ...@@ -117,6 +118,71 @@ struct ReferenceLayernorm : public device::BaseOperator
return 0; return 0;
} }
float Run4D(const Argument& arg)
{
int N = arg.lengths_[0];
int H = arg.lengths_[1];
int W = arg.lengths_[2];
int C = arg.lengths_[3];
Tensor<ComputeDataType> mean({N});
Tensor<ComputeDataType> var({N});
int reduce_length = H * W * C;
for(int n = 0; n < N; ++n)
{
mean(n) = 0;
var(n) = 0;
for(int h = 0; h < H; ++h)
for(int w = 0; w < W; ++w)
for(int c = 0; c < C; ++c)
{
auto x_val = ck::type_convert<ComputeDataType>(arg.x_m_n_(n, h, w, c));
mean(n) += x_val;
var(n) += x_val * x_val;
}
mean(n) = mean(n) / reduce_length;
var(n) = (var(n) / reduce_length) - (mean(n) * mean(n));
}
for(int n = 0; n < N; ++n)
{
ComputeDataType divisor =
static_cast<ComputeDataType>(1) / ck::math::sqrt(var(n) + arg.epsilon_);
for(int h = 0; h < H; ++h)
for(int w = 0; w < W; ++w)
for(int c = 0; c < C; ++c)
{
auto x_val = ck::type_convert<ComputeDataType>(arg.x_m_n_(n, h, w, c));
auto gamma_val =
ck::type_convert<ComputeDataType>(arg.gamma_n_(h, w, c));
auto beta_val = ck::type_convert<ComputeDataType>(arg.beta_n_(h, w, c));
auto y_val = (x_val - mean(n)) * divisor;
y_val = (y_val * gamma_val) + beta_val;
arg.y_elementwise_op_(y_val, y_val);
arg.y_m_n_(n, h, w, c) = ck::type_convert<YDataType>(y_val);
}
arg.save_mean_m_(n) = ck::type_convert<SaveMeanInvStdDataType>(mean(n));
arg.save_inv_std_m_(n) = ck::type_convert<SaveMeanInvStdDataType>(divisor);
}
return 0;
}
float Run(const Argument& arg)
{
if(arg.lengths_.size() == 2)
return Run2D(arg);
else if(arg.lengths_.size() == 4)
return Run4D(arg);
return 0;
}
float Run(const device::BaseArgument* p_arg, float Run(const device::BaseArgument* p_arg,
const StreamConfig& /* stream_config */ = StreamConfig{}) override const StreamConfig& /* stream_config */ = StreamConfig{}) override
{ {
...@@ -134,17 +200,16 @@ struct ReferenceLayernorm : public device::BaseOperator ...@@ -134,17 +200,16 @@ struct ReferenceLayernorm : public device::BaseOperator
{ {
const Argument* p_arg_ = dynamic_cast<const Argument*>(p_arg); const Argument* p_arg_ = dynamic_cast<const Argument*>(p_arg);
// TODO - support generic layernorm if(p_arg_->lengths_.size() == 2 && p_arg_->reduceDims_.size() == 1 &&
if(p_arg_->lengths_.size() != 2) p_arg_->reduceDims_[0] == 1)
return false; return true;
if(p_arg_->reduceDims_.size() != 1)
return false;
if(p_arg_->reduceDims_[0] != 1) else if(p_arg_->lengths_.size() == 4 && p_arg_->reduceDims_.size() == 3 &&
return false; p_arg_->reduceDims_[0] == 1 && p_arg_->reduceDims_[1] == 2 &&
p_arg_->reduceDims_[2] == 3)
return true;
return true; return false;
} }
static auto MakeArgument(const Tensor<XDataType>& x_m_n, static auto MakeArgument(const Tensor<XDataType>& x_m_n,
......
library/include/ck/library/reference_tensor_operation/cpu/reference_layernorm_bwd.hpp 0 → 100644
View file @ 29deceb6
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iostream>
#include <sstream>
#include <vector>
#include <algorithm>
#include "ck/tensor_operation/gpu/device/device_base.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
namespace ck {
namespace tensor_operation {
namespace host {
template <typename DYDataType,
typename XDataType,
typename GammaDataType,
typename MeanInvStdDataType,
typename DGammaDataType,
typename DBetaDataType,
typename DXDataType,
typename ComputeDataType>
struct ReferenceLayernormBwd : public device::BaseOperator
{
// Argument
struct Argument : public device::BaseArgument
{
Argument(const Tensor<DYDataType>& dy_m_n,
const Tensor<XDataType>& x_m_n,
const Tensor<GammaDataType>& gamma_n,
const Tensor<MeanInvStdDataType>& mean_m,
const Tensor<MeanInvStdDataType>& inv_std_m,
Tensor<DGammaDataType>& dgamma_n,
Tensor<DBetaDataType>& dbeta_n,
Tensor<DXDataType>& dx_m_n,
const std::vector<index_t> lengths)
: dy_m_n_(dy_m_n),
x_m_n_(x_m_n),
gamma_n_(gamma_n),
mean_m_(mean_m),
inv_std_m_(inv_std_m),
dgamma_n_(dgamma_n),
dbeta_n_(dbeta_n),
dx_m_n_(dx_m_n),
lengths_(lengths)
{
}
const Tensor<DYDataType>& dy_m_n_;
const Tensor<XDataType>& x_m_n_;
const Tensor<GammaDataType>& gamma_n_;
const Tensor<MeanInvStdDataType>& mean_m_;
const Tensor<MeanInvStdDataType>& inv_std_m_;
Tensor<DGammaDataType>& dgamma_n_;
Tensor<DBetaDataType>& dbeta_n_;
Tensor<DXDataType>& dx_m_n_;
std::vector<index_t> lengths_;
};
// Invoker
struct Invoker : public device::BaseInvoker
{
float Run(const Argument& arg)
{
int M = arg.lengths_[0];
int N = arg.lengths_[1];
// Calculate dgamma and dbeta
for(int n = 0; n < N; ++n)
{
ComputeDataType dgamma = 0;
ComputeDataType dbeta = 0;
for(int m = 0; m < M; ++m)
{
ComputeDataType dy = ck::type_convert<ComputeDataType>(arg.dy_m_n_(m, n));
ComputeDataType x = ck::type_convert<ComputeDataType>(arg.x_m_n_(m, n));
ComputeDataType mean = ck::type_convert<ComputeDataType>(arg.mean_m_(m));
ComputeDataType rstd = ck::type_convert<ComputeDataType>(arg.inv_std_m_(m));
dgamma += dy * rstd * (x - mean);
dbeta += dy;
}
arg.dgamma_n_(n) = ck::type_convert<DGammaDataType>(dgamma);
arg.dbeta_n_(n) = ck::type_convert<DBetaDataType>(dbeta);
}
// Calculate dx
for(int m = 0; m < M; ++m)
{
ComputeDataType ds = 0;
ComputeDataType db = 0;
ComputeDataType mean = ck::type_convert<ComputeDataType>(arg.mean_m_(m));
ComputeDataType rstd = ck::type_convert<ComputeDataType>(arg.inv_std_m_(m));
for(int n = 0; n < N; ++n)
{
ComputeDataType dy = ck::type_convert<ComputeDataType>(arg.dy_m_n_(m, n));
ComputeDataType x = ck::type_convert<ComputeDataType>(arg.x_m_n_(m, n));
ComputeDataType gamma = ck::type_convert<ComputeDataType>(arg.gamma_n_(n));
ds += dy * gamma * x;
db += dy * gamma;
}
for(int n = 0; n < N; ++n)
{
ComputeDataType dy = ck::type_convert<ComputeDataType>(arg.dy_m_n_(m, n));
ComputeDataType x = ck::type_convert<ComputeDataType>(arg.x_m_n_(m, n));
ComputeDataType gamma = ck::type_convert<ComputeDataType>(arg.gamma_n_(n));
ComputeDataType b = (db * mean - ds) * rstd * rstd * rstd / N;
ComputeDataType c = -b * mean - db * rstd / N;
arg.dx_m_n_(m, n) = ck::type_convert<DXDataType>(dy * gamma * rstd + b * x + c);
}
}
return 0;
}
float Run(const device::BaseArgument* p_arg,
const StreamConfig& /* stream_config */ = StreamConfig{}) override
{
return Run(*dynamic_cast<const Argument*>(p_arg));
}
};
static constexpr bool IsValidCompilationParameter()
{
// TODO: properly implement this check
return true;
}
bool IsSupportedArgument(const device::BaseArgument*) override { return true; }
static auto MakeArgument(const Tensor<DYDataType>& dy_m_n,
const Tensor<XDataType>& x_m_n,
const Tensor<GammaDataType>& gamma_n,
const Tensor<MeanInvStdDataType>& mean_m,
const Tensor<MeanInvStdDataType>& inv_std_m,
Tensor<DGammaDataType>& dgamma_n,
Tensor<DBetaDataType>& dbeta_n,
Tensor<DXDataType>& dx_m_n,
const std::vector<index_t> lengths)
{
return Argument{
dy_m_n, x_m_n, gamma_n, mean_m, inv_std_m, dgamma_n, dbeta_n, dx_m_n, lengths};
}
static auto MakeInvoker() { return Invoker{}; }
virtual std::unique_ptr<device::BaseInvoker> MakeInvokerPointer()
{
return std::make_unique<Invoker>(Invoker{});
}
std::string GetTypeString() const override
{
auto str = std::stringstream();
// clang-format off
str << "ReferenceLayernormBwd"
<< std::endl;
// clang-format on
return str.str();
}
};
} // namespace host
} // namespace tensor_operation
} // namespace ck
library/include/ck/library/tensor_operation_instance/device_operation_instance_factory.hpp
View file @ 29deceb6
...@@ -25,6 +25,8 @@ using BF8 = ck::bf8_t; ...@@ -25,6 +25,8 @@ using BF8 = ck::bf8_t;
using Empty_Tuple = ck::Tuple<>; using Empty_Tuple = ck::Tuple<>;
using BF16_Tuple = ck::Tuple<BF16>;
using F16_Tuple = ck::Tuple<F16>; using F16_Tuple = ck::Tuple<F16>;
using F16_F16_Tuple = ck::Tuple<F16, F16>; using F16_F16_Tuple = ck::Tuple<F16, F16>;
......
library/include/ck/library/tensor_operation_instance/gpu/contraction/device_contraction_instance.hpp 0 → 100644
View file @ 29deceb6
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_contraction_multiple_d_xdl_cshuffle.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using F16 = ck::half_t;
using BF16 = ck::bhalf_t;
using F32 = float;
using F64 = double;
using F16_Tuple = ck::Tuple<F16>;
using BF16_Tuple = ck::Tuple<BF16>;
using F32_Tuple = ck::Tuple<F32>;
using F64_Tuple = ck::Tuple<F64>;
using Empty_Tuple = ck::Tuple<>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using Bilinear = ck::tensor_operation::element_wise::Bilinear;
using Scale = ck::tensor_operation::element_wise::Scale;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
template <typename ADataType,
typename BDataType,
typename AccDataType,
typename CShuffleDataType,
typename DsDataType,
typename EDataType,
typename ComputeDataType,
typename AElementwiseOp,
typename BElementwiseOp,
typename CDEElementwiseOp>
using device_contraction_kk_instance = std::tuple<
// clang-format off
//#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Compute|
//#####################################| | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| Data|
//#####################################| | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| Type|
//#####################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 256, 128, 16, 4, 4, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 256, 16, 4, 4, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 128, 128, 16, 4, 4, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 128, 16, 4, 4, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 128, 64, 16, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 8>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 64, 128, 16, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 64, 64, 64, 16, 4, 4, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 64, 16, 4, 4, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 64, 128, 16, 4, 4, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 128, 32, 16, 4, 4, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 8>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 32, 128, 16, 4, 4, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 64, 64, 32, 16, 4, 4, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 64, 32, 64, 16, 4, 4, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4, ComputeDataType>
// clang-format on
>;
template <typename ADataType,
typename BDataType,
typename AccDataType,
typename CShuffleDataType,
typename DsDataType,
typename EDataType,
typename ComputeDataType,
typename AElementwiseOp,
typename BElementwiseOp,
typename CDEElementwiseOp>
using device_contraction_kn_instance = std::tuple<
// clang-format off
//#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Compute|
//#####################################| | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| Data|
//#####################################| | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| Type|
//#####################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 256, 128, 16, 4, 1, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 256, 128, 16, 4, 4, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 1, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 256, 16, 4, 1, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 256, 16, 4, 4, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 128, 128, 16, 4, 1, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, 1, 1, S<1, 8, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 128, 128, 16, 4, 4, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 128, 16, 4, 1, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 128, 16, 4, 4, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 1, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 128, 64, 16, 4, 1, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, 1, 1, S<1, 16, 1, 8>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 128, 64, 16, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 1, 1, 1, S<1, 16, 1, 8>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 64, 128, 16, 4, 1, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, 1, 1, S<1, 8, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 64, 128, 16, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 64, 16, 4, 1, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 64, 16, 4, 4, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 1, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 64, 128, 16, 4, 1, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 64, 128, 16, 4, 4, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 1, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>
// clang-format on
>;
template <typename ADataType,
typename BDataType,
typename AccDataType,
typename CShuffleDataType,
typename DsDataType,
typename EDataType,
typename ComputeDataType,
typename AElementwiseOp,
typename BElementwiseOp,
typename CDEElementwiseOp>
using device_contraction_mk_instance = std::tuple<
// clang-format off
//#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Compute|
//#####################################| | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| Data|
//#####################################| | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| Type|
//#####################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 256, 128, 16, 1, 4, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 256, 128, 16, 4, 4, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 256, 16, 1, 4, 32, 32, 2, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 256, 16, 4, 4, 32, 32, 2, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 128, 128, 16, 1, 4, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 128, 128, 16, 4, 4, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 128, 16, 1, 4, 32, 32, 2, 2, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 128, 16, 4, 4, 32, 32, 2, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 128, 64, 16, 1, 4, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 8>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 128, 64, 16, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 8>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 64, 128, 16, 1, 4, 32, 32, 2, 2, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 64, 128, 16, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 64, 16, 1, 4, 32, 32, 2, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 64, 16, 4, 4, 32, 32, 2, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 64, 128, 16, 1, 4, 32, 32, 1, 2, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 64, 128, 16, 4, 4, 32, 32, 1, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>
// clang-format on
>;
template <typename ADataType,
typename BDataType,
typename AccDataType,
typename CShuffleDataType,
typename DsDataType,
typename EDataType,
typename ComputeDataType,
typename AElementwiseOp,
typename BElementwiseOp,
typename CDEElementwiseOp>
using device_contraction_mn_instance = std::tuple<
// clang-format off
//#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Compute|
//#####################################| | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| Data|
//#####################################| | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| Type|
//#####################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 256, 128, 16, 1, 1, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 256, 128, 16, 4, 4, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 1, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 256, 16, 1, 1, 32, 32, 2, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 256, 16, 4, 4, 32, 32, 2, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 128, 128, 16, 1, 1, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, 1, 1, S<1, 8, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 128, 128, 16, 4, 4, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 128, 16, 1, 1, 32, 32, 2, 2, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 128, 16, 4, 4, 32, 32, 2, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 1, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 128, 64, 16, 1, 1, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, 1, 1, S<1, 16, 1, 8>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 128, 64, 16, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 1, 1, 1, S<1, 16, 1, 8>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 64, 128, 16, 1, 1, 32, 32, 2, 2, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, 1, 1, S<1, 8, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 64, 128, 16, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 64, 16, 1, 1, 32, 32, 2, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 64, 16, 4, 4, 32, 32, 2, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 1, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 64, 128, 16, 1, 1, 32, 32, 1, 2, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 64, 128, 16, 4, 4, 32, 32, 1, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 1, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>
// clang-format on
>;
template <typename ADataType,
typename BDataType,
typename AccDataType,
typename CShuffleDataType,
typename DsDataType,
typename EDataType,
typename ComputeDataType,
typename AElementwiseOp,
typename BElementwiseOp,
typename CDEElementwiseOp>
using device_contraction_f64_kk_instance = std::tuple<
// clang-format off
//#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Compute|
//#####################################| | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| Data|
//#####################################| | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| Type|
//#####################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 128, 16, 2, 2, 16, 16, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 16, 1, 16>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 128, 64, 16, 2, 2, 16, 16, 4, 4, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 16, 1, 8>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 64, 128, 16, 2, 2, 16, 16, 4, 4, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 8, 1, 16>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 64, 64, 64, 16, 2, 2, 16, 16, 4, 4, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 8, 1, 8>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 64, 16, 2, 2, 16, 16, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 16, 1, 16>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 64, 128, 16, 2, 2, 16, 16, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 16, 1, 16>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 128, 32, 16, 2, 2, 16, 16, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 16, 1, 8>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 32, 128, 16, 2, 2, 16, 16, 2, 4, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 8, 1, 16>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 64, 64, 32, 16, 2, 2, 16, 16, 4, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 8, 1, 8>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 64, 32, 64, 16, 2, 2, 16, 16, 2, 4, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 8, 1, 8>, 1, ComputeDataType>
// clang-format on
>;
template <typename ADataType,
typename BDataType,
typename AccDataType,
typename CShuffleDataType,
typename DsDataType,
typename EDataType,
typename ComputeDataType,
typename AElementwiseOp,
typename BElementwiseOp,
typename CDEElementwiseOp>
using device_contraction_f64_kn_instance = std::tuple<
// clang-format off
//#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Compute|
//#####################################| | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| Data|
//#####################################| | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| Type|
//#####################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 128, 16, 2, 1, 16, 16, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 0, 1, 1, S<1, 16, 1, 16>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 128, 16, 2, 2, 16, 16, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 1, 1, 1, S<1, 16, 1, 16>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 128, 64, 16, 2, 1, 16, 16, 4, 4, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 0, 1, 1, S<1, 16, 1, 8>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 128, 64, 16, 2, 2, 16, 16, 4, 4, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 1, 1, 1, S<1, 16, 1, 8>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 64, 128, 16, 2, 1, 16, 16, 4, 4, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 0, 1, 1, S<1, 8, 1, 16>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 64, 128, 16, 2, 2, 16, 16, 4, 4, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 1, 1, 1, S<1, 8, 1, 16>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 64, 16, 2, 1, 16, 16, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 0, 1, 1, S<1, 16, 1, 16>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 64, 16, 2, 2, 16, 16, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 1, 1, 1, S<1, 16, 1, 16>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 64, 128, 16, 2, 1, 16, 16, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 0, 1, 1, S<1, 16, 1, 16>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 64, 128, 16, 2, 2, 16, 16, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 1, 1, 1, S<1, 16, 1, 16>, 1, ComputeDataType>
// clang-format on
>;
template <typename ADataType,
typename BDataType,
typename AccDataType,
typename CShuffleDataType,
typename DsDataType,
typename EDataType,
typename ComputeDataType,
typename AElementwiseOp,
typename BElementwiseOp,
typename CDEElementwiseOp>
using device_contraction_f64_mk_instance = std::tuple<
// clang-format off
//#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Compute|
//#####################################| | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| Data|
//#####################################| | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| Type|
//#####################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 128, 16, 1, 2, 16, 16, 4, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 16, 1, 16>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 128, 16, 2, 2, 16, 16, 4, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 16, 1, 16>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 128, 64, 16, 1, 2, 16, 16, 4, 4, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 16, 1, 8>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 128, 64, 16, 2, 2, 16, 16, 4, 4, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 16, 1, 8>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 64, 128, 16, 1, 2, 16, 16, 4, 4, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 8, 1, 16>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 64, 128, 16, 2, 2, 16, 16, 4, 4, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 8, 1, 16>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 64, 16, 1, 2, 16, 16, 4, 2, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 16, 1, 16>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 64, 16, 2, 2, 16, 16, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 16, 1, 16>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 64, 128, 16, 1, 2, 16, 16, 2, 4, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 16, 1, 16>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 64, 128, 16, 2, 2, 16, 16, 2, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 16, 1, 16>, 1, ComputeDataType>
// clang-format on
>;
template <typename ADataType,
typename BDataType,
typename AccDataType,
typename CShuffleDataType,
typename DsDataType,
typename EDataType,
typename ComputeDataType,
typename AElementwiseOp,
typename BElementwiseOp,
typename CDEElementwiseOp>
using device_contraction_f64_mn_instance = std::tuple<
// clang-format off
//#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Compute|
//#####################################| | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| Data|
//#####################################| | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| Type|
//#####################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 128, 16, 1, 1, 16, 16, 4, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 0, 1, 1, S<1, 16, 1, 16>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 128, 16, 2, 2, 16, 16, 4, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 1, 1, 1, S<1, 16, 1, 16>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 128, 64, 16, 1, 1, 16, 16, 4, 4, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 0, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 0, 1, 1, S<1, 16, 1, 8>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 128, 64, 16, 2, 2, 16, 16, 4, 4, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 1, 1, 1, S<1, 16, 1, 8>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 64, 128, 16, 1, 1, 16, 16, 4, 4, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 0, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 0, 1, 1, S<1, 8, 1, 16>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 64, 128, 16, 2, 2, 16, 16, 4, 4, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 1, 1, 1, S<1, 8, 1, 16>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 64, 16, 1, 1, 16, 16, 4, 2, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 0, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 0, 1, 1, S<1, 16, 1, 16>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 64, 16, 2, 2, 16, 16, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 1, 1, 1, S<1, 16, 1, 16>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 64, 128, 16, 1, 1, 16, 16, 2, 4, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 0, 1, 1, S<1, 16, 1, 16>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 64, 128, 16, 2, 2, 16, 16, 2, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 1, 1, 1, S<1, 16, 1, 16>, 1, ComputeDataType>
// clang-format on
>;
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/include/ck/library/tensor_operation_instance/gpu/contraction_bilinear.hpp
View file @ 29deceb6
...@@ -17,7 +17,6 @@ namespace tensor_operation { ...@@ -17,7 +17,6 @@ namespace tensor_operation {
namespace device { namespace device {
namespace instance { namespace instance {
#ifdef CK_ENABLE_FP32 #ifdef CK_ENABLE_FP32
// float
void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_kknn_instance( void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_kknn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2, std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2, 2,
...@@ -28,7 +27,8 @@ void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_kknn ...@@ -28,7 +27,8 @@ void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_kknn
F32, F32,
PassThrough, PassThrough,
PassThrough, PassThrough,
Bilinear>>>& instances); Bilinear,
F32>>>& instances);
void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_knnn_instance( void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_knnn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2, std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
...@@ -40,7 +40,8 @@ void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_knnn ...@@ -40,7 +40,8 @@ void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_knnn
F32, F32,
PassThrough, PassThrough,
PassThrough, PassThrough,
Bilinear>>>& instances); Bilinear,
F32>>>& instances);
void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mknn_instance( void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mknn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2, std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
...@@ -52,7 +53,8 @@ void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mknn ...@@ -52,7 +53,8 @@ void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mknn
F32, F32,
PassThrough, PassThrough,
PassThrough, PassThrough,
Bilinear>>>& instances); Bilinear,
F32>>>& instances);
void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mnnn_instance( void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mnnn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2, std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
...@@ -64,10 +66,115 @@ void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mnnn ...@@ -64,10 +66,115 @@ void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mnnn
F32, F32,
PassThrough, PassThrough,
PassThrough, PassThrough,
Bilinear>>>& instances); Bilinear,
#endif F32>>>& instances);
void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_f16_kknn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F32,
F32,
F32_Tuple,
F32,
PassThrough,
PassThrough,
Bilinear,
F16>>>& instances);
void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_f16_knnn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F32,
F32,
F32_Tuple,
F32,
PassThrough,
PassThrough,
Bilinear,
F16>>>& instances);
void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_f16_mknn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F32,
F32,
F32_Tuple,
F32,
PassThrough,
PassThrough,
Bilinear,
F16>>>& instances);
void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_f16_mnnn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F32,
F32,
F32_Tuple,
F32,
PassThrough,
PassThrough,
Bilinear,
F16>>>& instances);
void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_bf16_kknn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F32,
F32,
F32_Tuple,
F32,
PassThrough,
PassThrough,
Bilinear,
BF16>>>& instances);
void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_bf16_knnn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F32,
F32,
F32_Tuple,
F32,
PassThrough,
PassThrough,
Bilinear,
BF16>>>& instances);
void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_bf16_mknn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F32,
F32,
F32_Tuple,
F32,
PassThrough,
PassThrough,
Bilinear,
BF16>>>& instances);
void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_bf16_mnnn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F32,
F32,
F32_Tuple,
F32,
PassThrough,
PassThrough,
Bilinear,
BF16>>>& instances);
#endif // CK_ENABLE_FP32
#ifdef CK_ENABLE_FP64 #ifdef CK_ENABLE_FP64
// double
void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_kknn_instance( void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_kknn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2, std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2, 2,
...@@ -78,7 +185,8 @@ void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_kknn ...@@ -78,7 +185,8 @@ void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_kknn
F64, F64,
PassThrough, PassThrough,
PassThrough, PassThrough,
Bilinear>>>& instances); Bilinear,
F64>>>& instances);
void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_knnn_instance( void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_knnn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2, std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
...@@ -90,7 +198,8 @@ void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_knnn ...@@ -90,7 +198,8 @@ void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_knnn
F64, F64,
PassThrough, PassThrough,
PassThrough, PassThrough,
Bilinear>>>& instances); Bilinear,
F64>>>& instances);
void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_mknn_instance( void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_mknn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2, std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
...@@ -102,7 +211,8 @@ void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_mknn ...@@ -102,7 +211,8 @@ void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_mknn
F64, F64,
PassThrough, PassThrough,
PassThrough, PassThrough,
Bilinear>>>& instances); Bilinear,
F64>>>& instances);
void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_mnnn_instance( void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_mnnn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2, std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
...@@ -114,8 +224,170 @@ void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_mnnn ...@@ -114,8 +224,170 @@ void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_mnnn
F64, F64,
PassThrough, PassThrough,
PassThrough, PassThrough,
Bilinear>>>& instances); Bilinear,
#endif F64>>>& instances);
void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_compute_f32_kknn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F64,
F64,
F64_Tuple,
F64,
PassThrough,
PassThrough,
Bilinear,
F32>>>& instances);
void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_compute_f32_knnn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F64,
F64,
F64_Tuple,
F64,
PassThrough,
PassThrough,
Bilinear,
F32>>>& instances);
void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_compute_f32_mknn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F64,
F64,
F64_Tuple,
F64,
PassThrough,
PassThrough,
Bilinear,
F32>>>& instances);
void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_compute_f32_mnnn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F64,
F64,
F64_Tuple,
F64,
PassThrough,
PassThrough,
Bilinear,
F32>>>& instances);
#endif // CK_ENABLE_FP64
#ifdef CK_ENABLE_FP16
void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_f16_compute_f32_kknn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F16,
F16,
F16_Tuple,
F16,
PassThrough,
PassThrough,
Bilinear,
F32>>>& instances);
void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_f16_compute_f32_knnn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F16,
F16,
F16_Tuple,
F16,
PassThrough,
PassThrough,
Bilinear,
F32>>>& instances);
void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_f16_compute_f32_mknn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F16,
F16,
F16_Tuple,
F16,
PassThrough,
PassThrough,
Bilinear,
F32>>>& instances);
void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_f16_compute_f32_mnnn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F16,
F16,
F16_Tuple,
F16,
PassThrough,
PassThrough,
Bilinear,
F32>>>& instances);
#endif // CK_ENABLE_FP16
#ifdef CK_ENABLE_BF16
void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_bf16_compute_f32_kknn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
BF16,
BF16,
BF16_Tuple,
BF16,
PassThrough,
PassThrough,
Bilinear,
F32>>>& instances);
void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_bf16_compute_f32_knnn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
BF16,
BF16,
BF16_Tuple,
BF16,
PassThrough,
PassThrough,
Bilinear,
F32>>>& instances);
void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_bf16_compute_f32_mknn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
BF16,
BF16,
BF16_Tuple,
BF16,
PassThrough,
PassThrough,
Bilinear,
F32>>>& instances);
void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_bf16_compute_f32_mnnn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
BF16,
BF16,
BF16_Tuple,
BF16,
PassThrough,
PassThrough,
Bilinear,
F32>>>& instances);
#endif // CK_ENABLE_FP16
// Contraction + Bilinear // Contraction + Bilinear
template <index_t NumDimM, template <index_t NumDimM,
index_t NumDimN, index_t NumDimN,
...@@ -123,7 +395,8 @@ template <index_t NumDimM, ...@@ -123,7 +395,8 @@ template <index_t NumDimM,
typename ADataType, typename ADataType,
typename BDataType, typename BDataType,
typename DDataType, typename DDataType,
typename EDataType> typename EDataType,
typename ComputeDataType>
struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceContractionMultipleD< struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceContractionMultipleD<
NumDimM, NumDimM,
NumDimN, NumDimN,
...@@ -134,7 +407,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceContra ...@@ -134,7 +407,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceContra
EDataType, EDataType,
ck::tensor_operation::element_wise::PassThrough, ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough, ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::Bilinear>> ck::tensor_operation::element_wise::Bilinear,
ComputeDataType>>
{ {
using DeviceOp = DeviceContractionMultipleD<NumDimM, using DeviceOp = DeviceContractionMultipleD<NumDimM,
NumDimN, NumDimN,
...@@ -145,45 +419,125 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceContra ...@@ -145,45 +419,125 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceContra
EDataType, EDataType,
ck::tensor_operation::element_wise::PassThrough, ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough, ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::Bilinear>; ck::tensor_operation::element_wise::Bilinear,
ComputeDataType>;
static auto GetInstances() static auto GetInstances()
{ {
std::vector<std::unique_ptr<DeviceOp>> op_ptrs; std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
#ifdef CK_ENABLE_FP32 #ifdef CK_ENABLE_FP32
if constexpr(is_same_v<ADataType, float> && is_same_v<BDataType, float> && if constexpr(is_same_v<ADataType, float> && is_same_v<BDataType, float> &&
is_same_v<DDataType, float> && is_same_v<EDataType, float>) is_same_v<EDataType, float>)
{ {
if constexpr(NumDimM == 2 && NumDimN == 2 && NumDimK == 2) if constexpr(NumDimM == 2 && NumDimN == 2 && NumDimK == 2)
{ {
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_knnn_instance( if constexpr(is_same_v<ComputeDataType, float>)
op_ptrs); {
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_kknn_instance( add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_kknn_instance(
op_ptrs); op_ptrs);
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mnnn_instance( add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_knnn_instance(
op_ptrs); op_ptrs);
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mknn_instance( add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mknn_instance(
op_ptrs); op_ptrs);
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mnnn_instance(
op_ptrs);
}
else if constexpr(is_same_v<ComputeDataType, ck::half_t>)
{
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_f16_kknn_instance(
op_ptrs);
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_f16_knnn_instance(
op_ptrs);
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_f16_mknn_instance(
op_ptrs);
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_f16_mnnn_instance(
op_ptrs);
}
else if constexpr(is_same_v<ComputeDataType, ck::bhalf_t>)
{
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_bf16_kknn_instance(
op_ptrs);
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_bf16_knnn_instance(
op_ptrs);
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_bf16_mknn_instance(
op_ptrs);
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_bf16_mnnn_instance(
op_ptrs);
}
} }
} }
#endif #endif // CK_ENABLE_FP32
#ifdef CK_ENABLE_FP64 #ifdef CK_ENABLE_FP64
if constexpr(is_same_v<ADataType, double> && is_same_v<BDataType, double> && if constexpr(is_same_v<ADataType, double> && is_same_v<BDataType, double> &&
is_same_v<DDataType, double> && is_same_v<EDataType, double>) is_same_v<EDataType, double>)
{
if constexpr(NumDimM == 2 && NumDimN == 2 && NumDimK == 2)
{
if constexpr(is_same_v<ComputeDataType, double>)
{
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_kknn_instance(
op_ptrs);
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_knnn_instance(
op_ptrs);
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_mknn_instance(
op_ptrs);
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_mnnn_instance(
op_ptrs);
}
else if constexpr(is_same_v<ComputeDataType, float>)
{
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_compute_f32_kknn_instance(
op_ptrs);
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_compute_f32_knnn_instance(
op_ptrs);
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_compute_f32_mknn_instance(
op_ptrs);
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_compute_f32_mnnn_instance(
op_ptrs);
}
}
}
#endif // CK_ENABLE_FP64
#ifdef CK_ENABLE_FP16
if constexpr(is_same_v<ADataType, ck::half_t> && is_same_v<BDataType, ck::half_t> &&
is_same_v<EDataType, ck::half_t>)
{
if constexpr(NumDimM == 2 && NumDimN == 2 && NumDimK == 2)
{
if constexpr(is_same_v<ComputeDataType, float>)
{
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_f16_compute_f32_kknn_instance(
op_ptrs);
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_f16_compute_f32_knnn_instance(
op_ptrs);
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_f16_compute_f32_mknn_instance(
op_ptrs);
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_f16_compute_f32_mnnn_instance(
op_ptrs);
}
}
}
#endif // CK_ENABLE_FP16
#ifdef CK_ENABLE_BF16
if constexpr(is_same_v<ADataType, ck::bhalf_t> && is_same_v<BDataType, ck::bhalf_t> &&
is_same_v<EDataType, ck::bhalf_t>)
{ {
if constexpr(NumDimM == 2 && NumDimN == 2 && NumDimK == 2) if constexpr(NumDimM == 2 && NumDimN == 2 && NumDimK == 2)
{ {
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_kknn_instance( if constexpr(is_same_v<ComputeDataType, float>)
op_ptrs); {
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_knnn_instance( add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_bf16_compute_f32_kknn_instance(
op_ptrs); op_ptrs);
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_mnnn_instance( add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_bf16_compute_f32_knnn_instance(
op_ptrs); op_ptrs);
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_mknn_instance( add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_bf16_compute_f32_mknn_instance(
op_ptrs); op_ptrs);
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_bf16_compute_f32_mnnn_instance(
op_ptrs);
}
} }
} }
#endif #endif // CK_ENABLE_BF16
return op_ptrs; return op_ptrs;
} }
}; };
......
library/include/ck/library/tensor_operation_instance/gpu/contraction_scale.hpp
View file @ 29deceb6
...@@ -17,7 +17,6 @@ namespace tensor_operation { ...@@ -17,7 +17,6 @@ namespace tensor_operation {
namespace device { namespace device {
namespace instance { namespace instance {
#ifdef CK_ENABLE_FP32 #ifdef CK_ENABLE_FP32
// float
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_kkn_instance( void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_kkn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2, std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2, 2,
...@@ -28,7 +27,8 @@ void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_kkn_instanc ...@@ -28,7 +27,8 @@ void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_kkn_instanc
F32, F32,
PassThrough, PassThrough,
PassThrough, PassThrough,
Scale>>>& instances); Scale,
F32>>>& instances);
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_knn_instance( void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_knn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2, std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
...@@ -40,7 +40,8 @@ void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_knn_instanc ...@@ -40,7 +40,8 @@ void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_knn_instanc
F32, F32,
PassThrough, PassThrough,
PassThrough, PassThrough,
Scale>>>& instances); Scale,
F32>>>& instances);
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mkn_instance( void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mkn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2, std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
...@@ -52,7 +53,8 @@ void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mkn_instanc ...@@ -52,7 +53,8 @@ void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mkn_instanc
F32, F32,
PassThrough, PassThrough,
PassThrough, PassThrough,
Scale>>>& instances); Scale,
F32>>>& instances);
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mnn_instance( void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mnn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2, std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
...@@ -64,10 +66,115 @@ void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mnn_instanc ...@@ -64,10 +66,115 @@ void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mnn_instanc
F32, F32,
PassThrough, PassThrough,
PassThrough, PassThrough,
Scale>>>& instances); Scale,
#endif F32>>>& instances);
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_f16_kkn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F32,
F32,
Empty_Tuple,
F32,
PassThrough,
PassThrough,
Scale,
F16>>>& instances);
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_f16_knn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F32,
F32,
Empty_Tuple,
F32,
PassThrough,
PassThrough,
Scale,
F16>>>& instances);
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_f16_mkn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F32,
F32,
Empty_Tuple,
F32,
PassThrough,
PassThrough,
Scale,
F16>>>& instances);
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_f16_mnn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F32,
F32,
Empty_Tuple,
F32,
PassThrough,
PassThrough,
Scale,
F16>>>& instances);
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_bf16_kkn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F32,
F32,
Empty_Tuple,
F32,
PassThrough,
PassThrough,
Scale,
BF16>>>& instances);
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_bf16_knn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F32,
F32,
Empty_Tuple,
F32,
PassThrough,
PassThrough,
Scale,
BF16>>>& instances);
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_bf16_mkn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F32,
F32,
Empty_Tuple,
F32,
PassThrough,
PassThrough,
Scale,
BF16>>>& instances);
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_bf16_mnn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F32,
F32,
Empty_Tuple,
F32,
PassThrough,
PassThrough,
Scale,
BF16>>>& instances);
#endif // CK_ENABLE_FP32
#ifdef CK_ENABLE_FP64 #ifdef CK_ENABLE_FP64
// double
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_kkn_instance( void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_kkn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2, std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2, 2,
...@@ -78,7 +185,8 @@ void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_kkn_instanc ...@@ -78,7 +185,8 @@ void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_kkn_instanc
F64, F64,
PassThrough, PassThrough,
PassThrough, PassThrough,
Scale>>>& instances); Scale,
F64>>>& instances);
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_knn_instance( void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_knn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2, std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
...@@ -90,7 +198,8 @@ void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_knn_instanc ...@@ -90,7 +198,8 @@ void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_knn_instanc
F64, F64,
PassThrough, PassThrough,
PassThrough, PassThrough,
Scale>>>& instances); Scale,
F64>>>& instances);
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mkn_instance( void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mkn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2, std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
...@@ -102,7 +211,8 @@ void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mkn_instanc ...@@ -102,7 +211,8 @@ void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mkn_instanc
F64, F64,
PassThrough, PassThrough,
PassThrough, PassThrough,
Scale>>>& instances); Scale,
F64>>>& instances);
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mnn_instance( void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mnn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2, std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
...@@ -114,15 +224,178 @@ void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mnn_instanc ...@@ -114,15 +224,178 @@ void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mnn_instanc
F64, F64,
PassThrough, PassThrough,
PassThrough, PassThrough,
Scale>>>& instances); Scale,
#endif F64>>>& instances);
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_compute_f32_kkn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F64,
F64,
Empty_Tuple,
F64,
PassThrough,
PassThrough,
Scale,
F32>>>& instances);
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_compute_f32_knn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F64,
F64,
Empty_Tuple,
F64,
PassThrough,
PassThrough,
Scale,
F32>>>& instances);
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_compute_f32_mkn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F64,
F64,
Empty_Tuple,
F64,
PassThrough,
PassThrough,
Scale,
F32>>>& instances);
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_compute_f32_mnn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F64,
F64,
Empty_Tuple,
F64,
PassThrough,
PassThrough,
Scale,
F32>>>& instances);
#endif // CK_ENABLE_FP64
#ifdef CK_ENABLE_FP16
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_compute_f32_kkn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
Scale,
F32>>>& instances);
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_compute_f32_knn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
Scale,
F32>>>& instances);
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_compute_f32_mkn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
Scale,
F32>>>& instances);
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_compute_f32_mnn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
Scale,
F32>>>& instances);
#endif // CK_ENABLE_FP16
#ifdef CK_ENABLE_BF16
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_compute_f32_kkn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
BF16,
BF16,
Empty_Tuple,
BF16,
PassThrough,
PassThrough,
Scale,
F32>>>& instances);
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_compute_f32_knn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
BF16,
BF16,
Empty_Tuple,
BF16,
PassThrough,
PassThrough,
Scale,
F32>>>& instances);
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_compute_f32_mkn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
BF16,
BF16,
Empty_Tuple,
BF16,
PassThrough,
PassThrough,
Scale,
F32>>>& instances);
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_compute_f32_mnn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
BF16,
BF16,
Empty_Tuple,
BF16,
PassThrough,
PassThrough,
Scale,
F32>>>& instances);
#endif // CK_ENABLE_FP16
// Contraction + Scale // Contraction + Scale
template <index_t NumDimM, template <index_t NumDimM,
index_t NumDimN, index_t NumDimN,
index_t NumDimK, index_t NumDimK,
typename ADataType, typename ADataType,
typename BDataType, typename BDataType,
typename EDataType> typename EDataType,
typename ComputeDataType>
struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceContractionMultipleD< struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceContractionMultipleD<
NumDimM, NumDimM,
NumDimN, NumDimN,
...@@ -133,7 +406,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceContra ...@@ -133,7 +406,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceContra
EDataType, EDataType,
ck::tensor_operation::element_wise::PassThrough, ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough, ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::Scale>> ck::tensor_operation::element_wise::Scale,
ComputeDataType>>
{ {
using DeviceOp = DeviceContractionMultipleD<NumDimM, using DeviceOp = DeviceContractionMultipleD<NumDimM,
NumDimN, NumDimN,
...@@ -144,7 +418,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceContra ...@@ -144,7 +418,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceContra
EDataType, EDataType,
ck::tensor_operation::element_wise::PassThrough, ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough, ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::Scale>; ck::tensor_operation::element_wise::Scale,
ComputeDataType>;
static auto GetInstances() static auto GetInstances()
{ {
...@@ -155,34 +430,113 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceContra ...@@ -155,34 +430,113 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceContra
{ {
if constexpr(NumDimM == 2 && NumDimN == 2 && NumDimK == 2) if constexpr(NumDimM == 2 && NumDimN == 2 && NumDimK == 2)
{ {
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_kkn_instance( if constexpr(is_same_v<ComputeDataType, float>)
op_ptrs); {
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_knn_instance( add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_kkn_instance(
op_ptrs); op_ptrs);
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mkn_instance( add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_knn_instance(
op_ptrs); op_ptrs);
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mnn_instance( add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mkn_instance(
op_ptrs); op_ptrs);
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mnn_instance(
op_ptrs);
}
else if constexpr(is_same_v<ComputeDataType, ck::half_t>)
{
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_f16_kkn_instance(
op_ptrs);
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_f16_knn_instance(
op_ptrs);
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_f16_mkn_instance(
op_ptrs);
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_f16_mnn_instance(
op_ptrs);
}
else if constexpr(is_same_v<ComputeDataType, ck::bhalf_t>)
{
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_bf16_kkn_instance(
op_ptrs);
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_bf16_knn_instance(
op_ptrs);
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_bf16_mkn_instance(
op_ptrs);
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_bf16_mnn_instance(
op_ptrs);
}
} }
} }
#endif #endif // CK_ENABLE_FP32
#ifdef CK_ENABLE_FP64 #ifdef CK_ENABLE_FP64
if constexpr(is_same_v<ADataType, double> && is_same_v<BDataType, double> && if constexpr(is_same_v<ADataType, double> && is_same_v<BDataType, double> &&
is_same_v<EDataType, double>) is_same_v<EDataType, double>)
{ {
if constexpr(NumDimM == 2 && NumDimN == 2 && NumDimK == 2) if constexpr(NumDimM == 2 && NumDimN == 2 && NumDimK == 2)
{ {
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_kkn_instance( if constexpr(is_same_v<ComputeDataType, double>)
op_ptrs); {
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_knn_instance( add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_kkn_instance(
op_ptrs); op_ptrs);
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mkn_instance( add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_knn_instance(
op_ptrs); op_ptrs);
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mnn_instance( add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mkn_instance(
op_ptrs); op_ptrs);
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mnn_instance(
op_ptrs);
}
else if constexpr(is_same_v<ComputeDataType, float>)
{
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_compute_f32_kkn_instance(
op_ptrs);
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_compute_f32_knn_instance(
op_ptrs);
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_compute_f32_mkn_instance(
op_ptrs);
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_compute_f32_mnn_instance(
op_ptrs);
}
}
}
#endif // CK_ENABLE_FP64
#ifdef CK_ENABLE_FP16
if constexpr(is_same_v<ADataType, ck::half_t> && is_same_v<BDataType, ck::half_t> &&
is_same_v<EDataType, ck::half_t>)
{
if constexpr(NumDimM == 2 && NumDimN == 2 && NumDimK == 2)
{
if constexpr(is_same_v<ComputeDataType, float>)
{
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_compute_f32_kkn_instance(
op_ptrs);
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_compute_f32_knn_instance(
op_ptrs);
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_compute_f32_mkn_instance(
op_ptrs);
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_compute_f32_mnn_instance(
op_ptrs);
}
}
}
#endif // CK_ENABLE_FP16
#ifdef CK_ENABLE_BF16
if constexpr(is_same_v<ADataType, ck::bhalf_t> && is_same_v<BDataType, ck::bhalf_t> &&
is_same_v<EDataType, ck::bhalf_t>)
{
if constexpr(NumDimM == 2 && NumDimN == 2 && NumDimK == 2)
{
if constexpr(is_same_v<ComputeDataType, float>)
{
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_compute_f32_kkn_instance(
op_ptrs);
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_compute_f32_knn_instance(
op_ptrs);
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_compute_f32_mkn_instance(
op_ptrs);
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_compute_f32_mnn_instance(
op_ptrs);
}
} }
} }
#endif #endif // CK_ENABLE_BF16
return op_ptrs; return op_ptrs;
} }
}; };
......
library/include/ck/library/tensor_operation_instance/gpu/gemm.hpp
View file @ 29deceb6
...@@ -227,6 +227,10 @@ void add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_instances( ...@@ -227,6 +227,10 @@ void add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(
DeviceGemm<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>& DeviceGemm<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances); instances);
void add_device_gemm_xdl_c_shuffle_lds_direct_load_f16_f16_f16_mk_nk_mn_instances(
std::vector<std::unique_ptr<
DeviceGemm<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
#endif #endif
#ifdef CK_ENABLE_BF16 #ifdef CK_ENABLE_BF16
void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_kn_mn_instances( void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_kn_mn_instances(
...@@ -289,6 +293,26 @@ void add_device_gemm_xdl_f32_f32_f32_mk_nk_mn_instances( ...@@ -289,6 +293,26 @@ void add_device_gemm_xdl_f32_f32_f32_mk_nk_mn_instances(
std::vector<std::unique_ptr< std::vector<std::unique_ptr<
DeviceGemm<Row, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>& DeviceGemm<Row, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
instances); instances);
void add_device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_km_kn_mn_instances(
std::vector<std::unique_ptr<
DeviceGemm<Col, Row, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_km_nk_mn_instances(
std::vector<std::unique_ptr<
DeviceGemm<Col, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_mk_kn_mn_instances(
std::vector<std::unique_ptr<
DeviceGemm<Row, Row, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_mk_nk_mn_instances(
std::vector<std::unique_ptr<
DeviceGemm<Row, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
instances);
#endif #endif
#ifdef CK_ENABLE_FP64 #ifdef CK_ENABLE_FP64
void add_device_gemm_xdl_f64_f64_f64_km_kn_mn_instances( void add_device_gemm_xdl_f64_f64_f64_km_kn_mn_instances(
...@@ -328,7 +352,18 @@ void add_device_gemm_xdl_c_shuffle_f8_f8_f8_mk_kn_mn_instances( ...@@ -328,7 +352,18 @@ void add_device_gemm_xdl_c_shuffle_f8_f8_f8_mk_kn_mn_instances(
void add_device_gemm_xdl_c_shuffle_f8_f8_f8_mk_nk_mn_instances( void add_device_gemm_xdl_c_shuffle_f8_f8_f8_mk_nk_mn_instances(
std::vector<std::unique_ptr< std::vector<std::unique_ptr<
DeviceGemm<Row, Col, Row, F8, F8, F8, PassThrough, PassThrough, PassThrough>>>& instances); DeviceGemm<Row, Col, Row, F8, F8, F8, PassThrough, PassThrough, PassThrough>>>& instances);
void add_device_gemm_xdl_c_shuffle_f16_f8_f16_mk_kn_mn_instances(
std::vector<std::unique_ptr<
DeviceGemm<Row, Row, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_c_shuffle_f16_f8_f16_mk_nk_mn_instances(
std::vector<std::unique_ptr<
DeviceGemm<Row, Col, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
#endif #endif
template <typename ALayout, template <typename ALayout,
typename BLayout, typename BLayout,
typename CLayout, typename CLayout,
...@@ -366,38 +401,46 @@ struct DeviceOperationInstanceFactory< ...@@ -366,38 +401,46 @@ struct DeviceOperationInstanceFactory<
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> && if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
is_same_v<CLayout, Row>) is_same_v<CLayout, Row>)
{ {
add_device_gemm_xdl_f32_f32_f32_mk_kn_mn_instances(op_ptrs); /// add_device_gemm_xdl_f32_f32_f32_mk_kn_mn_instances(op_ptrs);
#ifdef DL_KERNELS #ifdef DL_KERNELS
add_device_gemm_dl_f32_f32_f32_mk_kn_mn_instances(op_ptrs); add_device_gemm_dl_f32_f32_f32_mk_kn_mn_instances(op_ptrs);
#endif #endif
add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_kn_mn_instances(op_ptrs); add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_kn_mn_instances(op_ptrs);
add_device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_mk_kn_mn_instances(
op_ptrs);
} }
else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> && else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
is_same_v<CLayout, Row>) is_same_v<CLayout, Row>)
{ {
add_device_gemm_xdl_f32_f32_f32_mk_nk_mn_instances(op_ptrs); /// add_device_gemm_xdl_f32_f32_f32_mk_nk_mn_instances(op_ptrs);
#ifdef DL_KERNELS #ifdef DL_KERNELS
add_device_gemm_dl_f32_f32_f32_mk_nk_mn_instances(op_ptrs); add_device_gemm_dl_f32_f32_f32_mk_nk_mn_instances(op_ptrs);
#endif #endif
add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instances(op_ptrs); add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instances(op_ptrs);
add_device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_mk_nk_mn_instances(
op_ptrs);
} }
else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> && else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
is_same_v<CLayout, Row>) is_same_v<CLayout, Row>)
{ {
add_device_gemm_xdl_f32_f32_f32_km_kn_mn_instances(op_ptrs); /// add_device_gemm_xdl_f32_f32_f32_km_kn_mn_instances(op_ptrs);
#ifdef DL_KERNELS #ifdef DL_KERNELS
add_device_gemm_dl_f32_f32_f32_km_kn_mn_instances(op_ptrs); add_device_gemm_dl_f32_f32_f32_km_kn_mn_instances(op_ptrs);
#endif #endif
add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_kn_mn_instances(op_ptrs); add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_kn_mn_instances(op_ptrs);
add_device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_km_kn_mn_instances(
op_ptrs);
} }
else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> && else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
is_same_v<CLayout, Row>) is_same_v<CLayout, Row>)
{ {
add_device_gemm_xdl_f32_f32_f32_km_nk_mn_instances(op_ptrs); /// add_device_gemm_xdl_f32_f32_f32_km_nk_mn_instances(op_ptrs);
#ifdef DL_KERNELS #ifdef DL_KERNELS
add_device_gemm_dl_f32_f32_f32_km_nk_mn_instances(op_ptrs); add_device_gemm_dl_f32_f32_f32_km_nk_mn_instances(op_ptrs);
#endif #endif
add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_nk_mn_instances(op_ptrs); add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_nk_mn_instances(op_ptrs);
add_device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_km_nk_mn_instances(
op_ptrs);
} }
} }
#ifdef CK_ENABLE_FP16 #ifdef CK_ENABLE_FP16
...@@ -407,7 +450,7 @@ struct DeviceOperationInstanceFactory< ...@@ -407,7 +450,7 @@ struct DeviceOperationInstanceFactory<
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> && if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
is_same_v<CLayout, Row>) is_same_v<CLayout, Row>)
{ {
add_device_gemm_xdl_f16_f16_f16_mk_kn_mn_instances(op_ptrs); /// add_device_gemm_xdl_f16_f16_f16_mk_kn_mn_instances(op_ptrs);
#ifdef DL_KERNELS #ifdef DL_KERNELS
add_device_gemm_dl_f16_f16_f16_mk_kn_mn_instances(op_ptrs); add_device_gemm_dl_f16_f16_f16_mk_kn_mn_instances(op_ptrs);
add_device_gemm_dl_f16_f16_f16_mk_kn_mn_irregular_instances(op_ptrs); add_device_gemm_dl_f16_f16_f16_mk_kn_mn_irregular_instances(op_ptrs);
...@@ -419,7 +462,7 @@ struct DeviceOperationInstanceFactory< ...@@ -419,7 +462,7 @@ struct DeviceOperationInstanceFactory<
else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> && else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
is_same_v<CLayout, Row>) is_same_v<CLayout, Row>)
{ {
add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(op_ptrs); /// add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(op_ptrs);
#ifdef DL_KERNELS #ifdef DL_KERNELS
add_device_gemm_dl_f16_f16_f16_mk_nk_mn_instances(op_ptrs); add_device_gemm_dl_f16_f16_f16_mk_nk_mn_instances(op_ptrs);
add_device_gemm_dl_f16_f16_f16_mk_nk_mn_irregular_instances(op_ptrs); add_device_gemm_dl_f16_f16_f16_mk_nk_mn_irregular_instances(op_ptrs);
...@@ -428,11 +471,13 @@ struct DeviceOperationInstanceFactory< ...@@ -428,11 +471,13 @@ struct DeviceOperationInstanceFactory<
#endif #endif
add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances(op_ptrs); add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances(op_ptrs);
add_device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instances(op_ptrs); add_device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instances(op_ptrs);
add_device_gemm_xdl_c_shuffle_lds_direct_load_f16_f16_f16_mk_nk_mn_instances(
op_ptrs);
} }
else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> && else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
is_same_v<CLayout, Row>) is_same_v<CLayout, Row>)
{ {
add_device_gemm_xdl_f16_f16_f16_km_kn_mn_instances(op_ptrs); /// add_device_gemm_xdl_f16_f16_f16_km_kn_mn_instances(op_ptrs);
#ifdef DL_KERNELS #ifdef DL_KERNELS
add_device_gemm_dl_f16_f16_f16_km_kn_mn_instances(op_ptrs); add_device_gemm_dl_f16_f16_f16_km_kn_mn_instances(op_ptrs);
add_device_gemm_dl_f16_f16_f16_km_kn_mn_irregular_instances(op_ptrs); add_device_gemm_dl_f16_f16_f16_km_kn_mn_irregular_instances(op_ptrs);
...@@ -444,7 +489,7 @@ struct DeviceOperationInstanceFactory< ...@@ -444,7 +489,7 @@ struct DeviceOperationInstanceFactory<
else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> && else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
is_same_v<CLayout, Row>) is_same_v<CLayout, Row>)
{ {
add_device_gemm_xdl_f16_f16_f16_km_nk_mn_instances(op_ptrs); /// add_device_gemm_xdl_f16_f16_f16_km_nk_mn_instances(op_ptrs);
#ifdef DL_KERNELS #ifdef DL_KERNELS
add_device_gemm_dl_f16_f16_f16_km_nk_mn_instances(op_ptrs); add_device_gemm_dl_f16_f16_f16_km_nk_mn_instances(op_ptrs);
add_device_gemm_dl_f16_f16_f16_km_nk_mn_irregular_instances(op_ptrs); add_device_gemm_dl_f16_f16_f16_km_nk_mn_irregular_instances(op_ptrs);
...@@ -548,6 +593,20 @@ struct DeviceOperationInstanceFactory< ...@@ -548,6 +593,20 @@ struct DeviceOperationInstanceFactory<
add_device_gemm_xdl_c_shuffle_f8_f8_f8_km_nk_mn_instances(op_ptrs); add_device_gemm_xdl_c_shuffle_f8_f8_f8_km_nk_mn_instances(op_ptrs);
} }
} }
else if constexpr(is_same_v<ADataType, ck::half_t> && is_same_v<BDataType, ck::f8_t> &&
is_same_v<CDataType, ck::half_t>)
{
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
is_same_v<CLayout, Row>)
{
add_device_gemm_xdl_c_shuffle_f16_f8_f16_mk_kn_mn_instances(op_ptrs);
}
else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
is_same_v<CLayout, Row>)
{
add_device_gemm_xdl_c_shuffle_f16_f8_f16_mk_nk_mn_instances(op_ptrs);
}
}
#endif #endif
return op_ptrs; return op_ptrs;
} }
......
library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_instance.hpp
View file @ 29deceb6
...@@ -3,7 +3,7 @@ ...@@ -3,7 +3,7 @@
#include "ck/ck.hpp" #include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp" #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp" #include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_abd_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp" #include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp" #include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp" #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
...@@ -55,24 +55,24 @@ using device_grouped_conv_fwd_xdl_bf16_instances = std::tuple< ...@@ -55,24 +55,24 @@ using device_grouped_conv_fwd_xdl_bf16_instances = std::tuple<
//########################################| | | | | | | | | | | | Operation| Operation| Operation| | | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| //########################################| | | | | | | | | | | | Operation| Operation| Operation| | | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | //########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// generic instance // generic instance
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF16, BF16, F32, BF16, DsLayout, BF16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF16, BF16, F32, BF16, DsLayout, BF16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1>,
// instances for small conv.K and conv.C // instances for small conv.K and conv.C
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF16, BF16, F32, BF16, DsLayout, BF16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 1>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF16, BF16, F32, BF16, DsLayout, BF16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 1>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF16, BF16, F32, BF16, DsLayout, BF16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF16, BF16, F32, BF16, DsLayout, BF16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF16, BF16, F32, BF16, DsLayout, BF16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF16, BF16, F32, BF16, DsLayout, BF16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF16, BF16, F32, BF16, DsLayout, BF16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF16, BF16, F32, BF16, DsLayout, BF16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF16, BF16, F32, BF16, DsLayout, BF16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF16, BF16, F32, BF16, DsLayout, BF16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF16, BF16, F32, BF16, DsLayout, BF16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF16, BF16, F32, BF16, DsLayout, BF16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF16, BF16, F32, BF16, DsLayout, BF16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF16, BF16, F32, BF16, DsLayout, BF16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF16, BF16, F32, BF16, DsLayout, BF16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF16, BF16, F32, BF16, DsLayout, BF16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF16, BF16, F32, BF16, DsLayout, BF16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF16, BF16, F32, BF16, DsLayout, BF16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF16, BF16, F32, BF16, DsLayout, BF16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF16, BF16, F32, BF16, DsLayout, BF16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF16, BF16, F32, BF16, DsLayout, BF16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF16, BF16, F32, BF16, DsLayout, BF16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF16, BF16, F32, BF16, DsLayout, BF16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF16, BF16, F32, BF16, DsLayout, BF16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF16, BF16, F32, BF16, DsLayout, BF16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF16, BF16, F32, BF16, DsLayout, BF16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF16, BF16, F32, BF16, DsLayout, BF16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF16, BF16, F32, BF16, DsLayout, BF16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF16, BF16, F32, BF16, DsLayout, BF16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8> DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF16, BF16, F32, BF16, DsLayout, BF16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>
// clang-format on // clang-format on
>; >;
...@@ -89,24 +89,24 @@ using device_grouped_conv_fwd_xdl_f16_instances = std::tuple< ...@@ -89,24 +89,24 @@ using device_grouped_conv_fwd_xdl_f16_instances = std::tuple<
//########################################| | | | | | | | | | | | Operation| Operation| Operation| | | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| //########################################| | | | | | | | | | | | Operation| Operation| Operation| | | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | //########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// generic instance // generic instance
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1>,
// instances for small conv.K and conv.C // instances for small conv.K and conv.C
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 1>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 1>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8> DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>
// clang-format on // clang-format on
>; >;
...@@ -123,24 +123,24 @@ using device_grouped_conv_fwd_xdl_f32_instances = std::tuple< ...@@ -123,24 +123,24 @@ using device_grouped_conv_fwd_xdl_f32_instances = std::tuple<
//########################################| | | | | | | | | | | | Operation| Operation| Operation| | | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| //########################################| | | | | | | | | | | | Operation| Operation| Operation| | | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | //########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// generic instance // generic instance
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsLayout, F32, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 64, 16, 4, 4, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 8, 1, 8>, 1>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsLayout, F32, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 64, 16, 4, 4, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 8, 1, 8>, 1>,
// instances for small conv.K and conv.C // instances for small conv.K and conv.C
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsLayout, F32, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 32, 16, 4, 4, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 1>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsLayout, F32, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 32, 16, 4, 4, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 1>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsLayout, F32, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 128, 128, 16, 4, 4, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsLayout, F32, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 128, 128, 16, 4, 4, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsLayout, F32, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 256, 128, 16, 4, 4, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsLayout, F32, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 256, 128, 16, 4, 4, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsLayout, F32, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 128, 256, 16, 4, 4, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsLayout, F32, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 128, 256, 16, 4, 4, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsLayout, F32, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 128, 128, 128, 16, 4, 4, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsLayout, F32, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 128, 128, 128, 16, 4, 4, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsLayout, F32, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 128, 128, 16, 4, 4, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsLayout, F32, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 128, 128, 16, 4, 4, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsLayout, F32, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 128, 128, 64, 16, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 8>, 4>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsLayout, F32, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 128, 128, 64, 16, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 8>, 4>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsLayout, F32, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 128, 64, 128, 16, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsLayout, F32, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 128, 64, 128, 16, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsLayout, F32, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 64, 16, 4, 4, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsLayout, F32, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 64, 16, 4, 4, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsLayout, F32, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 128, 64, 16, 4, 4, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsLayout, F32, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 128, 64, 16, 4, 4, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsLayout, F32, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 64, 128, 16, 4, 4, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsLayout, F32, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 64, 128, 16, 4, 4, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsLayout, F32, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 128, 128, 32, 16, 4, 4, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 8>, 4>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsLayout, F32, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 128, 128, 32, 16, 4, 4, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 8>, 4>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsLayout, F32, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 128, 32, 128, 16, 4, 4, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsLayout, F32, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 128, 32, 128, 16, 4, 4, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsLayout, F32, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 32, 16, 4, 4, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsLayout, F32, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 32, 16, 4, 4, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsLayout, F32, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 32, 64, 16, 4, 4, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4> DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsLayout, F32, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 32, 64, 16, 4, 4, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4>
// clang-format on // clang-format on
>; >;
...@@ -157,24 +157,24 @@ using device_grouped_conv_fwd_xdl_int8_instances = std::tuple< ...@@ -157,24 +157,24 @@ using device_grouped_conv_fwd_xdl_int8_instances = std::tuple<
//########################################| | | | | | | | | | | | Operation| Operation| Operation| | | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| //########################################| | | | | | | | | | | | Operation| Operation| Operation| | | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | //########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// generic instance // generic instance
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, DsLayout, int8_t, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, DsLayout, int8_t, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1>,
// instances for small conv.K and conv.C // instances for small conv.K and conv.C
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, DsLayout, int8_t, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 1>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, DsLayout, int8_t, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 1>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, DsLayout, int8_t, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, DsLayout, int8_t, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, DsLayout, int8_t, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, DsLayout, int8_t, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, DsLayout, int8_t, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, DsLayout, int8_t, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, DsLayout, int8_t, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, DsLayout, int8_t, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, DsLayout, int8_t, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, DsLayout, int8_t, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, DsLayout, int8_t, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, DsLayout, int8_t, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, DsLayout, int8_t, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, DsLayout, int8_t, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, DsLayout, int8_t, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, DsLayout, int8_t, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, DsLayout, int8_t, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, DsLayout, int8_t, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, DsLayout, int8_t, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, DsLayout, int8_t, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, DsLayout, int8_t, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, DsLayout, int8_t, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, DsLayout, int8_t, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, DsLayout, int8_t, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, DsLayout, int8_t, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, DsLayout, int8_t, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, DsLayout, int8_t, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8> DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, DsLayout, int8_t, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>
// clang-format on // clang-format on
>; >;
...@@ -192,24 +192,24 @@ using device_grouped_conv_fwd_xdl_f16_comp_f8_instances = std::tuple< ...@@ -192,24 +192,24 @@ using device_grouped_conv_fwd_xdl_f16_comp_f8_instances = std::tuple<
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | //########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#ifdef CK_ENABLE_FP8 #ifdef CK_ENABLE_FP8
// generic instance // generic instance
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, F8>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, F8>,
// instances for small conv.K and conv.C // instances for small conv.K and conv.C
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, F8>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, F8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, F8>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, F8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, F8>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, F8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, F8>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, F8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, F8>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, F8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, F8>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, F8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, F8>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, F8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, F8>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, F8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, F8>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, F8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, F8>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, F8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, F8>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, F8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, F8>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, F8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, F8>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, F8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, F8>, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, F8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, F8> DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, F8>
#endif #endif
// clang-format on // clang-format on
>; >;
......
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