Skip to content

Commit d927221

Browse files
JeniferC99JeniferC99
authored
Merge pull request #3076 from ROCm/congma/release/rocm-rel-7.1/revert-2610
[7.1.1] cherry picks
2 parents 57d8d66 + 8f0182b commit d927221

File tree

5 files changed

+168
-290
lines changed

5 files changed

+168
-290
lines changed

‎include/ck/ck.hpp‎

Lines changed: 3 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -222,6 +222,9 @@
222222
// TODO: separate index calculation into "compile-time", "global", "block", "wave", "thread"
223223
#define CK_HACK_MERGE_CALCULATE_IDX_DIFF_LOW_CONST_USE_AMD_GCN_READ_FIRST_LANE 0
224224

225+
// workaround: conv crash when K, C is even
226+
#define CK_WORKAROUND_DISABLE_FILTER1x1STRIDE1PAD0_WHEN_K_C_IS_EVEN 1
227+
225228
// workaround: compiler crash when compiling recursive lambda
226229
#define CK_WORKAROUND_SWDEV_275126 1
227230

‎include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_xdl_cshuffle_v3.hpp‎

Lines changed: 9 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -331,8 +331,8 @@ struct DeviceGroupedConvBwdWeight_Xdl_CShuffleV3
331331
using CGridDesc_M_N = remove_cvref_t<decltype(ABCGridDescs{}[I2])>;
332332

333333
using GridwiseGemm = GridwiseGemm_xdl_cshuffle_conv_v3<
334-
tensor_layout::gemm::ColumnMajor,
335334
tensor_layout::gemm::RowMajor,
335+
tensor_layout::gemm::ColumnMajor,
336336
tensor_layout::gemm::RowMajor,
337337
ADataType,
338338
BDataType,
@@ -1299,6 +1299,13 @@ struct DeviceGroupedConvBwdWeight_Xdl_CShuffleV3
12991299
if constexpr(ConvBackwardWeightSpecialization ==
13001300
ConvolutionBackwardWeightSpecialization::Filter1x1Stride1Pad0)
13011301
{
1302+
// workaround: disable when K, C is even
1303+
#if CK_WORKAROUND_DISABLE_FILTER1x1STRIDE1PAD0_WHEN_K_C_IS_EVEN
1304+
if(arg.Conv_C_ % 2 == 0 || arg.Conv_K_ % 2 == 0)
1305+
{
1306+
return false;
1307+
}
1308+
#endif
13021309
// check if it's 1x1, stride=1 pad = 0 conv
13031310
for(int i = 0; i < NDimSpatial; i++)
13041311
{
@@ -1323,7 +1330,7 @@ struct DeviceGroupedConvBwdWeight_Xdl_CShuffleV3
13231330
}
13241331

13251332
// Gridwise GEMM size
1326-
return GridwiseGemm::CheckValidity(gemm_arg);
1333+
return true;
13271334
}
13281335

13291336
bool IsSupportedArgument(const BaseArgument* p_arg) override

‎include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_conv_v3.hpp‎

Lines changed: 0 additions & 198 deletions
Original file line numberDiff line numberDiff line change
@@ -4,7 +4,6 @@
44
#pragma once
55

66
#include "ck/utility/common_header.hpp"
7-
#include "ck/utility/env.hpp"
87
#include "ck/tensor_description/multi_index_transform_helper.hpp"
98
#include "ck/tensor_description/tensor_descriptor.hpp"
109
#include "ck/tensor_description/tensor_descriptor_helper.hpp"
@@ -607,203 +606,6 @@ struct GridwiseGemm_xdl_cshuffle_conv_v3
607606
c_block_size * sizeof(CShuffleDataType));
608607
}
609608

610-
// block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
611-
__host__ static constexpr bool CheckValidity(const Argument& karg)
612-
{
613-
static_assert((MPerBlock % (MPerXdl * MXdlPerWave) == 0) &&
614-
(NPerBlock % (NXdlPerWave * NPerXdl)) == 0,
615-
"Invalid tuning param!");
616-
617-
if constexpr(!(GemmSpec == tensor_operation::device::GemmSpecialization::MPadding ||
618-
GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding ||
619-
GemmSpec == tensor_operation::device::GemmSpecialization::MKPadding ||
620-
GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding) &&
621-
!(is_same<tensor_layout::gemm::RowMajor, ALayout>::value))
622-
{
623-
if(!(karg.M % MPerBlock == 0))
624-
{
625-
if(ck::EnvIsEnabled(CK_ENV(CK_LOGGING)))
626-
{
627-
std::cout << "Arg M value is not a multiple of MPerBlock! M: " << karg.M << " "
628-
<< __FILE__ << ":" << __LINE__ << ", in function: " << __func__
629-
<< std::endl;
630-
}
631-
return false;
632-
}
633-
}
634-
635-
if constexpr(!(GemmSpec == tensor_operation::device::GemmSpecialization::NPadding ||
636-
GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding ||
637-
GemmSpec == tensor_operation::device::GemmSpecialization::NKPadding ||
638-
GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding) &&
639-
(is_same<tensor_layout::gemm::RowMajor, BLayout>::value))
640-
{
641-
if(!(karg.N % NPerBlock == 0))
642-
{
643-
if(ck::EnvIsEnabled(CK_ENV(CK_LOGGING)))
644-
{
645-
std::cout << "Arg N value is not a multiple of NPerBlock! N: " << karg.N << " "
646-
<< __FILE__ << ":" << __LINE__ << ", in function: " << __func__
647-
<< std::endl;
648-
}
649-
return false;
650-
}
651-
}
652-
653-
if constexpr(!(GemmSpec == tensor_operation::device::GemmSpecialization::KPadding ||
654-
GemmSpec == tensor_operation::device::GemmSpecialization::MKPadding ||
655-
GemmSpec == tensor_operation::device::GemmSpecialization::NKPadding ||
656-
GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding))
657-
{
658-
659-
auto K_t = karg.KBatch * KPerBlock;
660-
if(!(karg.K % K_t == 0))
661-
{
662-
if(ck::EnvIsEnabled(CK_ENV(CK_LOGGING)))
663-
{
664-
std::cout << "Arg K value is not a multiple of K_Batch * K0PerBlock * K1! K: "
665-
<< karg.K << " " << __FILE__ << ":" << __LINE__
666-
<< ", in function: " << __func__ << std::endl;
667-
}
668-
return false;
669-
}
670-
}
671-
else
672-
{
673-
constexpr auto KReadVec = math::lcm(AK1Number, BK1Number);
674-
auto K_t = karg.KBatch * KReadVec;
675-
auto KReadPadSplited = math::integer_divide_ceil(karg.K, K_t) * KReadVec;
676-
if((KReadPadSplited * (karg.KBatch - 1)) >= karg.K)
677-
{
678-
return false;
679-
}
680-
}
681-
682-
if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
683-
{
684-
if(karg.K % ABlockTransferSrcScalarPerVector != 0)
685-
{
686-
if(ck::EnvIsEnabled(CK_ENV(CK_LOGGING)))
687-
{
688-
std::cout << "Arg K (" << karg.K
689-
<< ") value is not a multiple of ABlockTransferSrcScalarPerVector ("
690-
<< ABlockTransferSrcScalarPerVector << " )! " << __FILE__ << ":"
691-
<< __LINE__ << ", in function: " << __func__ << std::endl;
692-
}
693-
return false;
694-
}
695-
}
696-
else
697-
{
698-
if(karg.M % ABlockTransferSrcScalarPerVector != 0)
699-
{
700-
if(ck::EnvIsEnabled(CK_ENV(CK_LOGGING)))
701-
{
702-
std::cout << "Arg M (" << karg.M
703-
<< ") value is not a multiple of ABlockTransferSrcScalarPerVector ("
704-
<< ABlockTransferSrcScalarPerVector << " )! " << __FILE__ << ":"
705-
<< __LINE__ << ", in function: " << __func__ << std::endl;
706-
}
707-
return false;
708-
}
709-
}
710-
711-
if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
712-
{
713-
if(karg.N % BBlockTransferSrcScalarPerVector != 0)
714-
{
715-
if(ck::EnvIsEnabled(CK_ENV(CK_LOGGING)))
716-
{
717-
std::cout << "Arg N (" << karg.N
718-
<< ") value is not a multiple of BBlockTransferSrcScalarPerVector ("
719-
<< BBlockTransferSrcScalarPerVector << " )! " << __FILE__ << ":"
720-
<< __LINE__ << ", in function: " << __func__ << std::endl;
721-
}
722-
return false;
723-
}
724-
}
725-
else
726-
{
727-
if(karg.K % BBlockTransferSrcScalarPerVector != 0)
728-
{
729-
if(ck::EnvIsEnabled(CK_ENV(CK_LOGGING)))
730-
{
731-
std::cout << "Arg K (" << karg.K
732-
<< ") value is not a multiple of BBlockTransferSrcScalarPerVector ("
733-
<< BBlockTransferSrcScalarPerVector << " )! " << __FILE__ << ":"
734-
<< __LINE__ << ", in function: " << __func__ << std::endl;
735-
}
736-
return false;
737-
}
738-
}
739-
740-
if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
741-
{
742-
if(karg.N % CShuffleBlockTransferScalarPerVector_NPerBlock != 0)
743-
{
744-
if(ck::EnvIsEnabled(CK_ENV(CK_LOGGING)))
745-
{
746-
std::cout << "Arg N (" << karg.N
747-
<< ") value is not a multiple of "
748-
"CShuffleBlockTransferScalarPerVector_NPerBlock ("
749-
<< CShuffleBlockTransferScalarPerVector_NPerBlock << " )! "
750-
<< __FILE__ << ":" << __LINE__ << ", in function: " << __func__
751-
<< std::endl;
752-
}
753-
return false;
754-
}
755-
}
756-
else
757-
{
758-
if(karg.M % CShuffleBlockTransferScalarPerVector_NPerBlock != 0)
759-
{
760-
if(ck::EnvIsEnabled(CK_ENV(CK_LOGGING)))
761-
{
762-
std::cout << "Arg M (" << karg.M
763-
<< ") value is not a multiple of "
764-
"CShuffleBlockTransferScalarPerVector_NPerBlock ("
765-
<< CShuffleBlockTransferScalarPerVector_NPerBlock << " )! "
766-
<< __FILE__ << ":" << __LINE__ << ", in function: " << __func__
767-
<< std::endl;
768-
}
769-
return false;
770-
}
771-
}
772-
773-
if constexpr(!(is_same<remove_cvref_t<CDataType>, half_t>::value ||
774-
is_same<remove_cvref_t<CDataType>, float>::value ||
775-
is_same<remove_cvref_t<CDataType>, bhalf_t>::value ||
776-
is_same<remove_cvref_t<CDataType>, int32_t>::value))
777-
{
778-
if(!karg.IsReduceAdd())
779-
{
780-
if(ck::EnvIsEnabled(CK_ENV(CK_LOGGING)))
781-
{
782-
std::cout << " KBatch: " << karg.KBatch << " > 1 is not support yet" << __FILE__
783-
<< ":" << __LINE__ << ", in function: " << __func__ << std::endl;
784-
}
785-
if(karg.KBatch > 1)
786-
{
787-
return false;
788-
}
789-
}
790-
}
791-
792-
// check gridwise gemm pipeline
793-
const auto num_k_loop = karg.AK0 / (KPerBlock / AK1Value);
794-
795-
if constexpr(BlkGemmPipelineVer != BlockGemmPipelineVersion::v1)
796-
{
797-
if(num_k_loop <= BlockwiseGemmPipe::PrefetchStages)
798-
{
799-
return false;
800-
}
801-
}
802-
803-
// TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
804-
return true;
805-
}
806-
807609
__host__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
808610
{
809611
const index_t num_loop = K / KPerBlock;

0 commit comments

Comments
 (0)